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/Function.h"
24 #include "llvm/Intrinsics.h"
25 #include "llvm/ADT/VectorExtras.h"
26 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
27 #include "llvm/CodeGen/MachineFrameInfo.h"
28 #include "llvm/CodeGen/MachineFunction.h"
29 #include "llvm/CodeGen/MachineInstrBuilder.h"
30 #include "llvm/CodeGen/SelectionDAG.h"
31 #include "llvm/CodeGen/SSARegMap.h"
32 #include "llvm/Support/MathExtras.h"
33 #include "llvm/Target/TargetOptions.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/ADT/StringExtras.h"
39 static cl::opt<bool> EnableFastCC("enable-x86-fastcc", cl::Hidden,
40 cl::desc("Enable fastcc on X86"));
41 X86TargetLowering::X86TargetLowering(TargetMachine &TM)
42 : TargetLowering(TM) {
43 Subtarget = &TM.getSubtarget<X86Subtarget>();
44 X86ScalarSSE = Subtarget->hasSSE2();
45 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
47 // Set up the TargetLowering object.
49 // X86 is weird, it always uses i8 for shift amounts and setcc results.
50 setShiftAmountType(MVT::i8);
51 setSetCCResultType(MVT::i8);
52 setSetCCResultContents(ZeroOrOneSetCCResult);
53 setSchedulingPreference(SchedulingForRegPressure);
54 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
55 setStackPointerRegisterToSaveRestore(X86StackPtr);
57 if (Subtarget->isTargetDarwin()) {
58 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
59 setUseUnderscoreSetJmp(false);
60 setUseUnderscoreLongJmp(false);
61 } else if (Subtarget->isTargetMingw()) {
62 // MS runtime is weird: it exports _setjmp, but longjmp!
63 setUseUnderscoreSetJmp(true);
64 setUseUnderscoreLongJmp(false);
66 setUseUnderscoreSetJmp(true);
67 setUseUnderscoreLongJmp(true);
70 // Add legal addressing mode scale values.
71 addLegalAddressScale(8);
72 addLegalAddressScale(4);
73 addLegalAddressScale(2);
74 // Enter the ones which require both scale + index last. These are more
76 addLegalAddressScale(9);
77 addLegalAddressScale(5);
78 addLegalAddressScale(3);
80 // Set up the register classes.
81 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
82 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
83 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
84 if (Subtarget->is64Bit())
85 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
87 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Expand);
89 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
91 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
92 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
93 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
95 if (Subtarget->is64Bit()) {
96 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
97 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
100 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
101 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
103 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
106 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
108 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
109 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
110 // SSE has no i16 to fp conversion, only i32
112 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
114 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
115 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
118 if (!Subtarget->is64Bit()) {
119 // Custom lower SINT_TO_FP and FP_TO_SINT from/to i64 in 32-bit mode.
120 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
121 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
124 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
126 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
127 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
130 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
132 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
133 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
136 // Handle FP_TO_UINT by promoting the destination to a larger signed
138 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
139 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
140 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
142 if (Subtarget->is64Bit()) {
143 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
144 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
146 if (X86ScalarSSE && !Subtarget->hasSSE3())
147 // Expand FP_TO_UINT into a select.
148 // FIXME: We would like to use a Custom expander here eventually to do
149 // the optimal thing for SSE vs. the default expansion in the legalizer.
150 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
152 // With SSE3 we can use fisttpll to convert to a signed i64.
153 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
156 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
158 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
159 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
162 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
163 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
164 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
165 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
166 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
167 if (Subtarget->is64Bit())
168 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Expand);
169 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Expand);
170 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Expand);
171 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
172 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
173 setOperationAction(ISD::FREM , MVT::f64 , Expand);
175 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
176 setOperationAction(ISD::CTTZ , MVT::i8 , Expand);
177 setOperationAction(ISD::CTLZ , MVT::i8 , Expand);
178 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
179 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
180 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
181 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
182 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
183 setOperationAction(ISD::CTLZ , MVT::i32 , Expand);
184 if (Subtarget->is64Bit()) {
185 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
186 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
187 setOperationAction(ISD::CTLZ , MVT::i64 , Expand);
190 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
191 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
193 // These should be promoted to a larger select which is supported.
194 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
195 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
196 // X86 wants to expand cmov itself.
197 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
198 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
199 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
200 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
201 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
202 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
203 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
204 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
205 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
206 if (Subtarget->is64Bit()) {
207 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
208 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
210 // X86 ret instruction may pop stack.
211 setOperationAction(ISD::RET , MVT::Other, Custom);
213 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
214 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
215 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
216 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
217 if (Subtarget->is64Bit()) {
218 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
219 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
220 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
221 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
223 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
224 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
225 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
226 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
227 // X86 wants to expand memset / memcpy itself.
228 setOperationAction(ISD::MEMSET , MVT::Other, Custom);
229 setOperationAction(ISD::MEMCPY , MVT::Other, Custom);
231 // We don't have line number support yet.
232 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
233 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
234 // FIXME - use subtarget debug flags
235 if (!Subtarget->isTargetDarwin() &&
236 !Subtarget->isTargetELF() &&
237 !Subtarget->isTargetCygMing())
238 setOperationAction(ISD::LABEL, MVT::Other, Expand);
240 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
241 setOperationAction(ISD::VASTART , MVT::Other, Custom);
243 // Use the default implementation.
244 setOperationAction(ISD::VAARG , MVT::Other, Expand);
245 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
246 setOperationAction(ISD::VAEND , MVT::Other, Expand);
247 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
248 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
249 if (Subtarget->is64Bit())
250 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
251 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
254 // Set up the FP register classes.
255 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
256 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
258 // Use ANDPD to simulate FABS.
259 setOperationAction(ISD::FABS , MVT::f64, Custom);
260 setOperationAction(ISD::FABS , MVT::f32, Custom);
262 // Use XORP to simulate FNEG.
263 setOperationAction(ISD::FNEG , MVT::f64, Custom);
264 setOperationAction(ISD::FNEG , MVT::f32, Custom);
266 // Use ANDPD and ORPD to simulate FCOPYSIGN.
267 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
268 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
270 // We don't support sin/cos/fmod
271 setOperationAction(ISD::FSIN , MVT::f64, Expand);
272 setOperationAction(ISD::FCOS , MVT::f64, Expand);
273 setOperationAction(ISD::FREM , MVT::f64, Expand);
274 setOperationAction(ISD::FSIN , MVT::f32, Expand);
275 setOperationAction(ISD::FCOS , MVT::f32, Expand);
276 setOperationAction(ISD::FREM , MVT::f32, Expand);
278 // Expand FP immediates into loads from the stack, except for the special
280 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
281 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
282 addLegalFPImmediate(+0.0); // xorps / xorpd
284 // Set up the FP register classes.
285 addRegisterClass(MVT::f64, X86::RFPRegisterClass);
287 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
288 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
289 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
292 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
293 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
296 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
297 addLegalFPImmediate(+0.0); // FLD0
298 addLegalFPImmediate(+1.0); // FLD1
299 addLegalFPImmediate(-0.0); // FLD0/FCHS
300 addLegalFPImmediate(-1.0); // FLD1/FCHS
303 // First set operation action for all vector types to expand. Then we
304 // will selectively turn on ones that can be effectively codegen'd.
305 for (unsigned VT = (unsigned)MVT::Vector + 1;
306 VT != (unsigned)MVT::LAST_VALUETYPE; VT++) {
307 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
308 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
309 setOperationAction(ISD::FADD, (MVT::ValueType)VT, Expand);
310 setOperationAction(ISD::FSUB, (MVT::ValueType)VT, Expand);
311 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
312 setOperationAction(ISD::FMUL, (MVT::ValueType)VT, Expand);
313 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
314 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
315 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
316 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
317 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
318 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand);
319 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Expand);
320 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
321 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
324 if (Subtarget->hasMMX()) {
325 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
326 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
327 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
329 // FIXME: add MMX packed arithmetics
330 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Expand);
331 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Expand);
332 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Expand);
335 if (Subtarget->hasSSE1()) {
336 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
338 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
339 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
340 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
341 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
342 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
343 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
344 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
345 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
346 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
349 if (Subtarget->hasSSE2()) {
350 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
351 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
352 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
353 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
354 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
356 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
357 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
358 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
359 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
360 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
361 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
362 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
363 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
364 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
365 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
366 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
368 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
369 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
370 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
371 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
372 // Implement v4f32 insert_vector_elt in terms of SSE2 v8i16 ones.
373 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
375 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
376 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
377 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
378 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
379 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
381 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
382 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
383 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
384 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
385 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
386 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
388 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
389 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
390 setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote);
391 AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64);
392 setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote);
393 AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64);
394 setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote);
395 AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64);
396 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote);
397 AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64);
398 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
399 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64);
402 // Custom lower v2i64 and v2f64 selects.
403 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
404 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
405 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
406 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
409 // We want to custom lower some of our intrinsics.
410 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
412 // We have target-specific dag combine patterns for the following nodes:
413 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
414 setTargetDAGCombine(ISD::SELECT);
416 computeRegisterProperties();
418 // FIXME: These should be based on subtarget info. Plus, the values should
419 // be smaller when we are in optimizing for size mode.
420 maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
421 maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
422 maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
423 allowUnalignedMemoryAccesses = true; // x86 supports it!
426 //===----------------------------------------------------------------------===//
427 // C & StdCall Calling Convention implementation
428 //===----------------------------------------------------------------------===//
429 // StdCall calling convention seems to be standard for many Windows' API
430 // routines and around. It differs from C calling convention just a little:
431 // callee should clean up the stack, not caller. Symbols should be also
432 // decorated in some fancy way :) It doesn't support any vector arguments.
434 /// AddLiveIn - This helper function adds the specified physical register to the
435 /// MachineFunction as a live in value. It also creates a corresponding virtual
437 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
438 const TargetRegisterClass *RC) {
439 assert(RC->contains(PReg) && "Not the correct regclass!");
440 unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC);
441 MF.addLiveIn(PReg, VReg);
445 /// HowToPassArgument - Returns how an formal argument of the specified type
446 /// should be passed. If it is through stack, returns the size of the stack
447 /// slot; if it is through integer or XMM register, returns the number of
448 /// integer or XMM registers are needed.
450 HowToPassCallArgument(MVT::ValueType ObjectVT,
452 unsigned NumIntRegs, unsigned NumXMMRegs,
453 unsigned MaxNumIntRegs,
454 unsigned &ObjSize, unsigned &ObjIntRegs,
455 unsigned &ObjXMMRegs,
456 bool AllowVectors = true) {
461 if (MaxNumIntRegs>3) {
462 // We don't have too much registers on ia32! :)
467 default: assert(0 && "Unhandled argument type!");
469 if (ArgInReg && (NumIntRegs < MaxNumIntRegs))
475 if (ArgInReg && (NumIntRegs < MaxNumIntRegs))
481 if (ArgInReg && (NumIntRegs < MaxNumIntRegs))
487 if (ArgInReg && (NumIntRegs+2 <= MaxNumIntRegs)) {
489 } else if (ArgInReg && (NumIntRegs+1 <= MaxNumIntRegs)) {
513 assert(0 && "Unhandled argument type [vector]!");
517 SDOperand X86TargetLowering::LowerCCCArguments(SDOperand Op, SelectionDAG &DAG,
519 unsigned NumArgs = Op.Val->getNumValues() - 1;
520 MachineFunction &MF = DAG.getMachineFunction();
521 MachineFrameInfo *MFI = MF.getFrameInfo();
522 SDOperand Root = Op.getOperand(0);
523 SmallVector<SDOperand, 8> ArgValues;
524 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
526 // Add DAG nodes to load the arguments... On entry to a function on the X86,
527 // the stack frame looks like this:
529 // [ESP] -- return address
530 // [ESP + 4] -- first argument (leftmost lexically)
531 // [ESP + 8] -- second argument, if first argument is <= 4 bytes in size
534 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
535 unsigned NumSRetBytes= 0; // How much bytes on stack used for struct return
536 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
537 unsigned NumIntRegs = 0; // Integer regs used for parameter passing
539 static const unsigned XMMArgRegs[] = {
540 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
542 static const unsigned GPRArgRegs[][3] = {
543 { X86::AL, X86::DL, X86::CL },
544 { X86::AX, X86::DX, X86::CX },
545 { X86::EAX, X86::EDX, X86::ECX }
547 static const TargetRegisterClass* GPRClasses[3] = {
548 X86::GR8RegisterClass, X86::GR16RegisterClass, X86::GR32RegisterClass
551 // Handle regparm attribute
552 SmallVector<bool, 8> ArgInRegs(NumArgs, false);
553 SmallVector<bool, 8> SRetArgs(NumArgs, false);
555 for (unsigned i = 0; i<NumArgs; ++i) {
556 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(3+i))->getValue();
557 ArgInRegs[i] = (Flags >> 1) & 1;
558 SRetArgs[i] = (Flags >> 2) & 1;
562 for (unsigned i = 0; i < NumArgs; ++i) {
563 MVT::ValueType ObjectVT = Op.getValue(i).getValueType();
564 unsigned ArgIncrement = 4;
565 unsigned ObjSize = 0;
566 unsigned ObjXMMRegs = 0;
567 unsigned ObjIntRegs = 0;
571 HowToPassCallArgument(ObjectVT,
573 NumIntRegs, NumXMMRegs, 3,
574 ObjSize, ObjIntRegs, ObjXMMRegs,
578 ArgIncrement = ObjSize;
580 if (ObjIntRegs || ObjXMMRegs) {
582 default: assert(0 && "Unhandled argument type!");
586 unsigned RegToUse = GPRArgRegs[ObjectVT-MVT::i8][NumIntRegs];
587 Reg = AddLiveIn(MF, RegToUse, GPRClasses[ObjectVT-MVT::i8]);
588 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
597 assert(!isStdCall && "Unhandled argument type!");
598 Reg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs], X86::VR128RegisterClass);
599 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
602 NumIntRegs += ObjIntRegs;
603 NumXMMRegs += ObjXMMRegs;
606 // XMM arguments have to be aligned on 16-byte boundary.
608 ArgOffset = ((ArgOffset + 15) / 16) * 16;
609 // Create the SelectionDAG nodes corresponding to a load from this
611 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
612 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
613 ArgValue = DAG.getLoad(Op.Val->getValueType(i), Root, FIN, NULL, 0);
615 ArgOffset += ArgIncrement; // Move on to the next argument.
617 NumSRetBytes += ArgIncrement;
620 ArgValues.push_back(ArgValue);
623 ArgValues.push_back(Root);
625 // If the function takes variable number of arguments, make a frame index for
626 // the start of the first vararg value... for expansion of llvm.va_start.
628 VarArgsFrameIndex = MFI->CreateFixedObject(1, ArgOffset);
630 if (isStdCall && !isVarArg) {
631 BytesToPopOnReturn = ArgOffset; // Callee pops everything..
632 BytesCallerReserves = 0;
634 BytesToPopOnReturn = NumSRetBytes; // Callee pops hidden struct pointer.
635 BytesCallerReserves = ArgOffset;
638 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
639 ReturnAddrIndex = 0; // No return address slot generated yet.
642 MF.getInfo<X86FunctionInfo>()->setBytesToPopOnReturn(BytesToPopOnReturn);
644 // Return the new list of results.
645 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
646 &ArgValues[0], ArgValues.size());
649 SDOperand X86TargetLowering::LowerCCCCallTo(SDOperand Op, SelectionDAG &DAG,
651 SDOperand Chain = Op.getOperand(0);
652 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
653 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
654 SDOperand Callee = Op.getOperand(4);
655 MVT::ValueType RetVT= Op.Val->getValueType(0);
656 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
658 static const unsigned XMMArgRegs[] = {
659 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
661 static const unsigned GPR32ArgRegs[] = {
662 X86::EAX, X86::EDX, X86::ECX
665 // Count how many bytes are to be pushed on the stack.
666 unsigned NumBytes = 0;
667 // Keep track of the number of integer regs passed so far.
668 unsigned NumIntRegs = 0;
669 // Keep track of the number of XMM regs passed so far.
670 unsigned NumXMMRegs = 0;
671 // How much bytes on stack used for struct return
672 unsigned NumSRetBytes= 0;
674 // Handle regparm attribute
675 SmallVector<bool, 8> ArgInRegs(NumOps, false);
676 SmallVector<bool, 8> SRetArgs(NumOps, false);
677 for (unsigned i = 0; i<NumOps; ++i) {
679 dyn_cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
680 ArgInRegs[i] = (Flags >> 1) & 1;
681 SRetArgs[i] = (Flags >> 2) & 1;
684 // Calculate stack frame size
685 for (unsigned i = 0; i != NumOps; ++i) {
686 SDOperand Arg = Op.getOperand(5+2*i);
687 unsigned ArgIncrement = 4;
688 unsigned ObjSize = 0;
689 unsigned ObjIntRegs = 0;
690 unsigned ObjXMMRegs = 0;
692 HowToPassCallArgument(Arg.getValueType(),
694 NumIntRegs, NumXMMRegs, 3,
695 ObjSize, ObjIntRegs, ObjXMMRegs,
698 ArgIncrement = ObjSize;
700 NumIntRegs += ObjIntRegs;
701 NumXMMRegs += ObjXMMRegs;
703 // XMM arguments have to be aligned on 16-byte boundary.
705 NumBytes = ((NumBytes + 15) / 16) * 16;
706 NumBytes += ArgIncrement;
710 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
712 // Arguments go on the stack in reverse order, as specified by the ABI.
713 unsigned ArgOffset = 0;
716 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
717 SmallVector<SDOperand, 8> MemOpChains;
718 SDOperand StackPtr = DAG.getRegister(X86StackPtr, getPointerTy());
719 for (unsigned i = 0; i != NumOps; ++i) {
720 SDOperand Arg = Op.getOperand(5+2*i);
721 unsigned ArgIncrement = 4;
722 unsigned ObjSize = 0;
723 unsigned ObjIntRegs = 0;
724 unsigned ObjXMMRegs = 0;
726 HowToPassCallArgument(Arg.getValueType(),
728 NumIntRegs, NumXMMRegs, 3,
729 ObjSize, ObjIntRegs, ObjXMMRegs,
733 ArgIncrement = ObjSize;
735 if (Arg.getValueType() == MVT::i8 || Arg.getValueType() == MVT::i16) {
736 // Promote the integer to 32 bits. If the input type is signed use a
737 // sign extend, otherwise use a zero extend.
738 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
740 unsigned ExtOp = (Flags & 1) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
741 Arg = DAG.getNode(ExtOp, MVT::i32, Arg);
744 if (ObjIntRegs || ObjXMMRegs) {
745 switch (Arg.getValueType()) {
746 default: assert(0 && "Unhandled argument type!");
748 RegsToPass.push_back(std::make_pair(GPR32ArgRegs[NumIntRegs], Arg));
756 assert(!isStdCall && "Unhandled argument type!");
757 RegsToPass.push_back(std::make_pair(XMMArgRegs[NumXMMRegs], Arg));
761 NumIntRegs += ObjIntRegs;
762 NumXMMRegs += ObjXMMRegs;
765 // XMM arguments have to be aligned on 16-byte boundary.
767 ArgOffset = ((ArgOffset + 15) / 16) * 16;
769 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
770 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
771 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
773 ArgOffset += ArgIncrement; // Move on to the next argument.
775 NumSRetBytes += ArgIncrement;
779 // Sanity check: we haven't seen NumSRetBytes > 4
780 assert((NumSRetBytes<=4) &&
781 "Too much space for struct-return pointer requested");
783 if (!MemOpChains.empty())
784 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
785 &MemOpChains[0], MemOpChains.size());
787 // Build a sequence of copy-to-reg nodes chained together with token chain
788 // and flag operands which copy the outgoing args into registers.
790 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
791 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
793 InFlag = Chain.getValue(1);
796 // ELF / PIC requires GOT in the EBX register before function calls via PLT
798 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
799 Subtarget->isPICStyleGOT()) {
800 Chain = DAG.getCopyToReg(Chain, X86::EBX,
801 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
803 InFlag = Chain.getValue(1);
806 // If the callee is a GlobalAddress node (quite common, every direct call is)
807 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
808 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
809 // We should use extra load for direct calls to dllimported functions in
811 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
812 getTargetMachine(), true))
813 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
814 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
815 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
817 // Returns a chain & a flag for retval copy to use.
818 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
819 SmallVector<SDOperand, 8> Ops;
820 Ops.push_back(Chain);
821 Ops.push_back(Callee);
823 // Add argument registers to the end of the list so that they are known live
825 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
826 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
827 RegsToPass[i].second.getValueType()));
829 // Add an implicit use GOT pointer in EBX.
830 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
831 Subtarget->isPICStyleGOT())
832 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
835 Ops.push_back(InFlag);
837 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
838 NodeTys, &Ops[0], Ops.size());
839 InFlag = Chain.getValue(1);
841 // Create the CALLSEQ_END node.
842 unsigned NumBytesForCalleeToPush = 0;
846 NumBytesForCalleeToPush = NumSRetBytes;
848 NumBytesForCalleeToPush = NumBytes;
851 // If this is is a call to a struct-return function, the callee
852 // pops the hidden struct pointer, so we have to push it back.
853 // This is common for Darwin/X86, Linux & Mingw32 targets.
854 NumBytesForCalleeToPush = NumSRetBytes;
857 if (RetVT != MVT::Other)
858 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
860 NodeTys = DAG.getVTList(MVT::Other);
862 Ops.push_back(Chain);
863 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
864 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
865 Ops.push_back(InFlag);
866 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
867 if (RetVT != MVT::Other)
868 InFlag = Chain.getValue(1);
870 SmallVector<SDOperand, 8> ResultVals;
872 default: assert(0 && "Unknown value type to return!");
873 case MVT::Other: break;
875 Chain = DAG.getCopyFromReg(Chain, X86::AL, MVT::i8, InFlag).getValue(1);
876 ResultVals.push_back(Chain.getValue(0));
877 NodeTys = DAG.getVTList(MVT::i8, MVT::Other);
880 Chain = DAG.getCopyFromReg(Chain, X86::AX, MVT::i16, InFlag).getValue(1);
881 ResultVals.push_back(Chain.getValue(0));
882 NodeTys = DAG.getVTList(MVT::i16, MVT::Other);
885 if (Op.Val->getValueType(1) == MVT::i32) {
886 Chain = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag).getValue(1);
887 ResultVals.push_back(Chain.getValue(0));
888 Chain = DAG.getCopyFromReg(Chain, X86::EDX, MVT::i32,
889 Chain.getValue(2)).getValue(1);
890 ResultVals.push_back(Chain.getValue(0));
891 NodeTys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
893 Chain = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag).getValue(1);
894 ResultVals.push_back(Chain.getValue(0));
895 NodeTys = DAG.getVTList(MVT::i32, MVT::Other);
904 assert(!isStdCall && "Unknown value type to return!");
905 Chain = DAG.getCopyFromReg(Chain, X86::XMM0, RetVT, InFlag).getValue(1);
906 ResultVals.push_back(Chain.getValue(0));
907 NodeTys = DAG.getVTList(RetVT, MVT::Other);
911 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
912 SDOperand GROps[] = { Chain, InFlag };
913 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, GROps, 2);
914 Chain = RetVal.getValue(1);
915 InFlag = RetVal.getValue(2);
917 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
918 // shouldn't be necessary except that RFP cannot be live across
919 // multiple blocks. When stackifier is fixed, they can be uncoupled.
920 MachineFunction &MF = DAG.getMachineFunction();
921 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
922 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
923 Tys = DAG.getVTList(MVT::Other);
925 Chain, RetVal, StackSlot, DAG.getValueType(RetVT), InFlag
927 Chain = DAG.getNode(X86ISD::FST, Tys, Ops, 5);
928 RetVal = DAG.getLoad(RetVT, Chain, StackSlot, NULL, 0);
929 Chain = RetVal.getValue(1);
932 if (RetVT == MVT::f32 && !X86ScalarSSE)
933 // FIXME: we would really like to remember that this FP_ROUND
934 // operation is okay to eliminate if we allow excess FP precision.
935 RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal);
936 ResultVals.push_back(RetVal);
937 NodeTys = DAG.getVTList(RetVT, MVT::Other);
942 // If the function returns void, just return the chain.
943 if (ResultVals.empty())
946 // Otherwise, merge everything together with a MERGE_VALUES node.
947 ResultVals.push_back(Chain);
948 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys,
949 &ResultVals[0], ResultVals.size());
950 return Res.getValue(Op.ResNo);
954 //===----------------------------------------------------------------------===//
955 // X86-64 C Calling Convention implementation
956 //===----------------------------------------------------------------------===//
958 /// HowToPassX86_64CCCArgument - Returns how an formal argument of the specified
959 /// type should be passed. If it is through stack, returns the size of the stack
960 /// slot; if it is through integer or XMM register, returns the number of
961 /// integer or XMM registers are needed.
963 HowToPassX86_64CCCArgument(MVT::ValueType ObjectVT,
964 unsigned NumIntRegs, unsigned NumXMMRegs,
965 unsigned &ObjSize, unsigned &ObjIntRegs,
966 unsigned &ObjXMMRegs) {
972 default: assert(0 && "Unhandled argument type!");
982 case MVT::i8: ObjSize = 1; break;
983 case MVT::i16: ObjSize = 2; break;
984 case MVT::i32: ObjSize = 4; break;
985 case MVT::i64: ObjSize = 8; break;
1002 case MVT::f32: ObjSize = 4; break;
1003 case MVT::f64: ObjSize = 8; break;
1009 case MVT::v2f64: ObjSize = 16; break;
1017 X86TargetLowering::LowerX86_64CCCArguments(SDOperand Op, SelectionDAG &DAG) {
1018 unsigned NumArgs = Op.Val->getNumValues() - 1;
1019 MachineFunction &MF = DAG.getMachineFunction();
1020 MachineFrameInfo *MFI = MF.getFrameInfo();
1021 SDOperand Root = Op.getOperand(0);
1022 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1023 SmallVector<SDOperand, 8> ArgValues;
1025 // Add DAG nodes to load the arguments... On entry to a function on the X86,
1026 // the stack frame looks like this:
1028 // [RSP] -- return address
1029 // [RSP + 8] -- first nonreg argument (leftmost lexically)
1030 // [RSP +16] -- second nonreg argument, if 1st argument is <= 8 bytes in size
1033 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
1034 unsigned NumIntRegs = 0; // Int regs used for parameter passing.
1035 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1037 static const unsigned GPR8ArgRegs[] = {
1038 X86::DIL, X86::SIL, X86::DL, X86::CL, X86::R8B, X86::R9B
1040 static const unsigned GPR16ArgRegs[] = {
1041 X86::DI, X86::SI, X86::DX, X86::CX, X86::R8W, X86::R9W
1043 static const unsigned GPR32ArgRegs[] = {
1044 X86::EDI, X86::ESI, X86::EDX, X86::ECX, X86::R8D, X86::R9D
1046 static const unsigned GPR64ArgRegs[] = {
1047 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1049 static const unsigned XMMArgRegs[] = {
1050 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1051 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1054 for (unsigned i = 0; i < NumArgs; ++i) {
1055 MVT::ValueType ObjectVT = Op.getValue(i).getValueType();
1056 unsigned ArgIncrement = 8;
1057 unsigned ObjSize = 0;
1058 unsigned ObjIntRegs = 0;
1059 unsigned ObjXMMRegs = 0;
1061 // FIXME: __int128 and long double support?
1062 HowToPassX86_64CCCArgument(ObjectVT, NumIntRegs, NumXMMRegs,
1063 ObjSize, ObjIntRegs, ObjXMMRegs);
1065 ArgIncrement = ObjSize;
1069 if (ObjIntRegs || ObjXMMRegs) {
1071 default: assert(0 && "Unhandled argument type!");
1076 TargetRegisterClass *RC = NULL;
1080 RC = X86::GR8RegisterClass;
1081 Reg = GPR8ArgRegs[NumIntRegs];
1084 RC = X86::GR16RegisterClass;
1085 Reg = GPR16ArgRegs[NumIntRegs];
1088 RC = X86::GR32RegisterClass;
1089 Reg = GPR32ArgRegs[NumIntRegs];
1092 RC = X86::GR64RegisterClass;
1093 Reg = GPR64ArgRegs[NumIntRegs];
1096 Reg = AddLiveIn(MF, Reg, RC);
1097 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1108 TargetRegisterClass *RC= (ObjectVT == MVT::f32) ?
1109 X86::FR32RegisterClass : ((ObjectVT == MVT::f64) ?
1110 X86::FR64RegisterClass : X86::VR128RegisterClass);
1111 Reg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs], RC);
1112 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1116 NumIntRegs += ObjIntRegs;
1117 NumXMMRegs += ObjXMMRegs;
1118 } else if (ObjSize) {
1119 // XMM arguments have to be aligned on 16-byte boundary.
1121 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1122 // Create the SelectionDAG nodes corresponding to a load from this
1124 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
1125 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1126 ArgValue = DAG.getLoad(Op.Val->getValueType(i), Root, FIN, NULL, 0);
1127 ArgOffset += ArgIncrement; // Move on to the next argument.
1130 ArgValues.push_back(ArgValue);
1133 // If the function takes variable number of arguments, make a frame index for
1134 // the start of the first vararg value... for expansion of llvm.va_start.
1136 // For X86-64, if there are vararg parameters that are passed via
1137 // registers, then we must store them to their spots on the stack so they
1138 // may be loaded by deferencing the result of va_next.
1139 VarArgsGPOffset = NumIntRegs * 8;
1140 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1141 VarArgsFrameIndex = MFI->CreateFixedObject(1, ArgOffset);
1142 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1144 // Store the integer parameter registers.
1145 SmallVector<SDOperand, 8> MemOps;
1146 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1147 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1148 DAG.getConstant(VarArgsGPOffset, getPointerTy()));
1149 for (; NumIntRegs != 6; ++NumIntRegs) {
1150 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1151 X86::GR64RegisterClass);
1152 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1153 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1154 MemOps.push_back(Store);
1155 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1156 DAG.getConstant(8, getPointerTy()));
1159 // Now store the XMM (fp + vector) parameter registers.
1160 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1161 DAG.getConstant(VarArgsFPOffset, getPointerTy()));
1162 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1163 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1164 X86::VR128RegisterClass);
1165 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1166 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1167 MemOps.push_back(Store);
1168 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1169 DAG.getConstant(16, getPointerTy()));
1171 if (!MemOps.empty())
1172 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1173 &MemOps[0], MemOps.size());
1176 ArgValues.push_back(Root);
1178 ReturnAddrIndex = 0; // No return address slot generated yet.
1179 BytesToPopOnReturn = 0; // Callee pops nothing.
1180 BytesCallerReserves = ArgOffset;
1182 // Return the new list of results.
1183 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1184 &ArgValues[0], ArgValues.size());
1188 X86TargetLowering::LowerX86_64CCCCallTo(SDOperand Op, SelectionDAG &DAG) {
1189 SDOperand Chain = Op.getOperand(0);
1190 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1191 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1192 SDOperand Callee = Op.getOperand(4);
1193 MVT::ValueType RetVT= Op.Val->getValueType(0);
1194 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1196 // Count how many bytes are to be pushed on the stack.
1197 unsigned NumBytes = 0;
1198 unsigned NumIntRegs = 0; // Int regs used for parameter passing.
1199 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1201 static const unsigned GPR8ArgRegs[] = {
1202 X86::DIL, X86::SIL, X86::DL, X86::CL, X86::R8B, X86::R9B
1204 static const unsigned GPR16ArgRegs[] = {
1205 X86::DI, X86::SI, X86::DX, X86::CX, X86::R8W, X86::R9W
1207 static const unsigned GPR32ArgRegs[] = {
1208 X86::EDI, X86::ESI, X86::EDX, X86::ECX, X86::R8D, X86::R9D
1210 static const unsigned GPR64ArgRegs[] = {
1211 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1213 static const unsigned XMMArgRegs[] = {
1214 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1215 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1218 for (unsigned i = 0; i != NumOps; ++i) {
1219 SDOperand Arg = Op.getOperand(5+2*i);
1220 MVT::ValueType ArgVT = Arg.getValueType();
1223 default: assert(0 && "Unknown value type!");
1243 else if (ArgVT == MVT::f32 || ArgVT == MVT::f64)
1246 // XMM arguments have to be aligned on 16-byte boundary.
1247 NumBytes = ((NumBytes + 15) / 16) * 16;
1254 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1256 // Arguments go on the stack in reverse order, as specified by the ABI.
1257 unsigned ArgOffset = 0;
1260 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1261 SmallVector<SDOperand, 8> MemOpChains;
1262 SDOperand StackPtr = DAG.getRegister(X86StackPtr, getPointerTy());
1263 for (unsigned i = 0; i != NumOps; ++i) {
1264 SDOperand Arg = Op.getOperand(5+2*i);
1265 MVT::ValueType ArgVT = Arg.getValueType();
1268 default: assert(0 && "Unexpected ValueType for argument!");
1273 if (NumIntRegs < 6) {
1277 case MVT::i8: Reg = GPR8ArgRegs[NumIntRegs]; break;
1278 case MVT::i16: Reg = GPR16ArgRegs[NumIntRegs]; break;
1279 case MVT::i32: Reg = GPR32ArgRegs[NumIntRegs]; break;
1280 case MVT::i64: Reg = GPR64ArgRegs[NumIntRegs]; break;
1282 RegsToPass.push_back(std::make_pair(Reg, Arg));
1285 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1286 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1287 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1299 if (NumXMMRegs < 8) {
1300 RegsToPass.push_back(std::make_pair(XMMArgRegs[NumXMMRegs], Arg));
1303 if (ArgVT != MVT::f32 && ArgVT != MVT::f64) {
1304 // XMM arguments have to be aligned on 16-byte boundary.
1305 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1307 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1308 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1309 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1310 if (ArgVT == MVT::f32 || ArgVT == MVT::f64)
1318 if (!MemOpChains.empty())
1319 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1320 &MemOpChains[0], MemOpChains.size());
1322 // Build a sequence of copy-to-reg nodes chained together with token chain
1323 // and flag operands which copy the outgoing args into registers.
1325 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1326 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1328 InFlag = Chain.getValue(1);
1332 // From AMD64 ABI document:
1333 // For calls that may call functions that use varargs or stdargs
1334 // (prototype-less calls or calls to functions containing ellipsis (...) in
1335 // the declaration) %al is used as hidden argument to specify the number
1336 // of SSE registers used. The contents of %al do not need to match exactly
1337 // the number of registers, but must be an ubound on the number of SSE
1338 // registers used and is in the range 0 - 8 inclusive.
1339 Chain = DAG.getCopyToReg(Chain, X86::AL,
1340 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1341 InFlag = Chain.getValue(1);
1344 // If the callee is a GlobalAddress node (quite common, every direct call is)
1345 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1346 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1347 // We should use extra load for direct calls to dllimported functions in
1349 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1350 getTargetMachine(), true))
1351 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1352 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1353 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1355 // Returns a chain & a flag for retval copy to use.
1356 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1357 SmallVector<SDOperand, 8> Ops;
1358 Ops.push_back(Chain);
1359 Ops.push_back(Callee);
1361 // Add argument registers to the end of the list so that they are known live
1363 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1364 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1365 RegsToPass[i].second.getValueType()));
1368 Ops.push_back(InFlag);
1370 // FIXME: Do not generate X86ISD::TAILCALL for now.
1371 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1372 NodeTys, &Ops[0], Ops.size());
1373 InFlag = Chain.getValue(1);
1375 if (RetVT != MVT::Other)
1376 // Returns a flag for retval copy to use.
1377 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1379 NodeTys = DAG.getVTList(MVT::Other);
1381 Ops.push_back(Chain);
1382 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1383 Ops.push_back(DAG.getConstant(0, getPointerTy()));
1384 Ops.push_back(InFlag);
1385 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1386 if (RetVT != MVT::Other)
1387 InFlag = Chain.getValue(1);
1389 SmallVector<SDOperand, 8> ResultVals;
1391 default: assert(0 && "Unknown value type to return!");
1392 case MVT::Other: break;
1394 Chain = DAG.getCopyFromReg(Chain, X86::AL, MVT::i8, InFlag).getValue(1);
1395 ResultVals.push_back(Chain.getValue(0));
1396 NodeTys = DAG.getVTList(MVT::i8, MVT::Other);
1399 Chain = DAG.getCopyFromReg(Chain, X86::AX, MVT::i16, InFlag).getValue(1);
1400 ResultVals.push_back(Chain.getValue(0));
1401 NodeTys = DAG.getVTList(MVT::i16, MVT::Other);
1404 Chain = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag).getValue(1);
1405 ResultVals.push_back(Chain.getValue(0));
1406 NodeTys = DAG.getVTList(MVT::i32, MVT::Other);
1409 if (Op.Val->getValueType(1) == MVT::i64) {
1410 // FIXME: __int128 support?
1411 Chain = DAG.getCopyFromReg(Chain, X86::RAX, MVT::i64, InFlag).getValue(1);
1412 ResultVals.push_back(Chain.getValue(0));
1413 Chain = DAG.getCopyFromReg(Chain, X86::RDX, MVT::i64,
1414 Chain.getValue(2)).getValue(1);
1415 ResultVals.push_back(Chain.getValue(0));
1416 NodeTys = DAG.getVTList(MVT::i64, MVT::i64, MVT::Other);
1418 Chain = DAG.getCopyFromReg(Chain, X86::RAX, MVT::i64, InFlag).getValue(1);
1419 ResultVals.push_back(Chain.getValue(0));
1420 NodeTys = DAG.getVTList(MVT::i64, MVT::Other);
1431 // FIXME: long double support?
1432 Chain = DAG.getCopyFromReg(Chain, X86::XMM0, RetVT, InFlag).getValue(1);
1433 ResultVals.push_back(Chain.getValue(0));
1434 NodeTys = DAG.getVTList(RetVT, MVT::Other);
1438 // If the function returns void, just return the chain.
1439 if (ResultVals.empty())
1442 // Otherwise, merge everything together with a MERGE_VALUES node.
1443 ResultVals.push_back(Chain);
1444 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys,
1445 &ResultVals[0], ResultVals.size());
1446 return Res.getValue(Op.ResNo);
1449 //===----------------------------------------------------------------------===//
1450 // Fast & FastCall Calling Convention implementation
1451 //===----------------------------------------------------------------------===//
1453 // The X86 'fast' calling convention passes up to two integer arguments in
1454 // registers (an appropriate portion of EAX/EDX), passes arguments in C order,
1455 // and requires that the callee pop its arguments off the stack (allowing proper
1456 // tail calls), and has the same return value conventions as C calling convs.
1458 // This calling convention always arranges for the callee pop value to be 8n+4
1459 // bytes, which is needed for tail recursion elimination and stack alignment
1462 // Note that this can be enhanced in the future to pass fp vals in registers
1463 // (when we have a global fp allocator) and do other tricks.
1465 //===----------------------------------------------------------------------===//
1466 // The X86 'fastcall' calling convention passes up to two integer arguments in
1467 // registers (an appropriate portion of ECX/EDX), passes arguments in C order,
1468 // and requires that the callee pop its arguments off the stack (allowing proper
1469 // tail calls), and has the same return value conventions as C calling convs.
1471 // This calling convention always arranges for the callee pop value to be 8n+4
1472 // bytes, which is needed for tail recursion elimination and stack alignment
1477 X86TargetLowering::LowerFastCCArguments(SDOperand Op, SelectionDAG &DAG,
1479 unsigned NumArgs = Op.Val->getNumValues()-1;
1480 MachineFunction &MF = DAG.getMachineFunction();
1481 MachineFrameInfo *MFI = MF.getFrameInfo();
1482 SDOperand Root = Op.getOperand(0);
1483 SmallVector<SDOperand, 8> ArgValues;
1485 // Add DAG nodes to load the arguments... On entry to a function the stack
1486 // frame looks like this:
1488 // [ESP] -- return address
1489 // [ESP + 4] -- first nonreg argument (leftmost lexically)
1490 // [ESP + 8] -- second nonreg argument, if 1st argument is <= 4 bytes in size
1492 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
1494 // Keep track of the number of integer regs passed so far. This can be either
1495 // 0 (neither EAX/ECX or EDX used), 1 (EAX/ECX is used) or 2 (EAX/ECX and EDX
1497 unsigned NumIntRegs = 0;
1498 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1500 static const unsigned XMMArgRegs[] = {
1501 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1504 static const unsigned GPRArgRegs[][2][2] = {
1505 {{ X86::AL, X86::DL }, { X86::CL, X86::DL }},
1506 {{ X86::AX, X86::DX }, { X86::CX, X86::DX }},
1507 {{ X86::EAX, X86::EDX }, { X86::ECX, X86::EDX }}
1510 static const TargetRegisterClass* GPRClasses[3] = {
1511 X86::GR8RegisterClass, X86::GR16RegisterClass, X86::GR32RegisterClass
1514 unsigned GPRInd = (isFastCall ? 1 : 0);
1515 for (unsigned i = 0; i < NumArgs; ++i) {
1516 MVT::ValueType ObjectVT = Op.getValue(i).getValueType();
1517 unsigned ArgIncrement = 4;
1518 unsigned ObjSize = 0;
1519 unsigned ObjXMMRegs = 0;
1520 unsigned ObjIntRegs = 0;
1524 HowToPassCallArgument(ObjectVT,
1525 true, // Use as much registers as possible
1526 NumIntRegs, NumXMMRegs,
1527 (isFastCall ? 2 : FASTCC_NUM_INT_ARGS_INREGS),
1528 ObjSize, ObjIntRegs, ObjXMMRegs,
1532 ArgIncrement = ObjSize;
1534 if (ObjIntRegs || ObjXMMRegs) {
1536 default: assert(0 && "Unhandled argument type!");
1540 unsigned RegToUse = GPRArgRegs[ObjectVT-MVT::i8][GPRInd][NumIntRegs];
1541 Reg = AddLiveIn(MF, RegToUse, GPRClasses[ObjectVT-MVT::i8]);
1542 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1551 assert(!isFastCall && "Unhandled argument type!");
1552 Reg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs], X86::VR128RegisterClass);
1553 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1557 NumIntRegs += ObjIntRegs;
1558 NumXMMRegs += ObjXMMRegs;
1561 // XMM arguments have to be aligned on 16-byte boundary.
1563 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1564 // Create the SelectionDAG nodes corresponding to a load from this
1566 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
1567 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1568 ArgValue = DAG.getLoad(Op.Val->getValueType(i), Root, FIN, NULL, 0);
1570 ArgOffset += ArgIncrement; // Move on to the next argument.
1573 ArgValues.push_back(ArgValue);
1576 ArgValues.push_back(Root);
1578 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1579 // arguments and the arguments after the retaddr has been pushed are aligned.
1580 if ((ArgOffset & 7) == 0)
1583 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1584 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
1585 ReturnAddrIndex = 0; // No return address slot generated yet.
1586 BytesToPopOnReturn = ArgOffset; // Callee pops all stack arguments.
1587 BytesCallerReserves = 0;
1589 MF.getInfo<X86FunctionInfo>()->setBytesToPopOnReturn(BytesToPopOnReturn);
1591 // Finally, inform the code generator which regs we return values in.
1592 switch (getValueType(MF.getFunction()->getReturnType())) {
1593 default: assert(0 && "Unknown type!");
1594 case MVT::isVoid: break;
1599 MF.addLiveOut(X86::EAX);
1602 MF.addLiveOut(X86::EAX);
1603 MF.addLiveOut(X86::EDX);
1607 MF.addLiveOut(X86::ST0);
1615 assert(!isFastCall && "Unknown result type");
1616 MF.addLiveOut(X86::XMM0);
1620 // Return the new list of results.
1621 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1622 &ArgValues[0], ArgValues.size());
1625 SDOperand X86TargetLowering::LowerFastCCCallTo(SDOperand Op, SelectionDAG &DAG,
1627 SDOperand Chain = Op.getOperand(0);
1628 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1629 SDOperand Callee = Op.getOperand(4);
1630 MVT::ValueType RetVT= Op.Val->getValueType(0);
1631 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1633 // Count how many bytes are to be pushed on the stack.
1634 unsigned NumBytes = 0;
1636 // Keep track of the number of integer regs passed so far. This can be either
1637 // 0 (neither EAX/ECX or EDX used), 1 (EAX/ECX is used) or 2 (EAX/ECX and EDX
1639 unsigned NumIntRegs = 0;
1640 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1642 static const unsigned GPRArgRegs[][2][2] = {
1643 {{ X86::AL, X86::DL }, { X86::CL, X86::DL }},
1644 {{ X86::AX, X86::DX }, { X86::CX, X86::DX }},
1645 {{ X86::EAX, X86::EDX }, { X86::ECX, X86::EDX }}
1647 static const unsigned XMMArgRegs[] = {
1648 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1651 unsigned GPRInd = (isFastCall ? 1 : 0);
1652 for (unsigned i = 0; i != NumOps; ++i) {
1653 SDOperand Arg = Op.getOperand(5+2*i);
1655 switch (Arg.getValueType()) {
1656 default: assert(0 && "Unknown value type!");
1660 unsigned MaxNumIntRegs = (isFastCall ? 2 : FASTCC_NUM_INT_ARGS_INREGS);
1661 if (NumIntRegs < MaxNumIntRegs) {
1678 assert(!isFastCall && "Unknown value type!");
1682 // XMM arguments have to be aligned on 16-byte boundary.
1683 NumBytes = ((NumBytes + 15) / 16) * 16;
1690 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1691 // arguments and the arguments after the retaddr has been pushed are aligned.
1692 if ((NumBytes & 7) == 0)
1695 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1697 // Arguments go on the stack in reverse order, as specified by the ABI.
1698 unsigned ArgOffset = 0;
1700 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1701 SmallVector<SDOperand, 8> MemOpChains;
1702 SDOperand StackPtr = DAG.getRegister(X86StackPtr, getPointerTy());
1703 for (unsigned i = 0; i != NumOps; ++i) {
1704 SDOperand Arg = Op.getOperand(5+2*i);
1706 switch (Arg.getValueType()) {
1707 default: assert(0 && "Unexpected ValueType for argument!");
1711 unsigned MaxNumIntRegs = (isFastCall ? 2 : FASTCC_NUM_INT_ARGS_INREGS);
1712 if (NumIntRegs < MaxNumIntRegs) {
1714 GPRArgRegs[Arg.getValueType()-MVT::i8][GPRInd][NumIntRegs];
1715 RegsToPass.push_back(std::make_pair(RegToUse, Arg));
1721 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1722 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1723 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1728 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1729 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1730 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1740 assert(!isFastCall && "Unexpected ValueType for argument!");
1741 if (NumXMMRegs < 4) {
1742 RegsToPass.push_back(std::make_pair(XMMArgRegs[NumXMMRegs], Arg));
1745 // XMM arguments have to be aligned on 16-byte boundary.
1746 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1747 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1748 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1749 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1756 if (!MemOpChains.empty())
1757 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1758 &MemOpChains[0], MemOpChains.size());
1760 // Build a sequence of copy-to-reg nodes chained together with token chain
1761 // and flag operands which copy the outgoing args into registers.
1763 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1764 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1766 InFlag = Chain.getValue(1);
1769 // If the callee is a GlobalAddress node (quite common, every direct call is)
1770 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1771 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1772 // We should use extra load for direct calls to dllimported functions in
1774 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1775 getTargetMachine(), true))
1776 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1777 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1778 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1780 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1782 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1783 Subtarget->isPICStyleGOT()) {
1784 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1785 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1787 InFlag = Chain.getValue(1);
1790 // Returns a chain & a flag for retval copy to use.
1791 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1792 SmallVector<SDOperand, 8> Ops;
1793 Ops.push_back(Chain);
1794 Ops.push_back(Callee);
1796 // Add argument registers to the end of the list so that they are known live
1798 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1799 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1800 RegsToPass[i].second.getValueType()));
1802 // Add an implicit use GOT pointer in EBX.
1803 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1804 Subtarget->isPICStyleGOT())
1805 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1808 Ops.push_back(InFlag);
1810 // FIXME: Do not generate X86ISD::TAILCALL for now.
1811 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1812 NodeTys, &Ops[0], Ops.size());
1813 InFlag = Chain.getValue(1);
1815 if (RetVT != MVT::Other)
1816 // Returns a flag for retval copy to use.
1817 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1819 NodeTys = DAG.getVTList(MVT::Other);
1821 Ops.push_back(Chain);
1822 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1823 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1824 Ops.push_back(InFlag);
1825 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1826 if (RetVT != MVT::Other)
1827 InFlag = Chain.getValue(1);
1829 SmallVector<SDOperand, 8> ResultVals;
1831 default: assert(0 && "Unknown value type to return!");
1832 case MVT::Other: break;
1834 Chain = DAG.getCopyFromReg(Chain, X86::AL, MVT::i8, InFlag).getValue(1);
1835 ResultVals.push_back(Chain.getValue(0));
1836 NodeTys = DAG.getVTList(MVT::i8, MVT::Other);
1839 Chain = DAG.getCopyFromReg(Chain, X86::AX, MVT::i16, InFlag).getValue(1);
1840 ResultVals.push_back(Chain.getValue(0));
1841 NodeTys = DAG.getVTList(MVT::i16, MVT::Other);
1844 if (Op.Val->getValueType(1) == MVT::i32) {
1845 Chain = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag).getValue(1);
1846 ResultVals.push_back(Chain.getValue(0));
1847 Chain = DAG.getCopyFromReg(Chain, X86::EDX, MVT::i32,
1848 Chain.getValue(2)).getValue(1);
1849 ResultVals.push_back(Chain.getValue(0));
1850 NodeTys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
1852 Chain = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag).getValue(1);
1853 ResultVals.push_back(Chain.getValue(0));
1854 NodeTys = DAG.getVTList(MVT::i32, MVT::Other);
1864 assert(0 && "Unknown value type to return!");
1866 Chain = DAG.getCopyFromReg(Chain, X86::XMM0, RetVT, InFlag).getValue(1);
1867 ResultVals.push_back(Chain.getValue(0));
1868 NodeTys = DAG.getVTList(RetVT, MVT::Other);
1873 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
1874 SmallVector<SDOperand, 8> Ops;
1875 Ops.push_back(Chain);
1876 Ops.push_back(InFlag);
1877 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys,
1878 &Ops[0], Ops.size());
1879 Chain = RetVal.getValue(1);
1880 InFlag = RetVal.getValue(2);
1882 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
1883 // shouldn't be necessary except that RFP cannot be live across
1884 // multiple blocks. When stackifier is fixed, they can be uncoupled.
1885 MachineFunction &MF = DAG.getMachineFunction();
1886 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
1887 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
1888 Tys = DAG.getVTList(MVT::Other);
1890 Ops.push_back(Chain);
1891 Ops.push_back(RetVal);
1892 Ops.push_back(StackSlot);
1893 Ops.push_back(DAG.getValueType(RetVT));
1894 Ops.push_back(InFlag);
1895 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
1896 RetVal = DAG.getLoad(RetVT, Chain, StackSlot, NULL, 0);
1897 Chain = RetVal.getValue(1);
1900 if (RetVT == MVT::f32 && !X86ScalarSSE)
1901 // FIXME: we would really like to remember that this FP_ROUND
1902 // operation is okay to eliminate if we allow excess FP precision.
1903 RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal);
1904 ResultVals.push_back(RetVal);
1905 NodeTys = DAG.getVTList(RetVT, MVT::Other);
1912 // If the function returns void, just return the chain.
1913 if (ResultVals.empty())
1916 // Otherwise, merge everything together with a MERGE_VALUES node.
1917 ResultVals.push_back(Chain);
1918 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys,
1919 &ResultVals[0], ResultVals.size());
1920 return Res.getValue(Op.ResNo);
1923 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1924 if (ReturnAddrIndex == 0) {
1925 // Set up a frame object for the return address.
1926 MachineFunction &MF = DAG.getMachineFunction();
1927 if (Subtarget->is64Bit())
1928 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
1930 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
1933 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1938 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1939 /// specific condition code. It returns a false if it cannot do a direct
1940 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1942 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1943 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
1944 SelectionDAG &DAG) {
1945 X86CC = X86::COND_INVALID;
1947 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1948 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1949 // X > -1 -> X == 0, jump !sign.
1950 RHS = DAG.getConstant(0, RHS.getValueType());
1951 X86CC = X86::COND_NS;
1953 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1954 // X < 0 -> X == 0, jump on sign.
1955 X86CC = X86::COND_S;
1960 switch (SetCCOpcode) {
1962 case ISD::SETEQ: X86CC = X86::COND_E; break;
1963 case ISD::SETGT: X86CC = X86::COND_G; break;
1964 case ISD::SETGE: X86CC = X86::COND_GE; break;
1965 case ISD::SETLT: X86CC = X86::COND_L; break;
1966 case ISD::SETLE: X86CC = X86::COND_LE; break;
1967 case ISD::SETNE: X86CC = X86::COND_NE; break;
1968 case ISD::SETULT: X86CC = X86::COND_B; break;
1969 case ISD::SETUGT: X86CC = X86::COND_A; break;
1970 case ISD::SETULE: X86CC = X86::COND_BE; break;
1971 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1974 // On a floating point condition, the flags are set as follows:
1976 // 0 | 0 | 0 | X > Y
1977 // 0 | 0 | 1 | X < Y
1978 // 1 | 0 | 0 | X == Y
1979 // 1 | 1 | 1 | unordered
1981 switch (SetCCOpcode) {
1984 case ISD::SETEQ: X86CC = X86::COND_E; break;
1985 case ISD::SETOLT: Flip = true; // Fallthrough
1987 case ISD::SETGT: X86CC = X86::COND_A; break;
1988 case ISD::SETOLE: Flip = true; // Fallthrough
1990 case ISD::SETGE: X86CC = X86::COND_AE; break;
1991 case ISD::SETUGT: Flip = true; // Fallthrough
1993 case ISD::SETLT: X86CC = X86::COND_B; break;
1994 case ISD::SETUGE: Flip = true; // Fallthrough
1996 case ISD::SETLE: X86CC = X86::COND_BE; break;
1998 case ISD::SETNE: X86CC = X86::COND_NE; break;
1999 case ISD::SETUO: X86CC = X86::COND_P; break;
2000 case ISD::SETO: X86CC = X86::COND_NP; break;
2003 std::swap(LHS, RHS);
2006 return X86CC != X86::COND_INVALID;
2009 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2010 /// code. Current x86 isa includes the following FP cmov instructions:
2011 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2012 static bool hasFPCMov(unsigned X86CC) {
2028 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2029 /// true if Op is undef or if its value falls within the specified range (L, H].
2030 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
2031 if (Op.getOpcode() == ISD::UNDEF)
2034 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
2035 return (Val >= Low && Val < Hi);
2038 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2039 /// true if Op is undef or if its value equal to the specified value.
2040 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
2041 if (Op.getOpcode() == ISD::UNDEF)
2043 return cast<ConstantSDNode>(Op)->getValue() == Val;
2046 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2047 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2048 bool X86::isPSHUFDMask(SDNode *N) {
2049 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2051 if (N->getNumOperands() != 4)
2054 // Check if the value doesn't reference the second vector.
2055 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2056 SDOperand Arg = N->getOperand(i);
2057 if (Arg.getOpcode() == ISD::UNDEF) continue;
2058 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2059 if (cast<ConstantSDNode>(Arg)->getValue() >= 4)
2066 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2067 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2068 bool X86::isPSHUFHWMask(SDNode *N) {
2069 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2071 if (N->getNumOperands() != 8)
2074 // Lower quadword copied in order.
2075 for (unsigned i = 0; i != 4; ++i) {
2076 SDOperand Arg = N->getOperand(i);
2077 if (Arg.getOpcode() == ISD::UNDEF) continue;
2078 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2079 if (cast<ConstantSDNode>(Arg)->getValue() != i)
2083 // Upper quadword shuffled.
2084 for (unsigned i = 4; i != 8; ++i) {
2085 SDOperand Arg = N->getOperand(i);
2086 if (Arg.getOpcode() == ISD::UNDEF) continue;
2087 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2088 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2089 if (Val < 4 || Val > 7)
2096 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2097 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2098 bool X86::isPSHUFLWMask(SDNode *N) {
2099 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2101 if (N->getNumOperands() != 8)
2104 // Upper quadword copied in order.
2105 for (unsigned i = 4; i != 8; ++i)
2106 if (!isUndefOrEqual(N->getOperand(i), i))
2109 // Lower quadword shuffled.
2110 for (unsigned i = 0; i != 4; ++i)
2111 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2117 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2118 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2119 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
2120 if (NumElems != 2 && NumElems != 4) return false;
2122 unsigned Half = NumElems / 2;
2123 for (unsigned i = 0; i < Half; ++i)
2124 if (!isUndefOrInRange(Elems[i], 0, NumElems))
2126 for (unsigned i = Half; i < NumElems; ++i)
2127 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
2133 bool X86::isSHUFPMask(SDNode *N) {
2134 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2135 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
2138 /// isCommutedSHUFP - Returns true if the shuffle mask is except
2139 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2140 /// half elements to come from vector 1 (which would equal the dest.) and
2141 /// the upper half to come from vector 2.
2142 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
2143 if (NumOps != 2 && NumOps != 4) return false;
2145 unsigned Half = NumOps / 2;
2146 for (unsigned i = 0; i < Half; ++i)
2147 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
2149 for (unsigned i = Half; i < NumOps; ++i)
2150 if (!isUndefOrInRange(Ops[i], 0, NumOps))
2155 static bool isCommutedSHUFP(SDNode *N) {
2156 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2157 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
2160 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2161 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2162 bool X86::isMOVHLPSMask(SDNode *N) {
2163 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2165 if (N->getNumOperands() != 4)
2168 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2169 return isUndefOrEqual(N->getOperand(0), 6) &&
2170 isUndefOrEqual(N->getOperand(1), 7) &&
2171 isUndefOrEqual(N->getOperand(2), 2) &&
2172 isUndefOrEqual(N->getOperand(3), 3);
2175 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2176 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2178 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2179 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2181 if (N->getNumOperands() != 4)
2184 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2185 return isUndefOrEqual(N->getOperand(0), 2) &&
2186 isUndefOrEqual(N->getOperand(1), 3) &&
2187 isUndefOrEqual(N->getOperand(2), 2) &&
2188 isUndefOrEqual(N->getOperand(3), 3);
2191 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2192 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2193 bool X86::isMOVLPMask(SDNode *N) {
2194 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2196 unsigned NumElems = N->getNumOperands();
2197 if (NumElems != 2 && NumElems != 4)
2200 for (unsigned i = 0; i < NumElems/2; ++i)
2201 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2204 for (unsigned i = NumElems/2; i < NumElems; ++i)
2205 if (!isUndefOrEqual(N->getOperand(i), i))
2211 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2212 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2214 bool X86::isMOVHPMask(SDNode *N) {
2215 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2217 unsigned NumElems = N->getNumOperands();
2218 if (NumElems != 2 && NumElems != 4)
2221 for (unsigned i = 0; i < NumElems/2; ++i)
2222 if (!isUndefOrEqual(N->getOperand(i), i))
2225 for (unsigned i = 0; i < NumElems/2; ++i) {
2226 SDOperand Arg = N->getOperand(i + NumElems/2);
2227 if (!isUndefOrEqual(Arg, i + NumElems))
2234 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2235 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2236 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
2237 bool V2IsSplat = false) {
2238 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2241 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2242 SDOperand BitI = Elts[i];
2243 SDOperand BitI1 = Elts[i+1];
2244 if (!isUndefOrEqual(BitI, j))
2247 if (isUndefOrEqual(BitI1, NumElts))
2250 if (!isUndefOrEqual(BitI1, j + NumElts))
2258 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2259 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2260 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2263 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2264 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2265 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
2266 bool V2IsSplat = false) {
2267 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2270 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
2271 SDOperand BitI = Elts[i];
2272 SDOperand BitI1 = Elts[i+1];
2273 if (!isUndefOrEqual(BitI, j + NumElts/2))
2276 if (isUndefOrEqual(BitI1, NumElts))
2279 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2287 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2288 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2289 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
2292 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2293 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2295 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2296 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2298 unsigned NumElems = N->getNumOperands();
2299 if (NumElems != 4 && NumElems != 8 && NumElems != 16)
2302 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2303 SDOperand BitI = N->getOperand(i);
2304 SDOperand BitI1 = N->getOperand(i+1);
2306 if (!isUndefOrEqual(BitI, j))
2308 if (!isUndefOrEqual(BitI1, j))
2315 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2316 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2317 /// MOVSD, and MOVD, i.e. setting the lowest element.
2318 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
2319 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2322 if (!isUndefOrEqual(Elts[0], NumElts))
2325 for (unsigned i = 1; i < NumElts; ++i) {
2326 if (!isUndefOrEqual(Elts[i], i))
2333 bool X86::isMOVLMask(SDNode *N) {
2334 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2335 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
2338 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2339 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2340 /// element of vector 2 and the other elements to come from vector 1 in order.
2341 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
2342 bool V2IsSplat = false,
2343 bool V2IsUndef = false) {
2344 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2347 if (!isUndefOrEqual(Ops[0], 0))
2350 for (unsigned i = 1; i < NumOps; ++i) {
2351 SDOperand Arg = Ops[i];
2352 if (!(isUndefOrEqual(Arg, i+NumOps) ||
2353 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
2354 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
2361 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2362 bool V2IsUndef = false) {
2363 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2364 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
2365 V2IsSplat, V2IsUndef);
2368 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2369 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2370 bool X86::isMOVSHDUPMask(SDNode *N) {
2371 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2373 if (N->getNumOperands() != 4)
2376 // Expect 1, 1, 3, 3
2377 for (unsigned i = 0; i < 2; ++i) {
2378 SDOperand Arg = N->getOperand(i);
2379 if (Arg.getOpcode() == ISD::UNDEF) continue;
2380 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2381 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2382 if (Val != 1) return false;
2386 for (unsigned i = 2; i < 4; ++i) {
2387 SDOperand Arg = N->getOperand(i);
2388 if (Arg.getOpcode() == ISD::UNDEF) continue;
2389 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2390 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2391 if (Val != 3) return false;
2395 // Don't use movshdup if it can be done with a shufps.
2399 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2400 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2401 bool X86::isMOVSLDUPMask(SDNode *N) {
2402 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2404 if (N->getNumOperands() != 4)
2407 // Expect 0, 0, 2, 2
2408 for (unsigned i = 0; i < 2; ++i) {
2409 SDOperand Arg = N->getOperand(i);
2410 if (Arg.getOpcode() == ISD::UNDEF) continue;
2411 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2412 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2413 if (Val != 0) return false;
2417 for (unsigned i = 2; i < 4; ++i) {
2418 SDOperand Arg = N->getOperand(i);
2419 if (Arg.getOpcode() == ISD::UNDEF) continue;
2420 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2421 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2422 if (Val != 2) return false;
2426 // Don't use movshdup if it can be done with a shufps.
2430 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2431 /// a splat of a single element.
2432 static bool isSplatMask(SDNode *N) {
2433 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2435 // This is a splat operation if each element of the permute is the same, and
2436 // if the value doesn't reference the second vector.
2437 unsigned NumElems = N->getNumOperands();
2438 SDOperand ElementBase;
2440 for (; i != NumElems; ++i) {
2441 SDOperand Elt = N->getOperand(i);
2442 if (isa<ConstantSDNode>(Elt)) {
2448 if (!ElementBase.Val)
2451 for (; i != NumElems; ++i) {
2452 SDOperand Arg = N->getOperand(i);
2453 if (Arg.getOpcode() == ISD::UNDEF) continue;
2454 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2455 if (Arg != ElementBase) return false;
2458 // Make sure it is a splat of the first vector operand.
2459 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2462 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2463 /// a splat of a single element and it's a 2 or 4 element mask.
2464 bool X86::isSplatMask(SDNode *N) {
2465 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2467 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2468 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2470 return ::isSplatMask(N);
2473 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2474 /// specifies a splat of zero element.
2475 bool X86::isSplatLoMask(SDNode *N) {
2476 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2478 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2479 if (!isUndefOrEqual(N->getOperand(i), 0))
2484 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2485 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2487 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2488 unsigned NumOperands = N->getNumOperands();
2489 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2491 for (unsigned i = 0; i < NumOperands; ++i) {
2493 SDOperand Arg = N->getOperand(NumOperands-i-1);
2494 if (Arg.getOpcode() != ISD::UNDEF)
2495 Val = cast<ConstantSDNode>(Arg)->getValue();
2496 if (Val >= NumOperands) Val -= NumOperands;
2498 if (i != NumOperands - 1)
2505 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2506 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2508 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2510 // 8 nodes, but we only care about the last 4.
2511 for (unsigned i = 7; i >= 4; --i) {
2513 SDOperand Arg = N->getOperand(i);
2514 if (Arg.getOpcode() != ISD::UNDEF)
2515 Val = cast<ConstantSDNode>(Arg)->getValue();
2524 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2525 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2527 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2529 // 8 nodes, but we only care about the first 4.
2530 for (int i = 3; i >= 0; --i) {
2532 SDOperand Arg = N->getOperand(i);
2533 if (Arg.getOpcode() != ISD::UNDEF)
2534 Val = cast<ConstantSDNode>(Arg)->getValue();
2543 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2544 /// specifies a 8 element shuffle that can be broken into a pair of
2545 /// PSHUFHW and PSHUFLW.
2546 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2547 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2549 if (N->getNumOperands() != 8)
2552 // Lower quadword shuffled.
2553 for (unsigned i = 0; i != 4; ++i) {
2554 SDOperand Arg = N->getOperand(i);
2555 if (Arg.getOpcode() == ISD::UNDEF) continue;
2556 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2557 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2562 // Upper quadword shuffled.
2563 for (unsigned i = 4; i != 8; ++i) {
2564 SDOperand Arg = N->getOperand(i);
2565 if (Arg.getOpcode() == ISD::UNDEF) continue;
2566 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2567 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2568 if (Val < 4 || Val > 7)
2575 /// CommuteVectorShuffle - Swap vector_shuffle operandsas well as
2576 /// values in ther permute mask.
2577 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2578 SDOperand &V2, SDOperand &Mask,
2579 SelectionDAG &DAG) {
2580 MVT::ValueType VT = Op.getValueType();
2581 MVT::ValueType MaskVT = Mask.getValueType();
2582 MVT::ValueType EltVT = MVT::getVectorBaseType(MaskVT);
2583 unsigned NumElems = Mask.getNumOperands();
2584 SmallVector<SDOperand, 8> MaskVec;
2586 for (unsigned i = 0; i != NumElems; ++i) {
2587 SDOperand Arg = Mask.getOperand(i);
2588 if (Arg.getOpcode() == ISD::UNDEF) {
2589 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2592 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2593 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2595 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2597 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2601 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2602 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2605 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2606 /// match movhlps. The lower half elements should come from upper half of
2607 /// V1 (and in order), and the upper half elements should come from the upper
2608 /// half of V2 (and in order).
2609 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2610 unsigned NumElems = Mask->getNumOperands();
2613 for (unsigned i = 0, e = 2; i != e; ++i)
2614 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2616 for (unsigned i = 2; i != 4; ++i)
2617 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2622 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2623 /// is promoted to a vector.
2624 static inline bool isScalarLoadToVector(SDNode *N) {
2625 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2626 N = N->getOperand(0).Val;
2627 return ISD::isNON_EXTLoad(N);
2632 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2633 /// match movlp{s|d}. The lower half elements should come from lower half of
2634 /// V1 (and in order), and the upper half elements should come from the upper
2635 /// half of V2 (and in order). And since V1 will become the source of the
2636 /// MOVLP, it must be either a vector load or a scalar load to vector.
2637 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2638 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2640 // Is V2 is a vector load, don't do this transformation. We will try to use
2641 // load folding shufps op.
2642 if (ISD::isNON_EXTLoad(V2))
2645 unsigned NumElems = Mask->getNumOperands();
2646 if (NumElems != 2 && NumElems != 4)
2648 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2649 if (!isUndefOrEqual(Mask->getOperand(i), i))
2651 for (unsigned i = NumElems/2; i != NumElems; ++i)
2652 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2657 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2659 static bool isSplatVector(SDNode *N) {
2660 if (N->getOpcode() != ISD::BUILD_VECTOR)
2663 SDOperand SplatValue = N->getOperand(0);
2664 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2665 if (N->getOperand(i) != SplatValue)
2670 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2672 static bool isUndefShuffle(SDNode *N) {
2673 if (N->getOpcode() != ISD::BUILD_VECTOR)
2676 SDOperand V1 = N->getOperand(0);
2677 SDOperand V2 = N->getOperand(1);
2678 SDOperand Mask = N->getOperand(2);
2679 unsigned NumElems = Mask.getNumOperands();
2680 for (unsigned i = 0; i != NumElems; ++i) {
2681 SDOperand Arg = Mask.getOperand(i);
2682 if (Arg.getOpcode() != ISD::UNDEF) {
2683 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2684 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2686 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2693 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2694 /// that point to V2 points to its first element.
2695 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2696 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2698 bool Changed = false;
2699 SmallVector<SDOperand, 8> MaskVec;
2700 unsigned NumElems = Mask.getNumOperands();
2701 for (unsigned i = 0; i != NumElems; ++i) {
2702 SDOperand Arg = Mask.getOperand(i);
2703 if (Arg.getOpcode() != ISD::UNDEF) {
2704 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2705 if (Val > NumElems) {
2706 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2710 MaskVec.push_back(Arg);
2714 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2715 &MaskVec[0], MaskVec.size());
2719 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2720 /// operation of specified width.
2721 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2722 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2723 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2725 SmallVector<SDOperand, 8> MaskVec;
2726 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2727 for (unsigned i = 1; i != NumElems; ++i)
2728 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2729 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2732 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2733 /// of specified width.
2734 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2735 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2736 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2737 SmallVector<SDOperand, 8> MaskVec;
2738 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2739 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2740 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2742 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2745 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2746 /// of specified width.
2747 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2748 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2749 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2750 unsigned Half = NumElems/2;
2751 SmallVector<SDOperand, 8> MaskVec;
2752 for (unsigned i = 0; i != Half; ++i) {
2753 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2754 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2756 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2759 /// getZeroVector - Returns a vector of specified type with all zero elements.
2761 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2762 assert(MVT::isVector(VT) && "Expected a vector type");
2763 unsigned NumElems = getVectorNumElements(VT);
2764 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2765 bool isFP = MVT::isFloatingPoint(EVT);
2766 SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT);
2767 SmallVector<SDOperand, 8> ZeroVec(NumElems, Zero);
2768 return DAG.getNode(ISD::BUILD_VECTOR, VT, &ZeroVec[0], ZeroVec.size());
2771 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2773 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2774 SDOperand V1 = Op.getOperand(0);
2775 SDOperand Mask = Op.getOperand(2);
2776 MVT::ValueType VT = Op.getValueType();
2777 unsigned NumElems = Mask.getNumOperands();
2778 Mask = getUnpacklMask(NumElems, DAG);
2779 while (NumElems != 4) {
2780 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2783 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2785 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2786 Mask = getZeroVector(MaskVT, DAG);
2787 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2788 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2789 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2792 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2794 static inline bool isZeroNode(SDOperand Elt) {
2795 return ((isa<ConstantSDNode>(Elt) &&
2796 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2797 (isa<ConstantFPSDNode>(Elt) &&
2798 cast<ConstantFPSDNode>(Elt)->isExactlyValue(0.0)));
2801 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2802 /// vector and zero or undef vector.
2803 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2804 unsigned NumElems, unsigned Idx,
2805 bool isZero, SelectionDAG &DAG) {
2806 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2807 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2808 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2809 SDOperand Zero = DAG.getConstant(0, EVT);
2810 SmallVector<SDOperand, 8> MaskVec(NumElems, Zero);
2811 MaskVec[Idx] = DAG.getConstant(NumElems, EVT);
2812 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2813 &MaskVec[0], MaskVec.size());
2814 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2817 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2819 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
2820 unsigned NumNonZero, unsigned NumZero,
2821 SelectionDAG &DAG, TargetLowering &TLI) {
2827 for (unsigned i = 0; i < 16; ++i) {
2828 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
2829 if (ThisIsNonZero && First) {
2831 V = getZeroVector(MVT::v8i16, DAG);
2833 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2838 SDOperand ThisElt(0, 0), LastElt(0, 0);
2839 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
2840 if (LastIsNonZero) {
2841 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
2843 if (ThisIsNonZero) {
2844 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
2845 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
2846 ThisElt, DAG.getConstant(8, MVT::i8));
2848 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
2853 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
2854 DAG.getConstant(i/2, TLI.getPointerTy()));
2858 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
2861 /// LowerBuildVectorv16i8 - Custom lower build_vector of v8i16.
2863 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
2864 unsigned NumNonZero, unsigned NumZero,
2865 SelectionDAG &DAG, TargetLowering &TLI) {
2871 for (unsigned i = 0; i < 8; ++i) {
2872 bool isNonZero = (NonZeros & (1 << i)) != 0;
2876 V = getZeroVector(MVT::v8i16, DAG);
2878 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2881 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
2882 DAG.getConstant(i, TLI.getPointerTy()));
2890 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2891 // All zero's are handled with pxor.
2892 if (ISD::isBuildVectorAllZeros(Op.Val))
2895 // All one's are handled with pcmpeqd.
2896 if (ISD::isBuildVectorAllOnes(Op.Val))
2899 MVT::ValueType VT = Op.getValueType();
2900 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2901 unsigned EVTBits = MVT::getSizeInBits(EVT);
2903 unsigned NumElems = Op.getNumOperands();
2904 unsigned NumZero = 0;
2905 unsigned NumNonZero = 0;
2906 unsigned NonZeros = 0;
2907 std::set<SDOperand> Values;
2908 for (unsigned i = 0; i < NumElems; ++i) {
2909 SDOperand Elt = Op.getOperand(i);
2910 if (Elt.getOpcode() != ISD::UNDEF) {
2912 if (isZeroNode(Elt))
2915 NonZeros |= (1 << i);
2921 if (NumNonZero == 0)
2922 // Must be a mix of zero and undef. Return a zero vector.
2923 return getZeroVector(VT, DAG);
2925 // Splat is obviously ok. Let legalizer expand it to a shuffle.
2926 if (Values.size() == 1)
2929 // Special case for single non-zero element.
2930 if (NumNonZero == 1) {
2931 unsigned Idx = CountTrailingZeros_32(NonZeros);
2932 SDOperand Item = Op.getOperand(Idx);
2933 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
2935 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
2936 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
2939 if (EVTBits == 32) {
2940 // Turn it into a shuffle of zero and zero-extended scalar to vector.
2941 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
2943 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2944 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
2945 SmallVector<SDOperand, 8> MaskVec;
2946 for (unsigned i = 0; i < NumElems; i++)
2947 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
2948 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2949 &MaskVec[0], MaskVec.size());
2950 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
2951 DAG.getNode(ISD::UNDEF, VT), Mask);
2955 // Let legalizer expand 2-wide build_vector's.
2959 // If element VT is < 32 bits, convert it to inserts into a zero vector.
2961 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
2963 if (V.Val) return V;
2966 if (EVTBits == 16) {
2967 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
2969 if (V.Val) return V;
2972 // If element VT is == 32 bits, turn it into a number of shuffles.
2973 SmallVector<SDOperand, 8> V;
2975 if (NumElems == 4 && NumZero > 0) {
2976 for (unsigned i = 0; i < 4; ++i) {
2977 bool isZero = !(NonZeros & (1 << i));
2979 V[i] = getZeroVector(VT, DAG);
2981 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2984 for (unsigned i = 0; i < 2; ++i) {
2985 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
2988 V[i] = V[i*2]; // Must be a zero vector.
2991 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
2992 getMOVLMask(NumElems, DAG));
2995 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2996 getMOVLMask(NumElems, DAG));
2999 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3000 getUnpacklMask(NumElems, DAG));
3005 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
3006 // clears the upper bits.
3007 // FIXME: we can do the same for v4f32 case when we know both parts of
3008 // the lower half come from scalar_to_vector (loadf32). We should do
3009 // that in post legalizer dag combiner with target specific hooks.
3010 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
3012 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3013 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
3014 SmallVector<SDOperand, 8> MaskVec;
3015 bool Reverse = (NonZeros & 0x3) == 2;
3016 for (unsigned i = 0; i < 2; ++i)
3018 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3020 MaskVec.push_back(DAG.getConstant(i, EVT));
3021 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3022 for (unsigned i = 0; i < 2; ++i)
3024 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3026 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3027 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3028 &MaskVec[0], MaskVec.size());
3029 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3032 if (Values.size() > 2) {
3033 // Expand into a number of unpckl*.
3035 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3036 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3037 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3038 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
3039 for (unsigned i = 0; i < NumElems; ++i)
3040 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3042 while (NumElems != 0) {
3043 for (unsigned i = 0; i < NumElems; ++i)
3044 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3055 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
3056 SDOperand V1 = Op.getOperand(0);
3057 SDOperand V2 = Op.getOperand(1);
3058 SDOperand PermMask = Op.getOperand(2);
3059 MVT::ValueType VT = Op.getValueType();
3060 unsigned NumElems = PermMask.getNumOperands();
3061 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3062 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3063 bool V1IsSplat = false;
3064 bool V2IsSplat = false;
3066 if (isUndefShuffle(Op.Val))
3067 return DAG.getNode(ISD::UNDEF, VT);
3069 if (isSplatMask(PermMask.Val)) {
3070 if (NumElems <= 4) return Op;
3071 // Promote it to a v4i32 splat.
3072 return PromoteSplat(Op, DAG);
3075 if (X86::isMOVLMask(PermMask.Val))
3076 return (V1IsUndef) ? V2 : Op;
3078 if (X86::isMOVSHDUPMask(PermMask.Val) ||
3079 X86::isMOVSLDUPMask(PermMask.Val) ||
3080 X86::isMOVHLPSMask(PermMask.Val) ||
3081 X86::isMOVHPMask(PermMask.Val) ||
3082 X86::isMOVLPMask(PermMask.Val))
3085 if (ShouldXformToMOVHLPS(PermMask.Val) ||
3086 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
3087 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3089 bool Commuted = false;
3090 V1IsSplat = isSplatVector(V1.Val);
3091 V2IsSplat = isSplatVector(V2.Val);
3092 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
3093 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3094 std::swap(V1IsSplat, V2IsSplat);
3095 std::swap(V1IsUndef, V2IsUndef);
3099 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
3100 if (V2IsUndef) return V1;
3101 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3103 // V2 is a splat, so the mask may be malformed. That is, it may point
3104 // to any V2 element. The instruction selectior won't like this. Get
3105 // a corrected mask and commute to form a proper MOVS{S|D}.
3106 SDOperand NewMask = getMOVLMask(NumElems, DAG);
3107 if (NewMask.Val != PermMask.Val)
3108 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3113 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3114 X86::isUNPCKLMask(PermMask.Val) ||
3115 X86::isUNPCKHMask(PermMask.Val))
3119 // Normalize mask so all entries that point to V2 points to its first
3120 // element then try to match unpck{h|l} again. If match, return a
3121 // new vector_shuffle with the corrected mask.
3122 SDOperand NewMask = NormalizeMask(PermMask, DAG);
3123 if (NewMask.Val != PermMask.Val) {
3124 if (X86::isUNPCKLMask(PermMask.Val, true)) {
3125 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
3126 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3127 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
3128 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
3129 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3134 // Normalize the node to match x86 shuffle ops if needed
3135 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
3136 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3139 // Commute is back and try unpck* again.
3140 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3141 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3142 X86::isUNPCKLMask(PermMask.Val) ||
3143 X86::isUNPCKHMask(PermMask.Val))
3147 // If VT is integer, try PSHUF* first, then SHUFP*.
3148 if (MVT::isInteger(VT)) {
3149 if (X86::isPSHUFDMask(PermMask.Val) ||
3150 X86::isPSHUFHWMask(PermMask.Val) ||
3151 X86::isPSHUFLWMask(PermMask.Val)) {
3152 if (V2.getOpcode() != ISD::UNDEF)
3153 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3154 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3158 if (X86::isSHUFPMask(PermMask.Val))
3161 // Handle v8i16 shuffle high / low shuffle node pair.
3162 if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
3163 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3164 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
3165 SmallVector<SDOperand, 8> MaskVec;
3166 for (unsigned i = 0; i != 4; ++i)
3167 MaskVec.push_back(PermMask.getOperand(i));
3168 for (unsigned i = 4; i != 8; ++i)
3169 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3170 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3171 &MaskVec[0], MaskVec.size());
3172 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3174 for (unsigned i = 0; i != 4; ++i)
3175 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3176 for (unsigned i = 4; i != 8; ++i)
3177 MaskVec.push_back(PermMask.getOperand(i));
3178 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0],MaskVec.size());
3179 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3182 // Floating point cases in the other order.
3183 if (X86::isSHUFPMask(PermMask.Val))
3185 if (X86::isPSHUFDMask(PermMask.Val) ||
3186 X86::isPSHUFHWMask(PermMask.Val) ||
3187 X86::isPSHUFLWMask(PermMask.Val)) {
3188 if (V2.getOpcode() != ISD::UNDEF)
3189 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3190 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3195 if (NumElems == 4) {
3196 MVT::ValueType MaskVT = PermMask.getValueType();
3197 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
3198 SmallVector<std::pair<int, int>, 8> Locs;
3199 Locs.reserve(NumElems);
3200 SmallVector<SDOperand, 8> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3201 SmallVector<SDOperand, 8> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3204 // If no more than two elements come from either vector. This can be
3205 // implemented with two shuffles. First shuffle gather the elements.
3206 // The second shuffle, which takes the first shuffle as both of its
3207 // vector operands, put the elements into the right order.
3208 for (unsigned i = 0; i != NumElems; ++i) {
3209 SDOperand Elt = PermMask.getOperand(i);
3210 if (Elt.getOpcode() == ISD::UNDEF) {
3211 Locs[i] = std::make_pair(-1, -1);
3213 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3214 if (Val < NumElems) {
3215 Locs[i] = std::make_pair(0, NumLo);
3219 Locs[i] = std::make_pair(1, NumHi);
3220 if (2+NumHi < NumElems)
3221 Mask1[2+NumHi] = Elt;
3226 if (NumLo <= 2 && NumHi <= 2) {
3227 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3228 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3229 &Mask1[0], Mask1.size()));
3230 for (unsigned i = 0; i != NumElems; ++i) {
3231 if (Locs[i].first == -1)
3234 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
3235 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
3236 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3240 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3241 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3242 &Mask2[0], Mask2.size()));
3245 // Break it into (shuffle shuffle_hi, shuffle_lo).
3247 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3248 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3249 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
3250 unsigned MaskIdx = 0;
3252 unsigned HiIdx = NumElems/2;
3253 for (unsigned i = 0; i != NumElems; ++i) {
3254 if (i == NumElems/2) {
3260 SDOperand Elt = PermMask.getOperand(i);
3261 if (Elt.getOpcode() == ISD::UNDEF) {
3262 Locs[i] = std::make_pair(-1, -1);
3263 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
3264 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3265 (*MaskPtr)[LoIdx] = Elt;
3268 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3269 (*MaskPtr)[HiIdx] = Elt;
3274 SDOperand LoShuffle =
3275 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3276 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3277 &LoMask[0], LoMask.size()));
3278 SDOperand HiShuffle =
3279 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3280 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3281 &HiMask[0], HiMask.size()));
3282 SmallVector<SDOperand, 8> MaskOps;
3283 for (unsigned i = 0; i != NumElems; ++i) {
3284 if (Locs[i].first == -1) {
3285 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3287 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
3288 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3291 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3292 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3293 &MaskOps[0], MaskOps.size()));
3300 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3301 if (!isa<ConstantSDNode>(Op.getOperand(1)))
3304 MVT::ValueType VT = Op.getValueType();
3305 // TODO: handle v16i8.
3306 if (MVT::getSizeInBits(VT) == 16) {
3307 // Transform it so it match pextrw which produces a 32-bit result.
3308 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
3309 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
3310 Op.getOperand(0), Op.getOperand(1));
3311 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
3312 DAG.getValueType(VT));
3313 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3314 } else if (MVT::getSizeInBits(VT) == 32) {
3315 SDOperand Vec = Op.getOperand(0);
3316 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3319 // SHUFPS the element to the lowest double word, then movss.
3320 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3321 SmallVector<SDOperand, 8> IdxVec;
3322 IdxVec.push_back(DAG.getConstant(Idx, MVT::getVectorBaseType(MaskVT)));
3323 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3324 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3325 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3326 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3327 &IdxVec[0], IdxVec.size());
3328 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3329 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3330 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3331 DAG.getConstant(0, getPointerTy()));
3332 } else if (MVT::getSizeInBits(VT) == 64) {
3333 SDOperand Vec = Op.getOperand(0);
3334 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3338 // UNPCKHPD the element to the lowest double word, then movsd.
3339 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
3340 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
3341 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3342 SmallVector<SDOperand, 8> IdxVec;
3343 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorBaseType(MaskVT)));
3344 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3345 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3346 &IdxVec[0], IdxVec.size());
3347 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3348 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3349 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3350 DAG.getConstant(0, getPointerTy()));
3357 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3358 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
3359 // as its second argument.
3360 MVT::ValueType VT = Op.getValueType();
3361 MVT::ValueType BaseVT = MVT::getVectorBaseType(VT);
3362 SDOperand N0 = Op.getOperand(0);
3363 SDOperand N1 = Op.getOperand(1);
3364 SDOperand N2 = Op.getOperand(2);
3365 if (MVT::getSizeInBits(BaseVT) == 16) {
3366 if (N1.getValueType() != MVT::i32)
3367 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
3368 if (N2.getValueType() != MVT::i32)
3369 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(), MVT::i32);
3370 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
3371 } else if (MVT::getSizeInBits(BaseVT) == 32) {
3372 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
3375 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
3376 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3377 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
3378 SmallVector<SDOperand, 8> MaskVec;
3379 MaskVec.push_back(DAG.getConstant(4, BaseVT));
3380 for (unsigned i = 1; i <= 3; ++i)
3381 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3382 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
3383 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3384 &MaskVec[0], MaskVec.size()));
3386 // Use two pinsrw instructions to insert a 32 bit value.
3388 if (MVT::isFloatingPoint(N1.getValueType())) {
3389 if (ISD::isNON_EXTLoad(N1.Val)) {
3390 // Just load directly from f32mem to GR32.
3391 LoadSDNode *LD = cast<LoadSDNode>(N1);
3392 N1 = DAG.getLoad(MVT::i32, LD->getChain(), LD->getBasePtr(),
3393 LD->getSrcValue(), LD->getSrcValueOffset());
3395 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
3396 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
3397 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
3398 DAG.getConstant(0, getPointerTy()));
3401 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
3402 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3403 DAG.getConstant(Idx, getPointerTy()));
3404 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
3405 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3406 DAG.getConstant(Idx+1, getPointerTy()));
3407 return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
3415 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3416 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
3417 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
3420 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
3421 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
3422 // one of the above mentioned nodes. It has to be wrapped because otherwise
3423 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
3424 // be used to form addressing mode. These wrapped nodes will be selected
3427 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
3428 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3429 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
3431 CP->getAlignment());
3432 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3433 // With PIC, the address is actually $g + Offset.
3434 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3435 !Subtarget->isPICStyleRIPRel()) {
3436 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3437 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3445 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
3446 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3447 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
3448 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3449 // With PIC, the address is actually $g + Offset.
3450 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3451 !Subtarget->isPICStyleRIPRel()) {
3452 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3453 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3457 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
3458 // load the value at address GV, not the value of GV itself. This means that
3459 // the GlobalAddress must be in the base or index register of the address, not
3460 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
3461 // The same applies for external symbols during PIC codegen
3462 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
3463 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
3469 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
3470 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
3471 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
3472 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3473 // With PIC, the address is actually $g + Offset.
3474 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3475 !Subtarget->isPICStyleRIPRel()) {
3476 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3477 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3484 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
3485 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3486 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
3487 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3488 // With PIC, the address is actually $g + Offset.
3489 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3490 !Subtarget->isPICStyleRIPRel()) {
3491 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3492 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3499 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
3500 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
3501 "Not an i64 shift!");
3502 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
3503 SDOperand ShOpLo = Op.getOperand(0);
3504 SDOperand ShOpHi = Op.getOperand(1);
3505 SDOperand ShAmt = Op.getOperand(2);
3506 SDOperand Tmp1 = isSRA ?
3507 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
3508 DAG.getConstant(0, MVT::i32);
3510 SDOperand Tmp2, Tmp3;
3511 if (Op.getOpcode() == ISD::SHL_PARTS) {
3512 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
3513 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
3515 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
3516 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
3519 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3520 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
3521 DAG.getConstant(32, MVT::i8));
3522 SDOperand COps[]={DAG.getEntryNode(), AndNode, DAG.getConstant(0, MVT::i8)};
3523 SDOperand InFlag = DAG.getNode(X86ISD::CMP, VTs, 2, COps, 3).getValue(1);
3526 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3528 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
3529 SmallVector<SDOperand, 4> Ops;
3530 if (Op.getOpcode() == ISD::SHL_PARTS) {
3531 Ops.push_back(Tmp2);
3532 Ops.push_back(Tmp3);
3534 Ops.push_back(InFlag);
3535 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3536 InFlag = Hi.getValue(1);
3539 Ops.push_back(Tmp3);
3540 Ops.push_back(Tmp1);
3542 Ops.push_back(InFlag);
3543 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3545 Ops.push_back(Tmp2);
3546 Ops.push_back(Tmp3);
3548 Ops.push_back(InFlag);
3549 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3550 InFlag = Lo.getValue(1);
3553 Ops.push_back(Tmp3);
3554 Ops.push_back(Tmp1);
3556 Ops.push_back(InFlag);
3557 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3560 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
3564 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
3567 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
3568 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
3569 Op.getOperand(0).getValueType() >= MVT::i16 &&
3570 "Unknown SINT_TO_FP to lower!");
3573 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
3574 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
3575 MachineFunction &MF = DAG.getMachineFunction();
3576 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
3577 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3578 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
3579 StackSlot, NULL, 0);
3584 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
3586 Tys = DAG.getVTList(MVT::f64, MVT::Other);
3587 SmallVector<SDOperand, 8> Ops;
3588 Ops.push_back(Chain);
3589 Ops.push_back(StackSlot);
3590 Ops.push_back(DAG.getValueType(SrcVT));
3591 Result = DAG.getNode(X86ScalarSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
3592 Tys, &Ops[0], Ops.size());
3595 Chain = Result.getValue(1);
3596 SDOperand InFlag = Result.getValue(2);
3598 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
3599 // shouldn't be necessary except that RFP cannot be live across
3600 // multiple blocks. When stackifier is fixed, they can be uncoupled.
3601 MachineFunction &MF = DAG.getMachineFunction();
3602 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3603 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3604 Tys = DAG.getVTList(MVT::Other);
3605 SmallVector<SDOperand, 8> Ops;
3606 Ops.push_back(Chain);
3607 Ops.push_back(Result);
3608 Ops.push_back(StackSlot);
3609 Ops.push_back(DAG.getValueType(Op.getValueType()));
3610 Ops.push_back(InFlag);
3611 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
3612 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
3618 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
3619 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
3620 "Unknown FP_TO_SINT to lower!");
3621 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
3623 MachineFunction &MF = DAG.getMachineFunction();
3624 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
3625 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3626 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3629 switch (Op.getValueType()) {
3630 default: assert(0 && "Invalid FP_TO_SINT to lower!");
3631 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
3632 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
3633 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
3636 SDOperand Chain = DAG.getEntryNode();
3637 SDOperand Value = Op.getOperand(0);
3639 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
3640 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
3641 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
3643 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
3645 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
3646 Chain = Value.getValue(1);
3647 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3648 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3651 // Build the FP_TO_INT*_IN_MEM
3652 SDOperand Ops[] = { Chain, Value, StackSlot };
3653 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
3656 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
3659 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
3660 MVT::ValueType VT = Op.getValueType();
3661 const Type *OpNTy = MVT::getTypeForValueType(VT);
3662 std::vector<Constant*> CV;
3663 if (VT == MVT::f64) {
3664 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(~(1ULL << 63))));
3665 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3667 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(~(1U << 31))));
3668 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3669 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3670 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3672 Constant *CS = ConstantStruct::get(CV);
3673 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3674 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3675 SmallVector<SDOperand, 3> Ops;
3676 Ops.push_back(DAG.getEntryNode());
3677 Ops.push_back(CPIdx);
3678 Ops.push_back(DAG.getSrcValue(NULL));
3679 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3680 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
3683 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
3684 MVT::ValueType VT = Op.getValueType();
3685 const Type *OpNTy = MVT::getTypeForValueType(VT);
3686 std::vector<Constant*> CV;
3687 if (VT == MVT::f64) {
3688 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(1ULL << 63)));
3689 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3691 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(1U << 31)));
3692 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3693 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3694 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3696 Constant *CS = ConstantStruct::get(CV);
3697 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3698 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3699 SmallVector<SDOperand, 3> Ops;
3700 Ops.push_back(DAG.getEntryNode());
3701 Ops.push_back(CPIdx);
3702 Ops.push_back(DAG.getSrcValue(NULL));
3703 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3704 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
3707 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
3708 SDOperand Op0 = Op.getOperand(0);
3709 SDOperand Op1 = Op.getOperand(1);
3710 MVT::ValueType VT = Op.getValueType();
3711 MVT::ValueType SrcVT = Op1.getValueType();
3712 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
3714 // If second operand is smaller, extend it first.
3715 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
3716 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
3720 // First get the sign bit of second operand.
3721 std::vector<Constant*> CV;
3722 if (SrcVT == MVT::f64) {
3723 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(1ULL << 63)));
3724 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3726 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(1U << 31)));
3727 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3728 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3729 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3731 Constant *CS = ConstantStruct::get(CV);
3732 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3733 SDVTList Tys = DAG.getVTList(SrcVT, MVT::Other);
3734 SmallVector<SDOperand, 3> Ops;
3735 Ops.push_back(DAG.getEntryNode());
3736 Ops.push_back(CPIdx);
3737 Ops.push_back(DAG.getSrcValue(NULL));
3738 SDOperand Mask1 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3739 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
3741 // Shift sign bit right or left if the two operands have different types.
3742 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
3743 // Op0 is MVT::f32, Op1 is MVT::f64.
3744 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
3745 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
3746 DAG.getConstant(32, MVT::i32));
3747 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
3748 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
3749 DAG.getConstant(0, getPointerTy()));
3752 // Clear first operand sign bit.
3754 if (VT == MVT::f64) {
3755 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(~(1ULL << 63))));
3756 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3758 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(~(1U << 31))));
3759 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3760 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3761 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3763 CS = ConstantStruct::get(CV);
3764 CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3765 Tys = DAG.getVTList(VT, MVT::Other);
3767 Ops.push_back(DAG.getEntryNode());
3768 Ops.push_back(CPIdx);
3769 Ops.push_back(DAG.getSrcValue(NULL));
3770 SDOperand Mask2 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3771 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
3773 // Or the value with the sign bit.
3774 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
3777 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG,
3779 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
3781 SDOperand Op0 = Op.getOperand(0);
3782 SDOperand Op1 = Op.getOperand(1);
3783 SDOperand CC = Op.getOperand(2);
3784 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3785 const MVT::ValueType *VTs1 = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3786 const MVT::ValueType *VTs2 = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
3787 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
3790 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
3792 SDOperand Ops1[] = { Chain, Op0, Op1 };
3793 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, Ops1, 3).getValue(1);
3794 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
3795 return DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3798 assert(isFP && "Illegal integer SetCC!");
3800 SDOperand COps[] = { Chain, Op0, Op1 };
3801 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, COps, 3).getValue(1);
3803 switch (SetCCOpcode) {
3804 default: assert(false && "Illegal floating point SetCC!");
3805 case ISD::SETOEQ: { // !PF & ZF
3806 SDOperand Ops1[] = { DAG.getConstant(X86::COND_NP, MVT::i8), Cond };
3807 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3808 SDOperand Ops2[] = { DAG.getConstant(X86::COND_E, MVT::i8),
3810 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3811 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
3813 case ISD::SETUNE: { // PF | !ZF
3814 SDOperand Ops1[] = { DAG.getConstant(X86::COND_P, MVT::i8), Cond };
3815 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3816 SDOperand Ops2[] = { DAG.getConstant(X86::COND_NE, MVT::i8),
3818 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3819 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
3824 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
3825 bool addTest = true;
3826 SDOperand Chain = DAG.getEntryNode();
3827 SDOperand Cond = Op.getOperand(0);
3829 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3831 if (Cond.getOpcode() == ISD::SETCC)
3832 Cond = LowerSETCC(Cond, DAG, Chain);
3834 if (Cond.getOpcode() == X86ISD::SETCC) {
3835 CC = Cond.getOperand(0);
3837 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3838 // (since flag operand cannot be shared). Use it as the condition setting
3839 // operand in place of the X86ISD::SETCC.
3840 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3841 // to use a test instead of duplicating the X86ISD::CMP (for register
3842 // pressure reason)?
3843 SDOperand Cmp = Cond.getOperand(1);
3844 unsigned Opc = Cmp.getOpcode();
3845 bool IllegalFPCMov = !X86ScalarSSE &&
3846 MVT::isFloatingPoint(Op.getValueType()) &&
3847 !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
3848 if ((Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) &&
3850 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3851 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3857 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3858 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3859 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3862 VTs = DAG.getNodeValueTypes(Op.getValueType(), MVT::Flag);
3863 SmallVector<SDOperand, 4> Ops;
3864 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
3865 // condition is true.
3866 Ops.push_back(Op.getOperand(2));
3867 Ops.push_back(Op.getOperand(1));
3869 Ops.push_back(Cond.getValue(1));
3870 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3873 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
3874 bool addTest = true;
3875 SDOperand Chain = Op.getOperand(0);
3876 SDOperand Cond = Op.getOperand(1);
3877 SDOperand Dest = Op.getOperand(2);
3879 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3881 if (Cond.getOpcode() == ISD::SETCC)
3882 Cond = LowerSETCC(Cond, DAG, Chain);
3884 if (Cond.getOpcode() == X86ISD::SETCC) {
3885 CC = Cond.getOperand(0);
3887 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3888 // (since flag operand cannot be shared). Use it as the condition setting
3889 // operand in place of the X86ISD::SETCC.
3890 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3891 // to use a test instead of duplicating the X86ISD::CMP (for register
3892 // pressure reason)?
3893 SDOperand Cmp = Cond.getOperand(1);
3894 unsigned Opc = Cmp.getOpcode();
3895 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) {
3896 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3897 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3903 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3904 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3905 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3907 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
3908 Cond, Op.getOperand(2), CC, Cond.getValue(1));
3911 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
3912 unsigned CallingConv= cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3914 if (Subtarget->is64Bit())
3915 return LowerX86_64CCCCallTo(Op, DAG);
3917 switch (CallingConv) {
3919 assert(0 && "Unsupported calling convention");
3920 case CallingConv::Fast:
3922 return LowerFastCCCallTo(Op, DAG);
3925 case CallingConv::C:
3926 return LowerCCCCallTo(Op, DAG);
3927 case CallingConv::X86_StdCall:
3928 return LowerCCCCallTo(Op, DAG, true);
3929 case CallingConv::X86_FastCall:
3930 return LowerFastCCCallTo(Op, DAG, true);
3934 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
3937 switch(Op.getNumOperands()) {
3939 assert(0 && "Do not know how to return this many arguments!");
3941 case 1: // ret void.
3942 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Op.getOperand(0),
3943 DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
3945 MVT::ValueType ArgVT = Op.getOperand(1).getValueType();
3947 if (MVT::isVector(ArgVT) ||
3948 (Subtarget->is64Bit() && MVT::isFloatingPoint(ArgVT))) {
3949 // Integer or FP vector result -> XMM0.
3950 if (DAG.getMachineFunction().liveout_empty())
3951 DAG.getMachineFunction().addLiveOut(X86::XMM0);
3952 Copy = DAG.getCopyToReg(Op.getOperand(0), X86::XMM0, Op.getOperand(1),
3954 } else if (MVT::isInteger(ArgVT)) {
3955 // Integer result -> EAX / RAX.
3956 // The C calling convention guarantees the return value has been
3957 // promoted to at least MVT::i32. The X86-64 ABI doesn't require the
3958 // value to be promoted MVT::i64. So we don't have to extend it to
3959 // 64-bit. Return the value in EAX, but mark RAX as liveout.
3960 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
3961 if (DAG.getMachineFunction().liveout_empty())
3962 DAG.getMachineFunction().addLiveOut(Reg);
3964 Reg = (ArgVT == MVT::i64) ? X86::RAX : X86::EAX;
3965 Copy = DAG.getCopyToReg(Op.getOperand(0), Reg, Op.getOperand(1),
3967 } else if (!X86ScalarSSE) {
3968 // FP return with fp-stack value.
3969 if (DAG.getMachineFunction().liveout_empty())
3970 DAG.getMachineFunction().addLiveOut(X86::ST0);
3972 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3973 SDOperand Ops[] = { Op.getOperand(0), Op.getOperand(1) };
3974 Copy = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops, 2);
3976 // FP return with ScalarSSE (return on fp-stack).
3977 if (DAG.getMachineFunction().liveout_empty())
3978 DAG.getMachineFunction().addLiveOut(X86::ST0);
3981 SDOperand Chain = Op.getOperand(0);
3982 SDOperand Value = Op.getOperand(1);
3984 if (ISD::isNON_EXTLoad(Value.Val) &&
3985 (Chain == Value.getValue(1) || Chain == Value.getOperand(0))) {
3986 Chain = Value.getOperand(0);
3987 MemLoc = Value.getOperand(1);
3989 // Spill the value to memory and reload it into top of stack.
3990 unsigned Size = MVT::getSizeInBits(ArgVT)/8;
3991 MachineFunction &MF = DAG.getMachineFunction();
3992 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
3993 MemLoc = DAG.getFrameIndex(SSFI, getPointerTy());
3994 Chain = DAG.getStore(Op.getOperand(0), Value, MemLoc, NULL, 0);
3996 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
3997 SDOperand Ops[] = { Chain, MemLoc, DAG.getValueType(ArgVT) };
3998 Copy = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
4000 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4001 Ops[0] = Copy.getValue(1);
4003 Copy = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops, 2);
4008 unsigned Reg1 = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
4009 unsigned Reg2 = Subtarget->is64Bit() ? X86::RDX : X86::EDX;
4010 if (DAG.getMachineFunction().liveout_empty()) {
4011 DAG.getMachineFunction().addLiveOut(Reg1);
4012 DAG.getMachineFunction().addLiveOut(Reg2);
4015 Copy = DAG.getCopyToReg(Op.getOperand(0), Reg2, Op.getOperand(3),
4017 Copy = DAG.getCopyToReg(Copy, Reg1, Op.getOperand(1), Copy.getValue(1));
4021 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other,
4022 Copy, DAG.getConstant(getBytesToPopOnReturn(), MVT::i16),
4027 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
4028 MachineFunction &MF = DAG.getMachineFunction();
4029 const Function* Fn = MF.getFunction();
4030 if (Fn->hasExternalLinkage() &&
4031 Subtarget->isTargetCygMing() &&
4032 Fn->getName() == "main")
4033 MF.getInfo<X86FunctionInfo>()->setForceFramePointer(true);
4035 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4036 if (Subtarget->is64Bit())
4037 return LowerX86_64CCCArguments(Op, DAG);
4041 assert(0 && "Unsupported calling convention");
4042 case CallingConv::Fast:
4044 return LowerFastCCArguments(Op, DAG);
4047 case CallingConv::C:
4048 return LowerCCCArguments(Op, DAG);
4049 case CallingConv::X86_StdCall:
4050 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(StdCall);
4051 return LowerCCCArguments(Op, DAG, true);
4052 case CallingConv::X86_FastCall:
4053 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(FastCall);
4054 return LowerFastCCArguments(Op, DAG, true);
4058 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
4059 SDOperand InFlag(0, 0);
4060 SDOperand Chain = Op.getOperand(0);
4062 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
4063 if (Align == 0) Align = 1;
4065 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
4066 // If not DWORD aligned, call memset if size is less than the threshold.
4067 // It knows how to align to the right boundary first.
4068 if ((Align & 3) != 0 ||
4069 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
4070 MVT::ValueType IntPtr = getPointerTy();
4071 const Type *IntPtrTy = getTargetData()->getIntPtrType();
4072 TargetLowering::ArgListTy Args;
4073 TargetLowering::ArgListEntry Entry;
4074 Entry.Node = Op.getOperand(1);
4075 Entry.Ty = IntPtrTy;
4076 Entry.isSigned = false;
4077 Entry.isInReg = false;
4078 Entry.isSRet = false;
4079 Args.push_back(Entry);
4080 // Extend the unsigned i8 argument to be an int value for the call.
4081 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
4082 Entry.Ty = IntPtrTy;
4083 Entry.isSigned = false;
4084 Entry.isInReg = false;
4085 Entry.isSRet = false;
4086 Args.push_back(Entry);
4087 Entry.Node = Op.getOperand(3);
4088 Args.push_back(Entry);
4089 std::pair<SDOperand,SDOperand> CallResult =
4090 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
4091 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
4092 return CallResult.second;
4097 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
4098 unsigned BytesLeft = 0;
4099 bool TwoRepStos = false;
4102 uint64_t Val = ValC->getValue() & 255;
4104 // If the value is a constant, then we can potentially use larger sets.
4105 switch (Align & 3) {
4106 case 2: // WORD aligned
4109 Val = (Val << 8) | Val;
4111 case 0: // DWORD aligned
4114 Val = (Val << 8) | Val;
4115 Val = (Val << 16) | Val;
4116 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
4119 Val = (Val << 32) | Val;
4122 default: // Byte aligned
4125 Count = Op.getOperand(3);
4129 if (AVT > MVT::i8) {
4131 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4132 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
4133 BytesLeft = I->getValue() % UBytes;
4135 assert(AVT >= MVT::i32 &&
4136 "Do not use rep;stos if not at least DWORD aligned");
4137 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
4138 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
4143 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
4145 InFlag = Chain.getValue(1);
4148 Count = Op.getOperand(3);
4149 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
4150 InFlag = Chain.getValue(1);
4153 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4155 InFlag = Chain.getValue(1);
4156 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4157 Op.getOperand(1), InFlag);
4158 InFlag = Chain.getValue(1);
4160 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4161 SmallVector<SDOperand, 8> Ops;
4162 Ops.push_back(Chain);
4163 Ops.push_back(DAG.getValueType(AVT));
4164 Ops.push_back(InFlag);
4165 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4168 InFlag = Chain.getValue(1);
4169 Count = Op.getOperand(3);
4170 MVT::ValueType CVT = Count.getValueType();
4171 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4172 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4173 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4175 InFlag = Chain.getValue(1);
4176 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4178 Ops.push_back(Chain);
4179 Ops.push_back(DAG.getValueType(MVT::i8));
4180 Ops.push_back(InFlag);
4181 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4182 } else if (BytesLeft) {
4183 // Issue stores for the last 1 - 7 bytes.
4185 unsigned Val = ValC->getValue() & 255;
4186 unsigned Offset = I->getValue() - BytesLeft;
4187 SDOperand DstAddr = Op.getOperand(1);
4188 MVT::ValueType AddrVT = DstAddr.getValueType();
4189 if (BytesLeft >= 4) {
4190 Val = (Val << 8) | Val;
4191 Val = (Val << 16) | Val;
4192 Value = DAG.getConstant(Val, MVT::i32);
4193 Chain = DAG.getStore(Chain, Value,
4194 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4195 DAG.getConstant(Offset, AddrVT)),
4200 if (BytesLeft >= 2) {
4201 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
4202 Chain = DAG.getStore(Chain, Value,
4203 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4204 DAG.getConstant(Offset, AddrVT)),
4209 if (BytesLeft == 1) {
4210 Value = DAG.getConstant(Val, MVT::i8);
4211 Chain = DAG.getStore(Chain, Value,
4212 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4213 DAG.getConstant(Offset, AddrVT)),
4221 SDOperand X86TargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) {
4222 SDOperand Chain = Op.getOperand(0);
4224 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
4225 if (Align == 0) Align = 1;
4227 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
4228 // If not DWORD aligned, call memcpy if size is less than the threshold.
4229 // It knows how to align to the right boundary first.
4230 if ((Align & 3) != 0 ||
4231 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
4232 MVT::ValueType IntPtr = getPointerTy();
4233 TargetLowering::ArgListTy Args;
4234 TargetLowering::ArgListEntry Entry;
4235 Entry.Ty = getTargetData()->getIntPtrType();
4236 Entry.isSigned = false;
4237 Entry.isInReg = false;
4238 Entry.isSRet = false;
4239 Entry.Node = Op.getOperand(1); Args.push_back(Entry);
4240 Entry.Node = Op.getOperand(2); Args.push_back(Entry);
4241 Entry.Node = Op.getOperand(3); Args.push_back(Entry);
4242 std::pair<SDOperand,SDOperand> CallResult =
4243 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
4244 DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
4245 return CallResult.second;
4250 unsigned BytesLeft = 0;
4251 bool TwoRepMovs = false;
4252 switch (Align & 3) {
4253 case 2: // WORD aligned
4256 case 0: // DWORD aligned
4258 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
4261 default: // Byte aligned
4263 Count = Op.getOperand(3);
4267 if (AVT > MVT::i8) {
4269 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4270 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
4271 BytesLeft = I->getValue() % UBytes;
4273 assert(AVT >= MVT::i32 &&
4274 "Do not use rep;movs if not at least DWORD aligned");
4275 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
4276 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
4281 SDOperand InFlag(0, 0);
4282 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4284 InFlag = Chain.getValue(1);
4285 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4286 Op.getOperand(1), InFlag);
4287 InFlag = Chain.getValue(1);
4288 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
4289 Op.getOperand(2), InFlag);
4290 InFlag = Chain.getValue(1);
4292 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4293 SmallVector<SDOperand, 8> Ops;
4294 Ops.push_back(Chain);
4295 Ops.push_back(DAG.getValueType(AVT));
4296 Ops.push_back(InFlag);
4297 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4300 InFlag = Chain.getValue(1);
4301 Count = Op.getOperand(3);
4302 MVT::ValueType CVT = Count.getValueType();
4303 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4304 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4305 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4307 InFlag = Chain.getValue(1);
4308 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4310 Ops.push_back(Chain);
4311 Ops.push_back(DAG.getValueType(MVT::i8));
4312 Ops.push_back(InFlag);
4313 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4314 } else if (BytesLeft) {
4315 // Issue loads and stores for the last 1 - 7 bytes.
4316 unsigned Offset = I->getValue() - BytesLeft;
4317 SDOperand DstAddr = Op.getOperand(1);
4318 MVT::ValueType DstVT = DstAddr.getValueType();
4319 SDOperand SrcAddr = Op.getOperand(2);
4320 MVT::ValueType SrcVT = SrcAddr.getValueType();
4322 if (BytesLeft >= 4) {
4323 Value = DAG.getLoad(MVT::i32, Chain,
4324 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4325 DAG.getConstant(Offset, SrcVT)),
4327 Chain = Value.getValue(1);
4328 Chain = DAG.getStore(Chain, Value,
4329 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4330 DAG.getConstant(Offset, DstVT)),
4335 if (BytesLeft >= 2) {
4336 Value = DAG.getLoad(MVT::i16, Chain,
4337 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4338 DAG.getConstant(Offset, SrcVT)),
4340 Chain = Value.getValue(1);
4341 Chain = DAG.getStore(Chain, Value,
4342 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4343 DAG.getConstant(Offset, DstVT)),
4349 if (BytesLeft == 1) {
4350 Value = DAG.getLoad(MVT::i8, Chain,
4351 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4352 DAG.getConstant(Offset, SrcVT)),
4354 Chain = Value.getValue(1);
4355 Chain = DAG.getStore(Chain, Value,
4356 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4357 DAG.getConstant(Offset, DstVT)),
4366 X86TargetLowering::LowerREADCYCLCECOUNTER(SDOperand Op, SelectionDAG &DAG) {
4367 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
4368 SDOperand TheOp = Op.getOperand(0);
4369 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheOp, 1);
4370 if (Subtarget->is64Bit()) {
4371 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
4372 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::RDX,
4373 MVT::i64, Copy1.getValue(2));
4374 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, Copy2,
4375 DAG.getConstant(32, MVT::i8));
4377 DAG.getNode(ISD::OR, MVT::i64, Copy1, Tmp), Copy2.getValue(1)
4380 Tys = DAG.getVTList(MVT::i64, MVT::Other);
4381 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2);
4384 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
4385 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::EDX,
4386 MVT::i32, Copy1.getValue(2));
4387 SDOperand Ops[] = { Copy1, Copy2, Copy2.getValue(1) };
4388 Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
4389 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 3);
4392 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
4393 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
4395 if (!Subtarget->is64Bit()) {
4396 // vastart just stores the address of the VarArgsFrameIndex slot into the
4397 // memory location argument.
4398 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4399 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
4404 // gp_offset (0 - 6 * 8)
4405 // fp_offset (48 - 48 + 8 * 16)
4406 // overflow_arg_area (point to parameters coming in memory).
4408 SmallVector<SDOperand, 8> MemOps;
4409 SDOperand FIN = Op.getOperand(1);
4411 SDOperand Store = DAG.getStore(Op.getOperand(0),
4412 DAG.getConstant(VarArgsGPOffset, MVT::i32),
4413 FIN, SV->getValue(), SV->getOffset());
4414 MemOps.push_back(Store);
4417 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4418 DAG.getConstant(4, getPointerTy()));
4419 Store = DAG.getStore(Op.getOperand(0),
4420 DAG.getConstant(VarArgsFPOffset, MVT::i32),
4421 FIN, SV->getValue(), SV->getOffset());
4422 MemOps.push_back(Store);
4424 // Store ptr to overflow_arg_area
4425 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4426 DAG.getConstant(4, getPointerTy()));
4427 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4428 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
4430 MemOps.push_back(Store);
4432 // Store ptr to reg_save_area.
4433 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4434 DAG.getConstant(8, getPointerTy()));
4435 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
4436 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
4438 MemOps.push_back(Store);
4439 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
4443 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
4444 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
4446 default: return SDOperand(); // Don't custom lower most intrinsics.
4447 // Comparison intrinsics.
4448 case Intrinsic::x86_sse_comieq_ss:
4449 case Intrinsic::x86_sse_comilt_ss:
4450 case Intrinsic::x86_sse_comile_ss:
4451 case Intrinsic::x86_sse_comigt_ss:
4452 case Intrinsic::x86_sse_comige_ss:
4453 case Intrinsic::x86_sse_comineq_ss:
4454 case Intrinsic::x86_sse_ucomieq_ss:
4455 case Intrinsic::x86_sse_ucomilt_ss:
4456 case Intrinsic::x86_sse_ucomile_ss:
4457 case Intrinsic::x86_sse_ucomigt_ss:
4458 case Intrinsic::x86_sse_ucomige_ss:
4459 case Intrinsic::x86_sse_ucomineq_ss:
4460 case Intrinsic::x86_sse2_comieq_sd:
4461 case Intrinsic::x86_sse2_comilt_sd:
4462 case Intrinsic::x86_sse2_comile_sd:
4463 case Intrinsic::x86_sse2_comigt_sd:
4464 case Intrinsic::x86_sse2_comige_sd:
4465 case Intrinsic::x86_sse2_comineq_sd:
4466 case Intrinsic::x86_sse2_ucomieq_sd:
4467 case Intrinsic::x86_sse2_ucomilt_sd:
4468 case Intrinsic::x86_sse2_ucomile_sd:
4469 case Intrinsic::x86_sse2_ucomigt_sd:
4470 case Intrinsic::x86_sse2_ucomige_sd:
4471 case Intrinsic::x86_sse2_ucomineq_sd: {
4473 ISD::CondCode CC = ISD::SETCC_INVALID;
4476 case Intrinsic::x86_sse_comieq_ss:
4477 case Intrinsic::x86_sse2_comieq_sd:
4481 case Intrinsic::x86_sse_comilt_ss:
4482 case Intrinsic::x86_sse2_comilt_sd:
4486 case Intrinsic::x86_sse_comile_ss:
4487 case Intrinsic::x86_sse2_comile_sd:
4491 case Intrinsic::x86_sse_comigt_ss:
4492 case Intrinsic::x86_sse2_comigt_sd:
4496 case Intrinsic::x86_sse_comige_ss:
4497 case Intrinsic::x86_sse2_comige_sd:
4501 case Intrinsic::x86_sse_comineq_ss:
4502 case Intrinsic::x86_sse2_comineq_sd:
4506 case Intrinsic::x86_sse_ucomieq_ss:
4507 case Intrinsic::x86_sse2_ucomieq_sd:
4508 Opc = X86ISD::UCOMI;
4511 case Intrinsic::x86_sse_ucomilt_ss:
4512 case Intrinsic::x86_sse2_ucomilt_sd:
4513 Opc = X86ISD::UCOMI;
4516 case Intrinsic::x86_sse_ucomile_ss:
4517 case Intrinsic::x86_sse2_ucomile_sd:
4518 Opc = X86ISD::UCOMI;
4521 case Intrinsic::x86_sse_ucomigt_ss:
4522 case Intrinsic::x86_sse2_ucomigt_sd:
4523 Opc = X86ISD::UCOMI;
4526 case Intrinsic::x86_sse_ucomige_ss:
4527 case Intrinsic::x86_sse2_ucomige_sd:
4528 Opc = X86ISD::UCOMI;
4531 case Intrinsic::x86_sse_ucomineq_ss:
4532 case Intrinsic::x86_sse2_ucomineq_sd:
4533 Opc = X86ISD::UCOMI;
4539 SDOperand LHS = Op.getOperand(1);
4540 SDOperand RHS = Op.getOperand(2);
4541 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
4543 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
4544 SDOperand Ops1[] = { DAG.getEntryNode(), LHS, RHS };
4545 SDOperand Cond = DAG.getNode(Opc, VTs, 2, Ops1, 3);
4546 VTs = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
4547 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
4548 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, VTs, 2, Ops2, 2);
4549 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
4554 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
4555 // Depths > 0 not supported yet!
4556 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4559 // Just load the return address
4560 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4561 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
4564 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
4565 // Depths > 0 not supported yet!
4566 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4569 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4570 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
4571 DAG.getConstant(4, getPointerTy()));
4574 /// LowerOperation - Provide custom lowering hooks for some operations.
4576 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
4577 switch (Op.getOpcode()) {
4578 default: assert(0 && "Should not custom lower this!");
4579 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4580 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4581 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4582 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
4583 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4584 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4585 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4586 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
4587 case ISD::SHL_PARTS:
4588 case ISD::SRA_PARTS:
4589 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
4590 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4591 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
4592 case ISD::FABS: return LowerFABS(Op, DAG);
4593 case ISD::FNEG: return LowerFNEG(Op, DAG);
4594 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4595 case ISD::SETCC: return LowerSETCC(Op, DAG, DAG.getEntryNode());
4596 case ISD::SELECT: return LowerSELECT(Op, DAG);
4597 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
4598 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4599 case ISD::CALL: return LowerCALL(Op, DAG);
4600 case ISD::RET: return LowerRET(Op, DAG);
4601 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
4602 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
4603 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
4604 case ISD::READCYCLECOUNTER: return LowerREADCYCLCECOUNTER(Op, DAG);
4605 case ISD::VASTART: return LowerVASTART(Op, DAG);
4606 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4607 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4608 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4613 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
4615 default: return NULL;
4616 case X86ISD::SHLD: return "X86ISD::SHLD";
4617 case X86ISD::SHRD: return "X86ISD::SHRD";
4618 case X86ISD::FAND: return "X86ISD::FAND";
4619 case X86ISD::FOR: return "X86ISD::FOR";
4620 case X86ISD::FXOR: return "X86ISD::FXOR";
4621 case X86ISD::FSRL: return "X86ISD::FSRL";
4622 case X86ISD::FILD: return "X86ISD::FILD";
4623 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
4624 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
4625 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
4626 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
4627 case X86ISD::FLD: return "X86ISD::FLD";
4628 case X86ISD::FST: return "X86ISD::FST";
4629 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
4630 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
4631 case X86ISD::CALL: return "X86ISD::CALL";
4632 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
4633 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
4634 case X86ISD::CMP: return "X86ISD::CMP";
4635 case X86ISD::COMI: return "X86ISD::COMI";
4636 case X86ISD::UCOMI: return "X86ISD::UCOMI";
4637 case X86ISD::SETCC: return "X86ISD::SETCC";
4638 case X86ISD::CMOV: return "X86ISD::CMOV";
4639 case X86ISD::BRCOND: return "X86ISD::BRCOND";
4640 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
4641 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
4642 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
4643 case X86ISD::LOAD_PACK: return "X86ISD::LOAD_PACK";
4644 case X86ISD::LOAD_UA: return "X86ISD::LOAD_UA";
4645 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
4646 case X86ISD::Wrapper: return "X86ISD::Wrapper";
4647 case X86ISD::S2VEC: return "X86ISD::S2VEC";
4648 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
4649 case X86ISD::PINSRW: return "X86ISD::PINSRW";
4650 case X86ISD::FMAX: return "X86ISD::FMAX";
4651 case X86ISD::FMIN: return "X86ISD::FMIN";
4655 /// isLegalAddressImmediate - Return true if the integer value or
4656 /// GlobalValue can be used as the offset of the target addressing mode.
4657 bool X86TargetLowering::isLegalAddressImmediate(int64_t V) const {
4658 // X86 allows a sign-extended 32-bit immediate field.
4659 return (V > -(1LL << 32) && V < (1LL << 32)-1);
4662 bool X86TargetLowering::isLegalAddressImmediate(GlobalValue *GV) const {
4663 // In 64-bit mode, GV is 64-bit so it won't fit in the 32-bit displacement
4664 // field unless we are in small code model.
4665 if (Subtarget->is64Bit() &&
4666 getTargetMachine().getCodeModel() != CodeModel::Small)
4669 return (!Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false));
4672 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4673 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4674 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4675 /// are assumed to be legal.
4677 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
4678 // Only do shuffles on 128-bit vector types for now.
4679 if (MVT::getSizeInBits(VT) == 64) return false;
4680 return (Mask.Val->getNumOperands() <= 4 ||
4681 isSplatMask(Mask.Val) ||
4682 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
4683 X86::isUNPCKLMask(Mask.Val) ||
4684 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
4685 X86::isUNPCKHMask(Mask.Val));
4688 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
4690 SelectionDAG &DAG) const {
4691 unsigned NumElts = BVOps.size();
4692 // Only do shuffles on 128-bit vector types for now.
4693 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
4694 if (NumElts == 2) return true;
4696 return (isMOVLMask(&BVOps[0], 4) ||
4697 isCommutedMOVL(&BVOps[0], 4, true) ||
4698 isSHUFPMask(&BVOps[0], 4) ||
4699 isCommutedSHUFP(&BVOps[0], 4));
4704 //===----------------------------------------------------------------------===//
4705 // X86 Scheduler Hooks
4706 //===----------------------------------------------------------------------===//
4709 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
4710 MachineBasicBlock *BB) {
4711 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4712 switch (MI->getOpcode()) {
4713 default: assert(false && "Unexpected instr type to insert");
4714 case X86::CMOV_FR32:
4715 case X86::CMOV_FR64:
4716 case X86::CMOV_V4F32:
4717 case X86::CMOV_V2F64:
4718 case X86::CMOV_V2I64: {
4719 // To "insert" a SELECT_CC instruction, we actually have to insert the
4720 // diamond control-flow pattern. The incoming instruction knows the
4721 // destination vreg to set, the condition code register to branch on, the
4722 // true/false values to select between, and a branch opcode to use.
4723 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4724 ilist<MachineBasicBlock>::iterator It = BB;
4730 // cmpTY ccX, r1, r2
4732 // fallthrough --> copy0MBB
4733 MachineBasicBlock *thisMBB = BB;
4734 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
4735 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
4737 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
4738 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
4739 MachineFunction *F = BB->getParent();
4740 F->getBasicBlockList().insert(It, copy0MBB);
4741 F->getBasicBlockList().insert(It, sinkMBB);
4742 // Update machine-CFG edges by first adding all successors of the current
4743 // block to the new block which will contain the Phi node for the select.
4744 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
4745 e = BB->succ_end(); i != e; ++i)
4746 sinkMBB->addSuccessor(*i);
4747 // Next, remove all successors of the current block, and add the true
4748 // and fallthrough blocks as its successors.
4749 while(!BB->succ_empty())
4750 BB->removeSuccessor(BB->succ_begin());
4751 BB->addSuccessor(copy0MBB);
4752 BB->addSuccessor(sinkMBB);
4755 // %FalseValue = ...
4756 // # fallthrough to sinkMBB
4759 // Update machine-CFG edges
4760 BB->addSuccessor(sinkMBB);
4763 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4766 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
4767 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4768 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4770 delete MI; // The pseudo instruction is gone now.
4774 case X86::FP_TO_INT16_IN_MEM:
4775 case X86::FP_TO_INT32_IN_MEM:
4776 case X86::FP_TO_INT64_IN_MEM: {
4777 // Change the floating point control register to use "round towards zero"
4778 // mode when truncating to an integer value.
4779 MachineFunction *F = BB->getParent();
4780 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
4781 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
4783 // Load the old value of the high byte of the control word...
4785 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
4786 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
4788 // Set the high part to be round to zero...
4789 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
4792 // Reload the modified control word now...
4793 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4795 // Restore the memory image of control word to original value
4796 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
4799 // Get the X86 opcode to use.
4801 switch (MI->getOpcode()) {
4802 default: assert(0 && "illegal opcode!");
4803 case X86::FP_TO_INT16_IN_MEM: Opc = X86::FpIST16m; break;
4804 case X86::FP_TO_INT32_IN_MEM: Opc = X86::FpIST32m; break;
4805 case X86::FP_TO_INT64_IN_MEM: Opc = X86::FpIST64m; break;
4809 MachineOperand &Op = MI->getOperand(0);
4810 if (Op.isRegister()) {
4811 AM.BaseType = X86AddressMode::RegBase;
4812 AM.Base.Reg = Op.getReg();
4814 AM.BaseType = X86AddressMode::FrameIndexBase;
4815 AM.Base.FrameIndex = Op.getFrameIndex();
4817 Op = MI->getOperand(1);
4818 if (Op.isImmediate())
4819 AM.Scale = Op.getImm();
4820 Op = MI->getOperand(2);
4821 if (Op.isImmediate())
4822 AM.IndexReg = Op.getImm();
4823 Op = MI->getOperand(3);
4824 if (Op.isGlobalAddress()) {
4825 AM.GV = Op.getGlobal();
4827 AM.Disp = Op.getImm();
4829 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
4830 .addReg(MI->getOperand(4).getReg());
4832 // Reload the original control word now.
4833 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4835 delete MI; // The pseudo instruction is gone now.
4841 //===----------------------------------------------------------------------===//
4842 // X86 Optimization Hooks
4843 //===----------------------------------------------------------------------===//
4845 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
4847 uint64_t &KnownZero,
4849 unsigned Depth) const {
4850 unsigned Opc = Op.getOpcode();
4851 assert((Opc >= ISD::BUILTIN_OP_END ||
4852 Opc == ISD::INTRINSIC_WO_CHAIN ||
4853 Opc == ISD::INTRINSIC_W_CHAIN ||
4854 Opc == ISD::INTRINSIC_VOID) &&
4855 "Should use MaskedValueIsZero if you don't know whether Op"
4856 " is a target node!");
4858 KnownZero = KnownOne = 0; // Don't know anything.
4862 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
4867 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4868 /// element of the result of the vector shuffle.
4869 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
4870 MVT::ValueType VT = N->getValueType(0);
4871 SDOperand PermMask = N->getOperand(2);
4872 unsigned NumElems = PermMask.getNumOperands();
4873 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
4875 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4877 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4878 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
4879 SDOperand Idx = PermMask.getOperand(i);
4880 if (Idx.getOpcode() == ISD::UNDEF)
4881 return DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4882 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
4887 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
4888 /// node is a GlobalAddress + an offset.
4889 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
4890 unsigned Opc = N->getOpcode();
4891 if (Opc == X86ISD::Wrapper) {
4892 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
4893 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
4896 } else if (Opc == ISD::ADD) {
4897 SDOperand N1 = N->getOperand(0);
4898 SDOperand N2 = N->getOperand(1);
4899 if (isGAPlusOffset(N1.Val, GA, Offset)) {
4900 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
4902 Offset += V->getSignExtended();
4905 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
4906 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
4908 Offset += V->getSignExtended();
4916 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
4918 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
4919 MachineFrameInfo *MFI) {
4920 if (N->getOperand(0).Val != Base->getOperand(0).Val)
4923 SDOperand Loc = N->getOperand(1);
4924 SDOperand BaseLoc = Base->getOperand(1);
4925 if (Loc.getOpcode() == ISD::FrameIndex) {
4926 if (BaseLoc.getOpcode() != ISD::FrameIndex)
4928 int FI = dyn_cast<FrameIndexSDNode>(Loc)->getIndex();
4929 int BFI = dyn_cast<FrameIndexSDNode>(BaseLoc)->getIndex();
4930 int FS = MFI->getObjectSize(FI);
4931 int BFS = MFI->getObjectSize(BFI);
4932 if (FS != BFS || FS != Size) return false;
4933 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
4935 GlobalValue *GV1 = NULL;
4936 GlobalValue *GV2 = NULL;
4937 int64_t Offset1 = 0;
4938 int64_t Offset2 = 0;
4939 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
4940 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
4941 if (isGA1 && isGA2 && GV1 == GV2)
4942 return Offset1 == (Offset2 + Dist*Size);
4948 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
4949 const X86Subtarget *Subtarget) {
4952 if (isGAPlusOffset(Base, GV, Offset))
4953 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
4955 assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
4956 int BFI = dyn_cast<FrameIndexSDNode>(Base)->getIndex();
4958 // Fixed objects do not specify alignment, however the offsets are known.
4959 return ((Subtarget->getStackAlignment() % 16) == 0 &&
4960 (MFI->getObjectOffset(BFI) % 16) == 0);
4962 return MFI->getObjectAlignment(BFI) >= 16;
4968 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
4969 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
4970 /// if the load addresses are consecutive, non-overlapping, and in the right
4972 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
4973 const X86Subtarget *Subtarget) {
4974 MachineFunction &MF = DAG.getMachineFunction();
4975 MachineFrameInfo *MFI = MF.getFrameInfo();
4976 MVT::ValueType VT = N->getValueType(0);
4977 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
4978 SDOperand PermMask = N->getOperand(2);
4979 int NumElems = (int)PermMask.getNumOperands();
4980 SDNode *Base = NULL;
4981 for (int i = 0; i < NumElems; ++i) {
4982 SDOperand Idx = PermMask.getOperand(i);
4983 if (Idx.getOpcode() == ISD::UNDEF) {
4984 if (!Base) return SDOperand();
4987 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
4988 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
4992 else if (!isConsecutiveLoad(Arg.Val, Base,
4993 i, MVT::getSizeInBits(EVT)/8,MFI))
4998 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
5000 LoadSDNode *LD = cast<LoadSDNode>(Base);
5001 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
5002 LD->getSrcValueOffset());
5004 // Just use movups, it's shorter.
5005 SDVTList Tys = DAG.getVTList(MVT::v4f32, MVT::Other);
5006 SmallVector<SDOperand, 3> Ops;
5007 Ops.push_back(Base->getOperand(0));
5008 Ops.push_back(Base->getOperand(1));
5009 Ops.push_back(Base->getOperand(2));
5010 return DAG.getNode(ISD::BIT_CONVERT, VT,
5011 DAG.getNode(X86ISD::LOAD_UA, Tys, &Ops[0], Ops.size()));
5015 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
5016 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
5017 const X86Subtarget *Subtarget) {
5018 SDOperand Cond = N->getOperand(0);
5020 // If we have SSE[12] support, try to form min/max nodes.
5021 if (Subtarget->hasSSE2() &&
5022 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
5023 if (Cond.getOpcode() == ISD::SETCC) {
5024 // Get the LHS/RHS of the select.
5025 SDOperand LHS = N->getOperand(1);
5026 SDOperand RHS = N->getOperand(2);
5027 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
5029 unsigned Opcode = 0;
5030 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
5033 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
5036 if (!UnsafeFPMath) break;
5038 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
5040 Opcode = X86ISD::FMIN;
5043 case ISD::SETOGT: // (X > Y) ? X : Y -> max
5046 if (!UnsafeFPMath) break;
5048 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
5050 Opcode = X86ISD::FMAX;
5053 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
5056 case ISD::SETOGT: // (X > Y) ? Y : X -> min
5059 if (!UnsafeFPMath) break;
5061 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
5063 Opcode = X86ISD::FMIN;
5066 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
5069 if (!UnsafeFPMath) break;
5071 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
5073 Opcode = X86ISD::FMAX;
5079 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
5088 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
5089 DAGCombinerInfo &DCI) const {
5090 SelectionDAG &DAG = DCI.DAG;
5091 switch (N->getOpcode()) {
5093 case ISD::VECTOR_SHUFFLE:
5094 return PerformShuffleCombine(N, DAG, Subtarget);
5096 return PerformSELECTCombine(N, DAG, Subtarget);
5102 //===----------------------------------------------------------------------===//
5103 // X86 Inline Assembly Support
5104 //===----------------------------------------------------------------------===//
5106 /// getConstraintType - Given a constraint letter, return the type of
5107 /// constraint it is for this target.
5108 X86TargetLowering::ConstraintType
5109 X86TargetLowering::getConstraintType(char ConstraintLetter) const {
5110 switch (ConstraintLetter) {
5119 return C_RegisterClass;
5120 default: return TargetLowering::getConstraintType(ConstraintLetter);
5124 /// isOperandValidForConstraint - Return the specified operand (possibly
5125 /// modified) if the specified SDOperand is valid for the specified target
5126 /// constraint letter, otherwise return null.
5127 SDOperand X86TargetLowering::
5128 isOperandValidForConstraint(SDOperand Op, char Constraint, SelectionDAG &DAG) {
5129 switch (Constraint) {
5132 // Literal immediates are always ok.
5133 if (isa<ConstantSDNode>(Op)) return Op;
5135 // If we are in non-pic codegen mode, we allow the address of a global to
5136 // be used with 'i'.
5137 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op)) {
5138 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
5139 return SDOperand(0, 0);
5141 if (GA->getOpcode() != ISD::TargetGlobalAddress)
5142 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5147 // Otherwise, not valid for this mode.
5148 return SDOperand(0, 0);
5150 return TargetLowering::isOperandValidForConstraint(Op, Constraint, DAG);
5154 std::vector<unsigned> X86TargetLowering::
5155 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5156 MVT::ValueType VT) const {
5157 if (Constraint.size() == 1) {
5158 // FIXME: not handling fp-stack yet!
5159 // FIXME: not handling MMX registers yet ('y' constraint).
5160 switch (Constraint[0]) { // GCC X86 Constraint Letters
5161 default: break; // Unknown constraint letter
5162 case 'A': // EAX/EDX
5163 if (VT == MVT::i32 || VT == MVT::i64)
5164 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
5166 case 'r': // GENERAL_REGS
5167 case 'R': // LEGACY_REGS
5168 if (VT == MVT::i64 && Subtarget->is64Bit())
5169 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
5170 X86::RSI, X86::RDI, X86::RBP, X86::RSP,
5171 X86::R8, X86::R9, X86::R10, X86::R11,
5172 X86::R12, X86::R13, X86::R14, X86::R15, 0);
5174 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
5175 X86::ESI, X86::EDI, X86::EBP, X86::ESP, 0);
5176 else if (VT == MVT::i16)
5177 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
5178 X86::SI, X86::DI, X86::BP, X86::SP, 0);
5179 else if (VT == MVT::i8)
5180 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
5182 case 'l': // INDEX_REGS
5184 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
5185 X86::ESI, X86::EDI, X86::EBP, 0);
5186 else if (VT == MVT::i16)
5187 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
5188 X86::SI, X86::DI, X86::BP, 0);
5189 else if (VT == MVT::i8)
5190 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
5192 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
5195 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
5196 else if (VT == MVT::i16)
5197 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
5198 else if (VT == MVT::i8)
5199 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
5201 case 'x': // SSE_REGS if SSE1 allowed
5202 if (Subtarget->hasSSE1())
5203 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
5204 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
5206 return std::vector<unsigned>();
5207 case 'Y': // SSE_REGS if SSE2 allowed
5208 if (Subtarget->hasSSE2())
5209 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
5210 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
5212 return std::vector<unsigned>();
5216 return std::vector<unsigned>();
5219 std::pair<unsigned, const TargetRegisterClass*>
5220 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5221 MVT::ValueType VT) const {
5222 // Use the default implementation in TargetLowering to convert the register
5223 // constraint into a member of a register class.
5224 std::pair<unsigned, const TargetRegisterClass*> Res;
5225 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5227 // Not found as a standard register?
5228 if (Res.second == 0) {
5229 // GCC calls "st(0)" just plain "st".
5230 if (StringsEqualNoCase("{st}", Constraint)) {
5231 Res.first = X86::ST0;
5232 Res.second = X86::RSTRegisterClass;
5238 // Otherwise, check to see if this is a register class of the wrong value
5239 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
5240 // turn into {ax},{dx}.
5241 if (Res.second->hasType(VT))
5242 return Res; // Correct type already, nothing to do.
5244 // All of the single-register GCC register classes map their values onto
5245 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
5246 // really want an 8-bit or 32-bit register, map to the appropriate register
5247 // class and return the appropriate register.
5248 if (Res.second != X86::GR16RegisterClass)
5251 if (VT == MVT::i8) {
5252 unsigned DestReg = 0;
5253 switch (Res.first) {
5255 case X86::AX: DestReg = X86::AL; break;
5256 case X86::DX: DestReg = X86::DL; break;
5257 case X86::CX: DestReg = X86::CL; break;
5258 case X86::BX: DestReg = X86::BL; break;
5261 Res.first = DestReg;
5262 Res.second = Res.second = X86::GR8RegisterClass;
5264 } else if (VT == MVT::i32) {
5265 unsigned DestReg = 0;
5266 switch (Res.first) {
5268 case X86::AX: DestReg = X86::EAX; break;
5269 case X86::DX: DestReg = X86::EDX; break;
5270 case X86::CX: DestReg = X86::ECX; break;
5271 case X86::BX: DestReg = X86::EBX; break;
5272 case X86::SI: DestReg = X86::ESI; break;
5273 case X86::DI: DestReg = X86::EDI; break;
5274 case X86::BP: DestReg = X86::EBP; break;
5275 case X86::SP: DestReg = X86::ESP; break;
5278 Res.first = DestReg;
5279 Res.second = Res.second = X86::GR32RegisterClass;
5281 } else if (VT == MVT::i64) {
5282 unsigned DestReg = 0;
5283 switch (Res.first) {
5285 case X86::AX: DestReg = X86::RAX; break;
5286 case X86::DX: DestReg = X86::RDX; break;
5287 case X86::CX: DestReg = X86::RCX; break;
5288 case X86::BX: DestReg = X86::RBX; break;
5289 case X86::SI: DestReg = X86::RSI; break;
5290 case X86::DI: DestReg = X86::RDI; break;
5291 case X86::BP: DestReg = X86::RBP; break;
5292 case X86::SP: DestReg = X86::RSP; break;
5295 Res.first = DestReg;
5296 Res.second = Res.second = X86::GR64RegisterClass;