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 std::vector<SDOperand> 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 std::vector<bool> ArgInRegs(NumArgs, false);
553 std::vector<bool> 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 std::vector<MVT::ValueType> RetVTs(Op.Val->value_begin(),
646 Op.Val->value_end());
647 return DAG.getNode(ISD::MERGE_VALUES, RetVTs, &ArgValues[0],ArgValues.size());
650 SDOperand X86TargetLowering::LowerCCCCallTo(SDOperand Op, SelectionDAG &DAG,
652 SDOperand Chain = Op.getOperand(0);
653 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
654 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
655 SDOperand Callee = Op.getOperand(4);
656 MVT::ValueType RetVT= Op.Val->getValueType(0);
657 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
659 static const unsigned XMMArgRegs[] = {
660 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
662 static const unsigned GPR32ArgRegs[] = {
663 X86::EAX, X86::EDX, X86::ECX
666 // Count how many bytes are to be pushed on the stack.
667 unsigned NumBytes = 0;
668 // Keep track of the number of integer regs passed so far.
669 unsigned NumIntRegs = 0;
670 // Keep track of the number of XMM regs passed so far.
671 unsigned NumXMMRegs = 0;
672 // How much bytes on stack used for struct return
673 unsigned NumSRetBytes= 0;
675 // Handle regparm attribute
676 std::vector<bool> ArgInRegs(NumOps, false);
677 std::vector<bool> SRetArgs(NumOps, false);
678 for (unsigned i = 0; i<NumOps; ++i) {
680 dyn_cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
681 ArgInRegs[i] = (Flags >> 1) & 1;
682 SRetArgs[i] = (Flags >> 2) & 1;
685 // Calculate stack frame size
686 for (unsigned i = 0; i != NumOps; ++i) {
687 SDOperand Arg = Op.getOperand(5+2*i);
688 unsigned ArgIncrement = 4;
689 unsigned ObjSize = 0;
690 unsigned ObjIntRegs = 0;
691 unsigned ObjXMMRegs = 0;
693 HowToPassCallArgument(Arg.getValueType(),
695 NumIntRegs, NumXMMRegs, 3,
696 ObjSize, ObjIntRegs, ObjXMMRegs,
699 ArgIncrement = ObjSize;
701 NumIntRegs += ObjIntRegs;
702 NumXMMRegs += ObjXMMRegs;
704 // XMM arguments have to be aligned on 16-byte boundary.
706 NumBytes = ((NumBytes + 15) / 16) * 16;
707 NumBytes += ArgIncrement;
711 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
713 // Arguments go on the stack in reverse order, as specified by the ABI.
714 unsigned ArgOffset = 0;
717 std::vector<std::pair<unsigned, SDOperand> > RegsToPass;
718 std::vector<SDOperand> MemOpChains;
719 SDOperand StackPtr = DAG.getRegister(X86StackPtr, getPointerTy());
720 for (unsigned i = 0; i != NumOps; ++i) {
721 SDOperand Arg = Op.getOperand(5+2*i);
722 unsigned ArgIncrement = 4;
723 unsigned ObjSize = 0;
724 unsigned ObjIntRegs = 0;
725 unsigned ObjXMMRegs = 0;
727 HowToPassCallArgument(Arg.getValueType(),
729 NumIntRegs, NumXMMRegs, 3,
730 ObjSize, ObjIntRegs, ObjXMMRegs,
734 ArgIncrement = ObjSize;
736 if (Arg.getValueType() == MVT::i8 || Arg.getValueType() == MVT::i16) {
737 // Promote the integer to 32 bits. If the input type is signed use a
738 // sign extend, otherwise use a zero extend.
739 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
741 unsigned ExtOp = (Flags & 1) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
742 Arg = DAG.getNode(ExtOp, MVT::i32, Arg);
745 if (ObjIntRegs || ObjXMMRegs) {
746 switch (Arg.getValueType()) {
747 default: assert(0 && "Unhandled argument type!");
749 RegsToPass.push_back(std::make_pair(GPR32ArgRegs[NumIntRegs], Arg));
757 assert(!isStdCall && "Unhandled argument type!");
758 RegsToPass.push_back(std::make_pair(XMMArgRegs[NumXMMRegs], Arg));
762 NumIntRegs += ObjIntRegs;
763 NumXMMRegs += ObjXMMRegs;
766 // XMM arguments have to be aligned on 16-byte boundary.
768 ArgOffset = ((ArgOffset + 15) / 16) * 16;
770 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
771 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
772 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
774 ArgOffset += ArgIncrement; // Move on to the next argument.
776 NumSRetBytes += ArgIncrement;
780 if (!MemOpChains.empty())
781 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
782 &MemOpChains[0], MemOpChains.size());
784 // Build a sequence of copy-to-reg nodes chained together with token chain
785 // and flag operands which copy the outgoing args into registers.
787 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
788 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
790 InFlag = Chain.getValue(1);
793 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
794 Subtarget->isPICStyleGOT()) {
795 Chain = DAG.getCopyToReg(Chain, X86::EBX,
796 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
798 InFlag = Chain.getValue(1);
801 // If the callee is a GlobalAddress node (quite common, every direct call is)
802 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
803 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
804 // We should use extra load for direct calls to dllimported functions in
806 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
807 getTargetMachine(), true))
808 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
809 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
810 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
812 std::vector<MVT::ValueType> NodeTys;
813 NodeTys.push_back(MVT::Other); // Returns a chain
814 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
815 std::vector<SDOperand> Ops;
816 Ops.push_back(Chain);
817 Ops.push_back(Callee);
819 // Add argument registers to the end of the list so that they are known live
821 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
822 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
823 RegsToPass[i].second.getValueType()));
826 Ops.push_back(InFlag);
828 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
829 NodeTys, &Ops[0], Ops.size());
830 InFlag = Chain.getValue(1);
832 // Create the CALLSEQ_END node.
833 unsigned NumBytesForCalleeToPush = 0;
837 NumBytesForCalleeToPush = NumSRetBytes;
839 NumBytesForCalleeToPush = NumBytes;
842 // If this is is a call to a struct-return function, the callee
843 // pops the hidden struct pointer, so we have to push it back.
844 // This is common for Darwin/X86, Linux & Mingw32 targets.
845 NumBytesForCalleeToPush = NumSRetBytes;
849 NodeTys.push_back(MVT::Other); // Returns a chain
850 if (RetVT != MVT::Other)
851 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
853 Ops.push_back(Chain);
854 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
855 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
856 Ops.push_back(InFlag);
857 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
858 if (RetVT != MVT::Other)
859 InFlag = Chain.getValue(1);
861 std::vector<SDOperand> ResultVals;
864 default: assert(0 && "Unknown value type to return!");
865 case MVT::Other: break;
867 Chain = DAG.getCopyFromReg(Chain, X86::AL, MVT::i8, InFlag).getValue(1);
868 ResultVals.push_back(Chain.getValue(0));
869 NodeTys.push_back(MVT::i8);
872 Chain = DAG.getCopyFromReg(Chain, X86::AX, MVT::i16, InFlag).getValue(1);
873 ResultVals.push_back(Chain.getValue(0));
874 NodeTys.push_back(MVT::i16);
877 if (Op.Val->getValueType(1) == MVT::i32) {
878 Chain = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag).getValue(1);
879 ResultVals.push_back(Chain.getValue(0));
880 Chain = DAG.getCopyFromReg(Chain, X86::EDX, MVT::i32,
881 Chain.getValue(2)).getValue(1);
882 ResultVals.push_back(Chain.getValue(0));
883 NodeTys.push_back(MVT::i32);
885 Chain = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag).getValue(1);
886 ResultVals.push_back(Chain.getValue(0));
888 NodeTys.push_back(MVT::i32);
896 assert(!isStdCall && "Unknown value type to return!");
897 Chain = DAG.getCopyFromReg(Chain, X86::XMM0, RetVT, InFlag).getValue(1);
898 ResultVals.push_back(Chain.getValue(0));
899 NodeTys.push_back(RetVT);
903 std::vector<MVT::ValueType> Tys;
904 Tys.push_back(MVT::f64);
905 Tys.push_back(MVT::Other);
906 Tys.push_back(MVT::Flag);
907 std::vector<SDOperand> Ops;
908 Ops.push_back(Chain);
909 Ops.push_back(InFlag);
910 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys,
911 &Ops[0], Ops.size());
912 Chain = RetVal.getValue(1);
913 InFlag = RetVal.getValue(2);
915 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
916 // shouldn't be necessary except that RFP cannot be live across
917 // multiple blocks. When stackifier is fixed, they can be uncoupled.
918 MachineFunction &MF = DAG.getMachineFunction();
919 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
920 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
922 Tys.push_back(MVT::Other);
924 Ops.push_back(Chain);
925 Ops.push_back(RetVal);
926 Ops.push_back(StackSlot);
927 Ops.push_back(DAG.getValueType(RetVT));
928 Ops.push_back(InFlag);
929 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
930 RetVal = DAG.getLoad(RetVT, Chain, StackSlot, NULL, 0);
931 Chain = RetVal.getValue(1);
934 if (RetVT == MVT::f32 && !X86ScalarSSE)
935 // FIXME: we would really like to remember that this FP_ROUND
936 // operation is okay to eliminate if we allow excess FP precision.
937 RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal);
938 ResultVals.push_back(RetVal);
939 NodeTys.push_back(RetVT);
944 // If the function returns void, just return the chain.
945 if (ResultVals.empty())
948 // Otherwise, merge everything together with a MERGE_VALUES node.
949 NodeTys.push_back(MVT::Other);
950 ResultVals.push_back(Chain);
951 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys,
952 &ResultVals[0], ResultVals.size());
953 return Res.getValue(Op.ResNo);
957 //===----------------------------------------------------------------------===//
958 // X86-64 C Calling Convention implementation
959 //===----------------------------------------------------------------------===//
961 /// HowToPassX86_64CCCArgument - Returns how an formal argument of the specified
962 /// type should be passed. If it is through stack, returns the size of the stack
963 /// slot; if it is through integer or XMM register, returns the number of
964 /// integer or XMM registers are needed.
966 HowToPassX86_64CCCArgument(MVT::ValueType ObjectVT,
967 unsigned NumIntRegs, unsigned NumXMMRegs,
968 unsigned &ObjSize, unsigned &ObjIntRegs,
969 unsigned &ObjXMMRegs) {
975 default: assert(0 && "Unhandled argument type!");
985 case MVT::i8: ObjSize = 1; break;
986 case MVT::i16: ObjSize = 2; break;
987 case MVT::i32: ObjSize = 4; break;
988 case MVT::i64: ObjSize = 8; break;
1005 case MVT::f32: ObjSize = 4; break;
1006 case MVT::f64: ObjSize = 8; break;
1012 case MVT::v2f64: ObjSize = 16; break;
1020 X86TargetLowering::LowerX86_64CCCArguments(SDOperand Op, SelectionDAG &DAG) {
1021 unsigned NumArgs = Op.Val->getNumValues() - 1;
1022 MachineFunction &MF = DAG.getMachineFunction();
1023 MachineFrameInfo *MFI = MF.getFrameInfo();
1024 SDOperand Root = Op.getOperand(0);
1025 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1026 std::vector<SDOperand> ArgValues;
1028 // Add DAG nodes to load the arguments... On entry to a function on the X86,
1029 // the stack frame looks like this:
1031 // [RSP] -- return address
1032 // [RSP + 8] -- first nonreg argument (leftmost lexically)
1033 // [RSP +16] -- second nonreg argument, if 1st argument is <= 8 bytes in size
1036 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
1037 unsigned NumIntRegs = 0; // Int regs used for parameter passing.
1038 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1040 static const unsigned GPR8ArgRegs[] = {
1041 X86::DIL, X86::SIL, X86::DL, X86::CL, X86::R8B, X86::R9B
1043 static const unsigned GPR16ArgRegs[] = {
1044 X86::DI, X86::SI, X86::DX, X86::CX, X86::R8W, X86::R9W
1046 static const unsigned GPR32ArgRegs[] = {
1047 X86::EDI, X86::ESI, X86::EDX, X86::ECX, X86::R8D, X86::R9D
1049 static const unsigned GPR64ArgRegs[] = {
1050 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1052 static const unsigned XMMArgRegs[] = {
1053 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1054 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1057 for (unsigned i = 0; i < NumArgs; ++i) {
1058 MVT::ValueType ObjectVT = Op.getValue(i).getValueType();
1059 unsigned ArgIncrement = 8;
1060 unsigned ObjSize = 0;
1061 unsigned ObjIntRegs = 0;
1062 unsigned ObjXMMRegs = 0;
1064 // FIXME: __int128 and long double support?
1065 HowToPassX86_64CCCArgument(ObjectVT, NumIntRegs, NumXMMRegs,
1066 ObjSize, ObjIntRegs, ObjXMMRegs);
1068 ArgIncrement = ObjSize;
1072 if (ObjIntRegs || ObjXMMRegs) {
1074 default: assert(0 && "Unhandled argument type!");
1079 TargetRegisterClass *RC = NULL;
1083 RC = X86::GR8RegisterClass;
1084 Reg = GPR8ArgRegs[NumIntRegs];
1087 RC = X86::GR16RegisterClass;
1088 Reg = GPR16ArgRegs[NumIntRegs];
1091 RC = X86::GR32RegisterClass;
1092 Reg = GPR32ArgRegs[NumIntRegs];
1095 RC = X86::GR64RegisterClass;
1096 Reg = GPR64ArgRegs[NumIntRegs];
1099 Reg = AddLiveIn(MF, Reg, RC);
1100 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1111 TargetRegisterClass *RC= (ObjectVT == MVT::f32) ?
1112 X86::FR32RegisterClass : ((ObjectVT == MVT::f64) ?
1113 X86::FR64RegisterClass : X86::VR128RegisterClass);
1114 Reg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs], RC);
1115 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1119 NumIntRegs += ObjIntRegs;
1120 NumXMMRegs += ObjXMMRegs;
1121 } else if (ObjSize) {
1122 // XMM arguments have to be aligned on 16-byte boundary.
1124 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1125 // Create the SelectionDAG nodes corresponding to a load from this
1127 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
1128 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1129 ArgValue = DAG.getLoad(Op.Val->getValueType(i), Root, FIN, NULL, 0);
1130 ArgOffset += ArgIncrement; // Move on to the next argument.
1133 ArgValues.push_back(ArgValue);
1136 // If the function takes variable number of arguments, make a frame index for
1137 // the start of the first vararg value... for expansion of llvm.va_start.
1139 // For X86-64, if there are vararg parameters that are passed via
1140 // registers, then we must store them to their spots on the stack so they
1141 // may be loaded by deferencing the result of va_next.
1142 VarArgsGPOffset = NumIntRegs * 8;
1143 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1144 VarArgsFrameIndex = MFI->CreateFixedObject(1, ArgOffset);
1145 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1147 // Store the integer parameter registers.
1148 std::vector<SDOperand> MemOps;
1149 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1150 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1151 DAG.getConstant(VarArgsGPOffset, getPointerTy()));
1152 for (; NumIntRegs != 6; ++NumIntRegs) {
1153 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1154 X86::GR64RegisterClass);
1155 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1156 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1157 MemOps.push_back(Store);
1158 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1159 DAG.getConstant(8, getPointerTy()));
1162 // Now store the XMM (fp + vector) parameter registers.
1163 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1164 DAG.getConstant(VarArgsFPOffset, getPointerTy()));
1165 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1166 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1167 X86::VR128RegisterClass);
1168 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1169 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1170 MemOps.push_back(Store);
1171 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1172 DAG.getConstant(16, getPointerTy()));
1174 if (!MemOps.empty())
1175 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1176 &MemOps[0], MemOps.size());
1179 ArgValues.push_back(Root);
1181 ReturnAddrIndex = 0; // No return address slot generated yet.
1182 BytesToPopOnReturn = 0; // Callee pops nothing.
1183 BytesCallerReserves = ArgOffset;
1185 // Return the new list of results.
1186 std::vector<MVT::ValueType> RetVTs(Op.Val->value_begin(),
1187 Op.Val->value_end());
1188 return DAG.getNode(ISD::MERGE_VALUES, RetVTs, &ArgValues[0],ArgValues.size());
1192 X86TargetLowering::LowerX86_64CCCCallTo(SDOperand Op, SelectionDAG &DAG) {
1193 SDOperand Chain = Op.getOperand(0);
1194 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1195 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1196 SDOperand Callee = Op.getOperand(4);
1197 MVT::ValueType RetVT= Op.Val->getValueType(0);
1198 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1200 // Count how many bytes are to be pushed on the stack.
1201 unsigned NumBytes = 0;
1202 unsigned NumIntRegs = 0; // Int regs used for parameter passing.
1203 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1205 static const unsigned GPR8ArgRegs[] = {
1206 X86::DIL, X86::SIL, X86::DL, X86::CL, X86::R8B, X86::R9B
1208 static const unsigned GPR16ArgRegs[] = {
1209 X86::DI, X86::SI, X86::DX, X86::CX, X86::R8W, X86::R9W
1211 static const unsigned GPR32ArgRegs[] = {
1212 X86::EDI, X86::ESI, X86::EDX, X86::ECX, X86::R8D, X86::R9D
1214 static const unsigned GPR64ArgRegs[] = {
1215 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1217 static const unsigned XMMArgRegs[] = {
1218 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1219 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1222 for (unsigned i = 0; i != NumOps; ++i) {
1223 SDOperand Arg = Op.getOperand(5+2*i);
1224 MVT::ValueType ArgVT = Arg.getValueType();
1227 default: assert(0 && "Unknown value type!");
1247 else if (ArgVT == MVT::f32 || ArgVT == MVT::f64)
1250 // XMM arguments have to be aligned on 16-byte boundary.
1251 NumBytes = ((NumBytes + 15) / 16) * 16;
1258 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1260 // Arguments go on the stack in reverse order, as specified by the ABI.
1261 unsigned ArgOffset = 0;
1264 std::vector<std::pair<unsigned, SDOperand> > RegsToPass;
1265 std::vector<SDOperand> MemOpChains;
1266 SDOperand StackPtr = DAG.getRegister(X86StackPtr, getPointerTy());
1267 for (unsigned i = 0; i != NumOps; ++i) {
1268 SDOperand Arg = Op.getOperand(5+2*i);
1269 MVT::ValueType ArgVT = Arg.getValueType();
1272 default: assert(0 && "Unexpected ValueType for argument!");
1277 if (NumIntRegs < 6) {
1281 case MVT::i8: Reg = GPR8ArgRegs[NumIntRegs]; break;
1282 case MVT::i16: Reg = GPR16ArgRegs[NumIntRegs]; break;
1283 case MVT::i32: Reg = GPR32ArgRegs[NumIntRegs]; break;
1284 case MVT::i64: Reg = GPR64ArgRegs[NumIntRegs]; break;
1286 RegsToPass.push_back(std::make_pair(Reg, Arg));
1289 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1290 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1291 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1303 if (NumXMMRegs < 8) {
1304 RegsToPass.push_back(std::make_pair(XMMArgRegs[NumXMMRegs], Arg));
1307 if (ArgVT != MVT::f32 && ArgVT != MVT::f64) {
1308 // XMM arguments have to be aligned on 16-byte boundary.
1309 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1311 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1312 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1313 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1314 if (ArgVT == MVT::f32 || ArgVT == MVT::f64)
1322 if (!MemOpChains.empty())
1323 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1324 &MemOpChains[0], MemOpChains.size());
1326 // Build a sequence of copy-to-reg nodes chained together with token chain
1327 // and flag operands which copy the outgoing args into registers.
1329 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1330 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1332 InFlag = Chain.getValue(1);
1336 // From AMD64 ABI document:
1337 // For calls that may call functions that use varargs or stdargs
1338 // (prototype-less calls or calls to functions containing ellipsis (...) in
1339 // the declaration) %al is used as hidden argument to specify the number
1340 // of SSE registers used. The contents of %al do not need to match exactly
1341 // the number of registers, but must be an ubound on the number of SSE
1342 // registers used and is in the range 0 - 8 inclusive.
1343 Chain = DAG.getCopyToReg(Chain, X86::AL,
1344 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1345 InFlag = Chain.getValue(1);
1348 // If the callee is a GlobalAddress node (quite common, every direct call is)
1349 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1350 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1351 // We should use extra load for direct calls to dllimported functions in
1353 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1354 getTargetMachine(), true))
1355 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1356 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1357 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1359 std::vector<MVT::ValueType> NodeTys;
1360 NodeTys.push_back(MVT::Other); // Returns a chain
1361 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
1362 std::vector<SDOperand> Ops;
1363 Ops.push_back(Chain);
1364 Ops.push_back(Callee);
1366 // Add argument registers to the end of the list so that they are known live
1368 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1369 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1370 RegsToPass[i].second.getValueType()));
1373 Ops.push_back(InFlag);
1375 // FIXME: Do not generate X86ISD::TAILCALL for now.
1376 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1377 NodeTys, &Ops[0], Ops.size());
1378 InFlag = Chain.getValue(1);
1381 NodeTys.push_back(MVT::Other); // Returns a chain
1382 if (RetVT != MVT::Other)
1383 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
1385 Ops.push_back(Chain);
1386 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1387 Ops.push_back(DAG.getConstant(0, getPointerTy()));
1388 Ops.push_back(InFlag);
1389 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1390 if (RetVT != MVT::Other)
1391 InFlag = Chain.getValue(1);
1393 std::vector<SDOperand> ResultVals;
1396 default: assert(0 && "Unknown value type to return!");
1397 case MVT::Other: break;
1399 Chain = DAG.getCopyFromReg(Chain, X86::AL, MVT::i8, InFlag).getValue(1);
1400 ResultVals.push_back(Chain.getValue(0));
1401 NodeTys.push_back(MVT::i8);
1404 Chain = DAG.getCopyFromReg(Chain, X86::AX, MVT::i16, InFlag).getValue(1);
1405 ResultVals.push_back(Chain.getValue(0));
1406 NodeTys.push_back(MVT::i16);
1409 Chain = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag).getValue(1);
1410 ResultVals.push_back(Chain.getValue(0));
1411 NodeTys.push_back(MVT::i32);
1414 if (Op.Val->getValueType(1) == MVT::i64) {
1415 // FIXME: __int128 support?
1416 Chain = DAG.getCopyFromReg(Chain, X86::RAX, MVT::i64, InFlag).getValue(1);
1417 ResultVals.push_back(Chain.getValue(0));
1418 Chain = DAG.getCopyFromReg(Chain, X86::RDX, MVT::i64,
1419 Chain.getValue(2)).getValue(1);
1420 ResultVals.push_back(Chain.getValue(0));
1421 NodeTys.push_back(MVT::i64);
1423 Chain = DAG.getCopyFromReg(Chain, X86::RAX, MVT::i64, InFlag).getValue(1);
1424 ResultVals.push_back(Chain.getValue(0));
1426 NodeTys.push_back(MVT::i64);
1436 // FIXME: long double support?
1437 Chain = DAG.getCopyFromReg(Chain, X86::XMM0, RetVT, InFlag).getValue(1);
1438 ResultVals.push_back(Chain.getValue(0));
1439 NodeTys.push_back(RetVT);
1443 // If the function returns void, just return the chain.
1444 if (ResultVals.empty())
1447 // Otherwise, merge everything together with a MERGE_VALUES node.
1448 NodeTys.push_back(MVT::Other);
1449 ResultVals.push_back(Chain);
1450 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys,
1451 &ResultVals[0], ResultVals.size());
1452 return Res.getValue(Op.ResNo);
1455 //===----------------------------------------------------------------------===//
1456 // Fast & FastCall Calling Convention implementation
1457 //===----------------------------------------------------------------------===//
1459 // The X86 'fast' calling convention passes up to two integer arguments in
1460 // registers (an appropriate portion of EAX/EDX), passes arguments in C order,
1461 // and requires that the callee pop its arguments off the stack (allowing proper
1462 // tail calls), and has the same return value conventions as C calling convs.
1464 // This calling convention always arranges for the callee pop value to be 8n+4
1465 // bytes, which is needed for tail recursion elimination and stack alignment
1468 // Note that this can be enhanced in the future to pass fp vals in registers
1469 // (when we have a global fp allocator) and do other tricks.
1471 //===----------------------------------------------------------------------===//
1472 // The X86 'fastcall' calling convention passes up to two integer arguments in
1473 // registers (an appropriate portion of ECX/EDX), passes arguments in C order,
1474 // and requires that the callee pop its arguments off the stack (allowing proper
1475 // tail calls), and has the same return value conventions as C calling convs.
1477 // This calling convention always arranges for the callee pop value to be 8n+4
1478 // bytes, which is needed for tail recursion elimination and stack alignment
1483 X86TargetLowering::LowerFastCCArguments(SDOperand Op, SelectionDAG &DAG,
1485 unsigned NumArgs = Op.Val->getNumValues()-1;
1486 MachineFunction &MF = DAG.getMachineFunction();
1487 MachineFrameInfo *MFI = MF.getFrameInfo();
1488 SDOperand Root = Op.getOperand(0);
1489 std::vector<SDOperand> ArgValues;
1491 // Add DAG nodes to load the arguments... On entry to a function the stack
1492 // frame looks like this:
1494 // [ESP] -- return address
1495 // [ESP + 4] -- first nonreg argument (leftmost lexically)
1496 // [ESP + 8] -- second nonreg argument, if 1st argument is <= 4 bytes in size
1498 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
1500 // Keep track of the number of integer regs passed so far. This can be either
1501 // 0 (neither EAX/ECX or EDX used), 1 (EAX/ECX is used) or 2 (EAX/ECX and EDX
1503 unsigned NumIntRegs = 0;
1504 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1506 static const unsigned XMMArgRegs[] = {
1507 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1510 static const unsigned GPRArgRegs[][2][2] = {
1511 {{ X86::AL, X86::DL }, { X86::CL, X86::DL }},
1512 {{ X86::AX, X86::DX }, { X86::CX, X86::DX }},
1513 {{ X86::EAX, X86::EDX }, { X86::ECX, X86::EDX }}
1516 static const TargetRegisterClass* GPRClasses[3] = {
1517 X86::GR8RegisterClass, X86::GR16RegisterClass, X86::GR32RegisterClass
1520 unsigned GPRInd = (isFastCall ? 1 : 0);
1521 for (unsigned i = 0; i < NumArgs; ++i) {
1522 MVT::ValueType ObjectVT = Op.getValue(i).getValueType();
1523 unsigned ArgIncrement = 4;
1524 unsigned ObjSize = 0;
1525 unsigned ObjXMMRegs = 0;
1526 unsigned ObjIntRegs = 0;
1530 HowToPassCallArgument(ObjectVT,
1531 true, // Use as much registers as possible
1532 NumIntRegs, NumXMMRegs,
1533 (isFastCall ? 2 : FASTCC_NUM_INT_ARGS_INREGS),
1534 ObjSize, ObjIntRegs, ObjXMMRegs,
1538 ArgIncrement = ObjSize;
1540 if (ObjIntRegs || ObjXMMRegs) {
1542 default: assert(0 && "Unhandled argument type!");
1546 unsigned RegToUse = GPRArgRegs[ObjectVT-MVT::i8][GPRInd][NumIntRegs];
1547 Reg = AddLiveIn(MF, RegToUse, GPRClasses[ObjectVT-MVT::i8]);
1548 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1557 assert(!isFastCall && "Unhandled argument type!");
1558 Reg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs], X86::VR128RegisterClass);
1559 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1563 NumIntRegs += ObjIntRegs;
1564 NumXMMRegs += ObjXMMRegs;
1567 // XMM arguments have to be aligned on 16-byte boundary.
1569 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1570 // Create the SelectionDAG nodes corresponding to a load from this
1572 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
1573 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1574 ArgValue = DAG.getLoad(Op.Val->getValueType(i), Root, FIN, NULL, 0);
1576 ArgOffset += ArgIncrement; // Move on to the next argument.
1579 ArgValues.push_back(ArgValue);
1582 ArgValues.push_back(Root);
1584 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1585 // arguments and the arguments after the retaddr has been pushed are aligned.
1586 if ((ArgOffset & 7) == 0)
1589 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1590 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
1591 ReturnAddrIndex = 0; // No return address slot generated yet.
1592 BytesToPopOnReturn = ArgOffset; // Callee pops all stack arguments.
1593 BytesCallerReserves = 0;
1595 MF.getInfo<X86FunctionInfo>()->setBytesToPopOnReturn(BytesToPopOnReturn);
1597 // Finally, inform the code generator which regs we return values in.
1598 switch (getValueType(MF.getFunction()->getReturnType())) {
1599 default: assert(0 && "Unknown type!");
1600 case MVT::isVoid: break;
1605 MF.addLiveOut(X86::EAX);
1608 MF.addLiveOut(X86::EAX);
1609 MF.addLiveOut(X86::EDX);
1613 MF.addLiveOut(X86::ST0);
1621 assert(!isFastCall && "Unknown result type");
1622 MF.addLiveOut(X86::XMM0);
1626 // Return the new list of results.
1627 std::vector<MVT::ValueType> RetVTs(Op.Val->value_begin(),
1628 Op.Val->value_end());
1629 return DAG.getNode(ISD::MERGE_VALUES, RetVTs, &ArgValues[0],ArgValues.size());
1632 SDOperand X86TargetLowering::LowerFastCCCallTo(SDOperand Op, SelectionDAG &DAG,
1634 SDOperand Chain = Op.getOperand(0);
1635 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1636 SDOperand Callee = Op.getOperand(4);
1637 MVT::ValueType RetVT= Op.Val->getValueType(0);
1638 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1640 // Count how many bytes are to be pushed on the stack.
1641 unsigned NumBytes = 0;
1643 // Keep track of the number of integer regs passed so far. This can be either
1644 // 0 (neither EAX/ECX or EDX used), 1 (EAX/ECX is used) or 2 (EAX/ECX and EDX
1646 unsigned NumIntRegs = 0;
1647 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1649 static const unsigned GPRArgRegs[][2][2] = {
1650 {{ X86::AL, X86::DL }, { X86::CL, X86::DL }},
1651 {{ X86::AX, X86::DX }, { X86::CX, X86::DX }},
1652 {{ X86::EAX, X86::EDX }, { X86::ECX, X86::EDX }}
1654 static const unsigned XMMArgRegs[] = {
1655 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1658 unsigned GPRInd = (isFastCall ? 1 : 0);
1659 for (unsigned i = 0; i != NumOps; ++i) {
1660 SDOperand Arg = Op.getOperand(5+2*i);
1662 switch (Arg.getValueType()) {
1663 default: assert(0 && "Unknown value type!");
1667 unsigned MaxNumIntRegs = (isFastCall ? 2 : FASTCC_NUM_INT_ARGS_INREGS);
1668 if (NumIntRegs < MaxNumIntRegs) {
1685 assert(!isFastCall && "Unknown value type!");
1689 // XMM arguments have to be aligned on 16-byte boundary.
1690 NumBytes = ((NumBytes + 15) / 16) * 16;
1697 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1698 // arguments and the arguments after the retaddr has been pushed are aligned.
1699 if ((NumBytes & 7) == 0)
1702 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1704 // Arguments go on the stack in reverse order, as specified by the ABI.
1705 unsigned ArgOffset = 0;
1707 std::vector<std::pair<unsigned, SDOperand> > RegsToPass;
1708 std::vector<SDOperand> MemOpChains;
1709 SDOperand StackPtr = DAG.getRegister(X86StackPtr, getPointerTy());
1710 for (unsigned i = 0; i != NumOps; ++i) {
1711 SDOperand Arg = Op.getOperand(5+2*i);
1713 switch (Arg.getValueType()) {
1714 default: assert(0 && "Unexpected ValueType for argument!");
1718 unsigned MaxNumIntRegs = (isFastCall ? 2 : FASTCC_NUM_INT_ARGS_INREGS);
1719 if (NumIntRegs < MaxNumIntRegs) {
1721 GPRArgRegs[Arg.getValueType()-MVT::i8][GPRInd][NumIntRegs];
1722 RegsToPass.push_back(std::make_pair(RegToUse, Arg));
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));
1735 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1736 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1737 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1747 assert(!isFastCall && "Unexpected ValueType for argument!");
1748 if (NumXMMRegs < 4) {
1749 RegsToPass.push_back(std::make_pair(XMMArgRegs[NumXMMRegs], Arg));
1752 // XMM arguments have to be aligned on 16-byte boundary.
1753 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1754 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1755 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1756 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1763 if (!MemOpChains.empty())
1764 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1765 &MemOpChains[0], MemOpChains.size());
1767 // Build a sequence of copy-to-reg nodes chained together with token chain
1768 // and flag operands which copy the outgoing args into registers.
1770 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1771 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1773 InFlag = Chain.getValue(1);
1776 // If the callee is a GlobalAddress node (quite common, every direct call is)
1777 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1778 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1779 // We should use extra load for direct calls to dllimported functions in
1781 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1782 getTargetMachine(), true))
1783 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1784 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1785 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1787 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1788 Subtarget->isPICStyleGOT()) {
1789 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1790 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1792 InFlag = Chain.getValue(1);
1795 std::vector<MVT::ValueType> NodeTys;
1796 NodeTys.push_back(MVT::Other); // Returns a chain
1797 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
1798 std::vector<SDOperand> Ops;
1799 Ops.push_back(Chain);
1800 Ops.push_back(Callee);
1802 // Add argument registers to the end of the list so that they are known live
1804 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1805 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1806 RegsToPass[i].second.getValueType()));
1809 Ops.push_back(InFlag);
1811 // FIXME: Do not generate X86ISD::TAILCALL for now.
1812 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1813 NodeTys, &Ops[0], Ops.size());
1814 InFlag = Chain.getValue(1);
1817 NodeTys.push_back(MVT::Other); // Returns a chain
1818 if (RetVT != MVT::Other)
1819 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
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 std::vector<SDOperand> ResultVals;
1832 default: assert(0 && "Unknown value type to return!");
1833 case MVT::Other: break;
1835 Chain = DAG.getCopyFromReg(Chain, X86::AL, MVT::i8, InFlag).getValue(1);
1836 ResultVals.push_back(Chain.getValue(0));
1837 NodeTys.push_back(MVT::i8);
1840 Chain = DAG.getCopyFromReg(Chain, X86::AX, MVT::i16, InFlag).getValue(1);
1841 ResultVals.push_back(Chain.getValue(0));
1842 NodeTys.push_back(MVT::i16);
1845 if (Op.Val->getValueType(1) == MVT::i32) {
1846 Chain = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag).getValue(1);
1847 ResultVals.push_back(Chain.getValue(0));
1848 Chain = DAG.getCopyFromReg(Chain, X86::EDX, MVT::i32,
1849 Chain.getValue(2)).getValue(1);
1850 ResultVals.push_back(Chain.getValue(0));
1851 NodeTys.push_back(MVT::i32);
1853 Chain = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag).getValue(1);
1854 ResultVals.push_back(Chain.getValue(0));
1856 NodeTys.push_back(MVT::i32);
1865 assert(0 && "Unknown value type to return!");
1867 Chain = DAG.getCopyFromReg(Chain, X86::XMM0, RetVT, InFlag).getValue(1);
1868 ResultVals.push_back(Chain.getValue(0));
1869 NodeTys.push_back(RetVT);
1874 std::vector<MVT::ValueType> Tys;
1875 Tys.push_back(MVT::f64);
1876 Tys.push_back(MVT::Other);
1877 Tys.push_back(MVT::Flag);
1878 std::vector<SDOperand> Ops;
1879 Ops.push_back(Chain);
1880 Ops.push_back(InFlag);
1881 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys,
1882 &Ops[0], Ops.size());
1883 Chain = RetVal.getValue(1);
1884 InFlag = RetVal.getValue(2);
1886 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
1887 // shouldn't be necessary except that RFP cannot be live across
1888 // multiple blocks. When stackifier is fixed, they can be uncoupled.
1889 MachineFunction &MF = DAG.getMachineFunction();
1890 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
1891 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
1893 Tys.push_back(MVT::Other);
1895 Ops.push_back(Chain);
1896 Ops.push_back(RetVal);
1897 Ops.push_back(StackSlot);
1898 Ops.push_back(DAG.getValueType(RetVT));
1899 Ops.push_back(InFlag);
1900 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
1901 RetVal = DAG.getLoad(RetVT, Chain, StackSlot, NULL, 0);
1902 Chain = RetVal.getValue(1);
1905 if (RetVT == MVT::f32 && !X86ScalarSSE)
1906 // FIXME: we would really like to remember that this FP_ROUND
1907 // operation is okay to eliminate if we allow excess FP precision.
1908 RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal);
1909 ResultVals.push_back(RetVal);
1910 NodeTys.push_back(RetVT);
1916 // If the function returns void, just return the chain.
1917 if (ResultVals.empty())
1920 // Otherwise, merge everything together with a MERGE_VALUES node.
1921 NodeTys.push_back(MVT::Other);
1922 ResultVals.push_back(Chain);
1923 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys,
1924 &ResultVals[0], ResultVals.size());
1925 return Res.getValue(Op.ResNo);
1928 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1929 if (ReturnAddrIndex == 0) {
1930 // Set up a frame object for the return address.
1931 MachineFunction &MF = DAG.getMachineFunction();
1932 if (Subtarget->is64Bit())
1933 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
1935 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
1938 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1943 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1944 /// specific condition code. It returns a false if it cannot do a direct
1945 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1947 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1948 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
1949 SelectionDAG &DAG) {
1950 X86CC = X86::COND_INVALID;
1952 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1953 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1954 // X > -1 -> X == 0, jump !sign.
1955 RHS = DAG.getConstant(0, RHS.getValueType());
1956 X86CC = X86::COND_NS;
1958 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1959 // X < 0 -> X == 0, jump on sign.
1960 X86CC = X86::COND_S;
1965 switch (SetCCOpcode) {
1967 case ISD::SETEQ: X86CC = X86::COND_E; break;
1968 case ISD::SETGT: X86CC = X86::COND_G; break;
1969 case ISD::SETGE: X86CC = X86::COND_GE; break;
1970 case ISD::SETLT: X86CC = X86::COND_L; break;
1971 case ISD::SETLE: X86CC = X86::COND_LE; break;
1972 case ISD::SETNE: X86CC = X86::COND_NE; break;
1973 case ISD::SETULT: X86CC = X86::COND_B; break;
1974 case ISD::SETUGT: X86CC = X86::COND_A; break;
1975 case ISD::SETULE: X86CC = X86::COND_BE; break;
1976 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1979 // On a floating point condition, the flags are set as follows:
1981 // 0 | 0 | 0 | X > Y
1982 // 0 | 0 | 1 | X < Y
1983 // 1 | 0 | 0 | X == Y
1984 // 1 | 1 | 1 | unordered
1986 switch (SetCCOpcode) {
1989 case ISD::SETEQ: X86CC = X86::COND_E; break;
1990 case ISD::SETOLT: Flip = true; // Fallthrough
1992 case ISD::SETGT: X86CC = X86::COND_A; break;
1993 case ISD::SETOLE: Flip = true; // Fallthrough
1995 case ISD::SETGE: X86CC = X86::COND_AE; break;
1996 case ISD::SETUGT: Flip = true; // Fallthrough
1998 case ISD::SETLT: X86CC = X86::COND_B; break;
1999 case ISD::SETUGE: Flip = true; // Fallthrough
2001 case ISD::SETLE: X86CC = X86::COND_BE; break;
2003 case ISD::SETNE: X86CC = X86::COND_NE; break;
2004 case ISD::SETUO: X86CC = X86::COND_P; break;
2005 case ISD::SETO: X86CC = X86::COND_NP; break;
2008 std::swap(LHS, RHS);
2011 return X86CC != X86::COND_INVALID;
2014 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2015 /// code. Current x86 isa includes the following FP cmov instructions:
2016 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2017 static bool hasFPCMov(unsigned X86CC) {
2033 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
2034 /// true if Op is undef or if its value falls within the specified range (L, H].
2035 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
2036 if (Op.getOpcode() == ISD::UNDEF)
2039 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
2040 return (Val >= Low && Val < Hi);
2043 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
2044 /// true if Op is undef or if its value equal to the specified value.
2045 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
2046 if (Op.getOpcode() == ISD::UNDEF)
2048 return cast<ConstantSDNode>(Op)->getValue() == Val;
2051 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
2052 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
2053 bool X86::isPSHUFDMask(SDNode *N) {
2054 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2056 if (N->getNumOperands() != 4)
2059 // Check if the value doesn't reference the second vector.
2060 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
2061 SDOperand Arg = N->getOperand(i);
2062 if (Arg.getOpcode() == ISD::UNDEF) continue;
2063 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2064 if (cast<ConstantSDNode>(Arg)->getValue() >= 4)
2071 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
2072 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
2073 bool X86::isPSHUFHWMask(SDNode *N) {
2074 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2076 if (N->getNumOperands() != 8)
2079 // Lower quadword copied in order.
2080 for (unsigned i = 0; i != 4; ++i) {
2081 SDOperand Arg = N->getOperand(i);
2082 if (Arg.getOpcode() == ISD::UNDEF) continue;
2083 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2084 if (cast<ConstantSDNode>(Arg)->getValue() != i)
2088 // Upper quadword shuffled.
2089 for (unsigned i = 4; i != 8; ++i) {
2090 SDOperand Arg = N->getOperand(i);
2091 if (Arg.getOpcode() == ISD::UNDEF) continue;
2092 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2093 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2094 if (Val < 4 || Val > 7)
2101 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
2102 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
2103 bool X86::isPSHUFLWMask(SDNode *N) {
2104 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2106 if (N->getNumOperands() != 8)
2109 // Upper quadword copied in order.
2110 for (unsigned i = 4; i != 8; ++i)
2111 if (!isUndefOrEqual(N->getOperand(i), i))
2114 // Lower quadword shuffled.
2115 for (unsigned i = 0; i != 4; ++i)
2116 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
2122 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2123 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2124 static bool isSHUFPMask(std::vector<SDOperand> &N) {
2125 unsigned NumElems = N.size();
2126 if (NumElems != 2 && NumElems != 4) return false;
2128 unsigned Half = NumElems / 2;
2129 for (unsigned i = 0; i < Half; ++i)
2130 if (!isUndefOrInRange(N[i], 0, NumElems))
2132 for (unsigned i = Half; i < NumElems; ++i)
2133 if (!isUndefOrInRange(N[i], NumElems, NumElems*2))
2139 bool X86::isSHUFPMask(SDNode *N) {
2140 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2141 std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
2142 return ::isSHUFPMask(Ops);
2145 /// isCommutedSHUFP - Returns true if the shuffle mask is except
2146 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2147 /// half elements to come from vector 1 (which would equal the dest.) and
2148 /// the upper half to come from vector 2.
2149 static bool isCommutedSHUFP(std::vector<SDOperand> &Ops) {
2150 unsigned NumElems = Ops.size();
2151 if (NumElems != 2 && NumElems != 4) return false;
2153 unsigned Half = NumElems / 2;
2154 for (unsigned i = 0; i < Half; ++i)
2155 if (!isUndefOrInRange(Ops[i], NumElems, NumElems*2))
2157 for (unsigned i = Half; i < NumElems; ++i)
2158 if (!isUndefOrInRange(Ops[i], 0, NumElems))
2163 static bool isCommutedSHUFP(SDNode *N) {
2164 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2165 std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
2166 return isCommutedSHUFP(Ops);
2169 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2170 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2171 bool X86::isMOVHLPSMask(SDNode *N) {
2172 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2174 if (N->getNumOperands() != 4)
2177 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2178 return isUndefOrEqual(N->getOperand(0), 6) &&
2179 isUndefOrEqual(N->getOperand(1), 7) &&
2180 isUndefOrEqual(N->getOperand(2), 2) &&
2181 isUndefOrEqual(N->getOperand(3), 3);
2184 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2185 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2187 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
2188 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2190 if (N->getNumOperands() != 4)
2193 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
2194 return isUndefOrEqual(N->getOperand(0), 2) &&
2195 isUndefOrEqual(N->getOperand(1), 3) &&
2196 isUndefOrEqual(N->getOperand(2), 2) &&
2197 isUndefOrEqual(N->getOperand(3), 3);
2200 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2201 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2202 bool X86::isMOVLPMask(SDNode *N) {
2203 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2205 unsigned NumElems = N->getNumOperands();
2206 if (NumElems != 2 && NumElems != 4)
2209 for (unsigned i = 0; i < NumElems/2; ++i)
2210 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
2213 for (unsigned i = NumElems/2; i < NumElems; ++i)
2214 if (!isUndefOrEqual(N->getOperand(i), i))
2220 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2221 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2223 bool X86::isMOVHPMask(SDNode *N) {
2224 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2226 unsigned NumElems = N->getNumOperands();
2227 if (NumElems != 2 && NumElems != 4)
2230 for (unsigned i = 0; i < NumElems/2; ++i)
2231 if (!isUndefOrEqual(N->getOperand(i), i))
2234 for (unsigned i = 0; i < NumElems/2; ++i) {
2235 SDOperand Arg = N->getOperand(i + NumElems/2);
2236 if (!isUndefOrEqual(Arg, i + NumElems))
2243 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2244 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2245 bool static isUNPCKLMask(std::vector<SDOperand> &N, bool V2IsSplat = false) {
2246 unsigned NumElems = N.size();
2247 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2250 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2251 SDOperand BitI = N[i];
2252 SDOperand BitI1 = N[i+1];
2253 if (!isUndefOrEqual(BitI, j))
2256 if (isUndefOrEqual(BitI1, NumElems))
2259 if (!isUndefOrEqual(BitI1, j + NumElems))
2267 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
2268 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2269 std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
2270 return ::isUNPCKLMask(Ops, V2IsSplat);
2273 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2274 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2275 bool static isUNPCKHMask(std::vector<SDOperand> &N, bool V2IsSplat = false) {
2276 unsigned NumElems = N.size();
2277 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2280 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2281 SDOperand BitI = N[i];
2282 SDOperand BitI1 = N[i+1];
2283 if (!isUndefOrEqual(BitI, j + NumElems/2))
2286 if (isUndefOrEqual(BitI1, NumElems))
2289 if (!isUndefOrEqual(BitI1, j + NumElems/2 + NumElems))
2297 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
2298 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2299 std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
2300 return ::isUNPCKHMask(Ops, V2IsSplat);
2303 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2304 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2306 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
2307 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2309 unsigned NumElems = N->getNumOperands();
2310 if (NumElems != 4 && NumElems != 8 && NumElems != 16)
2313 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
2314 SDOperand BitI = N->getOperand(i);
2315 SDOperand BitI1 = N->getOperand(i+1);
2317 if (!isUndefOrEqual(BitI, j))
2319 if (!isUndefOrEqual(BitI1, j))
2326 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2327 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2328 /// MOVSD, and MOVD, i.e. setting the lowest element.
2329 static bool isMOVLMask(std::vector<SDOperand> &N) {
2330 unsigned NumElems = N.size();
2331 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2334 if (!isUndefOrEqual(N[0], NumElems))
2337 for (unsigned i = 1; i < NumElems; ++i) {
2338 SDOperand Arg = N[i];
2339 if (!isUndefOrEqual(Arg, i))
2346 bool X86::isMOVLMask(SDNode *N) {
2347 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2348 std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
2349 return ::isMOVLMask(Ops);
2352 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2353 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2354 /// element of vector 2 and the other elements to come from vector 1 in order.
2355 static bool isCommutedMOVL(std::vector<SDOperand> &Ops, bool V2IsSplat = false,
2356 bool V2IsUndef = false) {
2357 unsigned NumElems = Ops.size();
2358 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2361 if (!isUndefOrEqual(Ops[0], 0))
2364 for (unsigned i = 1; i < NumElems; ++i) {
2365 SDOperand Arg = Ops[i];
2366 if (!(isUndefOrEqual(Arg, i+NumElems) ||
2367 (V2IsUndef && isUndefOrInRange(Arg, NumElems, NumElems*2)) ||
2368 (V2IsSplat && isUndefOrEqual(Arg, NumElems))))
2375 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
2376 bool V2IsUndef = false) {
2377 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2378 std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
2379 return isCommutedMOVL(Ops, V2IsSplat, V2IsUndef);
2382 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2383 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2384 bool X86::isMOVSHDUPMask(SDNode *N) {
2385 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2387 if (N->getNumOperands() != 4)
2390 // Expect 1, 1, 3, 3
2391 for (unsigned i = 0; i < 2; ++i) {
2392 SDOperand Arg = N->getOperand(i);
2393 if (Arg.getOpcode() == ISD::UNDEF) continue;
2394 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2395 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2396 if (Val != 1) return false;
2400 for (unsigned i = 2; i < 4; ++i) {
2401 SDOperand Arg = N->getOperand(i);
2402 if (Arg.getOpcode() == ISD::UNDEF) continue;
2403 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2404 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2405 if (Val != 3) return false;
2409 // Don't use movshdup if it can be done with a shufps.
2413 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2414 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2415 bool X86::isMOVSLDUPMask(SDNode *N) {
2416 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2418 if (N->getNumOperands() != 4)
2421 // Expect 0, 0, 2, 2
2422 for (unsigned i = 0; i < 2; ++i) {
2423 SDOperand Arg = N->getOperand(i);
2424 if (Arg.getOpcode() == ISD::UNDEF) continue;
2425 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2426 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2427 if (Val != 0) return false;
2431 for (unsigned i = 2; i < 4; ++i) {
2432 SDOperand Arg = N->getOperand(i);
2433 if (Arg.getOpcode() == ISD::UNDEF) continue;
2434 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2435 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2436 if (Val != 2) return false;
2440 // Don't use movshdup if it can be done with a shufps.
2444 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2445 /// a splat of a single element.
2446 static bool isSplatMask(SDNode *N) {
2447 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2449 // This is a splat operation if each element of the permute is the same, and
2450 // if the value doesn't reference the second vector.
2451 unsigned NumElems = N->getNumOperands();
2452 SDOperand ElementBase;
2454 for (; i != NumElems; ++i) {
2455 SDOperand Elt = N->getOperand(i);
2456 if (isa<ConstantSDNode>(Elt)) {
2462 if (!ElementBase.Val)
2465 for (; i != NumElems; ++i) {
2466 SDOperand Arg = N->getOperand(i);
2467 if (Arg.getOpcode() == ISD::UNDEF) continue;
2468 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2469 if (Arg != ElementBase) return false;
2472 // Make sure it is a splat of the first vector operand.
2473 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2476 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2477 /// a splat of a single element and it's a 2 or 4 element mask.
2478 bool X86::isSplatMask(SDNode *N) {
2479 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2481 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2482 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2484 return ::isSplatMask(N);
2487 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2488 /// specifies a splat of zero element.
2489 bool X86::isSplatLoMask(SDNode *N) {
2490 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2492 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2493 if (!isUndefOrEqual(N->getOperand(i), 0))
2498 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2499 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2501 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2502 unsigned NumOperands = N->getNumOperands();
2503 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2505 for (unsigned i = 0; i < NumOperands; ++i) {
2507 SDOperand Arg = N->getOperand(NumOperands-i-1);
2508 if (Arg.getOpcode() != ISD::UNDEF)
2509 Val = cast<ConstantSDNode>(Arg)->getValue();
2510 if (Val >= NumOperands) Val -= NumOperands;
2512 if (i != NumOperands - 1)
2519 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2520 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2522 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2524 // 8 nodes, but we only care about the last 4.
2525 for (unsigned i = 7; i >= 4; --i) {
2527 SDOperand Arg = N->getOperand(i);
2528 if (Arg.getOpcode() != ISD::UNDEF)
2529 Val = cast<ConstantSDNode>(Arg)->getValue();
2538 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2539 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2541 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2543 // 8 nodes, but we only care about the first 4.
2544 for (int i = 3; i >= 0; --i) {
2546 SDOperand Arg = N->getOperand(i);
2547 if (Arg.getOpcode() != ISD::UNDEF)
2548 Val = cast<ConstantSDNode>(Arg)->getValue();
2557 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2558 /// specifies a 8 element shuffle that can be broken into a pair of
2559 /// PSHUFHW and PSHUFLW.
2560 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2561 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2563 if (N->getNumOperands() != 8)
2566 // Lower quadword shuffled.
2567 for (unsigned i = 0; i != 4; ++i) {
2568 SDOperand Arg = N->getOperand(i);
2569 if (Arg.getOpcode() == ISD::UNDEF) continue;
2570 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2571 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2576 // Upper quadword shuffled.
2577 for (unsigned i = 4; i != 8; ++i) {
2578 SDOperand Arg = N->getOperand(i);
2579 if (Arg.getOpcode() == ISD::UNDEF) continue;
2580 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2581 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2582 if (Val < 4 || Val > 7)
2589 /// CommuteVectorShuffle - Swap vector_shuffle operandsas well as
2590 /// values in ther permute mask.
2591 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2592 SDOperand &V2, SDOperand &Mask,
2593 SelectionDAG &DAG) {
2594 MVT::ValueType VT = Op.getValueType();
2595 MVT::ValueType MaskVT = Mask.getValueType();
2596 MVT::ValueType EltVT = MVT::getVectorBaseType(MaskVT);
2597 unsigned NumElems = Mask.getNumOperands();
2598 std::vector<SDOperand> MaskVec;
2600 for (unsigned i = 0; i != NumElems; ++i) {
2601 SDOperand Arg = Mask.getOperand(i);
2602 if (Arg.getOpcode() == ISD::UNDEF) {
2603 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2606 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2607 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2609 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2611 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2615 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2616 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2619 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2620 /// match movhlps. The lower half elements should come from upper half of
2621 /// V1 (and in order), and the upper half elements should come from the upper
2622 /// half of V2 (and in order).
2623 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2624 unsigned NumElems = Mask->getNumOperands();
2627 for (unsigned i = 0, e = 2; i != e; ++i)
2628 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2630 for (unsigned i = 2; i != 4; ++i)
2631 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2636 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2637 /// is promoted to a vector.
2638 static inline bool isScalarLoadToVector(SDNode *N) {
2639 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2640 N = N->getOperand(0).Val;
2641 return ISD::isNON_EXTLoad(N);
2646 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2647 /// match movlp{s|d}. The lower half elements should come from lower half of
2648 /// V1 (and in order), and the upper half elements should come from the upper
2649 /// half of V2 (and in order). And since V1 will become the source of the
2650 /// MOVLP, it must be either a vector load or a scalar load to vector.
2651 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2652 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2654 // Is V2 is a vector load, don't do this transformation. We will try to use
2655 // load folding shufps op.
2656 if (ISD::isNON_EXTLoad(V2))
2659 unsigned NumElems = Mask->getNumOperands();
2660 if (NumElems != 2 && NumElems != 4)
2662 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2663 if (!isUndefOrEqual(Mask->getOperand(i), i))
2665 for (unsigned i = NumElems/2; i != NumElems; ++i)
2666 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2671 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2673 static bool isSplatVector(SDNode *N) {
2674 if (N->getOpcode() != ISD::BUILD_VECTOR)
2677 SDOperand SplatValue = N->getOperand(0);
2678 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2679 if (N->getOperand(i) != SplatValue)
2684 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2686 static bool isUndefShuffle(SDNode *N) {
2687 if (N->getOpcode() != ISD::BUILD_VECTOR)
2690 SDOperand V1 = N->getOperand(0);
2691 SDOperand V2 = N->getOperand(1);
2692 SDOperand Mask = N->getOperand(2);
2693 unsigned NumElems = Mask.getNumOperands();
2694 for (unsigned i = 0; i != NumElems; ++i) {
2695 SDOperand Arg = Mask.getOperand(i);
2696 if (Arg.getOpcode() != ISD::UNDEF) {
2697 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2698 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2700 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2707 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2708 /// that point to V2 points to its first element.
2709 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2710 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2712 bool Changed = false;
2713 std::vector<SDOperand> MaskVec;
2714 unsigned NumElems = Mask.getNumOperands();
2715 for (unsigned i = 0; i != NumElems; ++i) {
2716 SDOperand Arg = Mask.getOperand(i);
2717 if (Arg.getOpcode() != ISD::UNDEF) {
2718 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2719 if (Val > NumElems) {
2720 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2724 MaskVec.push_back(Arg);
2728 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2729 &MaskVec[0], MaskVec.size());
2733 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2734 /// operation of specified width.
2735 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2736 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2737 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2739 std::vector<SDOperand> MaskVec;
2740 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2741 for (unsigned i = 1; i != NumElems; ++i)
2742 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2743 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2746 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2747 /// of specified width.
2748 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2749 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2750 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2751 std::vector<SDOperand> MaskVec;
2752 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2753 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2754 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2756 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2759 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2760 /// of specified width.
2761 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2762 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2763 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2764 unsigned Half = NumElems/2;
2765 std::vector<SDOperand> MaskVec;
2766 for (unsigned i = 0; i != Half; ++i) {
2767 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2768 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2770 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2773 /// getZeroVector - Returns a vector of specified type with all zero elements.
2775 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2776 assert(MVT::isVector(VT) && "Expected a vector type");
2777 unsigned NumElems = getVectorNumElements(VT);
2778 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2779 bool isFP = MVT::isFloatingPoint(EVT);
2780 SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT);
2781 std::vector<SDOperand> ZeroVec(NumElems, Zero);
2782 return DAG.getNode(ISD::BUILD_VECTOR, VT, &ZeroVec[0], ZeroVec.size());
2785 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2787 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2788 SDOperand V1 = Op.getOperand(0);
2789 SDOperand Mask = Op.getOperand(2);
2790 MVT::ValueType VT = Op.getValueType();
2791 unsigned NumElems = Mask.getNumOperands();
2792 Mask = getUnpacklMask(NumElems, DAG);
2793 while (NumElems != 4) {
2794 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2797 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2799 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2800 Mask = getZeroVector(MaskVT, DAG);
2801 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2802 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2803 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2806 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2808 static inline bool isZeroNode(SDOperand Elt) {
2809 return ((isa<ConstantSDNode>(Elt) &&
2810 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2811 (isa<ConstantFPSDNode>(Elt) &&
2812 cast<ConstantFPSDNode>(Elt)->isExactlyValue(0.0)));
2815 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2816 /// vector and zero or undef vector.
2817 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2818 unsigned NumElems, unsigned Idx,
2819 bool isZero, SelectionDAG &DAG) {
2820 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2821 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2822 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2823 SDOperand Zero = DAG.getConstant(0, EVT);
2824 std::vector<SDOperand> MaskVec(NumElems, Zero);
2825 MaskVec[Idx] = DAG.getConstant(NumElems, EVT);
2826 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2827 &MaskVec[0], MaskVec.size());
2828 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2831 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2833 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
2834 unsigned NumNonZero, unsigned NumZero,
2835 SelectionDAG &DAG, TargetLowering &TLI) {
2841 for (unsigned i = 0; i < 16; ++i) {
2842 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
2843 if (ThisIsNonZero && First) {
2845 V = getZeroVector(MVT::v8i16, DAG);
2847 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2852 SDOperand ThisElt(0, 0), LastElt(0, 0);
2853 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
2854 if (LastIsNonZero) {
2855 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
2857 if (ThisIsNonZero) {
2858 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
2859 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
2860 ThisElt, DAG.getConstant(8, MVT::i8));
2862 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
2867 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
2868 DAG.getConstant(i/2, TLI.getPointerTy()));
2872 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
2875 /// LowerBuildVectorv16i8 - Custom lower build_vector of v8i16.
2877 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
2878 unsigned NumNonZero, unsigned NumZero,
2879 SelectionDAG &DAG, TargetLowering &TLI) {
2885 for (unsigned i = 0; i < 8; ++i) {
2886 bool isNonZero = (NonZeros & (1 << i)) != 0;
2890 V = getZeroVector(MVT::v8i16, DAG);
2892 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2895 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
2896 DAG.getConstant(i, TLI.getPointerTy()));
2904 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2905 // All zero's are handled with pxor.
2906 if (ISD::isBuildVectorAllZeros(Op.Val))
2909 // All one's are handled with pcmpeqd.
2910 if (ISD::isBuildVectorAllOnes(Op.Val))
2913 MVT::ValueType VT = Op.getValueType();
2914 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2915 unsigned EVTBits = MVT::getSizeInBits(EVT);
2917 unsigned NumElems = Op.getNumOperands();
2918 unsigned NumZero = 0;
2919 unsigned NumNonZero = 0;
2920 unsigned NonZeros = 0;
2921 std::set<SDOperand> Values;
2922 for (unsigned i = 0; i < NumElems; ++i) {
2923 SDOperand Elt = Op.getOperand(i);
2924 if (Elt.getOpcode() != ISD::UNDEF) {
2926 if (isZeroNode(Elt))
2929 NonZeros |= (1 << i);
2935 if (NumNonZero == 0)
2936 // Must be a mix of zero and undef. Return a zero vector.
2937 return getZeroVector(VT, DAG);
2939 // Splat is obviously ok. Let legalizer expand it to a shuffle.
2940 if (Values.size() == 1)
2943 // Special case for single non-zero element.
2944 if (NumNonZero == 1) {
2945 unsigned Idx = CountTrailingZeros_32(NonZeros);
2946 SDOperand Item = Op.getOperand(Idx);
2947 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
2949 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
2950 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
2953 if (EVTBits == 32) {
2954 // Turn it into a shuffle of zero and zero-extended scalar to vector.
2955 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
2957 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2958 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
2959 std::vector<SDOperand> MaskVec;
2960 for (unsigned i = 0; i < NumElems; i++)
2961 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
2962 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2963 &MaskVec[0], MaskVec.size());
2964 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
2965 DAG.getNode(ISD::UNDEF, VT), Mask);
2969 // Let legalizer expand 2-wide build_vector's.
2973 // If element VT is < 32 bits, convert it to inserts into a zero vector.
2975 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
2977 if (V.Val) return V;
2980 if (EVTBits == 16) {
2981 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
2983 if (V.Val) return V;
2986 // If element VT is == 32 bits, turn it into a number of shuffles.
2987 std::vector<SDOperand> V(NumElems);
2988 if (NumElems == 4 && NumZero > 0) {
2989 for (unsigned i = 0; i < 4; ++i) {
2990 bool isZero = !(NonZeros & (1 << i));
2992 V[i] = getZeroVector(VT, DAG);
2994 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2997 for (unsigned i = 0; i < 2; ++i) {
2998 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3001 V[i] = V[i*2]; // Must be a zero vector.
3004 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
3005 getMOVLMask(NumElems, DAG));
3008 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3009 getMOVLMask(NumElems, DAG));
3012 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
3013 getUnpacklMask(NumElems, DAG));
3018 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
3019 // clears the upper bits.
3020 // FIXME: we can do the same for v4f32 case when we know both parts of
3021 // the lower half come from scalar_to_vector (loadf32). We should do
3022 // that in post legalizer dag combiner with target specific hooks.
3023 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
3025 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3026 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
3027 std::vector<SDOperand> MaskVec;
3028 bool Reverse = (NonZeros & 0x3) == 2;
3029 for (unsigned i = 0; i < 2; ++i)
3031 MaskVec.push_back(DAG.getConstant(1-i, EVT));
3033 MaskVec.push_back(DAG.getConstant(i, EVT));
3034 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3035 for (unsigned i = 0; i < 2; ++i)
3037 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
3039 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
3040 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3041 &MaskVec[0], MaskVec.size());
3042 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
3045 if (Values.size() > 2) {
3046 // Expand into a number of unpckl*.
3048 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3049 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3050 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3051 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
3052 for (unsigned i = 0; i < NumElems; ++i)
3053 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
3055 while (NumElems != 0) {
3056 for (unsigned i = 0; i < NumElems; ++i)
3057 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
3068 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
3069 SDOperand V1 = Op.getOperand(0);
3070 SDOperand V2 = Op.getOperand(1);
3071 SDOperand PermMask = Op.getOperand(2);
3072 MVT::ValueType VT = Op.getValueType();
3073 unsigned NumElems = PermMask.getNumOperands();
3074 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
3075 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
3076 bool V1IsSplat = false;
3077 bool V2IsSplat = false;
3079 if (isUndefShuffle(Op.Val))
3080 return DAG.getNode(ISD::UNDEF, VT);
3082 if (isSplatMask(PermMask.Val)) {
3083 if (NumElems <= 4) return Op;
3084 // Promote it to a v4i32 splat.
3085 return PromoteSplat(Op, DAG);
3088 if (X86::isMOVLMask(PermMask.Val))
3089 return (V1IsUndef) ? V2 : Op;
3091 if (X86::isMOVSHDUPMask(PermMask.Val) ||
3092 X86::isMOVSLDUPMask(PermMask.Val) ||
3093 X86::isMOVHLPSMask(PermMask.Val) ||
3094 X86::isMOVHPMask(PermMask.Val) ||
3095 X86::isMOVLPMask(PermMask.Val))
3098 if (ShouldXformToMOVHLPS(PermMask.Val) ||
3099 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
3100 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3102 bool Commuted = false;
3103 V1IsSplat = isSplatVector(V1.Val);
3104 V2IsSplat = isSplatVector(V2.Val);
3105 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
3106 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3107 std::swap(V1IsSplat, V2IsSplat);
3108 std::swap(V1IsUndef, V2IsUndef);
3112 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
3113 if (V2IsUndef) return V1;
3114 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3116 // V2 is a splat, so the mask may be malformed. That is, it may point
3117 // to any V2 element. The instruction selectior won't like this. Get
3118 // a corrected mask and commute to form a proper MOVS{S|D}.
3119 SDOperand NewMask = getMOVLMask(NumElems, DAG);
3120 if (NewMask.Val != PermMask.Val)
3121 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3126 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3127 X86::isUNPCKLMask(PermMask.Val) ||
3128 X86::isUNPCKHMask(PermMask.Val))
3132 // Normalize mask so all entries that point to V2 points to its first
3133 // element then try to match unpck{h|l} again. If match, return a
3134 // new vector_shuffle with the corrected mask.
3135 SDOperand NewMask = NormalizeMask(PermMask, DAG);
3136 if (NewMask.Val != PermMask.Val) {
3137 if (X86::isUNPCKLMask(PermMask.Val, true)) {
3138 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
3139 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3140 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
3141 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
3142 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
3147 // Normalize the node to match x86 shuffle ops if needed
3148 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
3149 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3152 // Commute is back and try unpck* again.
3153 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
3154 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
3155 X86::isUNPCKLMask(PermMask.Val) ||
3156 X86::isUNPCKHMask(PermMask.Val))
3160 // If VT is integer, try PSHUF* first, then SHUFP*.
3161 if (MVT::isInteger(VT)) {
3162 if (X86::isPSHUFDMask(PermMask.Val) ||
3163 X86::isPSHUFHWMask(PermMask.Val) ||
3164 X86::isPSHUFLWMask(PermMask.Val)) {
3165 if (V2.getOpcode() != ISD::UNDEF)
3166 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3167 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3171 if (X86::isSHUFPMask(PermMask.Val))
3174 // Handle v8i16 shuffle high / low shuffle node pair.
3175 if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
3176 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
3177 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
3178 std::vector<SDOperand> MaskVec;
3179 for (unsigned i = 0; i != 4; ++i)
3180 MaskVec.push_back(PermMask.getOperand(i));
3181 for (unsigned i = 4; i != 8; ++i)
3182 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3183 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3184 &MaskVec[0], MaskVec.size());
3185 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3187 for (unsigned i = 0; i != 4; ++i)
3188 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3189 for (unsigned i = 4; i != 8; ++i)
3190 MaskVec.push_back(PermMask.getOperand(i));
3191 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0],MaskVec.size());
3192 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
3195 // Floating point cases in the other order.
3196 if (X86::isSHUFPMask(PermMask.Val))
3198 if (X86::isPSHUFDMask(PermMask.Val) ||
3199 X86::isPSHUFHWMask(PermMask.Val) ||
3200 X86::isPSHUFLWMask(PermMask.Val)) {
3201 if (V2.getOpcode() != ISD::UNDEF)
3202 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
3203 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
3208 if (NumElems == 4) {
3209 MVT::ValueType MaskVT = PermMask.getValueType();
3210 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
3211 std::vector<std::pair<int, int> > Locs;
3212 Locs.reserve(NumElems);
3213 std::vector<SDOperand> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3214 std::vector<SDOperand> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3217 // If no more than two elements come from either vector. This can be
3218 // implemented with two shuffles. First shuffle gather the elements.
3219 // The second shuffle, which takes the first shuffle as both of its
3220 // vector operands, put the elements into the right order.
3221 for (unsigned i = 0; i != NumElems; ++i) {
3222 SDOperand Elt = PermMask.getOperand(i);
3223 if (Elt.getOpcode() == ISD::UNDEF) {
3224 Locs[i] = std::make_pair(-1, -1);
3226 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
3227 if (Val < NumElems) {
3228 Locs[i] = std::make_pair(0, NumLo);
3232 Locs[i] = std::make_pair(1, NumHi);
3233 if (2+NumHi < NumElems)
3234 Mask1[2+NumHi] = Elt;
3239 if (NumLo <= 2 && NumHi <= 2) {
3240 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3241 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3242 &Mask1[0], Mask1.size()));
3243 for (unsigned i = 0; i != NumElems; ++i) {
3244 if (Locs[i].first == -1)
3247 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
3248 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
3249 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
3253 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
3254 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3255 &Mask2[0], Mask2.size()));
3258 // Break it into (shuffle shuffle_hi, shuffle_lo).
3260 std::vector<SDOperand> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3261 std::vector<SDOperand> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
3262 std::vector<SDOperand> *MaskPtr = &LoMask;
3263 unsigned MaskIdx = 0;
3265 unsigned HiIdx = NumElems/2;
3266 for (unsigned i = 0; i != NumElems; ++i) {
3267 if (i == NumElems/2) {
3273 SDOperand Elt = PermMask.getOperand(i);
3274 if (Elt.getOpcode() == ISD::UNDEF) {
3275 Locs[i] = std::make_pair(-1, -1);
3276 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
3277 Locs[i] = std::make_pair(MaskIdx, LoIdx);
3278 (*MaskPtr)[LoIdx] = Elt;
3281 Locs[i] = std::make_pair(MaskIdx, HiIdx);
3282 (*MaskPtr)[HiIdx] = Elt;
3287 SDOperand LoShuffle =
3288 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3289 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3290 &LoMask[0], LoMask.size()));
3291 SDOperand HiShuffle =
3292 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
3293 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3294 &HiMask[0], HiMask.size()));
3295 std::vector<SDOperand> MaskOps;
3296 for (unsigned i = 0; i != NumElems; ++i) {
3297 if (Locs[i].first == -1) {
3298 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
3300 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
3301 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
3304 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
3305 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3306 &MaskOps[0], MaskOps.size()));
3313 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3314 if (!isa<ConstantSDNode>(Op.getOperand(1)))
3317 MVT::ValueType VT = Op.getValueType();
3318 // TODO: handle v16i8.
3319 if (MVT::getSizeInBits(VT) == 16) {
3320 // Transform it so it match pextrw which produces a 32-bit result.
3321 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
3322 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
3323 Op.getOperand(0), Op.getOperand(1));
3324 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
3325 DAG.getValueType(VT));
3326 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
3327 } else if (MVT::getSizeInBits(VT) == 32) {
3328 SDOperand Vec = Op.getOperand(0);
3329 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3332 // SHUFPS the element to the lowest double word, then movss.
3333 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3334 std::vector<SDOperand> IdxVec;
3335 IdxVec.push_back(DAG.getConstant(Idx, MVT::getVectorBaseType(MaskVT)));
3336 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3337 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3338 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3339 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3340 &IdxVec[0], IdxVec.size());
3341 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3342 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3343 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3344 DAG.getConstant(0, getPointerTy()));
3345 } else if (MVT::getSizeInBits(VT) == 64) {
3346 SDOperand Vec = Op.getOperand(0);
3347 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3351 // UNPCKHPD the element to the lowest double word, then movsd.
3352 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
3353 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
3354 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3355 std::vector<SDOperand> IdxVec;
3356 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorBaseType(MaskVT)));
3357 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
3358 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3359 &IdxVec[0], IdxVec.size());
3360 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
3361 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
3362 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
3363 DAG.getConstant(0, getPointerTy()));
3370 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
3371 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
3372 // as its second argument.
3373 MVT::ValueType VT = Op.getValueType();
3374 MVT::ValueType BaseVT = MVT::getVectorBaseType(VT);
3375 SDOperand N0 = Op.getOperand(0);
3376 SDOperand N1 = Op.getOperand(1);
3377 SDOperand N2 = Op.getOperand(2);
3378 if (MVT::getSizeInBits(BaseVT) == 16) {
3379 if (N1.getValueType() != MVT::i32)
3380 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
3381 if (N2.getValueType() != MVT::i32)
3382 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(), MVT::i32);
3383 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
3384 } else if (MVT::getSizeInBits(BaseVT) == 32) {
3385 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
3388 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
3389 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3390 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
3391 std::vector<SDOperand> MaskVec;
3392 MaskVec.push_back(DAG.getConstant(4, BaseVT));
3393 for (unsigned i = 1; i <= 3; ++i)
3394 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3395 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
3396 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3397 &MaskVec[0], MaskVec.size()));
3399 // Use two pinsrw instructions to insert a 32 bit value.
3401 if (MVT::isFloatingPoint(N1.getValueType())) {
3402 if (ISD::isNON_EXTLoad(N1.Val)) {
3403 // Just load directly from f32mem to GR32.
3404 LoadSDNode *LD = cast<LoadSDNode>(N1);
3405 N1 = DAG.getLoad(MVT::i32, LD->getChain(), LD->getBasePtr(),
3406 LD->getSrcValue(), LD->getSrcValueOffset());
3408 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
3409 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
3410 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
3411 DAG.getConstant(0, getPointerTy()));
3414 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
3415 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3416 DAG.getConstant(Idx, getPointerTy()));
3417 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
3418 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3419 DAG.getConstant(Idx+1, getPointerTy()));
3420 return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
3428 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3429 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
3430 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
3433 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
3434 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
3435 // one of the above mentioned nodes. It has to be wrapped because otherwise
3436 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
3437 // be used to form addressing mode. These wrapped nodes will be selected
3440 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
3441 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3442 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
3444 CP->getAlignment());
3445 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3446 // With PIC, the address is actually $g + Offset.
3447 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3448 !Subtarget->isPICStyleRIPRel()) {
3449 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3450 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3458 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
3459 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3460 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
3461 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3462 // With PIC, the address is actually $g + Offset.
3463 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3464 !Subtarget->isPICStyleRIPRel()) {
3465 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3466 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3470 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
3471 // load the value at address GV, not the value of GV itself. This means that
3472 // the GlobalAddress must be in the base or index register of the address, not
3473 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
3474 // The same applies for external symbols during PIC codegen
3475 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
3476 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
3482 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
3483 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
3484 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
3485 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3486 // With PIC, the address is actually $g + Offset.
3487 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3488 !Subtarget->isPICStyleRIPRel()) {
3489 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3490 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3497 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
3498 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3499 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
3500 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3501 // With PIC, the address is actually $g + Offset.
3502 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3503 !Subtarget->isPICStyleRIPRel()) {
3504 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3505 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3512 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
3513 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
3514 "Not an i64 shift!");
3515 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
3516 SDOperand ShOpLo = Op.getOperand(0);
3517 SDOperand ShOpHi = Op.getOperand(1);
3518 SDOperand ShAmt = Op.getOperand(2);
3519 SDOperand Tmp1 = isSRA ?
3520 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
3521 DAG.getConstant(0, MVT::i32);
3523 SDOperand Tmp2, Tmp3;
3524 if (Op.getOpcode() == ISD::SHL_PARTS) {
3525 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
3526 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
3528 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
3529 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
3532 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3533 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
3534 DAG.getConstant(32, MVT::i8));
3535 SDOperand COps[]={DAG.getEntryNode(), AndNode, DAG.getConstant(0, MVT::i8)};
3536 SDOperand InFlag = DAG.getNode(X86ISD::CMP, VTs, 2, COps, 3).getValue(1);
3539 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3541 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
3542 SmallVector<SDOperand, 4> Ops;
3543 if (Op.getOpcode() == ISD::SHL_PARTS) {
3544 Ops.push_back(Tmp2);
3545 Ops.push_back(Tmp3);
3547 Ops.push_back(InFlag);
3548 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3549 InFlag = Hi.getValue(1);
3552 Ops.push_back(Tmp3);
3553 Ops.push_back(Tmp1);
3555 Ops.push_back(InFlag);
3556 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3558 Ops.push_back(Tmp2);
3559 Ops.push_back(Tmp3);
3561 Ops.push_back(InFlag);
3562 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3563 InFlag = Lo.getValue(1);
3566 Ops.push_back(Tmp3);
3567 Ops.push_back(Tmp1);
3569 Ops.push_back(InFlag);
3570 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3573 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
3577 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
3580 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
3581 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
3582 Op.getOperand(0).getValueType() >= MVT::i16 &&
3583 "Unknown SINT_TO_FP to lower!");
3586 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
3587 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
3588 MachineFunction &MF = DAG.getMachineFunction();
3589 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
3590 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3591 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
3592 StackSlot, NULL, 0);
3595 std::vector<MVT::ValueType> Tys;
3596 Tys.push_back(MVT::f64);
3597 Tys.push_back(MVT::Other);
3598 if (X86ScalarSSE) Tys.push_back(MVT::Flag);
3599 std::vector<SDOperand> Ops;
3600 Ops.push_back(Chain);
3601 Ops.push_back(StackSlot);
3602 Ops.push_back(DAG.getValueType(SrcVT));
3603 Result = DAG.getNode(X86ScalarSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
3604 Tys, &Ops[0], Ops.size());
3607 Chain = Result.getValue(1);
3608 SDOperand InFlag = Result.getValue(2);
3610 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
3611 // shouldn't be necessary except that RFP cannot be live across
3612 // multiple blocks. When stackifier is fixed, they can be uncoupled.
3613 MachineFunction &MF = DAG.getMachineFunction();
3614 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3615 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3616 std::vector<MVT::ValueType> Tys;
3617 Tys.push_back(MVT::Other);
3618 std::vector<SDOperand> Ops;
3619 Ops.push_back(Chain);
3620 Ops.push_back(Result);
3621 Ops.push_back(StackSlot);
3622 Ops.push_back(DAG.getValueType(Op.getValueType()));
3623 Ops.push_back(InFlag);
3624 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
3625 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
3631 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
3632 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
3633 "Unknown FP_TO_SINT to lower!");
3634 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
3636 MachineFunction &MF = DAG.getMachineFunction();
3637 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
3638 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3639 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3642 switch (Op.getValueType()) {
3643 default: assert(0 && "Invalid FP_TO_SINT to lower!");
3644 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
3645 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
3646 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
3649 SDOperand Chain = DAG.getEntryNode();
3650 SDOperand Value = Op.getOperand(0);
3652 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
3653 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
3654 std::vector<MVT::ValueType> Tys;
3655 Tys.push_back(MVT::f64);
3656 Tys.push_back(MVT::Other);
3657 std::vector<SDOperand> Ops;
3658 Ops.push_back(Chain);
3659 Ops.push_back(StackSlot);
3660 Ops.push_back(DAG.getValueType(Op.getOperand(0).getValueType()));
3661 Value = DAG.getNode(X86ISD::FLD, Tys, &Ops[0], Ops.size());
3662 Chain = Value.getValue(1);
3663 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3664 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3667 // Build the FP_TO_INT*_IN_MEM
3668 std::vector<SDOperand> Ops;
3669 Ops.push_back(Chain);
3670 Ops.push_back(Value);
3671 Ops.push_back(StackSlot);
3672 SDOperand FIST = DAG.getNode(Opc, MVT::Other, &Ops[0], Ops.size());
3675 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
3678 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
3679 MVT::ValueType VT = Op.getValueType();
3680 const Type *OpNTy = MVT::getTypeForValueType(VT);
3681 std::vector<Constant*> CV;
3682 if (VT == MVT::f64) {
3683 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(~(1ULL << 63))));
3684 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3686 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(~(1U << 31))));
3687 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3688 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3689 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3691 Constant *CS = ConstantStruct::get(CV);
3692 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3693 std::vector<MVT::ValueType> Tys;
3695 Tys.push_back(MVT::Other);
3696 SmallVector<SDOperand, 3> Ops;
3697 Ops.push_back(DAG.getEntryNode());
3698 Ops.push_back(CPIdx);
3699 Ops.push_back(DAG.getSrcValue(NULL));
3700 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3701 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
3704 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
3705 MVT::ValueType VT = Op.getValueType();
3706 const Type *OpNTy = MVT::getTypeForValueType(VT);
3707 std::vector<Constant*> CV;
3708 if (VT == MVT::f64) {
3709 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(1ULL << 63)));
3710 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3712 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(1U << 31)));
3713 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3714 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3715 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3717 Constant *CS = ConstantStruct::get(CV);
3718 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3719 std::vector<MVT::ValueType> Tys;
3721 Tys.push_back(MVT::Other);
3722 SmallVector<SDOperand, 3> Ops;
3723 Ops.push_back(DAG.getEntryNode());
3724 Ops.push_back(CPIdx);
3725 Ops.push_back(DAG.getSrcValue(NULL));
3726 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3727 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
3730 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
3731 SDOperand Op0 = Op.getOperand(0);
3732 SDOperand Op1 = Op.getOperand(1);
3733 MVT::ValueType VT = Op.getValueType();
3734 MVT::ValueType SrcVT = Op1.getValueType();
3735 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
3737 // If second operand is smaller, extend it first.
3738 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
3739 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
3743 // First get the sign bit of second operand.
3744 std::vector<Constant*> CV;
3745 if (SrcVT == MVT::f64) {
3746 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(1ULL << 63)));
3747 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3749 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(1U << 31)));
3750 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3751 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3752 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3754 Constant *CS = ConstantStruct::get(CV);
3755 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3756 std::vector<MVT::ValueType> Tys;
3757 Tys.push_back(SrcVT);
3758 Tys.push_back(MVT::Other);
3759 SmallVector<SDOperand, 3> Ops;
3760 Ops.push_back(DAG.getEntryNode());
3761 Ops.push_back(CPIdx);
3762 Ops.push_back(DAG.getSrcValue(NULL));
3763 SDOperand Mask1 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3764 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
3766 // Shift sign bit right or left if the two operands have different types.
3767 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
3768 // Op0 is MVT::f32, Op1 is MVT::f64.
3769 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
3770 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
3771 DAG.getConstant(32, MVT::i32));
3772 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
3773 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
3774 DAG.getConstant(0, getPointerTy()));
3777 // Clear first operand sign bit.
3779 if (VT == MVT::f64) {
3780 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(~(1ULL << 63))));
3781 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3783 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(~(1U << 31))));
3784 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3785 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3786 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3788 CS = ConstantStruct::get(CV);
3789 CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3792 Tys.push_back(MVT::Other);
3794 Ops.push_back(DAG.getEntryNode());
3795 Ops.push_back(CPIdx);
3796 Ops.push_back(DAG.getSrcValue(NULL));
3797 SDOperand Mask2 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3798 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
3800 // Or the value with the sign bit.
3801 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
3804 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG,
3806 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
3808 SDOperand Op0 = Op.getOperand(0);
3809 SDOperand Op1 = Op.getOperand(1);
3810 SDOperand CC = Op.getOperand(2);
3811 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3812 const MVT::ValueType *VTs1 = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3813 const MVT::ValueType *VTs2 = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
3814 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
3817 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
3819 SDOperand Ops1[] = { Chain, Op0, Op1 };
3820 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, Ops1, 3).getValue(1);
3821 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
3822 return DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3825 assert(isFP && "Illegal integer SetCC!");
3827 SDOperand COps[] = { Chain, Op0, Op1 };
3828 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, COps, 3).getValue(1);
3830 switch (SetCCOpcode) {
3831 default: assert(false && "Illegal floating point SetCC!");
3832 case ISD::SETOEQ: { // !PF & ZF
3833 SDOperand Ops1[] = { DAG.getConstant(X86::COND_NP, MVT::i8), Cond };
3834 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3835 SDOperand Ops2[] = { DAG.getConstant(X86::COND_E, MVT::i8),
3837 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3838 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
3840 case ISD::SETUNE: { // PF | !ZF
3841 SDOperand Ops1[] = { DAG.getConstant(X86::COND_P, MVT::i8), Cond };
3842 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3843 SDOperand Ops2[] = { DAG.getConstant(X86::COND_NE, MVT::i8),
3845 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3846 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
3851 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
3852 bool addTest = true;
3853 SDOperand Chain = DAG.getEntryNode();
3854 SDOperand Cond = Op.getOperand(0);
3856 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3858 if (Cond.getOpcode() == ISD::SETCC)
3859 Cond = LowerSETCC(Cond, DAG, Chain);
3861 if (Cond.getOpcode() == X86ISD::SETCC) {
3862 CC = Cond.getOperand(0);
3864 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3865 // (since flag operand cannot be shared). Use it as the condition setting
3866 // operand in place of the X86ISD::SETCC.
3867 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3868 // to use a test instead of duplicating the X86ISD::CMP (for register
3869 // pressure reason)?
3870 SDOperand Cmp = Cond.getOperand(1);
3871 unsigned Opc = Cmp.getOpcode();
3872 bool IllegalFPCMov = !X86ScalarSSE &&
3873 MVT::isFloatingPoint(Op.getValueType()) &&
3874 !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
3875 if ((Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) &&
3877 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3878 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3884 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3885 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3886 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3889 VTs = DAG.getNodeValueTypes(Op.getValueType(), MVT::Flag);
3890 SmallVector<SDOperand, 4> Ops;
3891 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
3892 // condition is true.
3893 Ops.push_back(Op.getOperand(2));
3894 Ops.push_back(Op.getOperand(1));
3896 Ops.push_back(Cond.getValue(1));
3897 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3900 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
3901 bool addTest = true;
3902 SDOperand Chain = Op.getOperand(0);
3903 SDOperand Cond = Op.getOperand(1);
3904 SDOperand Dest = Op.getOperand(2);
3906 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3908 if (Cond.getOpcode() == ISD::SETCC)
3909 Cond = LowerSETCC(Cond, DAG, Chain);
3911 if (Cond.getOpcode() == X86ISD::SETCC) {
3912 CC = Cond.getOperand(0);
3914 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3915 // (since flag operand cannot be shared). Use it as the condition setting
3916 // operand in place of the X86ISD::SETCC.
3917 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3918 // to use a test instead of duplicating the X86ISD::CMP (for register
3919 // pressure reason)?
3920 SDOperand Cmp = Cond.getOperand(1);
3921 unsigned Opc = Cmp.getOpcode();
3922 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) {
3923 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3924 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3930 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3931 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3932 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3934 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
3935 Cond, Op.getOperand(2), CC, Cond.getValue(1));
3938 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
3939 unsigned CallingConv= cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3941 if (Subtarget->is64Bit())
3942 return LowerX86_64CCCCallTo(Op, DAG);
3944 switch (CallingConv) {
3946 assert(0 && "Unsupported calling convention");
3947 case CallingConv::Fast:
3949 return LowerFastCCCallTo(Op, DAG);
3952 case CallingConv::C:
3953 return LowerCCCCallTo(Op, DAG);
3954 case CallingConv::X86_StdCall:
3955 return LowerCCCCallTo(Op, DAG, true);
3956 case CallingConv::X86_FastCall:
3957 return LowerFastCCCallTo(Op, DAG, true);
3961 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
3964 switch(Op.getNumOperands()) {
3966 assert(0 && "Do not know how to return this many arguments!");
3968 case 1: // ret void.
3969 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Op.getOperand(0),
3970 DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
3972 MVT::ValueType ArgVT = Op.getOperand(1).getValueType();
3974 if (MVT::isVector(ArgVT) ||
3975 (Subtarget->is64Bit() && MVT::isFloatingPoint(ArgVT))) {
3976 // Integer or FP vector result -> XMM0.
3977 if (DAG.getMachineFunction().liveout_empty())
3978 DAG.getMachineFunction().addLiveOut(X86::XMM0);
3979 Copy = DAG.getCopyToReg(Op.getOperand(0), X86::XMM0, Op.getOperand(1),
3981 } else if (MVT::isInteger(ArgVT)) {
3982 // Integer result -> EAX / RAX.
3983 // The C calling convention guarantees the return value has been
3984 // promoted to at least MVT::i32. The X86-64 ABI doesn't require the
3985 // value to be promoted MVT::i64. So we don't have to extend it to
3986 // 64-bit. Return the value in EAX, but mark RAX as liveout.
3987 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
3988 if (DAG.getMachineFunction().liveout_empty())
3989 DAG.getMachineFunction().addLiveOut(Reg);
3991 Reg = (ArgVT == MVT::i64) ? X86::RAX : X86::EAX;
3992 Copy = DAG.getCopyToReg(Op.getOperand(0), Reg, Op.getOperand(1),
3994 } else if (!X86ScalarSSE) {
3995 // FP return with fp-stack value.
3996 if (DAG.getMachineFunction().liveout_empty())
3997 DAG.getMachineFunction().addLiveOut(X86::ST0);
3999 std::vector<MVT::ValueType> Tys;
4000 Tys.push_back(MVT::Other);
4001 Tys.push_back(MVT::Flag);
4002 std::vector<SDOperand> Ops;
4003 Ops.push_back(Op.getOperand(0));
4004 Ops.push_back(Op.getOperand(1));
4005 Copy = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, &Ops[0], Ops.size());
4007 // FP return with ScalarSSE (return on fp-stack).
4008 if (DAG.getMachineFunction().liveout_empty())
4009 DAG.getMachineFunction().addLiveOut(X86::ST0);
4012 SDOperand Chain = Op.getOperand(0);
4013 SDOperand Value = Op.getOperand(1);
4015 if (ISD::isNON_EXTLoad(Value.Val) &&
4016 (Chain == Value.getValue(1) || Chain == Value.getOperand(0))) {
4017 Chain = Value.getOperand(0);
4018 MemLoc = Value.getOperand(1);
4020 // Spill the value to memory and reload it into top of stack.
4021 unsigned Size = MVT::getSizeInBits(ArgVT)/8;
4022 MachineFunction &MF = DAG.getMachineFunction();
4023 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4024 MemLoc = DAG.getFrameIndex(SSFI, getPointerTy());
4025 Chain = DAG.getStore(Op.getOperand(0), Value, MemLoc, NULL, 0);
4027 std::vector<MVT::ValueType> Tys;
4028 Tys.push_back(MVT::f64);
4029 Tys.push_back(MVT::Other);
4030 std::vector<SDOperand> Ops;
4031 Ops.push_back(Chain);
4032 Ops.push_back(MemLoc);
4033 Ops.push_back(DAG.getValueType(ArgVT));
4034 Copy = DAG.getNode(X86ISD::FLD, Tys, &Ops[0], Ops.size());
4036 Tys.push_back(MVT::Other);
4037 Tys.push_back(MVT::Flag);
4039 Ops.push_back(Copy.getValue(1));
4040 Ops.push_back(Copy);
4041 Copy = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, &Ops[0], Ops.size());
4046 unsigned Reg1 = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
4047 unsigned Reg2 = Subtarget->is64Bit() ? X86::RDX : X86::EDX;
4048 if (DAG.getMachineFunction().liveout_empty()) {
4049 DAG.getMachineFunction().addLiveOut(Reg1);
4050 DAG.getMachineFunction().addLiveOut(Reg2);
4053 Copy = DAG.getCopyToReg(Op.getOperand(0), Reg2, Op.getOperand(3),
4055 Copy = DAG.getCopyToReg(Copy, Reg1, Op.getOperand(1), Copy.getValue(1));
4059 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other,
4060 Copy, DAG.getConstant(getBytesToPopOnReturn(), MVT::i16),
4065 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
4066 MachineFunction &MF = DAG.getMachineFunction();
4067 const Function* Fn = MF.getFunction();
4068 if (Fn->hasExternalLinkage() &&
4069 Subtarget->isTargetCygMing() &&
4070 Fn->getName() == "main")
4071 MF.getInfo<X86FunctionInfo>()->setForceFramePointer(true);
4073 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
4074 if (Subtarget->is64Bit())
4075 return LowerX86_64CCCArguments(Op, DAG);
4079 assert(0 && "Unsupported calling convention");
4080 case CallingConv::Fast:
4082 return LowerFastCCArguments(Op, DAG);
4085 case CallingConv::C:
4086 return LowerCCCArguments(Op, DAG);
4087 case CallingConv::X86_StdCall:
4088 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(StdCall);
4089 return LowerCCCArguments(Op, DAG, true);
4090 case CallingConv::X86_FastCall:
4091 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(FastCall);
4092 return LowerFastCCArguments(Op, DAG, true);
4096 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
4097 SDOperand InFlag(0, 0);
4098 SDOperand Chain = Op.getOperand(0);
4100 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
4101 if (Align == 0) Align = 1;
4103 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
4104 // If not DWORD aligned, call memset if size is less than the threshold.
4105 // It knows how to align to the right boundary first.
4106 if ((Align & 3) != 0 ||
4107 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
4108 MVT::ValueType IntPtr = getPointerTy();
4109 const Type *IntPtrTy = getTargetData()->getIntPtrType();
4110 TargetLowering::ArgListTy Args;
4111 TargetLowering::ArgListEntry Entry;
4112 Entry.Node = Op.getOperand(1);
4113 Entry.Ty = IntPtrTy;
4114 Entry.isSigned = false;
4115 Entry.isInReg = false;
4116 Entry.isSRet = false;
4117 Args.push_back(Entry);
4118 // Extend the unsigned i8 argument to be an int value for the call.
4119 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
4120 Entry.Ty = IntPtrTy;
4121 Entry.isSigned = false;
4122 Entry.isInReg = false;
4123 Entry.isSRet = false;
4124 Args.push_back(Entry);
4125 Entry.Node = Op.getOperand(3);
4126 Args.push_back(Entry);
4127 std::pair<SDOperand,SDOperand> CallResult =
4128 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
4129 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
4130 return CallResult.second;
4135 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
4136 unsigned BytesLeft = 0;
4137 bool TwoRepStos = false;
4140 uint64_t Val = ValC->getValue() & 255;
4142 // If the value is a constant, then we can potentially use larger sets.
4143 switch (Align & 3) {
4144 case 2: // WORD aligned
4147 Val = (Val << 8) | Val;
4149 case 0: // DWORD aligned
4152 Val = (Val << 8) | Val;
4153 Val = (Val << 16) | Val;
4154 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
4157 Val = (Val << 32) | Val;
4160 default: // Byte aligned
4163 Count = Op.getOperand(3);
4167 if (AVT > MVT::i8) {
4169 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4170 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
4171 BytesLeft = I->getValue() % UBytes;
4173 assert(AVT >= MVT::i32 &&
4174 "Do not use rep;stos if not at least DWORD aligned");
4175 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
4176 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
4181 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
4183 InFlag = Chain.getValue(1);
4186 Count = Op.getOperand(3);
4187 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
4188 InFlag = Chain.getValue(1);
4191 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4193 InFlag = Chain.getValue(1);
4194 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4195 Op.getOperand(1), InFlag);
4196 InFlag = Chain.getValue(1);
4198 std::vector<MVT::ValueType> Tys;
4199 Tys.push_back(MVT::Other);
4200 Tys.push_back(MVT::Flag);
4201 std::vector<SDOperand> Ops;
4202 Ops.push_back(Chain);
4203 Ops.push_back(DAG.getValueType(AVT));
4204 Ops.push_back(InFlag);
4205 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4208 InFlag = Chain.getValue(1);
4209 Count = Op.getOperand(3);
4210 MVT::ValueType CVT = Count.getValueType();
4211 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4212 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4213 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4215 InFlag = Chain.getValue(1);
4217 Tys.push_back(MVT::Other);
4218 Tys.push_back(MVT::Flag);
4220 Ops.push_back(Chain);
4221 Ops.push_back(DAG.getValueType(MVT::i8));
4222 Ops.push_back(InFlag);
4223 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
4224 } else if (BytesLeft) {
4225 // Issue stores for the last 1 - 7 bytes.
4227 unsigned Val = ValC->getValue() & 255;
4228 unsigned Offset = I->getValue() - BytesLeft;
4229 SDOperand DstAddr = Op.getOperand(1);
4230 MVT::ValueType AddrVT = DstAddr.getValueType();
4231 if (BytesLeft >= 4) {
4232 Val = (Val << 8) | Val;
4233 Val = (Val << 16) | Val;
4234 Value = DAG.getConstant(Val, MVT::i32);
4235 Chain = DAG.getStore(Chain, Value,
4236 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4237 DAG.getConstant(Offset, AddrVT)),
4242 if (BytesLeft >= 2) {
4243 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
4244 Chain = DAG.getStore(Chain, Value,
4245 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4246 DAG.getConstant(Offset, AddrVT)),
4251 if (BytesLeft == 1) {
4252 Value = DAG.getConstant(Val, MVT::i8);
4253 Chain = DAG.getStore(Chain, Value,
4254 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
4255 DAG.getConstant(Offset, AddrVT)),
4263 SDOperand X86TargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) {
4264 SDOperand Chain = Op.getOperand(0);
4266 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
4267 if (Align == 0) Align = 1;
4269 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
4270 // If not DWORD aligned, call memcpy if size is less than the threshold.
4271 // It knows how to align to the right boundary first.
4272 if ((Align & 3) != 0 ||
4273 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
4274 MVT::ValueType IntPtr = getPointerTy();
4275 TargetLowering::ArgListTy Args;
4276 TargetLowering::ArgListEntry Entry;
4277 Entry.Ty = getTargetData()->getIntPtrType();
4278 Entry.isSigned = false;
4279 Entry.isInReg = false;
4280 Entry.isSRet = false;
4281 Entry.Node = Op.getOperand(1); Args.push_back(Entry);
4282 Entry.Node = Op.getOperand(2); Args.push_back(Entry);
4283 Entry.Node = Op.getOperand(3); Args.push_back(Entry);
4284 std::pair<SDOperand,SDOperand> CallResult =
4285 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
4286 DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
4287 return CallResult.second;
4292 unsigned BytesLeft = 0;
4293 bool TwoRepMovs = false;
4294 switch (Align & 3) {
4295 case 2: // WORD aligned
4298 case 0: // DWORD aligned
4300 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
4303 default: // Byte aligned
4305 Count = Op.getOperand(3);
4309 if (AVT > MVT::i8) {
4311 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
4312 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
4313 BytesLeft = I->getValue() % UBytes;
4315 assert(AVT >= MVT::i32 &&
4316 "Do not use rep;movs if not at least DWORD aligned");
4317 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
4318 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
4323 SDOperand InFlag(0, 0);
4324 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
4326 InFlag = Chain.getValue(1);
4327 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
4328 Op.getOperand(1), InFlag);
4329 InFlag = Chain.getValue(1);
4330 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
4331 Op.getOperand(2), InFlag);
4332 InFlag = Chain.getValue(1);
4334 std::vector<MVT::ValueType> Tys;
4335 Tys.push_back(MVT::Other);
4336 Tys.push_back(MVT::Flag);
4337 std::vector<SDOperand> Ops;
4338 Ops.push_back(Chain);
4339 Ops.push_back(DAG.getValueType(AVT));
4340 Ops.push_back(InFlag);
4341 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4344 InFlag = Chain.getValue(1);
4345 Count = Op.getOperand(3);
4346 MVT::ValueType CVT = Count.getValueType();
4347 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
4348 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
4349 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
4351 InFlag = Chain.getValue(1);
4353 Tys.push_back(MVT::Other);
4354 Tys.push_back(MVT::Flag);
4356 Ops.push_back(Chain);
4357 Ops.push_back(DAG.getValueType(MVT::i8));
4358 Ops.push_back(InFlag);
4359 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
4360 } else if (BytesLeft) {
4361 // Issue loads and stores for the last 1 - 7 bytes.
4362 unsigned Offset = I->getValue() - BytesLeft;
4363 SDOperand DstAddr = Op.getOperand(1);
4364 MVT::ValueType DstVT = DstAddr.getValueType();
4365 SDOperand SrcAddr = Op.getOperand(2);
4366 MVT::ValueType SrcVT = SrcAddr.getValueType();
4368 if (BytesLeft >= 4) {
4369 Value = DAG.getLoad(MVT::i32, Chain,
4370 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4371 DAG.getConstant(Offset, SrcVT)),
4373 Chain = Value.getValue(1);
4374 Chain = DAG.getStore(Chain, Value,
4375 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4376 DAG.getConstant(Offset, DstVT)),
4381 if (BytesLeft >= 2) {
4382 Value = DAG.getLoad(MVT::i16, Chain,
4383 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4384 DAG.getConstant(Offset, SrcVT)),
4386 Chain = Value.getValue(1);
4387 Chain = DAG.getStore(Chain, Value,
4388 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4389 DAG.getConstant(Offset, DstVT)),
4395 if (BytesLeft == 1) {
4396 Value = DAG.getLoad(MVT::i8, Chain,
4397 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
4398 DAG.getConstant(Offset, SrcVT)),
4400 Chain = Value.getValue(1);
4401 Chain = DAG.getStore(Chain, Value,
4402 DAG.getNode(ISD::ADD, DstVT, DstAddr,
4403 DAG.getConstant(Offset, DstVT)),
4412 X86TargetLowering::LowerREADCYCLCECOUNTER(SDOperand Op, SelectionDAG &DAG) {
4413 std::vector<MVT::ValueType> Tys;
4414 Tys.push_back(MVT::Other);
4415 Tys.push_back(MVT::Flag);
4416 std::vector<SDOperand> Ops;
4417 Ops.push_back(Op.getOperand(0));
4418 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &Ops[0], Ops.size());
4420 if (Subtarget->is64Bit()) {
4421 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
4422 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::RDX,
4423 MVT::i64, Copy1.getValue(2));
4424 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, Copy2,
4425 DAG.getConstant(32, MVT::i8));
4426 Ops.push_back(DAG.getNode(ISD::OR, MVT::i64, Copy1, Tmp));
4427 Ops.push_back(Copy2.getValue(1));
4429 Tys[1] = MVT::Other;
4431 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
4432 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::EDX,
4433 MVT::i32, Copy1.getValue(2));
4434 Ops.push_back(Copy1);
4435 Ops.push_back(Copy2);
4436 Ops.push_back(Copy2.getValue(1));
4437 Tys[0] = Tys[1] = MVT::i32;
4438 Tys.push_back(MVT::Other);
4440 return DAG.getNode(ISD::MERGE_VALUES, Tys, &Ops[0], Ops.size());
4443 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
4444 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
4446 if (!Subtarget->is64Bit()) {
4447 // vastart just stores the address of the VarArgsFrameIndex slot into the
4448 // memory location argument.
4449 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4450 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
4455 // gp_offset (0 - 6 * 8)
4456 // fp_offset (48 - 48 + 8 * 16)
4457 // overflow_arg_area (point to parameters coming in memory).
4459 std::vector<SDOperand> MemOps;
4460 SDOperand FIN = Op.getOperand(1);
4462 SDOperand Store = DAG.getStore(Op.getOperand(0),
4463 DAG.getConstant(VarArgsGPOffset, MVT::i32),
4464 FIN, SV->getValue(), SV->getOffset());
4465 MemOps.push_back(Store);
4468 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4469 DAG.getConstant(4, getPointerTy()));
4470 Store = DAG.getStore(Op.getOperand(0),
4471 DAG.getConstant(VarArgsFPOffset, MVT::i32),
4472 FIN, SV->getValue(), SV->getOffset());
4473 MemOps.push_back(Store);
4475 // Store ptr to overflow_arg_area
4476 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4477 DAG.getConstant(4, getPointerTy()));
4478 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
4479 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
4481 MemOps.push_back(Store);
4483 // Store ptr to reg_save_area.
4484 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
4485 DAG.getConstant(8, getPointerTy()));
4486 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
4487 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
4489 MemOps.push_back(Store);
4490 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
4494 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
4495 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
4497 default: return SDOperand(); // Don't custom lower most intrinsics.
4498 // Comparison intrinsics.
4499 case Intrinsic::x86_sse_comieq_ss:
4500 case Intrinsic::x86_sse_comilt_ss:
4501 case Intrinsic::x86_sse_comile_ss:
4502 case Intrinsic::x86_sse_comigt_ss:
4503 case Intrinsic::x86_sse_comige_ss:
4504 case Intrinsic::x86_sse_comineq_ss:
4505 case Intrinsic::x86_sse_ucomieq_ss:
4506 case Intrinsic::x86_sse_ucomilt_ss:
4507 case Intrinsic::x86_sse_ucomile_ss:
4508 case Intrinsic::x86_sse_ucomigt_ss:
4509 case Intrinsic::x86_sse_ucomige_ss:
4510 case Intrinsic::x86_sse_ucomineq_ss:
4511 case Intrinsic::x86_sse2_comieq_sd:
4512 case Intrinsic::x86_sse2_comilt_sd:
4513 case Intrinsic::x86_sse2_comile_sd:
4514 case Intrinsic::x86_sse2_comigt_sd:
4515 case Intrinsic::x86_sse2_comige_sd:
4516 case Intrinsic::x86_sse2_comineq_sd:
4517 case Intrinsic::x86_sse2_ucomieq_sd:
4518 case Intrinsic::x86_sse2_ucomilt_sd:
4519 case Intrinsic::x86_sse2_ucomile_sd:
4520 case Intrinsic::x86_sse2_ucomigt_sd:
4521 case Intrinsic::x86_sse2_ucomige_sd:
4522 case Intrinsic::x86_sse2_ucomineq_sd: {
4524 ISD::CondCode CC = ISD::SETCC_INVALID;
4527 case Intrinsic::x86_sse_comieq_ss:
4528 case Intrinsic::x86_sse2_comieq_sd:
4532 case Intrinsic::x86_sse_comilt_ss:
4533 case Intrinsic::x86_sse2_comilt_sd:
4537 case Intrinsic::x86_sse_comile_ss:
4538 case Intrinsic::x86_sse2_comile_sd:
4542 case Intrinsic::x86_sse_comigt_ss:
4543 case Intrinsic::x86_sse2_comigt_sd:
4547 case Intrinsic::x86_sse_comige_ss:
4548 case Intrinsic::x86_sse2_comige_sd:
4552 case Intrinsic::x86_sse_comineq_ss:
4553 case Intrinsic::x86_sse2_comineq_sd:
4557 case Intrinsic::x86_sse_ucomieq_ss:
4558 case Intrinsic::x86_sse2_ucomieq_sd:
4559 Opc = X86ISD::UCOMI;
4562 case Intrinsic::x86_sse_ucomilt_ss:
4563 case Intrinsic::x86_sse2_ucomilt_sd:
4564 Opc = X86ISD::UCOMI;
4567 case Intrinsic::x86_sse_ucomile_ss:
4568 case Intrinsic::x86_sse2_ucomile_sd:
4569 Opc = X86ISD::UCOMI;
4572 case Intrinsic::x86_sse_ucomigt_ss:
4573 case Intrinsic::x86_sse2_ucomigt_sd:
4574 Opc = X86ISD::UCOMI;
4577 case Intrinsic::x86_sse_ucomige_ss:
4578 case Intrinsic::x86_sse2_ucomige_sd:
4579 Opc = X86ISD::UCOMI;
4582 case Intrinsic::x86_sse_ucomineq_ss:
4583 case Intrinsic::x86_sse2_ucomineq_sd:
4584 Opc = X86ISD::UCOMI;
4590 SDOperand LHS = Op.getOperand(1);
4591 SDOperand RHS = Op.getOperand(2);
4592 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
4594 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
4595 SDOperand Ops1[] = { DAG.getEntryNode(), LHS, RHS };
4596 SDOperand Cond = DAG.getNode(Opc, VTs, 2, Ops1, 3);
4597 VTs = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
4598 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
4599 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, VTs, 2, Ops2, 2);
4600 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
4605 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
4606 // Depths > 0 not supported yet!
4607 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4610 // Just load the return address
4611 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4612 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
4615 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
4616 // Depths > 0 not supported yet!
4617 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4620 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4621 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
4622 DAG.getConstant(4, getPointerTy()));
4625 /// LowerOperation - Provide custom lowering hooks for some operations.
4627 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
4628 switch (Op.getOpcode()) {
4629 default: assert(0 && "Should not custom lower this!");
4630 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4631 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4632 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4633 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
4634 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4635 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4636 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4637 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
4638 case ISD::SHL_PARTS:
4639 case ISD::SRA_PARTS:
4640 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
4641 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4642 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
4643 case ISD::FABS: return LowerFABS(Op, DAG);
4644 case ISD::FNEG: return LowerFNEG(Op, DAG);
4645 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4646 case ISD::SETCC: return LowerSETCC(Op, DAG, DAG.getEntryNode());
4647 case ISD::SELECT: return LowerSELECT(Op, DAG);
4648 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
4649 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4650 case ISD::CALL: return LowerCALL(Op, DAG);
4651 case ISD::RET: return LowerRET(Op, DAG);
4652 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
4653 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
4654 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
4655 case ISD::READCYCLECOUNTER: return LowerREADCYCLCECOUNTER(Op, DAG);
4656 case ISD::VASTART: return LowerVASTART(Op, DAG);
4657 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4658 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4659 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4663 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
4665 default: return NULL;
4666 case X86ISD::SHLD: return "X86ISD::SHLD";
4667 case X86ISD::SHRD: return "X86ISD::SHRD";
4668 case X86ISD::FAND: return "X86ISD::FAND";
4669 case X86ISD::FOR: return "X86ISD::FOR";
4670 case X86ISD::FXOR: return "X86ISD::FXOR";
4671 case X86ISD::FSRL: return "X86ISD::FSRL";
4672 case X86ISD::FILD: return "X86ISD::FILD";
4673 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
4674 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
4675 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
4676 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
4677 case X86ISD::FLD: return "X86ISD::FLD";
4678 case X86ISD::FST: return "X86ISD::FST";
4679 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
4680 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
4681 case X86ISD::CALL: return "X86ISD::CALL";
4682 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
4683 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
4684 case X86ISD::CMP: return "X86ISD::CMP";
4685 case X86ISD::COMI: return "X86ISD::COMI";
4686 case X86ISD::UCOMI: return "X86ISD::UCOMI";
4687 case X86ISD::SETCC: return "X86ISD::SETCC";
4688 case X86ISD::CMOV: return "X86ISD::CMOV";
4689 case X86ISD::BRCOND: return "X86ISD::BRCOND";
4690 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
4691 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
4692 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
4693 case X86ISD::LOAD_PACK: return "X86ISD::LOAD_PACK";
4694 case X86ISD::LOAD_UA: return "X86ISD::LOAD_UA";
4695 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
4696 case X86ISD::Wrapper: return "X86ISD::Wrapper";
4697 case X86ISD::S2VEC: return "X86ISD::S2VEC";
4698 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
4699 case X86ISD::PINSRW: return "X86ISD::PINSRW";
4700 case X86ISD::FMAX: return "X86ISD::FMAX";
4701 case X86ISD::FMIN: return "X86ISD::FMIN";
4705 /// isLegalAddressImmediate - Return true if the integer value or
4706 /// GlobalValue can be used as the offset of the target addressing mode.
4707 bool X86TargetLowering::isLegalAddressImmediate(int64_t V) const {
4708 // X86 allows a sign-extended 32-bit immediate field.
4709 return (V > -(1LL << 32) && V < (1LL << 32)-1);
4712 bool X86TargetLowering::isLegalAddressImmediate(GlobalValue *GV) const {
4713 // In 64-bit mode, GV is 64-bit so it won't fit in the 32-bit displacement
4714 // field unless we are in small code model.
4715 if (Subtarget->is64Bit() &&
4716 getTargetMachine().getCodeModel() != CodeModel::Small)
4719 return (!Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false));
4722 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4723 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4724 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4725 /// are assumed to be legal.
4727 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
4728 // Only do shuffles on 128-bit vector types for now.
4729 if (MVT::getSizeInBits(VT) == 64) return false;
4730 return (Mask.Val->getNumOperands() <= 4 ||
4731 isSplatMask(Mask.Val) ||
4732 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
4733 X86::isUNPCKLMask(Mask.Val) ||
4734 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
4735 X86::isUNPCKHMask(Mask.Val));
4738 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
4740 SelectionDAG &DAG) const {
4741 unsigned NumElts = BVOps.size();
4742 // Only do shuffles on 128-bit vector types for now.
4743 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
4744 if (NumElts == 2) return true;
4746 return (isMOVLMask(BVOps) || isCommutedMOVL(BVOps, true) ||
4747 isSHUFPMask(BVOps) || isCommutedSHUFP(BVOps));
4752 //===----------------------------------------------------------------------===//
4753 // X86 Scheduler Hooks
4754 //===----------------------------------------------------------------------===//
4757 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
4758 MachineBasicBlock *BB) {
4759 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4760 switch (MI->getOpcode()) {
4761 default: assert(false && "Unexpected instr type to insert");
4762 case X86::CMOV_FR32:
4763 case X86::CMOV_FR64:
4764 case X86::CMOV_V4F32:
4765 case X86::CMOV_V2F64:
4766 case X86::CMOV_V2I64: {
4767 // To "insert" a SELECT_CC instruction, we actually have to insert the
4768 // diamond control-flow pattern. The incoming instruction knows the
4769 // destination vreg to set, the condition code register to branch on, the
4770 // true/false values to select between, and a branch opcode to use.
4771 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4772 ilist<MachineBasicBlock>::iterator It = BB;
4778 // cmpTY ccX, r1, r2
4780 // fallthrough --> copy0MBB
4781 MachineBasicBlock *thisMBB = BB;
4782 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
4783 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
4785 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
4786 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
4787 MachineFunction *F = BB->getParent();
4788 F->getBasicBlockList().insert(It, copy0MBB);
4789 F->getBasicBlockList().insert(It, sinkMBB);
4790 // Update machine-CFG edges by first adding all successors of the current
4791 // block to the new block which will contain the Phi node for the select.
4792 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
4793 e = BB->succ_end(); i != e; ++i)
4794 sinkMBB->addSuccessor(*i);
4795 // Next, remove all successors of the current block, and add the true
4796 // and fallthrough blocks as its successors.
4797 while(!BB->succ_empty())
4798 BB->removeSuccessor(BB->succ_begin());
4799 BB->addSuccessor(copy0MBB);
4800 BB->addSuccessor(sinkMBB);
4803 // %FalseValue = ...
4804 // # fallthrough to sinkMBB
4807 // Update machine-CFG edges
4808 BB->addSuccessor(sinkMBB);
4811 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4814 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
4815 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4816 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4818 delete MI; // The pseudo instruction is gone now.
4822 case X86::FP_TO_INT16_IN_MEM:
4823 case X86::FP_TO_INT32_IN_MEM:
4824 case X86::FP_TO_INT64_IN_MEM: {
4825 // Change the floating point control register to use "round towards zero"
4826 // mode when truncating to an integer value.
4827 MachineFunction *F = BB->getParent();
4828 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
4829 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
4831 // Load the old value of the high byte of the control word...
4833 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
4834 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
4836 // Set the high part to be round to zero...
4837 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
4840 // Reload the modified control word now...
4841 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4843 // Restore the memory image of control word to original value
4844 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
4847 // Get the X86 opcode to use.
4849 switch (MI->getOpcode()) {
4850 default: assert(0 && "illegal opcode!");
4851 case X86::FP_TO_INT16_IN_MEM: Opc = X86::FpIST16m; break;
4852 case X86::FP_TO_INT32_IN_MEM: Opc = X86::FpIST32m; break;
4853 case X86::FP_TO_INT64_IN_MEM: Opc = X86::FpIST64m; break;
4857 MachineOperand &Op = MI->getOperand(0);
4858 if (Op.isRegister()) {
4859 AM.BaseType = X86AddressMode::RegBase;
4860 AM.Base.Reg = Op.getReg();
4862 AM.BaseType = X86AddressMode::FrameIndexBase;
4863 AM.Base.FrameIndex = Op.getFrameIndex();
4865 Op = MI->getOperand(1);
4866 if (Op.isImmediate())
4867 AM.Scale = Op.getImm();
4868 Op = MI->getOperand(2);
4869 if (Op.isImmediate())
4870 AM.IndexReg = Op.getImm();
4871 Op = MI->getOperand(3);
4872 if (Op.isGlobalAddress()) {
4873 AM.GV = Op.getGlobal();
4875 AM.Disp = Op.getImm();
4877 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
4878 .addReg(MI->getOperand(4).getReg());
4880 // Reload the original control word now.
4881 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4883 delete MI; // The pseudo instruction is gone now.
4889 //===----------------------------------------------------------------------===//
4890 // X86 Optimization Hooks
4891 //===----------------------------------------------------------------------===//
4893 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
4895 uint64_t &KnownZero,
4897 unsigned Depth) const {
4898 unsigned Opc = Op.getOpcode();
4899 assert((Opc >= ISD::BUILTIN_OP_END ||
4900 Opc == ISD::INTRINSIC_WO_CHAIN ||
4901 Opc == ISD::INTRINSIC_W_CHAIN ||
4902 Opc == ISD::INTRINSIC_VOID) &&
4903 "Should use MaskedValueIsZero if you don't know whether Op"
4904 " is a target node!");
4906 KnownZero = KnownOne = 0; // Don't know anything.
4910 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
4915 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4916 /// element of the result of the vector shuffle.
4917 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
4918 MVT::ValueType VT = N->getValueType(0);
4919 SDOperand PermMask = N->getOperand(2);
4920 unsigned NumElems = PermMask.getNumOperands();
4921 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
4923 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4925 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4926 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
4927 SDOperand Idx = PermMask.getOperand(i);
4928 if (Idx.getOpcode() == ISD::UNDEF)
4929 return DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4930 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
4935 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
4936 /// node is a GlobalAddress + an offset.
4937 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
4938 unsigned Opc = N->getOpcode();
4939 if (Opc == X86ISD::Wrapper) {
4940 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
4941 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
4944 } else if (Opc == ISD::ADD) {
4945 SDOperand N1 = N->getOperand(0);
4946 SDOperand N2 = N->getOperand(1);
4947 if (isGAPlusOffset(N1.Val, GA, Offset)) {
4948 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
4950 Offset += V->getSignExtended();
4953 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
4954 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
4956 Offset += V->getSignExtended();
4964 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
4966 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
4967 MachineFrameInfo *MFI) {
4968 if (N->getOperand(0).Val != Base->getOperand(0).Val)
4971 SDOperand Loc = N->getOperand(1);
4972 SDOperand BaseLoc = Base->getOperand(1);
4973 if (Loc.getOpcode() == ISD::FrameIndex) {
4974 if (BaseLoc.getOpcode() != ISD::FrameIndex)
4976 int FI = dyn_cast<FrameIndexSDNode>(Loc)->getIndex();
4977 int BFI = dyn_cast<FrameIndexSDNode>(BaseLoc)->getIndex();
4978 int FS = MFI->getObjectSize(FI);
4979 int BFS = MFI->getObjectSize(BFI);
4980 if (FS != BFS || FS != Size) return false;
4981 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
4983 GlobalValue *GV1 = NULL;
4984 GlobalValue *GV2 = NULL;
4985 int64_t Offset1 = 0;
4986 int64_t Offset2 = 0;
4987 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
4988 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
4989 if (isGA1 && isGA2 && GV1 == GV2)
4990 return Offset1 == (Offset2 + Dist*Size);
4996 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
4997 const X86Subtarget *Subtarget) {
5000 if (isGAPlusOffset(Base, GV, Offset))
5001 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
5003 assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
5004 int BFI = dyn_cast<FrameIndexSDNode>(Base)->getIndex();
5006 // Fixed objects do not specify alignment, however the offsets are known.
5007 return ((Subtarget->getStackAlignment() % 16) == 0 &&
5008 (MFI->getObjectOffset(BFI) % 16) == 0);
5010 return MFI->getObjectAlignment(BFI) >= 16;
5016 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
5017 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
5018 /// if the load addresses are consecutive, non-overlapping, and in the right
5020 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
5021 const X86Subtarget *Subtarget) {
5022 MachineFunction &MF = DAG.getMachineFunction();
5023 MachineFrameInfo *MFI = MF.getFrameInfo();
5024 MVT::ValueType VT = N->getValueType(0);
5025 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
5026 SDOperand PermMask = N->getOperand(2);
5027 int NumElems = (int)PermMask.getNumOperands();
5028 SDNode *Base = NULL;
5029 for (int i = 0; i < NumElems; ++i) {
5030 SDOperand Idx = PermMask.getOperand(i);
5031 if (Idx.getOpcode() == ISD::UNDEF) {
5032 if (!Base) return SDOperand();
5035 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
5036 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
5040 else if (!isConsecutiveLoad(Arg.Val, Base,
5041 i, MVT::getSizeInBits(EVT)/8,MFI))
5046 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
5048 LoadSDNode *LD = cast<LoadSDNode>(Base);
5049 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
5050 LD->getSrcValueOffset());
5052 // Just use movups, it's shorter.
5053 std::vector<MVT::ValueType> Tys;
5054 Tys.push_back(MVT::v4f32);
5055 Tys.push_back(MVT::Other);
5056 SmallVector<SDOperand, 3> Ops;
5057 Ops.push_back(Base->getOperand(0));
5058 Ops.push_back(Base->getOperand(1));
5059 Ops.push_back(Base->getOperand(2));
5060 return DAG.getNode(ISD::BIT_CONVERT, VT,
5061 DAG.getNode(X86ISD::LOAD_UA, Tys, &Ops[0], Ops.size()));
5065 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
5066 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
5067 const X86Subtarget *Subtarget) {
5068 SDOperand Cond = N->getOperand(0);
5070 // If we have SSE[12] support, try to form min/max nodes.
5071 if (Subtarget->hasSSE2() &&
5072 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
5073 if (Cond.getOpcode() == ISD::SETCC) {
5074 // Get the LHS/RHS of the select.
5075 SDOperand LHS = N->getOperand(1);
5076 SDOperand RHS = N->getOperand(2);
5077 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
5079 unsigned Opcode = 0;
5080 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
5083 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
5086 if (!UnsafeFPMath) break;
5088 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
5090 Opcode = X86ISD::FMIN;
5093 case ISD::SETOGT: // (X > Y) ? X : Y -> max
5096 if (!UnsafeFPMath) break;
5098 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
5100 Opcode = X86ISD::FMAX;
5103 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
5106 case ISD::SETOGT: // (X > Y) ? Y : X -> min
5109 if (!UnsafeFPMath) break;
5111 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
5113 Opcode = X86ISD::FMIN;
5116 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
5119 if (!UnsafeFPMath) break;
5121 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
5123 Opcode = X86ISD::FMAX;
5129 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
5138 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
5139 DAGCombinerInfo &DCI) const {
5140 SelectionDAG &DAG = DCI.DAG;
5141 switch (N->getOpcode()) {
5143 case ISD::VECTOR_SHUFFLE:
5144 return PerformShuffleCombine(N, DAG, Subtarget);
5146 return PerformSELECTCombine(N, DAG, Subtarget);
5152 //===----------------------------------------------------------------------===//
5153 // X86 Inline Assembly Support
5154 //===----------------------------------------------------------------------===//
5156 /// getConstraintType - Given a constraint letter, return the type of
5157 /// constraint it is for this target.
5158 X86TargetLowering::ConstraintType
5159 X86TargetLowering::getConstraintType(char ConstraintLetter) const {
5160 switch (ConstraintLetter) {
5169 return C_RegisterClass;
5170 default: return TargetLowering::getConstraintType(ConstraintLetter);
5174 /// isOperandValidForConstraint - Return the specified operand (possibly
5175 /// modified) if the specified SDOperand is valid for the specified target
5176 /// constraint letter, otherwise return null.
5177 SDOperand X86TargetLowering::
5178 isOperandValidForConstraint(SDOperand Op, char Constraint, SelectionDAG &DAG) {
5179 switch (Constraint) {
5182 // Literal immediates are always ok.
5183 if (isa<ConstantSDNode>(Op)) return Op;
5185 // If we are in non-pic codegen mode, we allow the address of a global to
5186 // be used with 'i'.
5187 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op)) {
5188 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
5189 return SDOperand(0, 0);
5191 if (GA->getOpcode() != ISD::TargetGlobalAddress)
5192 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5197 // Otherwise, not valid for this mode.
5198 return SDOperand(0, 0);
5200 return TargetLowering::isOperandValidForConstraint(Op, Constraint, DAG);
5204 std::vector<unsigned> X86TargetLowering::
5205 getRegClassForInlineAsmConstraint(const std::string &Constraint,
5206 MVT::ValueType VT) const {
5207 if (Constraint.size() == 1) {
5208 // FIXME: not handling fp-stack yet!
5209 // FIXME: not handling MMX registers yet ('y' constraint).
5210 switch (Constraint[0]) { // GCC X86 Constraint Letters
5211 default: break; // Unknown constraint letter
5212 case 'A': // EAX/EDX
5213 if (VT == MVT::i32 || VT == MVT::i64)
5214 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
5216 case 'r': // GENERAL_REGS
5217 case 'R': // LEGACY_REGS
5218 if (VT == MVT::i64 && Subtarget->is64Bit())
5219 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
5220 X86::RSI, X86::RDI, X86::RBP, X86::RSP,
5221 X86::R8, X86::R9, X86::R10, X86::R11,
5222 X86::R12, X86::R13, X86::R14, X86::R15, 0);
5224 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
5225 X86::ESI, X86::EDI, X86::EBP, X86::ESP, 0);
5226 else if (VT == MVT::i16)
5227 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
5228 X86::SI, X86::DI, X86::BP, X86::SP, 0);
5229 else if (VT == MVT::i8)
5230 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
5232 case 'l': // INDEX_REGS
5234 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
5235 X86::ESI, X86::EDI, X86::EBP, 0);
5236 else if (VT == MVT::i16)
5237 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
5238 X86::SI, X86::DI, X86::BP, 0);
5239 else if (VT == MVT::i8)
5240 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
5242 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
5245 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
5246 else if (VT == MVT::i16)
5247 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
5248 else if (VT == MVT::i8)
5249 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
5251 case 'x': // SSE_REGS if SSE1 allowed
5252 if (Subtarget->hasSSE1())
5253 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
5254 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
5256 return std::vector<unsigned>();
5257 case 'Y': // SSE_REGS if SSE2 allowed
5258 if (Subtarget->hasSSE2())
5259 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
5260 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
5262 return std::vector<unsigned>();
5266 return std::vector<unsigned>();
5269 std::pair<unsigned, const TargetRegisterClass*>
5270 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5271 MVT::ValueType VT) const {
5272 // Use the default implementation in TargetLowering to convert the register
5273 // constraint into a member of a register class.
5274 std::pair<unsigned, const TargetRegisterClass*> Res;
5275 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5277 // Not found as a standard register?
5278 if (Res.second == 0) {
5279 // GCC calls "st(0)" just plain "st".
5280 if (StringsEqualNoCase("{st}", Constraint)) {
5281 Res.first = X86::ST0;
5282 Res.second = X86::RSTRegisterClass;
5288 // Otherwise, check to see if this is a register class of the wrong value
5289 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
5290 // turn into {ax},{dx}.
5291 if (Res.second->hasType(VT))
5292 return Res; // Correct type already, nothing to do.
5294 // All of the single-register GCC register classes map their values onto
5295 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
5296 // really want an 8-bit or 32-bit register, map to the appropriate register
5297 // class and return the appropriate register.
5298 if (Res.second != X86::GR16RegisterClass)
5301 if (VT == MVT::i8) {
5302 unsigned DestReg = 0;
5303 switch (Res.first) {
5305 case X86::AX: DestReg = X86::AL; break;
5306 case X86::DX: DestReg = X86::DL; break;
5307 case X86::CX: DestReg = X86::CL; break;
5308 case X86::BX: DestReg = X86::BL; break;
5311 Res.first = DestReg;
5312 Res.second = Res.second = X86::GR8RegisterClass;
5314 } else if (VT == MVT::i32) {
5315 unsigned DestReg = 0;
5316 switch (Res.first) {
5318 case X86::AX: DestReg = X86::EAX; break;
5319 case X86::DX: DestReg = X86::EDX; break;
5320 case X86::CX: DestReg = X86::ECX; break;
5321 case X86::BX: DestReg = X86::EBX; break;
5322 case X86::SI: DestReg = X86::ESI; break;
5323 case X86::DI: DestReg = X86::EDI; break;
5324 case X86::BP: DestReg = X86::EBP; break;
5325 case X86::SP: DestReg = X86::ESP; break;
5328 Res.first = DestReg;
5329 Res.second = Res.second = X86::GR32RegisterClass;
5331 } else if (VT == MVT::i64) {
5332 unsigned DestReg = 0;
5333 switch (Res.first) {
5335 case X86::AX: DestReg = X86::RAX; break;
5336 case X86::DX: DestReg = X86::RDX; break;
5337 case X86::CX: DestReg = X86::RCX; break;
5338 case X86::BX: DestReg = X86::RBX; break;
5339 case X86::SI: DestReg = X86::RSI; break;
5340 case X86::DI: DestReg = X86::RDI; break;
5341 case X86::BP: DestReg = X86::RBP; break;
5342 case X86::SP: DestReg = X86::RSP; break;
5345 Res.first = DestReg;
5346 Res.second = Res.second = X86::GR64RegisterClass;