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/CallingConvLower.h"
28 #include "llvm/CodeGen/MachineFrameInfo.h"
29 #include "llvm/CodeGen/MachineFunction.h"
30 #include "llvm/CodeGen/MachineInstrBuilder.h"
31 #include "llvm/CodeGen/SelectionDAG.h"
32 #include "llvm/CodeGen/SSARegMap.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Target/TargetOptions.h"
35 #include "llvm/ADT/StringExtras.h"
38 X86TargetLowering::X86TargetLowering(TargetMachine &TM)
39 : TargetLowering(TM) {
40 Subtarget = &TM.getSubtarget<X86Subtarget>();
41 X86ScalarSSE = Subtarget->hasSSE2();
42 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
44 // Set up the TargetLowering object.
46 // X86 is weird, it always uses i8 for shift amounts and setcc results.
47 setShiftAmountType(MVT::i8);
48 setSetCCResultType(MVT::i8);
49 setSetCCResultContents(ZeroOrOneSetCCResult);
50 setSchedulingPreference(SchedulingForRegPressure);
51 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
52 setStackPointerRegisterToSaveRestore(X86StackPtr);
54 if (Subtarget->isTargetDarwin()) {
55 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
56 setUseUnderscoreSetJmp(false);
57 setUseUnderscoreLongJmp(false);
58 } else if (Subtarget->isTargetMingw()) {
59 // MS runtime is weird: it exports _setjmp, but longjmp!
60 setUseUnderscoreSetJmp(true);
61 setUseUnderscoreLongJmp(false);
63 setUseUnderscoreSetJmp(true);
64 setUseUnderscoreLongJmp(true);
67 // Add legal addressing mode scale values.
68 addLegalAddressScale(8);
69 addLegalAddressScale(4);
70 addLegalAddressScale(2);
71 // Enter the ones which require both scale + index last. These are more
73 addLegalAddressScale(9);
74 addLegalAddressScale(5);
75 addLegalAddressScale(3);
77 // Set up the register classes.
78 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
79 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
80 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
81 if (Subtarget->is64Bit())
82 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
84 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Expand);
86 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
88 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
89 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
90 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
92 if (Subtarget->is64Bit()) {
93 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
94 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
97 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
98 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
100 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
103 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
105 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
106 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
107 // SSE has no i16 to fp conversion, only i32
109 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
111 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
112 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
115 if (!Subtarget->is64Bit()) {
116 // Custom lower SINT_TO_FP and FP_TO_SINT from/to i64 in 32-bit mode.
117 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
118 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
121 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
123 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
124 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
127 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
129 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
130 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
133 // Handle FP_TO_UINT by promoting the destination to a larger signed
135 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
136 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
137 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
139 if (Subtarget->is64Bit()) {
140 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
141 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
143 if (X86ScalarSSE && !Subtarget->hasSSE3())
144 // Expand FP_TO_UINT into a select.
145 // FIXME: We would like to use a Custom expander here eventually to do
146 // the optimal thing for SSE vs. the default expansion in the legalizer.
147 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
149 // With SSE3 we can use fisttpll to convert to a signed i64.
150 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
153 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
155 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
156 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
159 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
160 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
161 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
162 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
163 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
164 if (Subtarget->is64Bit())
165 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Expand);
166 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Expand);
167 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Expand);
168 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
169 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
170 setOperationAction(ISD::FREM , MVT::f64 , Expand);
172 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
173 setOperationAction(ISD::CTTZ , MVT::i8 , Expand);
174 setOperationAction(ISD::CTLZ , MVT::i8 , Expand);
175 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
176 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
177 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
178 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
179 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
180 setOperationAction(ISD::CTLZ , MVT::i32 , Expand);
181 if (Subtarget->is64Bit()) {
182 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
183 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
184 setOperationAction(ISD::CTLZ , MVT::i64 , Expand);
187 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
188 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
190 // These should be promoted to a larger select which is supported.
191 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
192 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
193 // X86 wants to expand cmov itself.
194 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
195 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
196 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
197 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
198 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
199 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
200 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
201 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
202 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
203 if (Subtarget->is64Bit()) {
204 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
205 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
207 // X86 ret instruction may pop stack.
208 setOperationAction(ISD::RET , MVT::Other, Custom);
210 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
211 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
212 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
213 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
214 if (Subtarget->is64Bit()) {
215 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
216 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
217 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
218 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
220 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
221 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
222 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
223 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
224 // X86 wants to expand memset / memcpy itself.
225 setOperationAction(ISD::MEMSET , MVT::Other, Custom);
226 setOperationAction(ISD::MEMCPY , MVT::Other, Custom);
228 // We don't have line number support yet.
229 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
230 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
231 // FIXME - use subtarget debug flags
232 if (!Subtarget->isTargetDarwin() &&
233 !Subtarget->isTargetELF() &&
234 !Subtarget->isTargetCygMing())
235 setOperationAction(ISD::LABEL, MVT::Other, Expand);
237 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
238 setOperationAction(ISD::VASTART , MVT::Other, Custom);
240 // Use the default implementation.
241 setOperationAction(ISD::VAARG , MVT::Other, Expand);
242 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
243 setOperationAction(ISD::VAEND , MVT::Other, Expand);
244 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
245 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
246 if (Subtarget->is64Bit())
247 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
248 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
251 // Set up the FP register classes.
252 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
253 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
255 // Use ANDPD to simulate FABS.
256 setOperationAction(ISD::FABS , MVT::f64, Custom);
257 setOperationAction(ISD::FABS , MVT::f32, Custom);
259 // Use XORP to simulate FNEG.
260 setOperationAction(ISD::FNEG , MVT::f64, Custom);
261 setOperationAction(ISD::FNEG , MVT::f32, Custom);
263 // Use ANDPD and ORPD to simulate FCOPYSIGN.
264 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
265 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
267 // We don't support sin/cos/fmod
268 setOperationAction(ISD::FSIN , MVT::f64, Expand);
269 setOperationAction(ISD::FCOS , MVT::f64, Expand);
270 setOperationAction(ISD::FREM , MVT::f64, Expand);
271 setOperationAction(ISD::FSIN , MVT::f32, Expand);
272 setOperationAction(ISD::FCOS , MVT::f32, Expand);
273 setOperationAction(ISD::FREM , MVT::f32, Expand);
275 // Expand FP immediates into loads from the stack, except for the special
277 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
278 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
279 addLegalFPImmediate(+0.0); // xorps / xorpd
281 // Set up the FP register classes.
282 addRegisterClass(MVT::f64, X86::RFPRegisterClass);
284 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
285 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
286 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
289 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
290 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
293 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
294 addLegalFPImmediate(+0.0); // FLD0
295 addLegalFPImmediate(+1.0); // FLD1
296 addLegalFPImmediate(-0.0); // FLD0/FCHS
297 addLegalFPImmediate(-1.0); // FLD1/FCHS
300 // First set operation action for all vector types to expand. Then we
301 // will selectively turn on ones that can be effectively codegen'd.
302 for (unsigned VT = (unsigned)MVT::Vector + 1;
303 VT != (unsigned)MVT::LAST_VALUETYPE; VT++) {
304 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
305 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
306 setOperationAction(ISD::FADD, (MVT::ValueType)VT, Expand);
307 setOperationAction(ISD::FSUB, (MVT::ValueType)VT, Expand);
308 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
309 setOperationAction(ISD::FMUL, (MVT::ValueType)VT, Expand);
310 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
311 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
312 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
313 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
314 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
315 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand);
316 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Expand);
317 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
318 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
321 if (Subtarget->hasMMX()) {
322 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
323 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
324 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
326 // FIXME: add MMX packed arithmetics
327 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Expand);
328 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Expand);
329 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Expand);
332 if (Subtarget->hasSSE1()) {
333 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
335 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
336 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
337 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
338 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
339 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
340 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
341 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
342 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
343 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
346 if (Subtarget->hasSSE2()) {
347 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
348 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
349 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
350 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
351 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
353 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
354 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
355 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
356 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
357 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
358 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
359 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
360 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
361 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
362 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
363 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
365 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
366 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
367 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
368 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
369 // Implement v4f32 insert_vector_elt in terms of SSE2 v8i16 ones.
370 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
372 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
373 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
374 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
375 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
376 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
378 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
379 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
380 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
381 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
382 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
383 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
385 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
386 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
387 setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote);
388 AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64);
389 setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote);
390 AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64);
391 setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote);
392 AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64);
393 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote);
394 AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64);
395 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
396 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64);
399 // Custom lower v2i64 and v2f64 selects.
400 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
401 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
402 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
403 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
406 // We want to custom lower some of our intrinsics.
407 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
409 // We have target-specific dag combine patterns for the following nodes:
410 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
411 setTargetDAGCombine(ISD::SELECT);
413 computeRegisterProperties();
415 // FIXME: These should be based on subtarget info. Plus, the values should
416 // be smaller when we are in optimizing for size mode.
417 maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
418 maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
419 maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
420 allowUnalignedMemoryAccesses = true; // x86 supports it!
424 //===----------------------------------------------------------------------===//
425 // Return Value Calling Convention Implementation
426 //===----------------------------------------------------------------------===//
428 #include "X86GenCallingConv.inc"
430 /// LowerRET - Lower an ISD::RET node.
431 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
432 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
434 SmallVector<CCValAssign, 16> RVLocs;
435 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
436 CCState CCInfo(CC, getTargetMachine(), RVLocs);
437 CCInfo.AnalyzeReturn(Op.Val, RetCC_X86);
440 // If this is the first return lowered for this function, add the regs to the
441 // liveout set for the function.
442 if (DAG.getMachineFunction().liveout_empty()) {
443 for (unsigned i = 0; i != RVLocs.size(); ++i)
444 if (RVLocs[i].isRegLoc())
445 DAG.getMachineFunction().addLiveOut(RVLocs[i].getLocReg());
448 SDOperand Chain = Op.getOperand(0);
451 // Copy the result values into the output registers.
452 if (RVLocs.size() != 1 || !RVLocs[0].isRegLoc() ||
453 RVLocs[0].getLocReg() != X86::ST0) {
454 for (unsigned i = 0; i != RVLocs.size(); ++i) {
455 CCValAssign &VA = RVLocs[i];
456 assert(VA.isRegLoc() && "Can only return in registers!");
457 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1),
459 Flag = Chain.getValue(1);
462 // We need to handle a destination of ST0 specially, because it isn't really
464 SDOperand Value = Op.getOperand(1);
466 // If this is an FP return with ScalarSSE, we need to move the value from
467 // an XMM register onto the fp-stack.
471 // If this is a load into a scalarsse value, don't store the loaded value
472 // back to the stack, only to reload it: just replace the scalar-sse load.
473 if (ISD::isNON_EXTLoad(Value.Val) &&
474 (Chain == Value.getValue(1) || Chain == Value.getOperand(0))) {
475 Chain = Value.getOperand(0);
476 MemLoc = Value.getOperand(1);
478 // Spill the value to memory and reload it into top of stack.
479 unsigned Size = MVT::getSizeInBits(RVLocs[0].getValVT())/8;
480 MachineFunction &MF = DAG.getMachineFunction();
481 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
482 MemLoc = DAG.getFrameIndex(SSFI, getPointerTy());
483 Chain = DAG.getStore(Op.getOperand(0), Value, MemLoc, NULL, 0);
485 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
486 SDOperand Ops[] = {Chain, MemLoc, DAG.getValueType(RVLocs[0].getValVT())};
487 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
488 Chain = Value.getValue(1);
491 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
492 SDOperand Ops[] = { Chain, Value };
493 Chain = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops, 2);
494 Flag = Chain.getValue(1);
497 SDOperand BytesToPop = DAG.getConstant(getBytesToPopOnReturn(), MVT::i16);
499 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop, Flag);
501 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop);
505 /// LowerCallResult - Lower the result values of an ISD::CALL into the
506 /// appropriate copies out of appropriate physical registers. This assumes that
507 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
508 /// being lowered. The returns a SDNode with the same number of values as the
510 SDNode *X86TargetLowering::
511 LowerCallResult(SDOperand Chain, SDOperand InFlag, SDNode *TheCall,
512 unsigned CallingConv, SelectionDAG &DAG) {
514 // Assign locations to each value returned by this call.
515 SmallVector<CCValAssign, 16> RVLocs;
516 CCState CCInfo(CallingConv, getTargetMachine(), RVLocs);
517 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
520 SmallVector<SDOperand, 8> ResultVals;
522 // Copy all of the result registers out of their specified physreg.
523 if (RVLocs.size() != 1 || RVLocs[0].getLocReg() != X86::ST0) {
524 for (unsigned i = 0; i != RVLocs.size(); ++i) {
525 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
526 RVLocs[i].getValVT(), InFlag).getValue(1);
527 InFlag = Chain.getValue(2);
528 ResultVals.push_back(Chain.getValue(0));
531 // Copies from the FP stack are special, as ST0 isn't a valid register
532 // before the fp stackifier runs.
534 // Copy ST0 into an RFP register with FP_GET_RESULT.
535 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
536 SDOperand GROps[] = { Chain, InFlag };
537 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, GROps, 2);
538 Chain = RetVal.getValue(1);
539 InFlag = RetVal.getValue(2);
541 // If we are using ScalarSSE, store ST(0) to the stack and reload it into
544 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
545 // shouldn't be necessary except that RFP cannot be live across
546 // multiple blocks. When stackifier is fixed, they can be uncoupled.
547 MachineFunction &MF = DAG.getMachineFunction();
548 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
549 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
551 Chain, RetVal, StackSlot, DAG.getValueType(RVLocs[0].getValVT()), InFlag
553 Chain = DAG.getNode(X86ISD::FST, MVT::Other, Ops, 5);
554 RetVal = DAG.getLoad(RVLocs[0].getValVT(), Chain, StackSlot, NULL, 0);
555 Chain = RetVal.getValue(1);
558 if (RVLocs[0].getValVT() == MVT::f32 && !X86ScalarSSE)
559 // FIXME: we would really like to remember that this FP_ROUND
560 // operation is okay to eliminate if we allow excess FP precision.
561 RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal);
562 ResultVals.push_back(RetVal);
565 // Merge everything together with a MERGE_VALUES node.
566 ResultVals.push_back(Chain);
567 return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(),
568 &ResultVals[0], ResultVals.size()).Val;
572 //===----------------------------------------------------------------------===//
573 // C & StdCall Calling Convention implementation
574 //===----------------------------------------------------------------------===//
575 // StdCall calling convention seems to be standard for many Windows' API
576 // routines and around. It differs from C calling convention just a little:
577 // callee should clean up the stack, not caller. Symbols should be also
578 // decorated in some fancy way :) It doesn't support any vector arguments.
580 /// AddLiveIn - This helper function adds the specified physical register to the
581 /// MachineFunction as a live in value. It also creates a corresponding virtual
583 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
584 const TargetRegisterClass *RC) {
585 assert(RC->contains(PReg) && "Not the correct regclass!");
586 unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC);
587 MF.addLiveIn(PReg, VReg);
591 SDOperand X86TargetLowering::LowerCCCArguments(SDOperand Op, SelectionDAG &DAG,
593 unsigned NumArgs = Op.Val->getNumValues() - 1;
594 MachineFunction &MF = DAG.getMachineFunction();
595 MachineFrameInfo *MFI = MF.getFrameInfo();
596 SDOperand Root = Op.getOperand(0);
597 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
599 // Assign locations to all of the incoming arguments.
600 SmallVector<CCValAssign, 16> ArgLocs;
601 CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(),
603 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_C);
605 SmallVector<SDOperand, 8> ArgValues;
606 unsigned LastVal = ~0U;
607 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
608 CCValAssign &VA = ArgLocs[i];
609 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
611 assert(VA.getValNo() != LastVal &&
612 "Don't support value assigned to multiple locs yet");
613 LastVal = VA.getValNo();
616 MVT::ValueType RegVT = VA.getLocVT();
617 TargetRegisterClass *RC;
618 if (RegVT == MVT::i32)
619 RC = X86::GR32RegisterClass;
621 assert(MVT::isVector(RegVT));
622 RC = X86::VR128RegisterClass;
625 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
626 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
628 // If this is an 8 or 16-bit value, it is really passed promoted to 32
629 // bits. Insert an assert[sz]ext to capture this, then truncate to the
631 if (VA.getLocInfo() == CCValAssign::SExt)
632 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
633 DAG.getValueType(VA.getValVT()));
634 else if (VA.getLocInfo() == CCValAssign::ZExt)
635 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
636 DAG.getValueType(VA.getValVT()));
638 if (VA.getLocInfo() != CCValAssign::Full)
639 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
641 ArgValues.push_back(ArgValue);
643 assert(VA.isMemLoc());
645 // Create the nodes corresponding to a load from this parameter slot.
646 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
647 VA.getLocMemOffset());
648 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
649 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
653 unsigned StackSize = CCInfo.getNextStackOffset();
655 ArgValues.push_back(Root);
657 // If the function takes variable number of arguments, make a frame index for
658 // the start of the first vararg value... for expansion of llvm.va_start.
660 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
662 if (isStdCall && !isVarArg) {
663 BytesToPopOnReturn = StackSize; // Callee pops everything..
664 BytesCallerReserves = 0;
666 BytesToPopOnReturn = 0; // Callee pops hidden struct pointer.
668 // If this is an sret function, the return should pop the hidden pointer.
669 if (NumArgs && (cast<ConstantSDNode>(Op.getOperand(3))->getValue() & 4))
670 BytesToPopOnReturn = 4;
672 BytesCallerReserves = StackSize;
675 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
676 ReturnAddrIndex = 0; // No return address slot generated yet.
678 MF.getInfo<X86FunctionInfo>()->setBytesToPopOnReturn(BytesToPopOnReturn);
680 // Return the new list of results.
681 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
682 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
685 SDOperand X86TargetLowering::LowerCCCCallTo(SDOperand Op, SelectionDAG &DAG,
687 SDOperand Chain = Op.getOperand(0);
688 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
689 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
690 SDOperand Callee = Op.getOperand(4);
691 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
693 // Analyze operands of the call, assigning locations to each operand.
694 SmallVector<CCValAssign, 16> ArgLocs;
695 CCState CCInfo(CC, getTargetMachine(), ArgLocs);
696 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_C);
698 // Get a count of how many bytes are to be pushed on the stack.
699 unsigned NumBytes = CCInfo.getNextStackOffset();
701 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
703 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
704 SmallVector<SDOperand, 8> MemOpChains;
708 // Walk the register/memloc assignments, inserting copies/loads.
709 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
710 CCValAssign &VA = ArgLocs[i];
711 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
713 // Promote the value if needed.
714 switch (VA.getLocInfo()) {
715 default: assert(0 && "Unknown loc info!");
716 case CCValAssign::Full: break;
717 case CCValAssign::SExt:
718 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
720 case CCValAssign::ZExt:
721 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
723 case CCValAssign::AExt:
724 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
729 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
731 assert(VA.isMemLoc());
732 if (StackPtr.Val == 0)
733 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
734 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
735 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
736 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
740 // If the first argument is an sret pointer, remember it.
741 bool isSRet = NumOps &&(cast<ConstantSDNode>(Op.getOperand(6))->getValue()&4);
743 if (!MemOpChains.empty())
744 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
745 &MemOpChains[0], MemOpChains.size());
747 // Build a sequence of copy-to-reg nodes chained together with token chain
748 // and flag operands which copy the outgoing args into registers.
750 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
751 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
753 InFlag = Chain.getValue(1);
756 // ELF / PIC requires GOT in the EBX register before function calls via PLT
758 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
759 Subtarget->isPICStyleGOT()) {
760 Chain = DAG.getCopyToReg(Chain, X86::EBX,
761 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
763 InFlag = Chain.getValue(1);
766 // If the callee is a GlobalAddress node (quite common, every direct call is)
767 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
768 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
769 // We should use extra load for direct calls to dllimported functions in
771 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
772 getTargetMachine(), true))
773 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
774 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
775 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
777 // Returns a chain & a flag for retval copy to use.
778 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
779 SmallVector<SDOperand, 8> Ops;
780 Ops.push_back(Chain);
781 Ops.push_back(Callee);
783 // Add argument registers to the end of the list so that they are known live
785 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
786 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
787 RegsToPass[i].second.getValueType()));
789 // Add an implicit use GOT pointer in EBX.
790 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
791 Subtarget->isPICStyleGOT())
792 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
795 Ops.push_back(InFlag);
797 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
798 NodeTys, &Ops[0], Ops.size());
799 InFlag = Chain.getValue(1);
801 // Create the CALLSEQ_END node.
802 unsigned NumBytesForCalleeToPush = 0;
804 if (CC == CallingConv::X86_StdCall) {
806 NumBytesForCalleeToPush = isSRet ? 4 : 0;
808 NumBytesForCalleeToPush = NumBytes;
810 // If this is is a call to a struct-return function, the callee
811 // pops the hidden struct pointer, so we have to push it back.
812 // This is common for Darwin/X86, Linux & Mingw32 targets.
813 NumBytesForCalleeToPush = isSRet ? 4 : 0;
816 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
818 Ops.push_back(Chain);
819 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
820 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
821 Ops.push_back(InFlag);
822 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
823 InFlag = Chain.getValue(1);
825 // Handle result values, copying them out of physregs into vregs that we
827 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
831 //===----------------------------------------------------------------------===//
832 // FastCall Calling Convention implementation
833 //===----------------------------------------------------------------------===//
835 // The X86 'fastcall' calling convention passes up to two integer arguments in
836 // registers (an appropriate portion of ECX/EDX), passes arguments in C order,
837 // and requires that the callee pop its arguments off the stack (allowing proper
838 // tail calls), and has the same return value conventions as C calling convs.
840 // This calling convention always arranges for the callee pop value to be 8n+4
841 // bytes, which is needed for tail recursion elimination and stack alignment
844 X86TargetLowering::LowerFastCCArguments(SDOperand Op, SelectionDAG &DAG) {
845 MachineFunction &MF = DAG.getMachineFunction();
846 MachineFrameInfo *MFI = MF.getFrameInfo();
847 SDOperand Root = Op.getOperand(0);
849 // Assign locations to all of the incoming arguments.
850 SmallVector<CCValAssign, 16> ArgLocs;
851 CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(),
853 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_FastCall);
855 SmallVector<SDOperand, 8> ArgValues;
856 unsigned LastVal = ~0U;
857 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
858 CCValAssign &VA = ArgLocs[i];
859 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
861 assert(VA.getValNo() != LastVal &&
862 "Don't support value assigned to multiple locs yet");
863 LastVal = VA.getValNo();
866 MVT::ValueType RegVT = VA.getLocVT();
867 TargetRegisterClass *RC;
868 if (RegVT == MVT::i32)
869 RC = X86::GR32RegisterClass;
871 assert(MVT::isVector(RegVT));
872 RC = X86::VR128RegisterClass;
875 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
876 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
878 // If this is an 8 or 16-bit value, it is really passed promoted to 32
879 // bits. Insert an assert[sz]ext to capture this, then truncate to the
881 if (VA.getLocInfo() == CCValAssign::SExt)
882 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
883 DAG.getValueType(VA.getValVT()));
884 else if (VA.getLocInfo() == CCValAssign::ZExt)
885 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
886 DAG.getValueType(VA.getValVT()));
888 if (VA.getLocInfo() != CCValAssign::Full)
889 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
891 ArgValues.push_back(ArgValue);
893 assert(VA.isMemLoc());
895 // Create the nodes corresponding to a load from this parameter slot.
896 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
897 VA.getLocMemOffset());
898 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
899 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
903 ArgValues.push_back(Root);
905 unsigned StackSize = CCInfo.getNextStackOffset();
907 if (!Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows()) {
908 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
909 // arguments and the arguments after the retaddr has been pushed are aligned.
910 if ((StackSize & 7) == 0)
914 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
915 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
916 ReturnAddrIndex = 0; // No return address slot generated yet.
917 BytesToPopOnReturn = StackSize; // Callee pops all stack arguments.
918 BytesCallerReserves = 0;
920 MF.getInfo<X86FunctionInfo>()->setBytesToPopOnReturn(BytesToPopOnReturn);
922 // Return the new list of results.
923 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
924 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
927 SDOperand X86TargetLowering::LowerFastCCCallTo(SDOperand Op, SelectionDAG &DAG,
929 SDOperand Chain = Op.getOperand(0);
930 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
931 SDOperand Callee = Op.getOperand(4);
933 // Analyze operands of the call, assigning locations to each operand.
934 SmallVector<CCValAssign, 16> ArgLocs;
935 CCState CCInfo(CC, getTargetMachine(), ArgLocs);
936 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_FastCall);
938 // Get a count of how many bytes are to be pushed on the stack.
939 unsigned NumBytes = CCInfo.getNextStackOffset();
941 if (!Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows()) {
942 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
943 // arguments and the arguments after the retaddr has been pushed are aligned.
944 if ((NumBytes & 7) == 0)
948 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
950 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
951 SmallVector<SDOperand, 8> MemOpChains;
955 // Walk the register/memloc assignments, inserting copies/loads.
956 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
957 CCValAssign &VA = ArgLocs[i];
958 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
960 // Promote the value if needed.
961 switch (VA.getLocInfo()) {
962 default: assert(0 && "Unknown loc info!");
963 case CCValAssign::Full: break;
964 case CCValAssign::SExt:
965 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
967 case CCValAssign::ZExt:
968 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
970 case CCValAssign::AExt:
971 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
976 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
978 assert(VA.isMemLoc());
979 if (StackPtr.Val == 0)
980 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
981 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
982 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
983 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
987 if (!MemOpChains.empty())
988 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
989 &MemOpChains[0], MemOpChains.size());
991 // Build a sequence of copy-to-reg nodes chained together with token chain
992 // and flag operands which copy the outgoing args into registers.
994 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
995 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
997 InFlag = Chain.getValue(1);
1000 // If the callee is a GlobalAddress node (quite common, every direct call is)
1001 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1002 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1003 // We should use extra load for direct calls to dllimported functions in
1005 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1006 getTargetMachine(), true))
1007 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1008 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1009 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1011 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1013 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1014 Subtarget->isPICStyleGOT()) {
1015 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1016 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1018 InFlag = Chain.getValue(1);
1021 // Returns a chain & a flag for retval copy to use.
1022 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1023 SmallVector<SDOperand, 8> Ops;
1024 Ops.push_back(Chain);
1025 Ops.push_back(Callee);
1027 // Add argument registers to the end of the list so that they are known live
1029 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1030 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1031 RegsToPass[i].second.getValueType()));
1033 // Add an implicit use GOT pointer in EBX.
1034 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1035 Subtarget->isPICStyleGOT())
1036 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1039 Ops.push_back(InFlag);
1041 // FIXME: Do not generate X86ISD::TAILCALL for now.
1042 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1043 NodeTys, &Ops[0], Ops.size());
1044 InFlag = Chain.getValue(1);
1046 // Returns a flag for retval copy to use.
1047 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1049 Ops.push_back(Chain);
1050 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1051 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1052 Ops.push_back(InFlag);
1053 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1054 InFlag = Chain.getValue(1);
1056 // Handle result values, copying them out of physregs into vregs that we
1058 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1062 //===----------------------------------------------------------------------===//
1063 // X86-64 C Calling Convention implementation
1064 //===----------------------------------------------------------------------===//
1067 X86TargetLowering::LowerX86_64CCCArguments(SDOperand Op, SelectionDAG &DAG) {
1068 MachineFunction &MF = DAG.getMachineFunction();
1069 MachineFrameInfo *MFI = MF.getFrameInfo();
1070 SDOperand Root = Op.getOperand(0);
1071 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1073 static const unsigned GPR64ArgRegs[] = {
1074 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1076 static const unsigned XMMArgRegs[] = {
1077 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1078 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1082 // Assign locations to all of the incoming arguments.
1083 SmallVector<CCValAssign, 16> ArgLocs;
1084 CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(),
1086 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_64_C);
1088 SmallVector<SDOperand, 8> ArgValues;
1089 unsigned LastVal = ~0U;
1090 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1091 CCValAssign &VA = ArgLocs[i];
1092 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1094 assert(VA.getValNo() != LastVal &&
1095 "Don't support value assigned to multiple locs yet");
1096 LastVal = VA.getValNo();
1098 if (VA.isRegLoc()) {
1099 MVT::ValueType RegVT = VA.getLocVT();
1100 TargetRegisterClass *RC;
1101 if (RegVT == MVT::i32)
1102 RC = X86::GR32RegisterClass;
1103 else if (RegVT == MVT::i64)
1104 RC = X86::GR64RegisterClass;
1105 else if (RegVT == MVT::f32)
1106 RC = X86::FR32RegisterClass;
1107 else if (RegVT == MVT::f64)
1108 RC = X86::FR64RegisterClass;
1110 assert(MVT::isVector(RegVT));
1111 RC = X86::VR128RegisterClass;
1114 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1115 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1117 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1118 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1120 if (VA.getLocInfo() == CCValAssign::SExt)
1121 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1122 DAG.getValueType(VA.getValVT()));
1123 else if (VA.getLocInfo() == CCValAssign::ZExt)
1124 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1125 DAG.getValueType(VA.getValVT()));
1127 if (VA.getLocInfo() != CCValAssign::Full)
1128 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1130 ArgValues.push_back(ArgValue);
1132 assert(VA.isMemLoc());
1134 // Create the nodes corresponding to a load from this parameter slot.
1135 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
1136 VA.getLocMemOffset());
1137 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1138 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
1142 unsigned StackSize = CCInfo.getNextStackOffset();
1144 // If the function takes variable number of arguments, make a frame index for
1145 // the start of the first vararg value... for expansion of llvm.va_start.
1147 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs, 6);
1148 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1150 // For X86-64, if there are vararg parameters that are passed via
1151 // registers, then we must store them to their spots on the stack so they
1152 // may be loaded by deferencing the result of va_next.
1153 VarArgsGPOffset = NumIntRegs * 8;
1154 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1155 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1156 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1158 // Store the integer parameter registers.
1159 SmallVector<SDOperand, 8> MemOps;
1160 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1161 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1162 DAG.getConstant(VarArgsGPOffset, getPointerTy()));
1163 for (; NumIntRegs != 6; ++NumIntRegs) {
1164 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1165 X86::GR64RegisterClass);
1166 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1167 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1168 MemOps.push_back(Store);
1169 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1170 DAG.getConstant(8, getPointerTy()));
1173 // Now store the XMM (fp + vector) parameter registers.
1174 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1175 DAG.getConstant(VarArgsFPOffset, getPointerTy()));
1176 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1177 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1178 X86::VR128RegisterClass);
1179 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1180 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1181 MemOps.push_back(Store);
1182 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1183 DAG.getConstant(16, getPointerTy()));
1185 if (!MemOps.empty())
1186 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1187 &MemOps[0], MemOps.size());
1190 ArgValues.push_back(Root);
1192 ReturnAddrIndex = 0; // No return address slot generated yet.
1193 BytesToPopOnReturn = 0; // Callee pops nothing.
1194 BytesCallerReserves = StackSize;
1196 // Return the new list of results.
1197 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1198 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1202 X86TargetLowering::LowerX86_64CCCCallTo(SDOperand Op, SelectionDAG &DAG,
1204 SDOperand Chain = Op.getOperand(0);
1205 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1206 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1207 SDOperand Callee = Op.getOperand(4);
1209 // Analyze operands of the call, assigning locations to each operand.
1210 SmallVector<CCValAssign, 16> ArgLocs;
1211 CCState CCInfo(CC, getTargetMachine(), ArgLocs);
1212 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_C);
1214 // Get a count of how many bytes are to be pushed on the stack.
1215 unsigned NumBytes = CCInfo.getNextStackOffset();
1216 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1218 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1219 SmallVector<SDOperand, 8> MemOpChains;
1223 // Walk the register/memloc assignments, inserting copies/loads.
1224 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1225 CCValAssign &VA = ArgLocs[i];
1226 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1228 // Promote the value if needed.
1229 switch (VA.getLocInfo()) {
1230 default: assert(0 && "Unknown loc info!");
1231 case CCValAssign::Full: break;
1232 case CCValAssign::SExt:
1233 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1235 case CCValAssign::ZExt:
1236 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1238 case CCValAssign::AExt:
1239 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1243 if (VA.isRegLoc()) {
1244 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1246 assert(VA.isMemLoc());
1247 if (StackPtr.Val == 0)
1248 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1249 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
1250 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1251 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1255 if (!MemOpChains.empty())
1256 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1257 &MemOpChains[0], MemOpChains.size());
1259 // Build a sequence of copy-to-reg nodes chained together with token chain
1260 // and flag operands which copy the outgoing args into registers.
1262 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1263 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1265 InFlag = Chain.getValue(1);
1269 // From AMD64 ABI document:
1270 // For calls that may call functions that use varargs or stdargs
1271 // (prototype-less calls or calls to functions containing ellipsis (...) in
1272 // the declaration) %al is used as hidden argument to specify the number
1273 // of SSE registers used. The contents of %al do not need to match exactly
1274 // the number of registers, but must be an ubound on the number of SSE
1275 // registers used and is in the range 0 - 8 inclusive.
1277 // Count the number of XMM registers allocated.
1278 static const unsigned XMMArgRegs[] = {
1279 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1280 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1282 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1284 Chain = DAG.getCopyToReg(Chain, X86::AL,
1285 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1286 InFlag = Chain.getValue(1);
1289 // If the callee is a GlobalAddress node (quite common, every direct call is)
1290 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1291 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1292 // We should use extra load for direct calls to dllimported functions in
1294 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1295 getTargetMachine(), true))
1296 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1297 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1298 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1300 // Returns a chain & a flag for retval copy to use.
1301 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1302 SmallVector<SDOperand, 8> Ops;
1303 Ops.push_back(Chain);
1304 Ops.push_back(Callee);
1306 // Add argument registers to the end of the list so that they are known live
1308 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1309 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1310 RegsToPass[i].second.getValueType()));
1313 Ops.push_back(InFlag);
1315 // FIXME: Do not generate X86ISD::TAILCALL for now.
1316 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1317 NodeTys, &Ops[0], Ops.size());
1318 InFlag = Chain.getValue(1);
1320 // Returns a flag for retval copy to use.
1321 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1323 Ops.push_back(Chain);
1324 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1325 Ops.push_back(DAG.getConstant(0, getPointerTy()));
1326 Ops.push_back(InFlag);
1327 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1328 InFlag = Chain.getValue(1);
1330 // Handle result values, copying them out of physregs into vregs that we
1332 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1336 //===----------------------------------------------------------------------===//
1337 // Other Lowering Hooks
1338 //===----------------------------------------------------------------------===//
1341 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1342 if (ReturnAddrIndex == 0) {
1343 // Set up a frame object for the return address.
1344 MachineFunction &MF = DAG.getMachineFunction();
1345 if (Subtarget->is64Bit())
1346 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
1348 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
1351 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1356 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1357 /// specific condition code. It returns a false if it cannot do a direct
1358 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1360 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1361 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
1362 SelectionDAG &DAG) {
1363 X86CC = X86::COND_INVALID;
1365 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1366 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1367 // X > -1 -> X == 0, jump !sign.
1368 RHS = DAG.getConstant(0, RHS.getValueType());
1369 X86CC = X86::COND_NS;
1371 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1372 // X < 0 -> X == 0, jump on sign.
1373 X86CC = X86::COND_S;
1378 switch (SetCCOpcode) {
1380 case ISD::SETEQ: X86CC = X86::COND_E; break;
1381 case ISD::SETGT: X86CC = X86::COND_G; break;
1382 case ISD::SETGE: X86CC = X86::COND_GE; break;
1383 case ISD::SETLT: X86CC = X86::COND_L; break;
1384 case ISD::SETLE: X86CC = X86::COND_LE; break;
1385 case ISD::SETNE: X86CC = X86::COND_NE; break;
1386 case ISD::SETULT: X86CC = X86::COND_B; break;
1387 case ISD::SETUGT: X86CC = X86::COND_A; break;
1388 case ISD::SETULE: X86CC = X86::COND_BE; break;
1389 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1392 // On a floating point condition, the flags are set as follows:
1394 // 0 | 0 | 0 | X > Y
1395 // 0 | 0 | 1 | X < Y
1396 // 1 | 0 | 0 | X == Y
1397 // 1 | 1 | 1 | unordered
1399 switch (SetCCOpcode) {
1402 case ISD::SETEQ: X86CC = X86::COND_E; break;
1403 case ISD::SETOLT: Flip = true; // Fallthrough
1405 case ISD::SETGT: X86CC = X86::COND_A; break;
1406 case ISD::SETOLE: Flip = true; // Fallthrough
1408 case ISD::SETGE: X86CC = X86::COND_AE; break;
1409 case ISD::SETUGT: Flip = true; // Fallthrough
1411 case ISD::SETLT: X86CC = X86::COND_B; break;
1412 case ISD::SETUGE: Flip = true; // Fallthrough
1414 case ISD::SETLE: X86CC = X86::COND_BE; break;
1416 case ISD::SETNE: X86CC = X86::COND_NE; break;
1417 case ISD::SETUO: X86CC = X86::COND_P; break;
1418 case ISD::SETO: X86CC = X86::COND_NP; break;
1421 std::swap(LHS, RHS);
1424 return X86CC != X86::COND_INVALID;
1427 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
1428 /// code. Current x86 isa includes the following FP cmov instructions:
1429 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
1430 static bool hasFPCMov(unsigned X86CC) {
1446 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
1447 /// true if Op is undef or if its value falls within the specified range (L, H].
1448 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
1449 if (Op.getOpcode() == ISD::UNDEF)
1452 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
1453 return (Val >= Low && Val < Hi);
1456 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
1457 /// true if Op is undef or if its value equal to the specified value.
1458 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
1459 if (Op.getOpcode() == ISD::UNDEF)
1461 return cast<ConstantSDNode>(Op)->getValue() == Val;
1464 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
1465 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
1466 bool X86::isPSHUFDMask(SDNode *N) {
1467 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1469 if (N->getNumOperands() != 4)
1472 // Check if the value doesn't reference the second vector.
1473 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1474 SDOperand Arg = N->getOperand(i);
1475 if (Arg.getOpcode() == ISD::UNDEF) continue;
1476 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1477 if (cast<ConstantSDNode>(Arg)->getValue() >= 4)
1484 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
1485 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
1486 bool X86::isPSHUFHWMask(SDNode *N) {
1487 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1489 if (N->getNumOperands() != 8)
1492 // Lower quadword copied in order.
1493 for (unsigned i = 0; i != 4; ++i) {
1494 SDOperand Arg = N->getOperand(i);
1495 if (Arg.getOpcode() == ISD::UNDEF) continue;
1496 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1497 if (cast<ConstantSDNode>(Arg)->getValue() != i)
1501 // Upper quadword shuffled.
1502 for (unsigned i = 4; i != 8; ++i) {
1503 SDOperand Arg = N->getOperand(i);
1504 if (Arg.getOpcode() == ISD::UNDEF) continue;
1505 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1506 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1507 if (Val < 4 || Val > 7)
1514 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
1515 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
1516 bool X86::isPSHUFLWMask(SDNode *N) {
1517 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1519 if (N->getNumOperands() != 8)
1522 // Upper quadword copied in order.
1523 for (unsigned i = 4; i != 8; ++i)
1524 if (!isUndefOrEqual(N->getOperand(i), i))
1527 // Lower quadword shuffled.
1528 for (unsigned i = 0; i != 4; ++i)
1529 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
1535 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
1536 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
1537 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
1538 if (NumElems != 2 && NumElems != 4) return false;
1540 unsigned Half = NumElems / 2;
1541 for (unsigned i = 0; i < Half; ++i)
1542 if (!isUndefOrInRange(Elems[i], 0, NumElems))
1544 for (unsigned i = Half; i < NumElems; ++i)
1545 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
1551 bool X86::isSHUFPMask(SDNode *N) {
1552 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1553 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
1556 /// isCommutedSHUFP - Returns true if the shuffle mask is except
1557 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
1558 /// half elements to come from vector 1 (which would equal the dest.) and
1559 /// the upper half to come from vector 2.
1560 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
1561 if (NumOps != 2 && NumOps != 4) return false;
1563 unsigned Half = NumOps / 2;
1564 for (unsigned i = 0; i < Half; ++i)
1565 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
1567 for (unsigned i = Half; i < NumOps; ++i)
1568 if (!isUndefOrInRange(Ops[i], 0, NumOps))
1573 static bool isCommutedSHUFP(SDNode *N) {
1574 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1575 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
1578 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
1579 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
1580 bool X86::isMOVHLPSMask(SDNode *N) {
1581 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1583 if (N->getNumOperands() != 4)
1586 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
1587 return isUndefOrEqual(N->getOperand(0), 6) &&
1588 isUndefOrEqual(N->getOperand(1), 7) &&
1589 isUndefOrEqual(N->getOperand(2), 2) &&
1590 isUndefOrEqual(N->getOperand(3), 3);
1593 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
1594 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
1596 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
1597 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1599 if (N->getNumOperands() != 4)
1602 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
1603 return isUndefOrEqual(N->getOperand(0), 2) &&
1604 isUndefOrEqual(N->getOperand(1), 3) &&
1605 isUndefOrEqual(N->getOperand(2), 2) &&
1606 isUndefOrEqual(N->getOperand(3), 3);
1609 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
1610 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
1611 bool X86::isMOVLPMask(SDNode *N) {
1612 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1614 unsigned NumElems = N->getNumOperands();
1615 if (NumElems != 2 && NumElems != 4)
1618 for (unsigned i = 0; i < NumElems/2; ++i)
1619 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
1622 for (unsigned i = NumElems/2; i < NumElems; ++i)
1623 if (!isUndefOrEqual(N->getOperand(i), i))
1629 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
1630 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
1632 bool X86::isMOVHPMask(SDNode *N) {
1633 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1635 unsigned NumElems = N->getNumOperands();
1636 if (NumElems != 2 && NumElems != 4)
1639 for (unsigned i = 0; i < NumElems/2; ++i)
1640 if (!isUndefOrEqual(N->getOperand(i), i))
1643 for (unsigned i = 0; i < NumElems/2; ++i) {
1644 SDOperand Arg = N->getOperand(i + NumElems/2);
1645 if (!isUndefOrEqual(Arg, i + NumElems))
1652 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
1653 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
1654 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
1655 bool V2IsSplat = false) {
1656 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1659 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
1660 SDOperand BitI = Elts[i];
1661 SDOperand BitI1 = Elts[i+1];
1662 if (!isUndefOrEqual(BitI, j))
1665 if (isUndefOrEqual(BitI1, NumElts))
1668 if (!isUndefOrEqual(BitI1, j + NumElts))
1676 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
1677 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1678 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
1681 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
1682 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
1683 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
1684 bool V2IsSplat = false) {
1685 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1688 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
1689 SDOperand BitI = Elts[i];
1690 SDOperand BitI1 = Elts[i+1];
1691 if (!isUndefOrEqual(BitI, j + NumElts/2))
1694 if (isUndefOrEqual(BitI1, NumElts))
1697 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
1705 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
1706 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1707 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
1710 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
1711 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
1713 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
1714 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1716 unsigned NumElems = N->getNumOperands();
1717 if (NumElems != 4 && NumElems != 8 && NumElems != 16)
1720 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
1721 SDOperand BitI = N->getOperand(i);
1722 SDOperand BitI1 = N->getOperand(i+1);
1724 if (!isUndefOrEqual(BitI, j))
1726 if (!isUndefOrEqual(BitI1, j))
1733 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
1734 /// specifies a shuffle of elements that is suitable for input to MOVSS,
1735 /// MOVSD, and MOVD, i.e. setting the lowest element.
1736 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
1737 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1740 if (!isUndefOrEqual(Elts[0], NumElts))
1743 for (unsigned i = 1; i < NumElts; ++i) {
1744 if (!isUndefOrEqual(Elts[i], i))
1751 bool X86::isMOVLMask(SDNode *N) {
1752 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1753 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
1756 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
1757 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
1758 /// element of vector 2 and the other elements to come from vector 1 in order.
1759 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
1760 bool V2IsSplat = false,
1761 bool V2IsUndef = false) {
1762 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
1765 if (!isUndefOrEqual(Ops[0], 0))
1768 for (unsigned i = 1; i < NumOps; ++i) {
1769 SDOperand Arg = Ops[i];
1770 if (!(isUndefOrEqual(Arg, i+NumOps) ||
1771 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
1772 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
1779 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
1780 bool V2IsUndef = false) {
1781 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1782 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
1783 V2IsSplat, V2IsUndef);
1786 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
1787 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
1788 bool X86::isMOVSHDUPMask(SDNode *N) {
1789 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1791 if (N->getNumOperands() != 4)
1794 // Expect 1, 1, 3, 3
1795 for (unsigned i = 0; i < 2; ++i) {
1796 SDOperand Arg = N->getOperand(i);
1797 if (Arg.getOpcode() == ISD::UNDEF) continue;
1798 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1799 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1800 if (Val != 1) return false;
1804 for (unsigned i = 2; i < 4; ++i) {
1805 SDOperand Arg = N->getOperand(i);
1806 if (Arg.getOpcode() == ISD::UNDEF) continue;
1807 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1808 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1809 if (Val != 3) return false;
1813 // Don't use movshdup if it can be done with a shufps.
1817 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
1818 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
1819 bool X86::isMOVSLDUPMask(SDNode *N) {
1820 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1822 if (N->getNumOperands() != 4)
1825 // Expect 0, 0, 2, 2
1826 for (unsigned i = 0; i < 2; ++i) {
1827 SDOperand Arg = N->getOperand(i);
1828 if (Arg.getOpcode() == ISD::UNDEF) continue;
1829 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1830 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1831 if (Val != 0) return false;
1835 for (unsigned i = 2; i < 4; ++i) {
1836 SDOperand Arg = N->getOperand(i);
1837 if (Arg.getOpcode() == ISD::UNDEF) continue;
1838 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1839 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1840 if (Val != 2) return false;
1844 // Don't use movshdup if it can be done with a shufps.
1848 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
1849 /// a splat of a single element.
1850 static bool isSplatMask(SDNode *N) {
1851 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1853 // This is a splat operation if each element of the permute is the same, and
1854 // if the value doesn't reference the second vector.
1855 unsigned NumElems = N->getNumOperands();
1856 SDOperand ElementBase;
1858 for (; i != NumElems; ++i) {
1859 SDOperand Elt = N->getOperand(i);
1860 if (isa<ConstantSDNode>(Elt)) {
1866 if (!ElementBase.Val)
1869 for (; i != NumElems; ++i) {
1870 SDOperand Arg = N->getOperand(i);
1871 if (Arg.getOpcode() == ISD::UNDEF) continue;
1872 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1873 if (Arg != ElementBase) return false;
1876 // Make sure it is a splat of the first vector operand.
1877 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
1880 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
1881 /// a splat of a single element and it's a 2 or 4 element mask.
1882 bool X86::isSplatMask(SDNode *N) {
1883 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1885 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
1886 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
1888 return ::isSplatMask(N);
1891 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
1892 /// specifies a splat of zero element.
1893 bool X86::isSplatLoMask(SDNode *N) {
1894 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1896 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
1897 if (!isUndefOrEqual(N->getOperand(i), 0))
1902 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
1903 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
1905 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
1906 unsigned NumOperands = N->getNumOperands();
1907 unsigned Shift = (NumOperands == 4) ? 2 : 1;
1909 for (unsigned i = 0; i < NumOperands; ++i) {
1911 SDOperand Arg = N->getOperand(NumOperands-i-1);
1912 if (Arg.getOpcode() != ISD::UNDEF)
1913 Val = cast<ConstantSDNode>(Arg)->getValue();
1914 if (Val >= NumOperands) Val -= NumOperands;
1916 if (i != NumOperands - 1)
1923 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
1924 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
1926 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
1928 // 8 nodes, but we only care about the last 4.
1929 for (unsigned i = 7; i >= 4; --i) {
1931 SDOperand Arg = N->getOperand(i);
1932 if (Arg.getOpcode() != ISD::UNDEF)
1933 Val = cast<ConstantSDNode>(Arg)->getValue();
1942 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
1943 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
1945 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
1947 // 8 nodes, but we only care about the first 4.
1948 for (int i = 3; i >= 0; --i) {
1950 SDOperand Arg = N->getOperand(i);
1951 if (Arg.getOpcode() != ISD::UNDEF)
1952 Val = cast<ConstantSDNode>(Arg)->getValue();
1961 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
1962 /// specifies a 8 element shuffle that can be broken into a pair of
1963 /// PSHUFHW and PSHUFLW.
1964 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
1965 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1967 if (N->getNumOperands() != 8)
1970 // Lower quadword shuffled.
1971 for (unsigned i = 0; i != 4; ++i) {
1972 SDOperand Arg = N->getOperand(i);
1973 if (Arg.getOpcode() == ISD::UNDEF) continue;
1974 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1975 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1980 // Upper quadword shuffled.
1981 for (unsigned i = 4; i != 8; ++i) {
1982 SDOperand Arg = N->getOperand(i);
1983 if (Arg.getOpcode() == ISD::UNDEF) continue;
1984 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1985 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1986 if (Val < 4 || Val > 7)
1993 /// CommuteVectorShuffle - Swap vector_shuffle operandsas well as
1994 /// values in ther permute mask.
1995 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
1996 SDOperand &V2, SDOperand &Mask,
1997 SelectionDAG &DAG) {
1998 MVT::ValueType VT = Op.getValueType();
1999 MVT::ValueType MaskVT = Mask.getValueType();
2000 MVT::ValueType EltVT = MVT::getVectorBaseType(MaskVT);
2001 unsigned NumElems = Mask.getNumOperands();
2002 SmallVector<SDOperand, 8> MaskVec;
2004 for (unsigned i = 0; i != NumElems; ++i) {
2005 SDOperand Arg = Mask.getOperand(i);
2006 if (Arg.getOpcode() == ISD::UNDEF) {
2007 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2010 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2011 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2013 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2015 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2019 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2020 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2023 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2024 /// match movhlps. The lower half elements should come from upper half of
2025 /// V1 (and in order), and the upper half elements should come from the upper
2026 /// half of V2 (and in order).
2027 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2028 unsigned NumElems = Mask->getNumOperands();
2031 for (unsigned i = 0, e = 2; i != e; ++i)
2032 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2034 for (unsigned i = 2; i != 4; ++i)
2035 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2040 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2041 /// is promoted to a vector.
2042 static inline bool isScalarLoadToVector(SDNode *N) {
2043 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2044 N = N->getOperand(0).Val;
2045 return ISD::isNON_EXTLoad(N);
2050 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2051 /// match movlp{s|d}. The lower half elements should come from lower half of
2052 /// V1 (and in order), and the upper half elements should come from the upper
2053 /// half of V2 (and in order). And since V1 will become the source of the
2054 /// MOVLP, it must be either a vector load or a scalar load to vector.
2055 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2056 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2058 // Is V2 is a vector load, don't do this transformation. We will try to use
2059 // load folding shufps op.
2060 if (ISD::isNON_EXTLoad(V2))
2063 unsigned NumElems = Mask->getNumOperands();
2064 if (NumElems != 2 && NumElems != 4)
2066 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2067 if (!isUndefOrEqual(Mask->getOperand(i), i))
2069 for (unsigned i = NumElems/2; i != NumElems; ++i)
2070 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2075 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2077 static bool isSplatVector(SDNode *N) {
2078 if (N->getOpcode() != ISD::BUILD_VECTOR)
2081 SDOperand SplatValue = N->getOperand(0);
2082 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2083 if (N->getOperand(i) != SplatValue)
2088 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2090 static bool isUndefShuffle(SDNode *N) {
2091 if (N->getOpcode() != ISD::BUILD_VECTOR)
2094 SDOperand V1 = N->getOperand(0);
2095 SDOperand V2 = N->getOperand(1);
2096 SDOperand Mask = N->getOperand(2);
2097 unsigned NumElems = Mask.getNumOperands();
2098 for (unsigned i = 0; i != NumElems; ++i) {
2099 SDOperand Arg = Mask.getOperand(i);
2100 if (Arg.getOpcode() != ISD::UNDEF) {
2101 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2102 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2104 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2111 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2112 /// that point to V2 points to its first element.
2113 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2114 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2116 bool Changed = false;
2117 SmallVector<SDOperand, 8> MaskVec;
2118 unsigned NumElems = Mask.getNumOperands();
2119 for (unsigned i = 0; i != NumElems; ++i) {
2120 SDOperand Arg = Mask.getOperand(i);
2121 if (Arg.getOpcode() != ISD::UNDEF) {
2122 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2123 if (Val > NumElems) {
2124 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2128 MaskVec.push_back(Arg);
2132 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2133 &MaskVec[0], MaskVec.size());
2137 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2138 /// operation of specified width.
2139 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2140 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2141 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2143 SmallVector<SDOperand, 8> MaskVec;
2144 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2145 for (unsigned i = 1; i != NumElems; ++i)
2146 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2147 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2150 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2151 /// of specified width.
2152 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2153 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2154 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2155 SmallVector<SDOperand, 8> MaskVec;
2156 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2157 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2158 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2160 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2163 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2164 /// of specified width.
2165 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2166 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2167 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2168 unsigned Half = NumElems/2;
2169 SmallVector<SDOperand, 8> MaskVec;
2170 for (unsigned i = 0; i != Half; ++i) {
2171 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2172 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2174 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2177 /// getZeroVector - Returns a vector of specified type with all zero elements.
2179 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2180 assert(MVT::isVector(VT) && "Expected a vector type");
2181 unsigned NumElems = getVectorNumElements(VT);
2182 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2183 bool isFP = MVT::isFloatingPoint(EVT);
2184 SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT);
2185 SmallVector<SDOperand, 8> ZeroVec(NumElems, Zero);
2186 return DAG.getNode(ISD::BUILD_VECTOR, VT, &ZeroVec[0], ZeroVec.size());
2189 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2191 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2192 SDOperand V1 = Op.getOperand(0);
2193 SDOperand Mask = Op.getOperand(2);
2194 MVT::ValueType VT = Op.getValueType();
2195 unsigned NumElems = Mask.getNumOperands();
2196 Mask = getUnpacklMask(NumElems, DAG);
2197 while (NumElems != 4) {
2198 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2201 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2203 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2204 Mask = getZeroVector(MaskVT, DAG);
2205 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2206 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2207 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2210 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2212 static inline bool isZeroNode(SDOperand Elt) {
2213 return ((isa<ConstantSDNode>(Elt) &&
2214 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2215 (isa<ConstantFPSDNode>(Elt) &&
2216 cast<ConstantFPSDNode>(Elt)->isExactlyValue(0.0)));
2219 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2220 /// vector and zero or undef vector.
2221 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2222 unsigned NumElems, unsigned Idx,
2223 bool isZero, SelectionDAG &DAG) {
2224 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2225 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2226 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2227 SDOperand Zero = DAG.getConstant(0, EVT);
2228 SmallVector<SDOperand, 8> MaskVec(NumElems, Zero);
2229 MaskVec[Idx] = DAG.getConstant(NumElems, EVT);
2230 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2231 &MaskVec[0], MaskVec.size());
2232 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2235 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2237 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
2238 unsigned NumNonZero, unsigned NumZero,
2239 SelectionDAG &DAG, TargetLowering &TLI) {
2245 for (unsigned i = 0; i < 16; ++i) {
2246 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
2247 if (ThisIsNonZero && First) {
2249 V = getZeroVector(MVT::v8i16, DAG);
2251 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2256 SDOperand ThisElt(0, 0), LastElt(0, 0);
2257 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
2258 if (LastIsNonZero) {
2259 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
2261 if (ThisIsNonZero) {
2262 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
2263 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
2264 ThisElt, DAG.getConstant(8, MVT::i8));
2266 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
2271 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
2272 DAG.getConstant(i/2, TLI.getPointerTy()));
2276 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
2279 /// LowerBuildVectorv16i8 - Custom lower build_vector of v8i16.
2281 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
2282 unsigned NumNonZero, unsigned NumZero,
2283 SelectionDAG &DAG, TargetLowering &TLI) {
2289 for (unsigned i = 0; i < 8; ++i) {
2290 bool isNonZero = (NonZeros & (1 << i)) != 0;
2294 V = getZeroVector(MVT::v8i16, DAG);
2296 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2299 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
2300 DAG.getConstant(i, TLI.getPointerTy()));
2308 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2309 // All zero's are handled with pxor.
2310 if (ISD::isBuildVectorAllZeros(Op.Val))
2313 // All one's are handled with pcmpeqd.
2314 if (ISD::isBuildVectorAllOnes(Op.Val))
2317 MVT::ValueType VT = Op.getValueType();
2318 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2319 unsigned EVTBits = MVT::getSizeInBits(EVT);
2321 unsigned NumElems = Op.getNumOperands();
2322 unsigned NumZero = 0;
2323 unsigned NumNonZero = 0;
2324 unsigned NonZeros = 0;
2325 std::set<SDOperand> Values;
2326 for (unsigned i = 0; i < NumElems; ++i) {
2327 SDOperand Elt = Op.getOperand(i);
2328 if (Elt.getOpcode() != ISD::UNDEF) {
2330 if (isZeroNode(Elt))
2333 NonZeros |= (1 << i);
2339 if (NumNonZero == 0)
2340 // Must be a mix of zero and undef. Return a zero vector.
2341 return getZeroVector(VT, DAG);
2343 // Splat is obviously ok. Let legalizer expand it to a shuffle.
2344 if (Values.size() == 1)
2347 // Special case for single non-zero element.
2348 if (NumNonZero == 1) {
2349 unsigned Idx = CountTrailingZeros_32(NonZeros);
2350 SDOperand Item = Op.getOperand(Idx);
2351 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
2353 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
2354 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
2357 if (EVTBits == 32) {
2358 // Turn it into a shuffle of zero and zero-extended scalar to vector.
2359 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
2361 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2362 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
2363 SmallVector<SDOperand, 8> MaskVec;
2364 for (unsigned i = 0; i < NumElems; i++)
2365 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
2366 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2367 &MaskVec[0], MaskVec.size());
2368 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
2369 DAG.getNode(ISD::UNDEF, VT), Mask);
2373 // Let legalizer expand 2-wide build_vector's.
2377 // If element VT is < 32 bits, convert it to inserts into a zero vector.
2379 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
2381 if (V.Val) return V;
2384 if (EVTBits == 16) {
2385 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
2387 if (V.Val) return V;
2390 // If element VT is == 32 bits, turn it into a number of shuffles.
2391 SmallVector<SDOperand, 8> V;
2393 if (NumElems == 4 && NumZero > 0) {
2394 for (unsigned i = 0; i < 4; ++i) {
2395 bool isZero = !(NonZeros & (1 << i));
2397 V[i] = getZeroVector(VT, DAG);
2399 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2402 for (unsigned i = 0; i < 2; ++i) {
2403 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
2406 V[i] = V[i*2]; // Must be a zero vector.
2409 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
2410 getMOVLMask(NumElems, DAG));
2413 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2414 getMOVLMask(NumElems, DAG));
2417 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2418 getUnpacklMask(NumElems, DAG));
2423 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
2424 // clears the upper bits.
2425 // FIXME: we can do the same for v4f32 case when we know both parts of
2426 // the lower half come from scalar_to_vector (loadf32). We should do
2427 // that in post legalizer dag combiner with target specific hooks.
2428 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
2430 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2431 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2432 SmallVector<SDOperand, 8> MaskVec;
2433 bool Reverse = (NonZeros & 0x3) == 2;
2434 for (unsigned i = 0; i < 2; ++i)
2436 MaskVec.push_back(DAG.getConstant(1-i, EVT));
2438 MaskVec.push_back(DAG.getConstant(i, EVT));
2439 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
2440 for (unsigned i = 0; i < 2; ++i)
2442 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
2444 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
2445 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2446 &MaskVec[0], MaskVec.size());
2447 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
2450 if (Values.size() > 2) {
2451 // Expand into a number of unpckl*.
2453 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
2454 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
2455 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
2456 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
2457 for (unsigned i = 0; i < NumElems; ++i)
2458 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2460 while (NumElems != 0) {
2461 for (unsigned i = 0; i < NumElems; ++i)
2462 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
2473 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
2474 SDOperand V1 = Op.getOperand(0);
2475 SDOperand V2 = Op.getOperand(1);
2476 SDOperand PermMask = Op.getOperand(2);
2477 MVT::ValueType VT = Op.getValueType();
2478 unsigned NumElems = PermMask.getNumOperands();
2479 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
2480 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
2481 bool V1IsSplat = false;
2482 bool V2IsSplat = false;
2484 if (isUndefShuffle(Op.Val))
2485 return DAG.getNode(ISD::UNDEF, VT);
2487 if (isSplatMask(PermMask.Val)) {
2488 if (NumElems <= 4) return Op;
2489 // Promote it to a v4i32 splat.
2490 return PromoteSplat(Op, DAG);
2493 if (X86::isMOVLMask(PermMask.Val))
2494 return (V1IsUndef) ? V2 : Op;
2496 if (X86::isMOVSHDUPMask(PermMask.Val) ||
2497 X86::isMOVSLDUPMask(PermMask.Val) ||
2498 X86::isMOVHLPSMask(PermMask.Val) ||
2499 X86::isMOVHPMask(PermMask.Val) ||
2500 X86::isMOVLPMask(PermMask.Val))
2503 if (ShouldXformToMOVHLPS(PermMask.Val) ||
2504 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
2505 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2507 bool Commuted = false;
2508 V1IsSplat = isSplatVector(V1.Val);
2509 V2IsSplat = isSplatVector(V2.Val);
2510 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
2511 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2512 std::swap(V1IsSplat, V2IsSplat);
2513 std::swap(V1IsUndef, V2IsUndef);
2517 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
2518 if (V2IsUndef) return V1;
2519 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2521 // V2 is a splat, so the mask may be malformed. That is, it may point
2522 // to any V2 element. The instruction selectior won't like this. Get
2523 // a corrected mask and commute to form a proper MOVS{S|D}.
2524 SDOperand NewMask = getMOVLMask(NumElems, DAG);
2525 if (NewMask.Val != PermMask.Val)
2526 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2531 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
2532 X86::isUNPCKLMask(PermMask.Val) ||
2533 X86::isUNPCKHMask(PermMask.Val))
2537 // Normalize mask so all entries that point to V2 points to its first
2538 // element then try to match unpck{h|l} again. If match, return a
2539 // new vector_shuffle with the corrected mask.
2540 SDOperand NewMask = NormalizeMask(PermMask, DAG);
2541 if (NewMask.Val != PermMask.Val) {
2542 if (X86::isUNPCKLMask(PermMask.Val, true)) {
2543 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
2544 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2545 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
2546 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
2547 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2552 // Normalize the node to match x86 shuffle ops if needed
2553 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
2554 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2557 // Commute is back and try unpck* again.
2558 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2559 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
2560 X86::isUNPCKLMask(PermMask.Val) ||
2561 X86::isUNPCKHMask(PermMask.Val))
2565 // If VT is integer, try PSHUF* first, then SHUFP*.
2566 if (MVT::isInteger(VT)) {
2567 if (X86::isPSHUFDMask(PermMask.Val) ||
2568 X86::isPSHUFHWMask(PermMask.Val) ||
2569 X86::isPSHUFLWMask(PermMask.Val)) {
2570 if (V2.getOpcode() != ISD::UNDEF)
2571 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
2572 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
2576 if (X86::isSHUFPMask(PermMask.Val))
2579 // Handle v8i16 shuffle high / low shuffle node pair.
2580 if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
2581 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2582 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2583 SmallVector<SDOperand, 8> MaskVec;
2584 for (unsigned i = 0; i != 4; ++i)
2585 MaskVec.push_back(PermMask.getOperand(i));
2586 for (unsigned i = 4; i != 8; ++i)
2587 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2588 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2589 &MaskVec[0], MaskVec.size());
2590 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2592 for (unsigned i = 0; i != 4; ++i)
2593 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2594 for (unsigned i = 4; i != 8; ++i)
2595 MaskVec.push_back(PermMask.getOperand(i));
2596 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0],MaskVec.size());
2597 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2600 // Floating point cases in the other order.
2601 if (X86::isSHUFPMask(PermMask.Val))
2603 if (X86::isPSHUFDMask(PermMask.Val) ||
2604 X86::isPSHUFHWMask(PermMask.Val) ||
2605 X86::isPSHUFLWMask(PermMask.Val)) {
2606 if (V2.getOpcode() != ISD::UNDEF)
2607 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
2608 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
2613 if (NumElems == 4) {
2614 MVT::ValueType MaskVT = PermMask.getValueType();
2615 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
2616 SmallVector<std::pair<int, int>, 8> Locs;
2617 Locs.reserve(NumElems);
2618 SmallVector<SDOperand, 8> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2619 SmallVector<SDOperand, 8> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2622 // If no more than two elements come from either vector. This can be
2623 // implemented with two shuffles. First shuffle gather the elements.
2624 // The second shuffle, which takes the first shuffle as both of its
2625 // vector operands, put the elements into the right order.
2626 for (unsigned i = 0; i != NumElems; ++i) {
2627 SDOperand Elt = PermMask.getOperand(i);
2628 if (Elt.getOpcode() == ISD::UNDEF) {
2629 Locs[i] = std::make_pair(-1, -1);
2631 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
2632 if (Val < NumElems) {
2633 Locs[i] = std::make_pair(0, NumLo);
2637 Locs[i] = std::make_pair(1, NumHi);
2638 if (2+NumHi < NumElems)
2639 Mask1[2+NumHi] = Elt;
2644 if (NumLo <= 2 && NumHi <= 2) {
2645 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2646 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2647 &Mask1[0], Mask1.size()));
2648 for (unsigned i = 0; i != NumElems; ++i) {
2649 if (Locs[i].first == -1)
2652 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
2653 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
2654 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
2658 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
2659 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2660 &Mask2[0], Mask2.size()));
2663 // Break it into (shuffle shuffle_hi, shuffle_lo).
2665 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2666 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2667 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
2668 unsigned MaskIdx = 0;
2670 unsigned HiIdx = NumElems/2;
2671 for (unsigned i = 0; i != NumElems; ++i) {
2672 if (i == NumElems/2) {
2678 SDOperand Elt = PermMask.getOperand(i);
2679 if (Elt.getOpcode() == ISD::UNDEF) {
2680 Locs[i] = std::make_pair(-1, -1);
2681 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
2682 Locs[i] = std::make_pair(MaskIdx, LoIdx);
2683 (*MaskPtr)[LoIdx] = Elt;
2686 Locs[i] = std::make_pair(MaskIdx, HiIdx);
2687 (*MaskPtr)[HiIdx] = Elt;
2692 SDOperand LoShuffle =
2693 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2694 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2695 &LoMask[0], LoMask.size()));
2696 SDOperand HiShuffle =
2697 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2698 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2699 &HiMask[0], HiMask.size()));
2700 SmallVector<SDOperand, 8> MaskOps;
2701 for (unsigned i = 0; i != NumElems; ++i) {
2702 if (Locs[i].first == -1) {
2703 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
2705 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
2706 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
2709 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
2710 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2711 &MaskOps[0], MaskOps.size()));
2718 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2719 if (!isa<ConstantSDNode>(Op.getOperand(1)))
2722 MVT::ValueType VT = Op.getValueType();
2723 // TODO: handle v16i8.
2724 if (MVT::getSizeInBits(VT) == 16) {
2725 // Transform it so it match pextrw which produces a 32-bit result.
2726 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
2727 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
2728 Op.getOperand(0), Op.getOperand(1));
2729 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
2730 DAG.getValueType(VT));
2731 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
2732 } else if (MVT::getSizeInBits(VT) == 32) {
2733 SDOperand Vec = Op.getOperand(0);
2734 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
2737 // SHUFPS the element to the lowest double word, then movss.
2738 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2739 SmallVector<SDOperand, 8> IdxVec;
2740 IdxVec.push_back(DAG.getConstant(Idx, MVT::getVectorBaseType(MaskVT)));
2741 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2742 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2743 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2744 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2745 &IdxVec[0], IdxVec.size());
2746 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
2747 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
2748 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
2749 DAG.getConstant(0, getPointerTy()));
2750 } else if (MVT::getSizeInBits(VT) == 64) {
2751 SDOperand Vec = Op.getOperand(0);
2752 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
2756 // UNPCKHPD the element to the lowest double word, then movsd.
2757 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
2758 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
2759 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2760 SmallVector<SDOperand, 8> IdxVec;
2761 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorBaseType(MaskVT)));
2762 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2763 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2764 &IdxVec[0], IdxVec.size());
2765 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
2766 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
2767 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
2768 DAG.getConstant(0, getPointerTy()));
2775 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2776 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
2777 // as its second argument.
2778 MVT::ValueType VT = Op.getValueType();
2779 MVT::ValueType BaseVT = MVT::getVectorBaseType(VT);
2780 SDOperand N0 = Op.getOperand(0);
2781 SDOperand N1 = Op.getOperand(1);
2782 SDOperand N2 = Op.getOperand(2);
2783 if (MVT::getSizeInBits(BaseVT) == 16) {
2784 if (N1.getValueType() != MVT::i32)
2785 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
2786 if (N2.getValueType() != MVT::i32)
2787 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(), MVT::i32);
2788 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
2789 } else if (MVT::getSizeInBits(BaseVT) == 32) {
2790 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
2793 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
2794 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2795 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2796 SmallVector<SDOperand, 8> MaskVec;
2797 MaskVec.push_back(DAG.getConstant(4, BaseVT));
2798 for (unsigned i = 1; i <= 3; ++i)
2799 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2800 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
2801 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2802 &MaskVec[0], MaskVec.size()));
2804 // Use two pinsrw instructions to insert a 32 bit value.
2806 if (MVT::isFloatingPoint(N1.getValueType())) {
2807 if (ISD::isNON_EXTLoad(N1.Val)) {
2808 // Just load directly from f32mem to GR32.
2809 LoadSDNode *LD = cast<LoadSDNode>(N1);
2810 N1 = DAG.getLoad(MVT::i32, LD->getChain(), LD->getBasePtr(),
2811 LD->getSrcValue(), LD->getSrcValueOffset());
2813 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
2814 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
2815 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
2816 DAG.getConstant(0, getPointerTy()));
2819 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
2820 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
2821 DAG.getConstant(Idx, getPointerTy()));
2822 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
2823 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
2824 DAG.getConstant(Idx+1, getPointerTy()));
2825 return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
2833 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2834 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
2835 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
2838 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2839 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
2840 // one of the above mentioned nodes. It has to be wrapped because otherwise
2841 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2842 // be used to form addressing mode. These wrapped nodes will be selected
2845 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
2846 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2847 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
2849 CP->getAlignment());
2850 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
2851 // With PIC, the address is actually $g + Offset.
2852 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2853 !Subtarget->isPICStyleRIPRel()) {
2854 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2855 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
2863 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
2864 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2865 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
2866 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
2867 // With PIC, the address is actually $g + Offset.
2868 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2869 !Subtarget->isPICStyleRIPRel()) {
2870 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2871 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
2875 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
2876 // load the value at address GV, not the value of GV itself. This means that
2877 // the GlobalAddress must be in the base or index register of the address, not
2878 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
2879 // The same applies for external symbols during PIC codegen
2880 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
2881 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
2887 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
2888 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
2889 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
2890 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
2891 // With PIC, the address is actually $g + Offset.
2892 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2893 !Subtarget->isPICStyleRIPRel()) {
2894 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2895 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
2902 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
2903 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
2904 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
2905 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
2906 // With PIC, the address is actually $g + Offset.
2907 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2908 !Subtarget->isPICStyleRIPRel()) {
2909 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2910 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
2917 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
2918 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
2919 "Not an i64 shift!");
2920 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
2921 SDOperand ShOpLo = Op.getOperand(0);
2922 SDOperand ShOpHi = Op.getOperand(1);
2923 SDOperand ShAmt = Op.getOperand(2);
2924 SDOperand Tmp1 = isSRA ?
2925 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
2926 DAG.getConstant(0, MVT::i32);
2928 SDOperand Tmp2, Tmp3;
2929 if (Op.getOpcode() == ISD::SHL_PARTS) {
2930 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
2931 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
2933 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
2934 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
2937 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
2938 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
2939 DAG.getConstant(32, MVT::i8));
2940 SDOperand COps[]={DAG.getEntryNode(), AndNode, DAG.getConstant(0, MVT::i8)};
2941 SDOperand InFlag = DAG.getNode(X86ISD::CMP, VTs, 2, COps, 3).getValue(1);
2944 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
2946 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
2947 SmallVector<SDOperand, 4> Ops;
2948 if (Op.getOpcode() == ISD::SHL_PARTS) {
2949 Ops.push_back(Tmp2);
2950 Ops.push_back(Tmp3);
2952 Ops.push_back(InFlag);
2953 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
2954 InFlag = Hi.getValue(1);
2957 Ops.push_back(Tmp3);
2958 Ops.push_back(Tmp1);
2960 Ops.push_back(InFlag);
2961 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
2963 Ops.push_back(Tmp2);
2964 Ops.push_back(Tmp3);
2966 Ops.push_back(InFlag);
2967 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
2968 InFlag = Lo.getValue(1);
2971 Ops.push_back(Tmp3);
2972 Ops.push_back(Tmp1);
2974 Ops.push_back(InFlag);
2975 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
2978 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
2982 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
2985 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
2986 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
2987 Op.getOperand(0).getValueType() >= MVT::i16 &&
2988 "Unknown SINT_TO_FP to lower!");
2991 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
2992 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
2993 MachineFunction &MF = DAG.getMachineFunction();
2994 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
2995 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
2996 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
2997 StackSlot, NULL, 0);
3002 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
3004 Tys = DAG.getVTList(MVT::f64, MVT::Other);
3005 SmallVector<SDOperand, 8> Ops;
3006 Ops.push_back(Chain);
3007 Ops.push_back(StackSlot);
3008 Ops.push_back(DAG.getValueType(SrcVT));
3009 Result = DAG.getNode(X86ScalarSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
3010 Tys, &Ops[0], Ops.size());
3013 Chain = Result.getValue(1);
3014 SDOperand InFlag = Result.getValue(2);
3016 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
3017 // shouldn't be necessary except that RFP cannot be live across
3018 // multiple blocks. When stackifier is fixed, they can be uncoupled.
3019 MachineFunction &MF = DAG.getMachineFunction();
3020 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3021 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3022 Tys = DAG.getVTList(MVT::Other);
3023 SmallVector<SDOperand, 8> Ops;
3024 Ops.push_back(Chain);
3025 Ops.push_back(Result);
3026 Ops.push_back(StackSlot);
3027 Ops.push_back(DAG.getValueType(Op.getValueType()));
3028 Ops.push_back(InFlag);
3029 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
3030 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
3036 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
3037 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
3038 "Unknown FP_TO_SINT to lower!");
3039 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
3041 MachineFunction &MF = DAG.getMachineFunction();
3042 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
3043 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3044 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3047 switch (Op.getValueType()) {
3048 default: assert(0 && "Invalid FP_TO_SINT to lower!");
3049 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
3050 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
3051 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
3054 SDOperand Chain = DAG.getEntryNode();
3055 SDOperand Value = Op.getOperand(0);
3057 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
3058 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
3059 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
3061 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
3063 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
3064 Chain = Value.getValue(1);
3065 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3066 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3069 // Build the FP_TO_INT*_IN_MEM
3070 SDOperand Ops[] = { Chain, Value, StackSlot };
3071 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
3074 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
3077 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
3078 MVT::ValueType VT = Op.getValueType();
3079 const Type *OpNTy = MVT::getTypeForValueType(VT);
3080 std::vector<Constant*> CV;
3081 if (VT == MVT::f64) {
3082 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(~(1ULL << 63))));
3083 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3085 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(~(1U << 31))));
3086 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3087 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3088 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3090 Constant *CS = ConstantStruct::get(CV);
3091 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3092 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3093 SmallVector<SDOperand, 3> Ops;
3094 Ops.push_back(DAG.getEntryNode());
3095 Ops.push_back(CPIdx);
3096 Ops.push_back(DAG.getSrcValue(NULL));
3097 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3098 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
3101 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
3102 MVT::ValueType VT = Op.getValueType();
3103 const Type *OpNTy = MVT::getTypeForValueType(VT);
3104 std::vector<Constant*> CV;
3105 if (VT == MVT::f64) {
3106 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(1ULL << 63)));
3107 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3109 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(1U << 31)));
3110 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3111 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3112 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3114 Constant *CS = ConstantStruct::get(CV);
3115 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3116 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3117 SmallVector<SDOperand, 3> Ops;
3118 Ops.push_back(DAG.getEntryNode());
3119 Ops.push_back(CPIdx);
3120 Ops.push_back(DAG.getSrcValue(NULL));
3121 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3122 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
3125 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
3126 SDOperand Op0 = Op.getOperand(0);
3127 SDOperand Op1 = Op.getOperand(1);
3128 MVT::ValueType VT = Op.getValueType();
3129 MVT::ValueType SrcVT = Op1.getValueType();
3130 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
3132 // If second operand is smaller, extend it first.
3133 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
3134 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
3138 // First get the sign bit of second operand.
3139 std::vector<Constant*> CV;
3140 if (SrcVT == MVT::f64) {
3141 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(1ULL << 63)));
3142 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3144 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(1U << 31)));
3145 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3146 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3147 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3149 Constant *CS = ConstantStruct::get(CV);
3150 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3151 SDVTList Tys = DAG.getVTList(SrcVT, MVT::Other);
3152 SmallVector<SDOperand, 3> Ops;
3153 Ops.push_back(DAG.getEntryNode());
3154 Ops.push_back(CPIdx);
3155 Ops.push_back(DAG.getSrcValue(NULL));
3156 SDOperand Mask1 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3157 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
3159 // Shift sign bit right or left if the two operands have different types.
3160 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
3161 // Op0 is MVT::f32, Op1 is MVT::f64.
3162 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
3163 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
3164 DAG.getConstant(32, MVT::i32));
3165 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
3166 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
3167 DAG.getConstant(0, getPointerTy()));
3170 // Clear first operand sign bit.
3172 if (VT == MVT::f64) {
3173 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(~(1ULL << 63))));
3174 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3176 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(~(1U << 31))));
3177 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3178 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3179 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3181 CS = ConstantStruct::get(CV);
3182 CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3183 Tys = DAG.getVTList(VT, MVT::Other);
3185 Ops.push_back(DAG.getEntryNode());
3186 Ops.push_back(CPIdx);
3187 Ops.push_back(DAG.getSrcValue(NULL));
3188 SDOperand Mask2 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3189 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
3191 // Or the value with the sign bit.
3192 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
3195 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG,
3197 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
3199 SDOperand Op0 = Op.getOperand(0);
3200 SDOperand Op1 = Op.getOperand(1);
3201 SDOperand CC = Op.getOperand(2);
3202 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3203 const MVT::ValueType *VTs1 = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3204 const MVT::ValueType *VTs2 = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
3205 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
3208 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
3210 SDOperand Ops1[] = { Chain, Op0, Op1 };
3211 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, Ops1, 3).getValue(1);
3212 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
3213 return DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3216 assert(isFP && "Illegal integer SetCC!");
3218 SDOperand COps[] = { Chain, Op0, Op1 };
3219 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, COps, 3).getValue(1);
3221 switch (SetCCOpcode) {
3222 default: assert(false && "Illegal floating point SetCC!");
3223 case ISD::SETOEQ: { // !PF & ZF
3224 SDOperand Ops1[] = { DAG.getConstant(X86::COND_NP, MVT::i8), Cond };
3225 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3226 SDOperand Ops2[] = { DAG.getConstant(X86::COND_E, MVT::i8),
3228 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3229 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
3231 case ISD::SETUNE: { // PF | !ZF
3232 SDOperand Ops1[] = { DAG.getConstant(X86::COND_P, MVT::i8), Cond };
3233 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3234 SDOperand Ops2[] = { DAG.getConstant(X86::COND_NE, MVT::i8),
3236 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3237 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
3242 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
3243 bool addTest = true;
3244 SDOperand Chain = DAG.getEntryNode();
3245 SDOperand Cond = Op.getOperand(0);
3247 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3249 if (Cond.getOpcode() == ISD::SETCC)
3250 Cond = LowerSETCC(Cond, DAG, Chain);
3252 if (Cond.getOpcode() == X86ISD::SETCC) {
3253 CC = Cond.getOperand(0);
3255 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3256 // (since flag operand cannot be shared). Use it as the condition setting
3257 // operand in place of the X86ISD::SETCC.
3258 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3259 // to use a test instead of duplicating the X86ISD::CMP (for register
3260 // pressure reason)?
3261 SDOperand Cmp = Cond.getOperand(1);
3262 unsigned Opc = Cmp.getOpcode();
3263 bool IllegalFPCMov = !X86ScalarSSE &&
3264 MVT::isFloatingPoint(Op.getValueType()) &&
3265 !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
3266 if ((Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) &&
3268 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3269 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3275 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3276 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3277 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3280 VTs = DAG.getNodeValueTypes(Op.getValueType(), MVT::Flag);
3281 SmallVector<SDOperand, 4> Ops;
3282 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
3283 // condition is true.
3284 Ops.push_back(Op.getOperand(2));
3285 Ops.push_back(Op.getOperand(1));
3287 Ops.push_back(Cond.getValue(1));
3288 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3291 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
3292 bool addTest = true;
3293 SDOperand Chain = Op.getOperand(0);
3294 SDOperand Cond = Op.getOperand(1);
3295 SDOperand Dest = Op.getOperand(2);
3297 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3299 if (Cond.getOpcode() == ISD::SETCC)
3300 Cond = LowerSETCC(Cond, DAG, Chain);
3302 if (Cond.getOpcode() == X86ISD::SETCC) {
3303 CC = Cond.getOperand(0);
3305 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3306 // (since flag operand cannot be shared). Use it as the condition setting
3307 // operand in place of the X86ISD::SETCC.
3308 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3309 // to use a test instead of duplicating the X86ISD::CMP (for register
3310 // pressure reason)?
3311 SDOperand Cmp = Cond.getOperand(1);
3312 unsigned Opc = Cmp.getOpcode();
3313 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) {
3314 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3315 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3321 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3322 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3323 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3325 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
3326 Cond, Op.getOperand(2), CC, Cond.getValue(1));
3329 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
3330 unsigned CallingConv= cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3332 if (Subtarget->is64Bit())
3333 return LowerX86_64CCCCallTo(Op, DAG, CallingConv);
3335 switch (CallingConv) {
3337 assert(0 && "Unsupported calling convention");
3338 case CallingConv::Fast:
3339 // TODO: Implement fastcc
3341 case CallingConv::C:
3342 case CallingConv::X86_StdCall:
3343 return LowerCCCCallTo(Op, DAG, CallingConv);
3344 case CallingConv::X86_FastCall:
3345 return LowerFastCCCallTo(Op, DAG, CallingConv);
3350 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
3351 MachineFunction &MF = DAG.getMachineFunction();
3352 const Function* Fn = MF.getFunction();
3353 if (Fn->hasExternalLinkage() &&
3354 Subtarget->isTargetCygMing() &&
3355 Fn->getName() == "main")
3356 MF.getInfo<X86FunctionInfo>()->setForceFramePointer(true);
3358 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3359 if (Subtarget->is64Bit())
3360 return LowerX86_64CCCArguments(Op, DAG);
3364 assert(0 && "Unsupported calling convention");
3365 case CallingConv::Fast:
3366 // TODO: implement fastcc.
3369 case CallingConv::C:
3370 return LowerCCCArguments(Op, DAG);
3371 case CallingConv::X86_StdCall:
3372 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(StdCall);
3373 return LowerCCCArguments(Op, DAG, true);
3374 case CallingConv::X86_FastCall:
3375 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(FastCall);
3376 return LowerFastCCArguments(Op, DAG);
3380 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
3381 SDOperand InFlag(0, 0);
3382 SDOperand Chain = Op.getOperand(0);
3384 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3385 if (Align == 0) Align = 1;
3387 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3388 // If not DWORD aligned, call memset if size is less than the threshold.
3389 // It knows how to align to the right boundary first.
3390 if ((Align & 3) != 0 ||
3391 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
3392 MVT::ValueType IntPtr = getPointerTy();
3393 const Type *IntPtrTy = getTargetData()->getIntPtrType();
3394 TargetLowering::ArgListTy Args;
3395 TargetLowering::ArgListEntry Entry;
3396 Entry.Node = Op.getOperand(1);
3397 Entry.Ty = IntPtrTy;
3398 Entry.isSigned = false;
3399 Entry.isInReg = false;
3400 Entry.isSRet = false;
3401 Args.push_back(Entry);
3402 // Extend the unsigned i8 argument to be an int value for the call.
3403 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
3404 Entry.Ty = IntPtrTy;
3405 Entry.isSigned = false;
3406 Entry.isInReg = false;
3407 Entry.isSRet = false;
3408 Args.push_back(Entry);
3409 Entry.Node = Op.getOperand(3);
3410 Args.push_back(Entry);
3411 std::pair<SDOperand,SDOperand> CallResult =
3412 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
3413 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
3414 return CallResult.second;
3419 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
3420 unsigned BytesLeft = 0;
3421 bool TwoRepStos = false;
3424 uint64_t Val = ValC->getValue() & 255;
3426 // If the value is a constant, then we can potentially use larger sets.
3427 switch (Align & 3) {
3428 case 2: // WORD aligned
3431 Val = (Val << 8) | Val;
3433 case 0: // DWORD aligned
3436 Val = (Val << 8) | Val;
3437 Val = (Val << 16) | Val;
3438 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
3441 Val = (Val << 32) | Val;
3444 default: // Byte aligned
3447 Count = Op.getOperand(3);
3451 if (AVT > MVT::i8) {
3453 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
3454 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
3455 BytesLeft = I->getValue() % UBytes;
3457 assert(AVT >= MVT::i32 &&
3458 "Do not use rep;stos if not at least DWORD aligned");
3459 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
3460 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
3465 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
3467 InFlag = Chain.getValue(1);
3470 Count = Op.getOperand(3);
3471 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
3472 InFlag = Chain.getValue(1);
3475 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
3477 InFlag = Chain.getValue(1);
3478 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
3479 Op.getOperand(1), InFlag);
3480 InFlag = Chain.getValue(1);
3482 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3483 SmallVector<SDOperand, 8> Ops;
3484 Ops.push_back(Chain);
3485 Ops.push_back(DAG.getValueType(AVT));
3486 Ops.push_back(InFlag);
3487 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
3490 InFlag = Chain.getValue(1);
3491 Count = Op.getOperand(3);
3492 MVT::ValueType CVT = Count.getValueType();
3493 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
3494 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
3495 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
3497 InFlag = Chain.getValue(1);
3498 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3500 Ops.push_back(Chain);
3501 Ops.push_back(DAG.getValueType(MVT::i8));
3502 Ops.push_back(InFlag);
3503 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
3504 } else if (BytesLeft) {
3505 // Issue stores for the last 1 - 7 bytes.
3507 unsigned Val = ValC->getValue() & 255;
3508 unsigned Offset = I->getValue() - BytesLeft;
3509 SDOperand DstAddr = Op.getOperand(1);
3510 MVT::ValueType AddrVT = DstAddr.getValueType();
3511 if (BytesLeft >= 4) {
3512 Val = (Val << 8) | Val;
3513 Val = (Val << 16) | Val;
3514 Value = DAG.getConstant(Val, MVT::i32);
3515 Chain = DAG.getStore(Chain, Value,
3516 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3517 DAG.getConstant(Offset, AddrVT)),
3522 if (BytesLeft >= 2) {
3523 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
3524 Chain = DAG.getStore(Chain, Value,
3525 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3526 DAG.getConstant(Offset, AddrVT)),
3531 if (BytesLeft == 1) {
3532 Value = DAG.getConstant(Val, MVT::i8);
3533 Chain = DAG.getStore(Chain, Value,
3534 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3535 DAG.getConstant(Offset, AddrVT)),
3543 SDOperand X86TargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) {
3544 SDOperand Chain = Op.getOperand(0);
3546 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3547 if (Align == 0) Align = 1;
3549 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3550 // If not DWORD aligned, call memcpy if size is less than the threshold.
3551 // It knows how to align to the right boundary first.
3552 if ((Align & 3) != 0 ||
3553 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
3554 MVT::ValueType IntPtr = getPointerTy();
3555 TargetLowering::ArgListTy Args;
3556 TargetLowering::ArgListEntry Entry;
3557 Entry.Ty = getTargetData()->getIntPtrType();
3558 Entry.isSigned = false;
3559 Entry.isInReg = false;
3560 Entry.isSRet = false;
3561 Entry.Node = Op.getOperand(1); Args.push_back(Entry);
3562 Entry.Node = Op.getOperand(2); Args.push_back(Entry);
3563 Entry.Node = Op.getOperand(3); Args.push_back(Entry);
3564 std::pair<SDOperand,SDOperand> CallResult =
3565 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
3566 DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
3567 return CallResult.second;
3572 unsigned BytesLeft = 0;
3573 bool TwoRepMovs = false;
3574 switch (Align & 3) {
3575 case 2: // WORD aligned
3578 case 0: // DWORD aligned
3580 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
3583 default: // Byte aligned
3585 Count = Op.getOperand(3);
3589 if (AVT > MVT::i8) {
3591 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
3592 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
3593 BytesLeft = I->getValue() % UBytes;
3595 assert(AVT >= MVT::i32 &&
3596 "Do not use rep;movs if not at least DWORD aligned");
3597 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
3598 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
3603 SDOperand InFlag(0, 0);
3604 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
3606 InFlag = Chain.getValue(1);
3607 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
3608 Op.getOperand(1), InFlag);
3609 InFlag = Chain.getValue(1);
3610 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
3611 Op.getOperand(2), InFlag);
3612 InFlag = Chain.getValue(1);
3614 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3615 SmallVector<SDOperand, 8> Ops;
3616 Ops.push_back(Chain);
3617 Ops.push_back(DAG.getValueType(AVT));
3618 Ops.push_back(InFlag);
3619 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
3622 InFlag = Chain.getValue(1);
3623 Count = Op.getOperand(3);
3624 MVT::ValueType CVT = Count.getValueType();
3625 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
3626 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
3627 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
3629 InFlag = Chain.getValue(1);
3630 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3632 Ops.push_back(Chain);
3633 Ops.push_back(DAG.getValueType(MVT::i8));
3634 Ops.push_back(InFlag);
3635 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
3636 } else if (BytesLeft) {
3637 // Issue loads and stores for the last 1 - 7 bytes.
3638 unsigned Offset = I->getValue() - BytesLeft;
3639 SDOperand DstAddr = Op.getOperand(1);
3640 MVT::ValueType DstVT = DstAddr.getValueType();
3641 SDOperand SrcAddr = Op.getOperand(2);
3642 MVT::ValueType SrcVT = SrcAddr.getValueType();
3644 if (BytesLeft >= 4) {
3645 Value = DAG.getLoad(MVT::i32, Chain,
3646 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3647 DAG.getConstant(Offset, SrcVT)),
3649 Chain = Value.getValue(1);
3650 Chain = DAG.getStore(Chain, Value,
3651 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3652 DAG.getConstant(Offset, DstVT)),
3657 if (BytesLeft >= 2) {
3658 Value = DAG.getLoad(MVT::i16, Chain,
3659 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3660 DAG.getConstant(Offset, SrcVT)),
3662 Chain = Value.getValue(1);
3663 Chain = DAG.getStore(Chain, Value,
3664 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3665 DAG.getConstant(Offset, DstVT)),
3671 if (BytesLeft == 1) {
3672 Value = DAG.getLoad(MVT::i8, Chain,
3673 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3674 DAG.getConstant(Offset, SrcVT)),
3676 Chain = Value.getValue(1);
3677 Chain = DAG.getStore(Chain, Value,
3678 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3679 DAG.getConstant(Offset, DstVT)),
3688 X86TargetLowering::LowerREADCYCLCECOUNTER(SDOperand Op, SelectionDAG &DAG) {
3689 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3690 SDOperand TheOp = Op.getOperand(0);
3691 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheOp, 1);
3692 if (Subtarget->is64Bit()) {
3693 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
3694 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::RDX,
3695 MVT::i64, Copy1.getValue(2));
3696 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, Copy2,
3697 DAG.getConstant(32, MVT::i8));
3699 DAG.getNode(ISD::OR, MVT::i64, Copy1, Tmp), Copy2.getValue(1)
3702 Tys = DAG.getVTList(MVT::i64, MVT::Other);
3703 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2);
3706 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
3707 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::EDX,
3708 MVT::i32, Copy1.getValue(2));
3709 SDOperand Ops[] = { Copy1, Copy2, Copy2.getValue(1) };
3710 Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
3711 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 3);
3714 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
3715 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
3717 if (!Subtarget->is64Bit()) {
3718 // vastart just stores the address of the VarArgsFrameIndex slot into the
3719 // memory location argument.
3720 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
3721 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
3726 // gp_offset (0 - 6 * 8)
3727 // fp_offset (48 - 48 + 8 * 16)
3728 // overflow_arg_area (point to parameters coming in memory).
3730 SmallVector<SDOperand, 8> MemOps;
3731 SDOperand FIN = Op.getOperand(1);
3733 SDOperand Store = DAG.getStore(Op.getOperand(0),
3734 DAG.getConstant(VarArgsGPOffset, MVT::i32),
3735 FIN, SV->getValue(), SV->getOffset());
3736 MemOps.push_back(Store);
3739 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
3740 DAG.getConstant(4, getPointerTy()));
3741 Store = DAG.getStore(Op.getOperand(0),
3742 DAG.getConstant(VarArgsFPOffset, MVT::i32),
3743 FIN, SV->getValue(), SV->getOffset());
3744 MemOps.push_back(Store);
3746 // Store ptr to overflow_arg_area
3747 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
3748 DAG.getConstant(4, getPointerTy()));
3749 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
3750 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
3752 MemOps.push_back(Store);
3754 // Store ptr to reg_save_area.
3755 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
3756 DAG.getConstant(8, getPointerTy()));
3757 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
3758 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
3760 MemOps.push_back(Store);
3761 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
3765 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
3766 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
3768 default: return SDOperand(); // Don't custom lower most intrinsics.
3769 // Comparison intrinsics.
3770 case Intrinsic::x86_sse_comieq_ss:
3771 case Intrinsic::x86_sse_comilt_ss:
3772 case Intrinsic::x86_sse_comile_ss:
3773 case Intrinsic::x86_sse_comigt_ss:
3774 case Intrinsic::x86_sse_comige_ss:
3775 case Intrinsic::x86_sse_comineq_ss:
3776 case Intrinsic::x86_sse_ucomieq_ss:
3777 case Intrinsic::x86_sse_ucomilt_ss:
3778 case Intrinsic::x86_sse_ucomile_ss:
3779 case Intrinsic::x86_sse_ucomigt_ss:
3780 case Intrinsic::x86_sse_ucomige_ss:
3781 case Intrinsic::x86_sse_ucomineq_ss:
3782 case Intrinsic::x86_sse2_comieq_sd:
3783 case Intrinsic::x86_sse2_comilt_sd:
3784 case Intrinsic::x86_sse2_comile_sd:
3785 case Intrinsic::x86_sse2_comigt_sd:
3786 case Intrinsic::x86_sse2_comige_sd:
3787 case Intrinsic::x86_sse2_comineq_sd:
3788 case Intrinsic::x86_sse2_ucomieq_sd:
3789 case Intrinsic::x86_sse2_ucomilt_sd:
3790 case Intrinsic::x86_sse2_ucomile_sd:
3791 case Intrinsic::x86_sse2_ucomigt_sd:
3792 case Intrinsic::x86_sse2_ucomige_sd:
3793 case Intrinsic::x86_sse2_ucomineq_sd: {
3795 ISD::CondCode CC = ISD::SETCC_INVALID;
3798 case Intrinsic::x86_sse_comieq_ss:
3799 case Intrinsic::x86_sse2_comieq_sd:
3803 case Intrinsic::x86_sse_comilt_ss:
3804 case Intrinsic::x86_sse2_comilt_sd:
3808 case Intrinsic::x86_sse_comile_ss:
3809 case Intrinsic::x86_sse2_comile_sd:
3813 case Intrinsic::x86_sse_comigt_ss:
3814 case Intrinsic::x86_sse2_comigt_sd:
3818 case Intrinsic::x86_sse_comige_ss:
3819 case Intrinsic::x86_sse2_comige_sd:
3823 case Intrinsic::x86_sse_comineq_ss:
3824 case Intrinsic::x86_sse2_comineq_sd:
3828 case Intrinsic::x86_sse_ucomieq_ss:
3829 case Intrinsic::x86_sse2_ucomieq_sd:
3830 Opc = X86ISD::UCOMI;
3833 case Intrinsic::x86_sse_ucomilt_ss:
3834 case Intrinsic::x86_sse2_ucomilt_sd:
3835 Opc = X86ISD::UCOMI;
3838 case Intrinsic::x86_sse_ucomile_ss:
3839 case Intrinsic::x86_sse2_ucomile_sd:
3840 Opc = X86ISD::UCOMI;
3843 case Intrinsic::x86_sse_ucomigt_ss:
3844 case Intrinsic::x86_sse2_ucomigt_sd:
3845 Opc = X86ISD::UCOMI;
3848 case Intrinsic::x86_sse_ucomige_ss:
3849 case Intrinsic::x86_sse2_ucomige_sd:
3850 Opc = X86ISD::UCOMI;
3853 case Intrinsic::x86_sse_ucomineq_ss:
3854 case Intrinsic::x86_sse2_ucomineq_sd:
3855 Opc = X86ISD::UCOMI;
3861 SDOperand LHS = Op.getOperand(1);
3862 SDOperand RHS = Op.getOperand(2);
3863 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
3865 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3866 SDOperand Ops1[] = { DAG.getEntryNode(), LHS, RHS };
3867 SDOperand Cond = DAG.getNode(Opc, VTs, 2, Ops1, 3);
3868 VTs = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
3869 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
3870 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, VTs, 2, Ops2, 2);
3871 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
3876 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
3877 // Depths > 0 not supported yet!
3878 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
3881 // Just load the return address
3882 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
3883 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
3886 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
3887 // Depths > 0 not supported yet!
3888 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
3891 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
3892 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
3893 DAG.getConstant(4, getPointerTy()));
3896 /// LowerOperation - Provide custom lowering hooks for some operations.
3898 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
3899 switch (Op.getOpcode()) {
3900 default: assert(0 && "Should not custom lower this!");
3901 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3902 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3903 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3904 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
3905 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
3906 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3907 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
3908 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
3909 case ISD::SHL_PARTS:
3910 case ISD::SRA_PARTS:
3911 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
3912 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
3913 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
3914 case ISD::FABS: return LowerFABS(Op, DAG);
3915 case ISD::FNEG: return LowerFNEG(Op, DAG);
3916 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3917 case ISD::SETCC: return LowerSETCC(Op, DAG, DAG.getEntryNode());
3918 case ISD::SELECT: return LowerSELECT(Op, DAG);
3919 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
3920 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
3921 case ISD::CALL: return LowerCALL(Op, DAG);
3922 case ISD::RET: return LowerRET(Op, DAG);
3923 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
3924 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
3925 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
3926 case ISD::READCYCLECOUNTER: return LowerREADCYCLCECOUNTER(Op, DAG);
3927 case ISD::VASTART: return LowerVASTART(Op, DAG);
3928 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
3929 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3930 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3935 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
3937 default: return NULL;
3938 case X86ISD::SHLD: return "X86ISD::SHLD";
3939 case X86ISD::SHRD: return "X86ISD::SHRD";
3940 case X86ISD::FAND: return "X86ISD::FAND";
3941 case X86ISD::FOR: return "X86ISD::FOR";
3942 case X86ISD::FXOR: return "X86ISD::FXOR";
3943 case X86ISD::FSRL: return "X86ISD::FSRL";
3944 case X86ISD::FILD: return "X86ISD::FILD";
3945 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
3946 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
3947 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
3948 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
3949 case X86ISD::FLD: return "X86ISD::FLD";
3950 case X86ISD::FST: return "X86ISD::FST";
3951 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
3952 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
3953 case X86ISD::CALL: return "X86ISD::CALL";
3954 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
3955 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
3956 case X86ISD::CMP: return "X86ISD::CMP";
3957 case X86ISD::COMI: return "X86ISD::COMI";
3958 case X86ISD::UCOMI: return "X86ISD::UCOMI";
3959 case X86ISD::SETCC: return "X86ISD::SETCC";
3960 case X86ISD::CMOV: return "X86ISD::CMOV";
3961 case X86ISD::BRCOND: return "X86ISD::BRCOND";
3962 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
3963 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
3964 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
3965 case X86ISD::LOAD_PACK: return "X86ISD::LOAD_PACK";
3966 case X86ISD::LOAD_UA: return "X86ISD::LOAD_UA";
3967 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
3968 case X86ISD::Wrapper: return "X86ISD::Wrapper";
3969 case X86ISD::S2VEC: return "X86ISD::S2VEC";
3970 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
3971 case X86ISD::PINSRW: return "X86ISD::PINSRW";
3972 case X86ISD::FMAX: return "X86ISD::FMAX";
3973 case X86ISD::FMIN: return "X86ISD::FMIN";
3977 /// isLegalAddressImmediate - Return true if the integer value or
3978 /// GlobalValue can be used as the offset of the target addressing mode.
3979 bool X86TargetLowering::isLegalAddressImmediate(int64_t V) const {
3980 // X86 allows a sign-extended 32-bit immediate field.
3981 return (V > -(1LL << 32) && V < (1LL << 32)-1);
3984 bool X86TargetLowering::isLegalAddressImmediate(GlobalValue *GV) const {
3985 // In 64-bit mode, GV is 64-bit so it won't fit in the 32-bit displacement
3986 // field unless we are in small code model.
3987 if (Subtarget->is64Bit() &&
3988 getTargetMachine().getCodeModel() != CodeModel::Small)
3991 return (!Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false));
3994 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3995 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3996 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3997 /// are assumed to be legal.
3999 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
4000 // Only do shuffles on 128-bit vector types for now.
4001 if (MVT::getSizeInBits(VT) == 64) return false;
4002 return (Mask.Val->getNumOperands() <= 4 ||
4003 isSplatMask(Mask.Val) ||
4004 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
4005 X86::isUNPCKLMask(Mask.Val) ||
4006 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
4007 X86::isUNPCKHMask(Mask.Val));
4010 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
4012 SelectionDAG &DAG) const {
4013 unsigned NumElts = BVOps.size();
4014 // Only do shuffles on 128-bit vector types for now.
4015 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
4016 if (NumElts == 2) return true;
4018 return (isMOVLMask(&BVOps[0], 4) ||
4019 isCommutedMOVL(&BVOps[0], 4, true) ||
4020 isSHUFPMask(&BVOps[0], 4) ||
4021 isCommutedSHUFP(&BVOps[0], 4));
4026 //===----------------------------------------------------------------------===//
4027 // X86 Scheduler Hooks
4028 //===----------------------------------------------------------------------===//
4031 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
4032 MachineBasicBlock *BB) {
4033 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4034 switch (MI->getOpcode()) {
4035 default: assert(false && "Unexpected instr type to insert");
4036 case X86::CMOV_FR32:
4037 case X86::CMOV_FR64:
4038 case X86::CMOV_V4F32:
4039 case X86::CMOV_V2F64:
4040 case X86::CMOV_V2I64: {
4041 // To "insert" a SELECT_CC instruction, we actually have to insert the
4042 // diamond control-flow pattern. The incoming instruction knows the
4043 // destination vreg to set, the condition code register to branch on, the
4044 // true/false values to select between, and a branch opcode to use.
4045 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4046 ilist<MachineBasicBlock>::iterator It = BB;
4052 // cmpTY ccX, r1, r2
4054 // fallthrough --> copy0MBB
4055 MachineBasicBlock *thisMBB = BB;
4056 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
4057 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
4059 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
4060 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
4061 MachineFunction *F = BB->getParent();
4062 F->getBasicBlockList().insert(It, copy0MBB);
4063 F->getBasicBlockList().insert(It, sinkMBB);
4064 // Update machine-CFG edges by first adding all successors of the current
4065 // block to the new block which will contain the Phi node for the select.
4066 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
4067 e = BB->succ_end(); i != e; ++i)
4068 sinkMBB->addSuccessor(*i);
4069 // Next, remove all successors of the current block, and add the true
4070 // and fallthrough blocks as its successors.
4071 while(!BB->succ_empty())
4072 BB->removeSuccessor(BB->succ_begin());
4073 BB->addSuccessor(copy0MBB);
4074 BB->addSuccessor(sinkMBB);
4077 // %FalseValue = ...
4078 // # fallthrough to sinkMBB
4081 // Update machine-CFG edges
4082 BB->addSuccessor(sinkMBB);
4085 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4088 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
4089 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4090 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4092 delete MI; // The pseudo instruction is gone now.
4096 case X86::FP_TO_INT16_IN_MEM:
4097 case X86::FP_TO_INT32_IN_MEM:
4098 case X86::FP_TO_INT64_IN_MEM: {
4099 // Change the floating point control register to use "round towards zero"
4100 // mode when truncating to an integer value.
4101 MachineFunction *F = BB->getParent();
4102 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
4103 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
4105 // Load the old value of the high byte of the control word...
4107 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
4108 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
4110 // Set the high part to be round to zero...
4111 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
4114 // Reload the modified control word now...
4115 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4117 // Restore the memory image of control word to original value
4118 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
4121 // Get the X86 opcode to use.
4123 switch (MI->getOpcode()) {
4124 default: assert(0 && "illegal opcode!");
4125 case X86::FP_TO_INT16_IN_MEM: Opc = X86::FpIST16m; break;
4126 case X86::FP_TO_INT32_IN_MEM: Opc = X86::FpIST32m; break;
4127 case X86::FP_TO_INT64_IN_MEM: Opc = X86::FpIST64m; break;
4131 MachineOperand &Op = MI->getOperand(0);
4132 if (Op.isRegister()) {
4133 AM.BaseType = X86AddressMode::RegBase;
4134 AM.Base.Reg = Op.getReg();
4136 AM.BaseType = X86AddressMode::FrameIndexBase;
4137 AM.Base.FrameIndex = Op.getFrameIndex();
4139 Op = MI->getOperand(1);
4140 if (Op.isImmediate())
4141 AM.Scale = Op.getImm();
4142 Op = MI->getOperand(2);
4143 if (Op.isImmediate())
4144 AM.IndexReg = Op.getImm();
4145 Op = MI->getOperand(3);
4146 if (Op.isGlobalAddress()) {
4147 AM.GV = Op.getGlobal();
4149 AM.Disp = Op.getImm();
4151 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
4152 .addReg(MI->getOperand(4).getReg());
4154 // Reload the original control word now.
4155 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4157 delete MI; // The pseudo instruction is gone now.
4163 //===----------------------------------------------------------------------===//
4164 // X86 Optimization Hooks
4165 //===----------------------------------------------------------------------===//
4167 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
4169 uint64_t &KnownZero,
4171 unsigned Depth) const {
4172 unsigned Opc = Op.getOpcode();
4173 assert((Opc >= ISD::BUILTIN_OP_END ||
4174 Opc == ISD::INTRINSIC_WO_CHAIN ||
4175 Opc == ISD::INTRINSIC_W_CHAIN ||
4176 Opc == ISD::INTRINSIC_VOID) &&
4177 "Should use MaskedValueIsZero if you don't know whether Op"
4178 " is a target node!");
4180 KnownZero = KnownOne = 0; // Don't know anything.
4184 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
4189 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4190 /// element of the result of the vector shuffle.
4191 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
4192 MVT::ValueType VT = N->getValueType(0);
4193 SDOperand PermMask = N->getOperand(2);
4194 unsigned NumElems = PermMask.getNumOperands();
4195 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
4197 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4199 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4200 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
4201 SDOperand Idx = PermMask.getOperand(i);
4202 if (Idx.getOpcode() == ISD::UNDEF)
4203 return DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4204 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
4209 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
4210 /// node is a GlobalAddress + an offset.
4211 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
4212 unsigned Opc = N->getOpcode();
4213 if (Opc == X86ISD::Wrapper) {
4214 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
4215 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
4218 } else if (Opc == ISD::ADD) {
4219 SDOperand N1 = N->getOperand(0);
4220 SDOperand N2 = N->getOperand(1);
4221 if (isGAPlusOffset(N1.Val, GA, Offset)) {
4222 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
4224 Offset += V->getSignExtended();
4227 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
4228 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
4230 Offset += V->getSignExtended();
4238 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
4240 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
4241 MachineFrameInfo *MFI) {
4242 if (N->getOperand(0).Val != Base->getOperand(0).Val)
4245 SDOperand Loc = N->getOperand(1);
4246 SDOperand BaseLoc = Base->getOperand(1);
4247 if (Loc.getOpcode() == ISD::FrameIndex) {
4248 if (BaseLoc.getOpcode() != ISD::FrameIndex)
4250 int FI = dyn_cast<FrameIndexSDNode>(Loc)->getIndex();
4251 int BFI = dyn_cast<FrameIndexSDNode>(BaseLoc)->getIndex();
4252 int FS = MFI->getObjectSize(FI);
4253 int BFS = MFI->getObjectSize(BFI);
4254 if (FS != BFS || FS != Size) return false;
4255 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
4257 GlobalValue *GV1 = NULL;
4258 GlobalValue *GV2 = NULL;
4259 int64_t Offset1 = 0;
4260 int64_t Offset2 = 0;
4261 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
4262 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
4263 if (isGA1 && isGA2 && GV1 == GV2)
4264 return Offset1 == (Offset2 + Dist*Size);
4270 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
4271 const X86Subtarget *Subtarget) {
4274 if (isGAPlusOffset(Base, GV, Offset))
4275 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
4277 assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
4278 int BFI = dyn_cast<FrameIndexSDNode>(Base)->getIndex();
4280 // Fixed objects do not specify alignment, however the offsets are known.
4281 return ((Subtarget->getStackAlignment() % 16) == 0 &&
4282 (MFI->getObjectOffset(BFI) % 16) == 0);
4284 return MFI->getObjectAlignment(BFI) >= 16;
4290 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
4291 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
4292 /// if the load addresses are consecutive, non-overlapping, and in the right
4294 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
4295 const X86Subtarget *Subtarget) {
4296 MachineFunction &MF = DAG.getMachineFunction();
4297 MachineFrameInfo *MFI = MF.getFrameInfo();
4298 MVT::ValueType VT = N->getValueType(0);
4299 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
4300 SDOperand PermMask = N->getOperand(2);
4301 int NumElems = (int)PermMask.getNumOperands();
4302 SDNode *Base = NULL;
4303 for (int i = 0; i < NumElems; ++i) {
4304 SDOperand Idx = PermMask.getOperand(i);
4305 if (Idx.getOpcode() == ISD::UNDEF) {
4306 if (!Base) return SDOperand();
4309 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
4310 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
4314 else if (!isConsecutiveLoad(Arg.Val, Base,
4315 i, MVT::getSizeInBits(EVT)/8,MFI))
4320 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
4322 LoadSDNode *LD = cast<LoadSDNode>(Base);
4323 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
4324 LD->getSrcValueOffset());
4326 // Just use movups, it's shorter.
4327 SDVTList Tys = DAG.getVTList(MVT::v4f32, MVT::Other);
4328 SmallVector<SDOperand, 3> Ops;
4329 Ops.push_back(Base->getOperand(0));
4330 Ops.push_back(Base->getOperand(1));
4331 Ops.push_back(Base->getOperand(2));
4332 return DAG.getNode(ISD::BIT_CONVERT, VT,
4333 DAG.getNode(X86ISD::LOAD_UA, Tys, &Ops[0], Ops.size()));
4337 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
4338 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
4339 const X86Subtarget *Subtarget) {
4340 SDOperand Cond = N->getOperand(0);
4342 // If we have SSE[12] support, try to form min/max nodes.
4343 if (Subtarget->hasSSE2() &&
4344 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
4345 if (Cond.getOpcode() == ISD::SETCC) {
4346 // Get the LHS/RHS of the select.
4347 SDOperand LHS = N->getOperand(1);
4348 SDOperand RHS = N->getOperand(2);
4349 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
4351 unsigned Opcode = 0;
4352 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
4355 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
4358 if (!UnsafeFPMath) break;
4360 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
4362 Opcode = X86ISD::FMIN;
4365 case ISD::SETOGT: // (X > Y) ? X : Y -> max
4368 if (!UnsafeFPMath) break;
4370 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
4372 Opcode = X86ISD::FMAX;
4375 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
4378 case ISD::SETOGT: // (X > Y) ? Y : X -> min
4381 if (!UnsafeFPMath) break;
4383 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
4385 Opcode = X86ISD::FMIN;
4388 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
4391 if (!UnsafeFPMath) break;
4393 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
4395 Opcode = X86ISD::FMAX;
4401 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
4410 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
4411 DAGCombinerInfo &DCI) const {
4412 SelectionDAG &DAG = DCI.DAG;
4413 switch (N->getOpcode()) {
4415 case ISD::VECTOR_SHUFFLE:
4416 return PerformShuffleCombine(N, DAG, Subtarget);
4418 return PerformSELECTCombine(N, DAG, Subtarget);
4424 //===----------------------------------------------------------------------===//
4425 // X86 Inline Assembly Support
4426 //===----------------------------------------------------------------------===//
4428 /// getConstraintType - Given a constraint letter, return the type of
4429 /// constraint it is for this target.
4430 X86TargetLowering::ConstraintType
4431 X86TargetLowering::getConstraintType(char ConstraintLetter) const {
4432 switch (ConstraintLetter) {
4441 return C_RegisterClass;
4442 default: return TargetLowering::getConstraintType(ConstraintLetter);
4446 /// isOperandValidForConstraint - Return the specified operand (possibly
4447 /// modified) if the specified SDOperand is valid for the specified target
4448 /// constraint letter, otherwise return null.
4449 SDOperand X86TargetLowering::
4450 isOperandValidForConstraint(SDOperand Op, char Constraint, SelectionDAG &DAG) {
4451 switch (Constraint) {
4454 // Literal immediates are always ok.
4455 if (isa<ConstantSDNode>(Op)) return Op;
4457 // If we are in non-pic codegen mode, we allow the address of a global to
4458 // be used with 'i'.
4459 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op)) {
4460 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
4461 return SDOperand(0, 0);
4463 if (GA->getOpcode() != ISD::TargetGlobalAddress)
4464 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
4469 // Otherwise, not valid for this mode.
4470 return SDOperand(0, 0);
4472 return TargetLowering::isOperandValidForConstraint(Op, Constraint, DAG);
4476 std::vector<unsigned> X86TargetLowering::
4477 getRegClassForInlineAsmConstraint(const std::string &Constraint,
4478 MVT::ValueType VT) const {
4479 if (Constraint.size() == 1) {
4480 // FIXME: not handling fp-stack yet!
4481 // FIXME: not handling MMX registers yet ('y' constraint).
4482 switch (Constraint[0]) { // GCC X86 Constraint Letters
4483 default: break; // Unknown constraint letter
4484 case 'A': // EAX/EDX
4485 if (VT == MVT::i32 || VT == MVT::i64)
4486 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
4488 case 'r': // GENERAL_REGS
4489 case 'R': // LEGACY_REGS
4490 if (VT == MVT::i64 && Subtarget->is64Bit())
4491 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
4492 X86::RSI, X86::RDI, X86::RBP, X86::RSP,
4493 X86::R8, X86::R9, X86::R10, X86::R11,
4494 X86::R12, X86::R13, X86::R14, X86::R15, 0);
4496 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
4497 X86::ESI, X86::EDI, X86::EBP, X86::ESP, 0);
4498 else if (VT == MVT::i16)
4499 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
4500 X86::SI, X86::DI, X86::BP, X86::SP, 0);
4501 else if (VT == MVT::i8)
4502 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
4504 case 'l': // INDEX_REGS
4506 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
4507 X86::ESI, X86::EDI, X86::EBP, 0);
4508 else if (VT == MVT::i16)
4509 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
4510 X86::SI, X86::DI, X86::BP, 0);
4511 else if (VT == MVT::i8)
4512 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
4514 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
4517 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
4518 else if (VT == MVT::i16)
4519 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
4520 else if (VT == MVT::i8)
4521 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
4523 case 'x': // SSE_REGS if SSE1 allowed
4524 if (Subtarget->hasSSE1())
4525 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
4526 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
4528 return std::vector<unsigned>();
4529 case 'Y': // SSE_REGS if SSE2 allowed
4530 if (Subtarget->hasSSE2())
4531 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
4532 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
4534 return std::vector<unsigned>();
4538 return std::vector<unsigned>();
4541 std::pair<unsigned, const TargetRegisterClass*>
4542 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4543 MVT::ValueType VT) const {
4544 // Use the default implementation in TargetLowering to convert the register
4545 // constraint into a member of a register class.
4546 std::pair<unsigned, const TargetRegisterClass*> Res;
4547 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4549 // Not found as a standard register?
4550 if (Res.second == 0) {
4551 // GCC calls "st(0)" just plain "st".
4552 if (StringsEqualNoCase("{st}", Constraint)) {
4553 Res.first = X86::ST0;
4554 Res.second = X86::RSTRegisterClass;
4560 // Otherwise, check to see if this is a register class of the wrong value
4561 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
4562 // turn into {ax},{dx}.
4563 if (Res.second->hasType(VT))
4564 return Res; // Correct type already, nothing to do.
4566 // All of the single-register GCC register classes map their values onto
4567 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
4568 // really want an 8-bit or 32-bit register, map to the appropriate register
4569 // class and return the appropriate register.
4570 if (Res.second != X86::GR16RegisterClass)
4573 if (VT == MVT::i8) {
4574 unsigned DestReg = 0;
4575 switch (Res.first) {
4577 case X86::AX: DestReg = X86::AL; break;
4578 case X86::DX: DestReg = X86::DL; break;
4579 case X86::CX: DestReg = X86::CL; break;
4580 case X86::BX: DestReg = X86::BL; break;
4583 Res.first = DestReg;
4584 Res.second = Res.second = X86::GR8RegisterClass;
4586 } else if (VT == MVT::i32) {
4587 unsigned DestReg = 0;
4588 switch (Res.first) {
4590 case X86::AX: DestReg = X86::EAX; break;
4591 case X86::DX: DestReg = X86::EDX; break;
4592 case X86::CX: DestReg = X86::ECX; break;
4593 case X86::BX: DestReg = X86::EBX; break;
4594 case X86::SI: DestReg = X86::ESI; break;
4595 case X86::DI: DestReg = X86::EDI; break;
4596 case X86::BP: DestReg = X86::EBP; break;
4597 case X86::SP: DestReg = X86::ESP; break;
4600 Res.first = DestReg;
4601 Res.second = Res.second = X86::GR32RegisterClass;
4603 } else if (VT == MVT::i64) {
4604 unsigned DestReg = 0;
4605 switch (Res.first) {
4607 case X86::AX: DestReg = X86::RAX; break;
4608 case X86::DX: DestReg = X86::RDX; break;
4609 case X86::CX: DestReg = X86::RCX; break;
4610 case X86::BX: DestReg = X86::RBX; break;
4611 case X86::SI: DestReg = X86::RSI; break;
4612 case X86::DI: DestReg = X86::RDI; break;
4613 case X86::BP: DestReg = X86::RBP; break;
4614 case X86::SP: DestReg = X86::RSP; break;
4617 Res.first = DestReg;
4618 Res.second = Res.second = X86::GR64RegisterClass;