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/Support/CommandLine.h"
36 #include "llvm/ADT/StringExtras.h"
40 static cl::opt<bool> EnableFastCC("enable-x86-fastcc", cl::Hidden,
41 cl::desc("Enable fastcc on X86"));
42 X86TargetLowering::X86TargetLowering(TargetMachine &TM)
43 : TargetLowering(TM) {
44 Subtarget = &TM.getSubtarget<X86Subtarget>();
45 X86ScalarSSE = Subtarget->hasSSE2();
46 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
48 // Set up the TargetLowering object.
50 // X86 is weird, it always uses i8 for shift amounts and setcc results.
51 setShiftAmountType(MVT::i8);
52 setSetCCResultType(MVT::i8);
53 setSetCCResultContents(ZeroOrOneSetCCResult);
54 setSchedulingPreference(SchedulingForRegPressure);
55 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
56 setStackPointerRegisterToSaveRestore(X86StackPtr);
58 if (Subtarget->isTargetDarwin()) {
59 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
60 setUseUnderscoreSetJmp(false);
61 setUseUnderscoreLongJmp(false);
62 } else if (Subtarget->isTargetMingw()) {
63 // MS runtime is weird: it exports _setjmp, but longjmp!
64 setUseUnderscoreSetJmp(true);
65 setUseUnderscoreLongJmp(false);
67 setUseUnderscoreSetJmp(true);
68 setUseUnderscoreLongJmp(true);
71 // Add legal addressing mode scale values.
72 addLegalAddressScale(8);
73 addLegalAddressScale(4);
74 addLegalAddressScale(2);
75 // Enter the ones which require both scale + index last. These are more
77 addLegalAddressScale(9);
78 addLegalAddressScale(5);
79 addLegalAddressScale(3);
81 // Set up the register classes.
82 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
83 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
84 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
85 if (Subtarget->is64Bit())
86 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
88 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Expand);
90 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
92 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
93 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
94 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
96 if (Subtarget->is64Bit()) {
97 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
98 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
101 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
102 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
104 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
107 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
109 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
110 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
111 // SSE has no i16 to fp conversion, only i32
113 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
115 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
116 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
119 if (!Subtarget->is64Bit()) {
120 // Custom lower SINT_TO_FP and FP_TO_SINT from/to i64 in 32-bit mode.
121 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
122 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
125 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
127 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
128 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
131 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
133 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
134 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
137 // Handle FP_TO_UINT by promoting the destination to a larger signed
139 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
140 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
141 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
143 if (Subtarget->is64Bit()) {
144 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
145 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
147 if (X86ScalarSSE && !Subtarget->hasSSE3())
148 // Expand FP_TO_UINT into a select.
149 // FIXME: We would like to use a Custom expander here eventually to do
150 // the optimal thing for SSE vs. the default expansion in the legalizer.
151 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
153 // With SSE3 we can use fisttpll to convert to a signed i64.
154 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
157 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
159 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
160 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
163 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
164 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
165 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
166 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
167 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
168 if (Subtarget->is64Bit())
169 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Expand);
170 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Expand);
171 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Expand);
172 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
173 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
174 setOperationAction(ISD::FREM , MVT::f64 , Expand);
176 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
177 setOperationAction(ISD::CTTZ , MVT::i8 , Expand);
178 setOperationAction(ISD::CTLZ , MVT::i8 , Expand);
179 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
180 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
181 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
182 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
183 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
184 setOperationAction(ISD::CTLZ , MVT::i32 , Expand);
185 if (Subtarget->is64Bit()) {
186 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
187 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
188 setOperationAction(ISD::CTLZ , MVT::i64 , Expand);
191 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
192 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
194 // These should be promoted to a larger select which is supported.
195 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
196 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
197 // X86 wants to expand cmov itself.
198 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
199 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
200 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
201 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
202 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
203 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
204 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
205 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
206 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
207 if (Subtarget->is64Bit()) {
208 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
209 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
211 // X86 ret instruction may pop stack.
212 setOperationAction(ISD::RET , MVT::Other, Custom);
214 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
215 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
216 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
217 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
218 if (Subtarget->is64Bit()) {
219 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
220 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
221 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
222 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
224 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
225 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
226 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
227 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
228 // X86 wants to expand memset / memcpy itself.
229 setOperationAction(ISD::MEMSET , MVT::Other, Custom);
230 setOperationAction(ISD::MEMCPY , MVT::Other, Custom);
232 // We don't have line number support yet.
233 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
234 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
235 // FIXME - use subtarget debug flags
236 if (!Subtarget->isTargetDarwin() &&
237 !Subtarget->isTargetELF() &&
238 !Subtarget->isTargetCygMing())
239 setOperationAction(ISD::LABEL, MVT::Other, Expand);
241 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
242 setOperationAction(ISD::VASTART , MVT::Other, Custom);
244 // Use the default implementation.
245 setOperationAction(ISD::VAARG , MVT::Other, Expand);
246 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
247 setOperationAction(ISD::VAEND , MVT::Other, Expand);
248 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
249 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
250 if (Subtarget->is64Bit())
251 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
252 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
255 // Set up the FP register classes.
256 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
257 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
259 // Use ANDPD to simulate FABS.
260 setOperationAction(ISD::FABS , MVT::f64, Custom);
261 setOperationAction(ISD::FABS , MVT::f32, Custom);
263 // Use XORP to simulate FNEG.
264 setOperationAction(ISD::FNEG , MVT::f64, Custom);
265 setOperationAction(ISD::FNEG , MVT::f32, Custom);
267 // Use ANDPD and ORPD to simulate FCOPYSIGN.
268 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
269 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
271 // We don't support sin/cos/fmod
272 setOperationAction(ISD::FSIN , MVT::f64, Expand);
273 setOperationAction(ISD::FCOS , MVT::f64, Expand);
274 setOperationAction(ISD::FREM , MVT::f64, Expand);
275 setOperationAction(ISD::FSIN , MVT::f32, Expand);
276 setOperationAction(ISD::FCOS , MVT::f32, Expand);
277 setOperationAction(ISD::FREM , MVT::f32, Expand);
279 // Expand FP immediates into loads from the stack, except for the special
281 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
282 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
283 addLegalFPImmediate(+0.0); // xorps / xorpd
285 // Set up the FP register classes.
286 addRegisterClass(MVT::f64, X86::RFPRegisterClass);
288 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
289 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
290 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
293 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
294 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
297 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
298 addLegalFPImmediate(+0.0); // FLD0
299 addLegalFPImmediate(+1.0); // FLD1
300 addLegalFPImmediate(-0.0); // FLD0/FCHS
301 addLegalFPImmediate(-1.0); // FLD1/FCHS
304 // First set operation action for all vector types to expand. Then we
305 // will selectively turn on ones that can be effectively codegen'd.
306 for (unsigned VT = (unsigned)MVT::Vector + 1;
307 VT != (unsigned)MVT::LAST_VALUETYPE; VT++) {
308 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
309 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
310 setOperationAction(ISD::FADD, (MVT::ValueType)VT, Expand);
311 setOperationAction(ISD::FSUB, (MVT::ValueType)VT, Expand);
312 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
313 setOperationAction(ISD::FMUL, (MVT::ValueType)VT, Expand);
314 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
315 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
316 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
317 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
318 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
319 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand);
320 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Expand);
321 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
322 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
325 if (Subtarget->hasMMX()) {
326 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
327 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
328 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
330 // FIXME: add MMX packed arithmetics
331 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Expand);
332 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Expand);
333 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Expand);
336 if (Subtarget->hasSSE1()) {
337 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
339 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
340 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
341 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
342 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
343 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
344 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
345 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
346 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
347 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
350 if (Subtarget->hasSSE2()) {
351 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
352 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
353 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
354 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
355 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
357 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
358 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
359 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
360 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
361 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
362 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
363 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
364 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
365 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
366 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
367 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
369 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
370 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
371 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
372 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
373 // Implement v4f32 insert_vector_elt in terms of SSE2 v8i16 ones.
374 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
376 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
377 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
378 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
379 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
380 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
382 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
383 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
384 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
385 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
386 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
387 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
389 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
390 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
391 setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote);
392 AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64);
393 setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote);
394 AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64);
395 setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote);
396 AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64);
397 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote);
398 AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64);
399 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
400 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64);
403 // Custom lower v2i64 and v2f64 selects.
404 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
405 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
406 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
407 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
410 // We want to custom lower some of our intrinsics.
411 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
413 // We have target-specific dag combine patterns for the following nodes:
414 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
415 setTargetDAGCombine(ISD::SELECT);
417 computeRegisterProperties();
419 // FIXME: These should be based on subtarget info. Plus, the values should
420 // be smaller when we are in optimizing for size mode.
421 maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
422 maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
423 maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
424 allowUnalignedMemoryAccesses = true; // x86 supports it!
428 //===----------------------------------------------------------------------===//
429 // Return Value Calling Convention Implementation
430 //===----------------------------------------------------------------------===//
432 #include "X86GenCallingConv.inc"
434 /// LowerRET - Lower an ISD::RET node.
435 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
436 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
438 SmallVector<CCValAssign, 16> RVLocs;
439 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
440 CCState CCInfo(CC, getTargetMachine(), RVLocs);
441 CCInfo.AnalyzeReturn(Op.Val, RetCC_X86);
444 // If this is the first return lowered for this function, add the regs to the
445 // liveout set for the function.
446 if (DAG.getMachineFunction().liveout_empty()) {
447 for (unsigned i = 0; i != RVLocs.size(); ++i)
448 if (RVLocs[i].isRegLoc())
449 DAG.getMachineFunction().addLiveOut(RVLocs[i].getLocReg());
452 SDOperand Chain = Op.getOperand(0);
455 // Copy the result values into the output registers.
456 if (RVLocs.size() != 1 || !RVLocs[0].isRegLoc() ||
457 RVLocs[0].getLocReg() != X86::ST0) {
458 for (unsigned i = 0; i != RVLocs.size(); ++i) {
459 CCValAssign &VA = RVLocs[i];
460 assert(VA.isRegLoc() && "Can only return in registers!");
461 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1),
463 Flag = Chain.getValue(1);
466 // We need to handle a destination of ST0 specially, because it isn't really
468 SDOperand Value = Op.getOperand(1);
470 // If this is an FP return with ScalarSSE, we need to move the value from
471 // an XMM register onto the fp-stack.
475 // If this is a load into a scalarsse value, don't store the loaded value
476 // back to the stack, only to reload it: just replace the scalar-sse load.
477 if (ISD::isNON_EXTLoad(Value.Val) &&
478 (Chain == Value.getValue(1) || Chain == Value.getOperand(0))) {
479 Chain = Value.getOperand(0);
480 MemLoc = Value.getOperand(1);
482 // Spill the value to memory and reload it into top of stack.
483 unsigned Size = MVT::getSizeInBits(RVLocs[0].getValVT())/8;
484 MachineFunction &MF = DAG.getMachineFunction();
485 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
486 MemLoc = DAG.getFrameIndex(SSFI, getPointerTy());
487 Chain = DAG.getStore(Op.getOperand(0), Value, MemLoc, NULL, 0);
489 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
490 SDOperand Ops[] = {Chain, MemLoc, DAG.getValueType(RVLocs[0].getValVT())};
491 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
492 Chain = Value.getValue(1);
495 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
496 SDOperand Ops[] = { Chain, Value };
497 Chain = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops, 2);
498 Flag = Chain.getValue(1);
501 SDOperand BytesToPop = DAG.getConstant(getBytesToPopOnReturn(), MVT::i16);
503 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop, Flag);
505 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop);
509 /// LowerCallResult - Lower the result values of an ISD::CALL into the
510 /// appropriate copies out of appropriate physical registers. This assumes that
511 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
512 /// being lowered. The returns a SDNode with the same number of values as the
514 SDNode *X86TargetLowering::
515 LowerCallResult(SDOperand Chain, SDOperand InFlag, SDNode *TheCall,
516 unsigned CallingConv, SelectionDAG &DAG) {
518 // Assign locations to each value returned by this call.
519 SmallVector<CCValAssign, 16> RVLocs;
520 CCState CCInfo(CallingConv, getTargetMachine(), RVLocs);
521 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
524 SmallVector<SDOperand, 8> ResultVals;
526 // Copy all of the result registers out of their specified physreg.
527 if (RVLocs.size() != 1 || RVLocs[0].getLocReg() != X86::ST0) {
528 for (unsigned i = 0; i != RVLocs.size(); ++i) {
529 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
530 RVLocs[i].getValVT(), InFlag).getValue(1);
531 InFlag = Chain.getValue(2);
532 ResultVals.push_back(Chain.getValue(0));
535 // Copies from the FP stack are special, as ST0 isn't a valid register
536 // before the fp stackifier runs.
538 // Copy ST0 into an RFP register with FP_GET_RESULT.
539 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
540 SDOperand GROps[] = { Chain, InFlag };
541 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, GROps, 2);
542 Chain = RetVal.getValue(1);
543 InFlag = RetVal.getValue(2);
545 // If we are using ScalarSSE, store ST(0) to the stack and reload it into
548 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
549 // shouldn't be necessary except that RFP cannot be live across
550 // multiple blocks. When stackifier is fixed, they can be uncoupled.
551 MachineFunction &MF = DAG.getMachineFunction();
552 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
553 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
555 Chain, RetVal, StackSlot, DAG.getValueType(RVLocs[0].getValVT()), InFlag
557 Chain = DAG.getNode(X86ISD::FST, MVT::Other, Ops, 5);
558 RetVal = DAG.getLoad(RVLocs[0].getValVT(), Chain, StackSlot, NULL, 0);
559 Chain = RetVal.getValue(1);
562 if (RVLocs[0].getValVT() == MVT::f32 && !X86ScalarSSE)
563 // FIXME: we would really like to remember that this FP_ROUND
564 // operation is okay to eliminate if we allow excess FP precision.
565 RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal);
566 ResultVals.push_back(RetVal);
569 // Merge everything together with a MERGE_VALUES node.
570 ResultVals.push_back(Chain);
571 return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(),
572 &ResultVals[0], ResultVals.size()).Val;
576 //===----------------------------------------------------------------------===//
577 // C & StdCall Calling Convention implementation
578 //===----------------------------------------------------------------------===//
579 // StdCall calling convention seems to be standard for many Windows' API
580 // routines and around. It differs from C calling convention just a little:
581 // callee should clean up the stack, not caller. Symbols should be also
582 // decorated in some fancy way :) It doesn't support any vector arguments.
584 /// AddLiveIn - This helper function adds the specified physical register to the
585 /// MachineFunction as a live in value. It also creates a corresponding virtual
587 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
588 const TargetRegisterClass *RC) {
589 assert(RC->contains(PReg) && "Not the correct regclass!");
590 unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC);
591 MF.addLiveIn(PReg, VReg);
595 SDOperand X86TargetLowering::LowerCCCArguments(SDOperand Op, SelectionDAG &DAG,
597 unsigned NumArgs = Op.Val->getNumValues() - 1;
598 MachineFunction &MF = DAG.getMachineFunction();
599 MachineFrameInfo *MFI = MF.getFrameInfo();
600 SDOperand Root = Op.getOperand(0);
601 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
603 // Assign locations to all of the incoming arguments.
604 SmallVector<CCValAssign, 16> ArgLocs;
605 CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(),
607 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_C);
609 SmallVector<SDOperand, 8> ArgValues;
610 unsigned LastVal = ~0U;
611 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
612 CCValAssign &VA = ArgLocs[i];
613 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
615 assert(VA.getValNo() != LastVal &&
616 "Don't support value assigned to multiple locs yet");
617 LastVal = VA.getValNo();
620 MVT::ValueType RegVT = VA.getLocVT();
621 TargetRegisterClass *RC;
622 if (RegVT == MVT::i32)
623 RC = X86::GR32RegisterClass;
625 assert(MVT::isVector(RegVT));
626 RC = X86::VR128RegisterClass;
629 SDOperand ArgValue = DAG.getCopyFromReg(Root, VA.getLocReg(), RegVT);
630 AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
632 // If this is an 8 or 16-bit value, it is really passed promoted to 32
633 // bits. Insert an assert[sz]ext to capture this, then truncate to the
635 if (VA.getLocInfo() == CCValAssign::SExt)
636 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
637 DAG.getValueType(VA.getValVT()));
638 else if (VA.getLocInfo() == CCValAssign::ZExt)
639 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
640 DAG.getValueType(VA.getValVT()));
642 if (VA.getLocInfo() != CCValAssign::Full)
643 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
645 ArgValues.push_back(ArgValue);
647 assert(VA.isMemLoc());
649 // Create the nodes corresponding to a load from this parameter slot.
650 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
651 VA.getLocMemOffset());
652 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
653 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
657 unsigned StackSize = CCInfo.getNextStackOffset();
659 ArgValues.push_back(Root);
661 // If the function takes variable number of arguments, make a frame index for
662 // the start of the first vararg value... for expansion of llvm.va_start.
664 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
666 if (isStdCall && !isVarArg) {
667 BytesToPopOnReturn = StackSize; // Callee pops everything..
668 BytesCallerReserves = 0;
670 BytesToPopOnReturn = 0; // Callee pops hidden struct pointer.
672 // If this is an sret function, the return should pop the hidden pointer.
673 if (NumArgs && (cast<ConstantSDNode>(Op.getOperand(3))->getValue() & 4))
674 BytesToPopOnReturn = 4;
676 BytesCallerReserves = StackSize;
679 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
680 ReturnAddrIndex = 0; // No return address slot generated yet.
682 MF.getInfo<X86FunctionInfo>()->setBytesToPopOnReturn(BytesToPopOnReturn);
684 // Return the new list of results.
685 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
686 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
689 SDOperand X86TargetLowering::LowerCCCCallTo(SDOperand Op, SelectionDAG &DAG,
691 SDOperand Chain = Op.getOperand(0);
692 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
693 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
694 SDOperand Callee = Op.getOperand(4);
695 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
697 // Analyze operands of the call, assigning locations to each operand.
698 SmallVector<CCValAssign, 16> ArgLocs;
699 CCState CCInfo(CC, getTargetMachine(), ArgLocs);
700 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_C);
702 // Get a count of how many bytes are to be pushed on the stack.
703 unsigned NumBytes = CCInfo.getNextStackOffset();
705 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
707 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
708 SmallVector<SDOperand, 8> MemOpChains;
712 // Walk the register/memloc assignments, inserting copies/loads.
713 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
714 CCValAssign &VA = ArgLocs[i];
715 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
717 // Promote the value if needed.
718 switch (VA.getLocInfo()) {
719 default: assert(0 && "Unknown loc info!");
720 case CCValAssign::Full: break;
721 case CCValAssign::SExt:
722 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
724 case CCValAssign::ZExt:
725 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
727 case CCValAssign::AExt:
728 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
733 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
735 assert(VA.isMemLoc());
736 if (StackPtr.Val == 0)
737 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
738 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
739 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
740 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
744 // If the first argument is an sret pointer, remember it.
745 bool isSRet = NumOps &&(cast<ConstantSDNode>(Op.getOperand(6))->getValue()&4);
747 if (!MemOpChains.empty())
748 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
749 &MemOpChains[0], MemOpChains.size());
751 // Build a sequence of copy-to-reg nodes chained together with token chain
752 // and flag operands which copy the outgoing args into registers.
754 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
755 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
757 InFlag = Chain.getValue(1);
760 // ELF / PIC requires GOT in the EBX register before function calls via PLT
762 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
763 Subtarget->isPICStyleGOT()) {
764 Chain = DAG.getCopyToReg(Chain, X86::EBX,
765 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
767 InFlag = Chain.getValue(1);
770 // If the callee is a GlobalAddress node (quite common, every direct call is)
771 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
772 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
773 // We should use extra load for direct calls to dllimported functions in
775 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
776 getTargetMachine(), true))
777 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
778 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
779 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
781 // Returns a chain & a flag for retval copy to use.
782 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
783 SmallVector<SDOperand, 8> Ops;
784 Ops.push_back(Chain);
785 Ops.push_back(Callee);
787 // Add argument registers to the end of the list so that they are known live
789 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
790 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
791 RegsToPass[i].second.getValueType()));
793 // Add an implicit use GOT pointer in EBX.
794 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
795 Subtarget->isPICStyleGOT())
796 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
799 Ops.push_back(InFlag);
801 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
802 NodeTys, &Ops[0], Ops.size());
803 InFlag = Chain.getValue(1);
805 // Create the CALLSEQ_END node.
806 unsigned NumBytesForCalleeToPush = 0;
808 if (CC == CallingConv::X86_StdCall) {
810 NumBytesForCalleeToPush = isSRet ? 4 : 0;
812 NumBytesForCalleeToPush = NumBytes;
814 // If this is is a call to a struct-return function, the callee
815 // pops the hidden struct pointer, so we have to push it back.
816 // This is common for Darwin/X86, Linux & Mingw32 targets.
817 NumBytesForCalleeToPush = isSRet ? 4 : 0;
820 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
822 Ops.push_back(Chain);
823 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
824 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
825 Ops.push_back(InFlag);
826 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
827 InFlag = Chain.getValue(1);
829 // Handle result values, copying them out of physregs into vregs that we
831 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
835 //===----------------------------------------------------------------------===//
836 // FastCall Calling Convention implementation
837 //===----------------------------------------------------------------------===//
839 // The X86 'fastcall' calling convention passes up to two integer arguments in
840 // registers (an appropriate portion of ECX/EDX), passes arguments in C order,
841 // and requires that the callee pop its arguments off the stack (allowing proper
842 // tail calls), and has the same return value conventions as C calling convs.
844 // This calling convention always arranges for the callee pop value to be 8n+4
845 // bytes, which is needed for tail recursion elimination and stack alignment
848 X86TargetLowering::LowerFastCCArguments(SDOperand Op, SelectionDAG &DAG) {
849 MachineFunction &MF = DAG.getMachineFunction();
850 MachineFrameInfo *MFI = MF.getFrameInfo();
851 SDOperand Root = Op.getOperand(0);
853 // Assign locations to all of the incoming arguments.
854 SmallVector<CCValAssign, 16> ArgLocs;
855 CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(),
857 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_FastCall);
859 SmallVector<SDOperand, 8> ArgValues;
860 unsigned LastVal = ~0U;
861 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
862 CCValAssign &VA = ArgLocs[i];
863 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
865 assert(VA.getValNo() != LastVal &&
866 "Don't support value assigned to multiple locs yet");
867 LastVal = VA.getValNo();
870 MVT::ValueType RegVT = VA.getLocVT();
871 TargetRegisterClass *RC;
872 if (RegVT == MVT::i32)
873 RC = X86::GR32RegisterClass;
875 assert(MVT::isVector(RegVT));
876 RC = X86::VR128RegisterClass;
879 SDOperand ArgValue = DAG.getCopyFromReg(Root, VA.getLocReg(), RegVT);
880 AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
882 // If this is an 8 or 16-bit value, it is really passed promoted to 32
883 // bits. Insert an assert[sz]ext to capture this, then truncate to the
885 if (VA.getLocInfo() == CCValAssign::SExt)
886 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
887 DAG.getValueType(VA.getValVT()));
888 else if (VA.getLocInfo() == CCValAssign::ZExt)
889 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
890 DAG.getValueType(VA.getValVT()));
892 if (VA.getLocInfo() != CCValAssign::Full)
893 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
895 ArgValues.push_back(ArgValue);
897 assert(VA.isMemLoc());
899 // Create the nodes corresponding to a load from this parameter slot.
900 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
901 VA.getLocMemOffset());
902 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
903 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
907 ArgValues.push_back(Root);
909 unsigned StackSize = CCInfo.getNextStackOffset();
911 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
912 // arguments and the arguments after the retaddr has been pushed are aligned.
913 if ((StackSize & 7) == 0)
916 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
917 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
918 ReturnAddrIndex = 0; // No return address slot generated yet.
919 BytesToPopOnReturn = StackSize; // Callee pops all stack arguments.
920 BytesCallerReserves = 0;
922 MF.getInfo<X86FunctionInfo>()->setBytesToPopOnReturn(BytesToPopOnReturn);
924 // Return the new list of results.
925 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
926 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
929 SDOperand X86TargetLowering::LowerFastCCCallTo(SDOperand Op, SelectionDAG &DAG,
931 SDOperand Chain = Op.getOperand(0);
932 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
933 SDOperand Callee = Op.getOperand(4);
935 // Analyze operands of the call, assigning locations to each operand.
936 SmallVector<CCValAssign, 16> ArgLocs;
937 CCState CCInfo(CC, getTargetMachine(), ArgLocs);
938 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_FastCall);
940 // Get a count of how many bytes are to be pushed on the stack.
941 unsigned NumBytes = CCInfo.getNextStackOffset();
943 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
944 // arguments and the arguments after the retaddr has been pushed are aligned.
945 if ((NumBytes & 7) == 0)
948 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
951 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
952 SmallVector<SDOperand, 8> MemOpChains;
956 // Walk the register/memloc assignments, inserting copies/loads.
957 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
958 CCValAssign &VA = ArgLocs[i];
959 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
961 // Promote the value if needed.
962 switch (VA.getLocInfo()) {
963 default: assert(0 && "Unknown loc info!");
964 case CCValAssign::Full: break;
965 case CCValAssign::SExt:
966 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
968 case CCValAssign::ZExt:
969 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
971 case CCValAssign::AExt:
972 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
977 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
979 assert(VA.isMemLoc());
980 if (StackPtr.Val == 0)
981 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
982 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
983 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
984 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
988 if (!MemOpChains.empty())
989 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
990 &MemOpChains[0], MemOpChains.size());
992 // Build a sequence of copy-to-reg nodes chained together with token chain
993 // and flag operands which copy the outgoing args into registers.
995 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
996 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
998 InFlag = Chain.getValue(1);
1001 // If the callee is a GlobalAddress node (quite common, every direct call is)
1002 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1003 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1004 // We should use extra load for direct calls to dllimported functions in
1006 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1007 getTargetMachine(), true))
1008 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1009 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1010 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1012 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1014 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1015 Subtarget->isPICStyleGOT()) {
1016 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1017 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1019 InFlag = Chain.getValue(1);
1022 // Returns a chain & a flag for retval copy to use.
1023 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1024 SmallVector<SDOperand, 8> Ops;
1025 Ops.push_back(Chain);
1026 Ops.push_back(Callee);
1028 // Add argument registers to the end of the list so that they are known live
1030 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1031 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1032 RegsToPass[i].second.getValueType()));
1034 // Add an implicit use GOT pointer in EBX.
1035 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1036 Subtarget->isPICStyleGOT())
1037 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1040 Ops.push_back(InFlag);
1042 // FIXME: Do not generate X86ISD::TAILCALL for now.
1043 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1044 NodeTys, &Ops[0], Ops.size());
1045 InFlag = Chain.getValue(1);
1047 // Returns a flag for retval copy to use.
1048 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1050 Ops.push_back(Chain);
1051 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1052 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1053 Ops.push_back(InFlag);
1054 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1055 InFlag = Chain.getValue(1);
1057 // Handle result values, copying them out of physregs into vregs that we
1059 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1063 //===----------------------------------------------------------------------===//
1064 // X86-64 C Calling Convention implementation
1065 //===----------------------------------------------------------------------===//
1068 X86TargetLowering::LowerX86_64CCCArguments(SDOperand Op, SelectionDAG &DAG) {
1069 MachineFunction &MF = DAG.getMachineFunction();
1070 MachineFrameInfo *MFI = MF.getFrameInfo();
1071 SDOperand Root = Op.getOperand(0);
1072 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1074 static const unsigned GPR64ArgRegs[] = {
1075 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1077 static const unsigned XMMArgRegs[] = {
1078 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1079 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1083 // Assign locations to all of the incoming arguments.
1084 SmallVector<CCValAssign, 16> ArgLocs;
1085 CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(),
1087 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_64_C);
1089 SmallVector<SDOperand, 8> ArgValues;
1090 unsigned LastVal = ~0U;
1091 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1092 CCValAssign &VA = ArgLocs[i];
1093 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1095 assert(VA.getValNo() != LastVal &&
1096 "Don't support value assigned to multiple locs yet");
1097 LastVal = VA.getValNo();
1099 if (VA.isRegLoc()) {
1100 MVT::ValueType RegVT = VA.getLocVT();
1101 TargetRegisterClass *RC;
1102 if (RegVT == MVT::i32)
1103 RC = X86::GR32RegisterClass;
1104 else if (RegVT == MVT::i64)
1105 RC = X86::GR64RegisterClass;
1106 else if (RegVT == MVT::f32)
1107 RC = X86::FR32RegisterClass;
1108 else if (RegVT == MVT::f64)
1109 RC = X86::FR64RegisterClass;
1111 assert(MVT::isVector(RegVT));
1112 RC = X86::VR128RegisterClass;
1115 SDOperand ArgValue = DAG.getCopyFromReg(Root, VA.getLocReg(), RegVT);
1116 AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1118 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1119 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1121 if (VA.getLocInfo() == CCValAssign::SExt)
1122 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1123 DAG.getValueType(VA.getValVT()));
1124 else if (VA.getLocInfo() == CCValAssign::ZExt)
1125 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1126 DAG.getValueType(VA.getValVT()));
1128 if (VA.getLocInfo() != CCValAssign::Full)
1129 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1131 ArgValues.push_back(ArgValue);
1133 assert(VA.isMemLoc());
1135 // Create the nodes corresponding to a load from this parameter slot.
1136 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
1137 VA.getLocMemOffset());
1138 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1139 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
1143 unsigned StackSize = CCInfo.getNextStackOffset();
1145 // If the function takes variable number of arguments, make a frame index for
1146 // the start of the first vararg value... for expansion of llvm.va_start.
1148 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs, 6);
1149 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1151 // For X86-64, if there are vararg parameters that are passed via
1152 // registers, then we must store them to their spots on the stack so they
1153 // may be loaded by deferencing the result of va_next.
1154 VarArgsGPOffset = NumIntRegs * 8;
1155 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1156 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1157 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1159 // Store the integer parameter registers.
1160 SmallVector<SDOperand, 8> MemOps;
1161 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1162 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1163 DAG.getConstant(VarArgsGPOffset, getPointerTy()));
1164 for (; NumIntRegs != 6; ++NumIntRegs) {
1165 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1166 X86::GR64RegisterClass);
1167 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1168 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1169 MemOps.push_back(Store);
1170 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1171 DAG.getConstant(8, getPointerTy()));
1174 // Now store the XMM (fp + vector) parameter registers.
1175 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1176 DAG.getConstant(VarArgsFPOffset, getPointerTy()));
1177 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1178 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1179 X86::VR128RegisterClass);
1180 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1181 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1182 MemOps.push_back(Store);
1183 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1184 DAG.getConstant(16, getPointerTy()));
1186 if (!MemOps.empty())
1187 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1188 &MemOps[0], MemOps.size());
1191 ArgValues.push_back(Root);
1193 ReturnAddrIndex = 0; // No return address slot generated yet.
1194 BytesToPopOnReturn = 0; // Callee pops nothing.
1195 BytesCallerReserves = StackSize;
1197 // Return the new list of results.
1198 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1199 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1203 X86TargetLowering::LowerX86_64CCCCallTo(SDOperand Op, SelectionDAG &DAG,
1205 SDOperand Chain = Op.getOperand(0);
1206 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1207 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1208 SDOperand Callee = Op.getOperand(4);
1210 // Analyze operands of the call, assigning locations to each operand.
1211 SmallVector<CCValAssign, 16> ArgLocs;
1212 CCState CCInfo(CC, getTargetMachine(), ArgLocs);
1213 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_C);
1215 // Get a count of how many bytes are to be pushed on the stack.
1216 unsigned NumBytes = CCInfo.getNextStackOffset();
1217 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1219 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1220 SmallVector<SDOperand, 8> MemOpChains;
1224 // Walk the register/memloc assignments, inserting copies/loads.
1225 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1226 CCValAssign &VA = ArgLocs[i];
1227 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1229 // Promote the value if needed.
1230 switch (VA.getLocInfo()) {
1231 default: assert(0 && "Unknown loc info!");
1232 case CCValAssign::Full: break;
1233 case CCValAssign::SExt:
1234 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1236 case CCValAssign::ZExt:
1237 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1239 case CCValAssign::AExt:
1240 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1244 if (VA.isRegLoc()) {
1245 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1247 assert(VA.isMemLoc());
1248 if (StackPtr.Val == 0)
1249 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1250 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
1251 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1252 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1256 if (!MemOpChains.empty())
1257 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1258 &MemOpChains[0], MemOpChains.size());
1260 // Build a sequence of copy-to-reg nodes chained together with token chain
1261 // and flag operands which copy the outgoing args into registers.
1263 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1264 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1266 InFlag = Chain.getValue(1);
1270 // From AMD64 ABI document:
1271 // For calls that may call functions that use varargs or stdargs
1272 // (prototype-less calls or calls to functions containing ellipsis (...) in
1273 // the declaration) %al is used as hidden argument to specify the number
1274 // of SSE registers used. The contents of %al do not need to match exactly
1275 // the number of registers, but must be an ubound on the number of SSE
1276 // registers used and is in the range 0 - 8 inclusive.
1278 // Count the number of XMM registers allocated.
1279 static const unsigned XMMArgRegs[] = {
1280 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1281 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1283 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1285 Chain = DAG.getCopyToReg(Chain, X86::AL,
1286 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1287 InFlag = Chain.getValue(1);
1290 // If the callee is a GlobalAddress node (quite common, every direct call is)
1291 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1292 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1293 // We should use extra load for direct calls to dllimported functions in
1295 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1296 getTargetMachine(), true))
1297 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1298 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1299 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1301 // Returns a chain & a flag for retval copy to use.
1302 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1303 SmallVector<SDOperand, 8> Ops;
1304 Ops.push_back(Chain);
1305 Ops.push_back(Callee);
1307 // Add argument registers to the end of the list so that they are known live
1309 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1310 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1311 RegsToPass[i].second.getValueType()));
1314 Ops.push_back(InFlag);
1316 // FIXME: Do not generate X86ISD::TAILCALL for now.
1317 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1318 NodeTys, &Ops[0], Ops.size());
1319 InFlag = Chain.getValue(1);
1321 // Returns a flag for retval copy to use.
1322 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1324 Ops.push_back(Chain);
1325 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1326 Ops.push_back(DAG.getConstant(0, getPointerTy()));
1327 Ops.push_back(InFlag);
1328 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1329 InFlag = Chain.getValue(1);
1331 // Handle result values, copying them out of physregs into vregs that we
1333 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1337 //===----------------------------------------------------------------------===//
1338 // Other Lowering Hooks
1339 //===----------------------------------------------------------------------===//
1342 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1343 if (ReturnAddrIndex == 0) {
1344 // Set up a frame object for the return address.
1345 MachineFunction &MF = DAG.getMachineFunction();
1346 if (Subtarget->is64Bit())
1347 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
1349 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
1352 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1357 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1358 /// specific condition code. It returns a false if it cannot do a direct
1359 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1361 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1362 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
1363 SelectionDAG &DAG) {
1364 X86CC = X86::COND_INVALID;
1366 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1367 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1368 // X > -1 -> X == 0, jump !sign.
1369 RHS = DAG.getConstant(0, RHS.getValueType());
1370 X86CC = X86::COND_NS;
1372 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1373 // X < 0 -> X == 0, jump on sign.
1374 X86CC = X86::COND_S;
1379 switch (SetCCOpcode) {
1381 case ISD::SETEQ: X86CC = X86::COND_E; break;
1382 case ISD::SETGT: X86CC = X86::COND_G; break;
1383 case ISD::SETGE: X86CC = X86::COND_GE; break;
1384 case ISD::SETLT: X86CC = X86::COND_L; break;
1385 case ISD::SETLE: X86CC = X86::COND_LE; break;
1386 case ISD::SETNE: X86CC = X86::COND_NE; break;
1387 case ISD::SETULT: X86CC = X86::COND_B; break;
1388 case ISD::SETUGT: X86CC = X86::COND_A; break;
1389 case ISD::SETULE: X86CC = X86::COND_BE; break;
1390 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1393 // On a floating point condition, the flags are set as follows:
1395 // 0 | 0 | 0 | X > Y
1396 // 0 | 0 | 1 | X < Y
1397 // 1 | 0 | 0 | X == Y
1398 // 1 | 1 | 1 | unordered
1400 switch (SetCCOpcode) {
1403 case ISD::SETEQ: X86CC = X86::COND_E; break;
1404 case ISD::SETOLT: Flip = true; // Fallthrough
1406 case ISD::SETGT: X86CC = X86::COND_A; break;
1407 case ISD::SETOLE: Flip = true; // Fallthrough
1409 case ISD::SETGE: X86CC = X86::COND_AE; break;
1410 case ISD::SETUGT: Flip = true; // Fallthrough
1412 case ISD::SETLT: X86CC = X86::COND_B; break;
1413 case ISD::SETUGE: Flip = true; // Fallthrough
1415 case ISD::SETLE: X86CC = X86::COND_BE; break;
1417 case ISD::SETNE: X86CC = X86::COND_NE; break;
1418 case ISD::SETUO: X86CC = X86::COND_P; break;
1419 case ISD::SETO: X86CC = X86::COND_NP; break;
1422 std::swap(LHS, RHS);
1425 return X86CC != X86::COND_INVALID;
1428 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
1429 /// code. Current x86 isa includes the following FP cmov instructions:
1430 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
1431 static bool hasFPCMov(unsigned X86CC) {
1447 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
1448 /// true if Op is undef or if its value falls within the specified range (L, H].
1449 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
1450 if (Op.getOpcode() == ISD::UNDEF)
1453 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
1454 return (Val >= Low && Val < Hi);
1457 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
1458 /// true if Op is undef or if its value equal to the specified value.
1459 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
1460 if (Op.getOpcode() == ISD::UNDEF)
1462 return cast<ConstantSDNode>(Op)->getValue() == Val;
1465 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
1466 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
1467 bool X86::isPSHUFDMask(SDNode *N) {
1468 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1470 if (N->getNumOperands() != 4)
1473 // Check if the value doesn't reference the second vector.
1474 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1475 SDOperand Arg = N->getOperand(i);
1476 if (Arg.getOpcode() == ISD::UNDEF) continue;
1477 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1478 if (cast<ConstantSDNode>(Arg)->getValue() >= 4)
1485 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
1486 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
1487 bool X86::isPSHUFHWMask(SDNode *N) {
1488 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1490 if (N->getNumOperands() != 8)
1493 // Lower quadword copied in order.
1494 for (unsigned i = 0; i != 4; ++i) {
1495 SDOperand Arg = N->getOperand(i);
1496 if (Arg.getOpcode() == ISD::UNDEF) continue;
1497 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1498 if (cast<ConstantSDNode>(Arg)->getValue() != i)
1502 // Upper quadword shuffled.
1503 for (unsigned i = 4; i != 8; ++i) {
1504 SDOperand Arg = N->getOperand(i);
1505 if (Arg.getOpcode() == ISD::UNDEF) continue;
1506 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1507 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1508 if (Val < 4 || Val > 7)
1515 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
1516 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
1517 bool X86::isPSHUFLWMask(SDNode *N) {
1518 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1520 if (N->getNumOperands() != 8)
1523 // Upper quadword copied in order.
1524 for (unsigned i = 4; i != 8; ++i)
1525 if (!isUndefOrEqual(N->getOperand(i), i))
1528 // Lower quadword shuffled.
1529 for (unsigned i = 0; i != 4; ++i)
1530 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
1536 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
1537 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
1538 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
1539 if (NumElems != 2 && NumElems != 4) return false;
1541 unsigned Half = NumElems / 2;
1542 for (unsigned i = 0; i < Half; ++i)
1543 if (!isUndefOrInRange(Elems[i], 0, NumElems))
1545 for (unsigned i = Half; i < NumElems; ++i)
1546 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
1552 bool X86::isSHUFPMask(SDNode *N) {
1553 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1554 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
1557 /// isCommutedSHUFP - Returns true if the shuffle mask is except
1558 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
1559 /// half elements to come from vector 1 (which would equal the dest.) and
1560 /// the upper half to come from vector 2.
1561 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
1562 if (NumOps != 2 && NumOps != 4) return false;
1564 unsigned Half = NumOps / 2;
1565 for (unsigned i = 0; i < Half; ++i)
1566 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
1568 for (unsigned i = Half; i < NumOps; ++i)
1569 if (!isUndefOrInRange(Ops[i], 0, NumOps))
1574 static bool isCommutedSHUFP(SDNode *N) {
1575 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1576 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
1579 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
1580 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
1581 bool X86::isMOVHLPSMask(SDNode *N) {
1582 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1584 if (N->getNumOperands() != 4)
1587 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
1588 return isUndefOrEqual(N->getOperand(0), 6) &&
1589 isUndefOrEqual(N->getOperand(1), 7) &&
1590 isUndefOrEqual(N->getOperand(2), 2) &&
1591 isUndefOrEqual(N->getOperand(3), 3);
1594 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
1595 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
1597 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
1598 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1600 if (N->getNumOperands() != 4)
1603 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
1604 return isUndefOrEqual(N->getOperand(0), 2) &&
1605 isUndefOrEqual(N->getOperand(1), 3) &&
1606 isUndefOrEqual(N->getOperand(2), 2) &&
1607 isUndefOrEqual(N->getOperand(3), 3);
1610 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
1611 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
1612 bool X86::isMOVLPMask(SDNode *N) {
1613 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1615 unsigned NumElems = N->getNumOperands();
1616 if (NumElems != 2 && NumElems != 4)
1619 for (unsigned i = 0; i < NumElems/2; ++i)
1620 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
1623 for (unsigned i = NumElems/2; i < NumElems; ++i)
1624 if (!isUndefOrEqual(N->getOperand(i), i))
1630 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
1631 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
1633 bool X86::isMOVHPMask(SDNode *N) {
1634 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1636 unsigned NumElems = N->getNumOperands();
1637 if (NumElems != 2 && NumElems != 4)
1640 for (unsigned i = 0; i < NumElems/2; ++i)
1641 if (!isUndefOrEqual(N->getOperand(i), i))
1644 for (unsigned i = 0; i < NumElems/2; ++i) {
1645 SDOperand Arg = N->getOperand(i + NumElems/2);
1646 if (!isUndefOrEqual(Arg, i + NumElems))
1653 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
1654 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
1655 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
1656 bool V2IsSplat = false) {
1657 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1660 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
1661 SDOperand BitI = Elts[i];
1662 SDOperand BitI1 = Elts[i+1];
1663 if (!isUndefOrEqual(BitI, j))
1666 if (isUndefOrEqual(BitI1, NumElts))
1669 if (!isUndefOrEqual(BitI1, j + NumElts))
1677 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
1678 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1679 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
1682 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
1683 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
1684 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
1685 bool V2IsSplat = false) {
1686 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1689 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
1690 SDOperand BitI = Elts[i];
1691 SDOperand BitI1 = Elts[i+1];
1692 if (!isUndefOrEqual(BitI, j + NumElts/2))
1695 if (isUndefOrEqual(BitI1, NumElts))
1698 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
1706 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
1707 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1708 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
1711 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
1712 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
1714 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
1715 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1717 unsigned NumElems = N->getNumOperands();
1718 if (NumElems != 4 && NumElems != 8 && NumElems != 16)
1721 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
1722 SDOperand BitI = N->getOperand(i);
1723 SDOperand BitI1 = N->getOperand(i+1);
1725 if (!isUndefOrEqual(BitI, j))
1727 if (!isUndefOrEqual(BitI1, j))
1734 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
1735 /// specifies a shuffle of elements that is suitable for input to MOVSS,
1736 /// MOVSD, and MOVD, i.e. setting the lowest element.
1737 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
1738 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1741 if (!isUndefOrEqual(Elts[0], NumElts))
1744 for (unsigned i = 1; i < NumElts; ++i) {
1745 if (!isUndefOrEqual(Elts[i], i))
1752 bool X86::isMOVLMask(SDNode *N) {
1753 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1754 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
1757 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
1758 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
1759 /// element of vector 2 and the other elements to come from vector 1 in order.
1760 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
1761 bool V2IsSplat = false,
1762 bool V2IsUndef = false) {
1763 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
1766 if (!isUndefOrEqual(Ops[0], 0))
1769 for (unsigned i = 1; i < NumOps; ++i) {
1770 SDOperand Arg = Ops[i];
1771 if (!(isUndefOrEqual(Arg, i+NumOps) ||
1772 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
1773 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
1780 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
1781 bool V2IsUndef = false) {
1782 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1783 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
1784 V2IsSplat, V2IsUndef);
1787 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
1788 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
1789 bool X86::isMOVSHDUPMask(SDNode *N) {
1790 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1792 if (N->getNumOperands() != 4)
1795 // Expect 1, 1, 3, 3
1796 for (unsigned i = 0; i < 2; ++i) {
1797 SDOperand Arg = N->getOperand(i);
1798 if (Arg.getOpcode() == ISD::UNDEF) continue;
1799 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1800 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1801 if (Val != 1) return false;
1805 for (unsigned i = 2; i < 4; ++i) {
1806 SDOperand Arg = N->getOperand(i);
1807 if (Arg.getOpcode() == ISD::UNDEF) continue;
1808 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1809 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1810 if (Val != 3) return false;
1814 // Don't use movshdup if it can be done with a shufps.
1818 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
1819 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
1820 bool X86::isMOVSLDUPMask(SDNode *N) {
1821 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1823 if (N->getNumOperands() != 4)
1826 // Expect 0, 0, 2, 2
1827 for (unsigned i = 0; i < 2; ++i) {
1828 SDOperand Arg = N->getOperand(i);
1829 if (Arg.getOpcode() == ISD::UNDEF) continue;
1830 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1831 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1832 if (Val != 0) return false;
1836 for (unsigned i = 2; i < 4; ++i) {
1837 SDOperand Arg = N->getOperand(i);
1838 if (Arg.getOpcode() == ISD::UNDEF) continue;
1839 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1840 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1841 if (Val != 2) return false;
1845 // Don't use movshdup if it can be done with a shufps.
1849 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
1850 /// a splat of a single element.
1851 static bool isSplatMask(SDNode *N) {
1852 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1854 // This is a splat operation if each element of the permute is the same, and
1855 // if the value doesn't reference the second vector.
1856 unsigned NumElems = N->getNumOperands();
1857 SDOperand ElementBase;
1859 for (; i != NumElems; ++i) {
1860 SDOperand Elt = N->getOperand(i);
1861 if (isa<ConstantSDNode>(Elt)) {
1867 if (!ElementBase.Val)
1870 for (; i != NumElems; ++i) {
1871 SDOperand Arg = N->getOperand(i);
1872 if (Arg.getOpcode() == ISD::UNDEF) continue;
1873 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1874 if (Arg != ElementBase) return false;
1877 // Make sure it is a splat of the first vector operand.
1878 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
1881 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
1882 /// a splat of a single element and it's a 2 or 4 element mask.
1883 bool X86::isSplatMask(SDNode *N) {
1884 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1886 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
1887 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
1889 return ::isSplatMask(N);
1892 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
1893 /// specifies a splat of zero element.
1894 bool X86::isSplatLoMask(SDNode *N) {
1895 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1897 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
1898 if (!isUndefOrEqual(N->getOperand(i), 0))
1903 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
1904 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
1906 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
1907 unsigned NumOperands = N->getNumOperands();
1908 unsigned Shift = (NumOperands == 4) ? 2 : 1;
1910 for (unsigned i = 0; i < NumOperands; ++i) {
1912 SDOperand Arg = N->getOperand(NumOperands-i-1);
1913 if (Arg.getOpcode() != ISD::UNDEF)
1914 Val = cast<ConstantSDNode>(Arg)->getValue();
1915 if (Val >= NumOperands) Val -= NumOperands;
1917 if (i != NumOperands - 1)
1924 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
1925 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
1927 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
1929 // 8 nodes, but we only care about the last 4.
1930 for (unsigned i = 7; i >= 4; --i) {
1932 SDOperand Arg = N->getOperand(i);
1933 if (Arg.getOpcode() != ISD::UNDEF)
1934 Val = cast<ConstantSDNode>(Arg)->getValue();
1943 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
1944 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
1946 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
1948 // 8 nodes, but we only care about the first 4.
1949 for (int i = 3; i >= 0; --i) {
1951 SDOperand Arg = N->getOperand(i);
1952 if (Arg.getOpcode() != ISD::UNDEF)
1953 Val = cast<ConstantSDNode>(Arg)->getValue();
1962 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
1963 /// specifies a 8 element shuffle that can be broken into a pair of
1964 /// PSHUFHW and PSHUFLW.
1965 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
1966 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1968 if (N->getNumOperands() != 8)
1971 // Lower quadword shuffled.
1972 for (unsigned i = 0; i != 4; ++i) {
1973 SDOperand Arg = N->getOperand(i);
1974 if (Arg.getOpcode() == ISD::UNDEF) continue;
1975 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1976 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1981 // Upper quadword shuffled.
1982 for (unsigned i = 4; i != 8; ++i) {
1983 SDOperand Arg = N->getOperand(i);
1984 if (Arg.getOpcode() == ISD::UNDEF) continue;
1985 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1986 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1987 if (Val < 4 || Val > 7)
1994 /// CommuteVectorShuffle - Swap vector_shuffle operandsas well as
1995 /// values in ther permute mask.
1996 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
1997 SDOperand &V2, SDOperand &Mask,
1998 SelectionDAG &DAG) {
1999 MVT::ValueType VT = Op.getValueType();
2000 MVT::ValueType MaskVT = Mask.getValueType();
2001 MVT::ValueType EltVT = MVT::getVectorBaseType(MaskVT);
2002 unsigned NumElems = Mask.getNumOperands();
2003 SmallVector<SDOperand, 8> MaskVec;
2005 for (unsigned i = 0; i != NumElems; ++i) {
2006 SDOperand Arg = Mask.getOperand(i);
2007 if (Arg.getOpcode() == ISD::UNDEF) {
2008 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2011 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2012 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2014 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2016 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2020 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2021 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2024 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2025 /// match movhlps. The lower half elements should come from upper half of
2026 /// V1 (and in order), and the upper half elements should come from the upper
2027 /// half of V2 (and in order).
2028 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2029 unsigned NumElems = Mask->getNumOperands();
2032 for (unsigned i = 0, e = 2; i != e; ++i)
2033 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2035 for (unsigned i = 2; i != 4; ++i)
2036 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2041 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2042 /// is promoted to a vector.
2043 static inline bool isScalarLoadToVector(SDNode *N) {
2044 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2045 N = N->getOperand(0).Val;
2046 return ISD::isNON_EXTLoad(N);
2051 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2052 /// match movlp{s|d}. The lower half elements should come from lower half of
2053 /// V1 (and in order), and the upper half elements should come from the upper
2054 /// half of V2 (and in order). And since V1 will become the source of the
2055 /// MOVLP, it must be either a vector load or a scalar load to vector.
2056 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2057 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2059 // Is V2 is a vector load, don't do this transformation. We will try to use
2060 // load folding shufps op.
2061 if (ISD::isNON_EXTLoad(V2))
2064 unsigned NumElems = Mask->getNumOperands();
2065 if (NumElems != 2 && NumElems != 4)
2067 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2068 if (!isUndefOrEqual(Mask->getOperand(i), i))
2070 for (unsigned i = NumElems/2; i != NumElems; ++i)
2071 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2076 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2078 static bool isSplatVector(SDNode *N) {
2079 if (N->getOpcode() != ISD::BUILD_VECTOR)
2082 SDOperand SplatValue = N->getOperand(0);
2083 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2084 if (N->getOperand(i) != SplatValue)
2089 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2091 static bool isUndefShuffle(SDNode *N) {
2092 if (N->getOpcode() != ISD::BUILD_VECTOR)
2095 SDOperand V1 = N->getOperand(0);
2096 SDOperand V2 = N->getOperand(1);
2097 SDOperand Mask = N->getOperand(2);
2098 unsigned NumElems = Mask.getNumOperands();
2099 for (unsigned i = 0; i != NumElems; ++i) {
2100 SDOperand Arg = Mask.getOperand(i);
2101 if (Arg.getOpcode() != ISD::UNDEF) {
2102 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2103 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2105 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2112 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2113 /// that point to V2 points to its first element.
2114 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2115 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2117 bool Changed = false;
2118 SmallVector<SDOperand, 8> MaskVec;
2119 unsigned NumElems = Mask.getNumOperands();
2120 for (unsigned i = 0; i != NumElems; ++i) {
2121 SDOperand Arg = Mask.getOperand(i);
2122 if (Arg.getOpcode() != ISD::UNDEF) {
2123 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2124 if (Val > NumElems) {
2125 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2129 MaskVec.push_back(Arg);
2133 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2134 &MaskVec[0], MaskVec.size());
2138 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2139 /// operation of specified width.
2140 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2141 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2142 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2144 SmallVector<SDOperand, 8> MaskVec;
2145 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2146 for (unsigned i = 1; i != NumElems; ++i)
2147 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2148 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2151 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2152 /// of specified width.
2153 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2154 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2155 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2156 SmallVector<SDOperand, 8> MaskVec;
2157 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2158 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2159 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2161 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2164 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2165 /// of specified width.
2166 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2167 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2168 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2169 unsigned Half = NumElems/2;
2170 SmallVector<SDOperand, 8> MaskVec;
2171 for (unsigned i = 0; i != Half; ++i) {
2172 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2173 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2175 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2178 /// getZeroVector - Returns a vector of specified type with all zero elements.
2180 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2181 assert(MVT::isVector(VT) && "Expected a vector type");
2182 unsigned NumElems = getVectorNumElements(VT);
2183 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2184 bool isFP = MVT::isFloatingPoint(EVT);
2185 SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT);
2186 SmallVector<SDOperand, 8> ZeroVec(NumElems, Zero);
2187 return DAG.getNode(ISD::BUILD_VECTOR, VT, &ZeroVec[0], ZeroVec.size());
2190 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2192 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2193 SDOperand V1 = Op.getOperand(0);
2194 SDOperand Mask = Op.getOperand(2);
2195 MVT::ValueType VT = Op.getValueType();
2196 unsigned NumElems = Mask.getNumOperands();
2197 Mask = getUnpacklMask(NumElems, DAG);
2198 while (NumElems != 4) {
2199 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2202 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2204 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2205 Mask = getZeroVector(MaskVT, DAG);
2206 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2207 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2208 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2211 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2213 static inline bool isZeroNode(SDOperand Elt) {
2214 return ((isa<ConstantSDNode>(Elt) &&
2215 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2216 (isa<ConstantFPSDNode>(Elt) &&
2217 cast<ConstantFPSDNode>(Elt)->isExactlyValue(0.0)));
2220 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2221 /// vector and zero or undef vector.
2222 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2223 unsigned NumElems, unsigned Idx,
2224 bool isZero, SelectionDAG &DAG) {
2225 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2226 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2227 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2228 SDOperand Zero = DAG.getConstant(0, EVT);
2229 SmallVector<SDOperand, 8> MaskVec(NumElems, Zero);
2230 MaskVec[Idx] = DAG.getConstant(NumElems, EVT);
2231 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2232 &MaskVec[0], MaskVec.size());
2233 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2236 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2238 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
2239 unsigned NumNonZero, unsigned NumZero,
2240 SelectionDAG &DAG, TargetLowering &TLI) {
2246 for (unsigned i = 0; i < 16; ++i) {
2247 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
2248 if (ThisIsNonZero && First) {
2250 V = getZeroVector(MVT::v8i16, DAG);
2252 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2257 SDOperand ThisElt(0, 0), LastElt(0, 0);
2258 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
2259 if (LastIsNonZero) {
2260 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
2262 if (ThisIsNonZero) {
2263 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
2264 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
2265 ThisElt, DAG.getConstant(8, MVT::i8));
2267 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
2272 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
2273 DAG.getConstant(i/2, TLI.getPointerTy()));
2277 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
2280 /// LowerBuildVectorv16i8 - Custom lower build_vector of v8i16.
2282 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
2283 unsigned NumNonZero, unsigned NumZero,
2284 SelectionDAG &DAG, TargetLowering &TLI) {
2290 for (unsigned i = 0; i < 8; ++i) {
2291 bool isNonZero = (NonZeros & (1 << i)) != 0;
2295 V = getZeroVector(MVT::v8i16, DAG);
2297 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2300 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
2301 DAG.getConstant(i, TLI.getPointerTy()));
2309 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2310 // All zero's are handled with pxor.
2311 if (ISD::isBuildVectorAllZeros(Op.Val))
2314 // All one's are handled with pcmpeqd.
2315 if (ISD::isBuildVectorAllOnes(Op.Val))
2318 MVT::ValueType VT = Op.getValueType();
2319 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2320 unsigned EVTBits = MVT::getSizeInBits(EVT);
2322 unsigned NumElems = Op.getNumOperands();
2323 unsigned NumZero = 0;
2324 unsigned NumNonZero = 0;
2325 unsigned NonZeros = 0;
2326 std::set<SDOperand> Values;
2327 for (unsigned i = 0; i < NumElems; ++i) {
2328 SDOperand Elt = Op.getOperand(i);
2329 if (Elt.getOpcode() != ISD::UNDEF) {
2331 if (isZeroNode(Elt))
2334 NonZeros |= (1 << i);
2340 if (NumNonZero == 0)
2341 // Must be a mix of zero and undef. Return a zero vector.
2342 return getZeroVector(VT, DAG);
2344 // Splat is obviously ok. Let legalizer expand it to a shuffle.
2345 if (Values.size() == 1)
2348 // Special case for single non-zero element.
2349 if (NumNonZero == 1) {
2350 unsigned Idx = CountTrailingZeros_32(NonZeros);
2351 SDOperand Item = Op.getOperand(Idx);
2352 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
2354 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
2355 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
2358 if (EVTBits == 32) {
2359 // Turn it into a shuffle of zero and zero-extended scalar to vector.
2360 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
2362 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2363 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
2364 SmallVector<SDOperand, 8> MaskVec;
2365 for (unsigned i = 0; i < NumElems; i++)
2366 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
2367 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2368 &MaskVec[0], MaskVec.size());
2369 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
2370 DAG.getNode(ISD::UNDEF, VT), Mask);
2374 // Let legalizer expand 2-wide build_vector's.
2378 // If element VT is < 32 bits, convert it to inserts into a zero vector.
2380 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
2382 if (V.Val) return V;
2385 if (EVTBits == 16) {
2386 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
2388 if (V.Val) return V;
2391 // If element VT is == 32 bits, turn it into a number of shuffles.
2392 SmallVector<SDOperand, 8> V;
2394 if (NumElems == 4 && NumZero > 0) {
2395 for (unsigned i = 0; i < 4; ++i) {
2396 bool isZero = !(NonZeros & (1 << i));
2398 V[i] = getZeroVector(VT, DAG);
2400 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2403 for (unsigned i = 0; i < 2; ++i) {
2404 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
2407 V[i] = V[i*2]; // Must be a zero vector.
2410 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
2411 getMOVLMask(NumElems, DAG));
2414 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2415 getMOVLMask(NumElems, DAG));
2418 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2419 getUnpacklMask(NumElems, DAG));
2424 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
2425 // clears the upper bits.
2426 // FIXME: we can do the same for v4f32 case when we know both parts of
2427 // the lower half come from scalar_to_vector (loadf32). We should do
2428 // that in post legalizer dag combiner with target specific hooks.
2429 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
2431 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2432 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2433 SmallVector<SDOperand, 8> MaskVec;
2434 bool Reverse = (NonZeros & 0x3) == 2;
2435 for (unsigned i = 0; i < 2; ++i)
2437 MaskVec.push_back(DAG.getConstant(1-i, EVT));
2439 MaskVec.push_back(DAG.getConstant(i, EVT));
2440 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
2441 for (unsigned i = 0; i < 2; ++i)
2443 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
2445 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
2446 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2447 &MaskVec[0], MaskVec.size());
2448 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
2451 if (Values.size() > 2) {
2452 // Expand into a number of unpckl*.
2454 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
2455 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
2456 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
2457 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
2458 for (unsigned i = 0; i < NumElems; ++i)
2459 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2461 while (NumElems != 0) {
2462 for (unsigned i = 0; i < NumElems; ++i)
2463 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
2474 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
2475 SDOperand V1 = Op.getOperand(0);
2476 SDOperand V2 = Op.getOperand(1);
2477 SDOperand PermMask = Op.getOperand(2);
2478 MVT::ValueType VT = Op.getValueType();
2479 unsigned NumElems = PermMask.getNumOperands();
2480 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
2481 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
2482 bool V1IsSplat = false;
2483 bool V2IsSplat = false;
2485 if (isUndefShuffle(Op.Val))
2486 return DAG.getNode(ISD::UNDEF, VT);
2488 if (isSplatMask(PermMask.Val)) {
2489 if (NumElems <= 4) return Op;
2490 // Promote it to a v4i32 splat.
2491 return PromoteSplat(Op, DAG);
2494 if (X86::isMOVLMask(PermMask.Val))
2495 return (V1IsUndef) ? V2 : Op;
2497 if (X86::isMOVSHDUPMask(PermMask.Val) ||
2498 X86::isMOVSLDUPMask(PermMask.Val) ||
2499 X86::isMOVHLPSMask(PermMask.Val) ||
2500 X86::isMOVHPMask(PermMask.Val) ||
2501 X86::isMOVLPMask(PermMask.Val))
2504 if (ShouldXformToMOVHLPS(PermMask.Val) ||
2505 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
2506 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2508 bool Commuted = false;
2509 V1IsSplat = isSplatVector(V1.Val);
2510 V2IsSplat = isSplatVector(V2.Val);
2511 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
2512 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2513 std::swap(V1IsSplat, V2IsSplat);
2514 std::swap(V1IsUndef, V2IsUndef);
2518 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
2519 if (V2IsUndef) return V1;
2520 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2522 // V2 is a splat, so the mask may be malformed. That is, it may point
2523 // to any V2 element. The instruction selectior won't like this. Get
2524 // a corrected mask and commute to form a proper MOVS{S|D}.
2525 SDOperand NewMask = getMOVLMask(NumElems, DAG);
2526 if (NewMask.Val != PermMask.Val)
2527 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2532 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
2533 X86::isUNPCKLMask(PermMask.Val) ||
2534 X86::isUNPCKHMask(PermMask.Val))
2538 // Normalize mask so all entries that point to V2 points to its first
2539 // element then try to match unpck{h|l} again. If match, return a
2540 // new vector_shuffle with the corrected mask.
2541 SDOperand NewMask = NormalizeMask(PermMask, DAG);
2542 if (NewMask.Val != PermMask.Val) {
2543 if (X86::isUNPCKLMask(PermMask.Val, true)) {
2544 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
2545 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2546 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
2547 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
2548 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2553 // Normalize the node to match x86 shuffle ops if needed
2554 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
2555 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2558 // Commute is back and try unpck* again.
2559 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2560 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
2561 X86::isUNPCKLMask(PermMask.Val) ||
2562 X86::isUNPCKHMask(PermMask.Val))
2566 // If VT is integer, try PSHUF* first, then SHUFP*.
2567 if (MVT::isInteger(VT)) {
2568 if (X86::isPSHUFDMask(PermMask.Val) ||
2569 X86::isPSHUFHWMask(PermMask.Val) ||
2570 X86::isPSHUFLWMask(PermMask.Val)) {
2571 if (V2.getOpcode() != ISD::UNDEF)
2572 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
2573 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
2577 if (X86::isSHUFPMask(PermMask.Val))
2580 // Handle v8i16 shuffle high / low shuffle node pair.
2581 if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
2582 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2583 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2584 SmallVector<SDOperand, 8> MaskVec;
2585 for (unsigned i = 0; i != 4; ++i)
2586 MaskVec.push_back(PermMask.getOperand(i));
2587 for (unsigned i = 4; i != 8; ++i)
2588 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2589 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2590 &MaskVec[0], MaskVec.size());
2591 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2593 for (unsigned i = 0; i != 4; ++i)
2594 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2595 for (unsigned i = 4; i != 8; ++i)
2596 MaskVec.push_back(PermMask.getOperand(i));
2597 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0],MaskVec.size());
2598 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2601 // Floating point cases in the other order.
2602 if (X86::isSHUFPMask(PermMask.Val))
2604 if (X86::isPSHUFDMask(PermMask.Val) ||
2605 X86::isPSHUFHWMask(PermMask.Val) ||
2606 X86::isPSHUFLWMask(PermMask.Val)) {
2607 if (V2.getOpcode() != ISD::UNDEF)
2608 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
2609 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
2614 if (NumElems == 4) {
2615 MVT::ValueType MaskVT = PermMask.getValueType();
2616 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
2617 SmallVector<std::pair<int, int>, 8> Locs;
2618 Locs.reserve(NumElems);
2619 SmallVector<SDOperand, 8> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2620 SmallVector<SDOperand, 8> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2623 // If no more than two elements come from either vector. This can be
2624 // implemented with two shuffles. First shuffle gather the elements.
2625 // The second shuffle, which takes the first shuffle as both of its
2626 // vector operands, put the elements into the right order.
2627 for (unsigned i = 0; i != NumElems; ++i) {
2628 SDOperand Elt = PermMask.getOperand(i);
2629 if (Elt.getOpcode() == ISD::UNDEF) {
2630 Locs[i] = std::make_pair(-1, -1);
2632 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
2633 if (Val < NumElems) {
2634 Locs[i] = std::make_pair(0, NumLo);
2638 Locs[i] = std::make_pair(1, NumHi);
2639 if (2+NumHi < NumElems)
2640 Mask1[2+NumHi] = Elt;
2645 if (NumLo <= 2 && NumHi <= 2) {
2646 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2647 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2648 &Mask1[0], Mask1.size()));
2649 for (unsigned i = 0; i != NumElems; ++i) {
2650 if (Locs[i].first == -1)
2653 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
2654 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
2655 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
2659 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
2660 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2661 &Mask2[0], Mask2.size()));
2664 // Break it into (shuffle shuffle_hi, shuffle_lo).
2666 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2667 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2668 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
2669 unsigned MaskIdx = 0;
2671 unsigned HiIdx = NumElems/2;
2672 for (unsigned i = 0; i != NumElems; ++i) {
2673 if (i == NumElems/2) {
2679 SDOperand Elt = PermMask.getOperand(i);
2680 if (Elt.getOpcode() == ISD::UNDEF) {
2681 Locs[i] = std::make_pair(-1, -1);
2682 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
2683 Locs[i] = std::make_pair(MaskIdx, LoIdx);
2684 (*MaskPtr)[LoIdx] = Elt;
2687 Locs[i] = std::make_pair(MaskIdx, HiIdx);
2688 (*MaskPtr)[HiIdx] = Elt;
2693 SDOperand LoShuffle =
2694 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2695 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2696 &LoMask[0], LoMask.size()));
2697 SDOperand HiShuffle =
2698 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2699 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2700 &HiMask[0], HiMask.size()));
2701 SmallVector<SDOperand, 8> MaskOps;
2702 for (unsigned i = 0; i != NumElems; ++i) {
2703 if (Locs[i].first == -1) {
2704 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
2706 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
2707 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
2710 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
2711 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2712 &MaskOps[0], MaskOps.size()));
2719 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2720 if (!isa<ConstantSDNode>(Op.getOperand(1)))
2723 MVT::ValueType VT = Op.getValueType();
2724 // TODO: handle v16i8.
2725 if (MVT::getSizeInBits(VT) == 16) {
2726 // Transform it so it match pextrw which produces a 32-bit result.
2727 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
2728 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
2729 Op.getOperand(0), Op.getOperand(1));
2730 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
2731 DAG.getValueType(VT));
2732 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
2733 } else if (MVT::getSizeInBits(VT) == 32) {
2734 SDOperand Vec = Op.getOperand(0);
2735 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
2738 // SHUFPS the element to the lowest double word, then movss.
2739 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2740 SmallVector<SDOperand, 8> IdxVec;
2741 IdxVec.push_back(DAG.getConstant(Idx, 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 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2745 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2746 &IdxVec[0], IdxVec.size());
2747 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
2748 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
2749 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
2750 DAG.getConstant(0, getPointerTy()));
2751 } else if (MVT::getSizeInBits(VT) == 64) {
2752 SDOperand Vec = Op.getOperand(0);
2753 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
2757 // UNPCKHPD the element to the lowest double word, then movsd.
2758 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
2759 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
2760 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2761 SmallVector<SDOperand, 8> IdxVec;
2762 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorBaseType(MaskVT)));
2763 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2764 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2765 &IdxVec[0], IdxVec.size());
2766 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
2767 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
2768 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
2769 DAG.getConstant(0, getPointerTy()));
2776 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2777 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
2778 // as its second argument.
2779 MVT::ValueType VT = Op.getValueType();
2780 MVT::ValueType BaseVT = MVT::getVectorBaseType(VT);
2781 SDOperand N0 = Op.getOperand(0);
2782 SDOperand N1 = Op.getOperand(1);
2783 SDOperand N2 = Op.getOperand(2);
2784 if (MVT::getSizeInBits(BaseVT) == 16) {
2785 if (N1.getValueType() != MVT::i32)
2786 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
2787 if (N2.getValueType() != MVT::i32)
2788 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(), MVT::i32);
2789 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
2790 } else if (MVT::getSizeInBits(BaseVT) == 32) {
2791 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
2794 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
2795 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2796 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2797 SmallVector<SDOperand, 8> MaskVec;
2798 MaskVec.push_back(DAG.getConstant(4, BaseVT));
2799 for (unsigned i = 1; i <= 3; ++i)
2800 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2801 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
2802 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2803 &MaskVec[0], MaskVec.size()));
2805 // Use two pinsrw instructions to insert a 32 bit value.
2807 if (MVT::isFloatingPoint(N1.getValueType())) {
2808 if (ISD::isNON_EXTLoad(N1.Val)) {
2809 // Just load directly from f32mem to GR32.
2810 LoadSDNode *LD = cast<LoadSDNode>(N1);
2811 N1 = DAG.getLoad(MVT::i32, LD->getChain(), LD->getBasePtr(),
2812 LD->getSrcValue(), LD->getSrcValueOffset());
2814 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
2815 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
2816 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
2817 DAG.getConstant(0, getPointerTy()));
2820 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
2821 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
2822 DAG.getConstant(Idx, getPointerTy()));
2823 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
2824 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
2825 DAG.getConstant(Idx+1, getPointerTy()));
2826 return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
2834 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2835 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
2836 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
2839 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2840 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
2841 // one of the above mentioned nodes. It has to be wrapped because otherwise
2842 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2843 // be used to form addressing mode. These wrapped nodes will be selected
2846 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
2847 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2848 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
2850 CP->getAlignment());
2851 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
2852 // With PIC, the address is actually $g + Offset.
2853 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2854 !Subtarget->isPICStyleRIPRel()) {
2855 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2856 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
2864 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
2865 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2866 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
2867 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
2868 // With PIC, the address is actually $g + Offset.
2869 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2870 !Subtarget->isPICStyleRIPRel()) {
2871 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2872 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
2876 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
2877 // load the value at address GV, not the value of GV itself. This means that
2878 // the GlobalAddress must be in the base or index register of the address, not
2879 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
2880 // The same applies for external symbols during PIC codegen
2881 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
2882 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
2888 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
2889 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
2890 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
2891 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
2892 // With PIC, the address is actually $g + Offset.
2893 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2894 !Subtarget->isPICStyleRIPRel()) {
2895 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2896 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
2903 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
2904 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
2905 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
2906 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
2907 // With PIC, the address is actually $g + Offset.
2908 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2909 !Subtarget->isPICStyleRIPRel()) {
2910 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2911 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
2918 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
2919 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
2920 "Not an i64 shift!");
2921 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
2922 SDOperand ShOpLo = Op.getOperand(0);
2923 SDOperand ShOpHi = Op.getOperand(1);
2924 SDOperand ShAmt = Op.getOperand(2);
2925 SDOperand Tmp1 = isSRA ?
2926 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
2927 DAG.getConstant(0, MVT::i32);
2929 SDOperand Tmp2, Tmp3;
2930 if (Op.getOpcode() == ISD::SHL_PARTS) {
2931 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
2932 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
2934 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
2935 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
2938 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
2939 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
2940 DAG.getConstant(32, MVT::i8));
2941 SDOperand COps[]={DAG.getEntryNode(), AndNode, DAG.getConstant(0, MVT::i8)};
2942 SDOperand InFlag = DAG.getNode(X86ISD::CMP, VTs, 2, COps, 3).getValue(1);
2945 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
2947 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
2948 SmallVector<SDOperand, 4> Ops;
2949 if (Op.getOpcode() == ISD::SHL_PARTS) {
2950 Ops.push_back(Tmp2);
2951 Ops.push_back(Tmp3);
2953 Ops.push_back(InFlag);
2954 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
2955 InFlag = Hi.getValue(1);
2958 Ops.push_back(Tmp3);
2959 Ops.push_back(Tmp1);
2961 Ops.push_back(InFlag);
2962 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
2964 Ops.push_back(Tmp2);
2965 Ops.push_back(Tmp3);
2967 Ops.push_back(InFlag);
2968 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
2969 InFlag = Lo.getValue(1);
2972 Ops.push_back(Tmp3);
2973 Ops.push_back(Tmp1);
2975 Ops.push_back(InFlag);
2976 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
2979 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
2983 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
2986 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
2987 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
2988 Op.getOperand(0).getValueType() >= MVT::i16 &&
2989 "Unknown SINT_TO_FP to lower!");
2992 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
2993 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
2994 MachineFunction &MF = DAG.getMachineFunction();
2995 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
2996 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
2997 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
2998 StackSlot, NULL, 0);
3003 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
3005 Tys = DAG.getVTList(MVT::f64, MVT::Other);
3006 SmallVector<SDOperand, 8> Ops;
3007 Ops.push_back(Chain);
3008 Ops.push_back(StackSlot);
3009 Ops.push_back(DAG.getValueType(SrcVT));
3010 Result = DAG.getNode(X86ScalarSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
3011 Tys, &Ops[0], Ops.size());
3014 Chain = Result.getValue(1);
3015 SDOperand InFlag = Result.getValue(2);
3017 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
3018 // shouldn't be necessary except that RFP cannot be live across
3019 // multiple blocks. When stackifier is fixed, they can be uncoupled.
3020 MachineFunction &MF = DAG.getMachineFunction();
3021 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3022 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3023 Tys = DAG.getVTList(MVT::Other);
3024 SmallVector<SDOperand, 8> Ops;
3025 Ops.push_back(Chain);
3026 Ops.push_back(Result);
3027 Ops.push_back(StackSlot);
3028 Ops.push_back(DAG.getValueType(Op.getValueType()));
3029 Ops.push_back(InFlag);
3030 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
3031 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
3037 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
3038 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
3039 "Unknown FP_TO_SINT to lower!");
3040 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
3042 MachineFunction &MF = DAG.getMachineFunction();
3043 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
3044 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3045 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3048 switch (Op.getValueType()) {
3049 default: assert(0 && "Invalid FP_TO_SINT to lower!");
3050 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
3051 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
3052 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
3055 SDOperand Chain = DAG.getEntryNode();
3056 SDOperand Value = Op.getOperand(0);
3058 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
3059 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
3060 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
3062 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
3064 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
3065 Chain = Value.getValue(1);
3066 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3067 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3070 // Build the FP_TO_INT*_IN_MEM
3071 SDOperand Ops[] = { Chain, Value, StackSlot };
3072 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
3075 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
3078 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
3079 MVT::ValueType VT = Op.getValueType();
3080 const Type *OpNTy = MVT::getTypeForValueType(VT);
3081 std::vector<Constant*> CV;
3082 if (VT == MVT::f64) {
3083 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(~(1ULL << 63))));
3084 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3086 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(~(1U << 31))));
3087 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3088 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3089 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3091 Constant *CS = ConstantStruct::get(CV);
3092 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3093 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3094 SmallVector<SDOperand, 3> Ops;
3095 Ops.push_back(DAG.getEntryNode());
3096 Ops.push_back(CPIdx);
3097 Ops.push_back(DAG.getSrcValue(NULL));
3098 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3099 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
3102 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
3103 MVT::ValueType VT = Op.getValueType();
3104 const Type *OpNTy = MVT::getTypeForValueType(VT);
3105 std::vector<Constant*> CV;
3106 if (VT == MVT::f64) {
3107 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(1ULL << 63)));
3108 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3110 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(1U << 31)));
3111 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3112 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3113 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3115 Constant *CS = ConstantStruct::get(CV);
3116 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3117 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3118 SmallVector<SDOperand, 3> Ops;
3119 Ops.push_back(DAG.getEntryNode());
3120 Ops.push_back(CPIdx);
3121 Ops.push_back(DAG.getSrcValue(NULL));
3122 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3123 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
3126 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
3127 SDOperand Op0 = Op.getOperand(0);
3128 SDOperand Op1 = Op.getOperand(1);
3129 MVT::ValueType VT = Op.getValueType();
3130 MVT::ValueType SrcVT = Op1.getValueType();
3131 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
3133 // If second operand is smaller, extend it first.
3134 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
3135 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
3139 // First get the sign bit of second operand.
3140 std::vector<Constant*> CV;
3141 if (SrcVT == MVT::f64) {
3142 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(1ULL << 63)));
3143 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3145 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(1U << 31)));
3146 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3147 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3148 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3150 Constant *CS = ConstantStruct::get(CV);
3151 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3152 SDVTList Tys = DAG.getVTList(SrcVT, MVT::Other);
3153 SmallVector<SDOperand, 3> Ops;
3154 Ops.push_back(DAG.getEntryNode());
3155 Ops.push_back(CPIdx);
3156 Ops.push_back(DAG.getSrcValue(NULL));
3157 SDOperand Mask1 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3158 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
3160 // Shift sign bit right or left if the two operands have different types.
3161 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
3162 // Op0 is MVT::f32, Op1 is MVT::f64.
3163 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
3164 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
3165 DAG.getConstant(32, MVT::i32));
3166 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
3167 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
3168 DAG.getConstant(0, getPointerTy()));
3171 // Clear first operand sign bit.
3173 if (VT == MVT::f64) {
3174 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(~(1ULL << 63))));
3175 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3177 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(~(1U << 31))));
3178 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3179 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3180 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3182 CS = ConstantStruct::get(CV);
3183 CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3184 Tys = DAG.getVTList(VT, MVT::Other);
3186 Ops.push_back(DAG.getEntryNode());
3187 Ops.push_back(CPIdx);
3188 Ops.push_back(DAG.getSrcValue(NULL));
3189 SDOperand Mask2 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3190 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
3192 // Or the value with the sign bit.
3193 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
3196 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG,
3198 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
3200 SDOperand Op0 = Op.getOperand(0);
3201 SDOperand Op1 = Op.getOperand(1);
3202 SDOperand CC = Op.getOperand(2);
3203 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3204 const MVT::ValueType *VTs1 = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3205 const MVT::ValueType *VTs2 = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
3206 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
3209 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
3211 SDOperand Ops1[] = { Chain, Op0, Op1 };
3212 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, Ops1, 3).getValue(1);
3213 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
3214 return DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3217 assert(isFP && "Illegal integer SetCC!");
3219 SDOperand COps[] = { Chain, Op0, Op1 };
3220 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, COps, 3).getValue(1);
3222 switch (SetCCOpcode) {
3223 default: assert(false && "Illegal floating point SetCC!");
3224 case ISD::SETOEQ: { // !PF & ZF
3225 SDOperand Ops1[] = { DAG.getConstant(X86::COND_NP, MVT::i8), Cond };
3226 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3227 SDOperand Ops2[] = { DAG.getConstant(X86::COND_E, MVT::i8),
3229 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3230 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
3232 case ISD::SETUNE: { // PF | !ZF
3233 SDOperand Ops1[] = { DAG.getConstant(X86::COND_P, MVT::i8), Cond };
3234 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3235 SDOperand Ops2[] = { DAG.getConstant(X86::COND_NE, MVT::i8),
3237 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3238 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
3243 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
3244 bool addTest = true;
3245 SDOperand Chain = DAG.getEntryNode();
3246 SDOperand Cond = Op.getOperand(0);
3248 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3250 if (Cond.getOpcode() == ISD::SETCC)
3251 Cond = LowerSETCC(Cond, DAG, Chain);
3253 if (Cond.getOpcode() == X86ISD::SETCC) {
3254 CC = Cond.getOperand(0);
3256 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3257 // (since flag operand cannot be shared). Use it as the condition setting
3258 // operand in place of the X86ISD::SETCC.
3259 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3260 // to use a test instead of duplicating the X86ISD::CMP (for register
3261 // pressure reason)?
3262 SDOperand Cmp = Cond.getOperand(1);
3263 unsigned Opc = Cmp.getOpcode();
3264 bool IllegalFPCMov = !X86ScalarSSE &&
3265 MVT::isFloatingPoint(Op.getValueType()) &&
3266 !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
3267 if ((Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) &&
3269 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3270 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3276 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3277 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3278 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3281 VTs = DAG.getNodeValueTypes(Op.getValueType(), MVT::Flag);
3282 SmallVector<SDOperand, 4> Ops;
3283 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
3284 // condition is true.
3285 Ops.push_back(Op.getOperand(2));
3286 Ops.push_back(Op.getOperand(1));
3288 Ops.push_back(Cond.getValue(1));
3289 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3292 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
3293 bool addTest = true;
3294 SDOperand Chain = Op.getOperand(0);
3295 SDOperand Cond = Op.getOperand(1);
3296 SDOperand Dest = Op.getOperand(2);
3298 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3300 if (Cond.getOpcode() == ISD::SETCC)
3301 Cond = LowerSETCC(Cond, DAG, Chain);
3303 if (Cond.getOpcode() == X86ISD::SETCC) {
3304 CC = Cond.getOperand(0);
3306 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3307 // (since flag operand cannot be shared). Use it as the condition setting
3308 // operand in place of the X86ISD::SETCC.
3309 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3310 // to use a test instead of duplicating the X86ISD::CMP (for register
3311 // pressure reason)?
3312 SDOperand Cmp = Cond.getOperand(1);
3313 unsigned Opc = Cmp.getOpcode();
3314 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) {
3315 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3316 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3322 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3323 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3324 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3326 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
3327 Cond, Op.getOperand(2), CC, Cond.getValue(1));
3330 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
3331 unsigned CallingConv= cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3333 if (Subtarget->is64Bit())
3334 return LowerX86_64CCCCallTo(Op, DAG, CallingConv);
3336 switch (CallingConv) {
3338 assert(0 && "Unsupported calling convention");
3339 case CallingConv::Fast:
3340 // TODO: Implement fastcc
3342 case CallingConv::C:
3343 case CallingConv::X86_StdCall:
3344 return LowerCCCCallTo(Op, DAG, CallingConv);
3345 case CallingConv::X86_FastCall:
3346 return LowerFastCCCallTo(Op, DAG, CallingConv);
3351 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
3352 MachineFunction &MF = DAG.getMachineFunction();
3353 const Function* Fn = MF.getFunction();
3354 if (Fn->hasExternalLinkage() &&
3355 Subtarget->isTargetCygMing() &&
3356 Fn->getName() == "main")
3357 MF.getInfo<X86FunctionInfo>()->setForceFramePointer(true);
3359 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3360 if (Subtarget->is64Bit())
3361 return LowerX86_64CCCArguments(Op, DAG);
3365 assert(0 && "Unsupported calling convention");
3366 case CallingConv::Fast:
3367 // TODO: implement fastcc.
3370 case CallingConv::C:
3371 return LowerCCCArguments(Op, DAG);
3372 case CallingConv::X86_StdCall:
3373 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(StdCall);
3374 return LowerCCCArguments(Op, DAG, true);
3375 case CallingConv::X86_FastCall:
3376 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(FastCall);
3377 return LowerFastCCArguments(Op, DAG);
3381 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
3382 SDOperand InFlag(0, 0);
3383 SDOperand Chain = Op.getOperand(0);
3385 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3386 if (Align == 0) Align = 1;
3388 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3389 // If not DWORD aligned, call memset if size is less than the threshold.
3390 // It knows how to align to the right boundary first.
3391 if ((Align & 3) != 0 ||
3392 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
3393 MVT::ValueType IntPtr = getPointerTy();
3394 const Type *IntPtrTy = getTargetData()->getIntPtrType();
3395 TargetLowering::ArgListTy Args;
3396 TargetLowering::ArgListEntry Entry;
3397 Entry.Node = Op.getOperand(1);
3398 Entry.Ty = IntPtrTy;
3399 Entry.isSigned = false;
3400 Entry.isInReg = false;
3401 Entry.isSRet = false;
3402 Args.push_back(Entry);
3403 // Extend the unsigned i8 argument to be an int value for the call.
3404 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
3405 Entry.Ty = IntPtrTy;
3406 Entry.isSigned = false;
3407 Entry.isInReg = false;
3408 Entry.isSRet = false;
3409 Args.push_back(Entry);
3410 Entry.Node = Op.getOperand(3);
3411 Args.push_back(Entry);
3412 std::pair<SDOperand,SDOperand> CallResult =
3413 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
3414 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
3415 return CallResult.second;
3420 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
3421 unsigned BytesLeft = 0;
3422 bool TwoRepStos = false;
3425 uint64_t Val = ValC->getValue() & 255;
3427 // If the value is a constant, then we can potentially use larger sets.
3428 switch (Align & 3) {
3429 case 2: // WORD aligned
3432 Val = (Val << 8) | Val;
3434 case 0: // DWORD aligned
3437 Val = (Val << 8) | Val;
3438 Val = (Val << 16) | Val;
3439 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
3442 Val = (Val << 32) | Val;
3445 default: // Byte aligned
3448 Count = Op.getOperand(3);
3452 if (AVT > MVT::i8) {
3454 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
3455 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
3456 BytesLeft = I->getValue() % UBytes;
3458 assert(AVT >= MVT::i32 &&
3459 "Do not use rep;stos if not at least DWORD aligned");
3460 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
3461 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
3466 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
3468 InFlag = Chain.getValue(1);
3471 Count = Op.getOperand(3);
3472 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
3473 InFlag = Chain.getValue(1);
3476 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
3478 InFlag = Chain.getValue(1);
3479 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
3480 Op.getOperand(1), InFlag);
3481 InFlag = Chain.getValue(1);
3483 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3484 SmallVector<SDOperand, 8> Ops;
3485 Ops.push_back(Chain);
3486 Ops.push_back(DAG.getValueType(AVT));
3487 Ops.push_back(InFlag);
3488 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
3491 InFlag = Chain.getValue(1);
3492 Count = Op.getOperand(3);
3493 MVT::ValueType CVT = Count.getValueType();
3494 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
3495 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
3496 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
3498 InFlag = Chain.getValue(1);
3499 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3501 Ops.push_back(Chain);
3502 Ops.push_back(DAG.getValueType(MVT::i8));
3503 Ops.push_back(InFlag);
3504 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
3505 } else if (BytesLeft) {
3506 // Issue stores for the last 1 - 7 bytes.
3508 unsigned Val = ValC->getValue() & 255;
3509 unsigned Offset = I->getValue() - BytesLeft;
3510 SDOperand DstAddr = Op.getOperand(1);
3511 MVT::ValueType AddrVT = DstAddr.getValueType();
3512 if (BytesLeft >= 4) {
3513 Val = (Val << 8) | Val;
3514 Val = (Val << 16) | Val;
3515 Value = DAG.getConstant(Val, MVT::i32);
3516 Chain = DAG.getStore(Chain, Value,
3517 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3518 DAG.getConstant(Offset, AddrVT)),
3523 if (BytesLeft >= 2) {
3524 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
3525 Chain = DAG.getStore(Chain, Value,
3526 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3527 DAG.getConstant(Offset, AddrVT)),
3532 if (BytesLeft == 1) {
3533 Value = DAG.getConstant(Val, MVT::i8);
3534 Chain = DAG.getStore(Chain, Value,
3535 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3536 DAG.getConstant(Offset, AddrVT)),
3544 SDOperand X86TargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) {
3545 SDOperand Chain = Op.getOperand(0);
3547 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3548 if (Align == 0) Align = 1;
3550 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3551 // If not DWORD aligned, call memcpy if size is less than the threshold.
3552 // It knows how to align to the right boundary first.
3553 if ((Align & 3) != 0 ||
3554 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
3555 MVT::ValueType IntPtr = getPointerTy();
3556 TargetLowering::ArgListTy Args;
3557 TargetLowering::ArgListEntry Entry;
3558 Entry.Ty = getTargetData()->getIntPtrType();
3559 Entry.isSigned = false;
3560 Entry.isInReg = false;
3561 Entry.isSRet = false;
3562 Entry.Node = Op.getOperand(1); Args.push_back(Entry);
3563 Entry.Node = Op.getOperand(2); Args.push_back(Entry);
3564 Entry.Node = Op.getOperand(3); Args.push_back(Entry);
3565 std::pair<SDOperand,SDOperand> CallResult =
3566 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
3567 DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
3568 return CallResult.second;
3573 unsigned BytesLeft = 0;
3574 bool TwoRepMovs = false;
3575 switch (Align & 3) {
3576 case 2: // WORD aligned
3579 case 0: // DWORD aligned
3581 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
3584 default: // Byte aligned
3586 Count = Op.getOperand(3);
3590 if (AVT > MVT::i8) {
3592 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
3593 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
3594 BytesLeft = I->getValue() % UBytes;
3596 assert(AVT >= MVT::i32 &&
3597 "Do not use rep;movs if not at least DWORD aligned");
3598 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
3599 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
3604 SDOperand InFlag(0, 0);
3605 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
3607 InFlag = Chain.getValue(1);
3608 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
3609 Op.getOperand(1), InFlag);
3610 InFlag = Chain.getValue(1);
3611 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
3612 Op.getOperand(2), InFlag);
3613 InFlag = Chain.getValue(1);
3615 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3616 SmallVector<SDOperand, 8> Ops;
3617 Ops.push_back(Chain);
3618 Ops.push_back(DAG.getValueType(AVT));
3619 Ops.push_back(InFlag);
3620 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
3623 InFlag = Chain.getValue(1);
3624 Count = Op.getOperand(3);
3625 MVT::ValueType CVT = Count.getValueType();
3626 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
3627 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
3628 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
3630 InFlag = Chain.getValue(1);
3631 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3633 Ops.push_back(Chain);
3634 Ops.push_back(DAG.getValueType(MVT::i8));
3635 Ops.push_back(InFlag);
3636 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
3637 } else if (BytesLeft) {
3638 // Issue loads and stores for the last 1 - 7 bytes.
3639 unsigned Offset = I->getValue() - BytesLeft;
3640 SDOperand DstAddr = Op.getOperand(1);
3641 MVT::ValueType DstVT = DstAddr.getValueType();
3642 SDOperand SrcAddr = Op.getOperand(2);
3643 MVT::ValueType SrcVT = SrcAddr.getValueType();
3645 if (BytesLeft >= 4) {
3646 Value = DAG.getLoad(MVT::i32, Chain,
3647 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3648 DAG.getConstant(Offset, SrcVT)),
3650 Chain = Value.getValue(1);
3651 Chain = DAG.getStore(Chain, Value,
3652 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3653 DAG.getConstant(Offset, DstVT)),
3658 if (BytesLeft >= 2) {
3659 Value = DAG.getLoad(MVT::i16, Chain,
3660 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3661 DAG.getConstant(Offset, SrcVT)),
3663 Chain = Value.getValue(1);
3664 Chain = DAG.getStore(Chain, Value,
3665 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3666 DAG.getConstant(Offset, DstVT)),
3672 if (BytesLeft == 1) {
3673 Value = DAG.getLoad(MVT::i8, Chain,
3674 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3675 DAG.getConstant(Offset, SrcVT)),
3677 Chain = Value.getValue(1);
3678 Chain = DAG.getStore(Chain, Value,
3679 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3680 DAG.getConstant(Offset, DstVT)),
3689 X86TargetLowering::LowerREADCYCLCECOUNTER(SDOperand Op, SelectionDAG &DAG) {
3690 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3691 SDOperand TheOp = Op.getOperand(0);
3692 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheOp, 1);
3693 if (Subtarget->is64Bit()) {
3694 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
3695 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::RDX,
3696 MVT::i64, Copy1.getValue(2));
3697 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, Copy2,
3698 DAG.getConstant(32, MVT::i8));
3700 DAG.getNode(ISD::OR, MVT::i64, Copy1, Tmp), Copy2.getValue(1)
3703 Tys = DAG.getVTList(MVT::i64, MVT::Other);
3704 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2);
3707 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
3708 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::EDX,
3709 MVT::i32, Copy1.getValue(2));
3710 SDOperand Ops[] = { Copy1, Copy2, Copy2.getValue(1) };
3711 Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
3712 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 3);
3715 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
3716 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
3718 if (!Subtarget->is64Bit()) {
3719 // vastart just stores the address of the VarArgsFrameIndex slot into the
3720 // memory location argument.
3721 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
3722 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
3727 // gp_offset (0 - 6 * 8)
3728 // fp_offset (48 - 48 + 8 * 16)
3729 // overflow_arg_area (point to parameters coming in memory).
3731 SmallVector<SDOperand, 8> MemOps;
3732 SDOperand FIN = Op.getOperand(1);
3734 SDOperand Store = DAG.getStore(Op.getOperand(0),
3735 DAG.getConstant(VarArgsGPOffset, MVT::i32),
3736 FIN, SV->getValue(), SV->getOffset());
3737 MemOps.push_back(Store);
3740 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
3741 DAG.getConstant(4, getPointerTy()));
3742 Store = DAG.getStore(Op.getOperand(0),
3743 DAG.getConstant(VarArgsFPOffset, MVT::i32),
3744 FIN, SV->getValue(), SV->getOffset());
3745 MemOps.push_back(Store);
3747 // Store ptr to overflow_arg_area
3748 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
3749 DAG.getConstant(4, getPointerTy()));
3750 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
3751 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
3753 MemOps.push_back(Store);
3755 // Store ptr to reg_save_area.
3756 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
3757 DAG.getConstant(8, getPointerTy()));
3758 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
3759 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
3761 MemOps.push_back(Store);
3762 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
3766 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
3767 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
3769 default: return SDOperand(); // Don't custom lower most intrinsics.
3770 // Comparison intrinsics.
3771 case Intrinsic::x86_sse_comieq_ss:
3772 case Intrinsic::x86_sse_comilt_ss:
3773 case Intrinsic::x86_sse_comile_ss:
3774 case Intrinsic::x86_sse_comigt_ss:
3775 case Intrinsic::x86_sse_comige_ss:
3776 case Intrinsic::x86_sse_comineq_ss:
3777 case Intrinsic::x86_sse_ucomieq_ss:
3778 case Intrinsic::x86_sse_ucomilt_ss:
3779 case Intrinsic::x86_sse_ucomile_ss:
3780 case Intrinsic::x86_sse_ucomigt_ss:
3781 case Intrinsic::x86_sse_ucomige_ss:
3782 case Intrinsic::x86_sse_ucomineq_ss:
3783 case Intrinsic::x86_sse2_comieq_sd:
3784 case Intrinsic::x86_sse2_comilt_sd:
3785 case Intrinsic::x86_sse2_comile_sd:
3786 case Intrinsic::x86_sse2_comigt_sd:
3787 case Intrinsic::x86_sse2_comige_sd:
3788 case Intrinsic::x86_sse2_comineq_sd:
3789 case Intrinsic::x86_sse2_ucomieq_sd:
3790 case Intrinsic::x86_sse2_ucomilt_sd:
3791 case Intrinsic::x86_sse2_ucomile_sd:
3792 case Intrinsic::x86_sse2_ucomigt_sd:
3793 case Intrinsic::x86_sse2_ucomige_sd:
3794 case Intrinsic::x86_sse2_ucomineq_sd: {
3796 ISD::CondCode CC = ISD::SETCC_INVALID;
3799 case Intrinsic::x86_sse_comieq_ss:
3800 case Intrinsic::x86_sse2_comieq_sd:
3804 case Intrinsic::x86_sse_comilt_ss:
3805 case Intrinsic::x86_sse2_comilt_sd:
3809 case Intrinsic::x86_sse_comile_ss:
3810 case Intrinsic::x86_sse2_comile_sd:
3814 case Intrinsic::x86_sse_comigt_ss:
3815 case Intrinsic::x86_sse2_comigt_sd:
3819 case Intrinsic::x86_sse_comige_ss:
3820 case Intrinsic::x86_sse2_comige_sd:
3824 case Intrinsic::x86_sse_comineq_ss:
3825 case Intrinsic::x86_sse2_comineq_sd:
3829 case Intrinsic::x86_sse_ucomieq_ss:
3830 case Intrinsic::x86_sse2_ucomieq_sd:
3831 Opc = X86ISD::UCOMI;
3834 case Intrinsic::x86_sse_ucomilt_ss:
3835 case Intrinsic::x86_sse2_ucomilt_sd:
3836 Opc = X86ISD::UCOMI;
3839 case Intrinsic::x86_sse_ucomile_ss:
3840 case Intrinsic::x86_sse2_ucomile_sd:
3841 Opc = X86ISD::UCOMI;
3844 case Intrinsic::x86_sse_ucomigt_ss:
3845 case Intrinsic::x86_sse2_ucomigt_sd:
3846 Opc = X86ISD::UCOMI;
3849 case Intrinsic::x86_sse_ucomige_ss:
3850 case Intrinsic::x86_sse2_ucomige_sd:
3851 Opc = X86ISD::UCOMI;
3854 case Intrinsic::x86_sse_ucomineq_ss:
3855 case Intrinsic::x86_sse2_ucomineq_sd:
3856 Opc = X86ISD::UCOMI;
3862 SDOperand LHS = Op.getOperand(1);
3863 SDOperand RHS = Op.getOperand(2);
3864 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
3866 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3867 SDOperand Ops1[] = { DAG.getEntryNode(), LHS, RHS };
3868 SDOperand Cond = DAG.getNode(Opc, VTs, 2, Ops1, 3);
3869 VTs = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
3870 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
3871 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, VTs, 2, Ops2, 2);
3872 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
3877 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
3878 // Depths > 0 not supported yet!
3879 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
3882 // Just load the return address
3883 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
3884 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
3887 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
3888 // Depths > 0 not supported yet!
3889 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
3892 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
3893 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
3894 DAG.getConstant(4, getPointerTy()));
3897 /// LowerOperation - Provide custom lowering hooks for some operations.
3899 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
3900 switch (Op.getOpcode()) {
3901 default: assert(0 && "Should not custom lower this!");
3902 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3903 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3904 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3905 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
3906 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
3907 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3908 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
3909 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
3910 case ISD::SHL_PARTS:
3911 case ISD::SRA_PARTS:
3912 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
3913 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
3914 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
3915 case ISD::FABS: return LowerFABS(Op, DAG);
3916 case ISD::FNEG: return LowerFNEG(Op, DAG);
3917 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3918 case ISD::SETCC: return LowerSETCC(Op, DAG, DAG.getEntryNode());
3919 case ISD::SELECT: return LowerSELECT(Op, DAG);
3920 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
3921 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
3922 case ISD::CALL: return LowerCALL(Op, DAG);
3923 case ISD::RET: return LowerRET(Op, DAG);
3924 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
3925 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
3926 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
3927 case ISD::READCYCLECOUNTER: return LowerREADCYCLCECOUNTER(Op, DAG);
3928 case ISD::VASTART: return LowerVASTART(Op, DAG);
3929 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
3930 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3931 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3936 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
3938 default: return NULL;
3939 case X86ISD::SHLD: return "X86ISD::SHLD";
3940 case X86ISD::SHRD: return "X86ISD::SHRD";
3941 case X86ISD::FAND: return "X86ISD::FAND";
3942 case X86ISD::FOR: return "X86ISD::FOR";
3943 case X86ISD::FXOR: return "X86ISD::FXOR";
3944 case X86ISD::FSRL: return "X86ISD::FSRL";
3945 case X86ISD::FILD: return "X86ISD::FILD";
3946 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
3947 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
3948 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
3949 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
3950 case X86ISD::FLD: return "X86ISD::FLD";
3951 case X86ISD::FST: return "X86ISD::FST";
3952 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
3953 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
3954 case X86ISD::CALL: return "X86ISD::CALL";
3955 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
3956 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
3957 case X86ISD::CMP: return "X86ISD::CMP";
3958 case X86ISD::COMI: return "X86ISD::COMI";
3959 case X86ISD::UCOMI: return "X86ISD::UCOMI";
3960 case X86ISD::SETCC: return "X86ISD::SETCC";
3961 case X86ISD::CMOV: return "X86ISD::CMOV";
3962 case X86ISD::BRCOND: return "X86ISD::BRCOND";
3963 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
3964 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
3965 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
3966 case X86ISD::LOAD_PACK: return "X86ISD::LOAD_PACK";
3967 case X86ISD::LOAD_UA: return "X86ISD::LOAD_UA";
3968 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
3969 case X86ISD::Wrapper: return "X86ISD::Wrapper";
3970 case X86ISD::S2VEC: return "X86ISD::S2VEC";
3971 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
3972 case X86ISD::PINSRW: return "X86ISD::PINSRW";
3973 case X86ISD::FMAX: return "X86ISD::FMAX";
3974 case X86ISD::FMIN: return "X86ISD::FMIN";
3978 /// isLegalAddressImmediate - Return true if the integer value or
3979 /// GlobalValue can be used as the offset of the target addressing mode.
3980 bool X86TargetLowering::isLegalAddressImmediate(int64_t V) const {
3981 // X86 allows a sign-extended 32-bit immediate field.
3982 return (V > -(1LL << 32) && V < (1LL << 32)-1);
3985 bool X86TargetLowering::isLegalAddressImmediate(GlobalValue *GV) const {
3986 // In 64-bit mode, GV is 64-bit so it won't fit in the 32-bit displacement
3987 // field unless we are in small code model.
3988 if (Subtarget->is64Bit() &&
3989 getTargetMachine().getCodeModel() != CodeModel::Small)
3992 return (!Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false));
3995 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3996 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3997 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3998 /// are assumed to be legal.
4000 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
4001 // Only do shuffles on 128-bit vector types for now.
4002 if (MVT::getSizeInBits(VT) == 64) return false;
4003 return (Mask.Val->getNumOperands() <= 4 ||
4004 isSplatMask(Mask.Val) ||
4005 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
4006 X86::isUNPCKLMask(Mask.Val) ||
4007 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
4008 X86::isUNPCKHMask(Mask.Val));
4011 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
4013 SelectionDAG &DAG) const {
4014 unsigned NumElts = BVOps.size();
4015 // Only do shuffles on 128-bit vector types for now.
4016 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
4017 if (NumElts == 2) return true;
4019 return (isMOVLMask(&BVOps[0], 4) ||
4020 isCommutedMOVL(&BVOps[0], 4, true) ||
4021 isSHUFPMask(&BVOps[0], 4) ||
4022 isCommutedSHUFP(&BVOps[0], 4));
4027 //===----------------------------------------------------------------------===//
4028 // X86 Scheduler Hooks
4029 //===----------------------------------------------------------------------===//
4032 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
4033 MachineBasicBlock *BB) {
4034 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4035 switch (MI->getOpcode()) {
4036 default: assert(false && "Unexpected instr type to insert");
4037 case X86::CMOV_FR32:
4038 case X86::CMOV_FR64:
4039 case X86::CMOV_V4F32:
4040 case X86::CMOV_V2F64:
4041 case X86::CMOV_V2I64: {
4042 // To "insert" a SELECT_CC instruction, we actually have to insert the
4043 // diamond control-flow pattern. The incoming instruction knows the
4044 // destination vreg to set, the condition code register to branch on, the
4045 // true/false values to select between, and a branch opcode to use.
4046 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4047 ilist<MachineBasicBlock>::iterator It = BB;
4053 // cmpTY ccX, r1, r2
4055 // fallthrough --> copy0MBB
4056 MachineBasicBlock *thisMBB = BB;
4057 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
4058 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
4060 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
4061 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
4062 MachineFunction *F = BB->getParent();
4063 F->getBasicBlockList().insert(It, copy0MBB);
4064 F->getBasicBlockList().insert(It, sinkMBB);
4065 // Update machine-CFG edges by first adding all successors of the current
4066 // block to the new block which will contain the Phi node for the select.
4067 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
4068 e = BB->succ_end(); i != e; ++i)
4069 sinkMBB->addSuccessor(*i);
4070 // Next, remove all successors of the current block, and add the true
4071 // and fallthrough blocks as its successors.
4072 while(!BB->succ_empty())
4073 BB->removeSuccessor(BB->succ_begin());
4074 BB->addSuccessor(copy0MBB);
4075 BB->addSuccessor(sinkMBB);
4078 // %FalseValue = ...
4079 // # fallthrough to sinkMBB
4082 // Update machine-CFG edges
4083 BB->addSuccessor(sinkMBB);
4086 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4089 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
4090 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4091 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4093 delete MI; // The pseudo instruction is gone now.
4097 case X86::FP_TO_INT16_IN_MEM:
4098 case X86::FP_TO_INT32_IN_MEM:
4099 case X86::FP_TO_INT64_IN_MEM: {
4100 // Change the floating point control register to use "round towards zero"
4101 // mode when truncating to an integer value.
4102 MachineFunction *F = BB->getParent();
4103 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
4104 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
4106 // Load the old value of the high byte of the control word...
4108 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
4109 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
4111 // Set the high part to be round to zero...
4112 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
4115 // Reload the modified control word now...
4116 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4118 // Restore the memory image of control word to original value
4119 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
4122 // Get the X86 opcode to use.
4124 switch (MI->getOpcode()) {
4125 default: assert(0 && "illegal opcode!");
4126 case X86::FP_TO_INT16_IN_MEM: Opc = X86::FpIST16m; break;
4127 case X86::FP_TO_INT32_IN_MEM: Opc = X86::FpIST32m; break;
4128 case X86::FP_TO_INT64_IN_MEM: Opc = X86::FpIST64m; break;
4132 MachineOperand &Op = MI->getOperand(0);
4133 if (Op.isRegister()) {
4134 AM.BaseType = X86AddressMode::RegBase;
4135 AM.Base.Reg = Op.getReg();
4137 AM.BaseType = X86AddressMode::FrameIndexBase;
4138 AM.Base.FrameIndex = Op.getFrameIndex();
4140 Op = MI->getOperand(1);
4141 if (Op.isImmediate())
4142 AM.Scale = Op.getImm();
4143 Op = MI->getOperand(2);
4144 if (Op.isImmediate())
4145 AM.IndexReg = Op.getImm();
4146 Op = MI->getOperand(3);
4147 if (Op.isGlobalAddress()) {
4148 AM.GV = Op.getGlobal();
4150 AM.Disp = Op.getImm();
4152 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
4153 .addReg(MI->getOperand(4).getReg());
4155 // Reload the original control word now.
4156 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4158 delete MI; // The pseudo instruction is gone now.
4164 //===----------------------------------------------------------------------===//
4165 // X86 Optimization Hooks
4166 //===----------------------------------------------------------------------===//
4168 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
4170 uint64_t &KnownZero,
4172 unsigned Depth) const {
4173 unsigned Opc = Op.getOpcode();
4174 assert((Opc >= ISD::BUILTIN_OP_END ||
4175 Opc == ISD::INTRINSIC_WO_CHAIN ||
4176 Opc == ISD::INTRINSIC_W_CHAIN ||
4177 Opc == ISD::INTRINSIC_VOID) &&
4178 "Should use MaskedValueIsZero if you don't know whether Op"
4179 " is a target node!");
4181 KnownZero = KnownOne = 0; // Don't know anything.
4185 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
4190 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4191 /// element of the result of the vector shuffle.
4192 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
4193 MVT::ValueType VT = N->getValueType(0);
4194 SDOperand PermMask = N->getOperand(2);
4195 unsigned NumElems = PermMask.getNumOperands();
4196 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
4198 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4200 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4201 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
4202 SDOperand Idx = PermMask.getOperand(i);
4203 if (Idx.getOpcode() == ISD::UNDEF)
4204 return DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4205 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
4210 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
4211 /// node is a GlobalAddress + an offset.
4212 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
4213 unsigned Opc = N->getOpcode();
4214 if (Opc == X86ISD::Wrapper) {
4215 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
4216 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
4219 } else if (Opc == ISD::ADD) {
4220 SDOperand N1 = N->getOperand(0);
4221 SDOperand N2 = N->getOperand(1);
4222 if (isGAPlusOffset(N1.Val, GA, Offset)) {
4223 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
4225 Offset += V->getSignExtended();
4228 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
4229 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
4231 Offset += V->getSignExtended();
4239 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
4241 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
4242 MachineFrameInfo *MFI) {
4243 if (N->getOperand(0).Val != Base->getOperand(0).Val)
4246 SDOperand Loc = N->getOperand(1);
4247 SDOperand BaseLoc = Base->getOperand(1);
4248 if (Loc.getOpcode() == ISD::FrameIndex) {
4249 if (BaseLoc.getOpcode() != ISD::FrameIndex)
4251 int FI = dyn_cast<FrameIndexSDNode>(Loc)->getIndex();
4252 int BFI = dyn_cast<FrameIndexSDNode>(BaseLoc)->getIndex();
4253 int FS = MFI->getObjectSize(FI);
4254 int BFS = MFI->getObjectSize(BFI);
4255 if (FS != BFS || FS != Size) return false;
4256 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
4258 GlobalValue *GV1 = NULL;
4259 GlobalValue *GV2 = NULL;
4260 int64_t Offset1 = 0;
4261 int64_t Offset2 = 0;
4262 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
4263 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
4264 if (isGA1 && isGA2 && GV1 == GV2)
4265 return Offset1 == (Offset2 + Dist*Size);
4271 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
4272 const X86Subtarget *Subtarget) {
4275 if (isGAPlusOffset(Base, GV, Offset))
4276 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
4278 assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
4279 int BFI = dyn_cast<FrameIndexSDNode>(Base)->getIndex();
4281 // Fixed objects do not specify alignment, however the offsets are known.
4282 return ((Subtarget->getStackAlignment() % 16) == 0 &&
4283 (MFI->getObjectOffset(BFI) % 16) == 0);
4285 return MFI->getObjectAlignment(BFI) >= 16;
4291 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
4292 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
4293 /// if the load addresses are consecutive, non-overlapping, and in the right
4295 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
4296 const X86Subtarget *Subtarget) {
4297 MachineFunction &MF = DAG.getMachineFunction();
4298 MachineFrameInfo *MFI = MF.getFrameInfo();
4299 MVT::ValueType VT = N->getValueType(0);
4300 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
4301 SDOperand PermMask = N->getOperand(2);
4302 int NumElems = (int)PermMask.getNumOperands();
4303 SDNode *Base = NULL;
4304 for (int i = 0; i < NumElems; ++i) {
4305 SDOperand Idx = PermMask.getOperand(i);
4306 if (Idx.getOpcode() == ISD::UNDEF) {
4307 if (!Base) return SDOperand();
4310 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
4311 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
4315 else if (!isConsecutiveLoad(Arg.Val, Base,
4316 i, MVT::getSizeInBits(EVT)/8,MFI))
4321 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
4323 LoadSDNode *LD = cast<LoadSDNode>(Base);
4324 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
4325 LD->getSrcValueOffset());
4327 // Just use movups, it's shorter.
4328 SDVTList Tys = DAG.getVTList(MVT::v4f32, MVT::Other);
4329 SmallVector<SDOperand, 3> Ops;
4330 Ops.push_back(Base->getOperand(0));
4331 Ops.push_back(Base->getOperand(1));
4332 Ops.push_back(Base->getOperand(2));
4333 return DAG.getNode(ISD::BIT_CONVERT, VT,
4334 DAG.getNode(X86ISD::LOAD_UA, Tys, &Ops[0], Ops.size()));
4338 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
4339 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
4340 const X86Subtarget *Subtarget) {
4341 SDOperand Cond = N->getOperand(0);
4343 // If we have SSE[12] support, try to form min/max nodes.
4344 if (Subtarget->hasSSE2() &&
4345 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
4346 if (Cond.getOpcode() == ISD::SETCC) {
4347 // Get the LHS/RHS of the select.
4348 SDOperand LHS = N->getOperand(1);
4349 SDOperand RHS = N->getOperand(2);
4350 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
4352 unsigned Opcode = 0;
4353 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
4356 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
4359 if (!UnsafeFPMath) break;
4361 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
4363 Opcode = X86ISD::FMIN;
4366 case ISD::SETOGT: // (X > Y) ? X : Y -> max
4369 if (!UnsafeFPMath) break;
4371 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
4373 Opcode = X86ISD::FMAX;
4376 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
4379 case ISD::SETOGT: // (X > Y) ? Y : X -> min
4382 if (!UnsafeFPMath) break;
4384 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
4386 Opcode = X86ISD::FMIN;
4389 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
4392 if (!UnsafeFPMath) break;
4394 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
4396 Opcode = X86ISD::FMAX;
4402 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
4411 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
4412 DAGCombinerInfo &DCI) const {
4413 SelectionDAG &DAG = DCI.DAG;
4414 switch (N->getOpcode()) {
4416 case ISD::VECTOR_SHUFFLE:
4417 return PerformShuffleCombine(N, DAG, Subtarget);
4419 return PerformSELECTCombine(N, DAG, Subtarget);
4425 //===----------------------------------------------------------------------===//
4426 // X86 Inline Assembly Support
4427 //===----------------------------------------------------------------------===//
4429 /// getConstraintType - Given a constraint letter, return the type of
4430 /// constraint it is for this target.
4431 X86TargetLowering::ConstraintType
4432 X86TargetLowering::getConstraintType(char ConstraintLetter) const {
4433 switch (ConstraintLetter) {
4442 return C_RegisterClass;
4443 default: return TargetLowering::getConstraintType(ConstraintLetter);
4447 /// isOperandValidForConstraint - Return the specified operand (possibly
4448 /// modified) if the specified SDOperand is valid for the specified target
4449 /// constraint letter, otherwise return null.
4450 SDOperand X86TargetLowering::
4451 isOperandValidForConstraint(SDOperand Op, char Constraint, SelectionDAG &DAG) {
4452 switch (Constraint) {
4455 // Literal immediates are always ok.
4456 if (isa<ConstantSDNode>(Op)) return Op;
4458 // If we are in non-pic codegen mode, we allow the address of a global to
4459 // be used with 'i'.
4460 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op)) {
4461 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
4462 return SDOperand(0, 0);
4464 if (GA->getOpcode() != ISD::TargetGlobalAddress)
4465 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
4470 // Otherwise, not valid for this mode.
4471 return SDOperand(0, 0);
4473 return TargetLowering::isOperandValidForConstraint(Op, Constraint, DAG);
4477 std::vector<unsigned> X86TargetLowering::
4478 getRegClassForInlineAsmConstraint(const std::string &Constraint,
4479 MVT::ValueType VT) const {
4480 if (Constraint.size() == 1) {
4481 // FIXME: not handling fp-stack yet!
4482 // FIXME: not handling MMX registers yet ('y' constraint).
4483 switch (Constraint[0]) { // GCC X86 Constraint Letters
4484 default: break; // Unknown constraint letter
4485 case 'A': // EAX/EDX
4486 if (VT == MVT::i32 || VT == MVT::i64)
4487 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
4489 case 'r': // GENERAL_REGS
4490 case 'R': // LEGACY_REGS
4491 if (VT == MVT::i64 && Subtarget->is64Bit())
4492 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
4493 X86::RSI, X86::RDI, X86::RBP, X86::RSP,
4494 X86::R8, X86::R9, X86::R10, X86::R11,
4495 X86::R12, X86::R13, X86::R14, X86::R15, 0);
4497 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
4498 X86::ESI, X86::EDI, X86::EBP, X86::ESP, 0);
4499 else if (VT == MVT::i16)
4500 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
4501 X86::SI, X86::DI, X86::BP, X86::SP, 0);
4502 else if (VT == MVT::i8)
4503 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
4505 case 'l': // INDEX_REGS
4507 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
4508 X86::ESI, X86::EDI, X86::EBP, 0);
4509 else if (VT == MVT::i16)
4510 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
4511 X86::SI, X86::DI, X86::BP, 0);
4512 else if (VT == MVT::i8)
4513 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
4515 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
4518 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
4519 else if (VT == MVT::i16)
4520 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
4521 else if (VT == MVT::i8)
4522 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
4524 case 'x': // SSE_REGS if SSE1 allowed
4525 if (Subtarget->hasSSE1())
4526 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
4527 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
4529 return std::vector<unsigned>();
4530 case 'Y': // SSE_REGS if SSE2 allowed
4531 if (Subtarget->hasSSE2())
4532 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
4533 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
4535 return std::vector<unsigned>();
4539 return std::vector<unsigned>();
4542 std::pair<unsigned, const TargetRegisterClass*>
4543 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4544 MVT::ValueType VT) const {
4545 // Use the default implementation in TargetLowering to convert the register
4546 // constraint into a member of a register class.
4547 std::pair<unsigned, const TargetRegisterClass*> Res;
4548 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4550 // Not found as a standard register?
4551 if (Res.second == 0) {
4552 // GCC calls "st(0)" just plain "st".
4553 if (StringsEqualNoCase("{st}", Constraint)) {
4554 Res.first = X86::ST0;
4555 Res.second = X86::RSTRegisterClass;
4561 // Otherwise, check to see if this is a register class of the wrong value
4562 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
4563 // turn into {ax},{dx}.
4564 if (Res.second->hasType(VT))
4565 return Res; // Correct type already, nothing to do.
4567 // All of the single-register GCC register classes map their values onto
4568 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
4569 // really want an 8-bit or 32-bit register, map to the appropriate register
4570 // class and return the appropriate register.
4571 if (Res.second != X86::GR16RegisterClass)
4574 if (VT == MVT::i8) {
4575 unsigned DestReg = 0;
4576 switch (Res.first) {
4578 case X86::AX: DestReg = X86::AL; break;
4579 case X86::DX: DestReg = X86::DL; break;
4580 case X86::CX: DestReg = X86::CL; break;
4581 case X86::BX: DestReg = X86::BL; break;
4584 Res.first = DestReg;
4585 Res.second = Res.second = X86::GR8RegisterClass;
4587 } else if (VT == MVT::i32) {
4588 unsigned DestReg = 0;
4589 switch (Res.first) {
4591 case X86::AX: DestReg = X86::EAX; break;
4592 case X86::DX: DestReg = X86::EDX; break;
4593 case X86::CX: DestReg = X86::ECX; break;
4594 case X86::BX: DestReg = X86::EBX; break;
4595 case X86::SI: DestReg = X86::ESI; break;
4596 case X86::DI: DestReg = X86::EDI; break;
4597 case X86::BP: DestReg = X86::EBP; break;
4598 case X86::SP: DestReg = X86::ESP; break;
4601 Res.first = DestReg;
4602 Res.second = Res.second = X86::GR32RegisterClass;
4604 } else if (VT == MVT::i64) {
4605 unsigned DestReg = 0;
4606 switch (Res.first) {
4608 case X86::AX: DestReg = X86::RAX; break;
4609 case X86::DX: DestReg = X86::RDX; break;
4610 case X86::CX: DestReg = X86::RCX; break;
4611 case X86::BX: DestReg = X86::RBX; break;
4612 case X86::SI: DestReg = X86::RSI; break;
4613 case X86::DI: DestReg = X86::RDI; break;
4614 case X86::BP: DestReg = X86::RBP; break;
4615 case X86::SP: DestReg = X86::RSP; break;
4618 Res.first = DestReg;
4619 Res.second = Res.second = X86::GR64RegisterClass;