1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Chris Lattner and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
16 #include "X86InstrBuilder.h"
17 #include "X86ISelLowering.h"
18 #include "X86MachineFunctionInfo.h"
19 #include "X86TargetMachine.h"
20 #include "llvm/CallingConv.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/Function.h"
24 #include "llvm/Intrinsics.h"
25 #include "llvm/ADT/VectorExtras.h"
26 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
27 #include "llvm/CodeGen/CallingConvLower.h"
28 #include "llvm/CodeGen/MachineFrameInfo.h"
29 #include "llvm/CodeGen/MachineFunction.h"
30 #include "llvm/CodeGen/MachineInstrBuilder.h"
31 #include "llvm/CodeGen/SelectionDAG.h"
32 #include "llvm/CodeGen/SSARegMap.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Target/TargetOptions.h"
35 #include "llvm/ADT/StringExtras.h"
38 X86TargetLowering::X86TargetLowering(TargetMachine &TM)
39 : TargetLowering(TM) {
40 Subtarget = &TM.getSubtarget<X86Subtarget>();
41 X86ScalarSSE = Subtarget->hasSSE2();
42 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
44 // Set up the TargetLowering object.
46 // X86 is weird, it always uses i8 for shift amounts and setcc results.
47 setShiftAmountType(MVT::i8);
48 setSetCCResultType(MVT::i8);
49 setSetCCResultContents(ZeroOrOneSetCCResult);
50 setSchedulingPreference(SchedulingForRegPressure);
51 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
52 setStackPointerRegisterToSaveRestore(X86StackPtr);
54 if (Subtarget->isTargetDarwin()) {
55 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
56 setUseUnderscoreSetJmp(false);
57 setUseUnderscoreLongJmp(false);
58 } else if (Subtarget->isTargetMingw()) {
59 // MS runtime is weird: it exports _setjmp, but longjmp!
60 setUseUnderscoreSetJmp(true);
61 setUseUnderscoreLongJmp(false);
63 setUseUnderscoreSetJmp(true);
64 setUseUnderscoreLongJmp(true);
67 // Add legal addressing mode scale values.
68 addLegalAddressScale(8);
69 addLegalAddressScale(4);
70 addLegalAddressScale(2);
71 // Enter the ones which require both scale + index last. These are more
73 addLegalAddressScale(9);
74 addLegalAddressScale(5);
75 addLegalAddressScale(3);
77 // Set up the register classes.
78 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
79 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
80 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
81 if (Subtarget->is64Bit())
82 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
84 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Expand);
86 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
88 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
89 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
90 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
92 if (Subtarget->is64Bit()) {
93 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
94 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
97 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
98 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
100 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
103 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
105 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
106 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
107 // SSE has no i16 to fp conversion, only i32
109 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
111 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
112 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
115 if (!Subtarget->is64Bit()) {
116 // Custom lower SINT_TO_FP and FP_TO_SINT from/to i64 in 32-bit mode.
117 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
118 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
121 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
123 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
124 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
127 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
129 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
130 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
133 // Handle FP_TO_UINT by promoting the destination to a larger signed
135 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
136 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
137 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
139 if (Subtarget->is64Bit()) {
140 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
141 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
143 if (X86ScalarSSE && !Subtarget->hasSSE3())
144 // Expand FP_TO_UINT into a select.
145 // FIXME: We would like to use a Custom expander here eventually to do
146 // the optimal thing for SSE vs. the default expansion in the legalizer.
147 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
149 // With SSE3 we can use fisttpll to convert to a signed i64.
150 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
153 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
155 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
156 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
159 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
160 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
161 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
162 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
163 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
164 if (Subtarget->is64Bit())
165 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Expand);
166 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Expand);
167 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Expand);
168 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
169 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
170 setOperationAction(ISD::FREM , MVT::f64 , Expand);
172 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
173 setOperationAction(ISD::CTTZ , MVT::i8 , Expand);
174 setOperationAction(ISD::CTLZ , MVT::i8 , Expand);
175 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
176 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
177 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
178 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
179 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
180 setOperationAction(ISD::CTLZ , MVT::i32 , Expand);
181 if (Subtarget->is64Bit()) {
182 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
183 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
184 setOperationAction(ISD::CTLZ , MVT::i64 , Expand);
187 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
188 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
190 // These should be promoted to a larger select which is supported.
191 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
192 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
193 // X86 wants to expand cmov itself.
194 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
195 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
196 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
197 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
198 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
199 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
200 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
201 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
202 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
203 if (Subtarget->is64Bit()) {
204 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
205 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
207 // X86 ret instruction may pop stack.
208 setOperationAction(ISD::RET , MVT::Other, Custom);
210 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
211 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
212 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
213 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
214 if (Subtarget->is64Bit()) {
215 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
216 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
217 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
218 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
220 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
221 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
222 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
223 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
224 // X86 wants to expand memset / memcpy itself.
225 setOperationAction(ISD::MEMSET , MVT::Other, Custom);
226 setOperationAction(ISD::MEMCPY , MVT::Other, Custom);
228 // We don't have line number support yet.
229 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
230 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
231 // FIXME - use subtarget debug flags
232 if (!Subtarget->isTargetDarwin() &&
233 !Subtarget->isTargetELF() &&
234 !Subtarget->isTargetCygMing())
235 setOperationAction(ISD::LABEL, MVT::Other, Expand);
237 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
238 setOperationAction(ISD::VASTART , MVT::Other, Custom);
239 setOperationAction(ISD::VAARG , MVT::Other, Expand);
240 setOperationAction(ISD::VAEND , MVT::Other, Expand);
241 if (Subtarget->is64Bit())
242 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
244 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
246 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
247 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
248 if (Subtarget->is64Bit())
249 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
250 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
253 // Set up the FP register classes.
254 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
255 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
257 // Use ANDPD to simulate FABS.
258 setOperationAction(ISD::FABS , MVT::f64, Custom);
259 setOperationAction(ISD::FABS , MVT::f32, Custom);
261 // Use XORP to simulate FNEG.
262 setOperationAction(ISD::FNEG , MVT::f64, Custom);
263 setOperationAction(ISD::FNEG , MVT::f32, Custom);
265 // Use ANDPD and ORPD to simulate FCOPYSIGN.
266 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
267 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
269 // We don't support sin/cos/fmod
270 setOperationAction(ISD::FSIN , MVT::f64, Expand);
271 setOperationAction(ISD::FCOS , MVT::f64, Expand);
272 setOperationAction(ISD::FREM , MVT::f64, Expand);
273 setOperationAction(ISD::FSIN , MVT::f32, Expand);
274 setOperationAction(ISD::FCOS , MVT::f32, Expand);
275 setOperationAction(ISD::FREM , MVT::f32, Expand);
277 // Expand FP immediates into loads from the stack, except for the special
279 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
280 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
281 addLegalFPImmediate(+0.0); // xorps / xorpd
283 // Set up the FP register classes.
284 addRegisterClass(MVT::f64, X86::RFPRegisterClass);
286 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
287 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
288 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
291 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
292 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
295 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
296 addLegalFPImmediate(+0.0); // FLD0
297 addLegalFPImmediate(+1.0); // FLD1
298 addLegalFPImmediate(-0.0); // FLD0/FCHS
299 addLegalFPImmediate(-1.0); // FLD1/FCHS
302 // First set operation action for all vector types to expand. Then we
303 // will selectively turn on ones that can be effectively codegen'd.
304 for (unsigned VT = (unsigned)MVT::Vector + 1;
305 VT != (unsigned)MVT::LAST_VALUETYPE; VT++) {
306 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
307 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
308 setOperationAction(ISD::FADD, (MVT::ValueType)VT, Expand);
309 setOperationAction(ISD::FSUB, (MVT::ValueType)VT, Expand);
310 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
311 setOperationAction(ISD::FMUL, (MVT::ValueType)VT, Expand);
312 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
313 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
314 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
315 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
316 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
317 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand);
318 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Expand);
319 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
320 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
323 if (Subtarget->hasMMX()) {
324 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
325 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
326 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
328 // FIXME: add MMX packed arithmetics
329 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Expand);
330 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Expand);
331 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Expand);
334 if (Subtarget->hasSSE1()) {
335 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
337 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
338 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
339 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
340 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
341 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
342 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
343 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
344 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
345 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
348 if (Subtarget->hasSSE2()) {
349 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
350 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
351 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
352 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
353 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
355 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
356 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
357 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
358 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
359 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
360 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
361 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
362 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
363 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
364 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
365 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
367 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
368 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
369 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
370 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
371 // Implement v4f32 insert_vector_elt in terms of SSE2 v8i16 ones.
372 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
374 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
375 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
376 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
377 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
378 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
380 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
381 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
382 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
383 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
384 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
385 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
387 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
388 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
389 setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote);
390 AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64);
391 setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote);
392 AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64);
393 setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote);
394 AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64);
395 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote);
396 AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64);
397 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
398 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64);
401 // Custom lower v2i64 and v2f64 selects.
402 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
403 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
404 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
405 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
408 // We want to custom lower some of our intrinsics.
409 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
411 // We have target-specific dag combine patterns for the following nodes:
412 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
413 setTargetDAGCombine(ISD::SELECT);
415 computeRegisterProperties();
417 // FIXME: These should be based on subtarget info. Plus, the values should
418 // be smaller when we are in optimizing for size mode.
419 maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
420 maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
421 maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
422 allowUnalignedMemoryAccesses = true; // x86 supports it!
426 //===----------------------------------------------------------------------===//
427 // Return Value Calling Convention Implementation
428 //===----------------------------------------------------------------------===//
430 #include "X86GenCallingConv.inc"
432 /// LowerRET - Lower an ISD::RET node.
433 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
434 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
436 SmallVector<CCValAssign, 16> RVLocs;
437 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
438 CCState CCInfo(CC, getTargetMachine(), RVLocs);
439 CCInfo.AnalyzeReturn(Op.Val, RetCC_X86);
442 // If this is the first return lowered for this function, add the regs to the
443 // liveout set for the function.
444 if (DAG.getMachineFunction().liveout_empty()) {
445 for (unsigned i = 0; i != RVLocs.size(); ++i)
446 if (RVLocs[i].isRegLoc())
447 DAG.getMachineFunction().addLiveOut(RVLocs[i].getLocReg());
450 SDOperand Chain = Op.getOperand(0);
453 // Copy the result values into the output registers.
454 if (RVLocs.size() != 1 || !RVLocs[0].isRegLoc() ||
455 RVLocs[0].getLocReg() != X86::ST0) {
456 for (unsigned i = 0; i != RVLocs.size(); ++i) {
457 CCValAssign &VA = RVLocs[i];
458 assert(VA.isRegLoc() && "Can only return in registers!");
459 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1),
461 Flag = Chain.getValue(1);
464 // We need to handle a destination of ST0 specially, because it isn't really
466 SDOperand Value = Op.getOperand(1);
468 // If this is an FP return with ScalarSSE, we need to move the value from
469 // an XMM register onto the fp-stack.
473 // If this is a load into a scalarsse value, don't store the loaded value
474 // back to the stack, only to reload it: just replace the scalar-sse load.
475 if (ISD::isNON_EXTLoad(Value.Val) &&
476 (Chain == Value.getValue(1) || Chain == Value.getOperand(0))) {
477 Chain = Value.getOperand(0);
478 MemLoc = Value.getOperand(1);
480 // Spill the value to memory and reload it into top of stack.
481 unsigned Size = MVT::getSizeInBits(RVLocs[0].getValVT())/8;
482 MachineFunction &MF = DAG.getMachineFunction();
483 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
484 MemLoc = DAG.getFrameIndex(SSFI, getPointerTy());
485 Chain = DAG.getStore(Op.getOperand(0), Value, MemLoc, NULL, 0);
487 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
488 SDOperand Ops[] = {Chain, MemLoc, DAG.getValueType(RVLocs[0].getValVT())};
489 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
490 Chain = Value.getValue(1);
493 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
494 SDOperand Ops[] = { Chain, Value };
495 Chain = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops, 2);
496 Flag = Chain.getValue(1);
499 SDOperand BytesToPop = DAG.getConstant(getBytesToPopOnReturn(), MVT::i16);
501 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop, Flag);
503 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop);
507 /// LowerCallResult - Lower the result values of an ISD::CALL into the
508 /// appropriate copies out of appropriate physical registers. This assumes that
509 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
510 /// being lowered. The returns a SDNode with the same number of values as the
512 SDNode *X86TargetLowering::
513 LowerCallResult(SDOperand Chain, SDOperand InFlag, SDNode *TheCall,
514 unsigned CallingConv, SelectionDAG &DAG) {
516 // Assign locations to each value returned by this call.
517 SmallVector<CCValAssign, 16> RVLocs;
518 CCState CCInfo(CallingConv, getTargetMachine(), RVLocs);
519 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
522 SmallVector<SDOperand, 8> ResultVals;
524 // Copy all of the result registers out of their specified physreg.
525 if (RVLocs.size() != 1 || RVLocs[0].getLocReg() != X86::ST0) {
526 for (unsigned i = 0; i != RVLocs.size(); ++i) {
527 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
528 RVLocs[i].getValVT(), InFlag).getValue(1);
529 InFlag = Chain.getValue(2);
530 ResultVals.push_back(Chain.getValue(0));
533 // Copies from the FP stack are special, as ST0 isn't a valid register
534 // before the fp stackifier runs.
536 // Copy ST0 into an RFP register with FP_GET_RESULT.
537 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
538 SDOperand GROps[] = { Chain, InFlag };
539 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, GROps, 2);
540 Chain = RetVal.getValue(1);
541 InFlag = RetVal.getValue(2);
543 // If we are using ScalarSSE, store ST(0) to the stack and reload it into
546 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
547 // shouldn't be necessary except that RFP cannot be live across
548 // multiple blocks. When stackifier is fixed, they can be uncoupled.
549 MachineFunction &MF = DAG.getMachineFunction();
550 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
551 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
553 Chain, RetVal, StackSlot, DAG.getValueType(RVLocs[0].getValVT()), InFlag
555 Chain = DAG.getNode(X86ISD::FST, MVT::Other, Ops, 5);
556 RetVal = DAG.getLoad(RVLocs[0].getValVT(), Chain, StackSlot, NULL, 0);
557 Chain = RetVal.getValue(1);
560 if (RVLocs[0].getValVT() == MVT::f32 && !X86ScalarSSE)
561 // FIXME: we would really like to remember that this FP_ROUND
562 // operation is okay to eliminate if we allow excess FP precision.
563 RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal);
564 ResultVals.push_back(RetVal);
567 // Merge everything together with a MERGE_VALUES node.
568 ResultVals.push_back(Chain);
569 return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(),
570 &ResultVals[0], ResultVals.size()).Val;
574 //===----------------------------------------------------------------------===//
575 // C & StdCall Calling Convention implementation
576 //===----------------------------------------------------------------------===//
577 // StdCall calling convention seems to be standard for many Windows' API
578 // routines and around. It differs from C calling convention just a little:
579 // callee should clean up the stack, not caller. Symbols should be also
580 // decorated in some fancy way :) It doesn't support any vector arguments.
582 /// AddLiveIn - This helper function adds the specified physical register to the
583 /// MachineFunction as a live in value. It also creates a corresponding virtual
585 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
586 const TargetRegisterClass *RC) {
587 assert(RC->contains(PReg) && "Not the correct regclass!");
588 unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC);
589 MF.addLiveIn(PReg, VReg);
593 SDOperand X86TargetLowering::LowerCCCArguments(SDOperand Op, SelectionDAG &DAG,
595 unsigned NumArgs = Op.Val->getNumValues() - 1;
596 MachineFunction &MF = DAG.getMachineFunction();
597 MachineFrameInfo *MFI = MF.getFrameInfo();
598 SDOperand Root = Op.getOperand(0);
599 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
601 // Assign locations to all of the incoming arguments.
602 SmallVector<CCValAssign, 16> ArgLocs;
603 CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(),
605 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_C);
607 SmallVector<SDOperand, 8> ArgValues;
608 unsigned LastVal = ~0U;
609 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
610 CCValAssign &VA = ArgLocs[i];
611 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
613 assert(VA.getValNo() != LastVal &&
614 "Don't support value assigned to multiple locs yet");
615 LastVal = VA.getValNo();
618 MVT::ValueType RegVT = VA.getLocVT();
619 TargetRegisterClass *RC;
620 if (RegVT == MVT::i32)
621 RC = X86::GR32RegisterClass;
623 assert(MVT::isVector(RegVT));
624 RC = X86::VR128RegisterClass;
627 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
628 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
630 // If this is an 8 or 16-bit value, it is really passed promoted to 32
631 // bits. Insert an assert[sz]ext to capture this, then truncate to the
633 if (VA.getLocInfo() == CCValAssign::SExt)
634 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
635 DAG.getValueType(VA.getValVT()));
636 else if (VA.getLocInfo() == CCValAssign::ZExt)
637 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
638 DAG.getValueType(VA.getValVT()));
640 if (VA.getLocInfo() != CCValAssign::Full)
641 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
643 ArgValues.push_back(ArgValue);
645 assert(VA.isMemLoc());
647 // Create the nodes corresponding to a load from this parameter slot.
648 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
649 VA.getLocMemOffset());
650 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
651 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
655 unsigned StackSize = CCInfo.getNextStackOffset();
657 ArgValues.push_back(Root);
659 // If the function takes variable number of arguments, make a frame index for
660 // the start of the first vararg value... for expansion of llvm.va_start.
662 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
664 if (isStdCall && !isVarArg) {
665 BytesToPopOnReturn = StackSize; // Callee pops everything..
666 BytesCallerReserves = 0;
668 BytesToPopOnReturn = 0; // Callee pops hidden struct pointer.
670 // If this is an sret function, the return should pop the hidden pointer.
671 if (NumArgs && (cast<ConstantSDNode>(Op.getOperand(3))->getValue() & 4))
672 BytesToPopOnReturn = 4;
674 BytesCallerReserves = StackSize;
677 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
678 ReturnAddrIndex = 0; // No return address slot generated yet.
680 MF.getInfo<X86FunctionInfo>()->setBytesToPopOnReturn(BytesToPopOnReturn);
682 // Return the new list of results.
683 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
684 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
687 SDOperand X86TargetLowering::LowerCCCCallTo(SDOperand Op, SelectionDAG &DAG,
689 SDOperand Chain = Op.getOperand(0);
690 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
691 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
692 SDOperand Callee = Op.getOperand(4);
693 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
695 // Analyze operands of the call, assigning locations to each operand.
696 SmallVector<CCValAssign, 16> ArgLocs;
697 CCState CCInfo(CC, getTargetMachine(), ArgLocs);
698 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_C);
700 // Get a count of how many bytes are to be pushed on the stack.
701 unsigned NumBytes = CCInfo.getNextStackOffset();
703 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
705 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
706 SmallVector<SDOperand, 8> MemOpChains;
710 // Walk the register/memloc assignments, inserting copies/loads.
711 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
712 CCValAssign &VA = ArgLocs[i];
713 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
715 // Promote the value if needed.
716 switch (VA.getLocInfo()) {
717 default: assert(0 && "Unknown loc info!");
718 case CCValAssign::Full: break;
719 case CCValAssign::SExt:
720 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
722 case CCValAssign::ZExt:
723 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
725 case CCValAssign::AExt:
726 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
731 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
733 assert(VA.isMemLoc());
734 if (StackPtr.Val == 0)
735 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
736 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
737 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
738 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
742 // If the first argument is an sret pointer, remember it.
743 bool isSRet = NumOps &&(cast<ConstantSDNode>(Op.getOperand(6))->getValue()&4);
745 if (!MemOpChains.empty())
746 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
747 &MemOpChains[0], MemOpChains.size());
749 // Build a sequence of copy-to-reg nodes chained together with token chain
750 // and flag operands which copy the outgoing args into registers.
752 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
753 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
755 InFlag = Chain.getValue(1);
758 // ELF / PIC requires GOT in the EBX register before function calls via PLT
760 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
761 Subtarget->isPICStyleGOT()) {
762 Chain = DAG.getCopyToReg(Chain, X86::EBX,
763 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
765 InFlag = Chain.getValue(1);
768 // If the callee is a GlobalAddress node (quite common, every direct call is)
769 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
770 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
771 // We should use extra load for direct calls to dllimported functions in
773 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
774 getTargetMachine(), true))
775 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
776 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
777 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
779 // Returns a chain & a flag for retval copy to use.
780 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
781 SmallVector<SDOperand, 8> Ops;
782 Ops.push_back(Chain);
783 Ops.push_back(Callee);
785 // Add argument registers to the end of the list so that they are known live
787 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
788 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
789 RegsToPass[i].second.getValueType()));
791 // Add an implicit use GOT pointer in EBX.
792 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
793 Subtarget->isPICStyleGOT())
794 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
797 Ops.push_back(InFlag);
799 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
800 NodeTys, &Ops[0], Ops.size());
801 InFlag = Chain.getValue(1);
803 // Create the CALLSEQ_END node.
804 unsigned NumBytesForCalleeToPush = 0;
806 if (CC == CallingConv::X86_StdCall) {
808 NumBytesForCalleeToPush = isSRet ? 4 : 0;
810 NumBytesForCalleeToPush = NumBytes;
812 // If this is is a call to a struct-return function, the callee
813 // pops the hidden struct pointer, so we have to push it back.
814 // This is common for Darwin/X86, Linux & Mingw32 targets.
815 NumBytesForCalleeToPush = isSRet ? 4 : 0;
818 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
820 Ops.push_back(Chain);
821 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
822 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
823 Ops.push_back(InFlag);
824 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
825 InFlag = Chain.getValue(1);
827 // Handle result values, copying them out of physregs into vregs that we
829 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
833 //===----------------------------------------------------------------------===//
834 // FastCall Calling Convention implementation
835 //===----------------------------------------------------------------------===//
837 // The X86 'fastcall' calling convention passes up to two integer arguments in
838 // registers (an appropriate portion of ECX/EDX), passes arguments in C order,
839 // and requires that the callee pop its arguments off the stack (allowing proper
840 // tail calls), and has the same return value conventions as C calling convs.
842 // This calling convention always arranges for the callee pop value to be 8n+4
843 // bytes, which is needed for tail recursion elimination and stack alignment
846 X86TargetLowering::LowerFastCCArguments(SDOperand Op, SelectionDAG &DAG) {
847 MachineFunction &MF = DAG.getMachineFunction();
848 MachineFrameInfo *MFI = MF.getFrameInfo();
849 SDOperand Root = Op.getOperand(0);
851 // Assign locations to all of the incoming arguments.
852 SmallVector<CCValAssign, 16> ArgLocs;
853 CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(),
855 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_FastCall);
857 SmallVector<SDOperand, 8> ArgValues;
858 unsigned LastVal = ~0U;
859 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
860 CCValAssign &VA = ArgLocs[i];
861 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
863 assert(VA.getValNo() != LastVal &&
864 "Don't support value assigned to multiple locs yet");
865 LastVal = VA.getValNo();
868 MVT::ValueType RegVT = VA.getLocVT();
869 TargetRegisterClass *RC;
870 if (RegVT == MVT::i32)
871 RC = X86::GR32RegisterClass;
873 assert(MVT::isVector(RegVT));
874 RC = X86::VR128RegisterClass;
877 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
878 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
880 // If this is an 8 or 16-bit value, it is really passed promoted to 32
881 // bits. Insert an assert[sz]ext to capture this, then truncate to the
883 if (VA.getLocInfo() == CCValAssign::SExt)
884 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
885 DAG.getValueType(VA.getValVT()));
886 else if (VA.getLocInfo() == CCValAssign::ZExt)
887 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
888 DAG.getValueType(VA.getValVT()));
890 if (VA.getLocInfo() != CCValAssign::Full)
891 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
893 ArgValues.push_back(ArgValue);
895 assert(VA.isMemLoc());
897 // Create the nodes corresponding to a load from this parameter slot.
898 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
899 VA.getLocMemOffset());
900 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
901 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
905 ArgValues.push_back(Root);
907 unsigned StackSize = CCInfo.getNextStackOffset();
909 if (!Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows()) {
910 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
911 // arguments and the arguments after the retaddr has been pushed are aligned.
912 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 if (!Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows()) {
944 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
945 // arguments and the arguments after the retaddr has been pushed are aligned.
946 if ((NumBytes & 7) == 0)
950 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
952 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
953 SmallVector<SDOperand, 8> MemOpChains;
957 // Walk the register/memloc assignments, inserting copies/loads.
958 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
959 CCValAssign &VA = ArgLocs[i];
960 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
962 // Promote the value if needed.
963 switch (VA.getLocInfo()) {
964 default: assert(0 && "Unknown loc info!");
965 case CCValAssign::Full: break;
966 case CCValAssign::SExt:
967 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
969 case CCValAssign::ZExt:
970 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
972 case CCValAssign::AExt:
973 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
978 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
980 assert(VA.isMemLoc());
981 if (StackPtr.Val == 0)
982 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
983 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
984 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
985 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
989 if (!MemOpChains.empty())
990 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
991 &MemOpChains[0], MemOpChains.size());
993 // Build a sequence of copy-to-reg nodes chained together with token chain
994 // and flag operands which copy the outgoing args into registers.
996 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
997 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
999 InFlag = Chain.getValue(1);
1002 // If the callee is a GlobalAddress node (quite common, every direct call is)
1003 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1004 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1005 // We should use extra load for direct calls to dllimported functions in
1007 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1008 getTargetMachine(), true))
1009 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1010 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1011 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1013 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1015 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1016 Subtarget->isPICStyleGOT()) {
1017 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1018 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1020 InFlag = Chain.getValue(1);
1023 // Returns a chain & a flag for retval copy to use.
1024 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1025 SmallVector<SDOperand, 8> Ops;
1026 Ops.push_back(Chain);
1027 Ops.push_back(Callee);
1029 // Add argument registers to the end of the list so that they are known live
1031 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1032 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1033 RegsToPass[i].second.getValueType()));
1035 // Add an implicit use GOT pointer in EBX.
1036 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1037 Subtarget->isPICStyleGOT())
1038 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1041 Ops.push_back(InFlag);
1043 // FIXME: Do not generate X86ISD::TAILCALL for now.
1044 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1045 NodeTys, &Ops[0], Ops.size());
1046 InFlag = Chain.getValue(1);
1048 // Returns a flag for retval copy to use.
1049 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1051 Ops.push_back(Chain);
1052 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1053 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1054 Ops.push_back(InFlag);
1055 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1056 InFlag = Chain.getValue(1);
1058 // Handle result values, copying them out of physregs into vregs that we
1060 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1064 //===----------------------------------------------------------------------===//
1065 // X86-64 C Calling Convention implementation
1066 //===----------------------------------------------------------------------===//
1069 X86TargetLowering::LowerX86_64CCCArguments(SDOperand Op, SelectionDAG &DAG) {
1070 MachineFunction &MF = DAG.getMachineFunction();
1071 MachineFrameInfo *MFI = MF.getFrameInfo();
1072 SDOperand Root = Op.getOperand(0);
1073 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1075 static const unsigned GPR64ArgRegs[] = {
1076 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1078 static const unsigned XMMArgRegs[] = {
1079 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1080 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1084 // Assign locations to all of the incoming arguments.
1085 SmallVector<CCValAssign, 16> ArgLocs;
1086 CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(),
1088 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_64_C);
1090 SmallVector<SDOperand, 8> ArgValues;
1091 unsigned LastVal = ~0U;
1092 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1093 CCValAssign &VA = ArgLocs[i];
1094 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1096 assert(VA.getValNo() != LastVal &&
1097 "Don't support value assigned to multiple locs yet");
1098 LastVal = VA.getValNo();
1100 if (VA.isRegLoc()) {
1101 MVT::ValueType RegVT = VA.getLocVT();
1102 TargetRegisterClass *RC;
1103 if (RegVT == MVT::i32)
1104 RC = X86::GR32RegisterClass;
1105 else if (RegVT == MVT::i64)
1106 RC = X86::GR64RegisterClass;
1107 else if (RegVT == MVT::f32)
1108 RC = X86::FR32RegisterClass;
1109 else if (RegVT == MVT::f64)
1110 RC = X86::FR64RegisterClass;
1112 assert(MVT::isVector(RegVT));
1113 RC = X86::VR128RegisterClass;
1116 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1117 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT);
1119 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1120 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1122 if (VA.getLocInfo() == CCValAssign::SExt)
1123 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1124 DAG.getValueType(VA.getValVT()));
1125 else if (VA.getLocInfo() == CCValAssign::ZExt)
1126 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1127 DAG.getValueType(VA.getValVT()));
1129 if (VA.getLocInfo() != CCValAssign::Full)
1130 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1132 ArgValues.push_back(ArgValue);
1134 assert(VA.isMemLoc());
1136 // Create the nodes corresponding to a load from this parameter slot.
1137 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
1138 VA.getLocMemOffset());
1139 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1140 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
1144 unsigned StackSize = CCInfo.getNextStackOffset();
1146 // If the function takes variable number of arguments, make a frame index for
1147 // the start of the first vararg value... for expansion of llvm.va_start.
1149 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs, 6);
1150 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1152 // For X86-64, if there are vararg parameters that are passed via
1153 // registers, then we must store them to their spots on the stack so they
1154 // may be loaded by deferencing the result of va_next.
1155 VarArgsGPOffset = NumIntRegs * 8;
1156 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1157 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1158 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1160 // Store the integer parameter registers.
1161 SmallVector<SDOperand, 8> MemOps;
1162 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1163 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1164 DAG.getConstant(VarArgsGPOffset, getPointerTy()));
1165 for (; NumIntRegs != 6; ++NumIntRegs) {
1166 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1167 X86::GR64RegisterClass);
1168 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1169 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1170 MemOps.push_back(Store);
1171 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1172 DAG.getConstant(8, getPointerTy()));
1175 // Now store the XMM (fp + vector) parameter registers.
1176 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1177 DAG.getConstant(VarArgsFPOffset, getPointerTy()));
1178 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1179 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1180 X86::VR128RegisterClass);
1181 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1182 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1183 MemOps.push_back(Store);
1184 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1185 DAG.getConstant(16, getPointerTy()));
1187 if (!MemOps.empty())
1188 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1189 &MemOps[0], MemOps.size());
1192 ArgValues.push_back(Root);
1194 ReturnAddrIndex = 0; // No return address slot generated yet.
1195 BytesToPopOnReturn = 0; // Callee pops nothing.
1196 BytesCallerReserves = StackSize;
1198 // Return the new list of results.
1199 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1200 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1204 X86TargetLowering::LowerX86_64CCCCallTo(SDOperand Op, SelectionDAG &DAG,
1206 SDOperand Chain = Op.getOperand(0);
1207 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1208 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1209 SDOperand Callee = Op.getOperand(4);
1211 // Analyze operands of the call, assigning locations to each operand.
1212 SmallVector<CCValAssign, 16> ArgLocs;
1213 CCState CCInfo(CC, getTargetMachine(), ArgLocs);
1214 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_C);
1216 // Get a count of how many bytes are to be pushed on the stack.
1217 unsigned NumBytes = CCInfo.getNextStackOffset();
1218 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1220 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1221 SmallVector<SDOperand, 8> MemOpChains;
1225 // Walk the register/memloc assignments, inserting copies/loads.
1226 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1227 CCValAssign &VA = ArgLocs[i];
1228 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1230 // Promote the value if needed.
1231 switch (VA.getLocInfo()) {
1232 default: assert(0 && "Unknown loc info!");
1233 case CCValAssign::Full: break;
1234 case CCValAssign::SExt:
1235 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1237 case CCValAssign::ZExt:
1238 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1240 case CCValAssign::AExt:
1241 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1245 if (VA.isRegLoc()) {
1246 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1248 assert(VA.isMemLoc());
1249 if (StackPtr.Val == 0)
1250 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1251 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
1252 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1253 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1257 if (!MemOpChains.empty())
1258 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1259 &MemOpChains[0], MemOpChains.size());
1261 // Build a sequence of copy-to-reg nodes chained together with token chain
1262 // and flag operands which copy the outgoing args into registers.
1264 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1265 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1267 InFlag = Chain.getValue(1);
1271 // From AMD64 ABI document:
1272 // For calls that may call functions that use varargs or stdargs
1273 // (prototype-less calls or calls to functions containing ellipsis (...) in
1274 // the declaration) %al is used as hidden argument to specify the number
1275 // of SSE registers used. The contents of %al do not need to match exactly
1276 // the number of registers, but must be an ubound on the number of SSE
1277 // registers used and is in the range 0 - 8 inclusive.
1279 // Count the number of XMM registers allocated.
1280 static const unsigned XMMArgRegs[] = {
1281 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1282 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1284 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1286 Chain = DAG.getCopyToReg(Chain, X86::AL,
1287 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1288 InFlag = Chain.getValue(1);
1291 // If the callee is a GlobalAddress node (quite common, every direct call is)
1292 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1293 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1294 // We should use extra load for direct calls to dllimported functions in
1296 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1297 getTargetMachine(), true))
1298 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1299 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1300 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1302 // Returns a chain & a flag for retval copy to use.
1303 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1304 SmallVector<SDOperand, 8> Ops;
1305 Ops.push_back(Chain);
1306 Ops.push_back(Callee);
1308 // Add argument registers to the end of the list so that they are known live
1310 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1311 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1312 RegsToPass[i].second.getValueType()));
1315 Ops.push_back(InFlag);
1317 // FIXME: Do not generate X86ISD::TAILCALL for now.
1318 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1319 NodeTys, &Ops[0], Ops.size());
1320 InFlag = Chain.getValue(1);
1322 // Returns a flag for retval copy to use.
1323 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1325 Ops.push_back(Chain);
1326 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1327 Ops.push_back(DAG.getConstant(0, getPointerTy()));
1328 Ops.push_back(InFlag);
1329 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1330 InFlag = Chain.getValue(1);
1332 // Handle result values, copying them out of physregs into vregs that we
1334 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1338 //===----------------------------------------------------------------------===//
1339 // Other Lowering Hooks
1340 //===----------------------------------------------------------------------===//
1343 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1344 if (ReturnAddrIndex == 0) {
1345 // Set up a frame object for the return address.
1346 MachineFunction &MF = DAG.getMachineFunction();
1347 if (Subtarget->is64Bit())
1348 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
1350 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
1353 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1358 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1359 /// specific condition code. It returns a false if it cannot do a direct
1360 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1362 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1363 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
1364 SelectionDAG &DAG) {
1365 X86CC = X86::COND_INVALID;
1367 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1368 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1369 // X > -1 -> X == 0, jump !sign.
1370 RHS = DAG.getConstant(0, RHS.getValueType());
1371 X86CC = X86::COND_NS;
1373 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1374 // X < 0 -> X == 0, jump on sign.
1375 X86CC = X86::COND_S;
1380 switch (SetCCOpcode) {
1382 case ISD::SETEQ: X86CC = X86::COND_E; break;
1383 case ISD::SETGT: X86CC = X86::COND_G; break;
1384 case ISD::SETGE: X86CC = X86::COND_GE; break;
1385 case ISD::SETLT: X86CC = X86::COND_L; break;
1386 case ISD::SETLE: X86CC = X86::COND_LE; break;
1387 case ISD::SETNE: X86CC = X86::COND_NE; break;
1388 case ISD::SETULT: X86CC = X86::COND_B; break;
1389 case ISD::SETUGT: X86CC = X86::COND_A; break;
1390 case ISD::SETULE: X86CC = X86::COND_BE; break;
1391 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1394 // On a floating point condition, the flags are set as follows:
1396 // 0 | 0 | 0 | X > Y
1397 // 0 | 0 | 1 | X < Y
1398 // 1 | 0 | 0 | X == Y
1399 // 1 | 1 | 1 | unordered
1401 switch (SetCCOpcode) {
1404 case ISD::SETEQ: X86CC = X86::COND_E; break;
1405 case ISD::SETOLT: Flip = true; // Fallthrough
1407 case ISD::SETGT: X86CC = X86::COND_A; break;
1408 case ISD::SETOLE: Flip = true; // Fallthrough
1410 case ISD::SETGE: X86CC = X86::COND_AE; break;
1411 case ISD::SETUGT: Flip = true; // Fallthrough
1413 case ISD::SETLT: X86CC = X86::COND_B; break;
1414 case ISD::SETUGE: Flip = true; // Fallthrough
1416 case ISD::SETLE: X86CC = X86::COND_BE; break;
1418 case ISD::SETNE: X86CC = X86::COND_NE; break;
1419 case ISD::SETUO: X86CC = X86::COND_P; break;
1420 case ISD::SETO: X86CC = X86::COND_NP; break;
1423 std::swap(LHS, RHS);
1426 return X86CC != X86::COND_INVALID;
1429 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
1430 /// code. Current x86 isa includes the following FP cmov instructions:
1431 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
1432 static bool hasFPCMov(unsigned X86CC) {
1448 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
1449 /// true if Op is undef or if its value falls within the specified range (L, H].
1450 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
1451 if (Op.getOpcode() == ISD::UNDEF)
1454 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
1455 return (Val >= Low && Val < Hi);
1458 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
1459 /// true if Op is undef or if its value equal to the specified value.
1460 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
1461 if (Op.getOpcode() == ISD::UNDEF)
1463 return cast<ConstantSDNode>(Op)->getValue() == Val;
1466 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
1467 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
1468 bool X86::isPSHUFDMask(SDNode *N) {
1469 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1471 if (N->getNumOperands() != 4)
1474 // Check if the value doesn't reference the second vector.
1475 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1476 SDOperand Arg = N->getOperand(i);
1477 if (Arg.getOpcode() == ISD::UNDEF) continue;
1478 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1479 if (cast<ConstantSDNode>(Arg)->getValue() >= 4)
1486 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
1487 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
1488 bool X86::isPSHUFHWMask(SDNode *N) {
1489 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1491 if (N->getNumOperands() != 8)
1494 // Lower quadword copied in order.
1495 for (unsigned i = 0; i != 4; ++i) {
1496 SDOperand Arg = N->getOperand(i);
1497 if (Arg.getOpcode() == ISD::UNDEF) continue;
1498 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1499 if (cast<ConstantSDNode>(Arg)->getValue() != i)
1503 // Upper quadword shuffled.
1504 for (unsigned i = 4; i != 8; ++i) {
1505 SDOperand Arg = N->getOperand(i);
1506 if (Arg.getOpcode() == ISD::UNDEF) continue;
1507 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1508 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1509 if (Val < 4 || Val > 7)
1516 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
1517 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
1518 bool X86::isPSHUFLWMask(SDNode *N) {
1519 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1521 if (N->getNumOperands() != 8)
1524 // Upper quadword copied in order.
1525 for (unsigned i = 4; i != 8; ++i)
1526 if (!isUndefOrEqual(N->getOperand(i), i))
1529 // Lower quadword shuffled.
1530 for (unsigned i = 0; i != 4; ++i)
1531 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
1537 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
1538 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
1539 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
1540 if (NumElems != 2 && NumElems != 4) return false;
1542 unsigned Half = NumElems / 2;
1543 for (unsigned i = 0; i < Half; ++i)
1544 if (!isUndefOrInRange(Elems[i], 0, NumElems))
1546 for (unsigned i = Half; i < NumElems; ++i)
1547 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
1553 bool X86::isSHUFPMask(SDNode *N) {
1554 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1555 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
1558 /// isCommutedSHUFP - Returns true if the shuffle mask is except
1559 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
1560 /// half elements to come from vector 1 (which would equal the dest.) and
1561 /// the upper half to come from vector 2.
1562 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
1563 if (NumOps != 2 && NumOps != 4) return false;
1565 unsigned Half = NumOps / 2;
1566 for (unsigned i = 0; i < Half; ++i)
1567 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
1569 for (unsigned i = Half; i < NumOps; ++i)
1570 if (!isUndefOrInRange(Ops[i], 0, NumOps))
1575 static bool isCommutedSHUFP(SDNode *N) {
1576 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1577 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
1580 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
1581 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
1582 bool X86::isMOVHLPSMask(SDNode *N) {
1583 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1585 if (N->getNumOperands() != 4)
1588 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
1589 return isUndefOrEqual(N->getOperand(0), 6) &&
1590 isUndefOrEqual(N->getOperand(1), 7) &&
1591 isUndefOrEqual(N->getOperand(2), 2) &&
1592 isUndefOrEqual(N->getOperand(3), 3);
1595 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
1596 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
1598 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
1599 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1601 if (N->getNumOperands() != 4)
1604 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
1605 return isUndefOrEqual(N->getOperand(0), 2) &&
1606 isUndefOrEqual(N->getOperand(1), 3) &&
1607 isUndefOrEqual(N->getOperand(2), 2) &&
1608 isUndefOrEqual(N->getOperand(3), 3);
1611 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
1612 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
1613 bool X86::isMOVLPMask(SDNode *N) {
1614 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1616 unsigned NumElems = N->getNumOperands();
1617 if (NumElems != 2 && NumElems != 4)
1620 for (unsigned i = 0; i < NumElems/2; ++i)
1621 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
1624 for (unsigned i = NumElems/2; i < NumElems; ++i)
1625 if (!isUndefOrEqual(N->getOperand(i), i))
1631 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
1632 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
1634 bool X86::isMOVHPMask(SDNode *N) {
1635 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1637 unsigned NumElems = N->getNumOperands();
1638 if (NumElems != 2 && NumElems != 4)
1641 for (unsigned i = 0; i < NumElems/2; ++i)
1642 if (!isUndefOrEqual(N->getOperand(i), i))
1645 for (unsigned i = 0; i < NumElems/2; ++i) {
1646 SDOperand Arg = N->getOperand(i + NumElems/2);
1647 if (!isUndefOrEqual(Arg, i + NumElems))
1654 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
1655 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
1656 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
1657 bool V2IsSplat = false) {
1658 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1661 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
1662 SDOperand BitI = Elts[i];
1663 SDOperand BitI1 = Elts[i+1];
1664 if (!isUndefOrEqual(BitI, j))
1667 if (isUndefOrEqual(BitI1, NumElts))
1670 if (!isUndefOrEqual(BitI1, j + NumElts))
1678 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
1679 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1680 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
1683 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
1684 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
1685 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
1686 bool V2IsSplat = false) {
1687 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1690 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
1691 SDOperand BitI = Elts[i];
1692 SDOperand BitI1 = Elts[i+1];
1693 if (!isUndefOrEqual(BitI, j + NumElts/2))
1696 if (isUndefOrEqual(BitI1, NumElts))
1699 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
1707 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
1708 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1709 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
1712 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
1713 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
1715 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
1716 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1718 unsigned NumElems = N->getNumOperands();
1719 if (NumElems != 4 && NumElems != 8 && NumElems != 16)
1722 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
1723 SDOperand BitI = N->getOperand(i);
1724 SDOperand BitI1 = N->getOperand(i+1);
1726 if (!isUndefOrEqual(BitI, j))
1728 if (!isUndefOrEqual(BitI1, j))
1735 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
1736 /// specifies a shuffle of elements that is suitable for input to MOVSS,
1737 /// MOVSD, and MOVD, i.e. setting the lowest element.
1738 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
1739 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1742 if (!isUndefOrEqual(Elts[0], NumElts))
1745 for (unsigned i = 1; i < NumElts; ++i) {
1746 if (!isUndefOrEqual(Elts[i], i))
1753 bool X86::isMOVLMask(SDNode *N) {
1754 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1755 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
1758 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
1759 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
1760 /// element of vector 2 and the other elements to come from vector 1 in order.
1761 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
1762 bool V2IsSplat = false,
1763 bool V2IsUndef = false) {
1764 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
1767 if (!isUndefOrEqual(Ops[0], 0))
1770 for (unsigned i = 1; i < NumOps; ++i) {
1771 SDOperand Arg = Ops[i];
1772 if (!(isUndefOrEqual(Arg, i+NumOps) ||
1773 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
1774 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
1781 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
1782 bool V2IsUndef = false) {
1783 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1784 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
1785 V2IsSplat, V2IsUndef);
1788 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
1789 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
1790 bool X86::isMOVSHDUPMask(SDNode *N) {
1791 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1793 if (N->getNumOperands() != 4)
1796 // Expect 1, 1, 3, 3
1797 for (unsigned i = 0; i < 2; ++i) {
1798 SDOperand Arg = N->getOperand(i);
1799 if (Arg.getOpcode() == ISD::UNDEF) continue;
1800 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1801 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1802 if (Val != 1) return false;
1806 for (unsigned i = 2; i < 4; ++i) {
1807 SDOperand Arg = N->getOperand(i);
1808 if (Arg.getOpcode() == ISD::UNDEF) continue;
1809 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1810 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1811 if (Val != 3) return false;
1815 // Don't use movshdup if it can be done with a shufps.
1819 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
1820 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
1821 bool X86::isMOVSLDUPMask(SDNode *N) {
1822 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1824 if (N->getNumOperands() != 4)
1827 // Expect 0, 0, 2, 2
1828 for (unsigned i = 0; i < 2; ++i) {
1829 SDOperand Arg = N->getOperand(i);
1830 if (Arg.getOpcode() == ISD::UNDEF) continue;
1831 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1832 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1833 if (Val != 0) return false;
1837 for (unsigned i = 2; i < 4; ++i) {
1838 SDOperand Arg = N->getOperand(i);
1839 if (Arg.getOpcode() == ISD::UNDEF) continue;
1840 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1841 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1842 if (Val != 2) return false;
1846 // Don't use movshdup if it can be done with a shufps.
1850 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
1851 /// a splat of a single element.
1852 static bool isSplatMask(SDNode *N) {
1853 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1855 // This is a splat operation if each element of the permute is the same, and
1856 // if the value doesn't reference the second vector.
1857 unsigned NumElems = N->getNumOperands();
1858 SDOperand ElementBase;
1860 for (; i != NumElems; ++i) {
1861 SDOperand Elt = N->getOperand(i);
1862 if (isa<ConstantSDNode>(Elt)) {
1868 if (!ElementBase.Val)
1871 for (; i != NumElems; ++i) {
1872 SDOperand Arg = N->getOperand(i);
1873 if (Arg.getOpcode() == ISD::UNDEF) continue;
1874 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1875 if (Arg != ElementBase) return false;
1878 // Make sure it is a splat of the first vector operand.
1879 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
1882 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
1883 /// a splat of a single element and it's a 2 or 4 element mask.
1884 bool X86::isSplatMask(SDNode *N) {
1885 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1887 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
1888 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
1890 return ::isSplatMask(N);
1893 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
1894 /// specifies a splat of zero element.
1895 bool X86::isSplatLoMask(SDNode *N) {
1896 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1898 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
1899 if (!isUndefOrEqual(N->getOperand(i), 0))
1904 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
1905 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
1907 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
1908 unsigned NumOperands = N->getNumOperands();
1909 unsigned Shift = (NumOperands == 4) ? 2 : 1;
1911 for (unsigned i = 0; i < NumOperands; ++i) {
1913 SDOperand Arg = N->getOperand(NumOperands-i-1);
1914 if (Arg.getOpcode() != ISD::UNDEF)
1915 Val = cast<ConstantSDNode>(Arg)->getValue();
1916 if (Val >= NumOperands) Val -= NumOperands;
1918 if (i != NumOperands - 1)
1925 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
1926 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
1928 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
1930 // 8 nodes, but we only care about the last 4.
1931 for (unsigned i = 7; i >= 4; --i) {
1933 SDOperand Arg = N->getOperand(i);
1934 if (Arg.getOpcode() != ISD::UNDEF)
1935 Val = cast<ConstantSDNode>(Arg)->getValue();
1944 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
1945 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
1947 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
1949 // 8 nodes, but we only care about the first 4.
1950 for (int i = 3; i >= 0; --i) {
1952 SDOperand Arg = N->getOperand(i);
1953 if (Arg.getOpcode() != ISD::UNDEF)
1954 Val = cast<ConstantSDNode>(Arg)->getValue();
1963 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
1964 /// specifies a 8 element shuffle that can be broken into a pair of
1965 /// PSHUFHW and PSHUFLW.
1966 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
1967 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1969 if (N->getNumOperands() != 8)
1972 // Lower quadword shuffled.
1973 for (unsigned i = 0; i != 4; ++i) {
1974 SDOperand Arg = N->getOperand(i);
1975 if (Arg.getOpcode() == ISD::UNDEF) continue;
1976 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1977 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1982 // Upper quadword shuffled.
1983 for (unsigned i = 4; i != 8; ++i) {
1984 SDOperand Arg = N->getOperand(i);
1985 if (Arg.getOpcode() == ISD::UNDEF) continue;
1986 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1987 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1988 if (Val < 4 || Val > 7)
1995 /// CommuteVectorShuffle - Swap vector_shuffle operandsas well as
1996 /// values in ther permute mask.
1997 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
1998 SDOperand &V2, SDOperand &Mask,
1999 SelectionDAG &DAG) {
2000 MVT::ValueType VT = Op.getValueType();
2001 MVT::ValueType MaskVT = Mask.getValueType();
2002 MVT::ValueType EltVT = MVT::getVectorBaseType(MaskVT);
2003 unsigned NumElems = Mask.getNumOperands();
2004 SmallVector<SDOperand, 8> MaskVec;
2006 for (unsigned i = 0; i != NumElems; ++i) {
2007 SDOperand Arg = Mask.getOperand(i);
2008 if (Arg.getOpcode() == ISD::UNDEF) {
2009 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2012 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2013 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2015 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2017 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2021 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2022 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2025 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2026 /// match movhlps. The lower half elements should come from upper half of
2027 /// V1 (and in order), and the upper half elements should come from the upper
2028 /// half of V2 (and in order).
2029 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2030 unsigned NumElems = Mask->getNumOperands();
2033 for (unsigned i = 0, e = 2; i != e; ++i)
2034 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2036 for (unsigned i = 2; i != 4; ++i)
2037 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2042 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2043 /// is promoted to a vector.
2044 static inline bool isScalarLoadToVector(SDNode *N) {
2045 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2046 N = N->getOperand(0).Val;
2047 return ISD::isNON_EXTLoad(N);
2052 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2053 /// match movlp{s|d}. The lower half elements should come from lower half of
2054 /// V1 (and in order), and the upper half elements should come from the upper
2055 /// half of V2 (and in order). And since V1 will become the source of the
2056 /// MOVLP, it must be either a vector load or a scalar load to vector.
2057 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2058 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2060 // Is V2 is a vector load, don't do this transformation. We will try to use
2061 // load folding shufps op.
2062 if (ISD::isNON_EXTLoad(V2))
2065 unsigned NumElems = Mask->getNumOperands();
2066 if (NumElems != 2 && NumElems != 4)
2068 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2069 if (!isUndefOrEqual(Mask->getOperand(i), i))
2071 for (unsigned i = NumElems/2; i != NumElems; ++i)
2072 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2077 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2079 static bool isSplatVector(SDNode *N) {
2080 if (N->getOpcode() != ISD::BUILD_VECTOR)
2083 SDOperand SplatValue = N->getOperand(0);
2084 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2085 if (N->getOperand(i) != SplatValue)
2090 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2092 static bool isUndefShuffle(SDNode *N) {
2093 if (N->getOpcode() != ISD::BUILD_VECTOR)
2096 SDOperand V1 = N->getOperand(0);
2097 SDOperand V2 = N->getOperand(1);
2098 SDOperand Mask = N->getOperand(2);
2099 unsigned NumElems = Mask.getNumOperands();
2100 for (unsigned i = 0; i != NumElems; ++i) {
2101 SDOperand Arg = Mask.getOperand(i);
2102 if (Arg.getOpcode() != ISD::UNDEF) {
2103 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2104 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2106 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2113 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2114 /// that point to V2 points to its first element.
2115 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2116 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2118 bool Changed = false;
2119 SmallVector<SDOperand, 8> MaskVec;
2120 unsigned NumElems = Mask.getNumOperands();
2121 for (unsigned i = 0; i != NumElems; ++i) {
2122 SDOperand Arg = Mask.getOperand(i);
2123 if (Arg.getOpcode() != ISD::UNDEF) {
2124 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2125 if (Val > NumElems) {
2126 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2130 MaskVec.push_back(Arg);
2134 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2135 &MaskVec[0], MaskVec.size());
2139 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2140 /// operation of specified width.
2141 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2142 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2143 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2145 SmallVector<SDOperand, 8> MaskVec;
2146 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2147 for (unsigned i = 1; i != NumElems; ++i)
2148 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2149 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2152 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2153 /// of specified width.
2154 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2155 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2156 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2157 SmallVector<SDOperand, 8> MaskVec;
2158 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2159 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2160 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2162 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2165 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2166 /// of specified width.
2167 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2168 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2169 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2170 unsigned Half = NumElems/2;
2171 SmallVector<SDOperand, 8> MaskVec;
2172 for (unsigned i = 0; i != Half; ++i) {
2173 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2174 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2176 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2179 /// getZeroVector - Returns a vector of specified type with all zero elements.
2181 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2182 assert(MVT::isVector(VT) && "Expected a vector type");
2183 unsigned NumElems = getVectorNumElements(VT);
2184 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2185 bool isFP = MVT::isFloatingPoint(EVT);
2186 SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT);
2187 SmallVector<SDOperand, 8> ZeroVec(NumElems, Zero);
2188 return DAG.getNode(ISD::BUILD_VECTOR, VT, &ZeroVec[0], ZeroVec.size());
2191 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2193 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2194 SDOperand V1 = Op.getOperand(0);
2195 SDOperand Mask = Op.getOperand(2);
2196 MVT::ValueType VT = Op.getValueType();
2197 unsigned NumElems = Mask.getNumOperands();
2198 Mask = getUnpacklMask(NumElems, DAG);
2199 while (NumElems != 4) {
2200 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2203 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2205 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2206 Mask = getZeroVector(MaskVT, DAG);
2207 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2208 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2209 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2212 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2214 static inline bool isZeroNode(SDOperand Elt) {
2215 return ((isa<ConstantSDNode>(Elt) &&
2216 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2217 (isa<ConstantFPSDNode>(Elt) &&
2218 cast<ConstantFPSDNode>(Elt)->isExactlyValue(0.0)));
2221 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2222 /// vector and zero or undef vector.
2223 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2224 unsigned NumElems, unsigned Idx,
2225 bool isZero, SelectionDAG &DAG) {
2226 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2227 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2228 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2229 SDOperand Zero = DAG.getConstant(0, EVT);
2230 SmallVector<SDOperand, 8> MaskVec(NumElems, Zero);
2231 MaskVec[Idx] = DAG.getConstant(NumElems, EVT);
2232 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2233 &MaskVec[0], MaskVec.size());
2234 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2237 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2239 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
2240 unsigned NumNonZero, unsigned NumZero,
2241 SelectionDAG &DAG, TargetLowering &TLI) {
2247 for (unsigned i = 0; i < 16; ++i) {
2248 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
2249 if (ThisIsNonZero && First) {
2251 V = getZeroVector(MVT::v8i16, DAG);
2253 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2258 SDOperand ThisElt(0, 0), LastElt(0, 0);
2259 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
2260 if (LastIsNonZero) {
2261 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
2263 if (ThisIsNonZero) {
2264 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
2265 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
2266 ThisElt, DAG.getConstant(8, MVT::i8));
2268 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
2273 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
2274 DAG.getConstant(i/2, TLI.getPointerTy()));
2278 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
2281 /// LowerBuildVectorv16i8 - Custom lower build_vector of v8i16.
2283 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
2284 unsigned NumNonZero, unsigned NumZero,
2285 SelectionDAG &DAG, TargetLowering &TLI) {
2291 for (unsigned i = 0; i < 8; ++i) {
2292 bool isNonZero = (NonZeros & (1 << i)) != 0;
2296 V = getZeroVector(MVT::v8i16, DAG);
2298 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2301 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
2302 DAG.getConstant(i, TLI.getPointerTy()));
2310 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2311 // All zero's are handled with pxor.
2312 if (ISD::isBuildVectorAllZeros(Op.Val))
2315 // All one's are handled with pcmpeqd.
2316 if (ISD::isBuildVectorAllOnes(Op.Val))
2319 MVT::ValueType VT = Op.getValueType();
2320 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2321 unsigned EVTBits = MVT::getSizeInBits(EVT);
2323 unsigned NumElems = Op.getNumOperands();
2324 unsigned NumZero = 0;
2325 unsigned NumNonZero = 0;
2326 unsigned NonZeros = 0;
2327 std::set<SDOperand> Values;
2328 for (unsigned i = 0; i < NumElems; ++i) {
2329 SDOperand Elt = Op.getOperand(i);
2330 if (Elt.getOpcode() != ISD::UNDEF) {
2332 if (isZeroNode(Elt))
2335 NonZeros |= (1 << i);
2341 if (NumNonZero == 0)
2342 // Must be a mix of zero and undef. Return a zero vector.
2343 return getZeroVector(VT, DAG);
2345 // Splat is obviously ok. Let legalizer expand it to a shuffle.
2346 if (Values.size() == 1)
2349 // Special case for single non-zero element.
2350 if (NumNonZero == 1) {
2351 unsigned Idx = CountTrailingZeros_32(NonZeros);
2352 SDOperand Item = Op.getOperand(Idx);
2353 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
2355 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
2356 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
2359 if (EVTBits == 32) {
2360 // Turn it into a shuffle of zero and zero-extended scalar to vector.
2361 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
2363 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2364 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
2365 SmallVector<SDOperand, 8> MaskVec;
2366 for (unsigned i = 0; i < NumElems; i++)
2367 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
2368 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2369 &MaskVec[0], MaskVec.size());
2370 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
2371 DAG.getNode(ISD::UNDEF, VT), Mask);
2375 // Let legalizer expand 2-wide build_vector's.
2379 // If element VT is < 32 bits, convert it to inserts into a zero vector.
2381 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
2383 if (V.Val) return V;
2386 if (EVTBits == 16) {
2387 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
2389 if (V.Val) return V;
2392 // If element VT is == 32 bits, turn it into a number of shuffles.
2393 SmallVector<SDOperand, 8> V;
2395 if (NumElems == 4 && NumZero > 0) {
2396 for (unsigned i = 0; i < 4; ++i) {
2397 bool isZero = !(NonZeros & (1 << i));
2399 V[i] = getZeroVector(VT, DAG);
2401 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2404 for (unsigned i = 0; i < 2; ++i) {
2405 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
2408 V[i] = V[i*2]; // Must be a zero vector.
2411 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
2412 getMOVLMask(NumElems, DAG));
2415 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2416 getMOVLMask(NumElems, DAG));
2419 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2420 getUnpacklMask(NumElems, DAG));
2425 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
2426 // clears the upper bits.
2427 // FIXME: we can do the same for v4f32 case when we know both parts of
2428 // the lower half come from scalar_to_vector (loadf32). We should do
2429 // that in post legalizer dag combiner with target specific hooks.
2430 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
2432 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2433 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2434 SmallVector<SDOperand, 8> MaskVec;
2435 bool Reverse = (NonZeros & 0x3) == 2;
2436 for (unsigned i = 0; i < 2; ++i)
2438 MaskVec.push_back(DAG.getConstant(1-i, EVT));
2440 MaskVec.push_back(DAG.getConstant(i, EVT));
2441 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
2442 for (unsigned i = 0; i < 2; ++i)
2444 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
2446 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
2447 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2448 &MaskVec[0], MaskVec.size());
2449 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
2452 if (Values.size() > 2) {
2453 // Expand into a number of unpckl*.
2455 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
2456 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
2457 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
2458 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
2459 for (unsigned i = 0; i < NumElems; ++i)
2460 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2462 while (NumElems != 0) {
2463 for (unsigned i = 0; i < NumElems; ++i)
2464 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
2475 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
2476 SDOperand V1 = Op.getOperand(0);
2477 SDOperand V2 = Op.getOperand(1);
2478 SDOperand PermMask = Op.getOperand(2);
2479 MVT::ValueType VT = Op.getValueType();
2480 unsigned NumElems = PermMask.getNumOperands();
2481 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
2482 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
2483 bool V1IsSplat = false;
2484 bool V2IsSplat = false;
2486 if (isUndefShuffle(Op.Val))
2487 return DAG.getNode(ISD::UNDEF, VT);
2489 if (isSplatMask(PermMask.Val)) {
2490 if (NumElems <= 4) return Op;
2491 // Promote it to a v4i32 splat.
2492 return PromoteSplat(Op, DAG);
2495 if (X86::isMOVLMask(PermMask.Val))
2496 return (V1IsUndef) ? V2 : Op;
2498 if (X86::isMOVSHDUPMask(PermMask.Val) ||
2499 X86::isMOVSLDUPMask(PermMask.Val) ||
2500 X86::isMOVHLPSMask(PermMask.Val) ||
2501 X86::isMOVHPMask(PermMask.Val) ||
2502 X86::isMOVLPMask(PermMask.Val))
2505 if (ShouldXformToMOVHLPS(PermMask.Val) ||
2506 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
2507 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2509 bool Commuted = false;
2510 V1IsSplat = isSplatVector(V1.Val);
2511 V2IsSplat = isSplatVector(V2.Val);
2512 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
2513 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2514 std::swap(V1IsSplat, V2IsSplat);
2515 std::swap(V1IsUndef, V2IsUndef);
2519 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
2520 if (V2IsUndef) return V1;
2521 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2523 // V2 is a splat, so the mask may be malformed. That is, it may point
2524 // to any V2 element. The instruction selectior won't like this. Get
2525 // a corrected mask and commute to form a proper MOVS{S|D}.
2526 SDOperand NewMask = getMOVLMask(NumElems, DAG);
2527 if (NewMask.Val != PermMask.Val)
2528 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2533 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
2534 X86::isUNPCKLMask(PermMask.Val) ||
2535 X86::isUNPCKHMask(PermMask.Val))
2539 // Normalize mask so all entries that point to V2 points to its first
2540 // element then try to match unpck{h|l} again. If match, return a
2541 // new vector_shuffle with the corrected mask.
2542 SDOperand NewMask = NormalizeMask(PermMask, DAG);
2543 if (NewMask.Val != PermMask.Val) {
2544 if (X86::isUNPCKLMask(PermMask.Val, true)) {
2545 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
2546 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2547 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
2548 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
2549 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2554 // Normalize the node to match x86 shuffle ops if needed
2555 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
2556 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2559 // Commute is back and try unpck* again.
2560 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2561 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
2562 X86::isUNPCKLMask(PermMask.Val) ||
2563 X86::isUNPCKHMask(PermMask.Val))
2567 // If VT is integer, try PSHUF* first, then SHUFP*.
2568 if (MVT::isInteger(VT)) {
2569 if (X86::isPSHUFDMask(PermMask.Val) ||
2570 X86::isPSHUFHWMask(PermMask.Val) ||
2571 X86::isPSHUFLWMask(PermMask.Val)) {
2572 if (V2.getOpcode() != ISD::UNDEF)
2573 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
2574 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
2578 if (X86::isSHUFPMask(PermMask.Val))
2581 // Handle v8i16 shuffle high / low shuffle node pair.
2582 if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
2583 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2584 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2585 SmallVector<SDOperand, 8> MaskVec;
2586 for (unsigned i = 0; i != 4; ++i)
2587 MaskVec.push_back(PermMask.getOperand(i));
2588 for (unsigned i = 4; i != 8; ++i)
2589 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2590 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2591 &MaskVec[0], MaskVec.size());
2592 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2594 for (unsigned i = 0; i != 4; ++i)
2595 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2596 for (unsigned i = 4; i != 8; ++i)
2597 MaskVec.push_back(PermMask.getOperand(i));
2598 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0],MaskVec.size());
2599 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2602 // Floating point cases in the other order.
2603 if (X86::isSHUFPMask(PermMask.Val))
2605 if (X86::isPSHUFDMask(PermMask.Val) ||
2606 X86::isPSHUFHWMask(PermMask.Val) ||
2607 X86::isPSHUFLWMask(PermMask.Val)) {
2608 if (V2.getOpcode() != ISD::UNDEF)
2609 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
2610 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
2615 if (NumElems == 4) {
2616 MVT::ValueType MaskVT = PermMask.getValueType();
2617 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
2618 SmallVector<std::pair<int, int>, 8> Locs;
2619 Locs.reserve(NumElems);
2620 SmallVector<SDOperand, 8> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2621 SmallVector<SDOperand, 8> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2624 // If no more than two elements come from either vector. This can be
2625 // implemented with two shuffles. First shuffle gather the elements.
2626 // The second shuffle, which takes the first shuffle as both of its
2627 // vector operands, put the elements into the right order.
2628 for (unsigned i = 0; i != NumElems; ++i) {
2629 SDOperand Elt = PermMask.getOperand(i);
2630 if (Elt.getOpcode() == ISD::UNDEF) {
2631 Locs[i] = std::make_pair(-1, -1);
2633 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
2634 if (Val < NumElems) {
2635 Locs[i] = std::make_pair(0, NumLo);
2639 Locs[i] = std::make_pair(1, NumHi);
2640 if (2+NumHi < NumElems)
2641 Mask1[2+NumHi] = Elt;
2646 if (NumLo <= 2 && NumHi <= 2) {
2647 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2648 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2649 &Mask1[0], Mask1.size()));
2650 for (unsigned i = 0; i != NumElems; ++i) {
2651 if (Locs[i].first == -1)
2654 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
2655 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
2656 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
2660 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
2661 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2662 &Mask2[0], Mask2.size()));
2665 // Break it into (shuffle shuffle_hi, shuffle_lo).
2667 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2668 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2669 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
2670 unsigned MaskIdx = 0;
2672 unsigned HiIdx = NumElems/2;
2673 for (unsigned i = 0; i != NumElems; ++i) {
2674 if (i == NumElems/2) {
2680 SDOperand Elt = PermMask.getOperand(i);
2681 if (Elt.getOpcode() == ISD::UNDEF) {
2682 Locs[i] = std::make_pair(-1, -1);
2683 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
2684 Locs[i] = std::make_pair(MaskIdx, LoIdx);
2685 (*MaskPtr)[LoIdx] = Elt;
2688 Locs[i] = std::make_pair(MaskIdx, HiIdx);
2689 (*MaskPtr)[HiIdx] = Elt;
2694 SDOperand LoShuffle =
2695 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2696 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2697 &LoMask[0], LoMask.size()));
2698 SDOperand HiShuffle =
2699 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2700 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2701 &HiMask[0], HiMask.size()));
2702 SmallVector<SDOperand, 8> MaskOps;
2703 for (unsigned i = 0; i != NumElems; ++i) {
2704 if (Locs[i].first == -1) {
2705 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
2707 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
2708 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
2711 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
2712 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2713 &MaskOps[0], MaskOps.size()));
2720 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2721 if (!isa<ConstantSDNode>(Op.getOperand(1)))
2724 MVT::ValueType VT = Op.getValueType();
2725 // TODO: handle v16i8.
2726 if (MVT::getSizeInBits(VT) == 16) {
2727 // Transform it so it match pextrw which produces a 32-bit result.
2728 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
2729 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
2730 Op.getOperand(0), Op.getOperand(1));
2731 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
2732 DAG.getValueType(VT));
2733 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
2734 } else if (MVT::getSizeInBits(VT) == 32) {
2735 SDOperand Vec = Op.getOperand(0);
2736 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
2739 // SHUFPS the element to the lowest double word, then movss.
2740 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2741 SmallVector<SDOperand, 8> IdxVec;
2742 IdxVec.push_back(DAG.getConstant(Idx, 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 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2746 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2747 &IdxVec[0], IdxVec.size());
2748 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
2749 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
2750 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
2751 DAG.getConstant(0, getPointerTy()));
2752 } else if (MVT::getSizeInBits(VT) == 64) {
2753 SDOperand Vec = Op.getOperand(0);
2754 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
2758 // UNPCKHPD the element to the lowest double word, then movsd.
2759 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
2760 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
2761 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2762 SmallVector<SDOperand, 8> IdxVec;
2763 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorBaseType(MaskVT)));
2764 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2765 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2766 &IdxVec[0], IdxVec.size());
2767 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
2768 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
2769 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
2770 DAG.getConstant(0, getPointerTy()));
2777 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2778 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
2779 // as its second argument.
2780 MVT::ValueType VT = Op.getValueType();
2781 MVT::ValueType BaseVT = MVT::getVectorBaseType(VT);
2782 SDOperand N0 = Op.getOperand(0);
2783 SDOperand N1 = Op.getOperand(1);
2784 SDOperand N2 = Op.getOperand(2);
2785 if (MVT::getSizeInBits(BaseVT) == 16) {
2786 if (N1.getValueType() != MVT::i32)
2787 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
2788 if (N2.getValueType() != MVT::i32)
2789 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(), MVT::i32);
2790 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
2791 } else if (MVT::getSizeInBits(BaseVT) == 32) {
2792 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
2795 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
2796 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2797 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2798 SmallVector<SDOperand, 8> MaskVec;
2799 MaskVec.push_back(DAG.getConstant(4, BaseVT));
2800 for (unsigned i = 1; i <= 3; ++i)
2801 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2802 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
2803 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2804 &MaskVec[0], MaskVec.size()));
2806 // Use two pinsrw instructions to insert a 32 bit value.
2808 if (MVT::isFloatingPoint(N1.getValueType())) {
2809 if (ISD::isNON_EXTLoad(N1.Val)) {
2810 // Just load directly from f32mem to GR32.
2811 LoadSDNode *LD = cast<LoadSDNode>(N1);
2812 N1 = DAG.getLoad(MVT::i32, LD->getChain(), LD->getBasePtr(),
2813 LD->getSrcValue(), LD->getSrcValueOffset());
2815 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
2816 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
2817 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
2818 DAG.getConstant(0, getPointerTy()));
2821 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
2822 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
2823 DAG.getConstant(Idx, getPointerTy()));
2824 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
2825 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
2826 DAG.getConstant(Idx+1, getPointerTy()));
2827 return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
2835 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2836 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
2837 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
2840 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2841 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
2842 // one of the above mentioned nodes. It has to be wrapped because otherwise
2843 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2844 // be used to form addressing mode. These wrapped nodes will be selected
2847 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
2848 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2849 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
2851 CP->getAlignment());
2852 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
2853 // With PIC, the address is actually $g + Offset.
2854 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2855 !Subtarget->isPICStyleRIPRel()) {
2856 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2857 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
2865 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
2866 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2867 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
2868 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
2869 // With PIC, the address is actually $g + Offset.
2870 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2871 !Subtarget->isPICStyleRIPRel()) {
2872 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2873 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
2877 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
2878 // load the value at address GV, not the value of GV itself. This means that
2879 // the GlobalAddress must be in the base or index register of the address, not
2880 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
2881 // The same applies for external symbols during PIC codegen
2882 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
2883 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
2889 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
2890 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
2891 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
2892 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
2893 // With PIC, the address is actually $g + Offset.
2894 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2895 !Subtarget->isPICStyleRIPRel()) {
2896 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2897 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
2904 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
2905 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
2906 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
2907 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
2908 // With PIC, the address is actually $g + Offset.
2909 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2910 !Subtarget->isPICStyleRIPRel()) {
2911 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2912 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
2919 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
2920 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
2921 "Not an i64 shift!");
2922 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
2923 SDOperand ShOpLo = Op.getOperand(0);
2924 SDOperand ShOpHi = Op.getOperand(1);
2925 SDOperand ShAmt = Op.getOperand(2);
2926 SDOperand Tmp1 = isSRA ?
2927 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
2928 DAG.getConstant(0, MVT::i32);
2930 SDOperand Tmp2, Tmp3;
2931 if (Op.getOpcode() == ISD::SHL_PARTS) {
2932 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
2933 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
2935 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
2936 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
2939 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
2940 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
2941 DAG.getConstant(32, MVT::i8));
2942 SDOperand COps[]={DAG.getEntryNode(), AndNode, DAG.getConstant(0, MVT::i8)};
2943 SDOperand InFlag = DAG.getNode(X86ISD::CMP, VTs, 2, COps, 3).getValue(1);
2946 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
2948 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
2949 SmallVector<SDOperand, 4> Ops;
2950 if (Op.getOpcode() == ISD::SHL_PARTS) {
2951 Ops.push_back(Tmp2);
2952 Ops.push_back(Tmp3);
2954 Ops.push_back(InFlag);
2955 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
2956 InFlag = Hi.getValue(1);
2959 Ops.push_back(Tmp3);
2960 Ops.push_back(Tmp1);
2962 Ops.push_back(InFlag);
2963 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
2965 Ops.push_back(Tmp2);
2966 Ops.push_back(Tmp3);
2968 Ops.push_back(InFlag);
2969 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
2970 InFlag = Lo.getValue(1);
2973 Ops.push_back(Tmp3);
2974 Ops.push_back(Tmp1);
2976 Ops.push_back(InFlag);
2977 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
2980 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
2984 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
2987 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
2988 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
2989 Op.getOperand(0).getValueType() >= MVT::i16 &&
2990 "Unknown SINT_TO_FP to lower!");
2993 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
2994 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
2995 MachineFunction &MF = DAG.getMachineFunction();
2996 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
2997 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
2998 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
2999 StackSlot, NULL, 0);
3004 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
3006 Tys = DAG.getVTList(MVT::f64, MVT::Other);
3007 SmallVector<SDOperand, 8> Ops;
3008 Ops.push_back(Chain);
3009 Ops.push_back(StackSlot);
3010 Ops.push_back(DAG.getValueType(SrcVT));
3011 Result = DAG.getNode(X86ScalarSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
3012 Tys, &Ops[0], Ops.size());
3015 Chain = Result.getValue(1);
3016 SDOperand InFlag = Result.getValue(2);
3018 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
3019 // shouldn't be necessary except that RFP cannot be live across
3020 // multiple blocks. When stackifier is fixed, they can be uncoupled.
3021 MachineFunction &MF = DAG.getMachineFunction();
3022 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3023 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3024 Tys = DAG.getVTList(MVT::Other);
3025 SmallVector<SDOperand, 8> Ops;
3026 Ops.push_back(Chain);
3027 Ops.push_back(Result);
3028 Ops.push_back(StackSlot);
3029 Ops.push_back(DAG.getValueType(Op.getValueType()));
3030 Ops.push_back(InFlag);
3031 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
3032 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
3038 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
3039 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
3040 "Unknown FP_TO_SINT to lower!");
3041 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
3043 MachineFunction &MF = DAG.getMachineFunction();
3044 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
3045 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3046 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3049 switch (Op.getValueType()) {
3050 default: assert(0 && "Invalid FP_TO_SINT to lower!");
3051 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
3052 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
3053 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
3056 SDOperand Chain = DAG.getEntryNode();
3057 SDOperand Value = Op.getOperand(0);
3059 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
3060 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
3061 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
3063 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
3065 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
3066 Chain = Value.getValue(1);
3067 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3068 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3071 // Build the FP_TO_INT*_IN_MEM
3072 SDOperand Ops[] = { Chain, Value, StackSlot };
3073 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
3076 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
3079 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
3080 MVT::ValueType VT = Op.getValueType();
3081 const Type *OpNTy = MVT::getTypeForValueType(VT);
3082 std::vector<Constant*> CV;
3083 if (VT == MVT::f64) {
3084 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(~(1ULL << 63))));
3085 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3087 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(~(1U << 31))));
3088 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3089 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3090 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3092 Constant *CS = ConstantStruct::get(CV);
3093 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3094 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3095 SmallVector<SDOperand, 3> Ops;
3096 Ops.push_back(DAG.getEntryNode());
3097 Ops.push_back(CPIdx);
3098 Ops.push_back(DAG.getSrcValue(NULL));
3099 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3100 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
3103 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
3104 MVT::ValueType VT = Op.getValueType();
3105 const Type *OpNTy = MVT::getTypeForValueType(VT);
3106 std::vector<Constant*> CV;
3107 if (VT == MVT::f64) {
3108 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(1ULL << 63)));
3109 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3111 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(1U << 31)));
3112 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3113 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3114 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3116 Constant *CS = ConstantStruct::get(CV);
3117 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3118 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3119 SmallVector<SDOperand, 3> Ops;
3120 Ops.push_back(DAG.getEntryNode());
3121 Ops.push_back(CPIdx);
3122 Ops.push_back(DAG.getSrcValue(NULL));
3123 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3124 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
3127 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
3128 SDOperand Op0 = Op.getOperand(0);
3129 SDOperand Op1 = Op.getOperand(1);
3130 MVT::ValueType VT = Op.getValueType();
3131 MVT::ValueType SrcVT = Op1.getValueType();
3132 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
3134 // If second operand is smaller, extend it first.
3135 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
3136 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
3140 // First get the sign bit of second operand.
3141 std::vector<Constant*> CV;
3142 if (SrcVT == MVT::f64) {
3143 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(1ULL << 63)));
3144 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3146 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(1U << 31)));
3147 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3148 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3149 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3151 Constant *CS = ConstantStruct::get(CV);
3152 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3153 SDVTList Tys = DAG.getVTList(SrcVT, MVT::Other);
3154 SmallVector<SDOperand, 3> Ops;
3155 Ops.push_back(DAG.getEntryNode());
3156 Ops.push_back(CPIdx);
3157 Ops.push_back(DAG.getSrcValue(NULL));
3158 SDOperand Mask1 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3159 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
3161 // Shift sign bit right or left if the two operands have different types.
3162 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
3163 // Op0 is MVT::f32, Op1 is MVT::f64.
3164 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
3165 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
3166 DAG.getConstant(32, MVT::i32));
3167 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
3168 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
3169 DAG.getConstant(0, getPointerTy()));
3172 // Clear first operand sign bit.
3174 if (VT == MVT::f64) {
3175 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(~(1ULL << 63))));
3176 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3178 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(~(1U << 31))));
3179 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3180 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3181 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3183 CS = ConstantStruct::get(CV);
3184 CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3185 Tys = DAG.getVTList(VT, MVT::Other);
3187 Ops.push_back(DAG.getEntryNode());
3188 Ops.push_back(CPIdx);
3189 Ops.push_back(DAG.getSrcValue(NULL));
3190 SDOperand Mask2 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3191 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
3193 // Or the value with the sign bit.
3194 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
3197 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG,
3199 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
3201 SDOperand Op0 = Op.getOperand(0);
3202 SDOperand Op1 = Op.getOperand(1);
3203 SDOperand CC = Op.getOperand(2);
3204 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3205 const MVT::ValueType *VTs1 = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3206 const MVT::ValueType *VTs2 = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
3207 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
3210 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
3212 SDOperand Ops1[] = { Chain, Op0, Op1 };
3213 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, Ops1, 3).getValue(1);
3214 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
3215 return DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3218 assert(isFP && "Illegal integer SetCC!");
3220 SDOperand COps[] = { Chain, Op0, Op1 };
3221 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, COps, 3).getValue(1);
3223 switch (SetCCOpcode) {
3224 default: assert(false && "Illegal floating point SetCC!");
3225 case ISD::SETOEQ: { // !PF & ZF
3226 SDOperand Ops1[] = { DAG.getConstant(X86::COND_NP, MVT::i8), Cond };
3227 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3228 SDOperand Ops2[] = { DAG.getConstant(X86::COND_E, MVT::i8),
3230 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3231 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
3233 case ISD::SETUNE: { // PF | !ZF
3234 SDOperand Ops1[] = { DAG.getConstant(X86::COND_P, MVT::i8), Cond };
3235 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3236 SDOperand Ops2[] = { DAG.getConstant(X86::COND_NE, MVT::i8),
3238 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3239 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
3244 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
3245 bool addTest = true;
3246 SDOperand Chain = DAG.getEntryNode();
3247 SDOperand Cond = Op.getOperand(0);
3249 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3251 if (Cond.getOpcode() == ISD::SETCC)
3252 Cond = LowerSETCC(Cond, DAG, Chain);
3254 if (Cond.getOpcode() == X86ISD::SETCC) {
3255 CC = Cond.getOperand(0);
3257 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3258 // (since flag operand cannot be shared). Use it as the condition setting
3259 // operand in place of the X86ISD::SETCC.
3260 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3261 // to use a test instead of duplicating the X86ISD::CMP (for register
3262 // pressure reason)?
3263 SDOperand Cmp = Cond.getOperand(1);
3264 unsigned Opc = Cmp.getOpcode();
3265 bool IllegalFPCMov = !X86ScalarSSE &&
3266 MVT::isFloatingPoint(Op.getValueType()) &&
3267 !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
3268 if ((Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) &&
3270 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3271 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3277 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3278 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3279 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3282 VTs = DAG.getNodeValueTypes(Op.getValueType(), MVT::Flag);
3283 SmallVector<SDOperand, 4> Ops;
3284 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
3285 // condition is true.
3286 Ops.push_back(Op.getOperand(2));
3287 Ops.push_back(Op.getOperand(1));
3289 Ops.push_back(Cond.getValue(1));
3290 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3293 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
3294 bool addTest = true;
3295 SDOperand Chain = Op.getOperand(0);
3296 SDOperand Cond = Op.getOperand(1);
3297 SDOperand Dest = Op.getOperand(2);
3299 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3301 if (Cond.getOpcode() == ISD::SETCC)
3302 Cond = LowerSETCC(Cond, DAG, Chain);
3304 if (Cond.getOpcode() == X86ISD::SETCC) {
3305 CC = Cond.getOperand(0);
3307 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3308 // (since flag operand cannot be shared). Use it as the condition setting
3309 // operand in place of the X86ISD::SETCC.
3310 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3311 // to use a test instead of duplicating the X86ISD::CMP (for register
3312 // pressure reason)?
3313 SDOperand Cmp = Cond.getOperand(1);
3314 unsigned Opc = Cmp.getOpcode();
3315 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) {
3316 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3317 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3323 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3324 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3325 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3327 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
3328 Cond, Op.getOperand(2), CC, Cond.getValue(1));
3331 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
3332 unsigned CallingConv= cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3334 if (Subtarget->is64Bit())
3335 return LowerX86_64CCCCallTo(Op, DAG, CallingConv);
3337 switch (CallingConv) {
3339 assert(0 && "Unsupported calling convention");
3340 case CallingConv::Fast:
3341 // TODO: Implement fastcc
3343 case CallingConv::C:
3344 case CallingConv::X86_StdCall:
3345 return LowerCCCCallTo(Op, DAG, CallingConv);
3346 case CallingConv::X86_FastCall:
3347 return LowerFastCCCallTo(Op, DAG, CallingConv);
3352 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
3353 MachineFunction &MF = DAG.getMachineFunction();
3354 const Function* Fn = MF.getFunction();
3355 if (Fn->hasExternalLinkage() &&
3356 Subtarget->isTargetCygMing() &&
3357 Fn->getName() == "main")
3358 MF.getInfo<X86FunctionInfo>()->setForceFramePointer(true);
3360 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3361 if (Subtarget->is64Bit())
3362 return LowerX86_64CCCArguments(Op, DAG);
3366 assert(0 && "Unsupported calling convention");
3367 case CallingConv::Fast:
3368 // TODO: implement fastcc.
3371 case CallingConv::C:
3372 return LowerCCCArguments(Op, DAG);
3373 case CallingConv::X86_StdCall:
3374 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(StdCall);
3375 return LowerCCCArguments(Op, DAG, true);
3376 case CallingConv::X86_FastCall:
3377 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(FastCall);
3378 return LowerFastCCArguments(Op, DAG);
3382 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
3383 SDOperand InFlag(0, 0);
3384 SDOperand Chain = Op.getOperand(0);
3386 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3387 if (Align == 0) Align = 1;
3389 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3390 // If not DWORD aligned, call memset if size is less than the threshold.
3391 // It knows how to align to the right boundary first.
3392 if ((Align & 3) != 0 ||
3393 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
3394 MVT::ValueType IntPtr = getPointerTy();
3395 const Type *IntPtrTy = getTargetData()->getIntPtrType();
3396 TargetLowering::ArgListTy Args;
3397 TargetLowering::ArgListEntry Entry;
3398 Entry.Node = Op.getOperand(1);
3399 Entry.Ty = IntPtrTy;
3400 Entry.isSigned = false;
3401 Entry.isInReg = false;
3402 Entry.isSRet = false;
3403 Args.push_back(Entry);
3404 // Extend the unsigned i8 argument to be an int value for the call.
3405 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
3406 Entry.Ty = IntPtrTy;
3407 Entry.isSigned = false;
3408 Entry.isInReg = false;
3409 Entry.isSRet = false;
3410 Args.push_back(Entry);
3411 Entry.Node = Op.getOperand(3);
3412 Args.push_back(Entry);
3413 std::pair<SDOperand,SDOperand> CallResult =
3414 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
3415 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
3416 return CallResult.second;
3421 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
3422 unsigned BytesLeft = 0;
3423 bool TwoRepStos = false;
3426 uint64_t Val = ValC->getValue() & 255;
3428 // If the value is a constant, then we can potentially use larger sets.
3429 switch (Align & 3) {
3430 case 2: // WORD aligned
3433 Val = (Val << 8) | Val;
3435 case 0: // DWORD aligned
3438 Val = (Val << 8) | Val;
3439 Val = (Val << 16) | Val;
3440 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
3443 Val = (Val << 32) | Val;
3446 default: // Byte aligned
3449 Count = Op.getOperand(3);
3453 if (AVT > MVT::i8) {
3455 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
3456 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
3457 BytesLeft = I->getValue() % UBytes;
3459 assert(AVT >= MVT::i32 &&
3460 "Do not use rep;stos if not at least DWORD aligned");
3461 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
3462 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
3467 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
3469 InFlag = Chain.getValue(1);
3472 Count = Op.getOperand(3);
3473 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
3474 InFlag = Chain.getValue(1);
3477 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
3479 InFlag = Chain.getValue(1);
3480 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
3481 Op.getOperand(1), InFlag);
3482 InFlag = Chain.getValue(1);
3484 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3485 SmallVector<SDOperand, 8> Ops;
3486 Ops.push_back(Chain);
3487 Ops.push_back(DAG.getValueType(AVT));
3488 Ops.push_back(InFlag);
3489 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
3492 InFlag = Chain.getValue(1);
3493 Count = Op.getOperand(3);
3494 MVT::ValueType CVT = Count.getValueType();
3495 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
3496 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
3497 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
3499 InFlag = Chain.getValue(1);
3500 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3502 Ops.push_back(Chain);
3503 Ops.push_back(DAG.getValueType(MVT::i8));
3504 Ops.push_back(InFlag);
3505 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
3506 } else if (BytesLeft) {
3507 // Issue stores for the last 1 - 7 bytes.
3509 unsigned Val = ValC->getValue() & 255;
3510 unsigned Offset = I->getValue() - BytesLeft;
3511 SDOperand DstAddr = Op.getOperand(1);
3512 MVT::ValueType AddrVT = DstAddr.getValueType();
3513 if (BytesLeft >= 4) {
3514 Val = (Val << 8) | Val;
3515 Val = (Val << 16) | Val;
3516 Value = DAG.getConstant(Val, MVT::i32);
3517 Chain = DAG.getStore(Chain, Value,
3518 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3519 DAG.getConstant(Offset, AddrVT)),
3524 if (BytesLeft >= 2) {
3525 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
3526 Chain = DAG.getStore(Chain, Value,
3527 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3528 DAG.getConstant(Offset, AddrVT)),
3533 if (BytesLeft == 1) {
3534 Value = DAG.getConstant(Val, MVT::i8);
3535 Chain = DAG.getStore(Chain, Value,
3536 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3537 DAG.getConstant(Offset, AddrVT)),
3545 SDOperand X86TargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) {
3546 SDOperand Chain = Op.getOperand(0);
3548 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3549 if (Align == 0) Align = 1;
3551 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3552 // If not DWORD aligned, call memcpy if size is less than the threshold.
3553 // It knows how to align to the right boundary first.
3554 if ((Align & 3) != 0 ||
3555 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
3556 MVT::ValueType IntPtr = getPointerTy();
3557 TargetLowering::ArgListTy Args;
3558 TargetLowering::ArgListEntry Entry;
3559 Entry.Ty = getTargetData()->getIntPtrType();
3560 Entry.isSigned = false;
3561 Entry.isInReg = false;
3562 Entry.isSRet = false;
3563 Entry.Node = Op.getOperand(1); Args.push_back(Entry);
3564 Entry.Node = Op.getOperand(2); Args.push_back(Entry);
3565 Entry.Node = Op.getOperand(3); Args.push_back(Entry);
3566 std::pair<SDOperand,SDOperand> CallResult =
3567 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
3568 DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
3569 return CallResult.second;
3574 unsigned BytesLeft = 0;
3575 bool TwoRepMovs = false;
3576 switch (Align & 3) {
3577 case 2: // WORD aligned
3580 case 0: // DWORD aligned
3582 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
3585 default: // Byte aligned
3587 Count = Op.getOperand(3);
3591 if (AVT > MVT::i8) {
3593 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
3594 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
3595 BytesLeft = I->getValue() % UBytes;
3597 assert(AVT >= MVT::i32 &&
3598 "Do not use rep;movs if not at least DWORD aligned");
3599 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
3600 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
3605 SDOperand InFlag(0, 0);
3606 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
3608 InFlag = Chain.getValue(1);
3609 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
3610 Op.getOperand(1), InFlag);
3611 InFlag = Chain.getValue(1);
3612 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
3613 Op.getOperand(2), InFlag);
3614 InFlag = Chain.getValue(1);
3616 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3617 SmallVector<SDOperand, 8> Ops;
3618 Ops.push_back(Chain);
3619 Ops.push_back(DAG.getValueType(AVT));
3620 Ops.push_back(InFlag);
3621 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
3624 InFlag = Chain.getValue(1);
3625 Count = Op.getOperand(3);
3626 MVT::ValueType CVT = Count.getValueType();
3627 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
3628 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
3629 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
3631 InFlag = Chain.getValue(1);
3632 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3634 Ops.push_back(Chain);
3635 Ops.push_back(DAG.getValueType(MVT::i8));
3636 Ops.push_back(InFlag);
3637 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
3638 } else if (BytesLeft) {
3639 // Issue loads and stores for the last 1 - 7 bytes.
3640 unsigned Offset = I->getValue() - BytesLeft;
3641 SDOperand DstAddr = Op.getOperand(1);
3642 MVT::ValueType DstVT = DstAddr.getValueType();
3643 SDOperand SrcAddr = Op.getOperand(2);
3644 MVT::ValueType SrcVT = SrcAddr.getValueType();
3646 if (BytesLeft >= 4) {
3647 Value = DAG.getLoad(MVT::i32, Chain,
3648 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3649 DAG.getConstant(Offset, SrcVT)),
3651 Chain = Value.getValue(1);
3652 Chain = DAG.getStore(Chain, Value,
3653 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3654 DAG.getConstant(Offset, DstVT)),
3659 if (BytesLeft >= 2) {
3660 Value = DAG.getLoad(MVT::i16, Chain,
3661 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3662 DAG.getConstant(Offset, SrcVT)),
3664 Chain = Value.getValue(1);
3665 Chain = DAG.getStore(Chain, Value,
3666 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3667 DAG.getConstant(Offset, DstVT)),
3673 if (BytesLeft == 1) {
3674 Value = DAG.getLoad(MVT::i8, Chain,
3675 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3676 DAG.getConstant(Offset, SrcVT)),
3678 Chain = Value.getValue(1);
3679 Chain = DAG.getStore(Chain, Value,
3680 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3681 DAG.getConstant(Offset, DstVT)),
3690 X86TargetLowering::LowerREADCYCLCECOUNTER(SDOperand Op, SelectionDAG &DAG) {
3691 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3692 SDOperand TheOp = Op.getOperand(0);
3693 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheOp, 1);
3694 if (Subtarget->is64Bit()) {
3695 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
3696 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::RDX,
3697 MVT::i64, Copy1.getValue(2));
3698 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, Copy2,
3699 DAG.getConstant(32, MVT::i8));
3701 DAG.getNode(ISD::OR, MVT::i64, Copy1, Tmp), Copy2.getValue(1)
3704 Tys = DAG.getVTList(MVT::i64, MVT::Other);
3705 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2);
3708 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
3709 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::EDX,
3710 MVT::i32, Copy1.getValue(2));
3711 SDOperand Ops[] = { Copy1, Copy2, Copy2.getValue(1) };
3712 Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
3713 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 3);
3716 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
3717 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
3719 if (!Subtarget->is64Bit()) {
3720 // vastart just stores the address of the VarArgsFrameIndex slot into the
3721 // memory location argument.
3722 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
3723 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
3728 // gp_offset (0 - 6 * 8)
3729 // fp_offset (48 - 48 + 8 * 16)
3730 // overflow_arg_area (point to parameters coming in memory).
3732 SmallVector<SDOperand, 8> MemOps;
3733 SDOperand FIN = Op.getOperand(1);
3735 SDOperand Store = DAG.getStore(Op.getOperand(0),
3736 DAG.getConstant(VarArgsGPOffset, MVT::i32),
3737 FIN, SV->getValue(), SV->getOffset());
3738 MemOps.push_back(Store);
3741 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
3742 DAG.getConstant(4, getPointerTy()));
3743 Store = DAG.getStore(Op.getOperand(0),
3744 DAG.getConstant(VarArgsFPOffset, MVT::i32),
3745 FIN, SV->getValue(), SV->getOffset());
3746 MemOps.push_back(Store);
3748 // Store ptr to overflow_arg_area
3749 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
3750 DAG.getConstant(4, getPointerTy()));
3751 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
3752 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
3754 MemOps.push_back(Store);
3756 // Store ptr to reg_save_area.
3757 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
3758 DAG.getConstant(8, getPointerTy()));
3759 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
3760 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
3762 MemOps.push_back(Store);
3763 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
3766 SDOperand X86TargetLowering::LowerVACOPY(SDOperand Op, SelectionDAG &DAG) {
3767 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
3768 SDOperand Chain = Op.getOperand(0);
3769 SDOperand DstPtr = Op.getOperand(1);
3770 SDOperand SrcPtr = Op.getOperand(2);
3771 SrcValueSDNode *DstSV = cast<SrcValueSDNode>(Op.getOperand(3));
3772 SrcValueSDNode *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4));
3774 SrcPtr = DAG.getLoad(getPointerTy(), Chain, SrcPtr,
3775 SrcSV->getValue(), SrcSV->getOffset());
3776 Chain = SrcPtr.getValue(1);
3777 for (unsigned i = 0; i < 3; ++i) {
3778 SDOperand Val = DAG.getLoad(MVT::i64, Chain, SrcPtr,
3779 SrcSV->getValue(), SrcSV->getOffset());
3780 Chain = Val.getValue(1);
3781 Chain = DAG.getStore(Chain, Val, DstPtr,
3782 DstSV->getValue(), DstSV->getOffset());
3785 SrcPtr = DAG.getNode(ISD::ADD, getPointerTy(), SrcPtr,
3786 DAG.getConstant(8, getPointerTy()));
3787 DstPtr = DAG.getNode(ISD::ADD, getPointerTy(), DstPtr,
3788 DAG.getConstant(8, getPointerTy()));
3794 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
3795 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
3797 default: return SDOperand(); // Don't custom lower most intrinsics.
3798 // Comparison intrinsics.
3799 case Intrinsic::x86_sse_comieq_ss:
3800 case Intrinsic::x86_sse_comilt_ss:
3801 case Intrinsic::x86_sse_comile_ss:
3802 case Intrinsic::x86_sse_comigt_ss:
3803 case Intrinsic::x86_sse_comige_ss:
3804 case Intrinsic::x86_sse_comineq_ss:
3805 case Intrinsic::x86_sse_ucomieq_ss:
3806 case Intrinsic::x86_sse_ucomilt_ss:
3807 case Intrinsic::x86_sse_ucomile_ss:
3808 case Intrinsic::x86_sse_ucomigt_ss:
3809 case Intrinsic::x86_sse_ucomige_ss:
3810 case Intrinsic::x86_sse_ucomineq_ss:
3811 case Intrinsic::x86_sse2_comieq_sd:
3812 case Intrinsic::x86_sse2_comilt_sd:
3813 case Intrinsic::x86_sse2_comile_sd:
3814 case Intrinsic::x86_sse2_comigt_sd:
3815 case Intrinsic::x86_sse2_comige_sd:
3816 case Intrinsic::x86_sse2_comineq_sd:
3817 case Intrinsic::x86_sse2_ucomieq_sd:
3818 case Intrinsic::x86_sse2_ucomilt_sd:
3819 case Intrinsic::x86_sse2_ucomile_sd:
3820 case Intrinsic::x86_sse2_ucomigt_sd:
3821 case Intrinsic::x86_sse2_ucomige_sd:
3822 case Intrinsic::x86_sse2_ucomineq_sd: {
3824 ISD::CondCode CC = ISD::SETCC_INVALID;
3827 case Intrinsic::x86_sse_comieq_ss:
3828 case Intrinsic::x86_sse2_comieq_sd:
3832 case Intrinsic::x86_sse_comilt_ss:
3833 case Intrinsic::x86_sse2_comilt_sd:
3837 case Intrinsic::x86_sse_comile_ss:
3838 case Intrinsic::x86_sse2_comile_sd:
3842 case Intrinsic::x86_sse_comigt_ss:
3843 case Intrinsic::x86_sse2_comigt_sd:
3847 case Intrinsic::x86_sse_comige_ss:
3848 case Intrinsic::x86_sse2_comige_sd:
3852 case Intrinsic::x86_sse_comineq_ss:
3853 case Intrinsic::x86_sse2_comineq_sd:
3857 case Intrinsic::x86_sse_ucomieq_ss:
3858 case Intrinsic::x86_sse2_ucomieq_sd:
3859 Opc = X86ISD::UCOMI;
3862 case Intrinsic::x86_sse_ucomilt_ss:
3863 case Intrinsic::x86_sse2_ucomilt_sd:
3864 Opc = X86ISD::UCOMI;
3867 case Intrinsic::x86_sse_ucomile_ss:
3868 case Intrinsic::x86_sse2_ucomile_sd:
3869 Opc = X86ISD::UCOMI;
3872 case Intrinsic::x86_sse_ucomigt_ss:
3873 case Intrinsic::x86_sse2_ucomigt_sd:
3874 Opc = X86ISD::UCOMI;
3877 case Intrinsic::x86_sse_ucomige_ss:
3878 case Intrinsic::x86_sse2_ucomige_sd:
3879 Opc = X86ISD::UCOMI;
3882 case Intrinsic::x86_sse_ucomineq_ss:
3883 case Intrinsic::x86_sse2_ucomineq_sd:
3884 Opc = X86ISD::UCOMI;
3890 SDOperand LHS = Op.getOperand(1);
3891 SDOperand RHS = Op.getOperand(2);
3892 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
3894 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3895 SDOperand Ops1[] = { DAG.getEntryNode(), LHS, RHS };
3896 SDOperand Cond = DAG.getNode(Opc, VTs, 2, Ops1, 3);
3897 VTs = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
3898 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
3899 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, VTs, 2, Ops2, 2);
3900 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
3905 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
3906 // Depths > 0 not supported yet!
3907 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
3910 // Just load the return address
3911 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
3912 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
3915 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
3916 // Depths > 0 not supported yet!
3917 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
3920 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
3921 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
3922 DAG.getConstant(4, getPointerTy()));
3925 /// LowerOperation - Provide custom lowering hooks for some operations.
3927 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
3928 switch (Op.getOpcode()) {
3929 default: assert(0 && "Should not custom lower this!");
3930 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3931 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3932 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3933 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
3934 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
3935 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3936 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
3937 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
3938 case ISD::SHL_PARTS:
3939 case ISD::SRA_PARTS:
3940 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
3941 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
3942 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
3943 case ISD::FABS: return LowerFABS(Op, DAG);
3944 case ISD::FNEG: return LowerFNEG(Op, DAG);
3945 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3946 case ISD::SETCC: return LowerSETCC(Op, DAG, DAG.getEntryNode());
3947 case ISD::SELECT: return LowerSELECT(Op, DAG);
3948 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
3949 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
3950 case ISD::CALL: return LowerCALL(Op, DAG);
3951 case ISD::RET: return LowerRET(Op, DAG);
3952 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
3953 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
3954 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
3955 case ISD::READCYCLECOUNTER: return LowerREADCYCLCECOUNTER(Op, DAG);
3956 case ISD::VASTART: return LowerVASTART(Op, DAG);
3957 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
3958 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
3959 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3960 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3965 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
3967 default: return NULL;
3968 case X86ISD::SHLD: return "X86ISD::SHLD";
3969 case X86ISD::SHRD: return "X86ISD::SHRD";
3970 case X86ISD::FAND: return "X86ISD::FAND";
3971 case X86ISD::FOR: return "X86ISD::FOR";
3972 case X86ISD::FXOR: return "X86ISD::FXOR";
3973 case X86ISD::FSRL: return "X86ISD::FSRL";
3974 case X86ISD::FILD: return "X86ISD::FILD";
3975 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
3976 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
3977 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
3978 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
3979 case X86ISD::FLD: return "X86ISD::FLD";
3980 case X86ISD::FST: return "X86ISD::FST";
3981 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
3982 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
3983 case X86ISD::CALL: return "X86ISD::CALL";
3984 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
3985 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
3986 case X86ISD::CMP: return "X86ISD::CMP";
3987 case X86ISD::COMI: return "X86ISD::COMI";
3988 case X86ISD::UCOMI: return "X86ISD::UCOMI";
3989 case X86ISD::SETCC: return "X86ISD::SETCC";
3990 case X86ISD::CMOV: return "X86ISD::CMOV";
3991 case X86ISD::BRCOND: return "X86ISD::BRCOND";
3992 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
3993 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
3994 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
3995 case X86ISD::LOAD_PACK: return "X86ISD::LOAD_PACK";
3996 case X86ISD::LOAD_UA: return "X86ISD::LOAD_UA";
3997 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
3998 case X86ISD::Wrapper: return "X86ISD::Wrapper";
3999 case X86ISD::S2VEC: return "X86ISD::S2VEC";
4000 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
4001 case X86ISD::PINSRW: return "X86ISD::PINSRW";
4002 case X86ISD::FMAX: return "X86ISD::FMAX";
4003 case X86ISD::FMIN: return "X86ISD::FMIN";
4007 /// isLegalAddressImmediate - Return true if the integer value or
4008 /// GlobalValue can be used as the offset of the target addressing mode.
4009 bool X86TargetLowering::isLegalAddressImmediate(int64_t V) const {
4010 // X86 allows a sign-extended 32-bit immediate field.
4011 return (V > -(1LL << 32) && V < (1LL << 32)-1);
4014 bool X86TargetLowering::isLegalAddressImmediate(GlobalValue *GV) const {
4015 // In 64-bit mode, GV is 64-bit so it won't fit in the 32-bit displacement
4016 // field unless we are in small code model.
4017 if (Subtarget->is64Bit() &&
4018 getTargetMachine().getCodeModel() != CodeModel::Small)
4021 return (!Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false));
4024 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4025 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4026 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4027 /// are assumed to be legal.
4029 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
4030 // Only do shuffles on 128-bit vector types for now.
4031 if (MVT::getSizeInBits(VT) == 64) return false;
4032 return (Mask.Val->getNumOperands() <= 4 ||
4033 isSplatMask(Mask.Val) ||
4034 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
4035 X86::isUNPCKLMask(Mask.Val) ||
4036 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
4037 X86::isUNPCKHMask(Mask.Val));
4040 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
4042 SelectionDAG &DAG) const {
4043 unsigned NumElts = BVOps.size();
4044 // Only do shuffles on 128-bit vector types for now.
4045 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
4046 if (NumElts == 2) return true;
4048 return (isMOVLMask(&BVOps[0], 4) ||
4049 isCommutedMOVL(&BVOps[0], 4, true) ||
4050 isSHUFPMask(&BVOps[0], 4) ||
4051 isCommutedSHUFP(&BVOps[0], 4));
4056 //===----------------------------------------------------------------------===//
4057 // X86 Scheduler Hooks
4058 //===----------------------------------------------------------------------===//
4061 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
4062 MachineBasicBlock *BB) {
4063 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4064 switch (MI->getOpcode()) {
4065 default: assert(false && "Unexpected instr type to insert");
4066 case X86::CMOV_FR32:
4067 case X86::CMOV_FR64:
4068 case X86::CMOV_V4F32:
4069 case X86::CMOV_V2F64:
4070 case X86::CMOV_V2I64: {
4071 // To "insert" a SELECT_CC instruction, we actually have to insert the
4072 // diamond control-flow pattern. The incoming instruction knows the
4073 // destination vreg to set, the condition code register to branch on, the
4074 // true/false values to select between, and a branch opcode to use.
4075 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4076 ilist<MachineBasicBlock>::iterator It = BB;
4082 // cmpTY ccX, r1, r2
4084 // fallthrough --> copy0MBB
4085 MachineBasicBlock *thisMBB = BB;
4086 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
4087 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
4089 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
4090 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
4091 MachineFunction *F = BB->getParent();
4092 F->getBasicBlockList().insert(It, copy0MBB);
4093 F->getBasicBlockList().insert(It, sinkMBB);
4094 // Update machine-CFG edges by first adding all successors of the current
4095 // block to the new block which will contain the Phi node for the select.
4096 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
4097 e = BB->succ_end(); i != e; ++i)
4098 sinkMBB->addSuccessor(*i);
4099 // Next, remove all successors of the current block, and add the true
4100 // and fallthrough blocks as its successors.
4101 while(!BB->succ_empty())
4102 BB->removeSuccessor(BB->succ_begin());
4103 BB->addSuccessor(copy0MBB);
4104 BB->addSuccessor(sinkMBB);
4107 // %FalseValue = ...
4108 // # fallthrough to sinkMBB
4111 // Update machine-CFG edges
4112 BB->addSuccessor(sinkMBB);
4115 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4118 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
4119 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4120 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4122 delete MI; // The pseudo instruction is gone now.
4126 case X86::FP_TO_INT16_IN_MEM:
4127 case X86::FP_TO_INT32_IN_MEM:
4128 case X86::FP_TO_INT64_IN_MEM: {
4129 // Change the floating point control register to use "round towards zero"
4130 // mode when truncating to an integer value.
4131 MachineFunction *F = BB->getParent();
4132 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
4133 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
4135 // Load the old value of the high byte of the control word...
4137 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
4138 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
4140 // Set the high part to be round to zero...
4141 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
4144 // Reload the modified control word now...
4145 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4147 // Restore the memory image of control word to original value
4148 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
4151 // Get the X86 opcode to use.
4153 switch (MI->getOpcode()) {
4154 default: assert(0 && "illegal opcode!");
4155 case X86::FP_TO_INT16_IN_MEM: Opc = X86::FpIST16m; break;
4156 case X86::FP_TO_INT32_IN_MEM: Opc = X86::FpIST32m; break;
4157 case X86::FP_TO_INT64_IN_MEM: Opc = X86::FpIST64m; break;
4161 MachineOperand &Op = MI->getOperand(0);
4162 if (Op.isRegister()) {
4163 AM.BaseType = X86AddressMode::RegBase;
4164 AM.Base.Reg = Op.getReg();
4166 AM.BaseType = X86AddressMode::FrameIndexBase;
4167 AM.Base.FrameIndex = Op.getFrameIndex();
4169 Op = MI->getOperand(1);
4170 if (Op.isImmediate())
4171 AM.Scale = Op.getImm();
4172 Op = MI->getOperand(2);
4173 if (Op.isImmediate())
4174 AM.IndexReg = Op.getImm();
4175 Op = MI->getOperand(3);
4176 if (Op.isGlobalAddress()) {
4177 AM.GV = Op.getGlobal();
4179 AM.Disp = Op.getImm();
4181 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
4182 .addReg(MI->getOperand(4).getReg());
4184 // Reload the original control word now.
4185 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4187 delete MI; // The pseudo instruction is gone now.
4193 //===----------------------------------------------------------------------===//
4194 // X86 Optimization Hooks
4195 //===----------------------------------------------------------------------===//
4197 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
4199 uint64_t &KnownZero,
4201 unsigned Depth) const {
4202 unsigned Opc = Op.getOpcode();
4203 assert((Opc >= ISD::BUILTIN_OP_END ||
4204 Opc == ISD::INTRINSIC_WO_CHAIN ||
4205 Opc == ISD::INTRINSIC_W_CHAIN ||
4206 Opc == ISD::INTRINSIC_VOID) &&
4207 "Should use MaskedValueIsZero if you don't know whether Op"
4208 " is a target node!");
4210 KnownZero = KnownOne = 0; // Don't know anything.
4214 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
4219 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4220 /// element of the result of the vector shuffle.
4221 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
4222 MVT::ValueType VT = N->getValueType(0);
4223 SDOperand PermMask = N->getOperand(2);
4224 unsigned NumElems = PermMask.getNumOperands();
4225 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
4227 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4229 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4230 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
4231 SDOperand Idx = PermMask.getOperand(i);
4232 if (Idx.getOpcode() == ISD::UNDEF)
4233 return DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4234 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
4239 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
4240 /// node is a GlobalAddress + an offset.
4241 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
4242 unsigned Opc = N->getOpcode();
4243 if (Opc == X86ISD::Wrapper) {
4244 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
4245 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
4248 } else if (Opc == ISD::ADD) {
4249 SDOperand N1 = N->getOperand(0);
4250 SDOperand N2 = N->getOperand(1);
4251 if (isGAPlusOffset(N1.Val, GA, Offset)) {
4252 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
4254 Offset += V->getSignExtended();
4257 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
4258 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
4260 Offset += V->getSignExtended();
4268 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
4270 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
4271 MachineFrameInfo *MFI) {
4272 if (N->getOperand(0).Val != Base->getOperand(0).Val)
4275 SDOperand Loc = N->getOperand(1);
4276 SDOperand BaseLoc = Base->getOperand(1);
4277 if (Loc.getOpcode() == ISD::FrameIndex) {
4278 if (BaseLoc.getOpcode() != ISD::FrameIndex)
4280 int FI = dyn_cast<FrameIndexSDNode>(Loc)->getIndex();
4281 int BFI = dyn_cast<FrameIndexSDNode>(BaseLoc)->getIndex();
4282 int FS = MFI->getObjectSize(FI);
4283 int BFS = MFI->getObjectSize(BFI);
4284 if (FS != BFS || FS != Size) return false;
4285 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
4287 GlobalValue *GV1 = NULL;
4288 GlobalValue *GV2 = NULL;
4289 int64_t Offset1 = 0;
4290 int64_t Offset2 = 0;
4291 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
4292 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
4293 if (isGA1 && isGA2 && GV1 == GV2)
4294 return Offset1 == (Offset2 + Dist*Size);
4300 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
4301 const X86Subtarget *Subtarget) {
4304 if (isGAPlusOffset(Base, GV, Offset))
4305 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
4307 assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
4308 int BFI = dyn_cast<FrameIndexSDNode>(Base)->getIndex();
4310 // Fixed objects do not specify alignment, however the offsets are known.
4311 return ((Subtarget->getStackAlignment() % 16) == 0 &&
4312 (MFI->getObjectOffset(BFI) % 16) == 0);
4314 return MFI->getObjectAlignment(BFI) >= 16;
4320 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
4321 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
4322 /// if the load addresses are consecutive, non-overlapping, and in the right
4324 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
4325 const X86Subtarget *Subtarget) {
4326 MachineFunction &MF = DAG.getMachineFunction();
4327 MachineFrameInfo *MFI = MF.getFrameInfo();
4328 MVT::ValueType VT = N->getValueType(0);
4329 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
4330 SDOperand PermMask = N->getOperand(2);
4331 int NumElems = (int)PermMask.getNumOperands();
4332 SDNode *Base = NULL;
4333 for (int i = 0; i < NumElems; ++i) {
4334 SDOperand Idx = PermMask.getOperand(i);
4335 if (Idx.getOpcode() == ISD::UNDEF) {
4336 if (!Base) return SDOperand();
4339 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
4340 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
4344 else if (!isConsecutiveLoad(Arg.Val, Base,
4345 i, MVT::getSizeInBits(EVT)/8,MFI))
4350 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
4352 LoadSDNode *LD = cast<LoadSDNode>(Base);
4353 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
4354 LD->getSrcValueOffset());
4356 // Just use movups, it's shorter.
4357 SDVTList Tys = DAG.getVTList(MVT::v4f32, MVT::Other);
4358 SmallVector<SDOperand, 3> Ops;
4359 Ops.push_back(Base->getOperand(0));
4360 Ops.push_back(Base->getOperand(1));
4361 Ops.push_back(Base->getOperand(2));
4362 return DAG.getNode(ISD::BIT_CONVERT, VT,
4363 DAG.getNode(X86ISD::LOAD_UA, Tys, &Ops[0], Ops.size()));
4367 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
4368 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
4369 const X86Subtarget *Subtarget) {
4370 SDOperand Cond = N->getOperand(0);
4372 // If we have SSE[12] support, try to form min/max nodes.
4373 if (Subtarget->hasSSE2() &&
4374 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
4375 if (Cond.getOpcode() == ISD::SETCC) {
4376 // Get the LHS/RHS of the select.
4377 SDOperand LHS = N->getOperand(1);
4378 SDOperand RHS = N->getOperand(2);
4379 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
4381 unsigned Opcode = 0;
4382 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
4385 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
4388 if (!UnsafeFPMath) break;
4390 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
4392 Opcode = X86ISD::FMIN;
4395 case ISD::SETOGT: // (X > Y) ? X : Y -> max
4398 if (!UnsafeFPMath) break;
4400 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
4402 Opcode = X86ISD::FMAX;
4405 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
4408 case ISD::SETOGT: // (X > Y) ? Y : X -> min
4411 if (!UnsafeFPMath) break;
4413 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
4415 Opcode = X86ISD::FMIN;
4418 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
4421 if (!UnsafeFPMath) break;
4423 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
4425 Opcode = X86ISD::FMAX;
4431 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
4440 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
4441 DAGCombinerInfo &DCI) const {
4442 SelectionDAG &DAG = DCI.DAG;
4443 switch (N->getOpcode()) {
4445 case ISD::VECTOR_SHUFFLE:
4446 return PerformShuffleCombine(N, DAG, Subtarget);
4448 return PerformSELECTCombine(N, DAG, Subtarget);
4454 //===----------------------------------------------------------------------===//
4455 // X86 Inline Assembly Support
4456 //===----------------------------------------------------------------------===//
4458 /// getConstraintType - Given a constraint letter, return the type of
4459 /// constraint it is for this target.
4460 X86TargetLowering::ConstraintType
4461 X86TargetLowering::getConstraintType(char ConstraintLetter) const {
4462 switch (ConstraintLetter) {
4471 return C_RegisterClass;
4472 default: return TargetLowering::getConstraintType(ConstraintLetter);
4476 /// isOperandValidForConstraint - Return the specified operand (possibly
4477 /// modified) if the specified SDOperand is valid for the specified target
4478 /// constraint letter, otherwise return null.
4479 SDOperand X86TargetLowering::
4480 isOperandValidForConstraint(SDOperand Op, char Constraint, SelectionDAG &DAG) {
4481 switch (Constraint) {
4484 // Literal immediates are always ok.
4485 if (isa<ConstantSDNode>(Op)) return Op;
4487 // If we are in non-pic codegen mode, we allow the address of a global to
4488 // be used with 'i'.
4489 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op)) {
4490 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
4491 return SDOperand(0, 0);
4493 if (GA->getOpcode() != ISD::TargetGlobalAddress)
4494 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
4499 // Otherwise, not valid for this mode.
4500 return SDOperand(0, 0);
4502 return TargetLowering::isOperandValidForConstraint(Op, Constraint, DAG);
4506 std::vector<unsigned> X86TargetLowering::
4507 getRegClassForInlineAsmConstraint(const std::string &Constraint,
4508 MVT::ValueType VT) const {
4509 if (Constraint.size() == 1) {
4510 // FIXME: not handling fp-stack yet!
4511 // FIXME: not handling MMX registers yet ('y' constraint).
4512 switch (Constraint[0]) { // GCC X86 Constraint Letters
4513 default: break; // Unknown constraint letter
4514 case 'A': // EAX/EDX
4515 if (VT == MVT::i32 || VT == MVT::i64)
4516 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
4518 case 'r': // GENERAL_REGS
4519 case 'R': // LEGACY_REGS
4520 if (VT == MVT::i64 && Subtarget->is64Bit())
4521 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
4522 X86::RSI, X86::RDI, X86::RBP, X86::RSP,
4523 X86::R8, X86::R9, X86::R10, X86::R11,
4524 X86::R12, X86::R13, X86::R14, X86::R15, 0);
4526 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
4527 X86::ESI, X86::EDI, X86::EBP, X86::ESP, 0);
4528 else if (VT == MVT::i16)
4529 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
4530 X86::SI, X86::DI, X86::BP, X86::SP, 0);
4531 else if (VT == MVT::i8)
4532 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
4534 case 'l': // INDEX_REGS
4536 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
4537 X86::ESI, X86::EDI, X86::EBP, 0);
4538 else if (VT == MVT::i16)
4539 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
4540 X86::SI, X86::DI, X86::BP, 0);
4541 else if (VT == MVT::i8)
4542 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
4544 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
4547 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
4548 else if (VT == MVT::i16)
4549 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
4550 else if (VT == MVT::i8)
4551 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
4553 case 'x': // SSE_REGS if SSE1 allowed
4554 if (Subtarget->hasSSE1())
4555 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
4556 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
4558 return std::vector<unsigned>();
4559 case 'Y': // SSE_REGS if SSE2 allowed
4560 if (Subtarget->hasSSE2())
4561 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
4562 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
4564 return std::vector<unsigned>();
4568 return std::vector<unsigned>();
4571 std::pair<unsigned, const TargetRegisterClass*>
4572 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4573 MVT::ValueType VT) const {
4574 // Use the default implementation in TargetLowering to convert the register
4575 // constraint into a member of a register class.
4576 std::pair<unsigned, const TargetRegisterClass*> Res;
4577 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4579 // Not found as a standard register?
4580 if (Res.second == 0) {
4581 // GCC calls "st(0)" just plain "st".
4582 if (StringsEqualNoCase("{st}", Constraint)) {
4583 Res.first = X86::ST0;
4584 Res.second = X86::RSTRegisterClass;
4590 // Otherwise, check to see if this is a register class of the wrong value
4591 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
4592 // turn into {ax},{dx}.
4593 if (Res.second->hasType(VT))
4594 return Res; // Correct type already, nothing to do.
4596 // All of the single-register GCC register classes map their values onto
4597 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
4598 // really want an 8-bit or 32-bit register, map to the appropriate register
4599 // class and return the appropriate register.
4600 if (Res.second != X86::GR16RegisterClass)
4603 if (VT == MVT::i8) {
4604 unsigned DestReg = 0;
4605 switch (Res.first) {
4607 case X86::AX: DestReg = X86::AL; break;
4608 case X86::DX: DestReg = X86::DL; break;
4609 case X86::CX: DestReg = X86::CL; break;
4610 case X86::BX: DestReg = X86::BL; break;
4613 Res.first = DestReg;
4614 Res.second = Res.second = X86::GR8RegisterClass;
4616 } else if (VT == MVT::i32) {
4617 unsigned DestReg = 0;
4618 switch (Res.first) {
4620 case X86::AX: DestReg = X86::EAX; break;
4621 case X86::DX: DestReg = X86::EDX; break;
4622 case X86::CX: DestReg = X86::ECX; break;
4623 case X86::BX: DestReg = X86::EBX; break;
4624 case X86::SI: DestReg = X86::ESI; break;
4625 case X86::DI: DestReg = X86::EDI; break;
4626 case X86::BP: DestReg = X86::EBP; break;
4627 case X86::SP: DestReg = X86::ESP; break;
4630 Res.first = DestReg;
4631 Res.second = Res.second = X86::GR32RegisterClass;
4633 } else if (VT == MVT::i64) {
4634 unsigned DestReg = 0;
4635 switch (Res.first) {
4637 case X86::AX: DestReg = X86::RAX; break;
4638 case X86::DX: DestReg = X86::RDX; break;
4639 case X86::CX: DestReg = X86::RCX; break;
4640 case X86::BX: DestReg = X86::RBX; break;
4641 case X86::SI: DestReg = X86::RSI; break;
4642 case X86::DI: DestReg = X86::RDI; break;
4643 case X86::BP: DestReg = X86::RBP; break;
4644 case X86::SP: DestReg = X86::RSP; break;
4647 Res.first = DestReg;
4648 Res.second = Res.second = X86::GR64RegisterClass;