1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Chris Lattner and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
16 #include "X86InstrBuilder.h"
17 #include "X86ISelLowering.h"
18 #include "X86MachineFunctionInfo.h"
19 #include "X86TargetMachine.h"
20 #include "llvm/CallingConv.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/Function.h"
24 #include "llvm/Intrinsics.h"
25 #include "llvm/ADT/VectorExtras.h"
26 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
27 #include "llvm/CodeGen/CallingConvLower.h"
28 #include "llvm/CodeGen/MachineFrameInfo.h"
29 #include "llvm/CodeGen/MachineFunction.h"
30 #include "llvm/CodeGen/MachineInstrBuilder.h"
31 #include "llvm/CodeGen/SelectionDAG.h"
32 #include "llvm/CodeGen/SSARegMap.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Target/TargetOptions.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/ADT/StringExtras.h"
40 static cl::opt<bool> EnableFastCC("enable-x86-fastcc", cl::Hidden,
41 cl::desc("Enable fastcc on X86"));
42 X86TargetLowering::X86TargetLowering(TargetMachine &TM)
43 : TargetLowering(TM) {
44 Subtarget = &TM.getSubtarget<X86Subtarget>();
45 X86ScalarSSE = Subtarget->hasSSE2();
46 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
48 // Set up the TargetLowering object.
50 // X86 is weird, it always uses i8 for shift amounts and setcc results.
51 setShiftAmountType(MVT::i8);
52 setSetCCResultType(MVT::i8);
53 setSetCCResultContents(ZeroOrOneSetCCResult);
54 setSchedulingPreference(SchedulingForRegPressure);
55 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
56 setStackPointerRegisterToSaveRestore(X86StackPtr);
58 if (Subtarget->isTargetDarwin()) {
59 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
60 setUseUnderscoreSetJmp(false);
61 setUseUnderscoreLongJmp(false);
62 } else if (Subtarget->isTargetMingw()) {
63 // MS runtime is weird: it exports _setjmp, but longjmp!
64 setUseUnderscoreSetJmp(true);
65 setUseUnderscoreLongJmp(false);
67 setUseUnderscoreSetJmp(true);
68 setUseUnderscoreLongJmp(true);
71 // Add legal addressing mode scale values.
72 addLegalAddressScale(8);
73 addLegalAddressScale(4);
74 addLegalAddressScale(2);
75 // Enter the ones which require both scale + index last. These are more
77 addLegalAddressScale(9);
78 addLegalAddressScale(5);
79 addLegalAddressScale(3);
81 // Set up the register classes.
82 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
83 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
84 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
85 if (Subtarget->is64Bit())
86 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
88 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Expand);
90 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
92 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
93 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
94 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
96 if (Subtarget->is64Bit()) {
97 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
98 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
101 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP.
102 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
104 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
107 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
109 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
110 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
111 // SSE has no i16 to fp conversion, only i32
113 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
115 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
116 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
119 if (!Subtarget->is64Bit()) {
120 // Custom lower SINT_TO_FP and FP_TO_SINT from/to i64 in 32-bit mode.
121 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
122 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
125 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
127 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
128 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
131 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
133 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
134 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
137 // Handle FP_TO_UINT by promoting the destination to a larger signed
139 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
140 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
141 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
143 if (Subtarget->is64Bit()) {
144 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
145 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
147 if (X86ScalarSSE && !Subtarget->hasSSE3())
148 // Expand FP_TO_UINT into a select.
149 // FIXME: We would like to use a Custom expander here eventually to do
150 // the optimal thing for SSE vs. the default expansion in the legalizer.
151 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
153 // With SSE3 we can use fisttpll to convert to a signed i64.
154 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
157 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
159 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
160 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
163 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
164 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
165 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
166 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
167 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
168 if (Subtarget->is64Bit())
169 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Expand);
170 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Expand);
171 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Expand);
172 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
173 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
174 setOperationAction(ISD::FREM , MVT::f64 , Expand);
176 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
177 setOperationAction(ISD::CTTZ , MVT::i8 , Expand);
178 setOperationAction(ISD::CTLZ , MVT::i8 , Expand);
179 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
180 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
181 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
182 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
183 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
184 setOperationAction(ISD::CTLZ , MVT::i32 , Expand);
185 if (Subtarget->is64Bit()) {
186 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
187 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
188 setOperationAction(ISD::CTLZ , MVT::i64 , Expand);
191 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
192 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
194 // These should be promoted to a larger select which is supported.
195 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
196 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
197 // X86 wants to expand cmov itself.
198 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
199 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
200 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
201 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
202 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
203 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
204 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
205 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
206 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
207 if (Subtarget->is64Bit()) {
208 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
209 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
211 // X86 ret instruction may pop stack.
212 setOperationAction(ISD::RET , MVT::Other, Custom);
214 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
215 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
216 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
217 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
218 if (Subtarget->is64Bit()) {
219 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
220 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
221 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
222 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
224 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
225 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
226 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
227 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
228 // X86 wants to expand memset / memcpy itself.
229 setOperationAction(ISD::MEMSET , MVT::Other, Custom);
230 setOperationAction(ISD::MEMCPY , MVT::Other, Custom);
232 // We don't have line number support yet.
233 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
234 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
235 // FIXME - use subtarget debug flags
236 if (!Subtarget->isTargetDarwin() &&
237 !Subtarget->isTargetELF() &&
238 !Subtarget->isTargetCygMing())
239 setOperationAction(ISD::LABEL, MVT::Other, Expand);
241 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
242 setOperationAction(ISD::VASTART , MVT::Other, Custom);
244 // Use the default implementation.
245 setOperationAction(ISD::VAARG , MVT::Other, Expand);
246 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
247 setOperationAction(ISD::VAEND , MVT::Other, Expand);
248 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
249 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
250 if (Subtarget->is64Bit())
251 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
252 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
255 // Set up the FP register classes.
256 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
257 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
259 // Use ANDPD to simulate FABS.
260 setOperationAction(ISD::FABS , MVT::f64, Custom);
261 setOperationAction(ISD::FABS , MVT::f32, Custom);
263 // Use XORP to simulate FNEG.
264 setOperationAction(ISD::FNEG , MVT::f64, Custom);
265 setOperationAction(ISD::FNEG , MVT::f32, Custom);
267 // Use ANDPD and ORPD to simulate FCOPYSIGN.
268 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
269 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
271 // We don't support sin/cos/fmod
272 setOperationAction(ISD::FSIN , MVT::f64, Expand);
273 setOperationAction(ISD::FCOS , MVT::f64, Expand);
274 setOperationAction(ISD::FREM , MVT::f64, Expand);
275 setOperationAction(ISD::FSIN , MVT::f32, Expand);
276 setOperationAction(ISD::FCOS , MVT::f32, Expand);
277 setOperationAction(ISD::FREM , MVT::f32, Expand);
279 // Expand FP immediates into loads from the stack, except for the special
281 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
282 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
283 addLegalFPImmediate(+0.0); // xorps / xorpd
285 // Set up the FP register classes.
286 addRegisterClass(MVT::f64, X86::RFPRegisterClass);
288 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
289 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
290 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
293 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
294 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
297 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
298 addLegalFPImmediate(+0.0); // FLD0
299 addLegalFPImmediate(+1.0); // FLD1
300 addLegalFPImmediate(-0.0); // FLD0/FCHS
301 addLegalFPImmediate(-1.0); // FLD1/FCHS
304 // First set operation action for all vector types to expand. Then we
305 // will selectively turn on ones that can be effectively codegen'd.
306 for (unsigned VT = (unsigned)MVT::Vector + 1;
307 VT != (unsigned)MVT::LAST_VALUETYPE; VT++) {
308 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
309 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
310 setOperationAction(ISD::FADD, (MVT::ValueType)VT, Expand);
311 setOperationAction(ISD::FSUB, (MVT::ValueType)VT, Expand);
312 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
313 setOperationAction(ISD::FMUL, (MVT::ValueType)VT, Expand);
314 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
315 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
316 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
317 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
318 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
319 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand);
320 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Expand);
321 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
322 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
325 if (Subtarget->hasMMX()) {
326 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
327 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
328 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
330 // FIXME: add MMX packed arithmetics
331 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Expand);
332 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Expand);
333 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Expand);
336 if (Subtarget->hasSSE1()) {
337 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
339 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
340 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
341 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
342 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
343 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
344 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
345 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
346 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
347 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
350 if (Subtarget->hasSSE2()) {
351 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
352 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
353 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
354 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
355 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
357 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
358 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
359 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
360 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
361 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
362 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
363 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
364 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
365 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
366 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
367 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
369 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
370 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
371 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
372 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
373 // Implement v4f32 insert_vector_elt in terms of SSE2 v8i16 ones.
374 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
376 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
377 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
378 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
379 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
380 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
382 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
383 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
384 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
385 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
386 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
387 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
389 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
390 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
391 setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote);
392 AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64);
393 setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote);
394 AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64);
395 setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote);
396 AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64);
397 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote);
398 AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64);
399 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
400 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64);
403 // Custom lower v2i64 and v2f64 selects.
404 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
405 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
406 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
407 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
410 // We want to custom lower some of our intrinsics.
411 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
413 // We have target-specific dag combine patterns for the following nodes:
414 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
415 setTargetDAGCombine(ISD::SELECT);
417 computeRegisterProperties();
419 // FIXME: These should be based on subtarget info. Plus, the values should
420 // be smaller when we are in optimizing for size mode.
421 maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
422 maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
423 maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
424 allowUnalignedMemoryAccesses = true; // x86 supports it!
428 //===----------------------------------------------------------------------===//
429 // Return Value Calling Convention Implementation
430 //===----------------------------------------------------------------------===//
432 #include "X86GenCallingConv.inc"
434 /// LowerRET - Lower an ISD::RET node.
435 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
436 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
438 SmallVector<CCValAssign, 16> RVLocs;
439 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
440 CCState CCInfo(CC, getTargetMachine(), RVLocs);
442 // Determine which register each value should be copied into.
443 for (unsigned i = 0; i != Op.getNumOperands() / 2; ++i) {
444 MVT::ValueType VT = Op.getOperand(i*2+1).getValueType();
445 if (RetCC_X86(i, VT, VT, CCValAssign::Full,
446 cast<ConstantSDNode>(Op.getOperand(i*2+2))->getValue(),
448 assert(0 && "Unhandled result type!");
451 // If this is the first return lowered for this function, add the regs to the
452 // liveout set for the function.
453 if (DAG.getMachineFunction().liveout_empty()) {
454 for (unsigned i = 0; i != RVLocs.size(); ++i)
455 if (RVLocs[i].isRegLoc())
456 DAG.getMachineFunction().addLiveOut(RVLocs[i].getLocReg());
459 SDOperand Chain = Op.getOperand(0);
462 // Copy the result values into the output registers.
463 if (RVLocs.size() != 1 || !RVLocs[0].isRegLoc() ||
464 RVLocs[0].getLocReg() != X86::ST0) {
465 for (unsigned i = 0; i != RVLocs.size(); ++i) {
466 CCValAssign &VA = RVLocs[i];
467 assert(VA.isRegLoc() && "Can only return in registers!");
468 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1),
470 Flag = Chain.getValue(1);
473 // We need to handle a destination of ST0 specially, because it isn't really
475 SDOperand Value = Op.getOperand(1);
477 // If this is an FP return with ScalarSSE, we need to move the value from
478 // an XMM register onto the fp-stack.
482 // If this is a load into a scalarsse value, don't store the loaded value
483 // back to the stack, only to reload it: just replace the scalar-sse load.
484 if (ISD::isNON_EXTLoad(Value.Val) &&
485 (Chain == Value.getValue(1) || Chain == Value.getOperand(0))) {
486 Chain = Value.getOperand(0);
487 MemLoc = Value.getOperand(1);
489 // Spill the value to memory and reload it into top of stack.
490 unsigned Size = MVT::getSizeInBits(RVLocs[0].getValVT())/8;
491 MachineFunction &MF = DAG.getMachineFunction();
492 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
493 MemLoc = DAG.getFrameIndex(SSFI, getPointerTy());
494 Chain = DAG.getStore(Op.getOperand(0), Value, MemLoc, NULL, 0);
496 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
497 SDOperand Ops[] = {Chain, MemLoc, DAG.getValueType(RVLocs[0].getValVT())};
498 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
499 Chain = Value.getValue(1);
502 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
503 SDOperand Ops[] = { Chain, Value };
504 Chain = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops, 2);
505 Flag = Chain.getValue(1);
508 SDOperand BytesToPop = DAG.getConstant(getBytesToPopOnReturn(), MVT::i16);
510 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop, Flag);
512 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop);
516 /// LowerCallResult - Lower the result values of an ISD::CALL into the
517 /// appropriate copies out of appropriate physical registers. This assumes that
518 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
519 /// being lowered. The returns a SDNode with the same number of values as the
521 SDNode *X86TargetLowering::
522 LowerCallResult(SDOperand Chain, SDOperand InFlag, SDNode *TheCall,
523 unsigned CallingConv, SelectionDAG &DAG) {
524 SmallVector<SDOperand, 8> ResultVals;
526 SmallVector<CCValAssign, 16> RVLocs;
527 CCState CCInfo(CallingConv, getTargetMachine(), RVLocs);
529 for (unsigned i = 0, e = TheCall->getNumValues() - 1; i != e; ++i) {
530 MVT::ValueType VT = TheCall->getValueType(i);
531 if (RetCC_X86(i, VT, VT, CCValAssign::Full, 0, CCInfo))
532 assert(0 && "Unhandled result type!");
535 // Copy all of the result registers out of their specified physreg.
536 if (RVLocs.size() != 1 || RVLocs[0].getLocReg() != X86::ST0) {
537 for (unsigned i = 0; i != RVLocs.size(); ++i) {
538 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(),
539 RVLocs[i].getValVT(), InFlag).getValue(1);
540 InFlag = Chain.getValue(2);
541 ResultVals.push_back(Chain.getValue(0));
544 // Copies from the FP stack are special, as ST0 isn't a valid register
545 // before the fp stackifier runs.
547 // Copy ST0 into an RFP register with FP_GET_RESULT.
548 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
549 SDOperand GROps[] = { Chain, InFlag };
550 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, GROps, 2);
551 Chain = RetVal.getValue(1);
552 InFlag = RetVal.getValue(2);
554 // If we are using ScalarSSE, store ST(0) to the stack and reload it into
557 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
558 // shouldn't be necessary except that RFP cannot be live across
559 // multiple blocks. When stackifier is fixed, they can be uncoupled.
560 MachineFunction &MF = DAG.getMachineFunction();
561 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
562 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
564 Chain, RetVal, StackSlot, DAG.getValueType(RVLocs[0].getValVT()), InFlag
566 Chain = DAG.getNode(X86ISD::FST, MVT::Other, Ops, 5);
567 RetVal = DAG.getLoad(RVLocs[0].getValVT(), Chain, StackSlot, NULL, 0);
568 Chain = RetVal.getValue(1);
571 if (RVLocs[0].getValVT() == MVT::f32 && !X86ScalarSSE)
572 // FIXME: we would really like to remember that this FP_ROUND
573 // operation is okay to eliminate if we allow excess FP precision.
574 RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal);
575 ResultVals.push_back(RetVal);
578 // Merge everything together with a MERGE_VALUES node.
579 ResultVals.push_back(Chain);
580 return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(),
581 &ResultVals[0], ResultVals.size()).Val;
585 //===----------------------------------------------------------------------===//
586 // C & StdCall Calling Convention implementation
587 //===----------------------------------------------------------------------===//
588 // StdCall calling convention seems to be standard for many Windows' API
589 // routines and around. It differs from C calling convention just a little:
590 // callee should clean up the stack, not caller. Symbols should be also
591 // decorated in some fancy way :) It doesn't support any vector arguments.
593 /// AddLiveIn - This helper function adds the specified physical register to the
594 /// MachineFunction as a live in value. It also creates a corresponding virtual
596 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
597 const TargetRegisterClass *RC) {
598 assert(RC->contains(PReg) && "Not the correct regclass!");
599 unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC);
600 MF.addLiveIn(PReg, VReg);
604 SDOperand X86TargetLowering::LowerCCCArguments(SDOperand Op, SelectionDAG &DAG,
606 unsigned NumArgs = Op.Val->getNumValues() - 1;
607 MachineFunction &MF = DAG.getMachineFunction();
608 MachineFrameInfo *MFI = MF.getFrameInfo();
609 SDOperand Root = Op.getOperand(0);
610 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
612 // Assign locations to all of the incoming arguments.
613 SmallVector<CCValAssign, 16> ArgLocs;
614 CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(),
616 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_C);
618 SmallVector<SDOperand, 8> ArgValues;
619 unsigned LastVal = ~0U;
620 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
621 CCValAssign &VA = ArgLocs[i];
622 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
624 assert(VA.getValNo() != LastVal &&
625 "Don't support value assigned to multiple locs yet");
626 LastVal = VA.getValNo();
629 MVT::ValueType RegVT = VA.getLocVT();
630 TargetRegisterClass *RC;
631 if (RegVT == MVT::i32)
632 RC = X86::GR32RegisterClass;
634 assert(MVT::isVector(RegVT));
635 RC = X86::VR128RegisterClass;
638 SDOperand ArgValue = DAG.getCopyFromReg(Root, VA.getLocReg(), RegVT);
639 AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
641 // If this is an 8 or 16-bit value, it is really passed promoted to 32
642 // bits. Insert an assert[sz]ext to capture this, then truncate to the
644 if (VA.getLocInfo() == CCValAssign::SExt)
645 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
646 DAG.getValueType(VA.getValVT()));
647 else if (VA.getLocInfo() == CCValAssign::ZExt)
648 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
649 DAG.getValueType(VA.getValVT()));
651 if (VA.getLocInfo() != CCValAssign::Full)
652 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
654 ArgValues.push_back(ArgValue);
656 assert(VA.isMemLoc());
658 // Create the nodes corresponding to a load from this parameter slot.
659 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
660 VA.getLocMemOffset());
661 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
662 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
666 unsigned StackSize = CCInfo.getNextStackOffset();
668 ArgValues.push_back(Root);
670 // If the function takes variable number of arguments, make a frame index for
671 // the start of the first vararg value... for expansion of llvm.va_start.
673 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
675 if (isStdCall && !isVarArg) {
676 BytesToPopOnReturn = StackSize; // Callee pops everything..
677 BytesCallerReserves = 0;
679 BytesToPopOnReturn = 0; // Callee pops hidden struct pointer.
681 // If this is an sret function, the return should pop the hidden pointer.
682 if (NumArgs && (cast<ConstantSDNode>(Op.getOperand(3))->getValue() & 4))
683 BytesToPopOnReturn = 4;
685 BytesCallerReserves = StackSize;
688 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
689 ReturnAddrIndex = 0; // No return address slot generated yet.
691 MF.getInfo<X86FunctionInfo>()->setBytesToPopOnReturn(BytesToPopOnReturn);
693 // Return the new list of results.
694 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
695 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
698 SDOperand X86TargetLowering::LowerCCCCallTo(SDOperand Op, SelectionDAG &DAG,
700 SDOperand Chain = Op.getOperand(0);
701 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
702 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
703 SDOperand Callee = Op.getOperand(4);
704 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
706 // Analyze operands of the call, assigning locations to each operand.
707 SmallVector<CCValAssign, 16> ArgLocs;
708 CCState CCInfo(CC, getTargetMachine(), ArgLocs);
709 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_C);
711 // Get a count of how many bytes are to be pushed on the stack.
712 unsigned NumBytes = CCInfo.getNextStackOffset();
714 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
716 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
717 SmallVector<SDOperand, 8> MemOpChains;
721 // Walk the register/memloc assignments, inserting copies/loads.
722 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
723 CCValAssign &VA = ArgLocs[i];
724 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
726 // Promote the value if needed.
727 switch (VA.getLocInfo()) {
728 default: assert(0 && "Unknown loc info!");
729 case CCValAssign::Full: break;
730 case CCValAssign::SExt:
731 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
733 case CCValAssign::ZExt:
734 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
736 case CCValAssign::AExt:
737 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
742 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
744 assert(VA.isMemLoc());
745 if (StackPtr.Val == 0)
746 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
747 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
748 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
749 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
753 // If the first argument is an sret pointer, remember it.
754 bool isSRet = NumOps &&(cast<ConstantSDNode>(Op.getOperand(6))->getValue()&4);
756 if (!MemOpChains.empty())
757 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
758 &MemOpChains[0], MemOpChains.size());
760 // Build a sequence of copy-to-reg nodes chained together with token chain
761 // and flag operands which copy the outgoing args into registers.
763 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
764 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
766 InFlag = Chain.getValue(1);
769 // ELF / PIC requires GOT in the EBX register before function calls via PLT
771 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
772 Subtarget->isPICStyleGOT()) {
773 Chain = DAG.getCopyToReg(Chain, X86::EBX,
774 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
776 InFlag = Chain.getValue(1);
779 // If the callee is a GlobalAddress node (quite common, every direct call is)
780 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
781 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
782 // We should use extra load for direct calls to dllimported functions in
784 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
785 getTargetMachine(), true))
786 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
787 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
788 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
790 // Returns a chain & a flag for retval copy to use.
791 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
792 SmallVector<SDOperand, 8> Ops;
793 Ops.push_back(Chain);
794 Ops.push_back(Callee);
796 // Add argument registers to the end of the list so that they are known live
798 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
799 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
800 RegsToPass[i].second.getValueType()));
802 // Add an implicit use GOT pointer in EBX.
803 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
804 Subtarget->isPICStyleGOT())
805 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
808 Ops.push_back(InFlag);
810 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
811 NodeTys, &Ops[0], Ops.size());
812 InFlag = Chain.getValue(1);
814 // Create the CALLSEQ_END node.
815 unsigned NumBytesForCalleeToPush = 0;
817 if (CC == CallingConv::X86_StdCall) {
819 NumBytesForCalleeToPush = isSRet ? 4 : 0;
821 NumBytesForCalleeToPush = NumBytes;
823 // If this is is a call to a struct-return function, the callee
824 // pops the hidden struct pointer, so we have to push it back.
825 // This is common for Darwin/X86, Linux & Mingw32 targets.
826 NumBytesForCalleeToPush = isSRet ? 4 : 0;
829 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
831 Ops.push_back(Chain);
832 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
833 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
834 Ops.push_back(InFlag);
835 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
836 InFlag = Chain.getValue(1);
838 // Handle result values, copying them out of physregs into vregs that we
840 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
844 //===----------------------------------------------------------------------===//
845 // FastCall Calling Convention implementation
846 //===----------------------------------------------------------------------===//
848 // The X86 'fastcall' calling convention passes up to two integer arguments in
849 // registers (an appropriate portion of ECX/EDX), passes arguments in C order,
850 // and requires that the callee pop its arguments off the stack (allowing proper
851 // tail calls), and has the same return value conventions as C calling convs.
853 // This calling convention always arranges for the callee pop value to be 8n+4
854 // bytes, which is needed for tail recursion elimination and stack alignment
857 X86TargetLowering::LowerFastCCArguments(SDOperand Op, SelectionDAG &DAG) {
858 MachineFunction &MF = DAG.getMachineFunction();
859 MachineFrameInfo *MFI = MF.getFrameInfo();
860 SDOperand Root = Op.getOperand(0);
862 // Assign locations to all of the incoming arguments.
863 SmallVector<CCValAssign, 16> ArgLocs;
864 CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(),
866 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_FastCall);
868 SmallVector<SDOperand, 8> ArgValues;
869 unsigned LastVal = ~0U;
870 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
871 CCValAssign &VA = ArgLocs[i];
872 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
874 assert(VA.getValNo() != LastVal &&
875 "Don't support value assigned to multiple locs yet");
876 LastVal = VA.getValNo();
879 MVT::ValueType RegVT = VA.getLocVT();
880 TargetRegisterClass *RC;
881 if (RegVT == MVT::i32)
882 RC = X86::GR32RegisterClass;
884 assert(MVT::isVector(RegVT));
885 RC = X86::VR128RegisterClass;
888 SDOperand ArgValue = DAG.getCopyFromReg(Root, VA.getLocReg(), RegVT);
889 AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
891 // If this is an 8 or 16-bit value, it is really passed promoted to 32
892 // bits. Insert an assert[sz]ext to capture this, then truncate to the
894 if (VA.getLocInfo() == CCValAssign::SExt)
895 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
896 DAG.getValueType(VA.getValVT()));
897 else if (VA.getLocInfo() == CCValAssign::ZExt)
898 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
899 DAG.getValueType(VA.getValVT()));
901 if (VA.getLocInfo() != CCValAssign::Full)
902 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
904 ArgValues.push_back(ArgValue);
906 assert(VA.isMemLoc());
908 // Create the nodes corresponding to a load from this parameter slot.
909 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
910 VA.getLocMemOffset());
911 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
912 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
916 ArgValues.push_back(Root);
918 unsigned StackSize = CCInfo.getNextStackOffset();
920 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
921 // arguments and the arguments after the retaddr has been pushed are aligned.
922 if ((StackSize & 7) == 0)
925 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
926 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
927 ReturnAddrIndex = 0; // No return address slot generated yet.
928 BytesToPopOnReturn = StackSize; // Callee pops all stack arguments.
929 BytesCallerReserves = 0;
931 MF.getInfo<X86FunctionInfo>()->setBytesToPopOnReturn(BytesToPopOnReturn);
933 // Return the new list of results.
934 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
935 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
938 SDOperand X86TargetLowering::LowerFastCCCallTo(SDOperand Op, SelectionDAG &DAG,
940 SDOperand Chain = Op.getOperand(0);
941 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
942 SDOperand Callee = Op.getOperand(4);
944 // Analyze operands of the call, assigning locations to each operand.
945 SmallVector<CCValAssign, 16> ArgLocs;
946 CCState CCInfo(CC, getTargetMachine(), ArgLocs);
947 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_FastCall);
949 // Get a count of how many bytes are to be pushed on the stack.
950 unsigned NumBytes = CCInfo.getNextStackOffset();
952 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
953 // arguments and the arguments after the retaddr has been pushed are aligned.
954 if ((NumBytes & 7) == 0)
957 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
960 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
961 SmallVector<SDOperand, 8> MemOpChains;
965 // Walk the register/memloc assignments, inserting copies/loads.
966 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
967 CCValAssign &VA = ArgLocs[i];
968 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
970 // Promote the value if needed.
971 switch (VA.getLocInfo()) {
972 default: assert(0 && "Unknown loc info!");
973 case CCValAssign::Full: break;
974 case CCValAssign::SExt:
975 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
977 case CCValAssign::ZExt:
978 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
980 case CCValAssign::AExt:
981 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
986 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
988 assert(VA.isMemLoc());
989 if (StackPtr.Val == 0)
990 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
991 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
992 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
993 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
997 if (!MemOpChains.empty())
998 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
999 &MemOpChains[0], MemOpChains.size());
1001 // Build a sequence of copy-to-reg nodes chained together with token chain
1002 // and flag operands which copy the outgoing args into registers.
1004 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1005 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1007 InFlag = Chain.getValue(1);
1010 // If the callee is a GlobalAddress node (quite common, every direct call is)
1011 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1012 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1013 // We should use extra load for direct calls to dllimported functions in
1015 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1016 getTargetMachine(), true))
1017 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1018 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1019 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1021 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1023 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1024 Subtarget->isPICStyleGOT()) {
1025 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1026 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1028 InFlag = Chain.getValue(1);
1031 // Returns a chain & a flag for retval copy to use.
1032 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1033 SmallVector<SDOperand, 8> Ops;
1034 Ops.push_back(Chain);
1035 Ops.push_back(Callee);
1037 // Add argument registers to the end of the list so that they are known live
1039 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1040 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1041 RegsToPass[i].second.getValueType()));
1043 // Add an implicit use GOT pointer in EBX.
1044 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1045 Subtarget->isPICStyleGOT())
1046 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1049 Ops.push_back(InFlag);
1051 // FIXME: Do not generate X86ISD::TAILCALL for now.
1052 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1053 NodeTys, &Ops[0], Ops.size());
1054 InFlag = Chain.getValue(1);
1056 // Returns a flag for retval copy to use.
1057 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1059 Ops.push_back(Chain);
1060 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1061 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1062 Ops.push_back(InFlag);
1063 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1064 InFlag = Chain.getValue(1);
1066 // Handle result values, copying them out of physregs into vregs that we
1068 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1072 //===----------------------------------------------------------------------===//
1073 // X86-64 C Calling Convention implementation
1074 //===----------------------------------------------------------------------===//
1077 X86TargetLowering::LowerX86_64CCCArguments(SDOperand Op, SelectionDAG &DAG) {
1078 MachineFunction &MF = DAG.getMachineFunction();
1079 MachineFrameInfo *MFI = MF.getFrameInfo();
1080 SDOperand Root = Op.getOperand(0);
1081 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1083 static const unsigned GPR64ArgRegs[] = {
1084 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1086 static const unsigned XMMArgRegs[] = {
1087 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1088 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1092 // Assign locations to all of the incoming arguments.
1093 SmallVector<CCValAssign, 16> ArgLocs;
1094 CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(),
1096 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_64_C);
1098 SmallVector<SDOperand, 8> ArgValues;
1099 unsigned LastVal = ~0U;
1100 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1101 CCValAssign &VA = ArgLocs[i];
1102 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1104 assert(VA.getValNo() != LastVal &&
1105 "Don't support value assigned to multiple locs yet");
1106 LastVal = VA.getValNo();
1108 if (VA.isRegLoc()) {
1109 MVT::ValueType RegVT = VA.getLocVT();
1110 TargetRegisterClass *RC;
1111 if (RegVT == MVT::i32)
1112 RC = X86::GR32RegisterClass;
1113 else if (RegVT == MVT::i64)
1114 RC = X86::GR64RegisterClass;
1115 else if (RegVT == MVT::f32)
1116 RC = X86::FR32RegisterClass;
1117 else if (RegVT == MVT::f64)
1118 RC = X86::FR64RegisterClass;
1120 assert(MVT::isVector(RegVT));
1121 RC = X86::VR128RegisterClass;
1124 SDOperand ArgValue = DAG.getCopyFromReg(Root, VA.getLocReg(), RegVT);
1125 AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1127 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1128 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1130 if (VA.getLocInfo() == CCValAssign::SExt)
1131 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1132 DAG.getValueType(VA.getValVT()));
1133 else if (VA.getLocInfo() == CCValAssign::ZExt)
1134 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1135 DAG.getValueType(VA.getValVT()));
1137 if (VA.getLocInfo() != CCValAssign::Full)
1138 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1140 ArgValues.push_back(ArgValue);
1142 assert(VA.isMemLoc());
1144 // Create the nodes corresponding to a load from this parameter slot.
1145 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
1146 VA.getLocMemOffset());
1147 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1148 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
1152 unsigned StackSize = CCInfo.getNextStackOffset();
1154 // If the function takes variable number of arguments, make a frame index for
1155 // the start of the first vararg value... for expansion of llvm.va_start.
1157 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs, 6);
1158 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1160 // For X86-64, if there are vararg parameters that are passed via
1161 // registers, then we must store them to their spots on the stack so they
1162 // may be loaded by deferencing the result of va_next.
1163 VarArgsGPOffset = NumIntRegs * 8;
1164 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1165 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1166 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1168 // Store the integer parameter registers.
1169 SmallVector<SDOperand, 8> MemOps;
1170 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1171 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1172 DAG.getConstant(VarArgsGPOffset, getPointerTy()));
1173 for (; NumIntRegs != 6; ++NumIntRegs) {
1174 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1175 X86::GR64RegisterClass);
1176 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1177 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1178 MemOps.push_back(Store);
1179 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1180 DAG.getConstant(8, getPointerTy()));
1183 // Now store the XMM (fp + vector) parameter registers.
1184 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1185 DAG.getConstant(VarArgsFPOffset, getPointerTy()));
1186 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1187 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1188 X86::VR128RegisterClass);
1189 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1190 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1191 MemOps.push_back(Store);
1192 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1193 DAG.getConstant(16, getPointerTy()));
1195 if (!MemOps.empty())
1196 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1197 &MemOps[0], MemOps.size());
1200 ArgValues.push_back(Root);
1202 ReturnAddrIndex = 0; // No return address slot generated yet.
1203 BytesToPopOnReturn = 0; // Callee pops nothing.
1204 BytesCallerReserves = StackSize;
1206 // Return the new list of results.
1207 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1208 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1212 X86TargetLowering::LowerX86_64CCCCallTo(SDOperand Op, SelectionDAG &DAG,
1214 SDOperand Chain = Op.getOperand(0);
1215 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1216 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1217 SDOperand Callee = Op.getOperand(4);
1219 // Analyze operands of the call, assigning locations to each operand.
1220 SmallVector<CCValAssign, 16> ArgLocs;
1221 CCState CCInfo(CC, getTargetMachine(), ArgLocs);
1222 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_C);
1224 // Get a count of how many bytes are to be pushed on the stack.
1225 unsigned NumBytes = CCInfo.getNextStackOffset();
1226 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1228 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1229 SmallVector<SDOperand, 8> MemOpChains;
1233 // Walk the register/memloc assignments, inserting copies/loads.
1234 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1235 CCValAssign &VA = ArgLocs[i];
1236 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1238 // Promote the value if needed.
1239 switch (VA.getLocInfo()) {
1240 default: assert(0 && "Unknown loc info!");
1241 case CCValAssign::Full: break;
1242 case CCValAssign::SExt:
1243 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1245 case CCValAssign::ZExt:
1246 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1248 case CCValAssign::AExt:
1249 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1253 if (VA.isRegLoc()) {
1254 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1256 assert(VA.isMemLoc());
1257 if (StackPtr.Val == 0)
1258 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1259 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
1260 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1261 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1265 if (!MemOpChains.empty())
1266 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1267 &MemOpChains[0], MemOpChains.size());
1269 // Build a sequence of copy-to-reg nodes chained together with token chain
1270 // and flag operands which copy the outgoing args into registers.
1272 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1273 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1275 InFlag = Chain.getValue(1);
1279 // From AMD64 ABI document:
1280 // For calls that may call functions that use varargs or stdargs
1281 // (prototype-less calls or calls to functions containing ellipsis (...) in
1282 // the declaration) %al is used as hidden argument to specify the number
1283 // of SSE registers used. The contents of %al do not need to match exactly
1284 // the number of registers, but must be an ubound on the number of SSE
1285 // registers used and is in the range 0 - 8 inclusive.
1287 // Count the number of XMM registers allocated.
1288 static const unsigned XMMArgRegs[] = {
1289 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1290 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1292 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1294 Chain = DAG.getCopyToReg(Chain, X86::AL,
1295 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1296 InFlag = Chain.getValue(1);
1299 // If the callee is a GlobalAddress node (quite common, every direct call is)
1300 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1301 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1302 // We should use extra load for direct calls to dllimported functions in
1304 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1305 getTargetMachine(), true))
1306 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1307 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1308 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1310 // Returns a chain & a flag for retval copy to use.
1311 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1312 SmallVector<SDOperand, 8> Ops;
1313 Ops.push_back(Chain);
1314 Ops.push_back(Callee);
1316 // Add argument registers to the end of the list so that they are known live
1318 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1319 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1320 RegsToPass[i].second.getValueType()));
1323 Ops.push_back(InFlag);
1325 // FIXME: Do not generate X86ISD::TAILCALL for now.
1326 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1327 NodeTys, &Ops[0], Ops.size());
1328 InFlag = Chain.getValue(1);
1330 // Returns a flag for retval copy to use.
1331 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1333 Ops.push_back(Chain);
1334 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1335 Ops.push_back(DAG.getConstant(0, getPointerTy()));
1336 Ops.push_back(InFlag);
1337 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1338 InFlag = Chain.getValue(1);
1340 // Handle result values, copying them out of physregs into vregs that we
1342 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1346 //===----------------------------------------------------------------------===//
1347 // Other Lowering Hooks
1348 //===----------------------------------------------------------------------===//
1351 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1352 if (ReturnAddrIndex == 0) {
1353 // Set up a frame object for the return address.
1354 MachineFunction &MF = DAG.getMachineFunction();
1355 if (Subtarget->is64Bit())
1356 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
1358 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
1361 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1366 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1367 /// specific condition code. It returns a false if it cannot do a direct
1368 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1370 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1371 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
1372 SelectionDAG &DAG) {
1373 X86CC = X86::COND_INVALID;
1375 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1376 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1377 // X > -1 -> X == 0, jump !sign.
1378 RHS = DAG.getConstant(0, RHS.getValueType());
1379 X86CC = X86::COND_NS;
1381 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1382 // X < 0 -> X == 0, jump on sign.
1383 X86CC = X86::COND_S;
1388 switch (SetCCOpcode) {
1390 case ISD::SETEQ: X86CC = X86::COND_E; break;
1391 case ISD::SETGT: X86CC = X86::COND_G; break;
1392 case ISD::SETGE: X86CC = X86::COND_GE; break;
1393 case ISD::SETLT: X86CC = X86::COND_L; break;
1394 case ISD::SETLE: X86CC = X86::COND_LE; break;
1395 case ISD::SETNE: X86CC = X86::COND_NE; break;
1396 case ISD::SETULT: X86CC = X86::COND_B; break;
1397 case ISD::SETUGT: X86CC = X86::COND_A; break;
1398 case ISD::SETULE: X86CC = X86::COND_BE; break;
1399 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1402 // On a floating point condition, the flags are set as follows:
1404 // 0 | 0 | 0 | X > Y
1405 // 0 | 0 | 1 | X < Y
1406 // 1 | 0 | 0 | X == Y
1407 // 1 | 1 | 1 | unordered
1409 switch (SetCCOpcode) {
1412 case ISD::SETEQ: X86CC = X86::COND_E; break;
1413 case ISD::SETOLT: Flip = true; // Fallthrough
1415 case ISD::SETGT: X86CC = X86::COND_A; break;
1416 case ISD::SETOLE: Flip = true; // Fallthrough
1418 case ISD::SETGE: X86CC = X86::COND_AE; break;
1419 case ISD::SETUGT: Flip = true; // Fallthrough
1421 case ISD::SETLT: X86CC = X86::COND_B; break;
1422 case ISD::SETUGE: Flip = true; // Fallthrough
1424 case ISD::SETLE: X86CC = X86::COND_BE; break;
1426 case ISD::SETNE: X86CC = X86::COND_NE; break;
1427 case ISD::SETUO: X86CC = X86::COND_P; break;
1428 case ISD::SETO: X86CC = X86::COND_NP; break;
1431 std::swap(LHS, RHS);
1434 return X86CC != X86::COND_INVALID;
1437 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
1438 /// code. Current x86 isa includes the following FP cmov instructions:
1439 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
1440 static bool hasFPCMov(unsigned X86CC) {
1456 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
1457 /// true if Op is undef or if its value falls within the specified range (L, H].
1458 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
1459 if (Op.getOpcode() == ISD::UNDEF)
1462 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
1463 return (Val >= Low && Val < Hi);
1466 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
1467 /// true if Op is undef or if its value equal to the specified value.
1468 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
1469 if (Op.getOpcode() == ISD::UNDEF)
1471 return cast<ConstantSDNode>(Op)->getValue() == Val;
1474 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
1475 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
1476 bool X86::isPSHUFDMask(SDNode *N) {
1477 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1479 if (N->getNumOperands() != 4)
1482 // Check if the value doesn't reference the second vector.
1483 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1484 SDOperand Arg = N->getOperand(i);
1485 if (Arg.getOpcode() == ISD::UNDEF) continue;
1486 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1487 if (cast<ConstantSDNode>(Arg)->getValue() >= 4)
1494 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
1495 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
1496 bool X86::isPSHUFHWMask(SDNode *N) {
1497 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1499 if (N->getNumOperands() != 8)
1502 // Lower quadword copied in order.
1503 for (unsigned i = 0; i != 4; ++i) {
1504 SDOperand Arg = N->getOperand(i);
1505 if (Arg.getOpcode() == ISD::UNDEF) continue;
1506 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1507 if (cast<ConstantSDNode>(Arg)->getValue() != i)
1511 // Upper quadword shuffled.
1512 for (unsigned i = 4; i != 8; ++i) {
1513 SDOperand Arg = N->getOperand(i);
1514 if (Arg.getOpcode() == ISD::UNDEF) continue;
1515 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1516 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1517 if (Val < 4 || Val > 7)
1524 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
1525 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
1526 bool X86::isPSHUFLWMask(SDNode *N) {
1527 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1529 if (N->getNumOperands() != 8)
1532 // Upper quadword copied in order.
1533 for (unsigned i = 4; i != 8; ++i)
1534 if (!isUndefOrEqual(N->getOperand(i), i))
1537 // Lower quadword shuffled.
1538 for (unsigned i = 0; i != 4; ++i)
1539 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
1545 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
1546 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
1547 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
1548 if (NumElems != 2 && NumElems != 4) return false;
1550 unsigned Half = NumElems / 2;
1551 for (unsigned i = 0; i < Half; ++i)
1552 if (!isUndefOrInRange(Elems[i], 0, NumElems))
1554 for (unsigned i = Half; i < NumElems; ++i)
1555 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
1561 bool X86::isSHUFPMask(SDNode *N) {
1562 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1563 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
1566 /// isCommutedSHUFP - Returns true if the shuffle mask is except
1567 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
1568 /// half elements to come from vector 1 (which would equal the dest.) and
1569 /// the upper half to come from vector 2.
1570 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
1571 if (NumOps != 2 && NumOps != 4) return false;
1573 unsigned Half = NumOps / 2;
1574 for (unsigned i = 0; i < Half; ++i)
1575 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
1577 for (unsigned i = Half; i < NumOps; ++i)
1578 if (!isUndefOrInRange(Ops[i], 0, NumOps))
1583 static bool isCommutedSHUFP(SDNode *N) {
1584 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1585 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
1588 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
1589 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
1590 bool X86::isMOVHLPSMask(SDNode *N) {
1591 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1593 if (N->getNumOperands() != 4)
1596 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
1597 return isUndefOrEqual(N->getOperand(0), 6) &&
1598 isUndefOrEqual(N->getOperand(1), 7) &&
1599 isUndefOrEqual(N->getOperand(2), 2) &&
1600 isUndefOrEqual(N->getOperand(3), 3);
1603 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
1604 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
1606 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
1607 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1609 if (N->getNumOperands() != 4)
1612 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
1613 return isUndefOrEqual(N->getOperand(0), 2) &&
1614 isUndefOrEqual(N->getOperand(1), 3) &&
1615 isUndefOrEqual(N->getOperand(2), 2) &&
1616 isUndefOrEqual(N->getOperand(3), 3);
1619 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
1620 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
1621 bool X86::isMOVLPMask(SDNode *N) {
1622 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1624 unsigned NumElems = N->getNumOperands();
1625 if (NumElems != 2 && NumElems != 4)
1628 for (unsigned i = 0; i < NumElems/2; ++i)
1629 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
1632 for (unsigned i = NumElems/2; i < NumElems; ++i)
1633 if (!isUndefOrEqual(N->getOperand(i), i))
1639 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
1640 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
1642 bool X86::isMOVHPMask(SDNode *N) {
1643 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1645 unsigned NumElems = N->getNumOperands();
1646 if (NumElems != 2 && NumElems != 4)
1649 for (unsigned i = 0; i < NumElems/2; ++i)
1650 if (!isUndefOrEqual(N->getOperand(i), i))
1653 for (unsigned i = 0; i < NumElems/2; ++i) {
1654 SDOperand Arg = N->getOperand(i + NumElems/2);
1655 if (!isUndefOrEqual(Arg, i + NumElems))
1662 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
1663 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
1664 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
1665 bool V2IsSplat = false) {
1666 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1669 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
1670 SDOperand BitI = Elts[i];
1671 SDOperand BitI1 = Elts[i+1];
1672 if (!isUndefOrEqual(BitI, j))
1675 if (isUndefOrEqual(BitI1, NumElts))
1678 if (!isUndefOrEqual(BitI1, j + NumElts))
1686 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
1687 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1688 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
1691 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
1692 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
1693 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
1694 bool V2IsSplat = false) {
1695 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1698 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
1699 SDOperand BitI = Elts[i];
1700 SDOperand BitI1 = Elts[i+1];
1701 if (!isUndefOrEqual(BitI, j + NumElts/2))
1704 if (isUndefOrEqual(BitI1, NumElts))
1707 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
1715 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
1716 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1717 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
1720 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
1721 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
1723 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
1724 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1726 unsigned NumElems = N->getNumOperands();
1727 if (NumElems != 4 && NumElems != 8 && NumElems != 16)
1730 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
1731 SDOperand BitI = N->getOperand(i);
1732 SDOperand BitI1 = N->getOperand(i+1);
1734 if (!isUndefOrEqual(BitI, j))
1736 if (!isUndefOrEqual(BitI1, j))
1743 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
1744 /// specifies a shuffle of elements that is suitable for input to MOVSS,
1745 /// MOVSD, and MOVD, i.e. setting the lowest element.
1746 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
1747 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1750 if (!isUndefOrEqual(Elts[0], NumElts))
1753 for (unsigned i = 1; i < NumElts; ++i) {
1754 if (!isUndefOrEqual(Elts[i], i))
1761 bool X86::isMOVLMask(SDNode *N) {
1762 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1763 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
1766 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
1767 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
1768 /// element of vector 2 and the other elements to come from vector 1 in order.
1769 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
1770 bool V2IsSplat = false,
1771 bool V2IsUndef = false) {
1772 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
1775 if (!isUndefOrEqual(Ops[0], 0))
1778 for (unsigned i = 1; i < NumOps; ++i) {
1779 SDOperand Arg = Ops[i];
1780 if (!(isUndefOrEqual(Arg, i+NumOps) ||
1781 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
1782 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
1789 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
1790 bool V2IsUndef = false) {
1791 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1792 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
1793 V2IsSplat, V2IsUndef);
1796 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
1797 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
1798 bool X86::isMOVSHDUPMask(SDNode *N) {
1799 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1801 if (N->getNumOperands() != 4)
1804 // Expect 1, 1, 3, 3
1805 for (unsigned i = 0; i < 2; ++i) {
1806 SDOperand Arg = N->getOperand(i);
1807 if (Arg.getOpcode() == ISD::UNDEF) continue;
1808 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1809 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1810 if (Val != 1) return false;
1814 for (unsigned i = 2; i < 4; ++i) {
1815 SDOperand Arg = N->getOperand(i);
1816 if (Arg.getOpcode() == ISD::UNDEF) continue;
1817 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1818 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1819 if (Val != 3) return false;
1823 // Don't use movshdup if it can be done with a shufps.
1827 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
1828 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
1829 bool X86::isMOVSLDUPMask(SDNode *N) {
1830 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1832 if (N->getNumOperands() != 4)
1835 // Expect 0, 0, 2, 2
1836 for (unsigned i = 0; i < 2; ++i) {
1837 SDOperand Arg = N->getOperand(i);
1838 if (Arg.getOpcode() == ISD::UNDEF) continue;
1839 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1840 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1841 if (Val != 0) return false;
1845 for (unsigned i = 2; i < 4; ++i) {
1846 SDOperand Arg = N->getOperand(i);
1847 if (Arg.getOpcode() == ISD::UNDEF) continue;
1848 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1849 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1850 if (Val != 2) return false;
1854 // Don't use movshdup if it can be done with a shufps.
1858 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
1859 /// a splat of a single element.
1860 static bool isSplatMask(SDNode *N) {
1861 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1863 // This is a splat operation if each element of the permute is the same, and
1864 // if the value doesn't reference the second vector.
1865 unsigned NumElems = N->getNumOperands();
1866 SDOperand ElementBase;
1868 for (; i != NumElems; ++i) {
1869 SDOperand Elt = N->getOperand(i);
1870 if (isa<ConstantSDNode>(Elt)) {
1876 if (!ElementBase.Val)
1879 for (; i != NumElems; ++i) {
1880 SDOperand Arg = N->getOperand(i);
1881 if (Arg.getOpcode() == ISD::UNDEF) continue;
1882 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1883 if (Arg != ElementBase) return false;
1886 // Make sure it is a splat of the first vector operand.
1887 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
1890 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
1891 /// a splat of a single element and it's a 2 or 4 element mask.
1892 bool X86::isSplatMask(SDNode *N) {
1893 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1895 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
1896 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
1898 return ::isSplatMask(N);
1901 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
1902 /// specifies a splat of zero element.
1903 bool X86::isSplatLoMask(SDNode *N) {
1904 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1906 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
1907 if (!isUndefOrEqual(N->getOperand(i), 0))
1912 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
1913 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
1915 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
1916 unsigned NumOperands = N->getNumOperands();
1917 unsigned Shift = (NumOperands == 4) ? 2 : 1;
1919 for (unsigned i = 0; i < NumOperands; ++i) {
1921 SDOperand Arg = N->getOperand(NumOperands-i-1);
1922 if (Arg.getOpcode() != ISD::UNDEF)
1923 Val = cast<ConstantSDNode>(Arg)->getValue();
1924 if (Val >= NumOperands) Val -= NumOperands;
1926 if (i != NumOperands - 1)
1933 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
1934 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
1936 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
1938 // 8 nodes, but we only care about the last 4.
1939 for (unsigned i = 7; i >= 4; --i) {
1941 SDOperand Arg = N->getOperand(i);
1942 if (Arg.getOpcode() != ISD::UNDEF)
1943 Val = cast<ConstantSDNode>(Arg)->getValue();
1952 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
1953 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
1955 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
1957 // 8 nodes, but we only care about the first 4.
1958 for (int i = 3; i >= 0; --i) {
1960 SDOperand Arg = N->getOperand(i);
1961 if (Arg.getOpcode() != ISD::UNDEF)
1962 Val = cast<ConstantSDNode>(Arg)->getValue();
1971 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
1972 /// specifies a 8 element shuffle that can be broken into a pair of
1973 /// PSHUFHW and PSHUFLW.
1974 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
1975 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1977 if (N->getNumOperands() != 8)
1980 // Lower quadword shuffled.
1981 for (unsigned i = 0; i != 4; ++i) {
1982 SDOperand Arg = N->getOperand(i);
1983 if (Arg.getOpcode() == ISD::UNDEF) continue;
1984 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1985 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1990 // Upper quadword shuffled.
1991 for (unsigned i = 4; i != 8; ++i) {
1992 SDOperand Arg = N->getOperand(i);
1993 if (Arg.getOpcode() == ISD::UNDEF) continue;
1994 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1995 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1996 if (Val < 4 || Val > 7)
2003 /// CommuteVectorShuffle - Swap vector_shuffle operandsas well as
2004 /// values in ther permute mask.
2005 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2006 SDOperand &V2, SDOperand &Mask,
2007 SelectionDAG &DAG) {
2008 MVT::ValueType VT = Op.getValueType();
2009 MVT::ValueType MaskVT = Mask.getValueType();
2010 MVT::ValueType EltVT = MVT::getVectorBaseType(MaskVT);
2011 unsigned NumElems = Mask.getNumOperands();
2012 SmallVector<SDOperand, 8> MaskVec;
2014 for (unsigned i = 0; i != NumElems; ++i) {
2015 SDOperand Arg = Mask.getOperand(i);
2016 if (Arg.getOpcode() == ISD::UNDEF) {
2017 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2020 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2021 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2023 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2025 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2029 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2030 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2033 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2034 /// match movhlps. The lower half elements should come from upper half of
2035 /// V1 (and in order), and the upper half elements should come from the upper
2036 /// half of V2 (and in order).
2037 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2038 unsigned NumElems = Mask->getNumOperands();
2041 for (unsigned i = 0, e = 2; i != e; ++i)
2042 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2044 for (unsigned i = 2; i != 4; ++i)
2045 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2050 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2051 /// is promoted to a vector.
2052 static inline bool isScalarLoadToVector(SDNode *N) {
2053 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2054 N = N->getOperand(0).Val;
2055 return ISD::isNON_EXTLoad(N);
2060 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2061 /// match movlp{s|d}. The lower half elements should come from lower half of
2062 /// V1 (and in order), and the upper half elements should come from the upper
2063 /// half of V2 (and in order). And since V1 will become the source of the
2064 /// MOVLP, it must be either a vector load or a scalar load to vector.
2065 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2066 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2068 // Is V2 is a vector load, don't do this transformation. We will try to use
2069 // load folding shufps op.
2070 if (ISD::isNON_EXTLoad(V2))
2073 unsigned NumElems = Mask->getNumOperands();
2074 if (NumElems != 2 && NumElems != 4)
2076 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2077 if (!isUndefOrEqual(Mask->getOperand(i), i))
2079 for (unsigned i = NumElems/2; i != NumElems; ++i)
2080 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2085 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2087 static bool isSplatVector(SDNode *N) {
2088 if (N->getOpcode() != ISD::BUILD_VECTOR)
2091 SDOperand SplatValue = N->getOperand(0);
2092 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2093 if (N->getOperand(i) != SplatValue)
2098 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2100 static bool isUndefShuffle(SDNode *N) {
2101 if (N->getOpcode() != ISD::BUILD_VECTOR)
2104 SDOperand V1 = N->getOperand(0);
2105 SDOperand V2 = N->getOperand(1);
2106 SDOperand Mask = N->getOperand(2);
2107 unsigned NumElems = Mask.getNumOperands();
2108 for (unsigned i = 0; i != NumElems; ++i) {
2109 SDOperand Arg = Mask.getOperand(i);
2110 if (Arg.getOpcode() != ISD::UNDEF) {
2111 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2112 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2114 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2121 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2122 /// that point to V2 points to its first element.
2123 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2124 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2126 bool Changed = false;
2127 SmallVector<SDOperand, 8> MaskVec;
2128 unsigned NumElems = Mask.getNumOperands();
2129 for (unsigned i = 0; i != NumElems; ++i) {
2130 SDOperand Arg = Mask.getOperand(i);
2131 if (Arg.getOpcode() != ISD::UNDEF) {
2132 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2133 if (Val > NumElems) {
2134 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2138 MaskVec.push_back(Arg);
2142 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2143 &MaskVec[0], MaskVec.size());
2147 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2148 /// operation of specified width.
2149 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2150 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2151 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2153 SmallVector<SDOperand, 8> MaskVec;
2154 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2155 for (unsigned i = 1; i != NumElems; ++i)
2156 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2157 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2160 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2161 /// of specified width.
2162 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2163 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2164 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2165 SmallVector<SDOperand, 8> MaskVec;
2166 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2167 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2168 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2170 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2173 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2174 /// of specified width.
2175 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2176 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2177 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2178 unsigned Half = NumElems/2;
2179 SmallVector<SDOperand, 8> MaskVec;
2180 for (unsigned i = 0; i != Half; ++i) {
2181 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2182 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2184 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2187 /// getZeroVector - Returns a vector of specified type with all zero elements.
2189 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2190 assert(MVT::isVector(VT) && "Expected a vector type");
2191 unsigned NumElems = getVectorNumElements(VT);
2192 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2193 bool isFP = MVT::isFloatingPoint(EVT);
2194 SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT);
2195 SmallVector<SDOperand, 8> ZeroVec(NumElems, Zero);
2196 return DAG.getNode(ISD::BUILD_VECTOR, VT, &ZeroVec[0], ZeroVec.size());
2199 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2201 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2202 SDOperand V1 = Op.getOperand(0);
2203 SDOperand Mask = Op.getOperand(2);
2204 MVT::ValueType VT = Op.getValueType();
2205 unsigned NumElems = Mask.getNumOperands();
2206 Mask = getUnpacklMask(NumElems, DAG);
2207 while (NumElems != 4) {
2208 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2211 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2213 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2214 Mask = getZeroVector(MaskVT, DAG);
2215 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2216 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2217 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2220 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2222 static inline bool isZeroNode(SDOperand Elt) {
2223 return ((isa<ConstantSDNode>(Elt) &&
2224 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2225 (isa<ConstantFPSDNode>(Elt) &&
2226 cast<ConstantFPSDNode>(Elt)->isExactlyValue(0.0)));
2229 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2230 /// vector and zero or undef vector.
2231 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2232 unsigned NumElems, unsigned Idx,
2233 bool isZero, SelectionDAG &DAG) {
2234 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2235 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2236 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2237 SDOperand Zero = DAG.getConstant(0, EVT);
2238 SmallVector<SDOperand, 8> MaskVec(NumElems, Zero);
2239 MaskVec[Idx] = DAG.getConstant(NumElems, EVT);
2240 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2241 &MaskVec[0], MaskVec.size());
2242 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2245 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2247 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
2248 unsigned NumNonZero, unsigned NumZero,
2249 SelectionDAG &DAG, TargetLowering &TLI) {
2255 for (unsigned i = 0; i < 16; ++i) {
2256 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
2257 if (ThisIsNonZero && First) {
2259 V = getZeroVector(MVT::v8i16, DAG);
2261 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2266 SDOperand ThisElt(0, 0), LastElt(0, 0);
2267 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
2268 if (LastIsNonZero) {
2269 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
2271 if (ThisIsNonZero) {
2272 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
2273 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
2274 ThisElt, DAG.getConstant(8, MVT::i8));
2276 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
2281 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
2282 DAG.getConstant(i/2, TLI.getPointerTy()));
2286 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
2289 /// LowerBuildVectorv16i8 - Custom lower build_vector of v8i16.
2291 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
2292 unsigned NumNonZero, unsigned NumZero,
2293 SelectionDAG &DAG, TargetLowering &TLI) {
2299 for (unsigned i = 0; i < 8; ++i) {
2300 bool isNonZero = (NonZeros & (1 << i)) != 0;
2304 V = getZeroVector(MVT::v8i16, DAG);
2306 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2309 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
2310 DAG.getConstant(i, TLI.getPointerTy()));
2318 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2319 // All zero's are handled with pxor.
2320 if (ISD::isBuildVectorAllZeros(Op.Val))
2323 // All one's are handled with pcmpeqd.
2324 if (ISD::isBuildVectorAllOnes(Op.Val))
2327 MVT::ValueType VT = Op.getValueType();
2328 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2329 unsigned EVTBits = MVT::getSizeInBits(EVT);
2331 unsigned NumElems = Op.getNumOperands();
2332 unsigned NumZero = 0;
2333 unsigned NumNonZero = 0;
2334 unsigned NonZeros = 0;
2335 std::set<SDOperand> Values;
2336 for (unsigned i = 0; i < NumElems; ++i) {
2337 SDOperand Elt = Op.getOperand(i);
2338 if (Elt.getOpcode() != ISD::UNDEF) {
2340 if (isZeroNode(Elt))
2343 NonZeros |= (1 << i);
2349 if (NumNonZero == 0)
2350 // Must be a mix of zero and undef. Return a zero vector.
2351 return getZeroVector(VT, DAG);
2353 // Splat is obviously ok. Let legalizer expand it to a shuffle.
2354 if (Values.size() == 1)
2357 // Special case for single non-zero element.
2358 if (NumNonZero == 1) {
2359 unsigned Idx = CountTrailingZeros_32(NonZeros);
2360 SDOperand Item = Op.getOperand(Idx);
2361 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
2363 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
2364 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
2367 if (EVTBits == 32) {
2368 // Turn it into a shuffle of zero and zero-extended scalar to vector.
2369 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
2371 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2372 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
2373 SmallVector<SDOperand, 8> MaskVec;
2374 for (unsigned i = 0; i < NumElems; i++)
2375 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
2376 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2377 &MaskVec[0], MaskVec.size());
2378 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
2379 DAG.getNode(ISD::UNDEF, VT), Mask);
2383 // Let legalizer expand 2-wide build_vector's.
2387 // If element VT is < 32 bits, convert it to inserts into a zero vector.
2389 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
2391 if (V.Val) return V;
2394 if (EVTBits == 16) {
2395 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
2397 if (V.Val) return V;
2400 // If element VT is == 32 bits, turn it into a number of shuffles.
2401 SmallVector<SDOperand, 8> V;
2403 if (NumElems == 4 && NumZero > 0) {
2404 for (unsigned i = 0; i < 4; ++i) {
2405 bool isZero = !(NonZeros & (1 << i));
2407 V[i] = getZeroVector(VT, DAG);
2409 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2412 for (unsigned i = 0; i < 2; ++i) {
2413 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
2416 V[i] = V[i*2]; // Must be a zero vector.
2419 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
2420 getMOVLMask(NumElems, DAG));
2423 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2424 getMOVLMask(NumElems, DAG));
2427 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2428 getUnpacklMask(NumElems, DAG));
2433 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
2434 // clears the upper bits.
2435 // FIXME: we can do the same for v4f32 case when we know both parts of
2436 // the lower half come from scalar_to_vector (loadf32). We should do
2437 // that in post legalizer dag combiner with target specific hooks.
2438 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
2440 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2441 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2442 SmallVector<SDOperand, 8> MaskVec;
2443 bool Reverse = (NonZeros & 0x3) == 2;
2444 for (unsigned i = 0; i < 2; ++i)
2446 MaskVec.push_back(DAG.getConstant(1-i, EVT));
2448 MaskVec.push_back(DAG.getConstant(i, EVT));
2449 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
2450 for (unsigned i = 0; i < 2; ++i)
2452 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
2454 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
2455 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2456 &MaskVec[0], MaskVec.size());
2457 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
2460 if (Values.size() > 2) {
2461 // Expand into a number of unpckl*.
2463 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
2464 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
2465 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
2466 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
2467 for (unsigned i = 0; i < NumElems; ++i)
2468 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2470 while (NumElems != 0) {
2471 for (unsigned i = 0; i < NumElems; ++i)
2472 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
2483 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
2484 SDOperand V1 = Op.getOperand(0);
2485 SDOperand V2 = Op.getOperand(1);
2486 SDOperand PermMask = Op.getOperand(2);
2487 MVT::ValueType VT = Op.getValueType();
2488 unsigned NumElems = PermMask.getNumOperands();
2489 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
2490 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
2491 bool V1IsSplat = false;
2492 bool V2IsSplat = false;
2494 if (isUndefShuffle(Op.Val))
2495 return DAG.getNode(ISD::UNDEF, VT);
2497 if (isSplatMask(PermMask.Val)) {
2498 if (NumElems <= 4) return Op;
2499 // Promote it to a v4i32 splat.
2500 return PromoteSplat(Op, DAG);
2503 if (X86::isMOVLMask(PermMask.Val))
2504 return (V1IsUndef) ? V2 : Op;
2506 if (X86::isMOVSHDUPMask(PermMask.Val) ||
2507 X86::isMOVSLDUPMask(PermMask.Val) ||
2508 X86::isMOVHLPSMask(PermMask.Val) ||
2509 X86::isMOVHPMask(PermMask.Val) ||
2510 X86::isMOVLPMask(PermMask.Val))
2513 if (ShouldXformToMOVHLPS(PermMask.Val) ||
2514 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
2515 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2517 bool Commuted = false;
2518 V1IsSplat = isSplatVector(V1.Val);
2519 V2IsSplat = isSplatVector(V2.Val);
2520 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
2521 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2522 std::swap(V1IsSplat, V2IsSplat);
2523 std::swap(V1IsUndef, V2IsUndef);
2527 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
2528 if (V2IsUndef) return V1;
2529 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2531 // V2 is a splat, so the mask may be malformed. That is, it may point
2532 // to any V2 element. The instruction selectior won't like this. Get
2533 // a corrected mask and commute to form a proper MOVS{S|D}.
2534 SDOperand NewMask = getMOVLMask(NumElems, DAG);
2535 if (NewMask.Val != PermMask.Val)
2536 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2541 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
2542 X86::isUNPCKLMask(PermMask.Val) ||
2543 X86::isUNPCKHMask(PermMask.Val))
2547 // Normalize mask so all entries that point to V2 points to its first
2548 // element then try to match unpck{h|l} again. If match, return a
2549 // new vector_shuffle with the corrected mask.
2550 SDOperand NewMask = NormalizeMask(PermMask, DAG);
2551 if (NewMask.Val != PermMask.Val) {
2552 if (X86::isUNPCKLMask(PermMask.Val, true)) {
2553 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
2554 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2555 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
2556 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
2557 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2562 // Normalize the node to match x86 shuffle ops if needed
2563 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
2564 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2567 // Commute is back and try unpck* again.
2568 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2569 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
2570 X86::isUNPCKLMask(PermMask.Val) ||
2571 X86::isUNPCKHMask(PermMask.Val))
2575 // If VT is integer, try PSHUF* first, then SHUFP*.
2576 if (MVT::isInteger(VT)) {
2577 if (X86::isPSHUFDMask(PermMask.Val) ||
2578 X86::isPSHUFHWMask(PermMask.Val) ||
2579 X86::isPSHUFLWMask(PermMask.Val)) {
2580 if (V2.getOpcode() != ISD::UNDEF)
2581 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
2582 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
2586 if (X86::isSHUFPMask(PermMask.Val))
2589 // Handle v8i16 shuffle high / low shuffle node pair.
2590 if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
2591 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2592 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2593 SmallVector<SDOperand, 8> MaskVec;
2594 for (unsigned i = 0; i != 4; ++i)
2595 MaskVec.push_back(PermMask.getOperand(i));
2596 for (unsigned i = 4; i != 8; ++i)
2597 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2598 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2599 &MaskVec[0], MaskVec.size());
2600 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2602 for (unsigned i = 0; i != 4; ++i)
2603 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2604 for (unsigned i = 4; i != 8; ++i)
2605 MaskVec.push_back(PermMask.getOperand(i));
2606 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0],MaskVec.size());
2607 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2610 // Floating point cases in the other order.
2611 if (X86::isSHUFPMask(PermMask.Val))
2613 if (X86::isPSHUFDMask(PermMask.Val) ||
2614 X86::isPSHUFHWMask(PermMask.Val) ||
2615 X86::isPSHUFLWMask(PermMask.Val)) {
2616 if (V2.getOpcode() != ISD::UNDEF)
2617 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
2618 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
2623 if (NumElems == 4) {
2624 MVT::ValueType MaskVT = PermMask.getValueType();
2625 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
2626 SmallVector<std::pair<int, int>, 8> Locs;
2627 Locs.reserve(NumElems);
2628 SmallVector<SDOperand, 8> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2629 SmallVector<SDOperand, 8> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2632 // If no more than two elements come from either vector. This can be
2633 // implemented with two shuffles. First shuffle gather the elements.
2634 // The second shuffle, which takes the first shuffle as both of its
2635 // vector operands, put the elements into the right order.
2636 for (unsigned i = 0; i != NumElems; ++i) {
2637 SDOperand Elt = PermMask.getOperand(i);
2638 if (Elt.getOpcode() == ISD::UNDEF) {
2639 Locs[i] = std::make_pair(-1, -1);
2641 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
2642 if (Val < NumElems) {
2643 Locs[i] = std::make_pair(0, NumLo);
2647 Locs[i] = std::make_pair(1, NumHi);
2648 if (2+NumHi < NumElems)
2649 Mask1[2+NumHi] = Elt;
2654 if (NumLo <= 2 && NumHi <= 2) {
2655 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2656 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2657 &Mask1[0], Mask1.size()));
2658 for (unsigned i = 0; i != NumElems; ++i) {
2659 if (Locs[i].first == -1)
2662 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
2663 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
2664 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
2668 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
2669 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2670 &Mask2[0], Mask2.size()));
2673 // Break it into (shuffle shuffle_hi, shuffle_lo).
2675 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2676 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2677 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
2678 unsigned MaskIdx = 0;
2680 unsigned HiIdx = NumElems/2;
2681 for (unsigned i = 0; i != NumElems; ++i) {
2682 if (i == NumElems/2) {
2688 SDOperand Elt = PermMask.getOperand(i);
2689 if (Elt.getOpcode() == ISD::UNDEF) {
2690 Locs[i] = std::make_pair(-1, -1);
2691 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
2692 Locs[i] = std::make_pair(MaskIdx, LoIdx);
2693 (*MaskPtr)[LoIdx] = Elt;
2696 Locs[i] = std::make_pair(MaskIdx, HiIdx);
2697 (*MaskPtr)[HiIdx] = Elt;
2702 SDOperand LoShuffle =
2703 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2704 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2705 &LoMask[0], LoMask.size()));
2706 SDOperand HiShuffle =
2707 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2708 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2709 &HiMask[0], HiMask.size()));
2710 SmallVector<SDOperand, 8> MaskOps;
2711 for (unsigned i = 0; i != NumElems; ++i) {
2712 if (Locs[i].first == -1) {
2713 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
2715 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
2716 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
2719 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
2720 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2721 &MaskOps[0], MaskOps.size()));
2728 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2729 if (!isa<ConstantSDNode>(Op.getOperand(1)))
2732 MVT::ValueType VT = Op.getValueType();
2733 // TODO: handle v16i8.
2734 if (MVT::getSizeInBits(VT) == 16) {
2735 // Transform it so it match pextrw which produces a 32-bit result.
2736 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
2737 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
2738 Op.getOperand(0), Op.getOperand(1));
2739 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
2740 DAG.getValueType(VT));
2741 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
2742 } else if (MVT::getSizeInBits(VT) == 32) {
2743 SDOperand Vec = Op.getOperand(0);
2744 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
2747 // SHUFPS the element to the lowest double word, then movss.
2748 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2749 SmallVector<SDOperand, 8> IdxVec;
2750 IdxVec.push_back(DAG.getConstant(Idx, MVT::getVectorBaseType(MaskVT)));
2751 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2752 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2753 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2754 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2755 &IdxVec[0], IdxVec.size());
2756 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
2757 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
2758 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
2759 DAG.getConstant(0, getPointerTy()));
2760 } else if (MVT::getSizeInBits(VT) == 64) {
2761 SDOperand Vec = Op.getOperand(0);
2762 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
2766 // UNPCKHPD the element to the lowest double word, then movsd.
2767 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
2768 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
2769 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2770 SmallVector<SDOperand, 8> IdxVec;
2771 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorBaseType(MaskVT)));
2772 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2773 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2774 &IdxVec[0], IdxVec.size());
2775 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
2776 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
2777 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
2778 DAG.getConstant(0, getPointerTy()));
2785 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2786 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
2787 // as its second argument.
2788 MVT::ValueType VT = Op.getValueType();
2789 MVT::ValueType BaseVT = MVT::getVectorBaseType(VT);
2790 SDOperand N0 = Op.getOperand(0);
2791 SDOperand N1 = Op.getOperand(1);
2792 SDOperand N2 = Op.getOperand(2);
2793 if (MVT::getSizeInBits(BaseVT) == 16) {
2794 if (N1.getValueType() != MVT::i32)
2795 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
2796 if (N2.getValueType() != MVT::i32)
2797 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(), MVT::i32);
2798 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
2799 } else if (MVT::getSizeInBits(BaseVT) == 32) {
2800 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
2803 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
2804 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2805 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2806 SmallVector<SDOperand, 8> MaskVec;
2807 MaskVec.push_back(DAG.getConstant(4, BaseVT));
2808 for (unsigned i = 1; i <= 3; ++i)
2809 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2810 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
2811 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2812 &MaskVec[0], MaskVec.size()));
2814 // Use two pinsrw instructions to insert a 32 bit value.
2816 if (MVT::isFloatingPoint(N1.getValueType())) {
2817 if (ISD::isNON_EXTLoad(N1.Val)) {
2818 // Just load directly from f32mem to GR32.
2819 LoadSDNode *LD = cast<LoadSDNode>(N1);
2820 N1 = DAG.getLoad(MVT::i32, LD->getChain(), LD->getBasePtr(),
2821 LD->getSrcValue(), LD->getSrcValueOffset());
2823 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
2824 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
2825 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
2826 DAG.getConstant(0, getPointerTy()));
2829 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
2830 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
2831 DAG.getConstant(Idx, getPointerTy()));
2832 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
2833 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
2834 DAG.getConstant(Idx+1, getPointerTy()));
2835 return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
2843 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2844 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
2845 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
2848 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2849 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
2850 // one of the above mentioned nodes. It has to be wrapped because otherwise
2851 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2852 // be used to form addressing mode. These wrapped nodes will be selected
2855 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
2856 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2857 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
2859 CP->getAlignment());
2860 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
2861 // With PIC, the address is actually $g + Offset.
2862 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2863 !Subtarget->isPICStyleRIPRel()) {
2864 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2865 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
2873 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
2874 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2875 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
2876 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
2877 // With PIC, the address is actually $g + Offset.
2878 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2879 !Subtarget->isPICStyleRIPRel()) {
2880 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2881 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
2885 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
2886 // load the value at address GV, not the value of GV itself. This means that
2887 // the GlobalAddress must be in the base or index register of the address, not
2888 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
2889 // The same applies for external symbols during PIC codegen
2890 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
2891 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
2897 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
2898 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
2899 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
2900 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
2901 // With PIC, the address is actually $g + Offset.
2902 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2903 !Subtarget->isPICStyleRIPRel()) {
2904 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2905 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
2912 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
2913 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
2914 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
2915 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
2916 // With PIC, the address is actually $g + Offset.
2917 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2918 !Subtarget->isPICStyleRIPRel()) {
2919 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2920 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
2927 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
2928 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
2929 "Not an i64 shift!");
2930 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
2931 SDOperand ShOpLo = Op.getOperand(0);
2932 SDOperand ShOpHi = Op.getOperand(1);
2933 SDOperand ShAmt = Op.getOperand(2);
2934 SDOperand Tmp1 = isSRA ?
2935 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
2936 DAG.getConstant(0, MVT::i32);
2938 SDOperand Tmp2, Tmp3;
2939 if (Op.getOpcode() == ISD::SHL_PARTS) {
2940 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
2941 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
2943 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
2944 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
2947 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
2948 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
2949 DAG.getConstant(32, MVT::i8));
2950 SDOperand COps[]={DAG.getEntryNode(), AndNode, DAG.getConstant(0, MVT::i8)};
2951 SDOperand InFlag = DAG.getNode(X86ISD::CMP, VTs, 2, COps, 3).getValue(1);
2954 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
2956 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
2957 SmallVector<SDOperand, 4> Ops;
2958 if (Op.getOpcode() == ISD::SHL_PARTS) {
2959 Ops.push_back(Tmp2);
2960 Ops.push_back(Tmp3);
2962 Ops.push_back(InFlag);
2963 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
2964 InFlag = Hi.getValue(1);
2967 Ops.push_back(Tmp3);
2968 Ops.push_back(Tmp1);
2970 Ops.push_back(InFlag);
2971 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
2973 Ops.push_back(Tmp2);
2974 Ops.push_back(Tmp3);
2976 Ops.push_back(InFlag);
2977 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
2978 InFlag = Lo.getValue(1);
2981 Ops.push_back(Tmp3);
2982 Ops.push_back(Tmp1);
2984 Ops.push_back(InFlag);
2985 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
2988 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
2992 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
2995 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
2996 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
2997 Op.getOperand(0).getValueType() >= MVT::i16 &&
2998 "Unknown SINT_TO_FP to lower!");
3001 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
3002 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
3003 MachineFunction &MF = DAG.getMachineFunction();
3004 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
3005 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3006 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
3007 StackSlot, NULL, 0);
3012 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
3014 Tys = DAG.getVTList(MVT::f64, MVT::Other);
3015 SmallVector<SDOperand, 8> Ops;
3016 Ops.push_back(Chain);
3017 Ops.push_back(StackSlot);
3018 Ops.push_back(DAG.getValueType(SrcVT));
3019 Result = DAG.getNode(X86ScalarSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
3020 Tys, &Ops[0], Ops.size());
3023 Chain = Result.getValue(1);
3024 SDOperand InFlag = Result.getValue(2);
3026 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
3027 // shouldn't be necessary except that RFP cannot be live across
3028 // multiple blocks. When stackifier is fixed, they can be uncoupled.
3029 MachineFunction &MF = DAG.getMachineFunction();
3030 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3031 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3032 Tys = DAG.getVTList(MVT::Other);
3033 SmallVector<SDOperand, 8> Ops;
3034 Ops.push_back(Chain);
3035 Ops.push_back(Result);
3036 Ops.push_back(StackSlot);
3037 Ops.push_back(DAG.getValueType(Op.getValueType()));
3038 Ops.push_back(InFlag);
3039 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
3040 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
3046 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
3047 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
3048 "Unknown FP_TO_SINT to lower!");
3049 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
3051 MachineFunction &MF = DAG.getMachineFunction();
3052 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
3053 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3054 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3057 switch (Op.getValueType()) {
3058 default: assert(0 && "Invalid FP_TO_SINT to lower!");
3059 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
3060 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
3061 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
3064 SDOperand Chain = DAG.getEntryNode();
3065 SDOperand Value = Op.getOperand(0);
3067 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
3068 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
3069 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
3071 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
3073 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
3074 Chain = Value.getValue(1);
3075 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3076 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3079 // Build the FP_TO_INT*_IN_MEM
3080 SDOperand Ops[] = { Chain, Value, StackSlot };
3081 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
3084 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
3087 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
3088 MVT::ValueType VT = Op.getValueType();
3089 const Type *OpNTy = MVT::getTypeForValueType(VT);
3090 std::vector<Constant*> CV;
3091 if (VT == MVT::f64) {
3092 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(~(1ULL << 63))));
3093 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3095 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(~(1U << 31))));
3096 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3097 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3098 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3100 Constant *CS = ConstantStruct::get(CV);
3101 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3102 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3103 SmallVector<SDOperand, 3> Ops;
3104 Ops.push_back(DAG.getEntryNode());
3105 Ops.push_back(CPIdx);
3106 Ops.push_back(DAG.getSrcValue(NULL));
3107 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3108 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
3111 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
3112 MVT::ValueType VT = Op.getValueType();
3113 const Type *OpNTy = MVT::getTypeForValueType(VT);
3114 std::vector<Constant*> CV;
3115 if (VT == MVT::f64) {
3116 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(1ULL << 63)));
3117 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3119 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(1U << 31)));
3120 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3121 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3122 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3124 Constant *CS = ConstantStruct::get(CV);
3125 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3126 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3127 SmallVector<SDOperand, 3> Ops;
3128 Ops.push_back(DAG.getEntryNode());
3129 Ops.push_back(CPIdx);
3130 Ops.push_back(DAG.getSrcValue(NULL));
3131 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3132 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
3135 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
3136 SDOperand Op0 = Op.getOperand(0);
3137 SDOperand Op1 = Op.getOperand(1);
3138 MVT::ValueType VT = Op.getValueType();
3139 MVT::ValueType SrcVT = Op1.getValueType();
3140 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
3142 // If second operand is smaller, extend it first.
3143 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
3144 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
3148 // First get the sign bit of second operand.
3149 std::vector<Constant*> CV;
3150 if (SrcVT == MVT::f64) {
3151 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(1ULL << 63)));
3152 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3154 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(1U << 31)));
3155 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3156 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3157 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3159 Constant *CS = ConstantStruct::get(CV);
3160 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3161 SDVTList Tys = DAG.getVTList(SrcVT, MVT::Other);
3162 SmallVector<SDOperand, 3> Ops;
3163 Ops.push_back(DAG.getEntryNode());
3164 Ops.push_back(CPIdx);
3165 Ops.push_back(DAG.getSrcValue(NULL));
3166 SDOperand Mask1 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3167 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
3169 // Shift sign bit right or left if the two operands have different types.
3170 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
3171 // Op0 is MVT::f32, Op1 is MVT::f64.
3172 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
3173 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
3174 DAG.getConstant(32, MVT::i32));
3175 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
3176 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
3177 DAG.getConstant(0, getPointerTy()));
3180 // Clear first operand sign bit.
3182 if (VT == MVT::f64) {
3183 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(~(1ULL << 63))));
3184 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3186 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(~(1U << 31))));
3187 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3188 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3189 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3191 CS = ConstantStruct::get(CV);
3192 CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3193 Tys = DAG.getVTList(VT, MVT::Other);
3195 Ops.push_back(DAG.getEntryNode());
3196 Ops.push_back(CPIdx);
3197 Ops.push_back(DAG.getSrcValue(NULL));
3198 SDOperand Mask2 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3199 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
3201 // Or the value with the sign bit.
3202 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
3205 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG,
3207 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
3209 SDOperand Op0 = Op.getOperand(0);
3210 SDOperand Op1 = Op.getOperand(1);
3211 SDOperand CC = Op.getOperand(2);
3212 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3213 const MVT::ValueType *VTs1 = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3214 const MVT::ValueType *VTs2 = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
3215 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
3218 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
3220 SDOperand Ops1[] = { Chain, Op0, Op1 };
3221 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, Ops1, 3).getValue(1);
3222 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
3223 return DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3226 assert(isFP && "Illegal integer SetCC!");
3228 SDOperand COps[] = { Chain, Op0, Op1 };
3229 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, COps, 3).getValue(1);
3231 switch (SetCCOpcode) {
3232 default: assert(false && "Illegal floating point SetCC!");
3233 case ISD::SETOEQ: { // !PF & ZF
3234 SDOperand Ops1[] = { DAG.getConstant(X86::COND_NP, MVT::i8), Cond };
3235 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3236 SDOperand Ops2[] = { DAG.getConstant(X86::COND_E, MVT::i8),
3238 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3239 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
3241 case ISD::SETUNE: { // PF | !ZF
3242 SDOperand Ops1[] = { DAG.getConstant(X86::COND_P, MVT::i8), Cond };
3243 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3244 SDOperand Ops2[] = { DAG.getConstant(X86::COND_NE, MVT::i8),
3246 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3247 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
3252 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
3253 bool addTest = true;
3254 SDOperand Chain = DAG.getEntryNode();
3255 SDOperand Cond = Op.getOperand(0);
3257 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3259 if (Cond.getOpcode() == ISD::SETCC)
3260 Cond = LowerSETCC(Cond, DAG, Chain);
3262 if (Cond.getOpcode() == X86ISD::SETCC) {
3263 CC = Cond.getOperand(0);
3265 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3266 // (since flag operand cannot be shared). Use it as the condition setting
3267 // operand in place of the X86ISD::SETCC.
3268 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3269 // to use a test instead of duplicating the X86ISD::CMP (for register
3270 // pressure reason)?
3271 SDOperand Cmp = Cond.getOperand(1);
3272 unsigned Opc = Cmp.getOpcode();
3273 bool IllegalFPCMov = !X86ScalarSSE &&
3274 MVT::isFloatingPoint(Op.getValueType()) &&
3275 !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
3276 if ((Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) &&
3278 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3279 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3285 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3286 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3287 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3290 VTs = DAG.getNodeValueTypes(Op.getValueType(), MVT::Flag);
3291 SmallVector<SDOperand, 4> Ops;
3292 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
3293 // condition is true.
3294 Ops.push_back(Op.getOperand(2));
3295 Ops.push_back(Op.getOperand(1));
3297 Ops.push_back(Cond.getValue(1));
3298 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3301 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
3302 bool addTest = true;
3303 SDOperand Chain = Op.getOperand(0);
3304 SDOperand Cond = Op.getOperand(1);
3305 SDOperand Dest = Op.getOperand(2);
3307 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3309 if (Cond.getOpcode() == ISD::SETCC)
3310 Cond = LowerSETCC(Cond, DAG, Chain);
3312 if (Cond.getOpcode() == X86ISD::SETCC) {
3313 CC = Cond.getOperand(0);
3315 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3316 // (since flag operand cannot be shared). Use it as the condition setting
3317 // operand in place of the X86ISD::SETCC.
3318 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3319 // to use a test instead of duplicating the X86ISD::CMP (for register
3320 // pressure reason)?
3321 SDOperand Cmp = Cond.getOperand(1);
3322 unsigned Opc = Cmp.getOpcode();
3323 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) {
3324 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3325 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3331 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3332 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3333 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3335 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
3336 Cond, Op.getOperand(2), CC, Cond.getValue(1));
3339 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
3340 unsigned CallingConv= cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3342 if (Subtarget->is64Bit())
3343 return LowerX86_64CCCCallTo(Op, DAG, CallingConv);
3345 switch (CallingConv) {
3347 assert(0 && "Unsupported calling convention");
3348 case CallingConv::Fast:
3349 // TODO: Implement fastcc
3351 case CallingConv::C:
3352 case CallingConv::X86_StdCall:
3353 return LowerCCCCallTo(Op, DAG, CallingConv);
3354 case CallingConv::X86_FastCall:
3355 return LowerFastCCCallTo(Op, DAG, CallingConv);
3360 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
3361 MachineFunction &MF = DAG.getMachineFunction();
3362 const Function* Fn = MF.getFunction();
3363 if (Fn->hasExternalLinkage() &&
3364 Subtarget->isTargetCygMing() &&
3365 Fn->getName() == "main")
3366 MF.getInfo<X86FunctionInfo>()->setForceFramePointer(true);
3368 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3369 if (Subtarget->is64Bit())
3370 return LowerX86_64CCCArguments(Op, DAG);
3374 assert(0 && "Unsupported calling convention");
3375 case CallingConv::Fast:
3376 // TODO: implement fastcc.
3379 case CallingConv::C:
3380 return LowerCCCArguments(Op, DAG);
3381 case CallingConv::X86_StdCall:
3382 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(StdCall);
3383 return LowerCCCArguments(Op, DAG, true);
3384 case CallingConv::X86_FastCall:
3385 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(FastCall);
3386 return LowerFastCCArguments(Op, DAG);
3390 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
3391 SDOperand InFlag(0, 0);
3392 SDOperand Chain = Op.getOperand(0);
3394 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3395 if (Align == 0) Align = 1;
3397 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3398 // If not DWORD aligned, call memset if size is less than the threshold.
3399 // It knows how to align to the right boundary first.
3400 if ((Align & 3) != 0 ||
3401 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
3402 MVT::ValueType IntPtr = getPointerTy();
3403 const Type *IntPtrTy = getTargetData()->getIntPtrType();
3404 TargetLowering::ArgListTy Args;
3405 TargetLowering::ArgListEntry Entry;
3406 Entry.Node = Op.getOperand(1);
3407 Entry.Ty = IntPtrTy;
3408 Entry.isSigned = false;
3409 Entry.isInReg = false;
3410 Entry.isSRet = false;
3411 Args.push_back(Entry);
3412 // Extend the unsigned i8 argument to be an int value for the call.
3413 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
3414 Entry.Ty = IntPtrTy;
3415 Entry.isSigned = false;
3416 Entry.isInReg = false;
3417 Entry.isSRet = false;
3418 Args.push_back(Entry);
3419 Entry.Node = Op.getOperand(3);
3420 Args.push_back(Entry);
3421 std::pair<SDOperand,SDOperand> CallResult =
3422 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
3423 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
3424 return CallResult.second;
3429 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
3430 unsigned BytesLeft = 0;
3431 bool TwoRepStos = false;
3434 uint64_t Val = ValC->getValue() & 255;
3436 // If the value is a constant, then we can potentially use larger sets.
3437 switch (Align & 3) {
3438 case 2: // WORD aligned
3441 Val = (Val << 8) | Val;
3443 case 0: // DWORD aligned
3446 Val = (Val << 8) | Val;
3447 Val = (Val << 16) | Val;
3448 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
3451 Val = (Val << 32) | Val;
3454 default: // Byte aligned
3457 Count = Op.getOperand(3);
3461 if (AVT > MVT::i8) {
3463 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
3464 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
3465 BytesLeft = I->getValue() % UBytes;
3467 assert(AVT >= MVT::i32 &&
3468 "Do not use rep;stos if not at least DWORD aligned");
3469 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
3470 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
3475 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
3477 InFlag = Chain.getValue(1);
3480 Count = Op.getOperand(3);
3481 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
3482 InFlag = Chain.getValue(1);
3485 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
3487 InFlag = Chain.getValue(1);
3488 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
3489 Op.getOperand(1), InFlag);
3490 InFlag = Chain.getValue(1);
3492 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3493 SmallVector<SDOperand, 8> Ops;
3494 Ops.push_back(Chain);
3495 Ops.push_back(DAG.getValueType(AVT));
3496 Ops.push_back(InFlag);
3497 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
3500 InFlag = Chain.getValue(1);
3501 Count = Op.getOperand(3);
3502 MVT::ValueType CVT = Count.getValueType();
3503 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
3504 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
3505 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
3507 InFlag = Chain.getValue(1);
3508 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3510 Ops.push_back(Chain);
3511 Ops.push_back(DAG.getValueType(MVT::i8));
3512 Ops.push_back(InFlag);
3513 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
3514 } else if (BytesLeft) {
3515 // Issue stores for the last 1 - 7 bytes.
3517 unsigned Val = ValC->getValue() & 255;
3518 unsigned Offset = I->getValue() - BytesLeft;
3519 SDOperand DstAddr = Op.getOperand(1);
3520 MVT::ValueType AddrVT = DstAddr.getValueType();
3521 if (BytesLeft >= 4) {
3522 Val = (Val << 8) | Val;
3523 Val = (Val << 16) | Val;
3524 Value = DAG.getConstant(Val, MVT::i32);
3525 Chain = DAG.getStore(Chain, Value,
3526 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3527 DAG.getConstant(Offset, AddrVT)),
3532 if (BytesLeft >= 2) {
3533 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
3534 Chain = DAG.getStore(Chain, Value,
3535 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3536 DAG.getConstant(Offset, AddrVT)),
3541 if (BytesLeft == 1) {
3542 Value = DAG.getConstant(Val, MVT::i8);
3543 Chain = DAG.getStore(Chain, Value,
3544 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3545 DAG.getConstant(Offset, AddrVT)),
3553 SDOperand X86TargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) {
3554 SDOperand Chain = Op.getOperand(0);
3556 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3557 if (Align == 0) Align = 1;
3559 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3560 // If not DWORD aligned, call memcpy if size is less than the threshold.
3561 // It knows how to align to the right boundary first.
3562 if ((Align & 3) != 0 ||
3563 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
3564 MVT::ValueType IntPtr = getPointerTy();
3565 TargetLowering::ArgListTy Args;
3566 TargetLowering::ArgListEntry Entry;
3567 Entry.Ty = getTargetData()->getIntPtrType();
3568 Entry.isSigned = false;
3569 Entry.isInReg = false;
3570 Entry.isSRet = false;
3571 Entry.Node = Op.getOperand(1); Args.push_back(Entry);
3572 Entry.Node = Op.getOperand(2); Args.push_back(Entry);
3573 Entry.Node = Op.getOperand(3); Args.push_back(Entry);
3574 std::pair<SDOperand,SDOperand> CallResult =
3575 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
3576 DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
3577 return CallResult.second;
3582 unsigned BytesLeft = 0;
3583 bool TwoRepMovs = false;
3584 switch (Align & 3) {
3585 case 2: // WORD aligned
3588 case 0: // DWORD aligned
3590 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
3593 default: // Byte aligned
3595 Count = Op.getOperand(3);
3599 if (AVT > MVT::i8) {
3601 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
3602 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
3603 BytesLeft = I->getValue() % UBytes;
3605 assert(AVT >= MVT::i32 &&
3606 "Do not use rep;movs if not at least DWORD aligned");
3607 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
3608 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
3613 SDOperand InFlag(0, 0);
3614 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
3616 InFlag = Chain.getValue(1);
3617 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
3618 Op.getOperand(1), InFlag);
3619 InFlag = Chain.getValue(1);
3620 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
3621 Op.getOperand(2), InFlag);
3622 InFlag = Chain.getValue(1);
3624 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3625 SmallVector<SDOperand, 8> Ops;
3626 Ops.push_back(Chain);
3627 Ops.push_back(DAG.getValueType(AVT));
3628 Ops.push_back(InFlag);
3629 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
3632 InFlag = Chain.getValue(1);
3633 Count = Op.getOperand(3);
3634 MVT::ValueType CVT = Count.getValueType();
3635 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
3636 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
3637 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
3639 InFlag = Chain.getValue(1);
3640 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3642 Ops.push_back(Chain);
3643 Ops.push_back(DAG.getValueType(MVT::i8));
3644 Ops.push_back(InFlag);
3645 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
3646 } else if (BytesLeft) {
3647 // Issue loads and stores for the last 1 - 7 bytes.
3648 unsigned Offset = I->getValue() - BytesLeft;
3649 SDOperand DstAddr = Op.getOperand(1);
3650 MVT::ValueType DstVT = DstAddr.getValueType();
3651 SDOperand SrcAddr = Op.getOperand(2);
3652 MVT::ValueType SrcVT = SrcAddr.getValueType();
3654 if (BytesLeft >= 4) {
3655 Value = DAG.getLoad(MVT::i32, Chain,
3656 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3657 DAG.getConstant(Offset, SrcVT)),
3659 Chain = Value.getValue(1);
3660 Chain = DAG.getStore(Chain, Value,
3661 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3662 DAG.getConstant(Offset, DstVT)),
3667 if (BytesLeft >= 2) {
3668 Value = DAG.getLoad(MVT::i16, Chain,
3669 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3670 DAG.getConstant(Offset, SrcVT)),
3672 Chain = Value.getValue(1);
3673 Chain = DAG.getStore(Chain, Value,
3674 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3675 DAG.getConstant(Offset, DstVT)),
3681 if (BytesLeft == 1) {
3682 Value = DAG.getLoad(MVT::i8, Chain,
3683 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3684 DAG.getConstant(Offset, SrcVT)),
3686 Chain = Value.getValue(1);
3687 Chain = DAG.getStore(Chain, Value,
3688 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3689 DAG.getConstant(Offset, DstVT)),
3698 X86TargetLowering::LowerREADCYCLCECOUNTER(SDOperand Op, SelectionDAG &DAG) {
3699 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3700 SDOperand TheOp = Op.getOperand(0);
3701 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheOp, 1);
3702 if (Subtarget->is64Bit()) {
3703 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
3704 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::RDX,
3705 MVT::i64, Copy1.getValue(2));
3706 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, Copy2,
3707 DAG.getConstant(32, MVT::i8));
3709 DAG.getNode(ISD::OR, MVT::i64, Copy1, Tmp), Copy2.getValue(1)
3712 Tys = DAG.getVTList(MVT::i64, MVT::Other);
3713 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2);
3716 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
3717 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::EDX,
3718 MVT::i32, Copy1.getValue(2));
3719 SDOperand Ops[] = { Copy1, Copy2, Copy2.getValue(1) };
3720 Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
3721 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 3);
3724 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
3725 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
3727 if (!Subtarget->is64Bit()) {
3728 // vastart just stores the address of the VarArgsFrameIndex slot into the
3729 // memory location argument.
3730 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
3731 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
3736 // gp_offset (0 - 6 * 8)
3737 // fp_offset (48 - 48 + 8 * 16)
3738 // overflow_arg_area (point to parameters coming in memory).
3740 SmallVector<SDOperand, 8> MemOps;
3741 SDOperand FIN = Op.getOperand(1);
3743 SDOperand Store = DAG.getStore(Op.getOperand(0),
3744 DAG.getConstant(VarArgsGPOffset, MVT::i32),
3745 FIN, SV->getValue(), SV->getOffset());
3746 MemOps.push_back(Store);
3749 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
3750 DAG.getConstant(4, getPointerTy()));
3751 Store = DAG.getStore(Op.getOperand(0),
3752 DAG.getConstant(VarArgsFPOffset, MVT::i32),
3753 FIN, SV->getValue(), SV->getOffset());
3754 MemOps.push_back(Store);
3756 // Store ptr to overflow_arg_area
3757 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
3758 DAG.getConstant(4, getPointerTy()));
3759 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
3760 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
3762 MemOps.push_back(Store);
3764 // Store ptr to reg_save_area.
3765 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
3766 DAG.getConstant(8, getPointerTy()));
3767 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
3768 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
3770 MemOps.push_back(Store);
3771 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
3775 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
3776 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
3778 default: return SDOperand(); // Don't custom lower most intrinsics.
3779 // Comparison intrinsics.
3780 case Intrinsic::x86_sse_comieq_ss:
3781 case Intrinsic::x86_sse_comilt_ss:
3782 case Intrinsic::x86_sse_comile_ss:
3783 case Intrinsic::x86_sse_comigt_ss:
3784 case Intrinsic::x86_sse_comige_ss:
3785 case Intrinsic::x86_sse_comineq_ss:
3786 case Intrinsic::x86_sse_ucomieq_ss:
3787 case Intrinsic::x86_sse_ucomilt_ss:
3788 case Intrinsic::x86_sse_ucomile_ss:
3789 case Intrinsic::x86_sse_ucomigt_ss:
3790 case Intrinsic::x86_sse_ucomige_ss:
3791 case Intrinsic::x86_sse_ucomineq_ss:
3792 case Intrinsic::x86_sse2_comieq_sd:
3793 case Intrinsic::x86_sse2_comilt_sd:
3794 case Intrinsic::x86_sse2_comile_sd:
3795 case Intrinsic::x86_sse2_comigt_sd:
3796 case Intrinsic::x86_sse2_comige_sd:
3797 case Intrinsic::x86_sse2_comineq_sd:
3798 case Intrinsic::x86_sse2_ucomieq_sd:
3799 case Intrinsic::x86_sse2_ucomilt_sd:
3800 case Intrinsic::x86_sse2_ucomile_sd:
3801 case Intrinsic::x86_sse2_ucomigt_sd:
3802 case Intrinsic::x86_sse2_ucomige_sd:
3803 case Intrinsic::x86_sse2_ucomineq_sd: {
3805 ISD::CondCode CC = ISD::SETCC_INVALID;
3808 case Intrinsic::x86_sse_comieq_ss:
3809 case Intrinsic::x86_sse2_comieq_sd:
3813 case Intrinsic::x86_sse_comilt_ss:
3814 case Intrinsic::x86_sse2_comilt_sd:
3818 case Intrinsic::x86_sse_comile_ss:
3819 case Intrinsic::x86_sse2_comile_sd:
3823 case Intrinsic::x86_sse_comigt_ss:
3824 case Intrinsic::x86_sse2_comigt_sd:
3828 case Intrinsic::x86_sse_comige_ss:
3829 case Intrinsic::x86_sse2_comige_sd:
3833 case Intrinsic::x86_sse_comineq_ss:
3834 case Intrinsic::x86_sse2_comineq_sd:
3838 case Intrinsic::x86_sse_ucomieq_ss:
3839 case Intrinsic::x86_sse2_ucomieq_sd:
3840 Opc = X86ISD::UCOMI;
3843 case Intrinsic::x86_sse_ucomilt_ss:
3844 case Intrinsic::x86_sse2_ucomilt_sd:
3845 Opc = X86ISD::UCOMI;
3848 case Intrinsic::x86_sse_ucomile_ss:
3849 case Intrinsic::x86_sse2_ucomile_sd:
3850 Opc = X86ISD::UCOMI;
3853 case Intrinsic::x86_sse_ucomigt_ss:
3854 case Intrinsic::x86_sse2_ucomigt_sd:
3855 Opc = X86ISD::UCOMI;
3858 case Intrinsic::x86_sse_ucomige_ss:
3859 case Intrinsic::x86_sse2_ucomige_sd:
3860 Opc = X86ISD::UCOMI;
3863 case Intrinsic::x86_sse_ucomineq_ss:
3864 case Intrinsic::x86_sse2_ucomineq_sd:
3865 Opc = X86ISD::UCOMI;
3871 SDOperand LHS = Op.getOperand(1);
3872 SDOperand RHS = Op.getOperand(2);
3873 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
3875 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3876 SDOperand Ops1[] = { DAG.getEntryNode(), LHS, RHS };
3877 SDOperand Cond = DAG.getNode(Opc, VTs, 2, Ops1, 3);
3878 VTs = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
3879 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
3880 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, VTs, 2, Ops2, 2);
3881 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
3886 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
3887 // Depths > 0 not supported yet!
3888 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
3891 // Just load the return address
3892 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
3893 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
3896 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
3897 // Depths > 0 not supported yet!
3898 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
3901 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
3902 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
3903 DAG.getConstant(4, getPointerTy()));
3906 /// LowerOperation - Provide custom lowering hooks for some operations.
3908 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
3909 switch (Op.getOpcode()) {
3910 default: assert(0 && "Should not custom lower this!");
3911 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3912 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3913 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3914 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
3915 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
3916 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3917 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
3918 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
3919 case ISD::SHL_PARTS:
3920 case ISD::SRA_PARTS:
3921 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
3922 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
3923 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
3924 case ISD::FABS: return LowerFABS(Op, DAG);
3925 case ISD::FNEG: return LowerFNEG(Op, DAG);
3926 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3927 case ISD::SETCC: return LowerSETCC(Op, DAG, DAG.getEntryNode());
3928 case ISD::SELECT: return LowerSELECT(Op, DAG);
3929 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
3930 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
3931 case ISD::CALL: return LowerCALL(Op, DAG);
3932 case ISD::RET: return LowerRET(Op, DAG);
3933 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
3934 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
3935 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
3936 case ISD::READCYCLECOUNTER: return LowerREADCYCLCECOUNTER(Op, DAG);
3937 case ISD::VASTART: return LowerVASTART(Op, DAG);
3938 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
3939 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3940 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3945 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
3947 default: return NULL;
3948 case X86ISD::SHLD: return "X86ISD::SHLD";
3949 case X86ISD::SHRD: return "X86ISD::SHRD";
3950 case X86ISD::FAND: return "X86ISD::FAND";
3951 case X86ISD::FOR: return "X86ISD::FOR";
3952 case X86ISD::FXOR: return "X86ISD::FXOR";
3953 case X86ISD::FSRL: return "X86ISD::FSRL";
3954 case X86ISD::FILD: return "X86ISD::FILD";
3955 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
3956 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
3957 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
3958 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
3959 case X86ISD::FLD: return "X86ISD::FLD";
3960 case X86ISD::FST: return "X86ISD::FST";
3961 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
3962 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
3963 case X86ISD::CALL: return "X86ISD::CALL";
3964 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
3965 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
3966 case X86ISD::CMP: return "X86ISD::CMP";
3967 case X86ISD::COMI: return "X86ISD::COMI";
3968 case X86ISD::UCOMI: return "X86ISD::UCOMI";
3969 case X86ISD::SETCC: return "X86ISD::SETCC";
3970 case X86ISD::CMOV: return "X86ISD::CMOV";
3971 case X86ISD::BRCOND: return "X86ISD::BRCOND";
3972 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
3973 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
3974 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
3975 case X86ISD::LOAD_PACK: return "X86ISD::LOAD_PACK";
3976 case X86ISD::LOAD_UA: return "X86ISD::LOAD_UA";
3977 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
3978 case X86ISD::Wrapper: return "X86ISD::Wrapper";
3979 case X86ISD::S2VEC: return "X86ISD::S2VEC";
3980 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
3981 case X86ISD::PINSRW: return "X86ISD::PINSRW";
3982 case X86ISD::FMAX: return "X86ISD::FMAX";
3983 case X86ISD::FMIN: return "X86ISD::FMIN";
3987 /// isLegalAddressImmediate - Return true if the integer value or
3988 /// GlobalValue can be used as the offset of the target addressing mode.
3989 bool X86TargetLowering::isLegalAddressImmediate(int64_t V) const {
3990 // X86 allows a sign-extended 32-bit immediate field.
3991 return (V > -(1LL << 32) && V < (1LL << 32)-1);
3994 bool X86TargetLowering::isLegalAddressImmediate(GlobalValue *GV) const {
3995 // In 64-bit mode, GV is 64-bit so it won't fit in the 32-bit displacement
3996 // field unless we are in small code model.
3997 if (Subtarget->is64Bit() &&
3998 getTargetMachine().getCodeModel() != CodeModel::Small)
4001 return (!Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false));
4004 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4005 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4006 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4007 /// are assumed to be legal.
4009 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
4010 // Only do shuffles on 128-bit vector types for now.
4011 if (MVT::getSizeInBits(VT) == 64) return false;
4012 return (Mask.Val->getNumOperands() <= 4 ||
4013 isSplatMask(Mask.Val) ||
4014 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
4015 X86::isUNPCKLMask(Mask.Val) ||
4016 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
4017 X86::isUNPCKHMask(Mask.Val));
4020 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
4022 SelectionDAG &DAG) const {
4023 unsigned NumElts = BVOps.size();
4024 // Only do shuffles on 128-bit vector types for now.
4025 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
4026 if (NumElts == 2) return true;
4028 return (isMOVLMask(&BVOps[0], 4) ||
4029 isCommutedMOVL(&BVOps[0], 4, true) ||
4030 isSHUFPMask(&BVOps[0], 4) ||
4031 isCommutedSHUFP(&BVOps[0], 4));
4036 //===----------------------------------------------------------------------===//
4037 // X86 Scheduler Hooks
4038 //===----------------------------------------------------------------------===//
4041 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
4042 MachineBasicBlock *BB) {
4043 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4044 switch (MI->getOpcode()) {
4045 default: assert(false && "Unexpected instr type to insert");
4046 case X86::CMOV_FR32:
4047 case X86::CMOV_FR64:
4048 case X86::CMOV_V4F32:
4049 case X86::CMOV_V2F64:
4050 case X86::CMOV_V2I64: {
4051 // To "insert" a SELECT_CC instruction, we actually have to insert the
4052 // diamond control-flow pattern. The incoming instruction knows the
4053 // destination vreg to set, the condition code register to branch on, the
4054 // true/false values to select between, and a branch opcode to use.
4055 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4056 ilist<MachineBasicBlock>::iterator It = BB;
4062 // cmpTY ccX, r1, r2
4064 // fallthrough --> copy0MBB
4065 MachineBasicBlock *thisMBB = BB;
4066 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
4067 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
4069 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
4070 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
4071 MachineFunction *F = BB->getParent();
4072 F->getBasicBlockList().insert(It, copy0MBB);
4073 F->getBasicBlockList().insert(It, sinkMBB);
4074 // Update machine-CFG edges by first adding all successors of the current
4075 // block to the new block which will contain the Phi node for the select.
4076 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
4077 e = BB->succ_end(); i != e; ++i)
4078 sinkMBB->addSuccessor(*i);
4079 // Next, remove all successors of the current block, and add the true
4080 // and fallthrough blocks as its successors.
4081 while(!BB->succ_empty())
4082 BB->removeSuccessor(BB->succ_begin());
4083 BB->addSuccessor(copy0MBB);
4084 BB->addSuccessor(sinkMBB);
4087 // %FalseValue = ...
4088 // # fallthrough to sinkMBB
4091 // Update machine-CFG edges
4092 BB->addSuccessor(sinkMBB);
4095 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4098 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
4099 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4100 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4102 delete MI; // The pseudo instruction is gone now.
4106 case X86::FP_TO_INT16_IN_MEM:
4107 case X86::FP_TO_INT32_IN_MEM:
4108 case X86::FP_TO_INT64_IN_MEM: {
4109 // Change the floating point control register to use "round towards zero"
4110 // mode when truncating to an integer value.
4111 MachineFunction *F = BB->getParent();
4112 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
4113 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
4115 // Load the old value of the high byte of the control word...
4117 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
4118 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
4120 // Set the high part to be round to zero...
4121 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
4124 // Reload the modified control word now...
4125 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4127 // Restore the memory image of control word to original value
4128 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
4131 // Get the X86 opcode to use.
4133 switch (MI->getOpcode()) {
4134 default: assert(0 && "illegal opcode!");
4135 case X86::FP_TO_INT16_IN_MEM: Opc = X86::FpIST16m; break;
4136 case X86::FP_TO_INT32_IN_MEM: Opc = X86::FpIST32m; break;
4137 case X86::FP_TO_INT64_IN_MEM: Opc = X86::FpIST64m; break;
4141 MachineOperand &Op = MI->getOperand(0);
4142 if (Op.isRegister()) {
4143 AM.BaseType = X86AddressMode::RegBase;
4144 AM.Base.Reg = Op.getReg();
4146 AM.BaseType = X86AddressMode::FrameIndexBase;
4147 AM.Base.FrameIndex = Op.getFrameIndex();
4149 Op = MI->getOperand(1);
4150 if (Op.isImmediate())
4151 AM.Scale = Op.getImm();
4152 Op = MI->getOperand(2);
4153 if (Op.isImmediate())
4154 AM.IndexReg = Op.getImm();
4155 Op = MI->getOperand(3);
4156 if (Op.isGlobalAddress()) {
4157 AM.GV = Op.getGlobal();
4159 AM.Disp = Op.getImm();
4161 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
4162 .addReg(MI->getOperand(4).getReg());
4164 // Reload the original control word now.
4165 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4167 delete MI; // The pseudo instruction is gone now.
4173 //===----------------------------------------------------------------------===//
4174 // X86 Optimization Hooks
4175 //===----------------------------------------------------------------------===//
4177 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
4179 uint64_t &KnownZero,
4181 unsigned Depth) const {
4182 unsigned Opc = Op.getOpcode();
4183 assert((Opc >= ISD::BUILTIN_OP_END ||
4184 Opc == ISD::INTRINSIC_WO_CHAIN ||
4185 Opc == ISD::INTRINSIC_W_CHAIN ||
4186 Opc == ISD::INTRINSIC_VOID) &&
4187 "Should use MaskedValueIsZero if you don't know whether Op"
4188 " is a target node!");
4190 KnownZero = KnownOne = 0; // Don't know anything.
4194 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
4199 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4200 /// element of the result of the vector shuffle.
4201 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
4202 MVT::ValueType VT = N->getValueType(0);
4203 SDOperand PermMask = N->getOperand(2);
4204 unsigned NumElems = PermMask.getNumOperands();
4205 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
4207 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4209 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4210 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
4211 SDOperand Idx = PermMask.getOperand(i);
4212 if (Idx.getOpcode() == ISD::UNDEF)
4213 return DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4214 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
4219 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
4220 /// node is a GlobalAddress + an offset.
4221 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
4222 unsigned Opc = N->getOpcode();
4223 if (Opc == X86ISD::Wrapper) {
4224 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
4225 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
4228 } else if (Opc == ISD::ADD) {
4229 SDOperand N1 = N->getOperand(0);
4230 SDOperand N2 = N->getOperand(1);
4231 if (isGAPlusOffset(N1.Val, GA, Offset)) {
4232 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
4234 Offset += V->getSignExtended();
4237 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
4238 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
4240 Offset += V->getSignExtended();
4248 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
4250 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
4251 MachineFrameInfo *MFI) {
4252 if (N->getOperand(0).Val != Base->getOperand(0).Val)
4255 SDOperand Loc = N->getOperand(1);
4256 SDOperand BaseLoc = Base->getOperand(1);
4257 if (Loc.getOpcode() == ISD::FrameIndex) {
4258 if (BaseLoc.getOpcode() != ISD::FrameIndex)
4260 int FI = dyn_cast<FrameIndexSDNode>(Loc)->getIndex();
4261 int BFI = dyn_cast<FrameIndexSDNode>(BaseLoc)->getIndex();
4262 int FS = MFI->getObjectSize(FI);
4263 int BFS = MFI->getObjectSize(BFI);
4264 if (FS != BFS || FS != Size) return false;
4265 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
4267 GlobalValue *GV1 = NULL;
4268 GlobalValue *GV2 = NULL;
4269 int64_t Offset1 = 0;
4270 int64_t Offset2 = 0;
4271 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
4272 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
4273 if (isGA1 && isGA2 && GV1 == GV2)
4274 return Offset1 == (Offset2 + Dist*Size);
4280 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
4281 const X86Subtarget *Subtarget) {
4284 if (isGAPlusOffset(Base, GV, Offset))
4285 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
4287 assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
4288 int BFI = dyn_cast<FrameIndexSDNode>(Base)->getIndex();
4290 // Fixed objects do not specify alignment, however the offsets are known.
4291 return ((Subtarget->getStackAlignment() % 16) == 0 &&
4292 (MFI->getObjectOffset(BFI) % 16) == 0);
4294 return MFI->getObjectAlignment(BFI) >= 16;
4300 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
4301 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
4302 /// if the load addresses are consecutive, non-overlapping, and in the right
4304 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
4305 const X86Subtarget *Subtarget) {
4306 MachineFunction &MF = DAG.getMachineFunction();
4307 MachineFrameInfo *MFI = MF.getFrameInfo();
4308 MVT::ValueType VT = N->getValueType(0);
4309 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
4310 SDOperand PermMask = N->getOperand(2);
4311 int NumElems = (int)PermMask.getNumOperands();
4312 SDNode *Base = NULL;
4313 for (int i = 0; i < NumElems; ++i) {
4314 SDOperand Idx = PermMask.getOperand(i);
4315 if (Idx.getOpcode() == ISD::UNDEF) {
4316 if (!Base) return SDOperand();
4319 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
4320 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
4324 else if (!isConsecutiveLoad(Arg.Val, Base,
4325 i, MVT::getSizeInBits(EVT)/8,MFI))
4330 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
4332 LoadSDNode *LD = cast<LoadSDNode>(Base);
4333 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
4334 LD->getSrcValueOffset());
4336 // Just use movups, it's shorter.
4337 SDVTList Tys = DAG.getVTList(MVT::v4f32, MVT::Other);
4338 SmallVector<SDOperand, 3> Ops;
4339 Ops.push_back(Base->getOperand(0));
4340 Ops.push_back(Base->getOperand(1));
4341 Ops.push_back(Base->getOperand(2));
4342 return DAG.getNode(ISD::BIT_CONVERT, VT,
4343 DAG.getNode(X86ISD::LOAD_UA, Tys, &Ops[0], Ops.size()));
4347 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
4348 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
4349 const X86Subtarget *Subtarget) {
4350 SDOperand Cond = N->getOperand(0);
4352 // If we have SSE[12] support, try to form min/max nodes.
4353 if (Subtarget->hasSSE2() &&
4354 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
4355 if (Cond.getOpcode() == ISD::SETCC) {
4356 // Get the LHS/RHS of the select.
4357 SDOperand LHS = N->getOperand(1);
4358 SDOperand RHS = N->getOperand(2);
4359 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
4361 unsigned Opcode = 0;
4362 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
4365 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
4368 if (!UnsafeFPMath) break;
4370 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
4372 Opcode = X86ISD::FMIN;
4375 case ISD::SETOGT: // (X > Y) ? X : Y -> max
4378 if (!UnsafeFPMath) break;
4380 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
4382 Opcode = X86ISD::FMAX;
4385 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
4388 case ISD::SETOGT: // (X > Y) ? Y : X -> min
4391 if (!UnsafeFPMath) break;
4393 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
4395 Opcode = X86ISD::FMIN;
4398 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
4401 if (!UnsafeFPMath) break;
4403 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
4405 Opcode = X86ISD::FMAX;
4411 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
4420 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
4421 DAGCombinerInfo &DCI) const {
4422 SelectionDAG &DAG = DCI.DAG;
4423 switch (N->getOpcode()) {
4425 case ISD::VECTOR_SHUFFLE:
4426 return PerformShuffleCombine(N, DAG, Subtarget);
4428 return PerformSELECTCombine(N, DAG, Subtarget);
4434 //===----------------------------------------------------------------------===//
4435 // X86 Inline Assembly Support
4436 //===----------------------------------------------------------------------===//
4438 /// getConstraintType - Given a constraint letter, return the type of
4439 /// constraint it is for this target.
4440 X86TargetLowering::ConstraintType
4441 X86TargetLowering::getConstraintType(char ConstraintLetter) const {
4442 switch (ConstraintLetter) {
4451 return C_RegisterClass;
4452 default: return TargetLowering::getConstraintType(ConstraintLetter);
4456 /// isOperandValidForConstraint - Return the specified operand (possibly
4457 /// modified) if the specified SDOperand is valid for the specified target
4458 /// constraint letter, otherwise return null.
4459 SDOperand X86TargetLowering::
4460 isOperandValidForConstraint(SDOperand Op, char Constraint, SelectionDAG &DAG) {
4461 switch (Constraint) {
4464 // Literal immediates are always ok.
4465 if (isa<ConstantSDNode>(Op)) return Op;
4467 // If we are in non-pic codegen mode, we allow the address of a global to
4468 // be used with 'i'.
4469 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op)) {
4470 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
4471 return SDOperand(0, 0);
4473 if (GA->getOpcode() != ISD::TargetGlobalAddress)
4474 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
4479 // Otherwise, not valid for this mode.
4480 return SDOperand(0, 0);
4482 return TargetLowering::isOperandValidForConstraint(Op, Constraint, DAG);
4486 std::vector<unsigned> X86TargetLowering::
4487 getRegClassForInlineAsmConstraint(const std::string &Constraint,
4488 MVT::ValueType VT) const {
4489 if (Constraint.size() == 1) {
4490 // FIXME: not handling fp-stack yet!
4491 // FIXME: not handling MMX registers yet ('y' constraint).
4492 switch (Constraint[0]) { // GCC X86 Constraint Letters
4493 default: break; // Unknown constraint letter
4494 case 'A': // EAX/EDX
4495 if (VT == MVT::i32 || VT == MVT::i64)
4496 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
4498 case 'r': // GENERAL_REGS
4499 case 'R': // LEGACY_REGS
4500 if (VT == MVT::i64 && Subtarget->is64Bit())
4501 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
4502 X86::RSI, X86::RDI, X86::RBP, X86::RSP,
4503 X86::R8, X86::R9, X86::R10, X86::R11,
4504 X86::R12, X86::R13, X86::R14, X86::R15, 0);
4506 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
4507 X86::ESI, X86::EDI, X86::EBP, X86::ESP, 0);
4508 else if (VT == MVT::i16)
4509 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
4510 X86::SI, X86::DI, X86::BP, X86::SP, 0);
4511 else if (VT == MVT::i8)
4512 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
4514 case 'l': // INDEX_REGS
4516 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
4517 X86::ESI, X86::EDI, X86::EBP, 0);
4518 else if (VT == MVT::i16)
4519 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
4520 X86::SI, X86::DI, X86::BP, 0);
4521 else if (VT == MVT::i8)
4522 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
4524 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
4527 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
4528 else if (VT == MVT::i16)
4529 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
4530 else if (VT == MVT::i8)
4531 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
4533 case 'x': // SSE_REGS if SSE1 allowed
4534 if (Subtarget->hasSSE1())
4535 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
4536 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
4538 return std::vector<unsigned>();
4539 case 'Y': // SSE_REGS if SSE2 allowed
4540 if (Subtarget->hasSSE2())
4541 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
4542 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
4544 return std::vector<unsigned>();
4548 return std::vector<unsigned>();
4551 std::pair<unsigned, const TargetRegisterClass*>
4552 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4553 MVT::ValueType VT) const {
4554 // Use the default implementation in TargetLowering to convert the register
4555 // constraint into a member of a register class.
4556 std::pair<unsigned, const TargetRegisterClass*> Res;
4557 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4559 // Not found as a standard register?
4560 if (Res.second == 0) {
4561 // GCC calls "st(0)" just plain "st".
4562 if (StringsEqualNoCase("{st}", Constraint)) {
4563 Res.first = X86::ST0;
4564 Res.second = X86::RSTRegisterClass;
4570 // Otherwise, check to see if this is a register class of the wrong value
4571 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
4572 // turn into {ax},{dx}.
4573 if (Res.second->hasType(VT))
4574 return Res; // Correct type already, nothing to do.
4576 // All of the single-register GCC register classes map their values onto
4577 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
4578 // really want an 8-bit or 32-bit register, map to the appropriate register
4579 // class and return the appropriate register.
4580 if (Res.second != X86::GR16RegisterClass)
4583 if (VT == MVT::i8) {
4584 unsigned DestReg = 0;
4585 switch (Res.first) {
4587 case X86::AX: DestReg = X86::AL; break;
4588 case X86::DX: DestReg = X86::DL; break;
4589 case X86::CX: DestReg = X86::CL; break;
4590 case X86::BX: DestReg = X86::BL; break;
4593 Res.first = DestReg;
4594 Res.second = Res.second = X86::GR8RegisterClass;
4596 } else if (VT == MVT::i32) {
4597 unsigned DestReg = 0;
4598 switch (Res.first) {
4600 case X86::AX: DestReg = X86::EAX; break;
4601 case X86::DX: DestReg = X86::EDX; break;
4602 case X86::CX: DestReg = X86::ECX; break;
4603 case X86::BX: DestReg = X86::EBX; break;
4604 case X86::SI: DestReg = X86::ESI; break;
4605 case X86::DI: DestReg = X86::EDI; break;
4606 case X86::BP: DestReg = X86::EBP; break;
4607 case X86::SP: DestReg = X86::ESP; break;
4610 Res.first = DestReg;
4611 Res.second = Res.second = X86::GR32RegisterClass;
4613 } else if (VT == MVT::i64) {
4614 unsigned DestReg = 0;
4615 switch (Res.first) {
4617 case X86::AX: DestReg = X86::RAX; break;
4618 case X86::DX: DestReg = X86::RDX; break;
4619 case X86::CX: DestReg = X86::RCX; break;
4620 case X86::BX: DestReg = X86::RBX; break;
4621 case X86::SI: DestReg = X86::RSI; break;
4622 case X86::DI: DestReg = X86::RDI; break;
4623 case X86::BP: DestReg = X86::RBP; break;
4624 case X86::SP: DestReg = X86::RSP; break;
4627 Res.first = DestReg;
4628 Res.second = Res.second = X86::GR64RegisterClass;