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 /// HowToPassArgument - Returns how an formal argument of the specified type
605 /// should be passed. If it is through stack, returns the size of the stack
606 /// slot; if it is through integer or XMM register, returns the number of
607 /// integer or XMM registers are needed.
609 HowToPassCallArgument(MVT::ValueType ObjectVT,
611 unsigned NumIntRegs, unsigned NumXMMRegs,
612 unsigned MaxNumIntRegs,
613 unsigned &ObjSize, unsigned &ObjIntRegs,
614 unsigned &ObjXMMRegs) {
619 if (MaxNumIntRegs>3) {
620 // We don't have too much registers on ia32! :)
625 default: assert(0 && "Unhandled argument type!");
627 if (ArgInReg && (NumIntRegs < MaxNumIntRegs))
633 if (ArgInReg && (NumIntRegs < MaxNumIntRegs))
639 if (ArgInReg && (NumIntRegs < MaxNumIntRegs))
664 SDOperand X86TargetLowering::LowerCCCArguments(SDOperand Op, SelectionDAG &DAG,
666 unsigned NumArgs = Op.Val->getNumValues() - 1;
667 MachineFunction &MF = DAG.getMachineFunction();
668 MachineFrameInfo *MFI = MF.getFrameInfo();
669 SDOperand Root = Op.getOperand(0);
670 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
672 SmallVector<CCValAssign, 16> ArgLocs;
673 CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(),
676 for (unsigned i = 0; i != NumArgs; ++i) {
677 MVT::ValueType ArgVT = Op.getValue(i).getValueType();
678 unsigned ArgFlags = cast<ConstantSDNode>(Op.getOperand(3+i))->getValue();
679 if (CC_X86_32_C(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo))
680 assert(0 && "Unhandled argument type!");
683 SmallVector<SDOperand, 8> ArgValues;
684 unsigned LastVal = ~0U;
685 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
686 CCValAssign &VA = ArgLocs[i];
687 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
689 assert(VA.getValNo() != LastVal &&
690 "Don't support value assigned to multiple locs yet");
691 LastVal = VA.getValNo();
694 MVT::ValueType RegVT = VA.getLocVT();
695 TargetRegisterClass *RC;
696 if (RegVT == MVT::i32)
697 RC = X86::GR32RegisterClass;
699 assert(MVT::isVector(RegVT));
700 RC = X86::VR128RegisterClass;
703 SDOperand ArgValue = DAG.getCopyFromReg(Root, VA.getLocReg(), RegVT);
704 AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
706 // If this is an 8 or 16-bit value, it is really passed promoted to 32
707 // bits. Insert an assert[sz]ext to capture this, then truncate to the
709 if (VA.getLocInfo() == CCValAssign::SExt)
710 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
711 DAG.getValueType(VA.getValVT()));
712 else if (VA.getLocInfo() == CCValAssign::ZExt)
713 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
714 DAG.getValueType(VA.getValVT()));
716 if (VA.getLocInfo() != CCValAssign::Full)
717 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
719 ArgValues.push_back(ArgValue);
721 assert(VA.isMemLoc());
723 // Create the nodes corresponding to a load from this parameter slot.
724 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
725 VA.getLocMemOffset());
726 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
727 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
731 unsigned StackSize = CCInfo.getNextStackOffset();
733 ArgValues.push_back(Root);
735 // If the function takes variable number of arguments, make a frame index for
736 // the start of the first vararg value... for expansion of llvm.va_start.
738 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
740 if (isStdCall && !isVarArg) {
741 BytesToPopOnReturn = StackSize; // Callee pops everything..
742 BytesCallerReserves = 0;
744 BytesToPopOnReturn = 0; // Callee pops hidden struct pointer.
746 // If this is an sret function, the return should pop the hidden pointer.
747 if (NumArgs && (cast<ConstantSDNode>(Op.getOperand(3))->getValue() & 4))
748 BytesToPopOnReturn = 4;
750 BytesCallerReserves = StackSize;
753 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
754 ReturnAddrIndex = 0; // No return address slot generated yet.
756 MF.getInfo<X86FunctionInfo>()->setBytesToPopOnReturn(BytesToPopOnReturn);
758 // Return the new list of results.
759 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
760 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
763 SDOperand X86TargetLowering::LowerCCCCallTo(SDOperand Op, SelectionDAG &DAG,
765 SDOperand Chain = Op.getOperand(0);
766 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
767 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
768 SDOperand Callee = Op.getOperand(4);
769 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
771 SmallVector<CCValAssign, 16> ArgLocs;
772 CCState CCInfo(CC, getTargetMachine(), ArgLocs);
774 for (unsigned i = 0; i != NumOps; ++i) {
775 MVT::ValueType ArgVT = Op.getOperand(5+2*i).getValueType();
776 unsigned ArgFlags =cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
777 if (CC_X86_32_C(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo))
778 assert(0 && "Unhandled argument type!");
781 // Get a count of how many bytes are to be pushed on the stack.
782 unsigned NumBytes = CCInfo.getNextStackOffset();
784 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
786 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
787 SmallVector<SDOperand, 8> MemOpChains;
791 // Walk the register/memloc assignments, inserting copies/loads.
792 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
793 CCValAssign &VA = ArgLocs[i];
794 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
796 // Promote the value if needed.
797 switch (VA.getLocInfo()) {
798 default: assert(0 && "Unknown loc info!");
799 case CCValAssign::Full: break;
800 case CCValAssign::SExt:
801 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
803 case CCValAssign::ZExt:
804 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
806 case CCValAssign::AExt:
807 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
812 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
814 assert(VA.isMemLoc());
815 if (StackPtr.Val == 0)
816 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
817 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
818 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
819 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
823 // If the first argument is an sret pointer, remember it.
824 bool isSRet = NumOps &&(cast<ConstantSDNode>(Op.getOperand(6))->getValue()&4);
826 if (!MemOpChains.empty())
827 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
828 &MemOpChains[0], MemOpChains.size());
830 // Build a sequence of copy-to-reg nodes chained together with token chain
831 // and flag operands which copy the outgoing args into registers.
833 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
834 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
836 InFlag = Chain.getValue(1);
839 // ELF / PIC requires GOT in the EBX register before function calls via PLT
841 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
842 Subtarget->isPICStyleGOT()) {
843 Chain = DAG.getCopyToReg(Chain, X86::EBX,
844 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
846 InFlag = Chain.getValue(1);
849 // If the callee is a GlobalAddress node (quite common, every direct call is)
850 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
851 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
852 // We should use extra load for direct calls to dllimported functions in
854 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
855 getTargetMachine(), true))
856 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
857 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
858 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
860 // Returns a chain & a flag for retval copy to use.
861 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
862 SmallVector<SDOperand, 8> Ops;
863 Ops.push_back(Chain);
864 Ops.push_back(Callee);
866 // Add argument registers to the end of the list so that they are known live
868 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
869 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
870 RegsToPass[i].second.getValueType()));
872 // Add an implicit use GOT pointer in EBX.
873 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
874 Subtarget->isPICStyleGOT())
875 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
878 Ops.push_back(InFlag);
880 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
881 NodeTys, &Ops[0], Ops.size());
882 InFlag = Chain.getValue(1);
884 // Create the CALLSEQ_END node.
885 unsigned NumBytesForCalleeToPush = 0;
887 if (CC == CallingConv::X86_StdCall) {
889 NumBytesForCalleeToPush = isSRet ? 4 : 0;
891 NumBytesForCalleeToPush = NumBytes;
893 // If this is is a call to a struct-return function, the callee
894 // pops the hidden struct pointer, so we have to push it back.
895 // This is common for Darwin/X86, Linux & Mingw32 targets.
896 NumBytesForCalleeToPush = isSRet ? 4 : 0;
899 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
901 Ops.push_back(Chain);
902 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
903 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
904 Ops.push_back(InFlag);
905 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
906 InFlag = Chain.getValue(1);
908 // Handle result values, copying them out of physregs into vregs that we
910 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
914 //===----------------------------------------------------------------------===//
915 // FastCall Calling Convention implementation
916 //===----------------------------------------------------------------------===//
918 // The X86 'fastcall' calling convention passes up to two integer arguments in
919 // registers (an appropriate portion of ECX/EDX), passes arguments in C order,
920 // and requires that the callee pop its arguments off the stack (allowing proper
921 // tail calls), and has the same return value conventions as C calling convs.
923 // This calling convention always arranges for the callee pop value to be 8n+4
924 // bytes, which is needed for tail recursion elimination and stack alignment
927 X86TargetLowering::LowerFastCCArguments(SDOperand Op, SelectionDAG &DAG) {
928 unsigned NumArgs = Op.Val->getNumValues()-1;
929 MachineFunction &MF = DAG.getMachineFunction();
930 MachineFrameInfo *MFI = MF.getFrameInfo();
931 SDOperand Root = Op.getOperand(0);
932 SmallVector<SDOperand, 8> ArgValues;
934 // Add DAG nodes to load the arguments... On entry to a function the stack
935 // frame looks like this:
937 // [ESP] -- return address
938 // [ESP + 4] -- first nonreg argument (leftmost lexically)
939 // [ESP + 8] -- second nonreg argument, if 1st argument is <= 4 bytes in size
941 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
943 // Keep track of the number of integer regs passed so far. This can be either
944 // 0 (neither EAX/ECX or EDX used), 1 (EAX/ECX is used) or 2 (EAX/ECX and EDX
946 unsigned NumIntRegs = 0;
947 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
949 static const unsigned XMMArgRegs[] = {
950 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
953 static const unsigned GPRArgRegs[][2] = {
954 { X86::CL, X86::DL },
955 { X86::CX, X86::DX },
956 { X86::ECX, X86::EDX }
959 static const TargetRegisterClass* GPRClasses[3] = {
960 X86::GR8RegisterClass, X86::GR16RegisterClass, X86::GR32RegisterClass
963 for (unsigned i = 0; i < NumArgs; ++i) {
964 MVT::ValueType ObjectVT = Op.getValue(i).getValueType();
965 unsigned ArgIncrement = 4;
966 unsigned ObjSize = 0;
967 unsigned ObjXMMRegs = 0;
968 unsigned ObjIntRegs = 0;
972 HowToPassCallArgument(ObjectVT,
973 true, // Use as much registers as possible
974 NumIntRegs, NumXMMRegs, 2,
975 ObjSize, ObjIntRegs, ObjXMMRegs);
978 ArgIncrement = ObjSize;
980 if (ObjIntRegs || ObjXMMRegs) {
982 default: assert(0 && "Unhandled argument type!");
986 unsigned RegToUse = GPRArgRegs[ObjectVT-MVT::i8][NumIntRegs];
987 Reg = AddLiveIn(MF, RegToUse, GPRClasses[ObjectVT-MVT::i8]);
988 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
997 Reg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs], X86::VR128RegisterClass);
998 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
1002 NumIntRegs += ObjIntRegs;
1003 NumXMMRegs += ObjXMMRegs;
1006 // XMM arguments have to be aligned on 16-byte boundary.
1008 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1009 // Create the SelectionDAG nodes corresponding to a load from this
1011 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
1012 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1013 ArgValue = DAG.getLoad(Op.Val->getValueType(i), Root, FIN, NULL, 0);
1015 ArgOffset += ArgIncrement; // Move on to the next argument.
1018 ArgValues.push_back(ArgValue);
1021 ArgValues.push_back(Root);
1023 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1024 // arguments and the arguments after the retaddr has been pushed are aligned.
1025 if ((ArgOffset & 7) == 0)
1028 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1029 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only.
1030 ReturnAddrIndex = 0; // No return address slot generated yet.
1031 BytesToPopOnReturn = ArgOffset; // Callee pops all stack arguments.
1032 BytesCallerReserves = 0;
1034 MF.getInfo<X86FunctionInfo>()->setBytesToPopOnReturn(BytesToPopOnReturn);
1036 // Finally, inform the code generator which regs we return values in.
1037 switch (getValueType(MF.getFunction()->getReturnType())) {
1038 default: assert(0 && "Unknown type!");
1039 case MVT::isVoid: break;
1044 MF.addLiveOut(X86::EAX);
1047 MF.addLiveOut(X86::EAX);
1048 MF.addLiveOut(X86::EDX);
1052 MF.addLiveOut(X86::ST0);
1060 MF.addLiveOut(X86::XMM0);
1064 // Return the new list of results.
1065 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1066 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1069 SDOperand X86TargetLowering::LowerFastCCCallTo(SDOperand Op, SelectionDAG &DAG,
1071 SDOperand Chain = Op.getOperand(0);
1072 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1073 SDOperand Callee = Op.getOperand(4);
1074 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1076 // Count how many bytes are to be pushed on the stack.
1077 unsigned NumBytes = 0;
1079 // Keep track of the number of integer regs passed so far. This can be either
1080 // 0 (neither EAX/ECX or EDX used), 1 (EAX/ECX is used) or 2 (EAX/ECX and EDX
1082 unsigned NumIntRegs = 0;
1083 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1085 static const unsigned GPRArgRegs[][2] = {
1086 { X86::CL, X86::DL },
1087 { X86::CX, X86::DX },
1088 { X86::ECX, X86::EDX }
1090 static const unsigned XMMArgRegs[] = {
1091 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1094 for (unsigned i = 0; i != NumOps; ++i) {
1095 SDOperand Arg = Op.getOperand(5+2*i);
1097 switch (Arg.getValueType()) {
1098 default: assert(0 && "Unknown value type!");
1102 if (NumIntRegs < 2) {
1121 // XMM arguments have to be aligned on 16-byte boundary.
1122 NumBytes = ((NumBytes + 15) / 16) * 16;
1129 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1130 // arguments and the arguments after the retaddr has been pushed are aligned.
1131 if ((NumBytes & 7) == 0)
1134 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1136 // Arguments go on the stack in reverse order, as specified by the ABI.
1137 unsigned ArgOffset = 0;
1139 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1140 SmallVector<SDOperand, 8> MemOpChains;
1141 SDOperand StackPtr = DAG.getRegister(X86StackPtr, getPointerTy());
1142 for (unsigned i = 0; i != NumOps; ++i) {
1143 SDOperand Arg = Op.getOperand(5+2*i);
1145 switch (Arg.getValueType()) {
1146 default: assert(0 && "Unexpected ValueType for argument!");
1150 if (NumIntRegs < 2) {
1152 GPRArgRegs[Arg.getValueType()-MVT::i8][NumIntRegs];
1153 RegsToPass.push_back(std::make_pair(RegToUse, Arg));
1158 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1159 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1160 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1165 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1166 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1167 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1177 if (NumXMMRegs < 4) {
1178 RegsToPass.push_back(std::make_pair(XMMArgRegs[NumXMMRegs], Arg));
1181 // XMM arguments have to be aligned on 16-byte boundary.
1182 ArgOffset = ((ArgOffset + 15) / 16) * 16;
1183 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1184 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1185 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1192 if (!MemOpChains.empty())
1193 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1194 &MemOpChains[0], MemOpChains.size());
1196 // Build a sequence of copy-to-reg nodes chained together with token chain
1197 // and flag operands which copy the outgoing args into registers.
1199 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1200 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1202 InFlag = Chain.getValue(1);
1205 // If the callee is a GlobalAddress node (quite common, every direct call is)
1206 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1207 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1208 // We should use extra load for direct calls to dllimported functions in
1210 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1211 getTargetMachine(), true))
1212 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1213 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1214 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1216 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1218 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1219 Subtarget->isPICStyleGOT()) {
1220 Chain = DAG.getCopyToReg(Chain, X86::EBX,
1221 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
1223 InFlag = Chain.getValue(1);
1226 // Returns a chain & a flag for retval copy to use.
1227 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1228 SmallVector<SDOperand, 8> Ops;
1229 Ops.push_back(Chain);
1230 Ops.push_back(Callee);
1232 // Add argument registers to the end of the list so that they are known live
1234 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1235 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1236 RegsToPass[i].second.getValueType()));
1238 // Add an implicit use GOT pointer in EBX.
1239 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1240 Subtarget->isPICStyleGOT())
1241 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1244 Ops.push_back(InFlag);
1246 // FIXME: Do not generate X86ISD::TAILCALL for now.
1247 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1248 NodeTys, &Ops[0], Ops.size());
1249 InFlag = Chain.getValue(1);
1251 // Returns a flag for retval copy to use.
1252 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1254 Ops.push_back(Chain);
1255 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1256 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1257 Ops.push_back(InFlag);
1258 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1259 InFlag = Chain.getValue(1);
1261 // Handle result values, copying them out of physregs into vregs that we
1263 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1267 //===----------------------------------------------------------------------===//
1268 // X86-64 C Calling Convention implementation
1269 //===----------------------------------------------------------------------===//
1272 X86TargetLowering::LowerX86_64CCCArguments(SDOperand Op, SelectionDAG &DAG) {
1273 unsigned NumArgs = Op.Val->getNumValues() - 1;
1274 MachineFunction &MF = DAG.getMachineFunction();
1275 MachineFrameInfo *MFI = MF.getFrameInfo();
1276 SDOperand Root = Op.getOperand(0);
1277 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1279 static const unsigned GPR64ArgRegs[] = {
1280 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1282 static const unsigned XMMArgRegs[] = {
1283 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1284 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1287 SmallVector<CCValAssign, 16> ArgLocs;
1288 CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(),
1291 for (unsigned i = 0; i != NumArgs; ++i) {
1292 MVT::ValueType ArgVT = Op.getValue(i).getValueType();
1293 unsigned ArgFlags = cast<ConstantSDNode>(Op.getOperand(3+i))->getValue();
1294 if (CC_X86_64_C(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo))
1295 assert(0 && "Unhandled argument type!");
1298 SmallVector<SDOperand, 8> ArgValues;
1299 unsigned LastVal = ~0U;
1300 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1301 CCValAssign &VA = ArgLocs[i];
1302 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1304 assert(VA.getValNo() != LastVal &&
1305 "Don't support value assigned to multiple locs yet");
1306 LastVal = VA.getValNo();
1308 if (VA.isRegLoc()) {
1309 MVT::ValueType RegVT = VA.getLocVT();
1310 TargetRegisterClass *RC;
1311 if (RegVT == MVT::i32)
1312 RC = X86::GR32RegisterClass;
1313 else if (RegVT == MVT::i64)
1314 RC = X86::GR64RegisterClass;
1315 else if (RegVT == MVT::f32)
1316 RC = X86::FR32RegisterClass;
1317 else if (RegVT == MVT::f64)
1318 RC = X86::FR64RegisterClass;
1320 assert(MVT::isVector(RegVT));
1321 RC = X86::VR128RegisterClass;
1324 SDOperand ArgValue = DAG.getCopyFromReg(Root, VA.getLocReg(), RegVT);
1325 AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC);
1327 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1328 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1330 if (VA.getLocInfo() == CCValAssign::SExt)
1331 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue,
1332 DAG.getValueType(VA.getValVT()));
1333 else if (VA.getLocInfo() == CCValAssign::ZExt)
1334 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue,
1335 DAG.getValueType(VA.getValVT()));
1337 if (VA.getLocInfo() != CCValAssign::Full)
1338 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue);
1340 ArgValues.push_back(ArgValue);
1342 assert(VA.isMemLoc());
1344 // Create the nodes corresponding to a load from this parameter slot.
1345 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8,
1346 VA.getLocMemOffset());
1347 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
1348 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0));
1352 unsigned StackSize = CCInfo.getNextStackOffset();
1354 // If the function takes variable number of arguments, make a frame index for
1355 // the start of the first vararg value... for expansion of llvm.va_start.
1357 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs, 6);
1358 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1360 // For X86-64, if there are vararg parameters that are passed via
1361 // registers, then we must store them to their spots on the stack so they
1362 // may be loaded by deferencing the result of va_next.
1363 VarArgsGPOffset = NumIntRegs * 8;
1364 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16;
1365 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1366 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16);
1368 // Store the integer parameter registers.
1369 SmallVector<SDOperand, 8> MemOps;
1370 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1371 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1372 DAG.getConstant(VarArgsGPOffset, getPointerTy()));
1373 for (; NumIntRegs != 6; ++NumIntRegs) {
1374 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs],
1375 X86::GR64RegisterClass);
1376 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1377 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1378 MemOps.push_back(Store);
1379 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1380 DAG.getConstant(8, getPointerTy()));
1383 // Now store the XMM (fp + vector) parameter registers.
1384 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN,
1385 DAG.getConstant(VarArgsFPOffset, getPointerTy()));
1386 for (; NumXMMRegs != 8; ++NumXMMRegs) {
1387 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
1388 X86::VR128RegisterClass);
1389 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32);
1390 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1391 MemOps.push_back(Store);
1392 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
1393 DAG.getConstant(16, getPointerTy()));
1395 if (!MemOps.empty())
1396 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
1397 &MemOps[0], MemOps.size());
1400 ArgValues.push_back(Root);
1402 ReturnAddrIndex = 0; // No return address slot generated yet.
1403 BytesToPopOnReturn = 0; // Callee pops nothing.
1404 BytesCallerReserves = StackSize;
1406 // Return the new list of results.
1407 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
1408 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo);
1412 X86TargetLowering::LowerX86_64CCCCallTo(SDOperand Op, SelectionDAG &DAG,
1414 SDOperand Chain = Op.getOperand(0);
1415 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1416 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1417 SDOperand Callee = Op.getOperand(4);
1418 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1420 SmallVector<CCValAssign, 16> ArgLocs;
1421 CCState CCInfo(CC, getTargetMachine(), ArgLocs);
1423 for (unsigned i = 0; i != NumOps; ++i) {
1424 MVT::ValueType ArgVT = Op.getOperand(5+2*i).getValueType();
1425 unsigned ArgFlags =cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
1426 if (CC_X86_64_C(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo))
1427 assert(0 && "Unhandled argument type!");
1430 // Get a count of how many bytes are to be pushed on the stack.
1431 unsigned NumBytes = CCInfo.getNextStackOffset();
1432 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1434 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass;
1435 SmallVector<SDOperand, 8> MemOpChains;
1439 // Walk the register/memloc assignments, inserting copies/loads.
1440 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1441 CCValAssign &VA = ArgLocs[i];
1442 SDOperand Arg = Op.getOperand(5+2*VA.getValNo());
1444 // Promote the value if needed.
1445 switch (VA.getLocInfo()) {
1446 default: assert(0 && "Unknown loc info!");
1447 case CCValAssign::Full: break;
1448 case CCValAssign::SExt:
1449 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg);
1451 case CCValAssign::ZExt:
1452 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg);
1454 case CCValAssign::AExt:
1455 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg);
1459 if (VA.isRegLoc()) {
1460 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1462 assert(VA.isMemLoc());
1463 if (StackPtr.Val == 0)
1464 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy());
1465 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy());
1466 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1467 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1471 if (!MemOpChains.empty())
1472 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1473 &MemOpChains[0], MemOpChains.size());
1475 // Build a sequence of copy-to-reg nodes chained together with token chain
1476 // and flag operands which copy the outgoing args into registers.
1478 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1479 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1481 InFlag = Chain.getValue(1);
1485 // From AMD64 ABI document:
1486 // For calls that may call functions that use varargs or stdargs
1487 // (prototype-less calls or calls to functions containing ellipsis (...) in
1488 // the declaration) %al is used as hidden argument to specify the number
1489 // of SSE registers used. The contents of %al do not need to match exactly
1490 // the number of registers, but must be an ubound on the number of SSE
1491 // registers used and is in the range 0 - 8 inclusive.
1493 // Count the number of XMM registers allocated.
1494 static const unsigned XMMArgRegs[] = {
1495 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1496 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1498 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1500 Chain = DAG.getCopyToReg(Chain, X86::AL,
1501 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1502 InFlag = Chain.getValue(1);
1505 // If the callee is a GlobalAddress node (quite common, every direct call is)
1506 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1507 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1508 // We should use extra load for direct calls to dllimported functions in
1510 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(),
1511 getTargetMachine(), true))
1512 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1513 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1514 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1516 // Returns a chain & a flag for retval copy to use.
1517 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1518 SmallVector<SDOperand, 8> Ops;
1519 Ops.push_back(Chain);
1520 Ops.push_back(Callee);
1522 // Add argument registers to the end of the list so that they are known live
1524 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1525 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1526 RegsToPass[i].second.getValueType()));
1529 Ops.push_back(InFlag);
1531 // FIXME: Do not generate X86ISD::TAILCALL for now.
1532 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1533 NodeTys, &Ops[0], Ops.size());
1534 InFlag = Chain.getValue(1);
1536 // Returns a flag for retval copy to use.
1537 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1539 Ops.push_back(Chain);
1540 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1541 Ops.push_back(DAG.getConstant(0, getPointerTy()));
1542 Ops.push_back(InFlag);
1543 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size());
1544 InFlag = Chain.getValue(1);
1546 // Handle result values, copying them out of physregs into vregs that we
1548 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo);
1552 //===----------------------------------------------------------------------===//
1553 // Other Lowering Hooks
1554 //===----------------------------------------------------------------------===//
1557 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1558 if (ReturnAddrIndex == 0) {
1559 // Set up a frame object for the return address.
1560 MachineFunction &MF = DAG.getMachineFunction();
1561 if (Subtarget->is64Bit())
1562 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8);
1564 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
1567 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
1572 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1573 /// specific condition code. It returns a false if it cannot do a direct
1574 /// translation. X86CC is the translated CondCode. LHS/RHS are modified as
1576 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1577 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS,
1578 SelectionDAG &DAG) {
1579 X86CC = X86::COND_INVALID;
1581 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
1582 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
1583 // X > -1 -> X == 0, jump !sign.
1584 RHS = DAG.getConstant(0, RHS.getValueType());
1585 X86CC = X86::COND_NS;
1587 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
1588 // X < 0 -> X == 0, jump on sign.
1589 X86CC = X86::COND_S;
1594 switch (SetCCOpcode) {
1596 case ISD::SETEQ: X86CC = X86::COND_E; break;
1597 case ISD::SETGT: X86CC = X86::COND_G; break;
1598 case ISD::SETGE: X86CC = X86::COND_GE; break;
1599 case ISD::SETLT: X86CC = X86::COND_L; break;
1600 case ISD::SETLE: X86CC = X86::COND_LE; break;
1601 case ISD::SETNE: X86CC = X86::COND_NE; break;
1602 case ISD::SETULT: X86CC = X86::COND_B; break;
1603 case ISD::SETUGT: X86CC = X86::COND_A; break;
1604 case ISD::SETULE: X86CC = X86::COND_BE; break;
1605 case ISD::SETUGE: X86CC = X86::COND_AE; break;
1608 // On a floating point condition, the flags are set as follows:
1610 // 0 | 0 | 0 | X > Y
1611 // 0 | 0 | 1 | X < Y
1612 // 1 | 0 | 0 | X == Y
1613 // 1 | 1 | 1 | unordered
1615 switch (SetCCOpcode) {
1618 case ISD::SETEQ: X86CC = X86::COND_E; break;
1619 case ISD::SETOLT: Flip = true; // Fallthrough
1621 case ISD::SETGT: X86CC = X86::COND_A; break;
1622 case ISD::SETOLE: Flip = true; // Fallthrough
1624 case ISD::SETGE: X86CC = X86::COND_AE; break;
1625 case ISD::SETUGT: Flip = true; // Fallthrough
1627 case ISD::SETLT: X86CC = X86::COND_B; break;
1628 case ISD::SETUGE: Flip = true; // Fallthrough
1630 case ISD::SETLE: X86CC = X86::COND_BE; break;
1632 case ISD::SETNE: X86CC = X86::COND_NE; break;
1633 case ISD::SETUO: X86CC = X86::COND_P; break;
1634 case ISD::SETO: X86CC = X86::COND_NP; break;
1637 std::swap(LHS, RHS);
1640 return X86CC != X86::COND_INVALID;
1643 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
1644 /// code. Current x86 isa includes the following FP cmov instructions:
1645 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
1646 static bool hasFPCMov(unsigned X86CC) {
1662 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
1663 /// true if Op is undef or if its value falls within the specified range (L, H].
1664 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
1665 if (Op.getOpcode() == ISD::UNDEF)
1668 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
1669 return (Val >= Low && Val < Hi);
1672 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
1673 /// true if Op is undef or if its value equal to the specified value.
1674 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
1675 if (Op.getOpcode() == ISD::UNDEF)
1677 return cast<ConstantSDNode>(Op)->getValue() == Val;
1680 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
1681 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
1682 bool X86::isPSHUFDMask(SDNode *N) {
1683 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1685 if (N->getNumOperands() != 4)
1688 // Check if the value doesn't reference the second vector.
1689 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1690 SDOperand Arg = N->getOperand(i);
1691 if (Arg.getOpcode() == ISD::UNDEF) continue;
1692 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1693 if (cast<ConstantSDNode>(Arg)->getValue() >= 4)
1700 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
1701 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
1702 bool X86::isPSHUFHWMask(SDNode *N) {
1703 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1705 if (N->getNumOperands() != 8)
1708 // Lower quadword copied in order.
1709 for (unsigned i = 0; i != 4; ++i) {
1710 SDOperand Arg = N->getOperand(i);
1711 if (Arg.getOpcode() == ISD::UNDEF) continue;
1712 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1713 if (cast<ConstantSDNode>(Arg)->getValue() != i)
1717 // Upper quadword shuffled.
1718 for (unsigned i = 4; i != 8; ++i) {
1719 SDOperand Arg = N->getOperand(i);
1720 if (Arg.getOpcode() == ISD::UNDEF) continue;
1721 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1722 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1723 if (Val < 4 || Val > 7)
1730 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
1731 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
1732 bool X86::isPSHUFLWMask(SDNode *N) {
1733 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1735 if (N->getNumOperands() != 8)
1738 // Upper quadword copied in order.
1739 for (unsigned i = 4; i != 8; ++i)
1740 if (!isUndefOrEqual(N->getOperand(i), i))
1743 // Lower quadword shuffled.
1744 for (unsigned i = 0; i != 4; ++i)
1745 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
1751 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
1752 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
1753 static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) {
1754 if (NumElems != 2 && NumElems != 4) return false;
1756 unsigned Half = NumElems / 2;
1757 for (unsigned i = 0; i < Half; ++i)
1758 if (!isUndefOrInRange(Elems[i], 0, NumElems))
1760 for (unsigned i = Half; i < NumElems; ++i)
1761 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2))
1767 bool X86::isSHUFPMask(SDNode *N) {
1768 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1769 return ::isSHUFPMask(N->op_begin(), N->getNumOperands());
1772 /// isCommutedSHUFP - Returns true if the shuffle mask is except
1773 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
1774 /// half elements to come from vector 1 (which would equal the dest.) and
1775 /// the upper half to come from vector 2.
1776 static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) {
1777 if (NumOps != 2 && NumOps != 4) return false;
1779 unsigned Half = NumOps / 2;
1780 for (unsigned i = 0; i < Half; ++i)
1781 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2))
1783 for (unsigned i = Half; i < NumOps; ++i)
1784 if (!isUndefOrInRange(Ops[i], 0, NumOps))
1789 static bool isCommutedSHUFP(SDNode *N) {
1790 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1791 return isCommutedSHUFP(N->op_begin(), N->getNumOperands());
1794 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
1795 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
1796 bool X86::isMOVHLPSMask(SDNode *N) {
1797 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1799 if (N->getNumOperands() != 4)
1802 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
1803 return isUndefOrEqual(N->getOperand(0), 6) &&
1804 isUndefOrEqual(N->getOperand(1), 7) &&
1805 isUndefOrEqual(N->getOperand(2), 2) &&
1806 isUndefOrEqual(N->getOperand(3), 3);
1809 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
1810 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
1812 bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) {
1813 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1815 if (N->getNumOperands() != 4)
1818 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3
1819 return isUndefOrEqual(N->getOperand(0), 2) &&
1820 isUndefOrEqual(N->getOperand(1), 3) &&
1821 isUndefOrEqual(N->getOperand(2), 2) &&
1822 isUndefOrEqual(N->getOperand(3), 3);
1825 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
1826 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
1827 bool X86::isMOVLPMask(SDNode *N) {
1828 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1830 unsigned NumElems = N->getNumOperands();
1831 if (NumElems != 2 && NumElems != 4)
1834 for (unsigned i = 0; i < NumElems/2; ++i)
1835 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
1838 for (unsigned i = NumElems/2; i < NumElems; ++i)
1839 if (!isUndefOrEqual(N->getOperand(i), i))
1845 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
1846 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
1848 bool X86::isMOVHPMask(SDNode *N) {
1849 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1851 unsigned NumElems = N->getNumOperands();
1852 if (NumElems != 2 && NumElems != 4)
1855 for (unsigned i = 0; i < NumElems/2; ++i)
1856 if (!isUndefOrEqual(N->getOperand(i), i))
1859 for (unsigned i = 0; i < NumElems/2; ++i) {
1860 SDOperand Arg = N->getOperand(i + NumElems/2);
1861 if (!isUndefOrEqual(Arg, i + NumElems))
1868 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
1869 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
1870 bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts,
1871 bool V2IsSplat = false) {
1872 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1875 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
1876 SDOperand BitI = Elts[i];
1877 SDOperand BitI1 = Elts[i+1];
1878 if (!isUndefOrEqual(BitI, j))
1881 if (isUndefOrEqual(BitI1, NumElts))
1884 if (!isUndefOrEqual(BitI1, j + NumElts))
1892 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
1893 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1894 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
1897 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
1898 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
1899 bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts,
1900 bool V2IsSplat = false) {
1901 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1904 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) {
1905 SDOperand BitI = Elts[i];
1906 SDOperand BitI1 = Elts[i+1];
1907 if (!isUndefOrEqual(BitI, j + NumElts/2))
1910 if (isUndefOrEqual(BitI1, NumElts))
1913 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
1921 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
1922 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1923 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat);
1926 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
1927 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
1929 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
1930 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1932 unsigned NumElems = N->getNumOperands();
1933 if (NumElems != 4 && NumElems != 8 && NumElems != 16)
1936 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
1937 SDOperand BitI = N->getOperand(i);
1938 SDOperand BitI1 = N->getOperand(i+1);
1940 if (!isUndefOrEqual(BitI, j))
1942 if (!isUndefOrEqual(BitI1, j))
1949 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
1950 /// specifies a shuffle of elements that is suitable for input to MOVSS,
1951 /// MOVSD, and MOVD, i.e. setting the lowest element.
1952 static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) {
1953 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
1956 if (!isUndefOrEqual(Elts[0], NumElts))
1959 for (unsigned i = 1; i < NumElts; ++i) {
1960 if (!isUndefOrEqual(Elts[i], i))
1967 bool X86::isMOVLMask(SDNode *N) {
1968 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1969 return ::isMOVLMask(N->op_begin(), N->getNumOperands());
1972 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
1973 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
1974 /// element of vector 2 and the other elements to come from vector 1 in order.
1975 static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps,
1976 bool V2IsSplat = false,
1977 bool V2IsUndef = false) {
1978 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
1981 if (!isUndefOrEqual(Ops[0], 0))
1984 for (unsigned i = 1; i < NumOps; ++i) {
1985 SDOperand Arg = Ops[i];
1986 if (!(isUndefOrEqual(Arg, i+NumOps) ||
1987 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) ||
1988 (V2IsSplat && isUndefOrEqual(Arg, NumOps))))
1995 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false,
1996 bool V2IsUndef = false) {
1997 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1998 return isCommutedMOVL(N->op_begin(), N->getNumOperands(),
1999 V2IsSplat, V2IsUndef);
2002 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2003 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2004 bool X86::isMOVSHDUPMask(SDNode *N) {
2005 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2007 if (N->getNumOperands() != 4)
2010 // Expect 1, 1, 3, 3
2011 for (unsigned i = 0; i < 2; ++i) {
2012 SDOperand Arg = N->getOperand(i);
2013 if (Arg.getOpcode() == ISD::UNDEF) continue;
2014 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2015 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2016 if (Val != 1) return false;
2020 for (unsigned i = 2; i < 4; ++i) {
2021 SDOperand Arg = N->getOperand(i);
2022 if (Arg.getOpcode() == ISD::UNDEF) continue;
2023 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2024 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2025 if (Val != 3) return false;
2029 // Don't use movshdup if it can be done with a shufps.
2033 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2034 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2035 bool X86::isMOVSLDUPMask(SDNode *N) {
2036 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2038 if (N->getNumOperands() != 4)
2041 // Expect 0, 0, 2, 2
2042 for (unsigned i = 0; i < 2; ++i) {
2043 SDOperand Arg = N->getOperand(i);
2044 if (Arg.getOpcode() == ISD::UNDEF) continue;
2045 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2046 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2047 if (Val != 0) return false;
2051 for (unsigned i = 2; i < 4; ++i) {
2052 SDOperand Arg = N->getOperand(i);
2053 if (Arg.getOpcode() == ISD::UNDEF) continue;
2054 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2055 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2056 if (Val != 2) return false;
2060 // Don't use movshdup if it can be done with a shufps.
2064 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2065 /// a splat of a single element.
2066 static bool isSplatMask(SDNode *N) {
2067 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2069 // This is a splat operation if each element of the permute is the same, and
2070 // if the value doesn't reference the second vector.
2071 unsigned NumElems = N->getNumOperands();
2072 SDOperand ElementBase;
2074 for (; i != NumElems; ++i) {
2075 SDOperand Elt = N->getOperand(i);
2076 if (isa<ConstantSDNode>(Elt)) {
2082 if (!ElementBase.Val)
2085 for (; i != NumElems; ++i) {
2086 SDOperand Arg = N->getOperand(i);
2087 if (Arg.getOpcode() == ISD::UNDEF) continue;
2088 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2089 if (Arg != ElementBase) return false;
2092 // Make sure it is a splat of the first vector operand.
2093 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
2096 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
2097 /// a splat of a single element and it's a 2 or 4 element mask.
2098 bool X86::isSplatMask(SDNode *N) {
2099 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2101 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
2102 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
2104 return ::isSplatMask(N);
2107 /// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand
2108 /// specifies a splat of zero element.
2109 bool X86::isSplatLoMask(SDNode *N) {
2110 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2112 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i)
2113 if (!isUndefOrEqual(N->getOperand(i), 0))
2118 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2119 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2121 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2122 unsigned NumOperands = N->getNumOperands();
2123 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2125 for (unsigned i = 0; i < NumOperands; ++i) {
2127 SDOperand Arg = N->getOperand(NumOperands-i-1);
2128 if (Arg.getOpcode() != ISD::UNDEF)
2129 Val = cast<ConstantSDNode>(Arg)->getValue();
2130 if (Val >= NumOperands) Val -= NumOperands;
2132 if (i != NumOperands - 1)
2139 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2140 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2142 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2144 // 8 nodes, but we only care about the last 4.
2145 for (unsigned i = 7; i >= 4; --i) {
2147 SDOperand Arg = N->getOperand(i);
2148 if (Arg.getOpcode() != ISD::UNDEF)
2149 Val = cast<ConstantSDNode>(Arg)->getValue();
2158 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2159 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2161 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2163 // 8 nodes, but we only care about the first 4.
2164 for (int i = 3; i >= 0; --i) {
2166 SDOperand Arg = N->getOperand(i);
2167 if (Arg.getOpcode() != ISD::UNDEF)
2168 Val = cast<ConstantSDNode>(Arg)->getValue();
2177 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2178 /// specifies a 8 element shuffle that can be broken into a pair of
2179 /// PSHUFHW and PSHUFLW.
2180 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2181 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2183 if (N->getNumOperands() != 8)
2186 // Lower quadword shuffled.
2187 for (unsigned i = 0; i != 4; ++i) {
2188 SDOperand Arg = N->getOperand(i);
2189 if (Arg.getOpcode() == ISD::UNDEF) continue;
2190 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2191 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2196 // Upper quadword shuffled.
2197 for (unsigned i = 4; i != 8; ++i) {
2198 SDOperand Arg = N->getOperand(i);
2199 if (Arg.getOpcode() == ISD::UNDEF) continue;
2200 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2201 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2202 if (Val < 4 || Val > 7)
2209 /// CommuteVectorShuffle - Swap vector_shuffle operandsas well as
2210 /// values in ther permute mask.
2211 static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1,
2212 SDOperand &V2, SDOperand &Mask,
2213 SelectionDAG &DAG) {
2214 MVT::ValueType VT = Op.getValueType();
2215 MVT::ValueType MaskVT = Mask.getValueType();
2216 MVT::ValueType EltVT = MVT::getVectorBaseType(MaskVT);
2217 unsigned NumElems = Mask.getNumOperands();
2218 SmallVector<SDOperand, 8> MaskVec;
2220 for (unsigned i = 0; i != NumElems; ++i) {
2221 SDOperand Arg = Mask.getOperand(i);
2222 if (Arg.getOpcode() == ISD::UNDEF) {
2223 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2226 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2227 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2229 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2231 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2235 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2236 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2239 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2240 /// match movhlps. The lower half elements should come from upper half of
2241 /// V1 (and in order), and the upper half elements should come from the upper
2242 /// half of V2 (and in order).
2243 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2244 unsigned NumElems = Mask->getNumOperands();
2247 for (unsigned i = 0, e = 2; i != e; ++i)
2248 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2250 for (unsigned i = 2; i != 4; ++i)
2251 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2256 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2257 /// is promoted to a vector.
2258 static inline bool isScalarLoadToVector(SDNode *N) {
2259 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2260 N = N->getOperand(0).Val;
2261 return ISD::isNON_EXTLoad(N);
2266 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2267 /// match movlp{s|d}. The lower half elements should come from lower half of
2268 /// V1 (and in order), and the upper half elements should come from the upper
2269 /// half of V2 (and in order). And since V1 will become the source of the
2270 /// MOVLP, it must be either a vector load or a scalar load to vector.
2271 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) {
2272 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2274 // Is V2 is a vector load, don't do this transformation. We will try to use
2275 // load folding shufps op.
2276 if (ISD::isNON_EXTLoad(V2))
2279 unsigned NumElems = Mask->getNumOperands();
2280 if (NumElems != 2 && NumElems != 4)
2282 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2283 if (!isUndefOrEqual(Mask->getOperand(i), i))
2285 for (unsigned i = NumElems/2; i != NumElems; ++i)
2286 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2291 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2293 static bool isSplatVector(SDNode *N) {
2294 if (N->getOpcode() != ISD::BUILD_VECTOR)
2297 SDOperand SplatValue = N->getOperand(0);
2298 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2299 if (N->getOperand(i) != SplatValue)
2304 /// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2306 static bool isUndefShuffle(SDNode *N) {
2307 if (N->getOpcode() != ISD::BUILD_VECTOR)
2310 SDOperand V1 = N->getOperand(0);
2311 SDOperand V2 = N->getOperand(1);
2312 SDOperand Mask = N->getOperand(2);
2313 unsigned NumElems = Mask.getNumOperands();
2314 for (unsigned i = 0; i != NumElems; ++i) {
2315 SDOperand Arg = Mask.getOperand(i);
2316 if (Arg.getOpcode() != ISD::UNDEF) {
2317 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2318 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF)
2320 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF)
2327 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2328 /// that point to V2 points to its first element.
2329 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2330 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2332 bool Changed = false;
2333 SmallVector<SDOperand, 8> MaskVec;
2334 unsigned NumElems = Mask.getNumOperands();
2335 for (unsigned i = 0; i != NumElems; ++i) {
2336 SDOperand Arg = Mask.getOperand(i);
2337 if (Arg.getOpcode() != ISD::UNDEF) {
2338 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2339 if (Val > NumElems) {
2340 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2344 MaskVec.push_back(Arg);
2348 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
2349 &MaskVec[0], MaskVec.size());
2353 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2354 /// operation of specified width.
2355 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2356 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2357 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2359 SmallVector<SDOperand, 8> MaskVec;
2360 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2361 for (unsigned i = 1; i != NumElems; ++i)
2362 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2363 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2366 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2367 /// of specified width.
2368 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2369 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2370 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2371 SmallVector<SDOperand, 8> MaskVec;
2372 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2373 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2374 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2376 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2379 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2380 /// of specified width.
2381 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2382 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2383 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2384 unsigned Half = NumElems/2;
2385 SmallVector<SDOperand, 8> MaskVec;
2386 for (unsigned i = 0; i != Half; ++i) {
2387 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2388 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2390 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size());
2393 /// getZeroVector - Returns a vector of specified type with all zero elements.
2395 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2396 assert(MVT::isVector(VT) && "Expected a vector type");
2397 unsigned NumElems = getVectorNumElements(VT);
2398 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2399 bool isFP = MVT::isFloatingPoint(EVT);
2400 SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT);
2401 SmallVector<SDOperand, 8> ZeroVec(NumElems, Zero);
2402 return DAG.getNode(ISD::BUILD_VECTOR, VT, &ZeroVec[0], ZeroVec.size());
2405 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2407 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2408 SDOperand V1 = Op.getOperand(0);
2409 SDOperand Mask = Op.getOperand(2);
2410 MVT::ValueType VT = Op.getValueType();
2411 unsigned NumElems = Mask.getNumOperands();
2412 Mask = getUnpacklMask(NumElems, DAG);
2413 while (NumElems != 4) {
2414 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2417 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2419 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2420 Mask = getZeroVector(MaskVT, DAG);
2421 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2422 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2423 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2426 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2428 static inline bool isZeroNode(SDOperand Elt) {
2429 return ((isa<ConstantSDNode>(Elt) &&
2430 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2431 (isa<ConstantFPSDNode>(Elt) &&
2432 cast<ConstantFPSDNode>(Elt)->isExactlyValue(0.0)));
2435 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2436 /// vector and zero or undef vector.
2437 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2438 unsigned NumElems, unsigned Idx,
2439 bool isZero, SelectionDAG &DAG) {
2440 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2441 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2442 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2443 SDOperand Zero = DAG.getConstant(0, EVT);
2444 SmallVector<SDOperand, 8> MaskVec(NumElems, Zero);
2445 MaskVec[Idx] = DAG.getConstant(NumElems, EVT);
2446 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2447 &MaskVec[0], MaskVec.size());
2448 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2451 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2453 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
2454 unsigned NumNonZero, unsigned NumZero,
2455 SelectionDAG &DAG, TargetLowering &TLI) {
2461 for (unsigned i = 0; i < 16; ++i) {
2462 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
2463 if (ThisIsNonZero && First) {
2465 V = getZeroVector(MVT::v8i16, DAG);
2467 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2472 SDOperand ThisElt(0, 0), LastElt(0, 0);
2473 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
2474 if (LastIsNonZero) {
2475 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
2477 if (ThisIsNonZero) {
2478 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
2479 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
2480 ThisElt, DAG.getConstant(8, MVT::i8));
2482 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
2487 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
2488 DAG.getConstant(i/2, TLI.getPointerTy()));
2492 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
2495 /// LowerBuildVectorv16i8 - Custom lower build_vector of v8i16.
2497 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
2498 unsigned NumNonZero, unsigned NumZero,
2499 SelectionDAG &DAG, TargetLowering &TLI) {
2505 for (unsigned i = 0; i < 8; ++i) {
2506 bool isNonZero = (NonZeros & (1 << i)) != 0;
2510 V = getZeroVector(MVT::v8i16, DAG);
2512 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2515 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
2516 DAG.getConstant(i, TLI.getPointerTy()));
2524 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2525 // All zero's are handled with pxor.
2526 if (ISD::isBuildVectorAllZeros(Op.Val))
2529 // All one's are handled with pcmpeqd.
2530 if (ISD::isBuildVectorAllOnes(Op.Val))
2533 MVT::ValueType VT = Op.getValueType();
2534 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2535 unsigned EVTBits = MVT::getSizeInBits(EVT);
2537 unsigned NumElems = Op.getNumOperands();
2538 unsigned NumZero = 0;
2539 unsigned NumNonZero = 0;
2540 unsigned NonZeros = 0;
2541 std::set<SDOperand> Values;
2542 for (unsigned i = 0; i < NumElems; ++i) {
2543 SDOperand Elt = Op.getOperand(i);
2544 if (Elt.getOpcode() != ISD::UNDEF) {
2546 if (isZeroNode(Elt))
2549 NonZeros |= (1 << i);
2555 if (NumNonZero == 0)
2556 // Must be a mix of zero and undef. Return a zero vector.
2557 return getZeroVector(VT, DAG);
2559 // Splat is obviously ok. Let legalizer expand it to a shuffle.
2560 if (Values.size() == 1)
2563 // Special case for single non-zero element.
2564 if (NumNonZero == 1) {
2565 unsigned Idx = CountTrailingZeros_32(NonZeros);
2566 SDOperand Item = Op.getOperand(Idx);
2567 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
2569 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
2570 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
2573 if (EVTBits == 32) {
2574 // Turn it into a shuffle of zero and zero-extended scalar to vector.
2575 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
2577 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2578 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
2579 SmallVector<SDOperand, 8> MaskVec;
2580 for (unsigned i = 0; i < NumElems; i++)
2581 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
2582 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2583 &MaskVec[0], MaskVec.size());
2584 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
2585 DAG.getNode(ISD::UNDEF, VT), Mask);
2589 // Let legalizer expand 2-wide build_vector's.
2593 // If element VT is < 32 bits, convert it to inserts into a zero vector.
2595 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
2597 if (V.Val) return V;
2600 if (EVTBits == 16) {
2601 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
2603 if (V.Val) return V;
2606 // If element VT is == 32 bits, turn it into a number of shuffles.
2607 SmallVector<SDOperand, 8> V;
2609 if (NumElems == 4 && NumZero > 0) {
2610 for (unsigned i = 0; i < 4; ++i) {
2611 bool isZero = !(NonZeros & (1 << i));
2613 V[i] = getZeroVector(VT, DAG);
2615 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2618 for (unsigned i = 0; i < 2; ++i) {
2619 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
2622 V[i] = V[i*2]; // Must be a zero vector.
2625 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
2626 getMOVLMask(NumElems, DAG));
2629 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2630 getMOVLMask(NumElems, DAG));
2633 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2634 getUnpacklMask(NumElems, DAG));
2639 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
2640 // clears the upper bits.
2641 // FIXME: we can do the same for v4f32 case when we know both parts of
2642 // the lower half come from scalar_to_vector (loadf32). We should do
2643 // that in post legalizer dag combiner with target specific hooks.
2644 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
2646 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2647 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2648 SmallVector<SDOperand, 8> MaskVec;
2649 bool Reverse = (NonZeros & 0x3) == 2;
2650 for (unsigned i = 0; i < 2; ++i)
2652 MaskVec.push_back(DAG.getConstant(1-i, EVT));
2654 MaskVec.push_back(DAG.getConstant(i, EVT));
2655 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
2656 for (unsigned i = 0; i < 2; ++i)
2658 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
2660 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
2661 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2662 &MaskVec[0], MaskVec.size());
2663 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
2666 if (Values.size() > 2) {
2667 // Expand into a number of unpckl*.
2669 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
2670 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
2671 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
2672 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
2673 for (unsigned i = 0; i < NumElems; ++i)
2674 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2676 while (NumElems != 0) {
2677 for (unsigned i = 0; i < NumElems; ++i)
2678 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
2689 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
2690 SDOperand V1 = Op.getOperand(0);
2691 SDOperand V2 = Op.getOperand(1);
2692 SDOperand PermMask = Op.getOperand(2);
2693 MVT::ValueType VT = Op.getValueType();
2694 unsigned NumElems = PermMask.getNumOperands();
2695 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
2696 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
2697 bool V1IsSplat = false;
2698 bool V2IsSplat = false;
2700 if (isUndefShuffle(Op.Val))
2701 return DAG.getNode(ISD::UNDEF, VT);
2703 if (isSplatMask(PermMask.Val)) {
2704 if (NumElems <= 4) return Op;
2705 // Promote it to a v4i32 splat.
2706 return PromoteSplat(Op, DAG);
2709 if (X86::isMOVLMask(PermMask.Val))
2710 return (V1IsUndef) ? V2 : Op;
2712 if (X86::isMOVSHDUPMask(PermMask.Val) ||
2713 X86::isMOVSLDUPMask(PermMask.Val) ||
2714 X86::isMOVHLPSMask(PermMask.Val) ||
2715 X86::isMOVHPMask(PermMask.Val) ||
2716 X86::isMOVLPMask(PermMask.Val))
2719 if (ShouldXformToMOVHLPS(PermMask.Val) ||
2720 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val))
2721 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2723 bool Commuted = false;
2724 V1IsSplat = isSplatVector(V1.Val);
2725 V2IsSplat = isSplatVector(V2.Val);
2726 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
2727 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2728 std::swap(V1IsSplat, V2IsSplat);
2729 std::swap(V1IsUndef, V2IsUndef);
2733 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) {
2734 if (V2IsUndef) return V1;
2735 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2737 // V2 is a splat, so the mask may be malformed. That is, it may point
2738 // to any V2 element. The instruction selectior won't like this. Get
2739 // a corrected mask and commute to form a proper MOVS{S|D}.
2740 SDOperand NewMask = getMOVLMask(NumElems, DAG);
2741 if (NewMask.Val != PermMask.Val)
2742 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2747 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
2748 X86::isUNPCKLMask(PermMask.Val) ||
2749 X86::isUNPCKHMask(PermMask.Val))
2753 // Normalize mask so all entries that point to V2 points to its first
2754 // element then try to match unpck{h|l} again. If match, return a
2755 // new vector_shuffle with the corrected mask.
2756 SDOperand NewMask = NormalizeMask(PermMask, DAG);
2757 if (NewMask.Val != PermMask.Val) {
2758 if (X86::isUNPCKLMask(PermMask.Val, true)) {
2759 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
2760 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2761 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
2762 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
2763 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2768 // Normalize the node to match x86 shuffle ops if needed
2769 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val))
2770 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2773 // Commute is back and try unpck* again.
2774 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG);
2775 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
2776 X86::isUNPCKLMask(PermMask.Val) ||
2777 X86::isUNPCKHMask(PermMask.Val))
2781 // If VT is integer, try PSHUF* first, then SHUFP*.
2782 if (MVT::isInteger(VT)) {
2783 if (X86::isPSHUFDMask(PermMask.Val) ||
2784 X86::isPSHUFHWMask(PermMask.Val) ||
2785 X86::isPSHUFLWMask(PermMask.Val)) {
2786 if (V2.getOpcode() != ISD::UNDEF)
2787 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
2788 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
2792 if (X86::isSHUFPMask(PermMask.Val))
2795 // Handle v8i16 shuffle high / low shuffle node pair.
2796 if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
2797 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2798 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2799 SmallVector<SDOperand, 8> MaskVec;
2800 for (unsigned i = 0; i != 4; ++i)
2801 MaskVec.push_back(PermMask.getOperand(i));
2802 for (unsigned i = 4; i != 8; ++i)
2803 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2804 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2805 &MaskVec[0], MaskVec.size());
2806 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2808 for (unsigned i = 0; i != 4; ++i)
2809 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2810 for (unsigned i = 4; i != 8; ++i)
2811 MaskVec.push_back(PermMask.getOperand(i));
2812 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0],MaskVec.size());
2813 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2816 // Floating point cases in the other order.
2817 if (X86::isSHUFPMask(PermMask.Val))
2819 if (X86::isPSHUFDMask(PermMask.Val) ||
2820 X86::isPSHUFHWMask(PermMask.Val) ||
2821 X86::isPSHUFLWMask(PermMask.Val)) {
2822 if (V2.getOpcode() != ISD::UNDEF)
2823 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
2824 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
2829 if (NumElems == 4) {
2830 MVT::ValueType MaskVT = PermMask.getValueType();
2831 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
2832 SmallVector<std::pair<int, int>, 8> Locs;
2833 Locs.reserve(NumElems);
2834 SmallVector<SDOperand, 8> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2835 SmallVector<SDOperand, 8> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2838 // If no more than two elements come from either vector. This can be
2839 // implemented with two shuffles. First shuffle gather the elements.
2840 // The second shuffle, which takes the first shuffle as both of its
2841 // vector operands, put the elements into the right order.
2842 for (unsigned i = 0; i != NumElems; ++i) {
2843 SDOperand Elt = PermMask.getOperand(i);
2844 if (Elt.getOpcode() == ISD::UNDEF) {
2845 Locs[i] = std::make_pair(-1, -1);
2847 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
2848 if (Val < NumElems) {
2849 Locs[i] = std::make_pair(0, NumLo);
2853 Locs[i] = std::make_pair(1, NumHi);
2854 if (2+NumHi < NumElems)
2855 Mask1[2+NumHi] = Elt;
2860 if (NumLo <= 2 && NumHi <= 2) {
2861 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2862 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2863 &Mask1[0], Mask1.size()));
2864 for (unsigned i = 0; i != NumElems; ++i) {
2865 if (Locs[i].first == -1)
2868 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
2869 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
2870 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
2874 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
2875 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2876 &Mask2[0], Mask2.size()));
2879 // Break it into (shuffle shuffle_hi, shuffle_lo).
2881 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2882 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2883 SmallVector<SDOperand,8> *MaskPtr = &LoMask;
2884 unsigned MaskIdx = 0;
2886 unsigned HiIdx = NumElems/2;
2887 for (unsigned i = 0; i != NumElems; ++i) {
2888 if (i == NumElems/2) {
2894 SDOperand Elt = PermMask.getOperand(i);
2895 if (Elt.getOpcode() == ISD::UNDEF) {
2896 Locs[i] = std::make_pair(-1, -1);
2897 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
2898 Locs[i] = std::make_pair(MaskIdx, LoIdx);
2899 (*MaskPtr)[LoIdx] = Elt;
2902 Locs[i] = std::make_pair(MaskIdx, HiIdx);
2903 (*MaskPtr)[HiIdx] = Elt;
2908 SDOperand LoShuffle =
2909 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2910 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2911 &LoMask[0], LoMask.size()));
2912 SDOperand HiShuffle =
2913 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2914 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2915 &HiMask[0], HiMask.size()));
2916 SmallVector<SDOperand, 8> MaskOps;
2917 for (unsigned i = 0; i != NumElems; ++i) {
2918 if (Locs[i].first == -1) {
2919 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
2921 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
2922 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
2925 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
2926 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2927 &MaskOps[0], MaskOps.size()));
2934 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2935 if (!isa<ConstantSDNode>(Op.getOperand(1)))
2938 MVT::ValueType VT = Op.getValueType();
2939 // TODO: handle v16i8.
2940 if (MVT::getSizeInBits(VT) == 16) {
2941 // Transform it so it match pextrw which produces a 32-bit result.
2942 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
2943 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
2944 Op.getOperand(0), Op.getOperand(1));
2945 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
2946 DAG.getValueType(VT));
2947 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
2948 } else if (MVT::getSizeInBits(VT) == 32) {
2949 SDOperand Vec = Op.getOperand(0);
2950 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
2953 // SHUFPS the element to the lowest double word, then movss.
2954 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2955 SmallVector<SDOperand, 8> IdxVec;
2956 IdxVec.push_back(DAG.getConstant(Idx, MVT::getVectorBaseType(MaskVT)));
2957 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2958 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2959 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2960 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2961 &IdxVec[0], IdxVec.size());
2962 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
2963 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
2964 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
2965 DAG.getConstant(0, getPointerTy()));
2966 } else if (MVT::getSizeInBits(VT) == 64) {
2967 SDOperand Vec = Op.getOperand(0);
2968 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
2972 // UNPCKHPD the element to the lowest double word, then movsd.
2973 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
2974 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
2975 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2976 SmallVector<SDOperand, 8> IdxVec;
2977 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorBaseType(MaskVT)));
2978 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2979 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
2980 &IdxVec[0], IdxVec.size());
2981 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
2982 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
2983 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
2984 DAG.getConstant(0, getPointerTy()));
2991 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2992 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
2993 // as its second argument.
2994 MVT::ValueType VT = Op.getValueType();
2995 MVT::ValueType BaseVT = MVT::getVectorBaseType(VT);
2996 SDOperand N0 = Op.getOperand(0);
2997 SDOperand N1 = Op.getOperand(1);
2998 SDOperand N2 = Op.getOperand(2);
2999 if (MVT::getSizeInBits(BaseVT) == 16) {
3000 if (N1.getValueType() != MVT::i32)
3001 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
3002 if (N2.getValueType() != MVT::i32)
3003 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(), MVT::i32);
3004 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
3005 } else if (MVT::getSizeInBits(BaseVT) == 32) {
3006 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
3009 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
3010 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
3011 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
3012 SmallVector<SDOperand, 8> MaskVec;
3013 MaskVec.push_back(DAG.getConstant(4, BaseVT));
3014 for (unsigned i = 1; i <= 3; ++i)
3015 MaskVec.push_back(DAG.getConstant(i, BaseVT));
3016 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
3017 DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
3018 &MaskVec[0], MaskVec.size()));
3020 // Use two pinsrw instructions to insert a 32 bit value.
3022 if (MVT::isFloatingPoint(N1.getValueType())) {
3023 if (ISD::isNON_EXTLoad(N1.Val)) {
3024 // Just load directly from f32mem to GR32.
3025 LoadSDNode *LD = cast<LoadSDNode>(N1);
3026 N1 = DAG.getLoad(MVT::i32, LD->getChain(), LD->getBasePtr(),
3027 LD->getSrcValue(), LD->getSrcValueOffset());
3029 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
3030 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
3031 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
3032 DAG.getConstant(0, getPointerTy()));
3035 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
3036 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3037 DAG.getConstant(Idx, getPointerTy()));
3038 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
3039 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
3040 DAG.getConstant(Idx+1, getPointerTy()));
3041 return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
3049 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
3050 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
3051 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
3054 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
3055 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
3056 // one of the above mentioned nodes. It has to be wrapped because otherwise
3057 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
3058 // be used to form addressing mode. These wrapped nodes will be selected
3061 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
3062 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3063 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(),
3065 CP->getAlignment());
3066 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3067 // With PIC, the address is actually $g + Offset.
3068 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3069 !Subtarget->isPICStyleRIPRel()) {
3070 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3071 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3079 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
3080 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
3081 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
3082 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3083 // With PIC, the address is actually $g + Offset.
3084 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3085 !Subtarget->isPICStyleRIPRel()) {
3086 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3087 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3091 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to
3092 // load the value at address GV, not the value of GV itself. This means that
3093 // the GlobalAddress must be in the base or index register of the address, not
3094 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call
3095 // The same applies for external symbols during PIC codegen
3096 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
3097 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0);
3103 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
3104 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
3105 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
3106 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3107 // With PIC, the address is actually $g + Offset.
3108 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3109 !Subtarget->isPICStyleRIPRel()) {
3110 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3111 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3118 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
3119 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3120 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
3121 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
3122 // With PIC, the address is actually $g + Offset.
3123 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
3124 !Subtarget->isPICStyleRIPRel()) {
3125 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3126 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3133 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
3134 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
3135 "Not an i64 shift!");
3136 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
3137 SDOperand ShOpLo = Op.getOperand(0);
3138 SDOperand ShOpHi = Op.getOperand(1);
3139 SDOperand ShAmt = Op.getOperand(2);
3140 SDOperand Tmp1 = isSRA ?
3141 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) :
3142 DAG.getConstant(0, MVT::i32);
3144 SDOperand Tmp2, Tmp3;
3145 if (Op.getOpcode() == ISD::SHL_PARTS) {
3146 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
3147 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
3149 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
3150 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
3153 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3154 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
3155 DAG.getConstant(32, MVT::i8));
3156 SDOperand COps[]={DAG.getEntryNode(), AndNode, DAG.getConstant(0, MVT::i8)};
3157 SDOperand InFlag = DAG.getNode(X86ISD::CMP, VTs, 2, COps, 3).getValue(1);
3160 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3162 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag);
3163 SmallVector<SDOperand, 4> Ops;
3164 if (Op.getOpcode() == ISD::SHL_PARTS) {
3165 Ops.push_back(Tmp2);
3166 Ops.push_back(Tmp3);
3168 Ops.push_back(InFlag);
3169 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3170 InFlag = Hi.getValue(1);
3173 Ops.push_back(Tmp3);
3174 Ops.push_back(Tmp1);
3176 Ops.push_back(InFlag);
3177 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3179 Ops.push_back(Tmp2);
3180 Ops.push_back(Tmp3);
3182 Ops.push_back(InFlag);
3183 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3184 InFlag = Lo.getValue(1);
3187 Ops.push_back(Tmp3);
3188 Ops.push_back(Tmp1);
3190 Ops.push_back(InFlag);
3191 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3194 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32);
3198 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size());
3201 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
3202 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
3203 Op.getOperand(0).getValueType() >= MVT::i16 &&
3204 "Unknown SINT_TO_FP to lower!");
3207 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
3208 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
3209 MachineFunction &MF = DAG.getMachineFunction();
3210 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
3211 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3212 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
3213 StackSlot, NULL, 0);
3218 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
3220 Tys = DAG.getVTList(MVT::f64, MVT::Other);
3221 SmallVector<SDOperand, 8> Ops;
3222 Ops.push_back(Chain);
3223 Ops.push_back(StackSlot);
3224 Ops.push_back(DAG.getValueType(SrcVT));
3225 Result = DAG.getNode(X86ScalarSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
3226 Tys, &Ops[0], Ops.size());
3229 Chain = Result.getValue(1);
3230 SDOperand InFlag = Result.getValue(2);
3232 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
3233 // shouldn't be necessary except that RFP cannot be live across
3234 // multiple blocks. When stackifier is fixed, they can be uncoupled.
3235 MachineFunction &MF = DAG.getMachineFunction();
3236 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3237 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3238 Tys = DAG.getVTList(MVT::Other);
3239 SmallVector<SDOperand, 8> Ops;
3240 Ops.push_back(Chain);
3241 Ops.push_back(Result);
3242 Ops.push_back(StackSlot);
3243 Ops.push_back(DAG.getValueType(Op.getValueType()));
3244 Ops.push_back(InFlag);
3245 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size());
3246 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0);
3252 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
3253 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
3254 "Unknown FP_TO_SINT to lower!");
3255 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
3257 MachineFunction &MF = DAG.getMachineFunction();
3258 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
3259 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3260 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3263 switch (Op.getValueType()) {
3264 default: assert(0 && "Invalid FP_TO_SINT to lower!");
3265 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
3266 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
3267 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
3270 SDOperand Chain = DAG.getEntryNode();
3271 SDOperand Value = Op.getOperand(0);
3273 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
3274 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0);
3275 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other);
3277 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
3279 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
3280 Chain = Value.getValue(1);
3281 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3282 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3285 // Build the FP_TO_INT*_IN_MEM
3286 SDOperand Ops[] = { Chain, Value, StackSlot };
3287 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);
3290 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
3293 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
3294 MVT::ValueType VT = Op.getValueType();
3295 const Type *OpNTy = MVT::getTypeForValueType(VT);
3296 std::vector<Constant*> CV;
3297 if (VT == MVT::f64) {
3298 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(~(1ULL << 63))));
3299 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3301 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(~(1U << 31))));
3302 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3303 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3304 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3306 Constant *CS = ConstantStruct::get(CV);
3307 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3308 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3309 SmallVector<SDOperand, 3> Ops;
3310 Ops.push_back(DAG.getEntryNode());
3311 Ops.push_back(CPIdx);
3312 Ops.push_back(DAG.getSrcValue(NULL));
3313 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3314 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
3317 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
3318 MVT::ValueType VT = Op.getValueType();
3319 const Type *OpNTy = MVT::getTypeForValueType(VT);
3320 std::vector<Constant*> CV;
3321 if (VT == MVT::f64) {
3322 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(1ULL << 63)));
3323 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3325 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(1U << 31)));
3326 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3327 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3328 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3330 Constant *CS = ConstantStruct::get(CV);
3331 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3332 SDVTList Tys = DAG.getVTList(VT, MVT::Other);
3333 SmallVector<SDOperand, 3> Ops;
3334 Ops.push_back(DAG.getEntryNode());
3335 Ops.push_back(CPIdx);
3336 Ops.push_back(DAG.getSrcValue(NULL));
3337 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3338 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
3341 SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
3342 SDOperand Op0 = Op.getOperand(0);
3343 SDOperand Op1 = Op.getOperand(1);
3344 MVT::ValueType VT = Op.getValueType();
3345 MVT::ValueType SrcVT = Op1.getValueType();
3346 const Type *SrcTy = MVT::getTypeForValueType(SrcVT);
3348 // If second operand is smaller, extend it first.
3349 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) {
3350 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
3354 // First get the sign bit of second operand.
3355 std::vector<Constant*> CV;
3356 if (SrcVT == MVT::f64) {
3357 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(1ULL << 63)));
3358 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3360 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(1U << 31)));
3361 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3362 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3363 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3365 Constant *CS = ConstantStruct::get(CV);
3366 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3367 SDVTList Tys = DAG.getVTList(SrcVT, MVT::Other);
3368 SmallVector<SDOperand, 3> Ops;
3369 Ops.push_back(DAG.getEntryNode());
3370 Ops.push_back(CPIdx);
3371 Ops.push_back(DAG.getSrcValue(NULL));
3372 SDOperand Mask1 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3373 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);
3375 // Shift sign bit right or left if the two operands have different types.
3376 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) {
3377 // Op0 is MVT::f32, Op1 is MVT::f64.
3378 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
3379 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
3380 DAG.getConstant(32, MVT::i32));
3381 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
3382 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
3383 DAG.getConstant(0, getPointerTy()));
3386 // Clear first operand sign bit.
3388 if (VT == MVT::f64) {
3389 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(~(1ULL << 63))));
3390 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3392 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(~(1U << 31))));
3393 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3394 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3395 CV.push_back(ConstantFP::get(SrcTy, 0.0));
3397 CS = ConstantStruct::get(CV);
3398 CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3399 Tys = DAG.getVTList(VT, MVT::Other);
3401 Ops.push_back(DAG.getEntryNode());
3402 Ops.push_back(CPIdx);
3403 Ops.push_back(DAG.getSrcValue(NULL));
3404 SDOperand Mask2 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size());
3405 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);
3407 // Or the value with the sign bit.
3408 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
3411 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG,
3413 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
3415 SDOperand Op0 = Op.getOperand(0);
3416 SDOperand Op1 = Op.getOperand(1);
3417 SDOperand CC = Op.getOperand(2);
3418 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3419 const MVT::ValueType *VTs1 = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3420 const MVT::ValueType *VTs2 = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
3421 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
3424 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
3426 SDOperand Ops1[] = { Chain, Op0, Op1 };
3427 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, Ops1, 3).getValue(1);
3428 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
3429 return DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3432 assert(isFP && "Illegal integer SetCC!");
3434 SDOperand COps[] = { Chain, Op0, Op1 };
3435 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, COps, 3).getValue(1);
3437 switch (SetCCOpcode) {
3438 default: assert(false && "Illegal floating point SetCC!");
3439 case ISD::SETOEQ: { // !PF & ZF
3440 SDOperand Ops1[] = { DAG.getConstant(X86::COND_NP, MVT::i8), Cond };
3441 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3442 SDOperand Ops2[] = { DAG.getConstant(X86::COND_E, MVT::i8),
3444 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3445 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
3447 case ISD::SETUNE: { // PF | !ZF
3448 SDOperand Ops1[] = { DAG.getConstant(X86::COND_P, MVT::i8), Cond };
3449 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2);
3450 SDOperand Ops2[] = { DAG.getConstant(X86::COND_NE, MVT::i8),
3452 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2);
3453 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
3458 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
3459 bool addTest = true;
3460 SDOperand Chain = DAG.getEntryNode();
3461 SDOperand Cond = Op.getOperand(0);
3463 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3465 if (Cond.getOpcode() == ISD::SETCC)
3466 Cond = LowerSETCC(Cond, DAG, Chain);
3468 if (Cond.getOpcode() == X86ISD::SETCC) {
3469 CC = Cond.getOperand(0);
3471 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3472 // (since flag operand cannot be shared). Use it as the condition setting
3473 // operand in place of the X86ISD::SETCC.
3474 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3475 // to use a test instead of duplicating the X86ISD::CMP (for register
3476 // pressure reason)?
3477 SDOperand Cmp = Cond.getOperand(1);
3478 unsigned Opc = Cmp.getOpcode();
3479 bool IllegalFPCMov = !X86ScalarSSE &&
3480 MVT::isFloatingPoint(Op.getValueType()) &&
3481 !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
3482 if ((Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) &&
3484 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3485 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3491 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3492 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3493 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3496 VTs = DAG.getNodeValueTypes(Op.getValueType(), MVT::Flag);
3497 SmallVector<SDOperand, 4> Ops;
3498 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
3499 // condition is true.
3500 Ops.push_back(Op.getOperand(2));
3501 Ops.push_back(Op.getOperand(1));
3503 Ops.push_back(Cond.getValue(1));
3504 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size());
3507 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
3508 bool addTest = true;
3509 SDOperand Chain = Op.getOperand(0);
3510 SDOperand Cond = Op.getOperand(1);
3511 SDOperand Dest = Op.getOperand(2);
3513 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
3515 if (Cond.getOpcode() == ISD::SETCC)
3516 Cond = LowerSETCC(Cond, DAG, Chain);
3518 if (Cond.getOpcode() == X86ISD::SETCC) {
3519 CC = Cond.getOperand(0);
3521 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3522 // (since flag operand cannot be shared). Use it as the condition setting
3523 // operand in place of the X86ISD::SETCC.
3524 // If the X86ISD::SETCC has more than one use, then perhaps it's better
3525 // to use a test instead of duplicating the X86ISD::CMP (for register
3526 // pressure reason)?
3527 SDOperand Cmp = Cond.getOperand(1);
3528 unsigned Opc = Cmp.getOpcode();
3529 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) {
3530 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) };
3531 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3);
3537 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
3538 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) };
3539 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3);
3541 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
3542 Cond, Op.getOperand(2), CC, Cond.getValue(1));
3545 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
3546 unsigned CallingConv= cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3548 if (Subtarget->is64Bit())
3549 return LowerX86_64CCCCallTo(Op, DAG, CallingConv);
3551 switch (CallingConv) {
3553 assert(0 && "Unsupported calling convention");
3554 case CallingConv::Fast:
3555 // TODO: Implement fastcc
3557 case CallingConv::C:
3558 case CallingConv::X86_StdCall:
3559 return LowerCCCCallTo(Op, DAG, CallingConv);
3560 case CallingConv::X86_FastCall:
3561 return LowerFastCCCallTo(Op, DAG, CallingConv);
3566 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
3567 MachineFunction &MF = DAG.getMachineFunction();
3568 const Function* Fn = MF.getFunction();
3569 if (Fn->hasExternalLinkage() &&
3570 Subtarget->isTargetCygMing() &&
3571 Fn->getName() == "main")
3572 MF.getInfo<X86FunctionInfo>()->setForceFramePointer(true);
3574 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3575 if (Subtarget->is64Bit())
3576 return LowerX86_64CCCArguments(Op, DAG);
3580 assert(0 && "Unsupported calling convention");
3581 case CallingConv::Fast:
3582 // TODO: implement fastcc.
3585 case CallingConv::C:
3586 return LowerCCCArguments(Op, DAG);
3587 case CallingConv::X86_StdCall:
3588 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(StdCall);
3589 return LowerCCCArguments(Op, DAG, true);
3590 case CallingConv::X86_FastCall:
3591 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(FastCall);
3592 return LowerFastCCArguments(Op, DAG);
3596 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
3597 SDOperand InFlag(0, 0);
3598 SDOperand Chain = Op.getOperand(0);
3600 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3601 if (Align == 0) Align = 1;
3603 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3604 // If not DWORD aligned, call memset if size is less than the threshold.
3605 // It knows how to align to the right boundary first.
3606 if ((Align & 3) != 0 ||
3607 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
3608 MVT::ValueType IntPtr = getPointerTy();
3609 const Type *IntPtrTy = getTargetData()->getIntPtrType();
3610 TargetLowering::ArgListTy Args;
3611 TargetLowering::ArgListEntry Entry;
3612 Entry.Node = Op.getOperand(1);
3613 Entry.Ty = IntPtrTy;
3614 Entry.isSigned = false;
3615 Entry.isInReg = false;
3616 Entry.isSRet = false;
3617 Args.push_back(Entry);
3618 // Extend the unsigned i8 argument to be an int value for the call.
3619 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
3620 Entry.Ty = IntPtrTy;
3621 Entry.isSigned = false;
3622 Entry.isInReg = false;
3623 Entry.isSRet = false;
3624 Args.push_back(Entry);
3625 Entry.Node = Op.getOperand(3);
3626 Args.push_back(Entry);
3627 std::pair<SDOperand,SDOperand> CallResult =
3628 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
3629 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
3630 return CallResult.second;
3635 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
3636 unsigned BytesLeft = 0;
3637 bool TwoRepStos = false;
3640 uint64_t Val = ValC->getValue() & 255;
3642 // If the value is a constant, then we can potentially use larger sets.
3643 switch (Align & 3) {
3644 case 2: // WORD aligned
3647 Val = (Val << 8) | Val;
3649 case 0: // DWORD aligned
3652 Val = (Val << 8) | Val;
3653 Val = (Val << 16) | Val;
3654 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned
3657 Val = (Val << 32) | Val;
3660 default: // Byte aligned
3663 Count = Op.getOperand(3);
3667 if (AVT > MVT::i8) {
3669 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
3670 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
3671 BytesLeft = I->getValue() % UBytes;
3673 assert(AVT >= MVT::i32 &&
3674 "Do not use rep;stos if not at least DWORD aligned");
3675 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
3676 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
3681 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
3683 InFlag = Chain.getValue(1);
3686 Count = Op.getOperand(3);
3687 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
3688 InFlag = Chain.getValue(1);
3691 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
3693 InFlag = Chain.getValue(1);
3694 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
3695 Op.getOperand(1), InFlag);
3696 InFlag = Chain.getValue(1);
3698 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3699 SmallVector<SDOperand, 8> Ops;
3700 Ops.push_back(Chain);
3701 Ops.push_back(DAG.getValueType(AVT));
3702 Ops.push_back(InFlag);
3703 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
3706 InFlag = Chain.getValue(1);
3707 Count = Op.getOperand(3);
3708 MVT::ValueType CVT = Count.getValueType();
3709 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
3710 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
3711 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
3713 InFlag = Chain.getValue(1);
3714 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3716 Ops.push_back(Chain);
3717 Ops.push_back(DAG.getValueType(MVT::i8));
3718 Ops.push_back(InFlag);
3719 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size());
3720 } else if (BytesLeft) {
3721 // Issue stores for the last 1 - 7 bytes.
3723 unsigned Val = ValC->getValue() & 255;
3724 unsigned Offset = I->getValue() - BytesLeft;
3725 SDOperand DstAddr = Op.getOperand(1);
3726 MVT::ValueType AddrVT = DstAddr.getValueType();
3727 if (BytesLeft >= 4) {
3728 Val = (Val << 8) | Val;
3729 Val = (Val << 16) | Val;
3730 Value = DAG.getConstant(Val, MVT::i32);
3731 Chain = DAG.getStore(Chain, Value,
3732 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3733 DAG.getConstant(Offset, AddrVT)),
3738 if (BytesLeft >= 2) {
3739 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
3740 Chain = DAG.getStore(Chain, Value,
3741 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3742 DAG.getConstant(Offset, AddrVT)),
3747 if (BytesLeft == 1) {
3748 Value = DAG.getConstant(Val, MVT::i8);
3749 Chain = DAG.getStore(Chain, Value,
3750 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3751 DAG.getConstant(Offset, AddrVT)),
3759 SDOperand X86TargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) {
3760 SDOperand Chain = Op.getOperand(0);
3762 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3763 if (Align == 0) Align = 1;
3765 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3766 // If not DWORD aligned, call memcpy if size is less than the threshold.
3767 // It knows how to align to the right boundary first.
3768 if ((Align & 3) != 0 ||
3769 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
3770 MVT::ValueType IntPtr = getPointerTy();
3771 TargetLowering::ArgListTy Args;
3772 TargetLowering::ArgListEntry Entry;
3773 Entry.Ty = getTargetData()->getIntPtrType();
3774 Entry.isSigned = false;
3775 Entry.isInReg = false;
3776 Entry.isSRet = false;
3777 Entry.Node = Op.getOperand(1); Args.push_back(Entry);
3778 Entry.Node = Op.getOperand(2); Args.push_back(Entry);
3779 Entry.Node = Op.getOperand(3); Args.push_back(Entry);
3780 std::pair<SDOperand,SDOperand> CallResult =
3781 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false,
3782 DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
3783 return CallResult.second;
3788 unsigned BytesLeft = 0;
3789 bool TwoRepMovs = false;
3790 switch (Align & 3) {
3791 case 2: // WORD aligned
3794 case 0: // DWORD aligned
3796 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned
3799 default: // Byte aligned
3801 Count = Op.getOperand(3);
3805 if (AVT > MVT::i8) {
3807 unsigned UBytes = MVT::getSizeInBits(AVT) / 8;
3808 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy());
3809 BytesLeft = I->getValue() % UBytes;
3811 assert(AVT >= MVT::i32 &&
3812 "Do not use rep;movs if not at least DWORD aligned");
3813 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(),
3814 Op.getOperand(3), DAG.getConstant(2, MVT::i8));
3819 SDOperand InFlag(0, 0);
3820 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX,
3822 InFlag = Chain.getValue(1);
3823 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI,
3824 Op.getOperand(1), InFlag);
3825 InFlag = Chain.getValue(1);
3826 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI,
3827 Op.getOperand(2), InFlag);
3828 InFlag = Chain.getValue(1);
3830 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3831 SmallVector<SDOperand, 8> Ops;
3832 Ops.push_back(Chain);
3833 Ops.push_back(DAG.getValueType(AVT));
3834 Ops.push_back(InFlag);
3835 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
3838 InFlag = Chain.getValue(1);
3839 Count = Op.getOperand(3);
3840 MVT::ValueType CVT = Count.getValueType();
3841 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
3842 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
3843 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
3845 InFlag = Chain.getValue(1);
3846 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3848 Ops.push_back(Chain);
3849 Ops.push_back(DAG.getValueType(MVT::i8));
3850 Ops.push_back(InFlag);
3851 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size());
3852 } else if (BytesLeft) {
3853 // Issue loads and stores for the last 1 - 7 bytes.
3854 unsigned Offset = I->getValue() - BytesLeft;
3855 SDOperand DstAddr = Op.getOperand(1);
3856 MVT::ValueType DstVT = DstAddr.getValueType();
3857 SDOperand SrcAddr = Op.getOperand(2);
3858 MVT::ValueType SrcVT = SrcAddr.getValueType();
3860 if (BytesLeft >= 4) {
3861 Value = DAG.getLoad(MVT::i32, Chain,
3862 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3863 DAG.getConstant(Offset, SrcVT)),
3865 Chain = Value.getValue(1);
3866 Chain = DAG.getStore(Chain, Value,
3867 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3868 DAG.getConstant(Offset, DstVT)),
3873 if (BytesLeft >= 2) {
3874 Value = DAG.getLoad(MVT::i16, Chain,
3875 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3876 DAG.getConstant(Offset, SrcVT)),
3878 Chain = Value.getValue(1);
3879 Chain = DAG.getStore(Chain, Value,
3880 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3881 DAG.getConstant(Offset, DstVT)),
3887 if (BytesLeft == 1) {
3888 Value = DAG.getLoad(MVT::i8, Chain,
3889 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3890 DAG.getConstant(Offset, SrcVT)),
3892 Chain = Value.getValue(1);
3893 Chain = DAG.getStore(Chain, Value,
3894 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3895 DAG.getConstant(Offset, DstVT)),
3904 X86TargetLowering::LowerREADCYCLCECOUNTER(SDOperand Op, SelectionDAG &DAG) {
3905 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
3906 SDOperand TheOp = Op.getOperand(0);
3907 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheOp, 1);
3908 if (Subtarget->is64Bit()) {
3909 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1));
3910 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::RDX,
3911 MVT::i64, Copy1.getValue(2));
3912 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, Copy2,
3913 DAG.getConstant(32, MVT::i8));
3915 DAG.getNode(ISD::OR, MVT::i64, Copy1, Tmp), Copy2.getValue(1)
3918 Tys = DAG.getVTList(MVT::i64, MVT::Other);
3919 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2);
3922 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1));
3923 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::EDX,
3924 MVT::i32, Copy1.getValue(2));
3925 SDOperand Ops[] = { Copy1, Copy2, Copy2.getValue(1) };
3926 Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
3927 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 3);
3930 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
3931 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
3933 if (!Subtarget->is64Bit()) {
3934 // vastart just stores the address of the VarArgsFrameIndex slot into the
3935 // memory location argument.
3936 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
3937 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(),
3942 // gp_offset (0 - 6 * 8)
3943 // fp_offset (48 - 48 + 8 * 16)
3944 // overflow_arg_area (point to parameters coming in memory).
3946 SmallVector<SDOperand, 8> MemOps;
3947 SDOperand FIN = Op.getOperand(1);
3949 SDOperand Store = DAG.getStore(Op.getOperand(0),
3950 DAG.getConstant(VarArgsGPOffset, MVT::i32),
3951 FIN, SV->getValue(), SV->getOffset());
3952 MemOps.push_back(Store);
3955 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
3956 DAG.getConstant(4, getPointerTy()));
3957 Store = DAG.getStore(Op.getOperand(0),
3958 DAG.getConstant(VarArgsFPOffset, MVT::i32),
3959 FIN, SV->getValue(), SV->getOffset());
3960 MemOps.push_back(Store);
3962 // Store ptr to overflow_arg_area
3963 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
3964 DAG.getConstant(4, getPointerTy()));
3965 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
3966 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(),
3968 MemOps.push_back(Store);
3970 // Store ptr to reg_save_area.
3971 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
3972 DAG.getConstant(8, getPointerTy()));
3973 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
3974 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(),
3976 MemOps.push_back(Store);
3977 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size());
3981 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
3982 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
3984 default: return SDOperand(); // Don't custom lower most intrinsics.
3985 // Comparison intrinsics.
3986 case Intrinsic::x86_sse_comieq_ss:
3987 case Intrinsic::x86_sse_comilt_ss:
3988 case Intrinsic::x86_sse_comile_ss:
3989 case Intrinsic::x86_sse_comigt_ss:
3990 case Intrinsic::x86_sse_comige_ss:
3991 case Intrinsic::x86_sse_comineq_ss:
3992 case Intrinsic::x86_sse_ucomieq_ss:
3993 case Intrinsic::x86_sse_ucomilt_ss:
3994 case Intrinsic::x86_sse_ucomile_ss:
3995 case Intrinsic::x86_sse_ucomigt_ss:
3996 case Intrinsic::x86_sse_ucomige_ss:
3997 case Intrinsic::x86_sse_ucomineq_ss:
3998 case Intrinsic::x86_sse2_comieq_sd:
3999 case Intrinsic::x86_sse2_comilt_sd:
4000 case Intrinsic::x86_sse2_comile_sd:
4001 case Intrinsic::x86_sse2_comigt_sd:
4002 case Intrinsic::x86_sse2_comige_sd:
4003 case Intrinsic::x86_sse2_comineq_sd:
4004 case Intrinsic::x86_sse2_ucomieq_sd:
4005 case Intrinsic::x86_sse2_ucomilt_sd:
4006 case Intrinsic::x86_sse2_ucomile_sd:
4007 case Intrinsic::x86_sse2_ucomigt_sd:
4008 case Intrinsic::x86_sse2_ucomige_sd:
4009 case Intrinsic::x86_sse2_ucomineq_sd: {
4011 ISD::CondCode CC = ISD::SETCC_INVALID;
4014 case Intrinsic::x86_sse_comieq_ss:
4015 case Intrinsic::x86_sse2_comieq_sd:
4019 case Intrinsic::x86_sse_comilt_ss:
4020 case Intrinsic::x86_sse2_comilt_sd:
4024 case Intrinsic::x86_sse_comile_ss:
4025 case Intrinsic::x86_sse2_comile_sd:
4029 case Intrinsic::x86_sse_comigt_ss:
4030 case Intrinsic::x86_sse2_comigt_sd:
4034 case Intrinsic::x86_sse_comige_ss:
4035 case Intrinsic::x86_sse2_comige_sd:
4039 case Intrinsic::x86_sse_comineq_ss:
4040 case Intrinsic::x86_sse2_comineq_sd:
4044 case Intrinsic::x86_sse_ucomieq_ss:
4045 case Intrinsic::x86_sse2_ucomieq_sd:
4046 Opc = X86ISD::UCOMI;
4049 case Intrinsic::x86_sse_ucomilt_ss:
4050 case Intrinsic::x86_sse2_ucomilt_sd:
4051 Opc = X86ISD::UCOMI;
4054 case Intrinsic::x86_sse_ucomile_ss:
4055 case Intrinsic::x86_sse2_ucomile_sd:
4056 Opc = X86ISD::UCOMI;
4059 case Intrinsic::x86_sse_ucomigt_ss:
4060 case Intrinsic::x86_sse2_ucomigt_sd:
4061 Opc = X86ISD::UCOMI;
4064 case Intrinsic::x86_sse_ucomige_ss:
4065 case Intrinsic::x86_sse2_ucomige_sd:
4066 Opc = X86ISD::UCOMI;
4069 case Intrinsic::x86_sse_ucomineq_ss:
4070 case Intrinsic::x86_sse2_ucomineq_sd:
4071 Opc = X86ISD::UCOMI;
4077 SDOperand LHS = Op.getOperand(1);
4078 SDOperand RHS = Op.getOperand(2);
4079 translateX86CC(CC, true, X86CC, LHS, RHS, DAG);
4081 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag);
4082 SDOperand Ops1[] = { DAG.getEntryNode(), LHS, RHS };
4083 SDOperand Cond = DAG.getNode(Opc, VTs, 2, Ops1, 3);
4084 VTs = DAG.getNodeValueTypes(MVT::i8, MVT::Flag);
4085 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond };
4086 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, VTs, 2, Ops2, 2);
4087 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
4092 SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
4093 // Depths > 0 not supported yet!
4094 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4097 // Just load the return address
4098 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4099 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
4102 SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
4103 // Depths > 0 not supported yet!
4104 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4107 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
4108 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI,
4109 DAG.getConstant(4, getPointerTy()));
4112 /// LowerOperation - Provide custom lowering hooks for some operations.
4114 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
4115 switch (Op.getOpcode()) {
4116 default: assert(0 && "Should not custom lower this!");
4117 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4118 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4119 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4120 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
4121 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4122 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4123 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4124 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
4125 case ISD::SHL_PARTS:
4126 case ISD::SRA_PARTS:
4127 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
4128 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4129 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
4130 case ISD::FABS: return LowerFABS(Op, DAG);
4131 case ISD::FNEG: return LowerFNEG(Op, DAG);
4132 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4133 case ISD::SETCC: return LowerSETCC(Op, DAG, DAG.getEntryNode());
4134 case ISD::SELECT: return LowerSELECT(Op, DAG);
4135 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
4136 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4137 case ISD::CALL: return LowerCALL(Op, DAG);
4138 case ISD::RET: return LowerRET(Op, DAG);
4139 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
4140 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
4141 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
4142 case ISD::READCYCLECOUNTER: return LowerREADCYCLCECOUNTER(Op, DAG);
4143 case ISD::VASTART: return LowerVASTART(Op, DAG);
4144 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4145 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4146 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4151 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
4153 default: return NULL;
4154 case X86ISD::SHLD: return "X86ISD::SHLD";
4155 case X86ISD::SHRD: return "X86ISD::SHRD";
4156 case X86ISD::FAND: return "X86ISD::FAND";
4157 case X86ISD::FOR: return "X86ISD::FOR";
4158 case X86ISD::FXOR: return "X86ISD::FXOR";
4159 case X86ISD::FSRL: return "X86ISD::FSRL";
4160 case X86ISD::FILD: return "X86ISD::FILD";
4161 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
4162 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
4163 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
4164 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
4165 case X86ISD::FLD: return "X86ISD::FLD";
4166 case X86ISD::FST: return "X86ISD::FST";
4167 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
4168 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
4169 case X86ISD::CALL: return "X86ISD::CALL";
4170 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
4171 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
4172 case X86ISD::CMP: return "X86ISD::CMP";
4173 case X86ISD::COMI: return "X86ISD::COMI";
4174 case X86ISD::UCOMI: return "X86ISD::UCOMI";
4175 case X86ISD::SETCC: return "X86ISD::SETCC";
4176 case X86ISD::CMOV: return "X86ISD::CMOV";
4177 case X86ISD::BRCOND: return "X86ISD::BRCOND";
4178 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
4179 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
4180 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
4181 case X86ISD::LOAD_PACK: return "X86ISD::LOAD_PACK";
4182 case X86ISD::LOAD_UA: return "X86ISD::LOAD_UA";
4183 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
4184 case X86ISD::Wrapper: return "X86ISD::Wrapper";
4185 case X86ISD::S2VEC: return "X86ISD::S2VEC";
4186 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
4187 case X86ISD::PINSRW: return "X86ISD::PINSRW";
4188 case X86ISD::FMAX: return "X86ISD::FMAX";
4189 case X86ISD::FMIN: return "X86ISD::FMIN";
4193 /// isLegalAddressImmediate - Return true if the integer value or
4194 /// GlobalValue can be used as the offset of the target addressing mode.
4195 bool X86TargetLowering::isLegalAddressImmediate(int64_t V) const {
4196 // X86 allows a sign-extended 32-bit immediate field.
4197 return (V > -(1LL << 32) && V < (1LL << 32)-1);
4200 bool X86TargetLowering::isLegalAddressImmediate(GlobalValue *GV) const {
4201 // In 64-bit mode, GV is 64-bit so it won't fit in the 32-bit displacement
4202 // field unless we are in small code model.
4203 if (Subtarget->is64Bit() &&
4204 getTargetMachine().getCodeModel() != CodeModel::Small)
4207 return (!Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false));
4210 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4211 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4212 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4213 /// are assumed to be legal.
4215 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
4216 // Only do shuffles on 128-bit vector types for now.
4217 if (MVT::getSizeInBits(VT) == 64) return false;
4218 return (Mask.Val->getNumOperands() <= 4 ||
4219 isSplatMask(Mask.Val) ||
4220 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
4221 X86::isUNPCKLMask(Mask.Val) ||
4222 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
4223 X86::isUNPCKHMask(Mask.Val));
4226 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
4228 SelectionDAG &DAG) const {
4229 unsigned NumElts = BVOps.size();
4230 // Only do shuffles on 128-bit vector types for now.
4231 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
4232 if (NumElts == 2) return true;
4234 return (isMOVLMask(&BVOps[0], 4) ||
4235 isCommutedMOVL(&BVOps[0], 4, true) ||
4236 isSHUFPMask(&BVOps[0], 4) ||
4237 isCommutedSHUFP(&BVOps[0], 4));
4242 //===----------------------------------------------------------------------===//
4243 // X86 Scheduler Hooks
4244 //===----------------------------------------------------------------------===//
4247 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
4248 MachineBasicBlock *BB) {
4249 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4250 switch (MI->getOpcode()) {
4251 default: assert(false && "Unexpected instr type to insert");
4252 case X86::CMOV_FR32:
4253 case X86::CMOV_FR64:
4254 case X86::CMOV_V4F32:
4255 case X86::CMOV_V2F64:
4256 case X86::CMOV_V2I64: {
4257 // To "insert" a SELECT_CC instruction, we actually have to insert the
4258 // diamond control-flow pattern. The incoming instruction knows the
4259 // destination vreg to set, the condition code register to branch on, the
4260 // true/false values to select between, and a branch opcode to use.
4261 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4262 ilist<MachineBasicBlock>::iterator It = BB;
4268 // cmpTY ccX, r1, r2
4270 // fallthrough --> copy0MBB
4271 MachineBasicBlock *thisMBB = BB;
4272 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
4273 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
4275 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
4276 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB);
4277 MachineFunction *F = BB->getParent();
4278 F->getBasicBlockList().insert(It, copy0MBB);
4279 F->getBasicBlockList().insert(It, sinkMBB);
4280 // Update machine-CFG edges by first adding all successors of the current
4281 // block to the new block which will contain the Phi node for the select.
4282 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
4283 e = BB->succ_end(); i != e; ++i)
4284 sinkMBB->addSuccessor(*i);
4285 // Next, remove all successors of the current block, and add the true
4286 // and fallthrough blocks as its successors.
4287 while(!BB->succ_empty())
4288 BB->removeSuccessor(BB->succ_begin());
4289 BB->addSuccessor(copy0MBB);
4290 BB->addSuccessor(sinkMBB);
4293 // %FalseValue = ...
4294 // # fallthrough to sinkMBB
4297 // Update machine-CFG edges
4298 BB->addSuccessor(sinkMBB);
4301 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4304 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg())
4305 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
4306 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4308 delete MI; // The pseudo instruction is gone now.
4312 case X86::FP_TO_INT16_IN_MEM:
4313 case X86::FP_TO_INT32_IN_MEM:
4314 case X86::FP_TO_INT64_IN_MEM: {
4315 // Change the floating point control register to use "round towards zero"
4316 // mode when truncating to an integer value.
4317 MachineFunction *F = BB->getParent();
4318 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
4319 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx);
4321 // Load the old value of the high byte of the control word...
4323 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
4324 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx);
4326 // Set the high part to be round to zero...
4327 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx)
4330 // Reload the modified control word now...
4331 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4333 // Restore the memory image of control word to original value
4334 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx)
4337 // Get the X86 opcode to use.
4339 switch (MI->getOpcode()) {
4340 default: assert(0 && "illegal opcode!");
4341 case X86::FP_TO_INT16_IN_MEM: Opc = X86::FpIST16m; break;
4342 case X86::FP_TO_INT32_IN_MEM: Opc = X86::FpIST32m; break;
4343 case X86::FP_TO_INT64_IN_MEM: Opc = X86::FpIST64m; break;
4347 MachineOperand &Op = MI->getOperand(0);
4348 if (Op.isRegister()) {
4349 AM.BaseType = X86AddressMode::RegBase;
4350 AM.Base.Reg = Op.getReg();
4352 AM.BaseType = X86AddressMode::FrameIndexBase;
4353 AM.Base.FrameIndex = Op.getFrameIndex();
4355 Op = MI->getOperand(1);
4356 if (Op.isImmediate())
4357 AM.Scale = Op.getImm();
4358 Op = MI->getOperand(2);
4359 if (Op.isImmediate())
4360 AM.IndexReg = Op.getImm();
4361 Op = MI->getOperand(3);
4362 if (Op.isGlobalAddress()) {
4363 AM.GV = Op.getGlobal();
4365 AM.Disp = Op.getImm();
4367 addFullAddress(BuildMI(BB, TII->get(Opc)), AM)
4368 .addReg(MI->getOperand(4).getReg());
4370 // Reload the original control word now.
4371 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx);
4373 delete MI; // The pseudo instruction is gone now.
4379 //===----------------------------------------------------------------------===//
4380 // X86 Optimization Hooks
4381 //===----------------------------------------------------------------------===//
4383 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
4385 uint64_t &KnownZero,
4387 unsigned Depth) const {
4388 unsigned Opc = Op.getOpcode();
4389 assert((Opc >= ISD::BUILTIN_OP_END ||
4390 Opc == ISD::INTRINSIC_WO_CHAIN ||
4391 Opc == ISD::INTRINSIC_W_CHAIN ||
4392 Opc == ISD::INTRINSIC_VOID) &&
4393 "Should use MaskedValueIsZero if you don't know whether Op"
4394 " is a target node!");
4396 KnownZero = KnownOne = 0; // Don't know anything.
4400 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
4405 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4406 /// element of the result of the vector shuffle.
4407 static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
4408 MVT::ValueType VT = N->getValueType(0);
4409 SDOperand PermMask = N->getOperand(2);
4410 unsigned NumElems = PermMask.getNumOperands();
4411 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
4413 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4415 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4416 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
4417 SDOperand Idx = PermMask.getOperand(i);
4418 if (Idx.getOpcode() == ISD::UNDEF)
4419 return DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT));
4420 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
4425 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
4426 /// node is a GlobalAddress + an offset.
4427 static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) {
4428 unsigned Opc = N->getOpcode();
4429 if (Opc == X86ISD::Wrapper) {
4430 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) {
4431 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
4434 } else if (Opc == ISD::ADD) {
4435 SDOperand N1 = N->getOperand(0);
4436 SDOperand N2 = N->getOperand(1);
4437 if (isGAPlusOffset(N1.Val, GA, Offset)) {
4438 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
4440 Offset += V->getSignExtended();
4443 } else if (isGAPlusOffset(N2.Val, GA, Offset)) {
4444 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
4446 Offset += V->getSignExtended();
4454 /// isConsecutiveLoad - Returns true if N is loading from an address of Base
4456 static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size,
4457 MachineFrameInfo *MFI) {
4458 if (N->getOperand(0).Val != Base->getOperand(0).Val)
4461 SDOperand Loc = N->getOperand(1);
4462 SDOperand BaseLoc = Base->getOperand(1);
4463 if (Loc.getOpcode() == ISD::FrameIndex) {
4464 if (BaseLoc.getOpcode() != ISD::FrameIndex)
4466 int FI = dyn_cast<FrameIndexSDNode>(Loc)->getIndex();
4467 int BFI = dyn_cast<FrameIndexSDNode>(BaseLoc)->getIndex();
4468 int FS = MFI->getObjectSize(FI);
4469 int BFS = MFI->getObjectSize(BFI);
4470 if (FS != BFS || FS != Size) return false;
4471 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size);
4473 GlobalValue *GV1 = NULL;
4474 GlobalValue *GV2 = NULL;
4475 int64_t Offset1 = 0;
4476 int64_t Offset2 = 0;
4477 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1);
4478 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2);
4479 if (isGA1 && isGA2 && GV1 == GV2)
4480 return Offset1 == (Offset2 + Dist*Size);
4486 static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI,
4487 const X86Subtarget *Subtarget) {
4490 if (isGAPlusOffset(Base, GV, Offset))
4491 return (GV->getAlignment() >= 16 && (Offset % 16) == 0);
4493 assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!");
4494 int BFI = dyn_cast<FrameIndexSDNode>(Base)->getIndex();
4496 // Fixed objects do not specify alignment, however the offsets are known.
4497 return ((Subtarget->getStackAlignment() % 16) == 0 &&
4498 (MFI->getObjectOffset(BFI) % 16) == 0);
4500 return MFI->getObjectAlignment(BFI) >= 16;
4506 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
4507 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
4508 /// if the load addresses are consecutive, non-overlapping, and in the right
4510 static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
4511 const X86Subtarget *Subtarget) {
4512 MachineFunction &MF = DAG.getMachineFunction();
4513 MachineFrameInfo *MFI = MF.getFrameInfo();
4514 MVT::ValueType VT = N->getValueType(0);
4515 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
4516 SDOperand PermMask = N->getOperand(2);
4517 int NumElems = (int)PermMask.getNumOperands();
4518 SDNode *Base = NULL;
4519 for (int i = 0; i < NumElems; ++i) {
4520 SDOperand Idx = PermMask.getOperand(i);
4521 if (Idx.getOpcode() == ISD::UNDEF) {
4522 if (!Base) return SDOperand();
4525 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG);
4526 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val))
4530 else if (!isConsecutiveLoad(Arg.Val, Base,
4531 i, MVT::getSizeInBits(EVT)/8,MFI))
4536 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget);
4538 LoadSDNode *LD = cast<LoadSDNode>(Base);
4539 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(),
4540 LD->getSrcValueOffset());
4542 // Just use movups, it's shorter.
4543 SDVTList Tys = DAG.getVTList(MVT::v4f32, MVT::Other);
4544 SmallVector<SDOperand, 3> Ops;
4545 Ops.push_back(Base->getOperand(0));
4546 Ops.push_back(Base->getOperand(1));
4547 Ops.push_back(Base->getOperand(2));
4548 return DAG.getNode(ISD::BIT_CONVERT, VT,
4549 DAG.getNode(X86ISD::LOAD_UA, Tys, &Ops[0], Ops.size()));
4553 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
4554 static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
4555 const X86Subtarget *Subtarget) {
4556 SDOperand Cond = N->getOperand(0);
4558 // If we have SSE[12] support, try to form min/max nodes.
4559 if (Subtarget->hasSSE2() &&
4560 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) {
4561 if (Cond.getOpcode() == ISD::SETCC) {
4562 // Get the LHS/RHS of the select.
4563 SDOperand LHS = N->getOperand(1);
4564 SDOperand RHS = N->getOperand(2);
4565 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
4567 unsigned Opcode = 0;
4568 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
4571 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
4574 if (!UnsafeFPMath) break;
4576 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
4578 Opcode = X86ISD::FMIN;
4581 case ISD::SETOGT: // (X > Y) ? X : Y -> max
4584 if (!UnsafeFPMath) break;
4586 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
4588 Opcode = X86ISD::FMAX;
4591 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
4594 case ISD::SETOGT: // (X > Y) ? Y : X -> min
4597 if (!UnsafeFPMath) break;
4599 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
4601 Opcode = X86ISD::FMIN;
4604 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
4607 if (!UnsafeFPMath) break;
4609 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
4611 Opcode = X86ISD::FMAX;
4617 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS);
4626 SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N,
4627 DAGCombinerInfo &DCI) const {
4628 SelectionDAG &DAG = DCI.DAG;
4629 switch (N->getOpcode()) {
4631 case ISD::VECTOR_SHUFFLE:
4632 return PerformShuffleCombine(N, DAG, Subtarget);
4634 return PerformSELECTCombine(N, DAG, Subtarget);
4640 //===----------------------------------------------------------------------===//
4641 // X86 Inline Assembly Support
4642 //===----------------------------------------------------------------------===//
4644 /// getConstraintType - Given a constraint letter, return the type of
4645 /// constraint it is for this target.
4646 X86TargetLowering::ConstraintType
4647 X86TargetLowering::getConstraintType(char ConstraintLetter) const {
4648 switch (ConstraintLetter) {
4657 return C_RegisterClass;
4658 default: return TargetLowering::getConstraintType(ConstraintLetter);
4662 /// isOperandValidForConstraint - Return the specified operand (possibly
4663 /// modified) if the specified SDOperand is valid for the specified target
4664 /// constraint letter, otherwise return null.
4665 SDOperand X86TargetLowering::
4666 isOperandValidForConstraint(SDOperand Op, char Constraint, SelectionDAG &DAG) {
4667 switch (Constraint) {
4670 // Literal immediates are always ok.
4671 if (isa<ConstantSDNode>(Op)) return Op;
4673 // If we are in non-pic codegen mode, we allow the address of a global to
4674 // be used with 'i'.
4675 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op)) {
4676 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
4677 return SDOperand(0, 0);
4679 if (GA->getOpcode() != ISD::TargetGlobalAddress)
4680 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
4685 // Otherwise, not valid for this mode.
4686 return SDOperand(0, 0);
4688 return TargetLowering::isOperandValidForConstraint(Op, Constraint, DAG);
4692 std::vector<unsigned> X86TargetLowering::
4693 getRegClassForInlineAsmConstraint(const std::string &Constraint,
4694 MVT::ValueType VT) const {
4695 if (Constraint.size() == 1) {
4696 // FIXME: not handling fp-stack yet!
4697 // FIXME: not handling MMX registers yet ('y' constraint).
4698 switch (Constraint[0]) { // GCC X86 Constraint Letters
4699 default: break; // Unknown constraint letter
4700 case 'A': // EAX/EDX
4701 if (VT == MVT::i32 || VT == MVT::i64)
4702 return make_vector<unsigned>(X86::EAX, X86::EDX, 0);
4704 case 'r': // GENERAL_REGS
4705 case 'R': // LEGACY_REGS
4706 if (VT == MVT::i64 && Subtarget->is64Bit())
4707 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
4708 X86::RSI, X86::RDI, X86::RBP, X86::RSP,
4709 X86::R8, X86::R9, X86::R10, X86::R11,
4710 X86::R12, X86::R13, X86::R14, X86::R15, 0);
4712 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
4713 X86::ESI, X86::EDI, X86::EBP, X86::ESP, 0);
4714 else if (VT == MVT::i16)
4715 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
4716 X86::SI, X86::DI, X86::BP, X86::SP, 0);
4717 else if (VT == MVT::i8)
4718 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
4720 case 'l': // INDEX_REGS
4722 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
4723 X86::ESI, X86::EDI, X86::EBP, 0);
4724 else if (VT == MVT::i16)
4725 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
4726 X86::SI, X86::DI, X86::BP, 0);
4727 else if (VT == MVT::i8)
4728 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
4730 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
4733 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
4734 else if (VT == MVT::i16)
4735 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
4736 else if (VT == MVT::i8)
4737 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
4739 case 'x': // SSE_REGS if SSE1 allowed
4740 if (Subtarget->hasSSE1())
4741 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
4742 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
4744 return std::vector<unsigned>();
4745 case 'Y': // SSE_REGS if SSE2 allowed
4746 if (Subtarget->hasSSE2())
4747 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
4748 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
4750 return std::vector<unsigned>();
4754 return std::vector<unsigned>();
4757 std::pair<unsigned, const TargetRegisterClass*>
4758 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4759 MVT::ValueType VT) const {
4760 // Use the default implementation in TargetLowering to convert the register
4761 // constraint into a member of a register class.
4762 std::pair<unsigned, const TargetRegisterClass*> Res;
4763 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4765 // Not found as a standard register?
4766 if (Res.second == 0) {
4767 // GCC calls "st(0)" just plain "st".
4768 if (StringsEqualNoCase("{st}", Constraint)) {
4769 Res.first = X86::ST0;
4770 Res.second = X86::RSTRegisterClass;
4776 // Otherwise, check to see if this is a register class of the wrong value
4777 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
4778 // turn into {ax},{dx}.
4779 if (Res.second->hasType(VT))
4780 return Res; // Correct type already, nothing to do.
4782 // All of the single-register GCC register classes map their values onto
4783 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
4784 // really want an 8-bit or 32-bit register, map to the appropriate register
4785 // class and return the appropriate register.
4786 if (Res.second != X86::GR16RegisterClass)
4789 if (VT == MVT::i8) {
4790 unsigned DestReg = 0;
4791 switch (Res.first) {
4793 case X86::AX: DestReg = X86::AL; break;
4794 case X86::DX: DestReg = X86::DL; break;
4795 case X86::CX: DestReg = X86::CL; break;
4796 case X86::BX: DestReg = X86::BL; break;
4799 Res.first = DestReg;
4800 Res.second = Res.second = X86::GR8RegisterClass;
4802 } else if (VT == MVT::i32) {
4803 unsigned DestReg = 0;
4804 switch (Res.first) {
4806 case X86::AX: DestReg = X86::EAX; break;
4807 case X86::DX: DestReg = X86::EDX; break;
4808 case X86::CX: DestReg = X86::ECX; break;
4809 case X86::BX: DestReg = X86::EBX; break;
4810 case X86::SI: DestReg = X86::ESI; break;
4811 case X86::DI: DestReg = X86::EDI; break;
4812 case X86::BP: DestReg = X86::EBP; break;
4813 case X86::SP: DestReg = X86::ESP; break;
4816 Res.first = DestReg;
4817 Res.second = Res.second = X86::GR32RegisterClass;
4819 } else if (VT == MVT::i64) {
4820 unsigned DestReg = 0;
4821 switch (Res.first) {
4823 case X86::AX: DestReg = X86::RAX; break;
4824 case X86::DX: DestReg = X86::RDX; break;
4825 case X86::CX: DestReg = X86::RCX; break;
4826 case X86::BX: DestReg = X86::RBX; break;
4827 case X86::SI: DestReg = X86::RSI; break;
4828 case X86::DI: DestReg = X86::RDI; break;
4829 case X86::BP: DestReg = X86::RBP; break;
4830 case X86::SP: DestReg = X86::RSP; break;
4833 Res.first = DestReg;
4834 Res.second = Res.second = X86::GR64RegisterClass;