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 "X86TargetMachine.h"
19 #include "llvm/CallingConv.h"
20 #include "llvm/Constants.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/Function.h"
23 #include "llvm/Intrinsics.h"
24 #include "llvm/ADT/VectorExtras.h"
25 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
26 #include "llvm/CodeGen/MachineFrameInfo.h"
27 #include "llvm/CodeGen/MachineFunction.h"
28 #include "llvm/CodeGen/MachineInstrBuilder.h"
29 #include "llvm/CodeGen/SelectionDAG.h"
30 #include "llvm/CodeGen/SSARegMap.h"
31 #include "llvm/Support/MathExtras.h"
32 #include "llvm/Target/TargetOptions.h"
36 #include "llvm/Support/CommandLine.h"
37 static cl::opt<bool> EnableFastCC("enable-x86-fastcc", cl::Hidden,
38 cl::desc("Enable fastcc on X86"));
40 X86TargetLowering::X86TargetLowering(TargetMachine &TM)
41 : TargetLowering(TM) {
42 Subtarget = &TM.getSubtarget<X86Subtarget>();
43 X86ScalarSSE = Subtarget->hasSSE2();
45 // Set up the TargetLowering object.
47 // X86 is weird, it always uses i8 for shift amounts and setcc results.
48 setShiftAmountType(MVT::i8);
49 setSetCCResultType(MVT::i8);
50 setSetCCResultContents(ZeroOrOneSetCCResult);
51 setSchedulingPreference(SchedulingForRegPressure);
52 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0
53 setStackPointerRegisterToSaveRestore(X86::ESP);
55 if (!Subtarget->isTargetDarwin())
56 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
57 setUseUnderscoreSetJmpLongJmp(true);
59 // Add legal addressing mode scale values.
60 addLegalAddressScale(8);
61 addLegalAddressScale(4);
62 addLegalAddressScale(2);
63 // Enter the ones which require both scale + index last. These are more
65 addLegalAddressScale(9);
66 addLegalAddressScale(5);
67 addLegalAddressScale(3);
69 // Set up the register classes.
70 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
71 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
72 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
74 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
76 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
77 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
78 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
81 // No SSE i64 SINT_TO_FP, so expand i32 UINT_TO_FP instead.
82 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand);
84 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
86 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
88 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
89 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
90 // SSE has no i16 to fp conversion, only i32
92 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
94 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
95 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
98 // We can handle SINT_TO_FP and FP_TO_SINT from/to i64 even though i64
100 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
101 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
103 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
105 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
106 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
109 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
111 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
112 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
115 // Handle FP_TO_UINT by promoting the destination to a larger signed
117 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
118 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
119 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
121 if (X86ScalarSSE && !Subtarget->hasSSE3())
122 // Expand FP_TO_UINT into a select.
123 // FIXME: We would like to use a Custom expander here eventually to do
124 // the optimal thing for SSE vs. the default expansion in the legalizer.
125 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
127 // With SSE3 we can use fisttpll to convert to a signed i64.
128 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
130 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
131 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
133 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
134 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
135 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
136 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
137 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Expand);
138 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Expand);
139 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
140 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
141 setOperationAction(ISD::SEXTLOAD , MVT::i1 , Expand);
142 setOperationAction(ISD::FREM , MVT::f64 , Expand);
143 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
144 setOperationAction(ISD::CTTZ , MVT::i8 , Expand);
145 setOperationAction(ISD::CTLZ , MVT::i8 , Expand);
146 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
147 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
148 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
149 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
150 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
151 setOperationAction(ISD::CTLZ , MVT::i32 , Expand);
152 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
153 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
155 // These should be promoted to a larger select which is supported.
156 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
157 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
159 // X86 wants to expand cmov itself.
160 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
161 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
162 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
163 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
164 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
165 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
166 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
167 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
168 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
169 // X86 ret instruction may pop stack.
170 setOperationAction(ISD::RET , MVT::Other, Custom);
172 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
173 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
174 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
175 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
176 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
177 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
178 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
179 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
180 // X86 wants to expand memset / memcpy itself.
181 setOperationAction(ISD::MEMSET , MVT::Other, Custom);
182 setOperationAction(ISD::MEMCPY , MVT::Other, Custom);
184 // We don't have line number support yet.
185 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
186 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
187 // FIXME - use subtarget debug flags
188 if (!Subtarget->isTargetDarwin())
189 setOperationAction(ISD::DEBUG_LABEL, MVT::Other, Expand);
191 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
192 setOperationAction(ISD::VASTART , MVT::Other, Custom);
194 // Use the default implementation.
195 setOperationAction(ISD::VAARG , MVT::Other, Expand);
196 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
197 setOperationAction(ISD::VAEND , MVT::Other, Expand);
198 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
199 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
200 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
202 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
203 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
206 // Set up the FP register classes.
207 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
208 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
210 // Use ANDPD to simulate FABS.
211 setOperationAction(ISD::FABS , MVT::f64, Custom);
212 setOperationAction(ISD::FABS , MVT::f32, Custom);
214 // Use XORP to simulate FNEG.
215 setOperationAction(ISD::FNEG , MVT::f64, Custom);
216 setOperationAction(ISD::FNEG , MVT::f32, Custom);
218 // We don't support sin/cos/fmod
219 setOperationAction(ISD::FSIN , MVT::f64, Expand);
220 setOperationAction(ISD::FCOS , MVT::f64, Expand);
221 setOperationAction(ISD::FREM , MVT::f64, Expand);
222 setOperationAction(ISD::FSIN , MVT::f32, Expand);
223 setOperationAction(ISD::FCOS , MVT::f32, Expand);
224 setOperationAction(ISD::FREM , MVT::f32, Expand);
226 // Expand FP immediates into loads from the stack, except for the special
228 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
229 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
230 addLegalFPImmediate(+0.0); // xorps / xorpd
232 // Set up the FP register classes.
233 addRegisterClass(MVT::f64, X86::RFPRegisterClass);
235 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
238 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
239 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
242 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
243 addLegalFPImmediate(+0.0); // FLD0
244 addLegalFPImmediate(+1.0); // FLD1
245 addLegalFPImmediate(-0.0); // FLD0/FCHS
246 addLegalFPImmediate(-1.0); // FLD1/FCHS
249 // First set operation action for all vector types to expand. Then we
250 // will selectively turn on ones that can be effectively codegen'd.
251 for (unsigned VT = (unsigned)MVT::Vector + 1;
252 VT != (unsigned)MVT::LAST_VALUETYPE; VT++) {
253 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
254 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
255 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
256 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand);
257 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Expand);
258 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
259 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
262 if (Subtarget->hasMMX()) {
263 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
264 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
265 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
267 // FIXME: add MMX packed arithmetics
268 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Expand);
269 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Expand);
270 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Expand);
273 if (Subtarget->hasSSE1()) {
274 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
276 setOperationAction(ISD::AND, MVT::v4f32, Legal);
277 setOperationAction(ISD::OR, MVT::v4f32, Legal);
278 setOperationAction(ISD::XOR, MVT::v4f32, Legal);
279 setOperationAction(ISD::ADD, MVT::v4f32, Legal);
280 setOperationAction(ISD::SUB, MVT::v4f32, Legal);
281 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
282 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
283 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
284 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
285 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
286 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
289 if (Subtarget->hasSSE2()) {
290 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
291 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
292 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
293 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
294 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
296 setOperationAction(ISD::ADD, MVT::v2f64, Legal);
297 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
298 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
299 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
300 setOperationAction(ISD::SUB, MVT::v2f64, Legal);
301 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
302 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
303 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
304 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
305 setOperationAction(ISD::MUL, MVT::v2f64, Legal);
307 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
308 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
309 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
310 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
311 // Implement v4f32 insert_vector_elt in terms of SSE2 v8i16 ones.
312 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
314 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
315 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
316 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
317 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
318 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
320 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
321 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
322 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
323 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
324 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
325 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
327 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
328 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) {
329 setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote);
330 AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64);
331 setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote);
332 AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64);
333 setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote);
334 AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64);
335 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote);
336 AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64);
337 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
338 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64);
341 // Custom lower v2i64 and v2f64 selects.
342 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
343 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
344 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
345 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
348 // We want to custom lower some of our intrinsics.
349 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
351 computeRegisterProperties();
353 // FIXME: These should be based on subtarget info. Plus, the values should
354 // be smaller when we are in optimizing for size mode.
355 maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
356 maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
357 maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
358 allowUnalignedMemoryAccesses = true; // x86 supports it!
361 //===----------------------------------------------------------------------===//
362 // C Calling Convention implementation
363 //===----------------------------------------------------------------------===//
365 /// AddLiveIn - This helper function adds the specified physical register to the
366 /// MachineFunction as a live in value. It also creates a corresponding virtual
368 static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
369 TargetRegisterClass *RC) {
370 assert(RC->contains(PReg) && "Not the correct regclass!");
371 unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC);
372 MF.addLiveIn(PReg, VReg);
376 /// HowToPassCCCArgument - Returns how an formal argument of the specified type
377 /// should be passed. If it is through stack, returns the size of the stack
378 /// frame; if it is through XMM register, returns the number of XMM registers
381 HowToPassCCCArgument(MVT::ValueType ObjectVT, unsigned NumXMMRegs,
382 unsigned &ObjSize, unsigned &ObjXMMRegs) {
384 default: assert(0 && "Unhandled argument type!");
386 case MVT::i8: ObjSize = 1; break;
387 case MVT::i16: ObjSize = 2; break;
388 case MVT::i32: ObjSize = 4; break;
389 case MVT::i64: ObjSize = 8; break;
390 case MVT::f32: ObjSize = 4; break;
391 case MVT::f64: ObjSize = 8; break;
406 /// getFormalArgObjects - Returns itself if Op is a FORMAL_ARGUMENTS, otherwise
407 /// returns the FORMAL_ARGUMENTS node(s) that made up parts of the node.
408 static std::vector<SDOperand> getFormalArgObjects(SDOperand Op) {
409 unsigned Opc = Op.getOpcode();
410 std::vector<SDOperand> Objs;
411 if (Opc == ISD::TRUNCATE) {
412 Op = Op.getOperand(0);
413 assert(Op.getOpcode() == ISD::AssertSext ||
414 Op.getOpcode() == ISD::AssertZext);
415 Objs.push_back(Op.getOperand(0));
416 } else if (Opc == ISD::FP_ROUND || Opc == ISD::VBIT_CONVERT) {
417 Objs.push_back(Op.getOperand(0));
418 } else if (Opc == ISD::BUILD_PAIR) {
419 Objs.push_back(Op.getOperand(0));
420 Objs.push_back(Op.getOperand(1));
427 SDOperand X86TargetLowering::LowerCCCArguments(SDOperand Op, SelectionDAG &DAG) {
428 unsigned NumArgs = Op.Val->getNumValues() - 1;
429 MachineFunction &MF = DAG.getMachineFunction();
430 MachineFrameInfo *MFI = MF.getFrameInfo();
431 SDOperand Root = Op.getOperand(0);
432 std::vector<SDOperand> ArgValues;
434 // Add DAG nodes to load the arguments... On entry to a function on the X86,
435 // the stack frame looks like this:
437 // [ESP] -- return address
438 // [ESP + 4] -- first argument (leftmost lexically)
439 // [ESP + 8] -- second argument, if first argument is four bytes in size
442 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
443 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
444 unsigned XMMArgRegs[] = { X86::XMM0, X86::XMM1, X86::XMM2 };
445 for (unsigned i = 0; i < NumArgs; ++i) {
446 MVT::ValueType ObjectVT = Op.getValue(i).getValueType();
447 unsigned ArgIncrement = 4;
448 unsigned ObjSize = 0;
449 unsigned ObjXMMRegs = 0;
450 HowToPassCCCArgument(ObjectVT, NumXMMRegs, ObjSize, ObjXMMRegs);
452 ArgIncrement = ObjSize;
456 // Passed in a XMM register.
457 unsigned Reg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs],
458 X86::VR128RegisterClass);
459 ArgValue= DAG.getCopyFromReg(Root, Reg, ObjectVT);
460 ArgValues.push_back(ArgValue);
461 NumXMMRegs += ObjXMMRegs;
463 // Create the frame index object for this incoming parameter...
464 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
465 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
466 ArgValue = DAG.getLoad(Op.Val->getValueType(i), Root, FIN,
467 DAG.getSrcValue(NULL));
468 ArgValues.push_back(ArgValue);
469 ArgOffset += ArgIncrement; // Move on to the next argument...
473 ArgValues.push_back(Root);
475 // If the function takes variable number of arguments, make a frame index for
476 // the start of the first vararg value... for expansion of llvm.va_start.
477 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
479 VarArgsFrameIndex = MFI->CreateFixedObject(1, ArgOffset);
480 ReturnAddrIndex = 0; // No return address slot generated yet.
481 BytesToPopOnReturn = 0; // Callee pops nothing.
482 BytesCallerReserves = ArgOffset;
484 // If this is a struct return on Darwin/X86, the callee pops the hidden struct
486 if (MF.getFunction()->getCallingConv() == CallingConv::CSRet &&
487 Subtarget->isTargetDarwin())
488 BytesToPopOnReturn = 4;
490 // Return the new list of results.
491 std::vector<MVT::ValueType> RetVTs(Op.Val->value_begin(),
492 Op.Val->value_end());
493 return DAG.getNode(ISD::MERGE_VALUES, RetVTs, ArgValues);
497 SDOperand X86TargetLowering::LowerCCCCallTo(SDOperand Op, SelectionDAG &DAG) {
498 SDOperand Chain = Op.getOperand(0);
499 unsigned CallingConv= cast<ConstantSDNode>(Op.getOperand(1))->getValue();
500 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
501 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
502 SDOperand Callee = Op.getOperand(4);
503 MVT::ValueType RetVT= Op.Val->getValueType(0);
504 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
506 // Keep track of the number of XMM regs passed so far.
507 unsigned NumXMMRegs = 0;
508 static const unsigned XMMArgRegs[] = {
509 X86::XMM0, X86::XMM1, X86::XMM2
512 // Count how many bytes are to be pushed on the stack.
513 unsigned NumBytes = 0;
514 for (unsigned i = 0; i != NumOps; ++i) {
515 SDOperand Arg = Op.getOperand(5+2*i);
517 switch (Arg.getValueType()) {
518 default: assert(0 && "Unexpected ValueType for argument!");
544 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
546 // Arguments go on the stack in reverse order, as specified by the ABI.
547 unsigned ArgOffset = 0;
549 std::vector<std::pair<unsigned, SDOperand> > RegsToPass;
550 std::vector<SDOperand> MemOpChains;
551 SDOperand StackPtr = DAG.getRegister(X86::ESP, getPointerTy());
552 for (unsigned i = 0; i != NumOps; ++i) {
553 SDOperand Arg = Op.getOperand(5+2*i);
555 switch (Arg.getValueType()) {
556 default: assert(0 && "Unexpected ValueType for argument!");
559 // Promote the integer to 32 bits. If the input type is signed use a
560 // sign extend, otherwise use a zero extend.
562 dyn_cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue() ?
563 ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
564 Arg = DAG.getNode(ExtOp, MVT::i32, Arg);
569 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
570 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
571 MemOpChains.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
572 Arg, PtrOff, DAG.getSrcValue(NULL)));
578 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
579 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
580 MemOpChains.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
581 Arg, PtrOff, DAG.getSrcValue(NULL)));
591 if (NumXMMRegs < 3) {
592 RegsToPass.push_back(std::make_pair(XMMArgRegs[NumXMMRegs], Arg));
595 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
596 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
597 MemOpChains.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
598 Arg, PtrOff, DAG.getSrcValue(NULL)));
605 if (!MemOpChains.empty())
606 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, MemOpChains);
608 // Build a sequence of copy-to-reg nodes chained together with token chain
609 // and flag operands which copy the outgoing args into registers.
611 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
612 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
614 InFlag = Chain.getValue(1);
617 // If the callee is a GlobalAddress node (quite common, every direct call is)
618 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
619 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
620 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
621 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
622 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
624 std::vector<MVT::ValueType> NodeTys;
625 NodeTys.push_back(MVT::Other); // Returns a chain
626 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
627 std::vector<SDOperand> Ops;
628 Ops.push_back(Chain);
629 Ops.push_back(Callee);
631 Ops.push_back(InFlag);
633 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
635 InFlag = Chain.getValue(1);
637 // Create the CALLSEQ_END node.
638 unsigned NumBytesForCalleeToPush = 0;
640 // If this is is a call to a struct-return function on Darwin/X86, the callee
641 // pops the hidden struct pointer, so we have to push it back.
642 if (CallingConv == CallingConv::CSRet && Subtarget->isTargetDarwin())
643 NumBytesForCalleeToPush = 4;
646 NodeTys.push_back(MVT::Other); // Returns a chain
647 if (RetVT != MVT::Other)
648 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
650 Ops.push_back(Chain);
651 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
652 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy()));
653 Ops.push_back(InFlag);
654 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, Ops);
655 if (RetVT != MVT::Other)
656 InFlag = Chain.getValue(1);
658 std::vector<SDOperand> ResultVals;
661 default: assert(0 && "Unknown value type to return!");
662 case MVT::Other: break;
664 Chain = DAG.getCopyFromReg(Chain, X86::AL, MVT::i8, InFlag).getValue(1);
665 ResultVals.push_back(Chain.getValue(0));
666 NodeTys.push_back(MVT::i8);
669 Chain = DAG.getCopyFromReg(Chain, X86::AX, MVT::i16, InFlag).getValue(1);
670 ResultVals.push_back(Chain.getValue(0));
671 NodeTys.push_back(MVT::i16);
674 if (Op.Val->getValueType(1) == MVT::i32) {
675 Chain = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag).getValue(1);
676 ResultVals.push_back(Chain.getValue(0));
677 Chain = DAG.getCopyFromReg(Chain, X86::EDX, MVT::i32,
678 Chain.getValue(2)).getValue(1);
679 ResultVals.push_back(Chain.getValue(0));
680 NodeTys.push_back(MVT::i32);
682 Chain = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag).getValue(1);
683 ResultVals.push_back(Chain.getValue(0));
685 NodeTys.push_back(MVT::i32);
694 Chain = DAG.getCopyFromReg(Chain, X86::XMM0, RetVT, InFlag).getValue(1);
695 ResultVals.push_back(Chain.getValue(0));
696 NodeTys.push_back(RetVT);
700 std::vector<MVT::ValueType> Tys;
701 Tys.push_back(MVT::f64);
702 Tys.push_back(MVT::Other);
703 Tys.push_back(MVT::Flag);
704 std::vector<SDOperand> Ops;
705 Ops.push_back(Chain);
706 Ops.push_back(InFlag);
707 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, Ops);
708 Chain = RetVal.getValue(1);
709 InFlag = RetVal.getValue(2);
711 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
712 // shouldn't be necessary except that RFP cannot be live across
713 // multiple blocks. When stackifier is fixed, they can be uncoupled.
714 MachineFunction &MF = DAG.getMachineFunction();
715 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
716 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
718 Tys.push_back(MVT::Other);
720 Ops.push_back(Chain);
721 Ops.push_back(RetVal);
722 Ops.push_back(StackSlot);
723 Ops.push_back(DAG.getValueType(RetVT));
724 Ops.push_back(InFlag);
725 Chain = DAG.getNode(X86ISD::FST, Tys, Ops);
726 RetVal = DAG.getLoad(RetVT, Chain, StackSlot,
727 DAG.getSrcValue(NULL));
728 Chain = RetVal.getValue(1);
731 if (RetVT == MVT::f32 && !X86ScalarSSE)
732 // FIXME: we would really like to remember that this FP_ROUND
733 // operation is okay to eliminate if we allow excess FP precision.
734 RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal);
735 ResultVals.push_back(RetVal);
736 NodeTys.push_back(RetVT);
741 // If the function returns void, just return the chain.
742 if (ResultVals.empty())
745 // Otherwise, merge everything together with a MERGE_VALUES node.
746 NodeTys.push_back(MVT::Other);
747 ResultVals.push_back(Chain);
748 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys, ResultVals);
749 return Res.getValue(Op.ResNo);
752 //===----------------------------------------------------------------------===//
753 // Fast Calling Convention implementation
754 //===----------------------------------------------------------------------===//
756 // The X86 'fast' calling convention passes up to two integer arguments in
757 // registers (an appropriate portion of EAX/EDX), passes arguments in C order,
758 // and requires that the callee pop its arguments off the stack (allowing proper
759 // tail calls), and has the same return value conventions as C calling convs.
761 // This calling convention always arranges for the callee pop value to be 8n+4
762 // bytes, which is needed for tail recursion elimination and stack alignment
765 // Note that this can be enhanced in the future to pass fp vals in registers
766 // (when we have a global fp allocator) and do other tricks.
769 // FASTCC_NUM_INT_ARGS_INREGS - This is the max number of integer arguments
770 // to pass in registers. 0 is none, 1 is is "use EAX", 2 is "use EAX and
771 // EDX". Anything more is illegal.
773 // FIXME: The linscan register allocator currently has problem with
774 // coalescing. At the time of this writing, whenever it decides to coalesce
775 // a physreg with a virtreg, this increases the size of the physreg's live
776 // range, and the live range cannot ever be reduced. This causes problems if
777 // too many physregs are coaleced with virtregs, which can cause the register
778 // allocator to wedge itself.
780 // This code triggers this problem more often if we pass args in registers,
781 // so disable it until this is fixed.
783 // NOTE: this isn't marked const, so that GCC doesn't emit annoying warnings
784 // about code being dead.
786 static unsigned FASTCC_NUM_INT_ARGS_INREGS = 0;
789 /// HowToPassFastCCArgument - Returns how an formal argument of the specified
790 /// type should be passed. If it is through stack, returns the size of the stack
791 /// frame; if it is through integer or XMM register, returns the number of
792 /// integer or XMM registers are needed.
794 HowToPassFastCCArgument(MVT::ValueType ObjectVT,
795 unsigned NumIntRegs, unsigned NumXMMRegs,
796 unsigned &ObjSize, unsigned &ObjIntRegs,
797 unsigned &ObjXMMRegs) {
802 default: assert(0 && "Unhandled argument type!");
805 if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS)
811 if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS)
817 if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS)
823 if (NumIntRegs+2 <= FASTCC_NUM_INT_ARGS_INREGS) {
825 } else if (NumIntRegs+1 <= FASTCC_NUM_INT_ARGS_INREGS) {
851 X86TargetLowering::LowerFastCCArguments(SDOperand Op, SelectionDAG &DAG) {
852 unsigned NumArgs = Op.Val->getNumValues()-1;
853 MachineFunction &MF = DAG.getMachineFunction();
854 MachineFrameInfo *MFI = MF.getFrameInfo();
855 SDOperand Root = Op.getOperand(0);
856 std::vector<SDOperand> ArgValues;
858 // Add DAG nodes to load the arguments... On entry to a function the stack
859 // frame looks like this:
861 // [ESP] -- return address
862 // [ESP + 4] -- first nonreg argument (leftmost lexically)
863 // [ESP + 8] -- second nonreg argument, if first argument is 4 bytes in size
865 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
867 // Keep track of the number of integer regs passed so far. This can be either
868 // 0 (neither EAX or EDX used), 1 (EAX is used) or 2 (EAX and EDX are both
870 unsigned NumIntRegs = 0;
871 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
873 static const unsigned XMMArgRegs[] = {
874 X86::XMM0, X86::XMM1, X86::XMM2
877 for (unsigned i = 0; i < NumArgs; ++i) {
878 MVT::ValueType ObjectVT = Op.getValue(i).getValueType();
879 unsigned ArgIncrement = 4;
880 unsigned ObjSize = 0;
881 unsigned ObjIntRegs = 0;
882 unsigned ObjXMMRegs = 0;
884 HowToPassFastCCArgument(ObjectVT, NumIntRegs, NumXMMRegs,
885 ObjSize, ObjIntRegs, ObjXMMRegs);
887 ArgIncrement = ObjSize;
891 if (ObjIntRegs || ObjXMMRegs) {
893 default: assert(0 && "Unhandled argument type!");
896 Reg = AddLiveIn(MF, NumIntRegs ? X86::DL : X86::AL,
897 X86::GR8RegisterClass);
898 ArgValue = DAG.getCopyFromReg(Root, Reg, MVT::i8);
901 Reg = AddLiveIn(MF, NumIntRegs ? X86::DX : X86::AX,
902 X86::GR16RegisterClass);
903 ArgValue = DAG.getCopyFromReg(Root, Reg, MVT::i16);
906 Reg = AddLiveIn(MF, NumIntRegs ? X86::EDX : X86::EAX,
907 X86::GR32RegisterClass);
908 ArgValue = DAG.getCopyFromReg(Root, Reg, MVT::i32);
911 Reg = AddLiveIn(MF, NumIntRegs ? X86::EDX : X86::EAX,
912 X86::GR32RegisterClass);
913 ArgValue = DAG.getCopyFromReg(Root, Reg, MVT::i32);
914 if (ObjIntRegs == 2) {
915 Reg = AddLiveIn(MF, X86::EDX, X86::GR32RegisterClass);
916 SDOperand ArgValue2 = DAG.getCopyFromReg(Root, Reg, MVT::i32);
917 ArgValue= DAG.getNode(ISD::BUILD_PAIR, MVT::i64, ArgValue, ArgValue2);
926 Reg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs], X86::VR128RegisterClass);
927 ArgValue = DAG.getCopyFromReg(Root, Reg, ObjectVT);
930 NumIntRegs += ObjIntRegs;
931 NumXMMRegs += ObjXMMRegs;
935 // Create the SelectionDAG nodes corresponding to a load from this
937 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
938 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy());
939 if (ObjectVT == MVT::i64 && ObjIntRegs) {
940 SDOperand ArgValue2 = DAG.getLoad(Op.Val->getValueType(i), Root, FIN,
941 DAG.getSrcValue(NULL));
942 ArgValue = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, ArgValue, ArgValue2);
944 ArgValue = DAG.getLoad(Op.Val->getValueType(i), Root, FIN,
945 DAG.getSrcValue(NULL));
946 ArgOffset += ArgIncrement; // Move on to the next argument.
949 ArgValues.push_back(ArgValue);
952 ArgValues.push_back(Root);
954 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
955 // arguments and the arguments after the retaddr has been pushed are aligned.
956 if ((ArgOffset & 7) == 0)
959 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
960 ReturnAddrIndex = 0; // No return address slot generated yet.
961 BytesToPopOnReturn = ArgOffset; // Callee pops all stack arguments.
962 BytesCallerReserves = 0;
964 // Finally, inform the code generator which regs we return values in.
965 switch (getValueType(MF.getFunction()->getReturnType())) {
966 default: assert(0 && "Unknown type!");
967 case MVT::isVoid: break;
972 MF.addLiveOut(X86::EAX);
975 MF.addLiveOut(X86::EAX);
976 MF.addLiveOut(X86::EDX);
980 MF.addLiveOut(X86::ST0);
983 const PackedType *PTy = cast<PackedType>(MF.getFunction()->getReturnType());
986 unsigned NumRegs = getPackedTypeBreakdown(PTy, EVT, LVT);
987 assert(NumRegs == 1 && "Unsupported type!");
988 MF.addLiveOut(X86::XMM0);
993 // Return the new list of results.
994 std::vector<MVT::ValueType> RetVTs(Op.Val->value_begin(),
995 Op.Val->value_end());
996 return DAG.getNode(ISD::MERGE_VALUES, RetVTs, ArgValues);
999 SDOperand X86TargetLowering::LowerFastCCCallTo(SDOperand Op, SelectionDAG &DAG) {
1000 SDOperand Chain = Op.getOperand(0);
1001 unsigned CallingConv= cast<ConstantSDNode>(Op.getOperand(1))->getValue();
1002 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1003 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1004 SDOperand Callee = Op.getOperand(4);
1005 MVT::ValueType RetVT= Op.Val->getValueType(0);
1006 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1008 // Count how many bytes are to be pushed on the stack.
1009 unsigned NumBytes = 0;
1011 // Keep track of the number of integer regs passed so far. This can be either
1012 // 0 (neither EAX or EDX used), 1 (EAX is used) or 2 (EAX and EDX are both
1014 unsigned NumIntRegs = 0;
1015 unsigned NumXMMRegs = 0; // XMM regs used for parameter passing.
1017 static const unsigned GPRArgRegs[][2] = {
1018 { X86::AL, X86::DL },
1019 { X86::AX, X86::DX },
1020 { X86::EAX, X86::EDX }
1022 static const unsigned XMMArgRegs[] = {
1023 X86::XMM0, X86::XMM1, X86::XMM2
1026 for (unsigned i = 0; i != NumOps; ++i) {
1027 SDOperand Arg = Op.getOperand(5+2*i);
1029 switch (Arg.getValueType()) {
1030 default: assert(0 && "Unknown value type!");
1034 if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS) {
1059 // Make sure the instruction takes 8n+4 bytes to make sure the start of the
1060 // arguments and the arguments after the retaddr has been pushed are aligned.
1061 if ((NumBytes & 7) == 0)
1064 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));
1066 // Arguments go on the stack in reverse order, as specified by the ABI.
1067 unsigned ArgOffset = 0;
1069 std::vector<std::pair<unsigned, SDOperand> > RegsToPass;
1070 std::vector<SDOperand> MemOpChains;
1071 SDOperand StackPtr = DAG.getRegister(X86::ESP, getPointerTy());
1072 for (unsigned i = 0; i != NumOps; ++i) {
1073 SDOperand Arg = Op.getOperand(5+2*i);
1075 switch (Arg.getValueType()) {
1076 default: assert(0 && "Unexpected ValueType for argument!");
1080 if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS) {
1081 RegsToPass.push_back(
1082 std::make_pair(GPRArgRegs[Arg.getValueType()-MVT::i8][NumIntRegs],
1089 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1090 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1091 MemOpChains.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
1092 Arg, PtrOff, DAG.getSrcValue(NULL)));
1097 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1098 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1099 MemOpChains.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
1100 Arg, PtrOff, DAG.getSrcValue(NULL)));
1110 if (NumXMMRegs < 3) {
1111 RegsToPass.push_back(std::make_pair(XMMArgRegs[NumXMMRegs], Arg));
1114 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1115 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff);
1116 MemOpChains.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
1117 Arg, PtrOff, DAG.getSrcValue(NULL)));
1124 if (!MemOpChains.empty())
1125 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, MemOpChains);
1127 // Build a sequence of copy-to-reg nodes chained together with token chain
1128 // and flag operands which copy the outgoing args into registers.
1130 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1131 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1133 InFlag = Chain.getValue(1);
1136 // If the callee is a GlobalAddress node (quite common, every direct call is)
1137 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1138 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1139 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
1140 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1141 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());
1143 std::vector<MVT::ValueType> NodeTys;
1144 NodeTys.push_back(MVT::Other); // Returns a chain
1145 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
1146 std::vector<SDOperand> Ops;
1147 Ops.push_back(Chain);
1148 Ops.push_back(Callee);
1150 Ops.push_back(InFlag);
1152 // FIXME: Do not generate X86ISD::TAILCALL for now.
1153 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL,
1155 InFlag = Chain.getValue(1);
1158 NodeTys.push_back(MVT::Other); // Returns a chain
1159 if (RetVT != MVT::Other)
1160 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
1162 Ops.push_back(Chain);
1163 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1164 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
1165 Ops.push_back(InFlag);
1166 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, Ops);
1167 if (RetVT != MVT::Other)
1168 InFlag = Chain.getValue(1);
1170 std::vector<SDOperand> ResultVals;
1173 default: assert(0 && "Unknown value type to return!");
1174 case MVT::Other: break;
1176 Chain = DAG.getCopyFromReg(Chain, X86::AL, MVT::i8, InFlag).getValue(1);
1177 ResultVals.push_back(Chain.getValue(0));
1178 NodeTys.push_back(MVT::i8);
1181 Chain = DAG.getCopyFromReg(Chain, X86::AX, MVT::i16, InFlag).getValue(1);
1182 ResultVals.push_back(Chain.getValue(0));
1183 NodeTys.push_back(MVT::i16);
1186 if (Op.Val->getValueType(1) == MVT::i32) {
1187 Chain = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag).getValue(1);
1188 ResultVals.push_back(Chain.getValue(0));
1189 Chain = DAG.getCopyFromReg(Chain, X86::EDX, MVT::i32,
1190 Chain.getValue(2)).getValue(1);
1191 ResultVals.push_back(Chain.getValue(0));
1192 NodeTys.push_back(MVT::i32);
1194 Chain = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag).getValue(1);
1195 ResultVals.push_back(Chain.getValue(0));
1197 NodeTys.push_back(MVT::i32);
1206 Chain = DAG.getCopyFromReg(Chain, X86::XMM0, RetVT, InFlag).getValue(1);
1207 ResultVals.push_back(Chain.getValue(0));
1208 NodeTys.push_back(RetVT);
1212 std::vector<MVT::ValueType> Tys;
1213 Tys.push_back(MVT::f64);
1214 Tys.push_back(MVT::Other);
1215 Tys.push_back(MVT::Flag);
1216 std::vector<SDOperand> Ops;
1217 Ops.push_back(Chain);
1218 Ops.push_back(InFlag);
1219 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, Ops);
1220 Chain = RetVal.getValue(1);
1221 InFlag = RetVal.getValue(2);
1223 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
1224 // shouldn't be necessary except that RFP cannot be live across
1225 // multiple blocks. When stackifier is fixed, they can be uncoupled.
1226 MachineFunction &MF = DAG.getMachineFunction();
1227 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
1228 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
1230 Tys.push_back(MVT::Other);
1232 Ops.push_back(Chain);
1233 Ops.push_back(RetVal);
1234 Ops.push_back(StackSlot);
1235 Ops.push_back(DAG.getValueType(RetVT));
1236 Ops.push_back(InFlag);
1237 Chain = DAG.getNode(X86ISD::FST, Tys, Ops);
1238 RetVal = DAG.getLoad(RetVT, Chain, StackSlot,
1239 DAG.getSrcValue(NULL));
1240 Chain = RetVal.getValue(1);
1243 if (RetVT == MVT::f32 && !X86ScalarSSE)
1244 // FIXME: we would really like to remember that this FP_ROUND
1245 // operation is okay to eliminate if we allow excess FP precision.
1246 RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal);
1247 ResultVals.push_back(RetVal);
1248 NodeTys.push_back(RetVT);
1254 // If the function returns void, just return the chain.
1255 if (ResultVals.empty())
1258 // Otherwise, merge everything together with a MERGE_VALUES node.
1259 NodeTys.push_back(MVT::Other);
1260 ResultVals.push_back(Chain);
1261 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys, ResultVals);
1262 return Res.getValue(Op.ResNo);
1265 SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
1266 if (ReturnAddrIndex == 0) {
1267 // Set up a frame object for the return address.
1268 MachineFunction &MF = DAG.getMachineFunction();
1269 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
1272 return DAG.getFrameIndex(ReturnAddrIndex, MVT::i32);
1277 std::pair<SDOperand, SDOperand> X86TargetLowering::
1278 LowerFrameReturnAddress(bool isFrameAddress, SDOperand Chain, unsigned Depth,
1279 SelectionDAG &DAG) {
1281 if (Depth) // Depths > 0 not supported yet!
1282 Result = DAG.getConstant(0, getPointerTy());
1284 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
1285 if (!isFrameAddress)
1286 // Just load the return address
1287 Result = DAG.getLoad(MVT::i32, DAG.getEntryNode(), RetAddrFI,
1288 DAG.getSrcValue(NULL));
1290 Result = DAG.getNode(ISD::SUB, MVT::i32, RetAddrFI,
1291 DAG.getConstant(4, MVT::i32));
1293 return std::make_pair(Result, Chain);
1296 /// getCondBrOpcodeForX86CC - Returns the X86 conditional branch opcode
1297 /// which corresponds to the condition code.
1298 static unsigned getCondBrOpcodeForX86CC(unsigned X86CC) {
1300 default: assert(0 && "Unknown X86 conditional code!");
1301 case X86ISD::COND_A: return X86::JA;
1302 case X86ISD::COND_AE: return X86::JAE;
1303 case X86ISD::COND_B: return X86::JB;
1304 case X86ISD::COND_BE: return X86::JBE;
1305 case X86ISD::COND_E: return X86::JE;
1306 case X86ISD::COND_G: return X86::JG;
1307 case X86ISD::COND_GE: return X86::JGE;
1308 case X86ISD::COND_L: return X86::JL;
1309 case X86ISD::COND_LE: return X86::JLE;
1310 case X86ISD::COND_NE: return X86::JNE;
1311 case X86ISD::COND_NO: return X86::JNO;
1312 case X86ISD::COND_NP: return X86::JNP;
1313 case X86ISD::COND_NS: return X86::JNS;
1314 case X86ISD::COND_O: return X86::JO;
1315 case X86ISD::COND_P: return X86::JP;
1316 case X86ISD::COND_S: return X86::JS;
1320 /// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
1321 /// specific condition code. It returns a false if it cannot do a direct
1322 /// translation. X86CC is the translated CondCode. Flip is set to true if the
1323 /// the order of comparison operands should be flipped.
1324 static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
1325 unsigned &X86CC, bool &Flip) {
1327 X86CC = X86ISD::COND_INVALID;
1329 switch (SetCCOpcode) {
1331 case ISD::SETEQ: X86CC = X86ISD::COND_E; break;
1332 case ISD::SETGT: X86CC = X86ISD::COND_G; break;
1333 case ISD::SETGE: X86CC = X86ISD::COND_GE; break;
1334 case ISD::SETLT: X86CC = X86ISD::COND_L; break;
1335 case ISD::SETLE: X86CC = X86ISD::COND_LE; break;
1336 case ISD::SETNE: X86CC = X86ISD::COND_NE; break;
1337 case ISD::SETULT: X86CC = X86ISD::COND_B; break;
1338 case ISD::SETUGT: X86CC = X86ISD::COND_A; break;
1339 case ISD::SETULE: X86CC = X86ISD::COND_BE; break;
1340 case ISD::SETUGE: X86CC = X86ISD::COND_AE; break;
1343 // On a floating point condition, the flags are set as follows:
1345 // 0 | 0 | 0 | X > Y
1346 // 0 | 0 | 1 | X < Y
1347 // 1 | 0 | 0 | X == Y
1348 // 1 | 1 | 1 | unordered
1349 switch (SetCCOpcode) {
1352 case ISD::SETEQ: X86CC = X86ISD::COND_E; break;
1353 case ISD::SETOLT: Flip = true; // Fallthrough
1355 case ISD::SETGT: X86CC = X86ISD::COND_A; break;
1356 case ISD::SETOLE: Flip = true; // Fallthrough
1358 case ISD::SETGE: X86CC = X86ISD::COND_AE; break;
1359 case ISD::SETUGT: Flip = true; // Fallthrough
1361 case ISD::SETLT: X86CC = X86ISD::COND_B; break;
1362 case ISD::SETUGE: Flip = true; // Fallthrough
1364 case ISD::SETLE: X86CC = X86ISD::COND_BE; break;
1366 case ISD::SETNE: X86CC = X86ISD::COND_NE; break;
1367 case ISD::SETUO: X86CC = X86ISD::COND_P; break;
1368 case ISD::SETO: X86CC = X86ISD::COND_NP; break;
1372 return X86CC != X86ISD::COND_INVALID;
1375 static bool translateX86CC(SDOperand CC, bool isFP, unsigned &X86CC,
1377 return translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC, Flip);
1380 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
1381 /// code. Current x86 isa includes the following FP cmov instructions:
1382 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
1383 static bool hasFPCMov(unsigned X86CC) {
1387 case X86ISD::COND_B:
1388 case X86ISD::COND_BE:
1389 case X86ISD::COND_E:
1390 case X86ISD::COND_P:
1391 case X86ISD::COND_A:
1392 case X86ISD::COND_AE:
1393 case X86ISD::COND_NE:
1394 case X86ISD::COND_NP:
1400 X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
1401 MachineBasicBlock *BB) {
1402 switch (MI->getOpcode()) {
1403 default: assert(false && "Unexpected instr type to insert");
1404 case X86::CMOV_FR32:
1405 case X86::CMOV_FR64:
1406 case X86::CMOV_V4F32:
1407 case X86::CMOV_V2F64:
1408 case X86::CMOV_V2I64: {
1409 // To "insert" a SELECT_CC instruction, we actually have to insert the
1410 // diamond control-flow pattern. The incoming instruction knows the
1411 // destination vreg to set, the condition code register to branch on, the
1412 // true/false values to select between, and a branch opcode to use.
1413 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1414 ilist<MachineBasicBlock>::iterator It = BB;
1420 // cmpTY ccX, r1, r2
1422 // fallthrough --> copy0MBB
1423 MachineBasicBlock *thisMBB = BB;
1424 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
1425 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
1426 unsigned Opc = getCondBrOpcodeForX86CC(MI->getOperand(3).getImmedValue());
1427 BuildMI(BB, Opc, 1).addMBB(sinkMBB);
1428 MachineFunction *F = BB->getParent();
1429 F->getBasicBlockList().insert(It, copy0MBB);
1430 F->getBasicBlockList().insert(It, sinkMBB);
1431 // Update machine-CFG edges by first adding all successors of the current
1432 // block to the new block which will contain the Phi node for the select.
1433 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
1434 e = BB->succ_end(); i != e; ++i)
1435 sinkMBB->addSuccessor(*i);
1436 // Next, remove all successors of the current block, and add the true
1437 // and fallthrough blocks as its successors.
1438 while(!BB->succ_empty())
1439 BB->removeSuccessor(BB->succ_begin());
1440 BB->addSuccessor(copy0MBB);
1441 BB->addSuccessor(sinkMBB);
1444 // %FalseValue = ...
1445 // # fallthrough to sinkMBB
1448 // Update machine-CFG edges
1449 BB->addSuccessor(sinkMBB);
1452 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
1455 BuildMI(BB, X86::PHI, 4, MI->getOperand(0).getReg())
1456 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
1457 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
1459 delete MI; // The pseudo instruction is gone now.
1463 case X86::FP_TO_INT16_IN_MEM:
1464 case X86::FP_TO_INT32_IN_MEM:
1465 case X86::FP_TO_INT64_IN_MEM: {
1466 // Change the floating point control register to use "round towards zero"
1467 // mode when truncating to an integer value.
1468 MachineFunction *F = BB->getParent();
1469 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
1470 addFrameReference(BuildMI(BB, X86::FNSTCW16m, 4), CWFrameIdx);
1472 // Load the old value of the high byte of the control word...
1474 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass);
1475 addFrameReference(BuildMI(BB, X86::MOV16rm, 4, OldCW), CWFrameIdx);
1477 // Set the high part to be round to zero...
1478 addFrameReference(BuildMI(BB, X86::MOV16mi, 5), CWFrameIdx).addImm(0xC7F);
1480 // Reload the modified control word now...
1481 addFrameReference(BuildMI(BB, X86::FLDCW16m, 4), CWFrameIdx);
1483 // Restore the memory image of control word to original value
1484 addFrameReference(BuildMI(BB, X86::MOV16mr, 5), CWFrameIdx).addReg(OldCW);
1486 // Get the X86 opcode to use.
1488 switch (MI->getOpcode()) {
1489 default: assert(0 && "illegal opcode!");
1490 case X86::FP_TO_INT16_IN_MEM: Opc = X86::FpIST16m; break;
1491 case X86::FP_TO_INT32_IN_MEM: Opc = X86::FpIST32m; break;
1492 case X86::FP_TO_INT64_IN_MEM: Opc = X86::FpIST64m; break;
1496 MachineOperand &Op = MI->getOperand(0);
1497 if (Op.isRegister()) {
1498 AM.BaseType = X86AddressMode::RegBase;
1499 AM.Base.Reg = Op.getReg();
1501 AM.BaseType = X86AddressMode::FrameIndexBase;
1502 AM.Base.FrameIndex = Op.getFrameIndex();
1504 Op = MI->getOperand(1);
1505 if (Op.isImmediate())
1506 AM.Scale = Op.getImmedValue();
1507 Op = MI->getOperand(2);
1508 if (Op.isImmediate())
1509 AM.IndexReg = Op.getImmedValue();
1510 Op = MI->getOperand(3);
1511 if (Op.isGlobalAddress()) {
1512 AM.GV = Op.getGlobal();
1514 AM.Disp = Op.getImmedValue();
1516 addFullAddress(BuildMI(BB, Opc, 5), AM).addReg(MI->getOperand(4).getReg());
1518 // Reload the original control word now.
1519 addFrameReference(BuildMI(BB, X86::FLDCW16m, 4), CWFrameIdx);
1521 delete MI; // The pseudo instruction is gone now.
1528 //===----------------------------------------------------------------------===//
1529 // X86 Custom Lowering Hooks
1530 //===----------------------------------------------------------------------===//
1532 /// DarwinGVRequiresExtraLoad - true if accessing the GV requires an extra
1533 /// load. For Darwin, external and weak symbols are indirect, loading the value
1534 /// at address GV rather then the value of GV itself. This means that the
1535 /// GlobalAddress must be in the base or index register of the address, not the
1536 /// GV offset field.
1537 static bool DarwinGVRequiresExtraLoad(GlobalValue *GV) {
1538 return (GV->hasWeakLinkage() || GV->hasLinkOnceLinkage() ||
1539 (GV->isExternal() && !GV->hasNotBeenReadFromBytecode()));
1542 /// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return
1543 /// true if Op is undef or if its value falls within the specified range (L, H].
1544 static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) {
1545 if (Op.getOpcode() == ISD::UNDEF)
1548 unsigned Val = cast<ConstantSDNode>(Op)->getValue();
1549 return (Val >= Low && Val < Hi);
1552 /// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return
1553 /// true if Op is undef or if its value equal to the specified value.
1554 static bool isUndefOrEqual(SDOperand Op, unsigned Val) {
1555 if (Op.getOpcode() == ISD::UNDEF)
1557 return cast<ConstantSDNode>(Op)->getValue() == Val;
1560 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
1561 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
1562 bool X86::isPSHUFDMask(SDNode *N) {
1563 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1565 if (N->getNumOperands() != 4)
1568 // Check if the value doesn't reference the second vector.
1569 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1570 SDOperand Arg = N->getOperand(i);
1571 if (Arg.getOpcode() == ISD::UNDEF) continue;
1572 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1573 if (cast<ConstantSDNode>(Arg)->getValue() >= 4)
1580 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
1581 /// specifies a shuffle of elements that is suitable for input to PSHUFHW.
1582 bool X86::isPSHUFHWMask(SDNode *N) {
1583 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1585 if (N->getNumOperands() != 8)
1588 // Lower quadword copied in order.
1589 for (unsigned i = 0; i != 4; ++i) {
1590 SDOperand Arg = N->getOperand(i);
1591 if (Arg.getOpcode() == ISD::UNDEF) continue;
1592 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1593 if (cast<ConstantSDNode>(Arg)->getValue() != i)
1597 // Upper quadword shuffled.
1598 for (unsigned i = 4; i != 8; ++i) {
1599 SDOperand Arg = N->getOperand(i);
1600 if (Arg.getOpcode() == ISD::UNDEF) continue;
1601 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1602 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1603 if (Val < 4 || Val > 7)
1610 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
1611 /// specifies a shuffle of elements that is suitable for input to PSHUFLW.
1612 bool X86::isPSHUFLWMask(SDNode *N) {
1613 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1615 if (N->getNumOperands() != 8)
1618 // Upper quadword copied in order.
1619 for (unsigned i = 4; i != 8; ++i)
1620 if (!isUndefOrEqual(N->getOperand(i), i))
1623 // Lower quadword shuffled.
1624 for (unsigned i = 0; i != 4; ++i)
1625 if (!isUndefOrInRange(N->getOperand(i), 0, 4))
1631 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
1632 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
1633 static bool isSHUFPMask(std::vector<SDOperand> &N) {
1634 unsigned NumElems = N.size();
1635 if (NumElems != 2 && NumElems != 4) return false;
1637 unsigned Half = NumElems / 2;
1638 for (unsigned i = 0; i < Half; ++i)
1639 if (!isUndefOrInRange(N[i], 0, NumElems))
1641 for (unsigned i = Half; i < NumElems; ++i)
1642 if (!isUndefOrInRange(N[i], NumElems, NumElems*2))
1648 bool X86::isSHUFPMask(SDNode *N) {
1649 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1650 std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
1651 return ::isSHUFPMask(Ops);
1654 /// isCommutedSHUFP - Returns true if the shuffle mask is except
1655 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
1656 /// half elements to come from vector 1 (which would equal the dest.) and
1657 /// the upper half to come from vector 2.
1658 static bool isCommutedSHUFP(std::vector<SDOperand> &Ops) {
1659 unsigned NumElems = Ops.size();
1660 if (NumElems != 2 && NumElems != 4) return false;
1662 unsigned Half = NumElems / 2;
1663 for (unsigned i = 0; i < Half; ++i)
1664 if (!isUndefOrInRange(Ops[i], NumElems, NumElems*2))
1666 for (unsigned i = Half; i < NumElems; ++i)
1667 if (!isUndefOrInRange(Ops[i], 0, NumElems))
1672 static bool isCommutedSHUFP(SDNode *N) {
1673 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1674 std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
1675 return isCommutedSHUFP(Ops);
1678 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
1679 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
1680 bool X86::isMOVHLPSMask(SDNode *N) {
1681 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1683 if (N->getNumOperands() != 4)
1686 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
1687 return isUndefOrEqual(N->getOperand(0), 6) &&
1688 isUndefOrEqual(N->getOperand(1), 7) &&
1689 isUndefOrEqual(N->getOperand(2), 2) &&
1690 isUndefOrEqual(N->getOperand(3), 3);
1693 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
1694 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
1695 bool X86::isMOVLPMask(SDNode *N) {
1696 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1698 unsigned NumElems = N->getNumOperands();
1699 if (NumElems != 2 && NumElems != 4)
1702 for (unsigned i = 0; i < NumElems/2; ++i)
1703 if (!isUndefOrEqual(N->getOperand(i), i + NumElems))
1706 for (unsigned i = NumElems/2; i < NumElems; ++i)
1707 if (!isUndefOrEqual(N->getOperand(i), i))
1713 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
1714 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
1716 bool X86::isMOVHPMask(SDNode *N) {
1717 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1719 unsigned NumElems = N->getNumOperands();
1720 if (NumElems != 2 && NumElems != 4)
1723 for (unsigned i = 0; i < NumElems/2; ++i)
1724 if (!isUndefOrEqual(N->getOperand(i), i))
1727 for (unsigned i = 0; i < NumElems/2; ++i) {
1728 SDOperand Arg = N->getOperand(i + NumElems/2);
1729 if (!isUndefOrEqual(Arg, i + NumElems))
1736 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
1737 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
1738 bool static isUNPCKLMask(std::vector<SDOperand> &N, bool V2IsSplat = false) {
1739 unsigned NumElems = N.size();
1740 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
1743 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
1744 SDOperand BitI = N[i];
1745 SDOperand BitI1 = N[i+1];
1746 if (!isUndefOrEqual(BitI, j))
1749 if (isUndefOrEqual(BitI1, NumElems))
1752 if (!isUndefOrEqual(BitI1, j + NumElems))
1760 bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) {
1761 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1762 std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
1763 return ::isUNPCKLMask(Ops, V2IsSplat);
1766 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
1767 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
1768 bool static isUNPCKHMask(std::vector<SDOperand> &N, bool V2IsSplat = false) {
1769 unsigned NumElems = N.size();
1770 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
1773 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
1774 SDOperand BitI = N[i];
1775 SDOperand BitI1 = N[i+1];
1776 if (!isUndefOrEqual(BitI, j + NumElems/2))
1779 if (isUndefOrEqual(BitI1, NumElems))
1782 if (!isUndefOrEqual(BitI1, j + NumElems/2 + NumElems))
1790 bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) {
1791 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1792 std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
1793 return ::isUNPCKHMask(Ops, V2IsSplat);
1796 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
1797 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
1799 bool X86::isUNPCKL_v_undef_Mask(SDNode *N) {
1800 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1802 unsigned NumElems = N->getNumOperands();
1803 if (NumElems != 4 && NumElems != 8 && NumElems != 16)
1806 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) {
1807 SDOperand BitI = N->getOperand(i);
1808 SDOperand BitI1 = N->getOperand(i+1);
1810 if (!isUndefOrEqual(BitI, j))
1812 if (!isUndefOrEqual(BitI1, j))
1819 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
1820 /// specifies a shuffle of elements that is suitable for input to MOVSS,
1821 /// MOVSD, and MOVD, i.e. setting the lowest element.
1822 static bool isMOVLMask(std::vector<SDOperand> &N) {
1823 unsigned NumElems = N.size();
1824 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
1827 if (!isUndefOrEqual(N[0], NumElems))
1830 for (unsigned i = 1; i < NumElems; ++i) {
1831 SDOperand Arg = N[i];
1832 if (!isUndefOrEqual(Arg, i))
1839 bool X86::isMOVLMask(SDNode *N) {
1840 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1841 std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
1842 return ::isMOVLMask(Ops);
1845 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
1846 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
1847 /// element of vector 2 and the other elements to come from vector 1 in order.
1848 static bool isCommutedMOVL(std::vector<SDOperand> &Ops, bool V2IsSplat = false) {
1849 unsigned NumElems = Ops.size();
1850 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
1853 if (!isUndefOrEqual(Ops[0], 0))
1856 for (unsigned i = 1; i < NumElems; ++i) {
1857 SDOperand Arg = Ops[i];
1859 if (!isUndefOrEqual(Arg, NumElems))
1862 if (!isUndefOrEqual(Arg, i+NumElems))
1870 static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false) {
1871 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1872 std::vector<SDOperand> Ops(N->op_begin(), N->op_end());
1873 return isCommutedMOVL(Ops, V2IsSplat);
1876 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
1877 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
1878 bool X86::isMOVSHDUPMask(SDNode *N) {
1879 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1881 if (N->getNumOperands() != 4)
1884 // Expect 1, 1, 3, 3
1885 for (unsigned i = 0; i < 2; ++i) {
1886 SDOperand Arg = N->getOperand(i);
1887 if (Arg.getOpcode() == ISD::UNDEF) continue;
1888 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1889 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1890 if (Val != 1) return false;
1894 for (unsigned i = 2; i < 4; ++i) {
1895 SDOperand Arg = N->getOperand(i);
1896 if (Arg.getOpcode() == ISD::UNDEF) continue;
1897 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1898 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1899 if (Val != 3) return false;
1903 // Don't use movshdup if it can be done with a shufps.
1907 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
1908 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
1909 bool X86::isMOVSLDUPMask(SDNode *N) {
1910 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1912 if (N->getNumOperands() != 4)
1915 // Expect 0, 0, 2, 2
1916 for (unsigned i = 0; i < 2; ++i) {
1917 SDOperand Arg = N->getOperand(i);
1918 if (Arg.getOpcode() == ISD::UNDEF) continue;
1919 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1920 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1921 if (Val != 0) return false;
1925 for (unsigned i = 2; i < 4; ++i) {
1926 SDOperand Arg = N->getOperand(i);
1927 if (Arg.getOpcode() == ISD::UNDEF) continue;
1928 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1929 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
1930 if (Val != 2) return false;
1934 // Don't use movshdup if it can be done with a shufps.
1938 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
1939 /// a splat of a single element.
1940 static bool isSplatMask(SDNode *N) {
1941 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1943 // This is a splat operation if each element of the permute is the same, and
1944 // if the value doesn't reference the second vector.
1945 unsigned NumElems = N->getNumOperands();
1946 SDOperand ElementBase;
1948 for (; i != NumElems; ++i) {
1949 SDOperand Elt = N->getOperand(i);
1950 if (ConstantSDNode *EltV = dyn_cast<ConstantSDNode>(Elt)) {
1956 if (!ElementBase.Val)
1959 for (; i != NumElems; ++i) {
1960 SDOperand Arg = N->getOperand(i);
1961 if (Arg.getOpcode() == ISD::UNDEF) continue;
1962 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
1963 if (Arg != ElementBase) return false;
1966 // Make sure it is a splat of the first vector operand.
1967 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems;
1970 /// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies
1971 /// a splat of a single element and it's a 2 or 4 element mask.
1972 bool X86::isSplatMask(SDNode *N) {
1973 assert(N->getOpcode() == ISD::BUILD_VECTOR);
1975 // We can only splat 64-bit, and 32-bit quantities with a single instruction.
1976 if (N->getNumOperands() != 4 && N->getNumOperands() != 2)
1978 return ::isSplatMask(N);
1981 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
1982 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
1984 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
1985 unsigned NumOperands = N->getNumOperands();
1986 unsigned Shift = (NumOperands == 4) ? 2 : 1;
1988 for (unsigned i = 0; i < NumOperands; ++i) {
1990 SDOperand Arg = N->getOperand(NumOperands-i-1);
1991 if (Arg.getOpcode() != ISD::UNDEF)
1992 Val = cast<ConstantSDNode>(Arg)->getValue();
1993 if (Val >= NumOperands) Val -= NumOperands;
1995 if (i != NumOperands - 1)
2002 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2003 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2005 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2007 // 8 nodes, but we only care about the last 4.
2008 for (unsigned i = 7; i >= 4; --i) {
2010 SDOperand Arg = N->getOperand(i);
2011 if (Arg.getOpcode() != ISD::UNDEF)
2012 Val = cast<ConstantSDNode>(Arg)->getValue();
2021 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2022 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2024 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2026 // 8 nodes, but we only care about the first 4.
2027 for (int i = 3; i >= 0; --i) {
2029 SDOperand Arg = N->getOperand(i);
2030 if (Arg.getOpcode() != ISD::UNDEF)
2031 Val = cast<ConstantSDNode>(Arg)->getValue();
2040 /// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand
2041 /// specifies a 8 element shuffle that can be broken into a pair of
2042 /// PSHUFHW and PSHUFLW.
2043 static bool isPSHUFHW_PSHUFLWMask(SDNode *N) {
2044 assert(N->getOpcode() == ISD::BUILD_VECTOR);
2046 if (N->getNumOperands() != 8)
2049 // Lower quadword shuffled.
2050 for (unsigned i = 0; i != 4; ++i) {
2051 SDOperand Arg = N->getOperand(i);
2052 if (Arg.getOpcode() == ISD::UNDEF) continue;
2053 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2054 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2059 // Upper quadword shuffled.
2060 for (unsigned i = 4; i != 8; ++i) {
2061 SDOperand Arg = N->getOperand(i);
2062 if (Arg.getOpcode() == ISD::UNDEF) continue;
2063 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2064 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2065 if (Val < 4 || Val > 7)
2072 /// CommuteVectorShuffle - Swap vector_shuffle operandsas well as
2073 /// values in ther permute mask.
2074 static SDOperand CommuteVectorShuffle(SDOperand Op, SelectionDAG &DAG) {
2075 SDOperand V1 = Op.getOperand(0);
2076 SDOperand V2 = Op.getOperand(1);
2077 SDOperand Mask = Op.getOperand(2);
2078 MVT::ValueType VT = Op.getValueType();
2079 MVT::ValueType MaskVT = Mask.getValueType();
2080 MVT::ValueType EltVT = MVT::getVectorBaseType(MaskVT);
2081 unsigned NumElems = Mask.getNumOperands();
2082 std::vector<SDOperand> MaskVec;
2084 for (unsigned i = 0; i != NumElems; ++i) {
2085 SDOperand Arg = Mask.getOperand(i);
2086 if (Arg.getOpcode() == ISD::UNDEF) {
2087 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT));
2090 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2091 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2093 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT));
2095 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT));
2098 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
2099 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V2, V1, Mask);
2102 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2103 /// match movhlps. The lower half elements should come from upper half of
2104 /// V1 (and in order), and the upper half elements should come from the upper
2105 /// half of V2 (and in order).
2106 static bool ShouldXformToMOVHLPS(SDNode *Mask) {
2107 unsigned NumElems = Mask->getNumOperands();
2110 for (unsigned i = 0, e = 2; i != e; ++i)
2111 if (!isUndefOrEqual(Mask->getOperand(i), i+2))
2113 for (unsigned i = 2; i != 4; ++i)
2114 if (!isUndefOrEqual(Mask->getOperand(i), i+4))
2119 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2120 /// is promoted to a vector.
2121 static inline bool isScalarLoadToVector(SDNode *N) {
2122 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
2123 N = N->getOperand(0).Val;
2124 return (N->getOpcode() == ISD::LOAD);
2129 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2130 /// match movlp{s|d}. The lower half elements should come from lower half of
2131 /// V1 (and in order), and the upper half elements should come from the upper
2132 /// half of V2 (and in order). And since V1 will become the source of the
2133 /// MOVLP, it must be either a vector load or a scalar load to vector.
2134 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *Mask) {
2135 if (V1->getOpcode() != ISD::LOAD && !isScalarLoadToVector(V1))
2138 unsigned NumElems = Mask->getNumOperands();
2139 if (NumElems != 2 && NumElems != 4)
2141 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2142 if (!isUndefOrEqual(Mask->getOperand(i), i))
2144 for (unsigned i = NumElems/2; i != NumElems; ++i)
2145 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems))
2150 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2152 static bool isSplatVector(SDNode *N) {
2153 if (N->getOpcode() != ISD::BUILD_VECTOR)
2156 SDOperand SplatValue = N->getOperand(0);
2157 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2158 if (N->getOperand(i) != SplatValue)
2163 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2164 /// that point to V2 points to its first element.
2165 static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) {
2166 assert(Mask.getOpcode() == ISD::BUILD_VECTOR);
2168 bool Changed = false;
2169 std::vector<SDOperand> MaskVec;
2170 unsigned NumElems = Mask.getNumOperands();
2171 for (unsigned i = 0; i != NumElems; ++i) {
2172 SDOperand Arg = Mask.getOperand(i);
2173 if (Arg.getOpcode() != ISD::UNDEF) {
2174 unsigned Val = cast<ConstantSDNode>(Arg)->getValue();
2175 if (Val > NumElems) {
2176 Arg = DAG.getConstant(NumElems, Arg.getValueType());
2180 MaskVec.push_back(Arg);
2184 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(), MaskVec);
2188 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2189 /// operation of specified width.
2190 static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) {
2191 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2192 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2194 std::vector<SDOperand> MaskVec;
2195 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT));
2196 for (unsigned i = 1; i != NumElems; ++i)
2197 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2198 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
2201 /// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation
2202 /// of specified width.
2203 static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) {
2204 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2205 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2206 std::vector<SDOperand> MaskVec;
2207 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2208 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2209 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT));
2211 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
2214 /// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation
2215 /// of specified width.
2216 static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) {
2217 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2218 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2219 unsigned Half = NumElems/2;
2220 std::vector<SDOperand> MaskVec;
2221 for (unsigned i = 0; i != Half; ++i) {
2222 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT));
2223 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT));
2225 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
2228 /// getZeroVector - Returns a vector of specified type with all zero elements.
2230 static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) {
2231 assert(MVT::isVector(VT) && "Expected a vector type");
2232 unsigned NumElems = getVectorNumElements(VT);
2233 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2234 bool isFP = MVT::isFloatingPoint(EVT);
2235 SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT);
2236 std::vector<SDOperand> ZeroVec(NumElems, Zero);
2237 return DAG.getNode(ISD::BUILD_VECTOR, VT, ZeroVec);
2240 /// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32.
2242 static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) {
2243 SDOperand V1 = Op.getOperand(0);
2244 SDOperand Mask = Op.getOperand(2);
2245 MVT::ValueType VT = Op.getValueType();
2246 unsigned NumElems = Mask.getNumOperands();
2247 Mask = getUnpacklMask(NumElems, DAG);
2248 while (NumElems != 4) {
2249 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask);
2252 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1);
2254 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2255 Mask = getZeroVector(MaskVT, DAG);
2256 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1,
2257 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask);
2258 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
2261 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2263 static inline bool isZeroNode(SDOperand Elt) {
2264 return ((isa<ConstantSDNode>(Elt) &&
2265 cast<ConstantSDNode>(Elt)->getValue() == 0) ||
2266 (isa<ConstantFPSDNode>(Elt) &&
2267 cast<ConstantFPSDNode>(Elt)->isExactlyValue(0.0)));
2270 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2271 /// vector and zero or undef vector.
2272 static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT,
2273 unsigned NumElems, unsigned Idx,
2274 bool isZero, SelectionDAG &DAG) {
2275 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT);
2276 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2277 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2278 SDOperand Zero = DAG.getConstant(0, EVT);
2279 std::vector<SDOperand> MaskVec(NumElems, Zero);
2280 MaskVec[Idx] = DAG.getConstant(NumElems, EVT);
2281 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
2282 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2285 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
2287 static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros,
2288 unsigned NumNonZero, unsigned NumZero,
2289 SelectionDAG &DAG) {
2295 for (unsigned i = 0; i < 16; ++i) {
2296 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
2297 if (ThisIsNonZero && First) {
2299 V = getZeroVector(MVT::v8i16, DAG);
2301 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2306 SDOperand ThisElt(0, 0), LastElt(0, 0);
2307 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
2308 if (LastIsNonZero) {
2309 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1));
2311 if (ThisIsNonZero) {
2312 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i));
2313 ThisElt = DAG.getNode(ISD::SHL, MVT::i16,
2314 ThisElt, DAG.getConstant(8, MVT::i8));
2316 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt);
2321 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt,
2322 DAG.getConstant(i/2, MVT::i32));
2326 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V);
2329 /// LowerBuildVectorv16i8 - Custom lower build_vector of v8i16.
2331 static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros,
2332 unsigned NumNonZero, unsigned NumZero,
2333 SelectionDAG &DAG) {
2339 for (unsigned i = 0; i < 8; ++i) {
2340 bool isNonZero = (NonZeros & (1 << i)) != 0;
2344 V = getZeroVector(MVT::v8i16, DAG);
2346 V = DAG.getNode(ISD::UNDEF, MVT::v8i16);
2349 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i),
2350 DAG.getConstant(i, MVT::i32));
2358 X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2359 // All zero's are handled with pxor.
2360 if (ISD::isBuildVectorAllZeros(Op.Val))
2363 // All one's are handled with pcmpeqd.
2364 if (ISD::isBuildVectorAllOnes(Op.Val))
2367 MVT::ValueType VT = Op.getValueType();
2368 MVT::ValueType EVT = MVT::getVectorBaseType(VT);
2369 unsigned EVTBits = MVT::getSizeInBits(EVT);
2371 unsigned NumElems = Op.getNumOperands();
2372 unsigned NumZero = 0;
2373 unsigned NumNonZero = 0;
2374 unsigned NonZeros = 0;
2375 std::set<SDOperand> Values;
2376 for (unsigned i = 0; i < NumElems; ++i) {
2377 SDOperand Elt = Op.getOperand(i);
2378 if (Elt.getOpcode() != ISD::UNDEF) {
2380 if (isZeroNode(Elt))
2383 NonZeros |= (1 << i);
2389 if (NumNonZero == 0)
2390 // Must be a mix of zero and undef. Return a zero vector.
2391 return getZeroVector(VT, DAG);
2393 // Splat is obviously ok. Let legalizer expand it to a shuffle.
2394 if (Values.size() == 1)
2397 // Special case for single non-zero element.
2398 if (NumNonZero == 1) {
2399 unsigned Idx = CountTrailingZeros_32(NonZeros);
2400 SDOperand Item = Op.getOperand(Idx);
2401 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item);
2403 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
2404 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx,
2407 if (EVTBits == 32) {
2408 // Turn it into a shuffle of zero and zero-extended scalar to vector.
2409 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0,
2411 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2412 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
2413 std::vector<SDOperand> MaskVec;
2414 for (unsigned i = 0; i < NumElems; i++)
2415 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT));
2416 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
2417 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item,
2418 DAG.getNode(ISD::UNDEF, VT), Mask);
2422 // Let legalizer expand 2-widde build_vector's.
2426 // If element VT is < 32 bits, convert it to inserts into a zero vector.
2428 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG);
2429 if (V.Val) return V;
2432 if (EVTBits == 16) {
2433 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG);
2434 if (V.Val) return V;
2437 // If element VT is == 32 bits, turn it into a number of shuffles.
2438 std::vector<SDOperand> V(NumElems);
2439 if (NumElems == 4 && NumZero > 0) {
2440 for (unsigned i = 0; i < 4; ++i) {
2441 bool isZero = !(NonZeros & (1 << i));
2443 V[i] = getZeroVector(VT, DAG);
2445 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2448 for (unsigned i = 0; i < 2; ++i) {
2449 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
2452 V[i] = V[i*2]; // Must be a zero vector.
2455 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2],
2456 getMOVLMask(NumElems, DAG));
2459 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2460 getMOVLMask(NumElems, DAG));
2463 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1],
2464 getUnpacklMask(NumElems, DAG));
2469 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd)
2470 // clears the upper bits.
2471 // FIXME: we can do the same for v4f32 case when we know both parts of
2472 // the lower half come from scalar_to_vector (loadf32). We should do
2473 // that in post legalizer dag combiner with target specific hooks.
2474 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0)
2476 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2477 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT);
2478 std::vector<SDOperand> MaskVec;
2479 bool Reverse = (NonZeros & 0x3) == 2;
2480 for (unsigned i = 0; i < 2; ++i)
2482 MaskVec.push_back(DAG.getConstant(1-i, EVT));
2484 MaskVec.push_back(DAG.getConstant(i, EVT));
2485 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
2486 for (unsigned i = 0; i < 2; ++i)
2488 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT));
2490 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT));
2491 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
2492 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask);
2495 if (Values.size() > 2) {
2496 // Expand into a number of unpckl*.
2498 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
2499 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
2500 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
2501 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG);
2502 for (unsigned i = 0; i < NumElems; ++i)
2503 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i));
2505 while (NumElems != 0) {
2506 for (unsigned i = 0; i < NumElems; ++i)
2507 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems],
2518 X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
2519 SDOperand V1 = Op.getOperand(0);
2520 SDOperand V2 = Op.getOperand(1);
2521 SDOperand PermMask = Op.getOperand(2);
2522 MVT::ValueType VT = Op.getValueType();
2523 unsigned NumElems = PermMask.getNumOperands();
2524 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
2525 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
2527 if (isSplatMask(PermMask.Val)) {
2528 if (NumElems <= 4) return Op;
2529 // Promote it to a v4i32 splat.
2530 return PromoteSplat(Op, DAG);
2533 if (X86::isMOVLMask(PermMask.Val))
2534 return (V1IsUndef) ? V2 : Op;
2536 if (X86::isMOVSHDUPMask(PermMask.Val) ||
2537 X86::isMOVSLDUPMask(PermMask.Val) ||
2538 X86::isMOVHLPSMask(PermMask.Val) ||
2539 X86::isMOVHPMask(PermMask.Val) ||
2540 X86::isMOVLPMask(PermMask.Val))
2543 if (ShouldXformToMOVHLPS(PermMask.Val) ||
2544 ShouldXformToMOVLP(V1.Val, PermMask.Val))
2545 return CommuteVectorShuffle(Op, DAG);
2547 bool V1IsSplat = isSplatVector(V1.Val) || V1.getOpcode() == ISD::UNDEF;
2548 bool V2IsSplat = isSplatVector(V2.Val) || V2.getOpcode() == ISD::UNDEF;
2549 if (V1IsSplat && !V2IsSplat) {
2550 Op = CommuteVectorShuffle(Op, DAG);
2551 V1 = Op.getOperand(0);
2552 V2 = Op.getOperand(1);
2553 PermMask = Op.getOperand(2);
2557 if (isCommutedMOVL(PermMask.Val, V2IsSplat)) {
2558 if (V2IsUndef) return V1;
2559 Op = CommuteVectorShuffle(Op, DAG);
2560 V1 = Op.getOperand(0);
2561 V2 = Op.getOperand(1);
2562 PermMask = Op.getOperand(2);
2564 // V2 is a splat, so the mask may be malformed. That is, it may point
2565 // to any V2 element. The instruction selectior won't like this. Get
2566 // a corrected mask and commute to form a proper MOVS{S|D}.
2567 SDOperand NewMask = getMOVLMask(NumElems, DAG);
2568 if (NewMask.Val != PermMask.Val)
2569 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2574 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) ||
2575 X86::isUNPCKLMask(PermMask.Val) ||
2576 X86::isUNPCKHMask(PermMask.Val))
2580 // Normalize mask so all entries that point to V2 points to its first
2581 // element then try to match unpck{h|l} again. If match, return a
2582 // new vector_shuffle with the corrected mask.
2583 SDOperand NewMask = NormalizeMask(PermMask, DAG);
2584 if (NewMask.Val != PermMask.Val) {
2585 if (X86::isUNPCKLMask(PermMask.Val, true)) {
2586 SDOperand NewMask = getUnpacklMask(NumElems, DAG);
2587 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2588 } else if (X86::isUNPCKHMask(PermMask.Val, true)) {
2589 SDOperand NewMask = getUnpackhMask(NumElems, DAG);
2590 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask);
2595 // Normalize the node to match x86 shuffle ops if needed
2596 if (V2.getOpcode() != ISD::UNDEF)
2597 if (isCommutedSHUFP(PermMask.Val)) {
2598 Op = CommuteVectorShuffle(Op, DAG);
2599 V1 = Op.getOperand(0);
2600 V2 = Op.getOperand(1);
2601 PermMask = Op.getOperand(2);
2604 // If VT is integer, try PSHUF* first, then SHUFP*.
2605 if (MVT::isInteger(VT)) {
2606 if (X86::isPSHUFDMask(PermMask.Val) ||
2607 X86::isPSHUFHWMask(PermMask.Val) ||
2608 X86::isPSHUFLWMask(PermMask.Val)) {
2609 if (V2.getOpcode() != ISD::UNDEF)
2610 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
2611 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
2615 if (X86::isSHUFPMask(PermMask.Val))
2618 // Handle v8i16 shuffle high / low shuffle node pair.
2619 if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) {
2620 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems);
2621 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2622 std::vector<SDOperand> MaskVec;
2623 for (unsigned i = 0; i != 4; ++i)
2624 MaskVec.push_back(PermMask.getOperand(i));
2625 for (unsigned i = 4; i != 8; ++i)
2626 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2627 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
2628 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2630 for (unsigned i = 0; i != 4; ++i)
2631 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2632 for (unsigned i = 4; i != 8; ++i)
2633 MaskVec.push_back(PermMask.getOperand(i));
2634 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec);
2635 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask);
2638 // Floating point cases in the other order.
2639 if (X86::isSHUFPMask(PermMask.Val))
2641 if (X86::isPSHUFDMask(PermMask.Val) ||
2642 X86::isPSHUFHWMask(PermMask.Val) ||
2643 X86::isPSHUFLWMask(PermMask.Val)) {
2644 if (V2.getOpcode() != ISD::UNDEF)
2645 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
2646 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
2651 if (NumElems == 4) {
2652 MVT::ValueType MaskVT = PermMask.getValueType();
2653 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT);
2654 std::vector<std::pair<int, int> > Locs;
2655 Locs.reserve(NumElems);
2656 std::vector<SDOperand> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2657 std::vector<SDOperand> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2660 // If no more than two elements come from either vector. This can be
2661 // implemented with two shuffles. First shuffle gather the elements.
2662 // The second shuffle, which takes the first shuffle as both of its
2663 // vector operands, put the elements into the right order.
2664 for (unsigned i = 0; i != NumElems; ++i) {
2665 SDOperand Elt = PermMask.getOperand(i);
2666 if (Elt.getOpcode() == ISD::UNDEF) {
2667 Locs[i] = std::make_pair(-1, -1);
2669 unsigned Val = cast<ConstantSDNode>(Elt)->getValue();
2670 if (Val < NumElems) {
2671 Locs[i] = std::make_pair(0, NumLo);
2675 Locs[i] = std::make_pair(1, NumHi);
2676 if (2+NumHi < NumElems)
2677 Mask1[2+NumHi] = Elt;
2682 if (NumLo <= 2 && NumHi <= 2) {
2683 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2684 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, Mask1));
2685 for (unsigned i = 0; i != NumElems; ++i) {
2686 if (Locs[i].first == -1)
2689 unsigned Idx = (i < NumElems/2) ? 0 : NumElems;
2690 Idx += Locs[i].first * (NumElems/2) + Locs[i].second;
2691 Mask2[i] = DAG.getConstant(Idx, MaskEVT);
2695 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1,
2696 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, Mask2));
2699 // Break it into (shuffle shuffle_hi, shuffle_lo).
2701 std::vector<SDOperand> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2702 std::vector<SDOperand> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT));
2703 std::vector<SDOperand> *MaskPtr = &LoMask;
2704 unsigned MaskIdx = 0;
2706 unsigned HiIdx = NumElems/2;
2707 for (unsigned i = 0; i != NumElems; ++i) {
2708 if (i == NumElems/2) {
2714 SDOperand Elt = PermMask.getOperand(i);
2715 if (Elt.getOpcode() == ISD::UNDEF) {
2716 Locs[i] = std::make_pair(-1, -1);
2717 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) {
2718 Locs[i] = std::make_pair(MaskIdx, LoIdx);
2719 (*MaskPtr)[LoIdx] = Elt;
2722 Locs[i] = std::make_pair(MaskIdx, HiIdx);
2723 (*MaskPtr)[HiIdx] = Elt;
2728 SDOperand LoShuffle =
2729 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2730 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, LoMask));
2731 SDOperand HiShuffle =
2732 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2,
2733 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, HiMask));
2734 std::vector<SDOperand> MaskOps;
2735 for (unsigned i = 0; i != NumElems; ++i) {
2736 if (Locs[i].first == -1) {
2737 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT));
2739 unsigned Idx = Locs[i].first * NumElems + Locs[i].second;
2740 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT));
2743 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle,
2744 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskOps));
2751 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2752 if (!isa<ConstantSDNode>(Op.getOperand(1)))
2755 MVT::ValueType VT = Op.getValueType();
2756 // TODO: handle v16i8.
2757 if (MVT::getSizeInBits(VT) == 16) {
2758 // Transform it so it match pextrw which produces a 32-bit result.
2759 MVT::ValueType EVT = (MVT::ValueType)(VT+1);
2760 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
2761 Op.getOperand(0), Op.getOperand(1));
2762 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract,
2763 DAG.getValueType(VT));
2764 return DAG.getNode(ISD::TRUNCATE, VT, Assert);
2765 } else if (MVT::getSizeInBits(VT) == 32) {
2766 SDOperand Vec = Op.getOperand(0);
2767 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
2771 // SHUFPS the element to the lowest double word, then movss.
2772 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2773 SDOperand IdxNode = DAG.getConstant((Idx < 2) ? Idx : Idx+4,
2774 MVT::getVectorBaseType(MaskVT));
2775 std::vector<SDOperand> IdxVec;
2776 IdxVec.push_back(DAG.getConstant(Idx, MVT::getVectorBaseType(MaskVT)));
2777 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2778 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2779 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2780 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, IdxVec);
2781 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
2783 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
2784 DAG.getConstant(0, MVT::i32));
2785 } else if (MVT::getSizeInBits(VT) == 64) {
2786 SDOperand Vec = Op.getOperand(0);
2787 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
2791 // UNPCKHPD the element to the lowest double word, then movsd.
2792 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
2793 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
2794 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2795 std::vector<SDOperand> IdxVec;
2796 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorBaseType(MaskVT)));
2797 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT)));
2798 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, IdxVec);
2799 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
2800 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
2801 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
2802 DAG.getConstant(0, MVT::i32));
2809 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2810 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
2811 // as its second argument.
2812 MVT::ValueType VT = Op.getValueType();
2813 MVT::ValueType BaseVT = MVT::getVectorBaseType(VT);
2814 SDOperand N0 = Op.getOperand(0);
2815 SDOperand N1 = Op.getOperand(1);
2816 SDOperand N2 = Op.getOperand(2);
2817 if (MVT::getSizeInBits(BaseVT) == 16) {
2818 if (N1.getValueType() != MVT::i32)
2819 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
2820 if (N2.getValueType() != MVT::i32)
2821 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(), MVT::i32);
2822 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
2823 } else if (MVT::getSizeInBits(BaseVT) == 32) {
2824 unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
2827 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
2828 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
2829 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT);
2830 std::vector<SDOperand> MaskVec;
2831 MaskVec.push_back(DAG.getConstant(4, BaseVT));
2832 for (unsigned i = 1; i <= 3; ++i)
2833 MaskVec.push_back(DAG.getConstant(i, BaseVT));
2834 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
2835 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, MaskVec));
2837 // Use two pinsrw instructions to insert a 32 bit value.
2839 if (MVT::isFloatingPoint(N1.getValueType())) {
2840 if (N1.getOpcode() == ISD::LOAD) {
2841 // Just load directly from f32mem to GR32.
2842 N1 = DAG.getLoad(MVT::i32, N1.getOperand(0), N1.getOperand(1),
2845 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
2846 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
2847 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
2848 DAG.getConstant(0, MVT::i32));
2851 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
2852 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
2853 DAG.getConstant(Idx, MVT::i32));
2854 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
2855 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
2856 DAG.getConstant(Idx+1, MVT::i32));
2857 return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
2865 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2866 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
2867 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt);
2870 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2871 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
2872 // one of the above mentioned nodes. It has to be wrapped because otherwise
2873 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2874 // be used to form addressing mode. These wrapped nodes will be selected
2877 X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
2878 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2879 SDOperand Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(),
2880 DAG.getTargetConstantPool(CP->get(), getPointerTy(),
2881 CP->getAlignment()));
2882 if (Subtarget->isTargetDarwin()) {
2883 // With PIC, the address is actually $g + Offset.
2884 if (getTargetMachine().getRelocationModel() == Reloc::PIC)
2885 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2886 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()), Result);
2893 X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
2894 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2895 SDOperand Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(),
2896 DAG.getTargetGlobalAddress(GV,
2898 if (Subtarget->isTargetDarwin()) {
2899 // With PIC, the address is actually $g + Offset.
2900 if (getTargetMachine().getRelocationModel() == Reloc::PIC)
2901 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2902 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
2905 // For Darwin, external and weak symbols are indirect, so we want to load
2906 // the value at address GV, not the value of GV itself. This means that
2907 // the GlobalAddress must be in the base or index register of the address,
2908 // not the GV offset field.
2909 if (getTargetMachine().getRelocationModel() != Reloc::Static &&
2910 DarwinGVRequiresExtraLoad(GV))
2911 Result = DAG.getLoad(MVT::i32, DAG.getEntryNode(),
2912 Result, DAG.getSrcValue(NULL));
2919 X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) {
2920 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
2921 SDOperand Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(),
2922 DAG.getTargetExternalSymbol(Sym,
2924 if (Subtarget->isTargetDarwin()) {
2925 // With PIC, the address is actually $g + Offset.
2926 if (getTargetMachine().getRelocationModel() == Reloc::PIC)
2927 Result = DAG.getNode(ISD::ADD, getPointerTy(),
2928 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
2935 SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) {
2936 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
2937 "Not an i64 shift!");
2938 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
2939 SDOperand ShOpLo = Op.getOperand(0);
2940 SDOperand ShOpHi = Op.getOperand(1);
2941 SDOperand ShAmt = Op.getOperand(2);
2942 SDOperand Tmp1 = isSRA ? DAG.getNode(ISD::SRA, MVT::i32, ShOpHi,
2943 DAG.getConstant(31, MVT::i8))
2944 : DAG.getConstant(0, MVT::i32);
2946 SDOperand Tmp2, Tmp3;
2947 if (Op.getOpcode() == ISD::SHL_PARTS) {
2948 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
2949 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
2951 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
2952 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
2955 SDOperand InFlag = DAG.getNode(X86ISD::TEST, MVT::Flag,
2956 ShAmt, DAG.getConstant(32, MVT::i8));
2959 SDOperand CC = DAG.getConstant(X86ISD::COND_NE, MVT::i8);
2961 std::vector<MVT::ValueType> Tys;
2962 Tys.push_back(MVT::i32);
2963 Tys.push_back(MVT::Flag);
2964 std::vector<SDOperand> Ops;
2965 if (Op.getOpcode() == ISD::SHL_PARTS) {
2966 Ops.push_back(Tmp2);
2967 Ops.push_back(Tmp3);
2969 Ops.push_back(InFlag);
2970 Hi = DAG.getNode(X86ISD::CMOV, Tys, Ops);
2971 InFlag = Hi.getValue(1);
2974 Ops.push_back(Tmp3);
2975 Ops.push_back(Tmp1);
2977 Ops.push_back(InFlag);
2978 Lo = DAG.getNode(X86ISD::CMOV, Tys, Ops);
2980 Ops.push_back(Tmp2);
2981 Ops.push_back(Tmp3);
2983 Ops.push_back(InFlag);
2984 Lo = DAG.getNode(X86ISD::CMOV, Tys, Ops);
2985 InFlag = Lo.getValue(1);
2988 Ops.push_back(Tmp3);
2989 Ops.push_back(Tmp1);
2991 Ops.push_back(InFlag);
2992 Hi = DAG.getNode(X86ISD::CMOV, Tys, Ops);
2996 Tys.push_back(MVT::i32);
2997 Tys.push_back(MVT::i32);
3001 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops);
3004 SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
3005 assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
3006 Op.getOperand(0).getValueType() >= MVT::i16 &&
3007 "Unknown SINT_TO_FP to lower!");
3010 MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
3011 unsigned Size = MVT::getSizeInBits(SrcVT)/8;
3012 MachineFunction &MF = DAG.getMachineFunction();
3013 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
3014 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3015 SDOperand Chain = DAG.getNode(ISD::STORE, MVT::Other,
3016 DAG.getEntryNode(), Op.getOperand(0),
3017 StackSlot, DAG.getSrcValue(NULL));
3020 std::vector<MVT::ValueType> Tys;
3021 Tys.push_back(MVT::f64);
3022 Tys.push_back(MVT::Other);
3023 if (X86ScalarSSE) Tys.push_back(MVT::Flag);
3024 std::vector<SDOperand> Ops;
3025 Ops.push_back(Chain);
3026 Ops.push_back(StackSlot);
3027 Ops.push_back(DAG.getValueType(SrcVT));
3028 Result = DAG.getNode(X86ScalarSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
3032 Chain = Result.getValue(1);
3033 SDOperand InFlag = Result.getValue(2);
3035 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
3036 // shouldn't be necessary except that RFP cannot be live across
3037 // multiple blocks. When stackifier is fixed, they can be uncoupled.
3038 MachineFunction &MF = DAG.getMachineFunction();
3039 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3040 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3041 std::vector<MVT::ValueType> Tys;
3042 Tys.push_back(MVT::Other);
3043 std::vector<SDOperand> Ops;
3044 Ops.push_back(Chain);
3045 Ops.push_back(Result);
3046 Ops.push_back(StackSlot);
3047 Ops.push_back(DAG.getValueType(Op.getValueType()));
3048 Ops.push_back(InFlag);
3049 Chain = DAG.getNode(X86ISD::FST, Tys, Ops);
3050 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot,
3051 DAG.getSrcValue(NULL));
3057 SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
3058 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
3059 "Unknown FP_TO_SINT to lower!");
3060 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
3062 MachineFunction &MF = DAG.getMachineFunction();
3063 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
3064 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3065 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3068 switch (Op.getValueType()) {
3069 default: assert(0 && "Invalid FP_TO_SINT to lower!");
3070 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
3071 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
3072 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
3075 SDOperand Chain = DAG.getEntryNode();
3076 SDOperand Value = Op.getOperand(0);
3078 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
3079 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value, StackSlot,
3080 DAG.getSrcValue(0));
3081 std::vector<MVT::ValueType> Tys;
3082 Tys.push_back(MVT::f64);
3083 Tys.push_back(MVT::Other);
3084 std::vector<SDOperand> Ops;
3085 Ops.push_back(Chain);
3086 Ops.push_back(StackSlot);
3087 Ops.push_back(DAG.getValueType(Op.getOperand(0).getValueType()));
3088 Value = DAG.getNode(X86ISD::FLD, Tys, Ops);
3089 Chain = Value.getValue(1);
3090 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
3091 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
3094 // Build the FP_TO_INT*_IN_MEM
3095 std::vector<SDOperand> Ops;
3096 Ops.push_back(Chain);
3097 Ops.push_back(Value);
3098 Ops.push_back(StackSlot);
3099 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops);
3102 return DAG.getLoad(Op.getValueType(), FIST, StackSlot,
3103 DAG.getSrcValue(NULL));
3106 SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) {
3107 MVT::ValueType VT = Op.getValueType();
3108 const Type *OpNTy = MVT::getTypeForValueType(VT);
3109 std::vector<Constant*> CV;
3110 if (VT == MVT::f64) {
3111 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(~(1ULL << 63))));
3112 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3114 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(~(1U << 31))));
3115 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3116 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3117 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3119 Constant *CS = ConstantStruct::get(CV);
3120 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3122 = DAG.getNode(X86ISD::LOAD_PACK,
3123 VT, DAG.getEntryNode(), CPIdx, DAG.getSrcValue(NULL));
3124 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
3127 SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) {
3128 MVT::ValueType VT = Op.getValueType();
3129 const Type *OpNTy = MVT::getTypeForValueType(VT);
3130 std::vector<Constant*> CV;
3131 if (VT == MVT::f64) {
3132 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(1ULL << 63)));
3133 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3135 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(1U << 31)));
3136 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3137 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3138 CV.push_back(ConstantFP::get(OpNTy, 0.0));
3140 Constant *CS = ConstantStruct::get(CV);
3141 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
3142 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK,
3143 VT, DAG.getEntryNode(), CPIdx, DAG.getSrcValue(NULL));
3144 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
3147 SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
3148 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
3150 SDOperand CC = Op.getOperand(2);
3151 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3152 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
3155 if (translateX86CC(CC, isFP, X86CC, Flip)) {
3157 Cond = DAG.getNode(X86ISD::CMP, MVT::Flag,
3158 Op.getOperand(1), Op.getOperand(0));
3160 Cond = DAG.getNode(X86ISD::CMP, MVT::Flag,
3161 Op.getOperand(0), Op.getOperand(1));
3162 return DAG.getNode(X86ISD::SETCC, MVT::i8,
3163 DAG.getConstant(X86CC, MVT::i8), Cond);
3165 assert(isFP && "Illegal integer SetCC!");
3167 Cond = DAG.getNode(X86ISD::CMP, MVT::Flag,
3168 Op.getOperand(0), Op.getOperand(1));
3169 std::vector<MVT::ValueType> Tys;
3170 std::vector<SDOperand> Ops;
3171 switch (SetCCOpcode) {
3172 default: assert(false && "Illegal floating point SetCC!");
3173 case ISD::SETOEQ: { // !PF & ZF
3174 Tys.push_back(MVT::i8);
3175 Tys.push_back(MVT::Flag);
3176 Ops.push_back(DAG.getConstant(X86ISD::COND_NP, MVT::i8));
3177 Ops.push_back(Cond);
3178 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, Tys, Ops);
3179 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
3180 DAG.getConstant(X86ISD::COND_E, MVT::i8),
3182 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
3184 case ISD::SETUNE: { // PF | !ZF
3185 Tys.push_back(MVT::i8);
3186 Tys.push_back(MVT::Flag);
3187 Ops.push_back(DAG.getConstant(X86ISD::COND_P, MVT::i8));
3188 Ops.push_back(Cond);
3189 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, Tys, Ops);
3190 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
3191 DAG.getConstant(X86ISD::COND_NE, MVT::i8),
3193 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
3199 SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) {
3200 MVT::ValueType VT = Op.getValueType();
3201 bool isFPStack = MVT::isFloatingPoint(VT) && !X86ScalarSSE;
3202 bool addTest = false;
3203 SDOperand Op0 = Op.getOperand(0);
3205 if (Op0.getOpcode() == ISD::SETCC)
3206 Op0 = LowerOperation(Op0, DAG);
3208 if (Op0.getOpcode() == X86ISD::SETCC) {
3209 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3210 // (since flag operand cannot be shared). If the X86ISD::SETCC does not
3211 // have another use it will be eliminated.
3212 // If the X86ISD::SETCC has more than one use, then it's probably better
3213 // to use a test instead of duplicating the X86ISD::CMP (for register
3214 // pressure reason).
3215 unsigned CmpOpc = Op0.getOperand(1).getOpcode();
3216 if (CmpOpc == X86ISD::CMP || CmpOpc == X86ISD::COMI ||
3217 CmpOpc == X86ISD::UCOMI) {
3218 if (!Op0.hasOneUse()) {
3219 std::vector<MVT::ValueType> Tys;
3220 for (unsigned i = 0; i < Op0.Val->getNumValues(); ++i)
3221 Tys.push_back(Op0.Val->getValueType(i));
3222 std::vector<SDOperand> Ops;
3223 for (unsigned i = 0; i < Op0.getNumOperands(); ++i)
3224 Ops.push_back(Op0.getOperand(i));
3225 Op0 = DAG.getNode(X86ISD::SETCC, Tys, Ops);
3228 CC = Op0.getOperand(0);
3229 Cond = Op0.getOperand(1);
3230 // Make a copy as flag result cannot be used by more than one.
3231 Cond = DAG.getNode(CmpOpc, MVT::Flag,
3232 Cond.getOperand(0), Cond.getOperand(1));
3234 isFPStack && !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
3241 CC = DAG.getConstant(X86ISD::COND_NE, MVT::i8);
3242 Cond = DAG.getNode(X86ISD::TEST, MVT::Flag, Op0, Op0);
3245 std::vector<MVT::ValueType> Tys;
3246 Tys.push_back(Op.getValueType());
3247 Tys.push_back(MVT::Flag);
3248 std::vector<SDOperand> Ops;
3249 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
3250 // condition is true.
3251 Ops.push_back(Op.getOperand(2));
3252 Ops.push_back(Op.getOperand(1));
3254 Ops.push_back(Cond);
3255 return DAG.getNode(X86ISD::CMOV, Tys, Ops);
3258 SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) {
3259 bool addTest = false;
3260 SDOperand Cond = Op.getOperand(1);
3261 SDOperand Dest = Op.getOperand(2);
3263 if (Cond.getOpcode() == ISD::SETCC)
3264 Cond = LowerOperation(Cond, DAG);
3266 if (Cond.getOpcode() == X86ISD::SETCC) {
3267 // If condition flag is set by a X86ISD::CMP, then make a copy of it
3268 // (since flag operand cannot be shared). If the X86ISD::SETCC does not
3269 // have another use it will be eliminated.
3270 // If the X86ISD::SETCC has more than one use, then it's probably better
3271 // to use a test instead of duplicating the X86ISD::CMP (for register
3272 // pressure reason).
3273 unsigned CmpOpc = Cond.getOperand(1).getOpcode();
3274 if (CmpOpc == X86ISD::CMP || CmpOpc == X86ISD::COMI ||
3275 CmpOpc == X86ISD::UCOMI) {
3276 if (!Cond.hasOneUse()) {
3277 std::vector<MVT::ValueType> Tys;
3278 for (unsigned i = 0; i < Cond.Val->getNumValues(); ++i)
3279 Tys.push_back(Cond.Val->getValueType(i));
3280 std::vector<SDOperand> Ops;
3281 for (unsigned i = 0; i < Cond.getNumOperands(); ++i)
3282 Ops.push_back(Cond.getOperand(i));
3283 Cond = DAG.getNode(X86ISD::SETCC, Tys, Ops);
3286 CC = Cond.getOperand(0);
3287 Cond = Cond.getOperand(1);
3288 // Make a copy as flag result cannot be used by more than one.
3289 Cond = DAG.getNode(CmpOpc, MVT::Flag,
3290 Cond.getOperand(0), Cond.getOperand(1));
3297 CC = DAG.getConstant(X86ISD::COND_NE, MVT::i8);
3298 Cond = DAG.getNode(X86ISD::TEST, MVT::Flag, Cond, Cond);
3300 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
3301 Op.getOperand(0), Op.getOperand(2), CC, Cond);
3304 SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
3305 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3306 SDOperand Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(),
3307 DAG.getTargetJumpTable(JT->getIndex(),
3309 if (Subtarget->isTargetDarwin()) {
3310 // With PIC, the address is actually $g + Offset.
3311 if (getTargetMachine().getRelocationModel() == Reloc::PIC)
3312 Result = DAG.getNode(ISD::ADD, getPointerTy(),
3313 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
3320 SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
3321 unsigned CallingConv= cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3322 if (CallingConv == CallingConv::Fast && EnableFastCC)
3323 return LowerFastCCCallTo(Op, DAG);
3325 return LowerCCCCallTo(Op, DAG);
3328 SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) {
3331 switch(Op.getNumOperands()) {
3333 assert(0 && "Do not know how to return this many arguments!");
3335 case 1: // ret void.
3336 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Op.getOperand(0),
3337 DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
3339 MVT::ValueType ArgVT = Op.getOperand(1).getValueType();
3341 if (MVT::isVector(ArgVT)) {
3342 // Integer or FP vector result -> XMM0.
3343 if (DAG.getMachineFunction().liveout_empty())
3344 DAG.getMachineFunction().addLiveOut(X86::XMM0);
3345 Copy = DAG.getCopyToReg(Op.getOperand(0), X86::XMM0, Op.getOperand(1),
3347 } else if (MVT::isInteger(ArgVT)) {
3348 // Integer result -> EAX
3349 if (DAG.getMachineFunction().liveout_empty())
3350 DAG.getMachineFunction().addLiveOut(X86::EAX);
3352 Copy = DAG.getCopyToReg(Op.getOperand(0), X86::EAX, Op.getOperand(1),
3354 } else if (!X86ScalarSSE) {
3355 // FP return with fp-stack value.
3356 if (DAG.getMachineFunction().liveout_empty())
3357 DAG.getMachineFunction().addLiveOut(X86::ST0);
3359 std::vector<MVT::ValueType> Tys;
3360 Tys.push_back(MVT::Other);
3361 Tys.push_back(MVT::Flag);
3362 std::vector<SDOperand> Ops;
3363 Ops.push_back(Op.getOperand(0));
3364 Ops.push_back(Op.getOperand(1));
3365 Copy = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops);
3367 // FP return with ScalarSSE (return on fp-stack).
3368 if (DAG.getMachineFunction().liveout_empty())
3369 DAG.getMachineFunction().addLiveOut(X86::ST0);
3372 SDOperand Chain = Op.getOperand(0);
3373 SDOperand Value = Op.getOperand(1);
3375 if (Value.getOpcode() == ISD::LOAD &&
3376 (Chain == Value.getValue(1) || Chain == Value.getOperand(0))) {
3377 Chain = Value.getOperand(0);
3378 MemLoc = Value.getOperand(1);
3380 // Spill the value to memory and reload it into top of stack.
3381 unsigned Size = MVT::getSizeInBits(ArgVT)/8;
3382 MachineFunction &MF = DAG.getMachineFunction();
3383 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
3384 MemLoc = DAG.getFrameIndex(SSFI, getPointerTy());
3385 Chain = DAG.getNode(ISD::STORE, MVT::Other, Op.getOperand(0),
3386 Value, MemLoc, DAG.getSrcValue(0));
3388 std::vector<MVT::ValueType> Tys;
3389 Tys.push_back(MVT::f64);
3390 Tys.push_back(MVT::Other);
3391 std::vector<SDOperand> Ops;
3392 Ops.push_back(Chain);
3393 Ops.push_back(MemLoc);
3394 Ops.push_back(DAG.getValueType(ArgVT));
3395 Copy = DAG.getNode(X86ISD::FLD, Tys, Ops);
3397 Tys.push_back(MVT::Other);
3398 Tys.push_back(MVT::Flag);
3400 Ops.push_back(Copy.getValue(1));
3401 Ops.push_back(Copy);
3402 Copy = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops);
3407 if (DAG.getMachineFunction().liveout_empty()) {
3408 DAG.getMachineFunction().addLiveOut(X86::EAX);
3409 DAG.getMachineFunction().addLiveOut(X86::EDX);
3412 Copy = DAG.getCopyToReg(Op.getOperand(0), X86::EDX, Op.getOperand(2),
3414 Copy = DAG.getCopyToReg(Copy, X86::EAX,Op.getOperand(1),Copy.getValue(1));
3417 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other,
3418 Copy, DAG.getConstant(getBytesToPopOnReturn(), MVT::i16),
3423 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
3424 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
3425 if (CC == CallingConv::Fast && EnableFastCC)
3426 return LowerFastCCArguments(Op, DAG);
3428 return LowerCCCArguments(Op, DAG);
3431 SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) {
3432 SDOperand InFlag(0, 0);
3433 SDOperand Chain = Op.getOperand(0);
3435 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3436 if (Align == 0) Align = 1;
3438 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3439 // If not DWORD aligned, call memset if size is less than the threshold.
3440 // It knows how to align to the right boundary first.
3441 if ((Align & 3) != 0 ||
3442 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
3443 MVT::ValueType IntPtr = getPointerTy();
3444 const Type *IntPtrTy = getTargetData()->getIntPtrType();
3445 std::vector<std::pair<SDOperand, const Type*> > Args;
3446 Args.push_back(std::make_pair(Op.getOperand(1), IntPtrTy));
3447 // Extend the ubyte argument to be an int value for the call.
3448 SDOperand Val = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
3449 Args.push_back(std::make_pair(Val, IntPtrTy));
3450 Args.push_back(std::make_pair(Op.getOperand(3), IntPtrTy));
3451 std::pair<SDOperand,SDOperand> CallResult =
3452 LowerCallTo(Chain, Type::VoidTy, false, CallingConv::C, false,
3453 DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
3454 return CallResult.second;
3459 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
3460 unsigned BytesLeft = 0;
3461 bool TwoRepStos = false;
3464 unsigned Val = ValC->getValue() & 255;
3466 // If the value is a constant, then we can potentially use larger sets.
3467 switch (Align & 3) {
3468 case 2: // WORD aligned
3470 Count = DAG.getConstant(I->getValue() / 2, MVT::i32);
3471 BytesLeft = I->getValue() % 2;
3472 Val = (Val << 8) | Val;
3475 case 0: // DWORD aligned
3478 Count = DAG.getConstant(I->getValue() / 4, MVT::i32);
3479 BytesLeft = I->getValue() % 4;
3481 Count = DAG.getNode(ISD::SRL, MVT::i32, Op.getOperand(3),
3482 DAG.getConstant(2, MVT::i8));
3485 Val = (Val << 8) | Val;
3486 Val = (Val << 16) | Val;
3489 default: // Byte aligned
3491 Count = Op.getOperand(3);
3496 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
3498 InFlag = Chain.getValue(1);
3501 Count = Op.getOperand(3);
3502 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
3503 InFlag = Chain.getValue(1);
3506 Chain = DAG.getCopyToReg(Chain, X86::ECX, Count, InFlag);
3507 InFlag = Chain.getValue(1);
3508 Chain = DAG.getCopyToReg(Chain, X86::EDI, Op.getOperand(1), InFlag);
3509 InFlag = Chain.getValue(1);
3511 std::vector<MVT::ValueType> Tys;
3512 Tys.push_back(MVT::Other);
3513 Tys.push_back(MVT::Flag);
3514 std::vector<SDOperand> Ops;
3515 Ops.push_back(Chain);
3516 Ops.push_back(DAG.getValueType(AVT));
3517 Ops.push_back(InFlag);
3518 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, Ops);
3521 InFlag = Chain.getValue(1);
3522 Count = Op.getOperand(3);
3523 MVT::ValueType CVT = Count.getValueType();
3524 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
3525 DAG.getConstant(3, CVT));
3526 Chain = DAG.getCopyToReg(Chain, X86::ECX, Left, InFlag);
3527 InFlag = Chain.getValue(1);
3529 Tys.push_back(MVT::Other);
3530 Tys.push_back(MVT::Flag);
3532 Ops.push_back(Chain);
3533 Ops.push_back(DAG.getValueType(MVT::i8));
3534 Ops.push_back(InFlag);
3535 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, Ops);
3536 } else if (BytesLeft) {
3537 // Issue stores for the last 1 - 3 bytes.
3539 unsigned Val = ValC->getValue() & 255;
3540 unsigned Offset = I->getValue() - BytesLeft;
3541 SDOperand DstAddr = Op.getOperand(1);
3542 MVT::ValueType AddrVT = DstAddr.getValueType();
3543 if (BytesLeft >= 2) {
3544 Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
3545 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value,
3546 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3547 DAG.getConstant(Offset, AddrVT)),
3548 DAG.getSrcValue(NULL));
3553 if (BytesLeft == 1) {
3554 Value = DAG.getConstant(Val, MVT::i8);
3555 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value,
3556 DAG.getNode(ISD::ADD, AddrVT, DstAddr,
3557 DAG.getConstant(Offset, AddrVT)),
3558 DAG.getSrcValue(NULL));
3565 SDOperand X86TargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) {
3566 SDOperand Chain = Op.getOperand(0);
3568 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
3569 if (Align == 0) Align = 1;
3571 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
3572 // If not DWORD aligned, call memcpy if size is less than the threshold.
3573 // It knows how to align to the right boundary first.
3574 if ((Align & 3) != 0 ||
3575 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) {
3576 MVT::ValueType IntPtr = getPointerTy();
3577 const Type *IntPtrTy = getTargetData()->getIntPtrType();
3578 std::vector<std::pair<SDOperand, const Type*> > Args;
3579 Args.push_back(std::make_pair(Op.getOperand(1), IntPtrTy));
3580 Args.push_back(std::make_pair(Op.getOperand(2), IntPtrTy));
3581 Args.push_back(std::make_pair(Op.getOperand(3), IntPtrTy));
3582 std::pair<SDOperand,SDOperand> CallResult =
3583 LowerCallTo(Chain, Type::VoidTy, false, CallingConv::C, false,
3584 DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
3585 return CallResult.second;
3590 unsigned BytesLeft = 0;
3591 bool TwoRepMovs = false;
3592 switch (Align & 3) {
3593 case 2: // WORD aligned
3595 Count = DAG.getConstant(I->getValue() / 2, MVT::i32);
3596 BytesLeft = I->getValue() % 2;
3598 case 0: // DWORD aligned
3601 Count = DAG.getConstant(I->getValue() / 4, MVT::i32);
3602 BytesLeft = I->getValue() % 4;
3604 Count = DAG.getNode(ISD::SRL, MVT::i32, Op.getOperand(3),
3605 DAG.getConstant(2, MVT::i8));
3609 default: // Byte aligned
3611 Count = Op.getOperand(3);
3615 SDOperand InFlag(0, 0);
3616 Chain = DAG.getCopyToReg(Chain, X86::ECX, Count, InFlag);
3617 InFlag = Chain.getValue(1);
3618 Chain = DAG.getCopyToReg(Chain, X86::EDI, Op.getOperand(1), InFlag);
3619 InFlag = Chain.getValue(1);
3620 Chain = DAG.getCopyToReg(Chain, X86::ESI, Op.getOperand(2), InFlag);
3621 InFlag = Chain.getValue(1);
3623 std::vector<MVT::ValueType> Tys;
3624 Tys.push_back(MVT::Other);
3625 Tys.push_back(MVT::Flag);
3626 std::vector<SDOperand> Ops;
3627 Ops.push_back(Chain);
3628 Ops.push_back(DAG.getValueType(AVT));
3629 Ops.push_back(InFlag);
3630 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, Ops);
3633 InFlag = Chain.getValue(1);
3634 Count = Op.getOperand(3);
3635 MVT::ValueType CVT = Count.getValueType();
3636 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count,
3637 DAG.getConstant(3, CVT));
3638 Chain = DAG.getCopyToReg(Chain, X86::ECX, Left, InFlag);
3639 InFlag = Chain.getValue(1);
3641 Tys.push_back(MVT::Other);
3642 Tys.push_back(MVT::Flag);
3644 Ops.push_back(Chain);
3645 Ops.push_back(DAG.getValueType(MVT::i8));
3646 Ops.push_back(InFlag);
3647 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, Ops);
3648 } else if (BytesLeft) {
3649 // Issue loads and stores for the last 1 - 3 bytes.
3650 unsigned Offset = I->getValue() - BytesLeft;
3651 SDOperand DstAddr = Op.getOperand(1);
3652 MVT::ValueType DstVT = DstAddr.getValueType();
3653 SDOperand SrcAddr = Op.getOperand(2);
3654 MVT::ValueType SrcVT = SrcAddr.getValueType();
3656 if (BytesLeft >= 2) {
3657 Value = DAG.getLoad(MVT::i16, Chain,
3658 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3659 DAG.getConstant(Offset, SrcVT)),
3660 DAG.getSrcValue(NULL));
3661 Chain = Value.getValue(1);
3662 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value,
3663 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3664 DAG.getConstant(Offset, DstVT)),
3665 DAG.getSrcValue(NULL));
3670 if (BytesLeft == 1) {
3671 Value = DAG.getLoad(MVT::i8, Chain,
3672 DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
3673 DAG.getConstant(Offset, SrcVT)),
3674 DAG.getSrcValue(NULL));
3675 Chain = Value.getValue(1);
3676 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value,
3677 DAG.getNode(ISD::ADD, DstVT, DstAddr,
3678 DAG.getConstant(Offset, DstVT)),
3679 DAG.getSrcValue(NULL));
3687 X86TargetLowering::LowerREADCYCLCECOUNTER(SDOperand Op, SelectionDAG &DAG) {
3688 std::vector<MVT::ValueType> Tys;
3689 Tys.push_back(MVT::Other);
3690 Tys.push_back(MVT::Flag);
3691 std::vector<SDOperand> Ops;
3692 Ops.push_back(Op.getOperand(0));
3693 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, Ops);
3695 Ops.push_back(DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1)));
3696 Ops.push_back(DAG.getCopyFromReg(Ops[0].getValue(1), X86::EDX,
3697 MVT::i32, Ops[0].getValue(2)));
3698 Ops.push_back(Ops[1].getValue(1));
3699 Tys[0] = Tys[1] = MVT::i32;
3700 Tys.push_back(MVT::Other);
3701 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops);
3704 SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) {
3705 // vastart just stores the address of the VarArgsFrameIndex slot into the
3706 // memory location argument.
3707 // FIXME: Replace MVT::i32 with PointerTy
3708 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
3709 return DAG.getNode(ISD::STORE, MVT::Other, Op.getOperand(0), FR,
3710 Op.getOperand(1), Op.getOperand(2));
3714 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
3715 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
3717 default: return SDOperand(); // Don't custom lower most intrinsics.
3718 // Comparison intrinsics.
3719 case Intrinsic::x86_sse_comieq_ss:
3720 case Intrinsic::x86_sse_comilt_ss:
3721 case Intrinsic::x86_sse_comile_ss:
3722 case Intrinsic::x86_sse_comigt_ss:
3723 case Intrinsic::x86_sse_comige_ss:
3724 case Intrinsic::x86_sse_comineq_ss:
3725 case Intrinsic::x86_sse_ucomieq_ss:
3726 case Intrinsic::x86_sse_ucomilt_ss:
3727 case Intrinsic::x86_sse_ucomile_ss:
3728 case Intrinsic::x86_sse_ucomigt_ss:
3729 case Intrinsic::x86_sse_ucomige_ss:
3730 case Intrinsic::x86_sse_ucomineq_ss:
3731 case Intrinsic::x86_sse2_comieq_sd:
3732 case Intrinsic::x86_sse2_comilt_sd:
3733 case Intrinsic::x86_sse2_comile_sd:
3734 case Intrinsic::x86_sse2_comigt_sd:
3735 case Intrinsic::x86_sse2_comige_sd:
3736 case Intrinsic::x86_sse2_comineq_sd:
3737 case Intrinsic::x86_sse2_ucomieq_sd:
3738 case Intrinsic::x86_sse2_ucomilt_sd:
3739 case Intrinsic::x86_sse2_ucomile_sd:
3740 case Intrinsic::x86_sse2_ucomigt_sd:
3741 case Intrinsic::x86_sse2_ucomige_sd:
3742 case Intrinsic::x86_sse2_ucomineq_sd: {
3744 ISD::CondCode CC = ISD::SETCC_INVALID;
3747 case Intrinsic::x86_sse_comieq_ss:
3748 case Intrinsic::x86_sse2_comieq_sd:
3752 case Intrinsic::x86_sse_comilt_ss:
3753 case Intrinsic::x86_sse2_comilt_sd:
3757 case Intrinsic::x86_sse_comile_ss:
3758 case Intrinsic::x86_sse2_comile_sd:
3762 case Intrinsic::x86_sse_comigt_ss:
3763 case Intrinsic::x86_sse2_comigt_sd:
3767 case Intrinsic::x86_sse_comige_ss:
3768 case Intrinsic::x86_sse2_comige_sd:
3772 case Intrinsic::x86_sse_comineq_ss:
3773 case Intrinsic::x86_sse2_comineq_sd:
3777 case Intrinsic::x86_sse_ucomieq_ss:
3778 case Intrinsic::x86_sse2_ucomieq_sd:
3779 Opc = X86ISD::UCOMI;
3782 case Intrinsic::x86_sse_ucomilt_ss:
3783 case Intrinsic::x86_sse2_ucomilt_sd:
3784 Opc = X86ISD::UCOMI;
3787 case Intrinsic::x86_sse_ucomile_ss:
3788 case Intrinsic::x86_sse2_ucomile_sd:
3789 Opc = X86ISD::UCOMI;
3792 case Intrinsic::x86_sse_ucomigt_ss:
3793 case Intrinsic::x86_sse2_ucomigt_sd:
3794 Opc = X86ISD::UCOMI;
3797 case Intrinsic::x86_sse_ucomige_ss:
3798 case Intrinsic::x86_sse2_ucomige_sd:
3799 Opc = X86ISD::UCOMI;
3802 case Intrinsic::x86_sse_ucomineq_ss:
3803 case Intrinsic::x86_sse2_ucomineq_sd:
3804 Opc = X86ISD::UCOMI;
3810 translateX86CC(CC, true, X86CC, Flip);
3811 SDOperand Cond = DAG.getNode(Opc, MVT::Flag, Op.getOperand(Flip?2:1),
3812 Op.getOperand(Flip?1:2));
3813 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, MVT::i8,
3814 DAG.getConstant(X86CC, MVT::i8), Cond);
3815 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC);
3820 /// LowerOperation - Provide custom lowering hooks for some operations.
3822 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
3823 switch (Op.getOpcode()) {
3824 default: assert(0 && "Should not custom lower this!");
3825 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3826 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3827 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3828 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
3829 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
3830 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3831 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
3832 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
3833 case ISD::SHL_PARTS:
3834 case ISD::SRA_PARTS:
3835 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
3836 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
3837 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
3838 case ISD::FABS: return LowerFABS(Op, DAG);
3839 case ISD::FNEG: return LowerFNEG(Op, DAG);
3840 case ISD::SETCC: return LowerSETCC(Op, DAG);
3841 case ISD::SELECT: return LowerSELECT(Op, DAG);
3842 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
3843 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
3844 case ISD::CALL: return LowerCALL(Op, DAG);
3845 case ISD::RET: return LowerRET(Op, DAG);
3846 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
3847 case ISD::MEMSET: return LowerMEMSET(Op, DAG);
3848 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG);
3849 case ISD::READCYCLECOUNTER: return LowerREADCYCLCECOUNTER(Op, DAG);
3850 case ISD::VASTART: return LowerVASTART(Op, DAG);
3851 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
3855 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
3857 default: return NULL;
3858 case X86ISD::SHLD: return "X86ISD::SHLD";
3859 case X86ISD::SHRD: return "X86ISD::SHRD";
3860 case X86ISD::FAND: return "X86ISD::FAND";
3861 case X86ISD::FXOR: return "X86ISD::FXOR";
3862 case X86ISD::FILD: return "X86ISD::FILD";
3863 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
3864 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
3865 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
3866 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
3867 case X86ISD::FLD: return "X86ISD::FLD";
3868 case X86ISD::FST: return "X86ISD::FST";
3869 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT";
3870 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT";
3871 case X86ISD::CALL: return "X86ISD::CALL";
3872 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
3873 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
3874 case X86ISD::CMP: return "X86ISD::CMP";
3875 case X86ISD::TEST: return "X86ISD::TEST";
3876 case X86ISD::COMI: return "X86ISD::COMI";
3877 case X86ISD::UCOMI: return "X86ISD::UCOMI";
3878 case X86ISD::SETCC: return "X86ISD::SETCC";
3879 case X86ISD::CMOV: return "X86ISD::CMOV";
3880 case X86ISD::BRCOND: return "X86ISD::BRCOND";
3881 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
3882 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
3883 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
3884 case X86ISD::LOAD_PACK: return "X86ISD::LOAD_PACK";
3885 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
3886 case X86ISD::Wrapper: return "X86ISD::Wrapper";
3887 case X86ISD::S2VEC: return "X86ISD::S2VEC";
3888 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
3889 case X86ISD::PINSRW: return "X86ISD::PINSRW";
3893 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
3895 uint64_t &KnownZero,
3897 unsigned Depth) const {
3898 unsigned Opc = Op.getOpcode();
3899 assert((Opc >= ISD::BUILTIN_OP_END ||
3900 Opc == ISD::INTRINSIC_WO_CHAIN ||
3901 Opc == ISD::INTRINSIC_W_CHAIN ||
3902 Opc == ISD::INTRINSIC_VOID) &&
3903 "Should use MaskedValueIsZero if you don't know whether Op"
3904 " is a target node!");
3906 KnownZero = KnownOne = 0; // Don't know anything.
3910 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL);
3915 std::vector<unsigned> X86TargetLowering::
3916 getRegClassForInlineAsmConstraint(const std::string &Constraint,
3917 MVT::ValueType VT) const {
3918 if (Constraint.size() == 1) {
3919 // FIXME: not handling fp-stack yet!
3920 // FIXME: not handling MMX registers yet ('y' constraint).
3921 switch (Constraint[0]) { // GCC X86 Constraint Letters
3922 default: break; // Unknown constriant letter
3923 case 'r': // GENERAL_REGS
3924 case 'R': // LEGACY_REGS
3926 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
3927 X86::ESI, X86::EDI, X86::EBP, X86::ESP, 0);
3928 else if (VT == MVT::i16)
3929 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
3930 X86::SI, X86::DI, X86::BP, X86::SP, 0);
3931 else if (VT == MVT::i8)
3932 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
3934 case 'l': // INDEX_REGS
3936 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
3937 X86::ESI, X86::EDI, X86::EBP, 0);
3938 else if (VT == MVT::i16)
3939 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
3940 X86::SI, X86::DI, X86::BP, 0);
3941 else if (VT == MVT::i8)
3942 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
3944 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode)
3947 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
3948 else if (VT == MVT::i16)
3949 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
3950 else if (VT == MVT::i8)
3951 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0);
3953 case 'x': // SSE_REGS if SSE1 allowed
3954 if (Subtarget->hasSSE1())
3955 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3956 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
3958 return std::vector<unsigned>();
3959 case 'Y': // SSE_REGS if SSE2 allowed
3960 if (Subtarget->hasSSE2())
3961 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3962 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7,
3964 return std::vector<unsigned>();
3968 return std::vector<unsigned>();
3971 /// isLegalAddressImmediate - Return true if the integer value or
3972 /// GlobalValue can be used as the offset of the target addressing mode.
3973 bool X86TargetLowering::isLegalAddressImmediate(int64_t V) const {
3974 // X86 allows a sign-extended 32-bit immediate field.
3975 return (V > -(1LL << 32) && V < (1LL << 32)-1);
3978 bool X86TargetLowering::isLegalAddressImmediate(GlobalValue *GV) const {
3979 if (Subtarget->isTargetDarwin()) {
3980 Reloc::Model RModel = getTargetMachine().getRelocationModel();
3981 if (RModel == Reloc::Static)
3983 else if (RModel == Reloc::DynamicNoPIC)
3984 return !DarwinGVRequiresExtraLoad(GV);
3991 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3992 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3993 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3994 /// are assumed to be legal.
3996 X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const {
3997 // Only do shuffles on 128-bit vector types for now.
3998 if (MVT::getSizeInBits(VT) == 64) return false;
3999 return (Mask.Val->getNumOperands() <= 4 ||
4000 isSplatMask(Mask.Val) ||
4001 isPSHUFHW_PSHUFLWMask(Mask.Val) ||
4002 X86::isUNPCKLMask(Mask.Val) ||
4003 X86::isUNPCKL_v_undef_Mask(Mask.Val) ||
4004 X86::isUNPCKHMask(Mask.Val));
4007 bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps,
4009 SelectionDAG &DAG) const {
4010 unsigned NumElts = BVOps.size();
4011 // Only do shuffles on 128-bit vector types for now.
4012 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false;
4013 if (NumElts == 2) return true;
4015 return (isMOVLMask(BVOps) || isCommutedMOVL(BVOps, true) ||
4016 isSHUFPMask(BVOps) || isCommutedSHUFP(BVOps));