1 //===-- SelectionDAGBuilder.cpp - Selection-DAG building ------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements routines for translating from LLVM IR into SelectionDAG IR.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "isel"
15 #include "SelectionDAGBuilder.h"
16 #include "SDNodeDbgValue.h"
17 #include "llvm/ADT/BitVector.h"
18 #include "llvm/ADT/Optional.h"
19 #include "llvm/ADT/SmallSet.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/Analysis/BranchProbabilityInfo.h"
22 #include "llvm/Analysis/ConstantFolding.h"
23 #include "llvm/Analysis/ValueTracking.h"
24 #include "llvm/CodeGen/Analysis.h"
25 #include "llvm/CodeGen/FastISel.h"
26 #include "llvm/CodeGen/FunctionLoweringInfo.h"
27 #include "llvm/CodeGen/GCMetadata.h"
28 #include "llvm/CodeGen/GCStrategy.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineJumpTableInfo.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/CodeGen/SelectionDAG.h"
36 #include "llvm/CodeGen/StackMaps.h"
37 #include "llvm/DebugInfo.h"
38 #include "llvm/IR/CallingConv.h"
39 #include "llvm/IR/Constants.h"
40 #include "llvm/IR/DataLayout.h"
41 #include "llvm/IR/DerivedTypes.h"
42 #include "llvm/IR/Function.h"
43 #include "llvm/IR/GlobalVariable.h"
44 #include "llvm/IR/InlineAsm.h"
45 #include "llvm/IR/Instructions.h"
46 #include "llvm/IR/IntrinsicInst.h"
47 #include "llvm/IR/Intrinsics.h"
48 #include "llvm/IR/LLVMContext.h"
49 #include "llvm/IR/Module.h"
50 #include "llvm/Support/CommandLine.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/ErrorHandling.h"
53 #include "llvm/Support/MathExtras.h"
54 #include "llvm/Support/raw_ostream.h"
55 #include "llvm/Target/TargetFrameLowering.h"
56 #include "llvm/Target/TargetInstrInfo.h"
57 #include "llvm/Target/TargetIntrinsicInfo.h"
58 #include "llvm/Target/TargetLibraryInfo.h"
59 #include "llvm/Target/TargetLowering.h"
60 #include "llvm/Target/TargetOptions.h"
61 #include "llvm/Target/TargetSelectionDAGInfo.h"
65 /// LimitFloatPrecision - Generate low-precision inline sequences for
66 /// some float libcalls (6, 8 or 12 bits).
67 static unsigned LimitFloatPrecision;
69 static cl::opt<unsigned, true>
70 LimitFPPrecision("limit-float-precision",
71 cl::desc("Generate low-precision inline sequences "
72 "for some float libcalls"),
73 cl::location(LimitFloatPrecision),
76 // Limit the width of DAG chains. This is important in general to prevent
77 // prevent DAG-based analysis from blowing up. For example, alias analysis and
78 // load clustering may not complete in reasonable time. It is difficult to
79 // recognize and avoid this situation within each individual analysis, and
80 // future analyses are likely to have the same behavior. Limiting DAG width is
81 // the safe approach, and will be especially important with global DAGs.
83 // MaxParallelChains default is arbitrarily high to avoid affecting
84 // optimization, but could be lowered to improve compile time. Any ld-ld-st-st
85 // sequence over this should have been converted to llvm.memcpy by the
86 // frontend. It easy to induce this behavior with .ll code such as:
87 // %buffer = alloca [4096 x i8]
88 // %data = load [4096 x i8]* %argPtr
89 // store [4096 x i8] %data, [4096 x i8]* %buffer
90 static const unsigned MaxParallelChains = 64;
92 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, SDLoc DL,
93 const SDValue *Parts, unsigned NumParts,
94 MVT PartVT, EVT ValueVT, const Value *V);
96 /// getCopyFromParts - Create a value that contains the specified legal parts
97 /// combined into the value they represent. If the parts combine to a type
98 /// larger then ValueVT then AssertOp can be used to specify whether the extra
99 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
100 /// (ISD::AssertSext).
101 static SDValue getCopyFromParts(SelectionDAG &DAG, SDLoc DL,
102 const SDValue *Parts,
103 unsigned NumParts, MVT PartVT, EVT ValueVT,
105 ISD::NodeType AssertOp = ISD::DELETED_NODE) {
106 if (ValueVT.isVector())
107 return getCopyFromPartsVector(DAG, DL, Parts, NumParts,
110 assert(NumParts > 0 && "No parts to assemble!");
111 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
112 SDValue Val = Parts[0];
115 // Assemble the value from multiple parts.
116 if (ValueVT.isInteger()) {
117 unsigned PartBits = PartVT.getSizeInBits();
118 unsigned ValueBits = ValueVT.getSizeInBits();
120 // Assemble the power of 2 part.
121 unsigned RoundParts = NumParts & (NumParts - 1) ?
122 1 << Log2_32(NumParts) : NumParts;
123 unsigned RoundBits = PartBits * RoundParts;
124 EVT RoundVT = RoundBits == ValueBits ?
125 ValueVT : EVT::getIntegerVT(*DAG.getContext(), RoundBits);
128 EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2);
130 if (RoundParts > 2) {
131 Lo = getCopyFromParts(DAG, DL, Parts, RoundParts / 2,
133 Hi = getCopyFromParts(DAG, DL, Parts + RoundParts / 2,
134 RoundParts / 2, PartVT, HalfVT, V);
136 Lo = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[0]);
137 Hi = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[1]);
140 if (TLI.isBigEndian())
143 Val = DAG.getNode(ISD::BUILD_PAIR, DL, RoundVT, Lo, Hi);
145 if (RoundParts < NumParts) {
146 // Assemble the trailing non-power-of-2 part.
147 unsigned OddParts = NumParts - RoundParts;
148 EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits);
149 Hi = getCopyFromParts(DAG, DL,
150 Parts + RoundParts, OddParts, PartVT, OddVT, V);
152 // Combine the round and odd parts.
154 if (TLI.isBigEndian())
156 EVT TotalVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
157 Hi = DAG.getNode(ISD::ANY_EXTEND, DL, TotalVT, Hi);
158 Hi = DAG.getNode(ISD::SHL, DL, TotalVT, Hi,
159 DAG.getConstant(Lo.getValueType().getSizeInBits(),
160 TLI.getPointerTy()));
161 Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, TotalVT, Lo);
162 Val = DAG.getNode(ISD::OR, DL, TotalVT, Lo, Hi);
164 } else if (PartVT.isFloatingPoint()) {
165 // FP split into multiple FP parts (for ppcf128)
166 assert(ValueVT == EVT(MVT::ppcf128) && PartVT == MVT::f64 &&
169 Lo = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[0]);
170 Hi = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[1]);
171 if (TLI.isBigEndian())
173 Val = DAG.getNode(ISD::BUILD_PAIR, DL, ValueVT, Lo, Hi);
175 // FP split into integer parts (soft fp)
176 assert(ValueVT.isFloatingPoint() && PartVT.isInteger() &&
177 !PartVT.isVector() && "Unexpected split");
178 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits());
179 Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT, V);
183 // There is now one part, held in Val. Correct it to match ValueVT.
184 EVT PartEVT = Val.getValueType();
186 if (PartEVT == ValueVT)
189 if (PartEVT.isInteger() && ValueVT.isInteger()) {
190 if (ValueVT.bitsLT(PartEVT)) {
191 // For a truncate, see if we have any information to
192 // indicate whether the truncated bits will always be
193 // zero or sign-extension.
194 if (AssertOp != ISD::DELETED_NODE)
195 Val = DAG.getNode(AssertOp, DL, PartEVT, Val,
196 DAG.getValueType(ValueVT));
197 return DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
199 return DAG.getNode(ISD::ANY_EXTEND, DL, ValueVT, Val);
202 if (PartEVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
203 // FP_ROUND's are always exact here.
204 if (ValueVT.bitsLT(Val.getValueType()))
205 return DAG.getNode(ISD::FP_ROUND, DL, ValueVT, Val,
206 DAG.getTargetConstant(1, TLI.getPointerTy()));
208 return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val);
211 if (PartEVT.getSizeInBits() == ValueVT.getSizeInBits())
212 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
214 llvm_unreachable("Unknown mismatch!");
217 /// getCopyFromPartsVector - Create a value that contains the specified legal
218 /// parts combined into the value they represent. If the parts combine to a
219 /// type larger then ValueVT then AssertOp can be used to specify whether the
220 /// extra bits are known to be zero (ISD::AssertZext) or sign extended from
221 /// ValueVT (ISD::AssertSext).
222 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, SDLoc DL,
223 const SDValue *Parts, unsigned NumParts,
224 MVT PartVT, EVT ValueVT, const Value *V) {
225 assert(ValueVT.isVector() && "Not a vector value");
226 assert(NumParts > 0 && "No parts to assemble!");
227 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
228 SDValue Val = Parts[0];
230 // Handle a multi-element vector.
234 unsigned NumIntermediates;
236 TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT,
237 NumIntermediates, RegisterVT);
238 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
239 NumParts = NumRegs; // Silence a compiler warning.
240 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
241 assert(RegisterVT == Parts[0].getSimpleValueType() &&
242 "Part type doesn't match part!");
244 // Assemble the parts into intermediate operands.
245 SmallVector<SDValue, 8> Ops(NumIntermediates);
246 if (NumIntermediates == NumParts) {
247 // If the register was not expanded, truncate or copy the value,
249 for (unsigned i = 0; i != NumParts; ++i)
250 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i], 1,
251 PartVT, IntermediateVT, V);
252 } else if (NumParts > 0) {
253 // If the intermediate type was expanded, build the intermediate
254 // operands from the parts.
255 assert(NumParts % NumIntermediates == 0 &&
256 "Must expand into a divisible number of parts!");
257 unsigned Factor = NumParts / NumIntermediates;
258 for (unsigned i = 0; i != NumIntermediates; ++i)
259 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor,
260 PartVT, IntermediateVT, V);
263 // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the
264 // intermediate operands.
265 Val = DAG.getNode(IntermediateVT.isVector() ?
266 ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR, DL,
267 ValueVT, &Ops[0], NumIntermediates);
270 // There is now one part, held in Val. Correct it to match ValueVT.
271 EVT PartEVT = Val.getValueType();
273 if (PartEVT == ValueVT)
276 if (PartEVT.isVector()) {
277 // If the element type of the source/dest vectors are the same, but the
278 // parts vector has more elements than the value vector, then we have a
279 // vector widening case (e.g. <2 x float> -> <4 x float>). Extract the
281 if (PartEVT.getVectorElementType() == ValueVT.getVectorElementType()) {
282 assert(PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements() &&
283 "Cannot narrow, it would be a lossy transformation");
284 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val,
285 DAG.getConstant(0, TLI.getVectorIdxTy()));
288 // Vector/Vector bitcast.
289 if (ValueVT.getSizeInBits() == PartEVT.getSizeInBits())
290 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
292 assert(PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements() &&
293 "Cannot handle this kind of promotion");
294 // Promoted vector extract
295 bool Smaller = ValueVT.bitsLE(PartEVT);
296 return DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
301 // Trivial bitcast if the types are the same size and the destination
302 // vector type is legal.
303 if (PartEVT.getSizeInBits() == ValueVT.getSizeInBits() &&
304 TLI.isTypeLegal(ValueVT))
305 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
307 // Handle cases such as i8 -> <1 x i1>
308 if (ValueVT.getVectorNumElements() != 1) {
309 LLVMContext &Ctx = *DAG.getContext();
310 Twine ErrMsg("non-trivial scalar-to-vector conversion");
311 if (const Instruction *I = dyn_cast_or_null<Instruction>(V)) {
312 if (const CallInst *CI = dyn_cast<CallInst>(I))
313 if (isa<InlineAsm>(CI->getCalledValue()))
314 ErrMsg = ErrMsg + ", possible invalid constraint for vector type";
315 Ctx.emitError(I, ErrMsg);
317 Ctx.emitError(ErrMsg);
319 return DAG.getUNDEF(ValueVT);
322 if (ValueVT.getVectorNumElements() == 1 &&
323 ValueVT.getVectorElementType() != PartEVT) {
324 bool Smaller = ValueVT.bitsLE(PartEVT);
325 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
326 DL, ValueVT.getScalarType(), Val);
329 return DAG.getNode(ISD::BUILD_VECTOR, DL, ValueVT, Val);
332 static void getCopyToPartsVector(SelectionDAG &DAG, SDLoc dl,
333 SDValue Val, SDValue *Parts, unsigned NumParts,
334 MVT PartVT, const Value *V);
336 /// getCopyToParts - Create a series of nodes that contain the specified value
337 /// split into legal parts. If the parts contain more bits than Val, then, for
338 /// integers, ExtendKind can be used to specify how to generate the extra bits.
339 static void getCopyToParts(SelectionDAG &DAG, SDLoc DL,
340 SDValue Val, SDValue *Parts, unsigned NumParts,
341 MVT PartVT, const Value *V,
342 ISD::NodeType ExtendKind = ISD::ANY_EXTEND) {
343 EVT ValueVT = Val.getValueType();
345 // Handle the vector case separately.
346 if (ValueVT.isVector())
347 return getCopyToPartsVector(DAG, DL, Val, Parts, NumParts, PartVT, V);
349 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
350 unsigned PartBits = PartVT.getSizeInBits();
351 unsigned OrigNumParts = NumParts;
352 assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!");
357 assert(!ValueVT.isVector() && "Vector case handled elsewhere");
358 EVT PartEVT = PartVT;
359 if (PartEVT == ValueVT) {
360 assert(NumParts == 1 && "No-op copy with multiple parts!");
365 if (NumParts * PartBits > ValueVT.getSizeInBits()) {
366 // If the parts cover more bits than the value has, promote the value.
367 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
368 assert(NumParts == 1 && "Do not know what to promote to!");
369 Val = DAG.getNode(ISD::FP_EXTEND, DL, PartVT, Val);
371 assert((PartVT.isInteger() || PartVT == MVT::x86mmx) &&
372 ValueVT.isInteger() &&
373 "Unknown mismatch!");
374 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
375 Val = DAG.getNode(ExtendKind, DL, ValueVT, Val);
376 if (PartVT == MVT::x86mmx)
377 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
379 } else if (PartBits == ValueVT.getSizeInBits()) {
380 // Different types of the same size.
381 assert(NumParts == 1 && PartEVT != ValueVT);
382 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
383 } else if (NumParts * PartBits < ValueVT.getSizeInBits()) {
384 // If the parts cover less bits than value has, truncate the value.
385 assert((PartVT.isInteger() || PartVT == MVT::x86mmx) &&
386 ValueVT.isInteger() &&
387 "Unknown mismatch!");
388 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
389 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
390 if (PartVT == MVT::x86mmx)
391 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
394 // The value may have changed - recompute ValueVT.
395 ValueVT = Val.getValueType();
396 assert(NumParts * PartBits == ValueVT.getSizeInBits() &&
397 "Failed to tile the value with PartVT!");
400 if (PartEVT != ValueVT) {
401 LLVMContext &Ctx = *DAG.getContext();
402 Twine ErrMsg("scalar-to-vector conversion failed");
403 if (const Instruction *I = dyn_cast_or_null<Instruction>(V)) {
404 if (const CallInst *CI = dyn_cast<CallInst>(I))
405 if (isa<InlineAsm>(CI->getCalledValue()))
406 ErrMsg = ErrMsg + ", possible invalid constraint for vector type";
407 Ctx.emitError(I, ErrMsg);
409 Ctx.emitError(ErrMsg);
417 // Expand the value into multiple parts.
418 if (NumParts & (NumParts - 1)) {
419 // The number of parts is not a power of 2. Split off and copy the tail.
420 assert(PartVT.isInteger() && ValueVT.isInteger() &&
421 "Do not know what to expand to!");
422 unsigned RoundParts = 1 << Log2_32(NumParts);
423 unsigned RoundBits = RoundParts * PartBits;
424 unsigned OddParts = NumParts - RoundParts;
425 SDValue OddVal = DAG.getNode(ISD::SRL, DL, ValueVT, Val,
426 DAG.getIntPtrConstant(RoundBits));
427 getCopyToParts(DAG, DL, OddVal, Parts + RoundParts, OddParts, PartVT, V);
429 if (TLI.isBigEndian())
430 // The odd parts were reversed by getCopyToParts - unreverse them.
431 std::reverse(Parts + RoundParts, Parts + NumParts);
433 NumParts = RoundParts;
434 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
435 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
438 // The number of parts is a power of 2. Repeatedly bisect the value using
440 Parts[0] = DAG.getNode(ISD::BITCAST, DL,
441 EVT::getIntegerVT(*DAG.getContext(),
442 ValueVT.getSizeInBits()),
445 for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) {
446 for (unsigned i = 0; i < NumParts; i += StepSize) {
447 unsigned ThisBits = StepSize * PartBits / 2;
448 EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits);
449 SDValue &Part0 = Parts[i];
450 SDValue &Part1 = Parts[i+StepSize/2];
452 Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
453 ThisVT, Part0, DAG.getIntPtrConstant(1));
454 Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
455 ThisVT, Part0, DAG.getIntPtrConstant(0));
457 if (ThisBits == PartBits && ThisVT != PartVT) {
458 Part0 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part0);
459 Part1 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part1);
464 if (TLI.isBigEndian())
465 std::reverse(Parts, Parts + OrigNumParts);
469 /// getCopyToPartsVector - Create a series of nodes that contain the specified
470 /// value split into legal parts.
471 static void getCopyToPartsVector(SelectionDAG &DAG, SDLoc DL,
472 SDValue Val, SDValue *Parts, unsigned NumParts,
473 MVT PartVT, const Value *V) {
474 EVT ValueVT = Val.getValueType();
475 assert(ValueVT.isVector() && "Not a vector");
476 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
479 EVT PartEVT = PartVT;
480 if (PartEVT == ValueVT) {
482 } else if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) {
483 // Bitconvert vector->vector case.
484 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
485 } else if (PartVT.isVector() &&
486 PartEVT.getVectorElementType() == ValueVT.getVectorElementType() &&
487 PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements()) {
488 EVT ElementVT = PartVT.getVectorElementType();
489 // Vector widening case, e.g. <2 x float> -> <4 x float>. Shuffle in
491 SmallVector<SDValue, 16> Ops;
492 for (unsigned i = 0, e = ValueVT.getVectorNumElements(); i != e; ++i)
493 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
494 ElementVT, Val, DAG.getConstant(i,
495 TLI.getVectorIdxTy())));
497 for (unsigned i = ValueVT.getVectorNumElements(),
498 e = PartVT.getVectorNumElements(); i != e; ++i)
499 Ops.push_back(DAG.getUNDEF(ElementVT));
501 Val = DAG.getNode(ISD::BUILD_VECTOR, DL, PartVT, &Ops[0], Ops.size());
503 // FIXME: Use CONCAT for 2x -> 4x.
505 //SDValue UndefElts = DAG.getUNDEF(VectorTy);
506 //Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts);
507 } else if (PartVT.isVector() &&
508 PartEVT.getVectorElementType().bitsGE(
509 ValueVT.getVectorElementType()) &&
510 PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements()) {
512 // Promoted vector extract
513 bool Smaller = PartEVT.bitsLE(ValueVT);
514 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
517 // Vector -> scalar conversion.
518 assert(ValueVT.getVectorNumElements() == 1 &&
519 "Only trivial vector-to-scalar conversions should get here!");
520 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
521 PartVT, Val, DAG.getConstant(0, TLI.getVectorIdxTy()));
523 bool Smaller = ValueVT.bitsLE(PartVT);
524 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
532 // Handle a multi-element vector.
535 unsigned NumIntermediates;
536 unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT,
538 NumIntermediates, RegisterVT);
539 unsigned NumElements = ValueVT.getVectorNumElements();
541 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
542 NumParts = NumRegs; // Silence a compiler warning.
543 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
545 // Split the vector into intermediate operands.
546 SmallVector<SDValue, 8> Ops(NumIntermediates);
547 for (unsigned i = 0; i != NumIntermediates; ++i) {
548 if (IntermediateVT.isVector())
549 Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL,
551 DAG.getConstant(i * (NumElements / NumIntermediates),
552 TLI.getVectorIdxTy()));
554 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
556 DAG.getConstant(i, TLI.getVectorIdxTy()));
559 // Split the intermediate operands into legal parts.
560 if (NumParts == NumIntermediates) {
561 // If the register was not expanded, promote or copy the value,
563 for (unsigned i = 0; i != NumParts; ++i)
564 getCopyToParts(DAG, DL, Ops[i], &Parts[i], 1, PartVT, V);
565 } else if (NumParts > 0) {
566 // If the intermediate type was expanded, split each the value into
568 assert(NumParts % NumIntermediates == 0 &&
569 "Must expand into a divisible number of parts!");
570 unsigned Factor = NumParts / NumIntermediates;
571 for (unsigned i = 0; i != NumIntermediates; ++i)
572 getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT, V);
577 /// RegsForValue - This struct represents the registers (physical or virtual)
578 /// that a particular set of values is assigned, and the type information
579 /// about the value. The most common situation is to represent one value at a
580 /// time, but struct or array values are handled element-wise as multiple
581 /// values. The splitting of aggregates is performed recursively, so that we
582 /// never have aggregate-typed registers. The values at this point do not
583 /// necessarily have legal types, so each value may require one or more
584 /// registers of some legal type.
586 struct RegsForValue {
587 /// ValueVTs - The value types of the values, which may not be legal, and
588 /// may need be promoted or synthesized from one or more registers.
590 SmallVector<EVT, 4> ValueVTs;
592 /// RegVTs - The value types of the registers. This is the same size as
593 /// ValueVTs and it records, for each value, what the type of the assigned
594 /// register or registers are. (Individual values are never synthesized
595 /// from more than one type of register.)
597 /// With virtual registers, the contents of RegVTs is redundant with TLI's
598 /// getRegisterType member function, however when with physical registers
599 /// it is necessary to have a separate record of the types.
601 SmallVector<MVT, 4> RegVTs;
603 /// Regs - This list holds the registers assigned to the values.
604 /// Each legal or promoted value requires one register, and each
605 /// expanded value requires multiple registers.
607 SmallVector<unsigned, 4> Regs;
611 RegsForValue(const SmallVector<unsigned, 4> ®s,
612 MVT regvt, EVT valuevt)
613 : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {}
615 RegsForValue(LLVMContext &Context, const TargetLowering &tli,
616 unsigned Reg, Type *Ty) {
617 ComputeValueVTs(tli, Ty, ValueVTs);
619 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
620 EVT ValueVT = ValueVTs[Value];
621 unsigned NumRegs = tli.getNumRegisters(Context, ValueVT);
622 MVT RegisterVT = tli.getRegisterType(Context, ValueVT);
623 for (unsigned i = 0; i != NumRegs; ++i)
624 Regs.push_back(Reg + i);
625 RegVTs.push_back(RegisterVT);
630 /// areValueTypesLegal - Return true if types of all the values are legal.
631 bool areValueTypesLegal(const TargetLowering &TLI) {
632 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
633 MVT RegisterVT = RegVTs[Value];
634 if (!TLI.isTypeLegal(RegisterVT))
640 /// append - Add the specified values to this one.
641 void append(const RegsForValue &RHS) {
642 ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end());
643 RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end());
644 Regs.append(RHS.Regs.begin(), RHS.Regs.end());
647 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
648 /// this value and returns the result as a ValueVTs value. This uses
649 /// Chain/Flag as the input and updates them for the output Chain/Flag.
650 /// If the Flag pointer is NULL, no flag is used.
651 SDValue getCopyFromRegs(SelectionDAG &DAG, FunctionLoweringInfo &FuncInfo,
653 SDValue &Chain, SDValue *Flag,
654 const Value *V = 0) const;
656 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
657 /// specified value into the registers specified by this object. This uses
658 /// Chain/Flag as the input and updates them for the output Chain/Flag.
659 /// If the Flag pointer is NULL, no flag is used.
660 void getCopyToRegs(SDValue Val, SelectionDAG &DAG, SDLoc dl,
661 SDValue &Chain, SDValue *Flag, const Value *V) const;
663 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
664 /// operand list. This adds the code marker, matching input operand index
665 /// (if applicable), and includes the number of values added into it.
666 void AddInlineAsmOperands(unsigned Kind,
667 bool HasMatching, unsigned MatchingIdx,
669 std::vector<SDValue> &Ops) const;
673 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
674 /// this value and returns the result as a ValueVT value. This uses
675 /// Chain/Flag as the input and updates them for the output Chain/Flag.
676 /// If the Flag pointer is NULL, no flag is used.
677 SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
678 FunctionLoweringInfo &FuncInfo,
680 SDValue &Chain, SDValue *Flag,
681 const Value *V) const {
682 // A Value with type {} or [0 x %t] needs no registers.
683 if (ValueVTs.empty())
686 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
688 // Assemble the legal parts into the final values.
689 SmallVector<SDValue, 4> Values(ValueVTs.size());
690 SmallVector<SDValue, 8> Parts;
691 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
692 // Copy the legal parts from the registers.
693 EVT ValueVT = ValueVTs[Value];
694 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
695 MVT RegisterVT = RegVTs[Value];
697 Parts.resize(NumRegs);
698 for (unsigned i = 0; i != NumRegs; ++i) {
701 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT);
703 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag);
704 *Flag = P.getValue(2);
707 Chain = P.getValue(1);
710 // If the source register was virtual and if we know something about it,
711 // add an assert node.
712 if (!TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) ||
713 !RegisterVT.isInteger() || RegisterVT.isVector())
716 const FunctionLoweringInfo::LiveOutInfo *LOI =
717 FuncInfo.GetLiveOutRegInfo(Regs[Part+i]);
721 unsigned RegSize = RegisterVT.getSizeInBits();
722 unsigned NumSignBits = LOI->NumSignBits;
723 unsigned NumZeroBits = LOI->KnownZero.countLeadingOnes();
725 if (NumZeroBits == RegSize) {
726 // The current value is a zero.
727 // Explicitly express that as it would be easier for
728 // optimizations to kick in.
729 Parts[i] = DAG.getConstant(0, RegisterVT);
733 // FIXME: We capture more information than the dag can represent. For
734 // now, just use the tightest assertzext/assertsext possible.
736 EVT FromVT(MVT::Other);
737 if (NumSignBits == RegSize)
738 isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1
739 else if (NumZeroBits >= RegSize-1)
740 isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1
741 else if (NumSignBits > RegSize-8)
742 isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8
743 else if (NumZeroBits >= RegSize-8)
744 isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8
745 else if (NumSignBits > RegSize-16)
746 isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16
747 else if (NumZeroBits >= RegSize-16)
748 isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16
749 else if (NumSignBits > RegSize-32)
750 isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32
751 else if (NumZeroBits >= RegSize-32)
752 isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32
756 // Add an assertion node.
757 assert(FromVT != MVT::Other);
758 Parts[i] = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl,
759 RegisterVT, P, DAG.getValueType(FromVT));
762 Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(),
763 NumRegs, RegisterVT, ValueVT, V);
768 return DAG.getNode(ISD::MERGE_VALUES, dl,
769 DAG.getVTList(&ValueVTs[0], ValueVTs.size()),
770 &Values[0], ValueVTs.size());
773 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
774 /// specified value into the registers specified by this object. This uses
775 /// Chain/Flag as the input and updates them for the output Chain/Flag.
776 /// If the Flag pointer is NULL, no flag is used.
777 void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, SDLoc dl,
778 SDValue &Chain, SDValue *Flag,
779 const Value *V) const {
780 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
782 // Get the list of the values's legal parts.
783 unsigned NumRegs = Regs.size();
784 SmallVector<SDValue, 8> Parts(NumRegs);
785 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
786 EVT ValueVT = ValueVTs[Value];
787 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
788 MVT RegisterVT = RegVTs[Value];
789 ISD::NodeType ExtendKind =
790 TLI.isZExtFree(Val, RegisterVT)? ISD::ZERO_EXTEND: ISD::ANY_EXTEND;
792 getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value),
793 &Parts[Part], NumParts, RegisterVT, V, ExtendKind);
797 // Copy the parts into the registers.
798 SmallVector<SDValue, 8> Chains(NumRegs);
799 for (unsigned i = 0; i != NumRegs; ++i) {
802 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]);
804 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag);
805 *Flag = Part.getValue(1);
808 Chains[i] = Part.getValue(0);
811 if (NumRegs == 1 || Flag)
812 // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is
813 // flagged to it. That is the CopyToReg nodes and the user are considered
814 // a single scheduling unit. If we create a TokenFactor and return it as
815 // chain, then the TokenFactor is both a predecessor (operand) of the
816 // user as well as a successor (the TF operands are flagged to the user).
817 // c1, f1 = CopyToReg
818 // c2, f2 = CopyToReg
819 // c3 = TokenFactor c1, c2
822 Chain = Chains[NumRegs-1];
824 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0], NumRegs);
827 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
828 /// operand list. This adds the code marker and includes the number of
829 /// values added into it.
830 void RegsForValue::AddInlineAsmOperands(unsigned Code, bool HasMatching,
831 unsigned MatchingIdx,
833 std::vector<SDValue> &Ops) const {
834 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
836 unsigned Flag = InlineAsm::getFlagWord(Code, Regs.size());
838 Flag = InlineAsm::getFlagWordForMatchingOp(Flag, MatchingIdx);
839 else if (!Regs.empty() &&
840 TargetRegisterInfo::isVirtualRegister(Regs.front())) {
841 // Put the register class of the virtual registers in the flag word. That
842 // way, later passes can recompute register class constraints for inline
843 // assembly as well as normal instructions.
844 // Don't do this for tied operands that can use the regclass information
846 const MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
847 const TargetRegisterClass *RC = MRI.getRegClass(Regs.front());
848 Flag = InlineAsm::getFlagWordForRegClass(Flag, RC->getID());
851 SDValue Res = DAG.getTargetConstant(Flag, MVT::i32);
854 unsigned SP = TLI.getStackPointerRegisterToSaveRestore();
855 for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) {
856 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVTs[Value]);
857 MVT RegisterVT = RegVTs[Value];
858 for (unsigned i = 0; i != NumRegs; ++i) {
859 assert(Reg < Regs.size() && "Mismatch in # registers expected");
860 unsigned TheReg = Regs[Reg++];
861 Ops.push_back(DAG.getRegister(TheReg, RegisterVT));
863 // Notice if we clobbered the stack pointer. Yes, inline asm can do this.
864 if (TheReg == SP && Code == InlineAsm::Kind_Clobber) {
865 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
866 MFI->setHasInlineAsmWithSPAdjust(true);
872 void SelectionDAGBuilder::init(GCFunctionInfo *gfi, AliasAnalysis &aa,
873 const TargetLibraryInfo *li) {
877 TD = DAG.getTarget().getDataLayout();
878 Context = DAG.getContext();
879 LPadToCallSiteMap.clear();
882 /// clear - Clear out the current SelectionDAG and the associated
883 /// state and prepare this SelectionDAGBuilder object to be used
884 /// for a new block. This doesn't clear out information about
885 /// additional blocks that are needed to complete switch lowering
886 /// or PHI node updating; that information is cleared out as it is
888 void SelectionDAGBuilder::clear() {
890 UnusedArgNodeMap.clear();
891 PendingLoads.clear();
892 PendingExports.clear();
897 /// clearDanglingDebugInfo - Clear the dangling debug information
898 /// map. This function is separated from the clear so that debug
899 /// information that is dangling in a basic block can be properly
900 /// resolved in a different basic block. This allows the
901 /// SelectionDAG to resolve dangling debug information attached
903 void SelectionDAGBuilder::clearDanglingDebugInfo() {
904 DanglingDebugInfoMap.clear();
907 /// getRoot - Return the current virtual root of the Selection DAG,
908 /// flushing any PendingLoad items. This must be done before emitting
909 /// a store or any other node that may need to be ordered after any
910 /// prior load instructions.
912 SDValue SelectionDAGBuilder::getRoot() {
913 if (PendingLoads.empty())
914 return DAG.getRoot();
916 if (PendingLoads.size() == 1) {
917 SDValue Root = PendingLoads[0];
919 PendingLoads.clear();
923 // Otherwise, we have to make a token factor node.
924 SDValue Root = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
925 &PendingLoads[0], PendingLoads.size());
926 PendingLoads.clear();
931 /// getControlRoot - Similar to getRoot, but instead of flushing all the
932 /// PendingLoad items, flush all the PendingExports items. It is necessary
933 /// to do this before emitting a terminator instruction.
935 SDValue SelectionDAGBuilder::getControlRoot() {
936 SDValue Root = DAG.getRoot();
938 if (PendingExports.empty())
941 // Turn all of the CopyToReg chains into one factored node.
942 if (Root.getOpcode() != ISD::EntryToken) {
943 unsigned i = 0, e = PendingExports.size();
944 for (; i != e; ++i) {
945 assert(PendingExports[i].getNode()->getNumOperands() > 1);
946 if (PendingExports[i].getNode()->getOperand(0) == Root)
947 break; // Don't add the root if we already indirectly depend on it.
951 PendingExports.push_back(Root);
954 Root = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
956 PendingExports.size());
957 PendingExports.clear();
962 void SelectionDAGBuilder::visit(const Instruction &I) {
963 // Set up outgoing PHI node register values before emitting the terminator.
964 if (isa<TerminatorInst>(&I))
965 HandlePHINodesInSuccessorBlocks(I.getParent());
971 visit(I.getOpcode(), I);
973 if (!isa<TerminatorInst>(&I) && !HasTailCall)
974 CopyToExportRegsIfNeeded(&I);
979 void SelectionDAGBuilder::visitPHI(const PHINode &) {
980 llvm_unreachable("SelectionDAGBuilder shouldn't visit PHI nodes!");
983 void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) {
984 // Note: this doesn't use InstVisitor, because it has to work with
985 // ConstantExpr's in addition to instructions.
987 default: llvm_unreachable("Unknown instruction type encountered!");
988 // Build the switch statement using the Instruction.def file.
989 #define HANDLE_INST(NUM, OPCODE, CLASS) \
990 case Instruction::OPCODE: visit##OPCODE((const CLASS&)I); break;
991 #include "llvm/IR/Instruction.def"
995 // resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V,
996 // generate the debug data structures now that we've seen its definition.
997 void SelectionDAGBuilder::resolveDanglingDebugInfo(const Value *V,
999 DanglingDebugInfo &DDI = DanglingDebugInfoMap[V];
1001 const DbgValueInst *DI = DDI.getDI();
1002 DebugLoc dl = DDI.getdl();
1003 unsigned DbgSDNodeOrder = DDI.getSDNodeOrder();
1004 MDNode *Variable = DI->getVariable();
1005 uint64_t Offset = DI->getOffset();
1007 if (Val.getNode()) {
1008 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, Val)) {
1009 SDV = DAG.getDbgValue(Variable, Val.getNode(),
1010 Val.getResNo(), Offset, dl, DbgSDNodeOrder);
1011 DAG.AddDbgValue(SDV, Val.getNode(), false);
1014 DEBUG(dbgs() << "Dropping debug info for " << *DI << "\n");
1015 DanglingDebugInfoMap[V] = DanglingDebugInfo();
1019 /// getValue - Return an SDValue for the given Value.
1020 SDValue SelectionDAGBuilder::getValue(const Value *V) {
1021 // If we already have an SDValue for this value, use it. It's important
1022 // to do this first, so that we don't create a CopyFromReg if we already
1023 // have a regular SDValue.
1024 SDValue &N = NodeMap[V];
1025 if (N.getNode()) return N;
1027 // If there's a virtual register allocated and initialized for this
1029 DenseMap<const Value *, unsigned>::iterator It = FuncInfo.ValueMap.find(V);
1030 if (It != FuncInfo.ValueMap.end()) {
1031 unsigned InReg = It->second;
1032 RegsForValue RFV(*DAG.getContext(), *TM.getTargetLowering(),
1033 InReg, V->getType());
1034 SDValue Chain = DAG.getEntryNode();
1035 N = RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, NULL, V);
1036 resolveDanglingDebugInfo(V, N);
1040 // Otherwise create a new SDValue and remember it.
1041 SDValue Val = getValueImpl(V);
1043 resolveDanglingDebugInfo(V, Val);
1047 /// getNonRegisterValue - Return an SDValue for the given Value, but
1048 /// don't look in FuncInfo.ValueMap for a virtual register.
1049 SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) {
1050 // If we already have an SDValue for this value, use it.
1051 SDValue &N = NodeMap[V];
1052 if (N.getNode()) return N;
1054 // Otherwise create a new SDValue and remember it.
1055 SDValue Val = getValueImpl(V);
1057 resolveDanglingDebugInfo(V, Val);
1061 /// getValueImpl - Helper function for getValue and getNonRegisterValue.
1062 /// Create an SDValue for the given value.
1063 SDValue SelectionDAGBuilder::getValueImpl(const Value *V) {
1064 const TargetLowering *TLI = TM.getTargetLowering();
1066 if (const Constant *C = dyn_cast<Constant>(V)) {
1067 EVT VT = TLI->getValueType(V->getType(), true);
1069 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C))
1070 return DAG.getConstant(*CI, VT);
1072 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
1073 return DAG.getGlobalAddress(GV, getCurSDLoc(), VT);
1075 if (isa<ConstantPointerNull>(C)) {
1076 unsigned AS = V->getType()->getPointerAddressSpace();
1077 return DAG.getConstant(0, TLI->getPointerTy(AS));
1080 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
1081 return DAG.getConstantFP(*CFP, VT);
1083 if (isa<UndefValue>(C) && !V->getType()->isAggregateType())
1084 return DAG.getUNDEF(VT);
1086 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
1087 visit(CE->getOpcode(), *CE);
1088 SDValue N1 = NodeMap[V];
1089 assert(N1.getNode() && "visit didn't populate the NodeMap!");
1093 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) {
1094 SmallVector<SDValue, 4> Constants;
1095 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end();
1097 SDNode *Val = getValue(*OI).getNode();
1098 // If the operand is an empty aggregate, there are no values.
1100 // Add each leaf value from the operand to the Constants list
1101 // to form a flattened list of all the values.
1102 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
1103 Constants.push_back(SDValue(Val, i));
1106 return DAG.getMergeValues(&Constants[0], Constants.size(),
1110 if (const ConstantDataSequential *CDS =
1111 dyn_cast<ConstantDataSequential>(C)) {
1112 SmallVector<SDValue, 4> Ops;
1113 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1114 SDNode *Val = getValue(CDS->getElementAsConstant(i)).getNode();
1115 // Add each leaf value from the operand to the Constants list
1116 // to form a flattened list of all the values.
1117 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
1118 Ops.push_back(SDValue(Val, i));
1121 if (isa<ArrayType>(CDS->getType()))
1122 return DAG.getMergeValues(&Ops[0], Ops.size(), getCurSDLoc());
1123 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(),
1124 VT, &Ops[0], Ops.size());
1127 if (C->getType()->isStructTy() || C->getType()->isArrayTy()) {
1128 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
1129 "Unknown struct or array constant!");
1131 SmallVector<EVT, 4> ValueVTs;
1132 ComputeValueVTs(*TLI, C->getType(), ValueVTs);
1133 unsigned NumElts = ValueVTs.size();
1135 return SDValue(); // empty struct
1136 SmallVector<SDValue, 4> Constants(NumElts);
1137 for (unsigned i = 0; i != NumElts; ++i) {
1138 EVT EltVT = ValueVTs[i];
1139 if (isa<UndefValue>(C))
1140 Constants[i] = DAG.getUNDEF(EltVT);
1141 else if (EltVT.isFloatingPoint())
1142 Constants[i] = DAG.getConstantFP(0, EltVT);
1144 Constants[i] = DAG.getConstant(0, EltVT);
1147 return DAG.getMergeValues(&Constants[0], NumElts,
1151 if (const BlockAddress *BA = dyn_cast<BlockAddress>(C))
1152 return DAG.getBlockAddress(BA, VT);
1154 VectorType *VecTy = cast<VectorType>(V->getType());
1155 unsigned NumElements = VecTy->getNumElements();
1157 // Now that we know the number and type of the elements, get that number of
1158 // elements into the Ops array based on what kind of constant it is.
1159 SmallVector<SDValue, 16> Ops;
1160 if (const ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
1161 for (unsigned i = 0; i != NumElements; ++i)
1162 Ops.push_back(getValue(CV->getOperand(i)));
1164 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!");
1165 EVT EltVT = TLI->getValueType(VecTy->getElementType());
1168 if (EltVT.isFloatingPoint())
1169 Op = DAG.getConstantFP(0, EltVT);
1171 Op = DAG.getConstant(0, EltVT);
1172 Ops.assign(NumElements, Op);
1175 // Create a BUILD_VECTOR node.
1176 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(),
1177 VT, &Ops[0], Ops.size());
1180 // If this is a static alloca, generate it as the frameindex instead of
1182 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
1183 DenseMap<const AllocaInst*, int>::iterator SI =
1184 FuncInfo.StaticAllocaMap.find(AI);
1185 if (SI != FuncInfo.StaticAllocaMap.end())
1186 return DAG.getFrameIndex(SI->second, TLI->getPointerTy());
1189 // If this is an instruction which fast-isel has deferred, select it now.
1190 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
1191 unsigned InReg = FuncInfo.InitializeRegForValue(Inst);
1192 RegsForValue RFV(*DAG.getContext(), *TLI, InReg, Inst->getType());
1193 SDValue Chain = DAG.getEntryNode();
1194 return RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, NULL, V);
1197 llvm_unreachable("Can't get register for value!");
1200 void SelectionDAGBuilder::visitRet(const ReturnInst &I) {
1201 const TargetLowering *TLI = TM.getTargetLowering();
1202 SDValue Chain = getControlRoot();
1203 SmallVector<ISD::OutputArg, 8> Outs;
1204 SmallVector<SDValue, 8> OutVals;
1206 if (!FuncInfo.CanLowerReturn) {
1207 unsigned DemoteReg = FuncInfo.DemoteRegister;
1208 const Function *F = I.getParent()->getParent();
1210 // Emit a store of the return value through the virtual register.
1211 // Leave Outs empty so that LowerReturn won't try to load return
1212 // registers the usual way.
1213 SmallVector<EVT, 1> PtrValueVTs;
1214 ComputeValueVTs(*TLI, PointerType::getUnqual(F->getReturnType()),
1217 SDValue RetPtr = DAG.getRegister(DemoteReg, PtrValueVTs[0]);
1218 SDValue RetOp = getValue(I.getOperand(0));
1220 SmallVector<EVT, 4> ValueVTs;
1221 SmallVector<uint64_t, 4> Offsets;
1222 ComputeValueVTs(*TLI, I.getOperand(0)->getType(), ValueVTs, &Offsets);
1223 unsigned NumValues = ValueVTs.size();
1225 SmallVector<SDValue, 4> Chains(NumValues);
1226 for (unsigned i = 0; i != NumValues; ++i) {
1227 SDValue Add = DAG.getNode(ISD::ADD, getCurSDLoc(),
1228 RetPtr.getValueType(), RetPtr,
1229 DAG.getIntPtrConstant(Offsets[i]));
1231 DAG.getStore(Chain, getCurSDLoc(),
1232 SDValue(RetOp.getNode(), RetOp.getResNo() + i),
1233 // FIXME: better loc info would be nice.
1234 Add, MachinePointerInfo(), false, false, 0);
1237 Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(),
1238 MVT::Other, &Chains[0], NumValues);
1239 } else if (I.getNumOperands() != 0) {
1240 SmallVector<EVT, 4> ValueVTs;
1241 ComputeValueVTs(*TLI, I.getOperand(0)->getType(), ValueVTs);
1242 unsigned NumValues = ValueVTs.size();
1244 SDValue RetOp = getValue(I.getOperand(0));
1245 for (unsigned j = 0, f = NumValues; j != f; ++j) {
1246 EVT VT = ValueVTs[j];
1248 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1250 const Function *F = I.getParent()->getParent();
1251 if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1253 ExtendKind = ISD::SIGN_EXTEND;
1254 else if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1256 ExtendKind = ISD::ZERO_EXTEND;
1258 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger())
1259 VT = TLI->getTypeForExtArgOrReturn(VT.getSimpleVT(), ExtendKind);
1261 unsigned NumParts = TLI->getNumRegisters(*DAG.getContext(), VT);
1262 MVT PartVT = TLI->getRegisterType(*DAG.getContext(), VT);
1263 SmallVector<SDValue, 4> Parts(NumParts);
1264 getCopyToParts(DAG, getCurSDLoc(),
1265 SDValue(RetOp.getNode(), RetOp.getResNo() + j),
1266 &Parts[0], NumParts, PartVT, &I, ExtendKind);
1268 // 'inreg' on function refers to return value
1269 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1270 if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1274 // Propagate extension type if any
1275 if (ExtendKind == ISD::SIGN_EXTEND)
1277 else if (ExtendKind == ISD::ZERO_EXTEND)
1280 for (unsigned i = 0; i < NumParts; ++i) {
1281 Outs.push_back(ISD::OutputArg(Flags, Parts[i].getValueType(),
1282 VT, /*isfixed=*/true, 0, 0));
1283 OutVals.push_back(Parts[i]);
1289 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1290 CallingConv::ID CallConv =
1291 DAG.getMachineFunction().getFunction()->getCallingConv();
1292 Chain = TM.getTargetLowering()->LowerReturn(Chain, CallConv, isVarArg,
1293 Outs, OutVals, getCurSDLoc(),
1296 // Verify that the target's LowerReturn behaved as expected.
1297 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
1298 "LowerReturn didn't return a valid chain!");
1300 // Update the DAG with the new chain value resulting from return lowering.
1304 /// CopyToExportRegsIfNeeded - If the given value has virtual registers
1305 /// created for it, emit nodes to copy the value into the virtual
1307 void SelectionDAGBuilder::CopyToExportRegsIfNeeded(const Value *V) {
1309 if (V->getType()->isEmptyTy())
1312 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
1313 if (VMI != FuncInfo.ValueMap.end()) {
1314 assert(!V->use_empty() && "Unused value assigned virtual registers!");
1315 CopyValueToVirtualRegister(V, VMI->second);
1319 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
1320 /// the current basic block, add it to ValueMap now so that we'll get a
1322 void SelectionDAGBuilder::ExportFromCurrentBlock(const Value *V) {
1323 // No need to export constants.
1324 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
1326 // Already exported?
1327 if (FuncInfo.isExportedInst(V)) return;
1329 unsigned Reg = FuncInfo.InitializeRegForValue(V);
1330 CopyValueToVirtualRegister(V, Reg);
1333 bool SelectionDAGBuilder::isExportableFromCurrentBlock(const Value *V,
1334 const BasicBlock *FromBB) {
1335 // The operands of the setcc have to be in this block. We don't know
1336 // how to export them from some other block.
1337 if (const Instruction *VI = dyn_cast<Instruction>(V)) {
1338 // Can export from current BB.
1339 if (VI->getParent() == FromBB)
1342 // Is already exported, noop.
1343 return FuncInfo.isExportedInst(V);
1346 // If this is an argument, we can export it if the BB is the entry block or
1347 // if it is already exported.
1348 if (isa<Argument>(V)) {
1349 if (FromBB == &FromBB->getParent()->getEntryBlock())
1352 // Otherwise, can only export this if it is already exported.
1353 return FuncInfo.isExportedInst(V);
1356 // Otherwise, constants can always be exported.
1360 /// Return branch probability calculated by BranchProbabilityInfo for IR blocks.
1361 uint32_t SelectionDAGBuilder::getEdgeWeight(const MachineBasicBlock *Src,
1362 const MachineBasicBlock *Dst) const {
1363 BranchProbabilityInfo *BPI = FuncInfo.BPI;
1366 const BasicBlock *SrcBB = Src->getBasicBlock();
1367 const BasicBlock *DstBB = Dst->getBasicBlock();
1368 return BPI->getEdgeWeight(SrcBB, DstBB);
1371 void SelectionDAGBuilder::
1372 addSuccessorWithWeight(MachineBasicBlock *Src, MachineBasicBlock *Dst,
1373 uint32_t Weight /* = 0 */) {
1375 Weight = getEdgeWeight(Src, Dst);
1376 Src->addSuccessor(Dst, Weight);
1380 static bool InBlock(const Value *V, const BasicBlock *BB) {
1381 if (const Instruction *I = dyn_cast<Instruction>(V))
1382 return I->getParent() == BB;
1386 /// EmitBranchForMergedCondition - Helper method for FindMergedConditions.
1387 /// This function emits a branch and is used at the leaves of an OR or an
1388 /// AND operator tree.
1391 SelectionDAGBuilder::EmitBranchForMergedCondition(const Value *Cond,
1392 MachineBasicBlock *TBB,
1393 MachineBasicBlock *FBB,
1394 MachineBasicBlock *CurBB,
1395 MachineBasicBlock *SwitchBB) {
1396 const BasicBlock *BB = CurBB->getBasicBlock();
1398 // If the leaf of the tree is a comparison, merge the condition into
1400 if (const CmpInst *BOp = dyn_cast<CmpInst>(Cond)) {
1401 // The operands of the cmp have to be in this block. We don't know
1402 // how to export them from some other block. If this is the first block
1403 // of the sequence, no exporting is needed.
1404 if (CurBB == SwitchBB ||
1405 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
1406 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) {
1407 ISD::CondCode Condition;
1408 if (const ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
1409 Condition = getICmpCondCode(IC->getPredicate());
1410 } else if (const FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) {
1411 Condition = getFCmpCondCode(FC->getPredicate());
1412 if (TM.Options.NoNaNsFPMath)
1413 Condition = getFCmpCodeWithoutNaN(Condition);
1415 Condition = ISD::SETEQ; // silence warning.
1416 llvm_unreachable("Unknown compare instruction");
1419 CaseBlock CB(Condition, BOp->getOperand(0),
1420 BOp->getOperand(1), NULL, TBB, FBB, CurBB);
1421 SwitchCases.push_back(CB);
1426 // Create a CaseBlock record representing this branch.
1427 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()),
1428 NULL, TBB, FBB, CurBB);
1429 SwitchCases.push_back(CB);
1432 /// FindMergedConditions - If Cond is an expression like
1433 void SelectionDAGBuilder::FindMergedConditions(const Value *Cond,
1434 MachineBasicBlock *TBB,
1435 MachineBasicBlock *FBB,
1436 MachineBasicBlock *CurBB,
1437 MachineBasicBlock *SwitchBB,
1439 // If this node is not part of the or/and tree, emit it as a branch.
1440 const Instruction *BOp = dyn_cast<Instruction>(Cond);
1441 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
1442 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
1443 BOp->getParent() != CurBB->getBasicBlock() ||
1444 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
1445 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
1446 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB);
1450 // Create TmpBB after CurBB.
1451 MachineFunction::iterator BBI = CurBB;
1452 MachineFunction &MF = DAG.getMachineFunction();
1453 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock());
1454 CurBB->getParent()->insert(++BBI, TmpBB);
1456 if (Opc == Instruction::Or) {
1457 // Codegen X | Y as:
1465 // Emit the LHS condition.
1466 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, SwitchBB, Opc);
1468 // Emit the RHS condition into TmpBB.
1469 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
1471 assert(Opc == Instruction::And && "Unknown merge op!");
1472 // Codegen X & Y as:
1479 // This requires creation of TmpBB after CurBB.
1481 // Emit the LHS condition.
1482 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, SwitchBB, Opc);
1484 // Emit the RHS condition into TmpBB.
1485 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
1489 /// If the set of cases should be emitted as a series of branches, return true.
1490 /// If we should emit this as a bunch of and/or'd together conditions, return
1493 SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases) {
1494 if (Cases.size() != 2) return true;
1496 // If this is two comparisons of the same values or'd or and'd together, they
1497 // will get folded into a single comparison, so don't emit two blocks.
1498 if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
1499 Cases[0].CmpRHS == Cases[1].CmpRHS) ||
1500 (Cases[0].CmpRHS == Cases[1].CmpLHS &&
1501 Cases[0].CmpLHS == Cases[1].CmpRHS)) {
1505 // Handle: (X != null) | (Y != null) --> (X|Y) != 0
1506 // Handle: (X == null) & (Y == null) --> (X|Y) == 0
1507 if (Cases[0].CmpRHS == Cases[1].CmpRHS &&
1508 Cases[0].CC == Cases[1].CC &&
1509 isa<Constant>(Cases[0].CmpRHS) &&
1510 cast<Constant>(Cases[0].CmpRHS)->isNullValue()) {
1511 if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB)
1513 if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB)
1520 void SelectionDAGBuilder::visitBr(const BranchInst &I) {
1521 MachineBasicBlock *BrMBB = FuncInfo.MBB;
1523 // Update machine-CFG edges.
1524 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
1526 // Figure out which block is immediately after the current one.
1527 MachineBasicBlock *NextBlock = 0;
1528 MachineFunction::iterator BBI = BrMBB;
1529 if (++BBI != FuncInfo.MF->end())
1532 if (I.isUnconditional()) {
1533 // Update machine-CFG edges.
1534 BrMBB->addSuccessor(Succ0MBB);
1536 // If this is not a fall-through branch, emit the branch.
1537 if (Succ0MBB != NextBlock)
1538 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(),
1539 MVT::Other, getControlRoot(),
1540 DAG.getBasicBlock(Succ0MBB)));
1545 // If this condition is one of the special cases we handle, do special stuff
1547 const Value *CondVal = I.getCondition();
1548 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
1550 // If this is a series of conditions that are or'd or and'd together, emit
1551 // this as a sequence of branches instead of setcc's with and/or operations.
1552 // As long as jumps are not expensive, this should improve performance.
1553 // For example, instead of something like:
1566 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
1567 if (!TM.getTargetLowering()->isJumpExpensive() &&
1569 (BOp->getOpcode() == Instruction::And ||
1570 BOp->getOpcode() == Instruction::Or)) {
1571 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB,
1573 // If the compares in later blocks need to use values not currently
1574 // exported from this block, export them now. This block should always
1575 // be the first entry.
1576 assert(SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!");
1578 // Allow some cases to be rejected.
1579 if (ShouldEmitAsBranches(SwitchCases)) {
1580 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1581 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1582 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1585 // Emit the branch for this block.
1586 visitSwitchCase(SwitchCases[0], BrMBB);
1587 SwitchCases.erase(SwitchCases.begin());
1591 // Okay, we decided not to do this, remove any inserted MBB's and clear
1593 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
1594 FuncInfo.MF->erase(SwitchCases[i].ThisBB);
1596 SwitchCases.clear();
1600 // Create a CaseBlock record representing this branch.
1601 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()),
1602 NULL, Succ0MBB, Succ1MBB, BrMBB);
1604 // Use visitSwitchCase to actually insert the fast branch sequence for this
1606 visitSwitchCase(CB, BrMBB);
1609 /// visitSwitchCase - Emits the necessary code to represent a single node in
1610 /// the binary search tree resulting from lowering a switch instruction.
1611 void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB,
1612 MachineBasicBlock *SwitchBB) {
1614 SDValue CondLHS = getValue(CB.CmpLHS);
1615 SDLoc dl = getCurSDLoc();
1617 // Build the setcc now.
1618 if (CB.CmpMHS == NULL) {
1619 // Fold "(X == true)" to X and "(X == false)" to !X to
1620 // handle common cases produced by branch lowering.
1621 if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) &&
1622 CB.CC == ISD::SETEQ)
1624 else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) &&
1625 CB.CC == ISD::SETEQ) {
1626 SDValue True = DAG.getConstant(1, CondLHS.getValueType());
1627 Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True);
1629 Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1631 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now");
1633 const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue();
1634 const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue();
1636 SDValue CmpOp = getValue(CB.CmpMHS);
1637 EVT VT = CmpOp.getValueType();
1639 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) {
1640 Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, VT),
1643 SDValue SUB = DAG.getNode(ISD::SUB, dl,
1644 VT, CmpOp, DAG.getConstant(Low, VT));
1645 Cond = DAG.getSetCC(dl, MVT::i1, SUB,
1646 DAG.getConstant(High-Low, VT), ISD::SETULE);
1650 // Update successor info
1651 addSuccessorWithWeight(SwitchBB, CB.TrueBB, CB.TrueWeight);
1652 // TrueBB and FalseBB are always different unless the incoming IR is
1653 // degenerate. This only happens when running llc on weird IR.
1654 if (CB.TrueBB != CB.FalseBB)
1655 addSuccessorWithWeight(SwitchBB, CB.FalseBB, CB.FalseWeight);
1657 // Set NextBlock to be the MBB immediately after the current one, if any.
1658 // This is used to avoid emitting unnecessary branches to the next block.
1659 MachineBasicBlock *NextBlock = 0;
1660 MachineFunction::iterator BBI = SwitchBB;
1661 if (++BBI != FuncInfo.MF->end())
1664 // If the lhs block is the next block, invert the condition so that we can
1665 // fall through to the lhs instead of the rhs block.
1666 if (CB.TrueBB == NextBlock) {
1667 std::swap(CB.TrueBB, CB.FalseBB);
1668 SDValue True = DAG.getConstant(1, Cond.getValueType());
1669 Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True);
1672 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1673 MVT::Other, getControlRoot(), Cond,
1674 DAG.getBasicBlock(CB.TrueBB));
1676 // Insert the false branch. Do this even if it's a fall through branch,
1677 // this makes it easier to do DAG optimizations which require inverting
1678 // the branch condition.
1679 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1680 DAG.getBasicBlock(CB.FalseBB));
1682 DAG.setRoot(BrCond);
1685 /// visitJumpTable - Emit JumpTable node in the current MBB
1686 void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) {
1687 // Emit the code for the jump table
1688 assert(JT.Reg != -1U && "Should lower JT Header first!");
1689 EVT PTy = TM.getTargetLowering()->getPointerTy();
1690 SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurSDLoc(),
1692 SDValue Table = DAG.getJumpTable(JT.JTI, PTy);
1693 SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurSDLoc(),
1694 MVT::Other, Index.getValue(1),
1696 DAG.setRoot(BrJumpTable);
1699 /// visitJumpTableHeader - This function emits necessary code to produce index
1700 /// in the JumpTable from switch case.
1701 void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT,
1702 JumpTableHeader &JTH,
1703 MachineBasicBlock *SwitchBB) {
1704 // Subtract the lowest switch case value from the value being switched on and
1705 // conditional branch to default mbb if the result is greater than the
1706 // difference between smallest and largest cases.
1707 SDValue SwitchOp = getValue(JTH.SValue);
1708 EVT VT = SwitchOp.getValueType();
1709 SDValue Sub = DAG.getNode(ISD::SUB, getCurSDLoc(), VT, SwitchOp,
1710 DAG.getConstant(JTH.First, VT));
1712 // The SDNode we just created, which holds the value being switched on minus
1713 // the smallest case value, needs to be copied to a virtual register so it
1714 // can be used as an index into the jump table in a subsequent basic block.
1715 // This value may be smaller or larger than the target's pointer type, and
1716 // therefore require extension or truncating.
1717 const TargetLowering *TLI = TM.getTargetLowering();
1718 SwitchOp = DAG.getZExtOrTrunc(Sub, getCurSDLoc(), TLI->getPointerTy());
1720 unsigned JumpTableReg = FuncInfo.CreateReg(TLI->getPointerTy());
1721 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurSDLoc(),
1722 JumpTableReg, SwitchOp);
1723 JT.Reg = JumpTableReg;
1725 // Emit the range check for the jump table, and branch to the default block
1726 // for the switch statement if the value being switched on exceeds the largest
1727 // case in the switch.
1728 SDValue CMP = DAG.getSetCC(getCurSDLoc(),
1729 TLI->getSetCCResultType(*DAG.getContext(),
1730 Sub.getValueType()),
1732 DAG.getConstant(JTH.Last - JTH.First,VT),
1735 // Set NextBlock to be the MBB immediately after the current one, if any.
1736 // This is used to avoid emitting unnecessary branches to the next block.
1737 MachineBasicBlock *NextBlock = 0;
1738 MachineFunction::iterator BBI = SwitchBB;
1740 if (++BBI != FuncInfo.MF->end())
1743 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurSDLoc(),
1744 MVT::Other, CopyTo, CMP,
1745 DAG.getBasicBlock(JT.Default));
1747 if (JT.MBB != NextBlock)
1748 BrCond = DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, BrCond,
1749 DAG.getBasicBlock(JT.MBB));
1751 DAG.setRoot(BrCond);
1754 /// Codegen a new tail for a stack protector check ParentMBB which has had its
1755 /// tail spliced into a stack protector check success bb.
1757 /// For a high level explanation of how this fits into the stack protector
1758 /// generation see the comment on the declaration of class
1759 /// StackProtectorDescriptor.
1760 void SelectionDAGBuilder::visitSPDescriptorParent(StackProtectorDescriptor &SPD,
1761 MachineBasicBlock *ParentBB) {
1763 // First create the loads to the guard/stack slot for the comparison.
1764 const TargetLowering *TLI = TM.getTargetLowering();
1765 EVT PtrTy = TLI->getPointerTy();
1767 MachineFrameInfo *MFI = ParentBB->getParent()->getFrameInfo();
1768 int FI = MFI->getStackProtectorIndex();
1770 const Value *IRGuard = SPD.getGuard();
1771 SDValue GuardPtr = getValue(IRGuard);
1772 SDValue StackSlotPtr = DAG.getFrameIndex(FI, PtrTy);
1775 TLI->getDataLayout()->getPrefTypeAlignment(IRGuard->getType());
1776 SDValue Guard = DAG.getLoad(PtrTy, getCurSDLoc(), DAG.getEntryNode(),
1777 GuardPtr, MachinePointerInfo(IRGuard, 0),
1778 true, false, false, Align);
1780 SDValue StackSlot = DAG.getLoad(PtrTy, getCurSDLoc(), DAG.getEntryNode(),
1782 MachinePointerInfo::getFixedStack(FI),
1783 true, false, false, Align);
1785 // Perform the comparison via a subtract/getsetcc.
1786 EVT VT = Guard.getValueType();
1787 SDValue Sub = DAG.getNode(ISD::SUB, getCurSDLoc(), VT, Guard, StackSlot);
1789 SDValue Cmp = DAG.getSetCC(getCurSDLoc(),
1790 TLI->getSetCCResultType(*DAG.getContext(),
1791 Sub.getValueType()),
1792 Sub, DAG.getConstant(0, VT),
1795 // If the sub is not 0, then we know the guard/stackslot do not equal, so
1796 // branch to failure MBB.
1797 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurSDLoc(),
1798 MVT::Other, StackSlot.getOperand(0),
1799 Cmp, DAG.getBasicBlock(SPD.getFailureMBB()));
1800 // Otherwise branch to success MBB.
1801 SDValue Br = DAG.getNode(ISD::BR, getCurSDLoc(),
1803 DAG.getBasicBlock(SPD.getSuccessMBB()));
1808 /// Codegen the failure basic block for a stack protector check.
1810 /// A failure stack protector machine basic block consists simply of a call to
1811 /// __stack_chk_fail().
1813 /// For a high level explanation of how this fits into the stack protector
1814 /// generation see the comment on the declaration of class
1815 /// StackProtectorDescriptor.
1817 SelectionDAGBuilder::visitSPDescriptorFailure(StackProtectorDescriptor &SPD) {
1818 const TargetLowering *TLI = TM.getTargetLowering();
1819 SDValue Chain = TLI->makeLibCall(DAG, RTLIB::STACKPROTECTOR_CHECK_FAIL,
1820 MVT::isVoid, 0, 0, false, getCurSDLoc(),
1821 false, false).second;
1825 /// visitBitTestHeader - This function emits necessary code to produce value
1826 /// suitable for "bit tests"
1827 void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B,
1828 MachineBasicBlock *SwitchBB) {
1829 // Subtract the minimum value
1830 SDValue SwitchOp = getValue(B.SValue);
1831 EVT VT = SwitchOp.getValueType();
1832 SDValue Sub = DAG.getNode(ISD::SUB, getCurSDLoc(), VT, SwitchOp,
1833 DAG.getConstant(B.First, VT));
1836 const TargetLowering *TLI = TM.getTargetLowering();
1837 SDValue RangeCmp = DAG.getSetCC(getCurSDLoc(),
1838 TLI->getSetCCResultType(*DAG.getContext(),
1839 Sub.getValueType()),
1840 Sub, DAG.getConstant(B.Range, VT),
1843 // Determine the type of the test operands.
1844 bool UsePtrType = false;
1845 if (!TLI->isTypeLegal(VT))
1848 for (unsigned i = 0, e = B.Cases.size(); i != e; ++i)
1849 if (!isUIntN(VT.getSizeInBits(), B.Cases[i].Mask)) {
1850 // Switch table case range are encoded into series of masks.
1851 // Just use pointer type, it's guaranteed to fit.
1857 VT = TLI->getPointerTy();
1858 Sub = DAG.getZExtOrTrunc(Sub, getCurSDLoc(), VT);
1861 B.RegVT = VT.getSimpleVT();
1862 B.Reg = FuncInfo.CreateReg(B.RegVT);
1863 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurSDLoc(),
1866 // Set NextBlock to be the MBB immediately after the current one, if any.
1867 // This is used to avoid emitting unnecessary branches to the next block.
1868 MachineBasicBlock *NextBlock = 0;
1869 MachineFunction::iterator BBI = SwitchBB;
1870 if (++BBI != FuncInfo.MF->end())
1873 MachineBasicBlock* MBB = B.Cases[0].ThisBB;
1875 addSuccessorWithWeight(SwitchBB, B.Default);
1876 addSuccessorWithWeight(SwitchBB, MBB);
1878 SDValue BrRange = DAG.getNode(ISD::BRCOND, getCurSDLoc(),
1879 MVT::Other, CopyTo, RangeCmp,
1880 DAG.getBasicBlock(B.Default));
1882 if (MBB != NextBlock)
1883 BrRange = DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, CopyTo,
1884 DAG.getBasicBlock(MBB));
1886 DAG.setRoot(BrRange);
1889 /// visitBitTestCase - this function produces one "bit test"
1890 void SelectionDAGBuilder::visitBitTestCase(BitTestBlock &BB,
1891 MachineBasicBlock* NextMBB,
1892 uint32_t BranchWeightToNext,
1895 MachineBasicBlock *SwitchBB) {
1897 SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurSDLoc(),
1900 unsigned PopCount = CountPopulation_64(B.Mask);
1901 const TargetLowering *TLI = TM.getTargetLowering();
1902 if (PopCount == 1) {
1903 // Testing for a single bit; just compare the shift count with what it
1904 // would need to be to shift a 1 bit in that position.
1905 Cmp = DAG.getSetCC(getCurSDLoc(),
1906 TLI->getSetCCResultType(*DAG.getContext(), VT),
1908 DAG.getConstant(countTrailingZeros(B.Mask), VT),
1910 } else if (PopCount == BB.Range) {
1911 // There is only one zero bit in the range, test for it directly.
1912 Cmp = DAG.getSetCC(getCurSDLoc(),
1913 TLI->getSetCCResultType(*DAG.getContext(), VT),
1915 DAG.getConstant(CountTrailingOnes_64(B.Mask), VT),
1918 // Make desired shift
1919 SDValue SwitchVal = DAG.getNode(ISD::SHL, getCurSDLoc(), VT,
1920 DAG.getConstant(1, VT), ShiftOp);
1922 // Emit bit tests and jumps
1923 SDValue AndOp = DAG.getNode(ISD::AND, getCurSDLoc(),
1924 VT, SwitchVal, DAG.getConstant(B.Mask, VT));
1925 Cmp = DAG.getSetCC(getCurSDLoc(),
1926 TLI->getSetCCResultType(*DAG.getContext(), VT),
1927 AndOp, DAG.getConstant(0, VT),
1931 // The branch weight from SwitchBB to B.TargetBB is B.ExtraWeight.
1932 addSuccessorWithWeight(SwitchBB, B.TargetBB, B.ExtraWeight);
1933 // The branch weight from SwitchBB to NextMBB is BranchWeightToNext.
1934 addSuccessorWithWeight(SwitchBB, NextMBB, BranchWeightToNext);
1936 SDValue BrAnd = DAG.getNode(ISD::BRCOND, getCurSDLoc(),
1937 MVT::Other, getControlRoot(),
1938 Cmp, DAG.getBasicBlock(B.TargetBB));
1940 // Set NextBlock to be the MBB immediately after the current one, if any.
1941 // This is used to avoid emitting unnecessary branches to the next block.
1942 MachineBasicBlock *NextBlock = 0;
1943 MachineFunction::iterator BBI = SwitchBB;
1944 if (++BBI != FuncInfo.MF->end())
1947 if (NextMBB != NextBlock)
1948 BrAnd = DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, BrAnd,
1949 DAG.getBasicBlock(NextMBB));
1954 void SelectionDAGBuilder::visitInvoke(const InvokeInst &I) {
1955 MachineBasicBlock *InvokeMBB = FuncInfo.MBB;
1957 // Retrieve successors.
1958 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
1959 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)];
1961 const Value *Callee(I.getCalledValue());
1962 const Function *Fn = dyn_cast<Function>(Callee);
1963 if (isa<InlineAsm>(Callee))
1965 else if (Fn && Fn->isIntrinsic()) {
1966 assert(Fn->getIntrinsicID() == Intrinsic::donothing);
1967 // Ignore invokes to @llvm.donothing: jump directly to the next BB.
1969 LowerCallTo(&I, getValue(Callee), false, LandingPad);
1971 // If the value of the invoke is used outside of its defining block, make it
1972 // available as a virtual register.
1973 CopyToExportRegsIfNeeded(&I);
1975 // Update successor info
1976 addSuccessorWithWeight(InvokeMBB, Return);
1977 addSuccessorWithWeight(InvokeMBB, LandingPad);
1979 // Drop into normal successor.
1980 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(),
1981 MVT::Other, getControlRoot(),
1982 DAG.getBasicBlock(Return)));
1985 void SelectionDAGBuilder::visitResume(const ResumeInst &RI) {
1986 llvm_unreachable("SelectionDAGBuilder shouldn't visit resume instructions!");
1989 void SelectionDAGBuilder::visitLandingPad(const LandingPadInst &LP) {
1990 assert(FuncInfo.MBB->isLandingPad() &&
1991 "Call to landingpad not in landing pad!");
1993 MachineBasicBlock *MBB = FuncInfo.MBB;
1994 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
1995 AddLandingPadInfo(LP, MMI, MBB);
1997 // If there aren't registers to copy the values into (e.g., during SjLj
1998 // exceptions), then don't bother to create these DAG nodes.
1999 const TargetLowering *TLI = TM.getTargetLowering();
2000 if (TLI->getExceptionPointerRegister() == 0 &&
2001 TLI->getExceptionSelectorRegister() == 0)
2004 SmallVector<EVT, 2> ValueVTs;
2005 ComputeValueVTs(*TLI, LP.getType(), ValueVTs);
2006 assert(ValueVTs.size() == 2 && "Only two-valued landingpads are supported");
2008 // Get the two live-in registers as SDValues. The physregs have already been
2009 // copied into virtual registers.
2011 Ops[0] = DAG.getZExtOrTrunc(
2012 DAG.getCopyFromReg(DAG.getEntryNode(), getCurSDLoc(),
2013 FuncInfo.ExceptionPointerVirtReg, TLI->getPointerTy()),
2014 getCurSDLoc(), ValueVTs[0]);
2015 Ops[1] = DAG.getZExtOrTrunc(
2016 DAG.getCopyFromReg(DAG.getEntryNode(), getCurSDLoc(),
2017 FuncInfo.ExceptionSelectorVirtReg, TLI->getPointerTy()),
2018 getCurSDLoc(), ValueVTs[1]);
2021 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
2022 DAG.getVTList(&ValueVTs[0], ValueVTs.size()),
2027 /// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for
2028 /// small case ranges).
2029 bool SelectionDAGBuilder::handleSmallSwitchRange(CaseRec& CR,
2030 CaseRecVector& WorkList,
2032 MachineBasicBlock *Default,
2033 MachineBasicBlock *SwitchBB) {
2034 // Size is the number of Cases represented by this range.
2035 size_t Size = CR.Range.second - CR.Range.first;
2039 // Get the MachineFunction which holds the current MBB. This is used when
2040 // inserting any additional MBBs necessary to represent the switch.
2041 MachineFunction *CurMF = FuncInfo.MF;
2043 // Figure out which block is immediately after the current one.
2044 MachineBasicBlock *NextBlock = 0;
2045 MachineFunction::iterator BBI = CR.CaseBB;
2047 if (++BBI != FuncInfo.MF->end())
2050 BranchProbabilityInfo *BPI = FuncInfo.BPI;
2051 // If any two of the cases has the same destination, and if one value
2052 // is the same as the other, but has one bit unset that the other has set,
2053 // use bit manipulation to do two compares at once. For example:
2054 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
2055 // TODO: This could be extended to merge any 2 cases in switches with 3 cases.
2056 // TODO: Handle cases where CR.CaseBB != SwitchBB.
2057 if (Size == 2 && CR.CaseBB == SwitchBB) {
2058 Case &Small = *CR.Range.first;
2059 Case &Big = *(CR.Range.second-1);
2061 if (Small.Low == Small.High && Big.Low == Big.High && Small.BB == Big.BB) {
2062 const APInt& SmallValue = cast<ConstantInt>(Small.Low)->getValue();
2063 const APInt& BigValue = cast<ConstantInt>(Big.Low)->getValue();
2065 // Check that there is only one bit different.
2066 if (BigValue.countPopulation() == SmallValue.countPopulation() + 1 &&
2067 (SmallValue | BigValue) == BigValue) {
2068 // Isolate the common bit.
2069 APInt CommonBit = BigValue & ~SmallValue;
2070 assert((SmallValue | CommonBit) == BigValue &&
2071 CommonBit.countPopulation() == 1 && "Not a common bit?");
2073 SDValue CondLHS = getValue(SV);
2074 EVT VT = CondLHS.getValueType();
2075 SDLoc DL = getCurSDLoc();
2077 SDValue Or = DAG.getNode(ISD::OR, DL, VT, CondLHS,
2078 DAG.getConstant(CommonBit, VT));
2079 SDValue Cond = DAG.getSetCC(DL, MVT::i1,
2080 Or, DAG.getConstant(BigValue, VT),
2083 // Update successor info.
2084 // Both Small and Big will jump to Small.BB, so we sum up the weights.
2085 addSuccessorWithWeight(SwitchBB, Small.BB,
2086 Small.ExtraWeight + Big.ExtraWeight);
2087 addSuccessorWithWeight(SwitchBB, Default,
2088 // The default destination is the first successor in IR.
2089 BPI ? BPI->getEdgeWeight(SwitchBB->getBasicBlock(), (unsigned)0) : 0);
2091 // Insert the true branch.
2092 SDValue BrCond = DAG.getNode(ISD::BRCOND, DL, MVT::Other,
2093 getControlRoot(), Cond,
2094 DAG.getBasicBlock(Small.BB));
2096 // Insert the false branch.
2097 BrCond = DAG.getNode(ISD::BR, DL, MVT::Other, BrCond,
2098 DAG.getBasicBlock(Default));
2100 DAG.setRoot(BrCond);
2106 // Order cases by weight so the most likely case will be checked first.
2107 uint32_t UnhandledWeights = 0;
2109 for (CaseItr I = CR.Range.first, IE = CR.Range.second; I != IE; ++I) {
2110 uint32_t IWeight = I->ExtraWeight;
2111 UnhandledWeights += IWeight;
2112 for (CaseItr J = CR.Range.first; J < I; ++J) {
2113 uint32_t JWeight = J->ExtraWeight;
2114 if (IWeight > JWeight)
2119 // Rearrange the case blocks so that the last one falls through if possible.
2120 Case &BackCase = *(CR.Range.second-1);
2122 NextBlock && Default != NextBlock && BackCase.BB != NextBlock) {
2123 // The last case block won't fall through into 'NextBlock' if we emit the
2124 // branches in this order. See if rearranging a case value would help.
2125 // We start at the bottom as it's the case with the least weight.
2126 for (Case *I = &*(CR.Range.second-2), *E = &*CR.Range.first-1; I != E; --I)
2127 if (I->BB == NextBlock) {
2128 std::swap(*I, BackCase);
2133 // Create a CaseBlock record representing a conditional branch to
2134 // the Case's target mbb if the value being switched on SV is equal
2136 MachineBasicBlock *CurBlock = CR.CaseBB;
2137 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
2138 MachineBasicBlock *FallThrough;
2140 FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock());
2141 CurMF->insert(BBI, FallThrough);
2143 // Put SV in a virtual register to make it available from the new blocks.
2144 ExportFromCurrentBlock(SV);
2146 // If the last case doesn't match, go to the default block.
2147 FallThrough = Default;
2150 const Value *RHS, *LHS, *MHS;
2152 if (I->High == I->Low) {
2153 // This is just small small case range :) containing exactly 1 case
2155 LHS = SV; RHS = I->High; MHS = NULL;
2158 LHS = I->Low; MHS = SV; RHS = I->High;
2161 // The false weight should be sum of all un-handled cases.
2162 UnhandledWeights -= I->ExtraWeight;
2163 CaseBlock CB(CC, LHS, RHS, MHS, /* truebb */ I->BB, /* falsebb */ FallThrough,
2165 /* trueweight */ I->ExtraWeight,
2166 /* falseweight */ UnhandledWeights);
2168 // If emitting the first comparison, just call visitSwitchCase to emit the
2169 // code into the current block. Otherwise, push the CaseBlock onto the
2170 // vector to be later processed by SDISel, and insert the node's MBB
2171 // before the next MBB.
2172 if (CurBlock == SwitchBB)
2173 visitSwitchCase(CB, SwitchBB);
2175 SwitchCases.push_back(CB);
2177 CurBlock = FallThrough;
2183 static inline bool areJTsAllowed(const TargetLowering &TLI) {
2184 return TLI.supportJumpTables() &&
2185 (TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
2186 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other));
2189 static APInt ComputeRange(const APInt &First, const APInt &Last) {
2190 uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1;
2191 APInt LastExt = Last.sext(BitWidth), FirstExt = First.sext(BitWidth);
2192 return (LastExt - FirstExt + 1ULL);
2195 /// handleJTSwitchCase - Emit jumptable for current switch case range
2196 bool SelectionDAGBuilder::handleJTSwitchCase(CaseRec &CR,
2197 CaseRecVector &WorkList,
2199 MachineBasicBlock *Default,
2200 MachineBasicBlock *SwitchBB) {
2201 Case& FrontCase = *CR.Range.first;
2202 Case& BackCase = *(CR.Range.second-1);
2204 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
2205 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
2207 APInt TSize(First.getBitWidth(), 0);
2208 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I)
2211 const TargetLowering *TLI = TM.getTargetLowering();
2212 if (!areJTsAllowed(*TLI) || TSize.ult(TLI->getMinimumJumpTableEntries()))
2215 APInt Range = ComputeRange(First, Last);
2216 // The density is TSize / Range. Require at least 40%.
2217 // It should not be possible for IntTSize to saturate for sane code, but make
2218 // sure we handle Range saturation correctly.
2219 uint64_t IntRange = Range.getLimitedValue(UINT64_MAX/10);
2220 uint64_t IntTSize = TSize.getLimitedValue(UINT64_MAX/10);
2221 if (IntTSize * 10 < IntRange * 4)
2224 DEBUG(dbgs() << "Lowering jump table\n"
2225 << "First entry: " << First << ". Last entry: " << Last << '\n'
2226 << "Range: " << Range << ". Size: " << TSize << ".\n\n");
2228 // Get the MachineFunction which holds the current MBB. This is used when
2229 // inserting any additional MBBs necessary to represent the switch.
2230 MachineFunction *CurMF = FuncInfo.MF;
2232 // Figure out which block is immediately after the current one.
2233 MachineFunction::iterator BBI = CR.CaseBB;
2236 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2238 // Create a new basic block to hold the code for loading the address
2239 // of the jump table, and jumping to it. Update successor information;
2240 // we will either branch to the default case for the switch, or the jump
2242 MachineBasicBlock *JumpTableBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2243 CurMF->insert(BBI, JumpTableBB);
2245 addSuccessorWithWeight(CR.CaseBB, Default);
2246 addSuccessorWithWeight(CR.CaseBB, JumpTableBB);
2248 // Build a vector of destination BBs, corresponding to each target
2249 // of the jump table. If the value of the jump table slot corresponds to
2250 // a case statement, push the case's BB onto the vector, otherwise, push
2252 std::vector<MachineBasicBlock*> DestBBs;
2254 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) {
2255 const APInt &Low = cast<ConstantInt>(I->Low)->getValue();
2256 const APInt &High = cast<ConstantInt>(I->High)->getValue();
2258 if (Low.sle(TEI) && TEI.sle(High)) {
2259 DestBBs.push_back(I->BB);
2263 DestBBs.push_back(Default);
2267 // Calculate weight for each unique destination in CR.
2268 DenseMap<MachineBasicBlock*, uint32_t> DestWeights;
2270 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
2271 DenseMap<MachineBasicBlock*, uint32_t>::iterator Itr =
2272 DestWeights.find(I->BB);
2273 if (Itr != DestWeights.end())
2274 Itr->second += I->ExtraWeight;
2276 DestWeights[I->BB] = I->ExtraWeight;
2279 // Update successor info. Add one edge to each unique successor.
2280 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs());
2281 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(),
2282 E = DestBBs.end(); I != E; ++I) {
2283 if (!SuccsHandled[(*I)->getNumber()]) {
2284 SuccsHandled[(*I)->getNumber()] = true;
2285 DenseMap<MachineBasicBlock*, uint32_t>::iterator Itr =
2286 DestWeights.find(*I);
2287 addSuccessorWithWeight(JumpTableBB, *I,
2288 Itr != DestWeights.end() ? Itr->second : 0);
2292 // Create a jump table index for this jump table.
2293 unsigned JTEncoding = TLI->getJumpTableEncoding();
2294 unsigned JTI = CurMF->getOrCreateJumpTableInfo(JTEncoding)
2295 ->createJumpTableIndex(DestBBs);
2297 // Set the jump table information so that we can codegen it as a second
2298 // MachineBasicBlock
2299 JumpTable JT(-1U, JTI, JumpTableBB, Default);
2300 JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == SwitchBB));
2301 if (CR.CaseBB == SwitchBB)
2302 visitJumpTableHeader(JT, JTH, SwitchBB);
2304 JTCases.push_back(JumpTableBlock(JTH, JT));
2308 /// handleBTSplitSwitchCase - emit comparison and split binary search tree into
2310 bool SelectionDAGBuilder::handleBTSplitSwitchCase(CaseRec& CR,
2311 CaseRecVector& WorkList,
2313 MachineBasicBlock* Default,
2314 MachineBasicBlock* SwitchBB) {
2315 // Get the MachineFunction which holds the current MBB. This is used when
2316 // inserting any additional MBBs necessary to represent the switch.
2317 MachineFunction *CurMF = FuncInfo.MF;
2319 // Figure out which block is immediately after the current one.
2320 MachineFunction::iterator BBI = CR.CaseBB;
2323 Case& FrontCase = *CR.Range.first;
2324 Case& BackCase = *(CR.Range.second-1);
2325 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2327 // Size is the number of Cases represented by this range.
2328 unsigned Size = CR.Range.second - CR.Range.first;
2330 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
2331 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
2333 CaseItr Pivot = CR.Range.first + Size/2;
2335 // Select optimal pivot, maximizing sum density of LHS and RHS. This will
2336 // (heuristically) allow us to emit JumpTable's later.
2337 APInt TSize(First.getBitWidth(), 0);
2338 for (CaseItr I = CR.Range.first, E = CR.Range.second;
2342 APInt LSize = FrontCase.size();
2343 APInt RSize = TSize-LSize;
2344 DEBUG(dbgs() << "Selecting best pivot: \n"
2345 << "First: " << First << ", Last: " << Last <<'\n'
2346 << "LSize: " << LSize << ", RSize: " << RSize << '\n');
2347 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second;
2349 const APInt &LEnd = cast<ConstantInt>(I->High)->getValue();
2350 const APInt &RBegin = cast<ConstantInt>(J->Low)->getValue();
2351 APInt Range = ComputeRange(LEnd, RBegin);
2352 assert((Range - 2ULL).isNonNegative() &&
2353 "Invalid case distance");
2354 // Use volatile double here to avoid excess precision issues on some hosts,
2355 // e.g. that use 80-bit X87 registers.
2356 volatile double LDensity =
2357 (double)LSize.roundToDouble() /
2358 (LEnd - First + 1ULL).roundToDouble();
2359 volatile double RDensity =
2360 (double)RSize.roundToDouble() /
2361 (Last - RBegin + 1ULL).roundToDouble();
2362 volatile double Metric = Range.logBase2()*(LDensity+RDensity);
2363 // Should always split in some non-trivial place
2364 DEBUG(dbgs() <<"=>Step\n"
2365 << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n'
2366 << "LDensity: " << LDensity
2367 << ", RDensity: " << RDensity << '\n'
2368 << "Metric: " << Metric << '\n');
2369 if (FMetric < Metric) {
2372 DEBUG(dbgs() << "Current metric set to: " << FMetric << '\n');
2379 const TargetLowering *TLI = TM.getTargetLowering();
2380 if (areJTsAllowed(*TLI)) {
2381 // If our case is dense we *really* should handle it earlier!
2382 assert((FMetric > 0) && "Should handle dense range earlier!");
2384 Pivot = CR.Range.first + Size/2;
2387 CaseRange LHSR(CR.Range.first, Pivot);
2388 CaseRange RHSR(Pivot, CR.Range.second);
2389 const Constant *C = Pivot->Low;
2390 MachineBasicBlock *FalseBB = 0, *TrueBB = 0;
2392 // We know that we branch to the LHS if the Value being switched on is
2393 // less than the Pivot value, C. We use this to optimize our binary
2394 // tree a bit, by recognizing that if SV is greater than or equal to the
2395 // LHS's Case Value, and that Case Value is exactly one less than the
2396 // Pivot's Value, then we can branch directly to the LHS's Target,
2397 // rather than creating a leaf node for it.
2398 if ((LHSR.second - LHSR.first) == 1 &&
2399 LHSR.first->High == CR.GE &&
2400 cast<ConstantInt>(C)->getValue() ==
2401 (cast<ConstantInt>(CR.GE)->getValue() + 1LL)) {
2402 TrueBB = LHSR.first->BB;
2404 TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2405 CurMF->insert(BBI, TrueBB);
2406 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR));
2408 // Put SV in a virtual register to make it available from the new blocks.
2409 ExportFromCurrentBlock(SV);
2412 // Similar to the optimization above, if the Value being switched on is
2413 // known to be less than the Constant CR.LT, and the current Case Value
2414 // is CR.LT - 1, then we can branch directly to the target block for
2415 // the current Case Value, rather than emitting a RHS leaf node for it.
2416 if ((RHSR.second - RHSR.first) == 1 && CR.LT &&
2417 cast<ConstantInt>(RHSR.first->Low)->getValue() ==
2418 (cast<ConstantInt>(CR.LT)->getValue() - 1LL)) {
2419 FalseBB = RHSR.first->BB;
2421 FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2422 CurMF->insert(BBI, FalseBB);
2423 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR));
2425 // Put SV in a virtual register to make it available from the new blocks.
2426 ExportFromCurrentBlock(SV);
2429 // Create a CaseBlock record representing a conditional branch to
2430 // the LHS node if the value being switched on SV is less than C.
2431 // Otherwise, branch to LHS.
2432 CaseBlock CB(ISD::SETLT, SV, C, NULL, TrueBB, FalseBB, CR.CaseBB);
2434 if (CR.CaseBB == SwitchBB)
2435 visitSwitchCase(CB, SwitchBB);
2437 SwitchCases.push_back(CB);
2442 /// handleBitTestsSwitchCase - if current case range has few destination and
2443 /// range span less, than machine word bitwidth, encode case range into series
2444 /// of masks and emit bit tests with these masks.
2445 bool SelectionDAGBuilder::handleBitTestsSwitchCase(CaseRec& CR,
2446 CaseRecVector& WorkList,
2448 MachineBasicBlock* Default,
2449 MachineBasicBlock* SwitchBB) {
2450 const TargetLowering *TLI = TM.getTargetLowering();
2451 EVT PTy = TLI->getPointerTy();
2452 unsigned IntPtrBits = PTy.getSizeInBits();
2454 Case& FrontCase = *CR.Range.first;
2455 Case& BackCase = *(CR.Range.second-1);
2457 // Get the MachineFunction which holds the current MBB. This is used when
2458 // inserting any additional MBBs necessary to represent the switch.
2459 MachineFunction *CurMF = FuncInfo.MF;
2461 // If target does not have legal shift left, do not emit bit tests at all.
2462 if (!TLI->isOperationLegal(ISD::SHL, PTy))
2466 for (CaseItr I = CR.Range.first, E = CR.Range.second;
2468 // Single case counts one, case range - two.
2469 numCmps += (I->Low == I->High ? 1 : 2);
2472 // Count unique destinations
2473 SmallSet<MachineBasicBlock*, 4> Dests;
2474 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2475 Dests.insert(I->BB);
2476 if (Dests.size() > 3)
2477 // Don't bother the code below, if there are too much unique destinations
2480 DEBUG(dbgs() << "Total number of unique destinations: "
2481 << Dests.size() << '\n'
2482 << "Total number of comparisons: " << numCmps << '\n');
2484 // Compute span of values.
2485 const APInt& minValue = cast<ConstantInt>(FrontCase.Low)->getValue();
2486 const APInt& maxValue = cast<ConstantInt>(BackCase.High)->getValue();
2487 APInt cmpRange = maxValue - minValue;
2489 DEBUG(dbgs() << "Compare range: " << cmpRange << '\n'
2490 << "Low bound: " << minValue << '\n'
2491 << "High bound: " << maxValue << '\n');
2493 if (cmpRange.uge(IntPtrBits) ||
2494 (!(Dests.size() == 1 && numCmps >= 3) &&
2495 !(Dests.size() == 2 && numCmps >= 5) &&
2496 !(Dests.size() >= 3 && numCmps >= 6)))
2499 DEBUG(dbgs() << "Emitting bit tests\n");
2500 APInt lowBound = APInt::getNullValue(cmpRange.getBitWidth());
2502 // Optimize the case where all the case values fit in a
2503 // word without having to subtract minValue. In this case,
2504 // we can optimize away the subtraction.
2505 if (minValue.isNonNegative() && maxValue.slt(IntPtrBits)) {
2506 cmpRange = maxValue;
2508 lowBound = minValue;
2511 CaseBitsVector CasesBits;
2512 unsigned i, count = 0;
2514 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2515 MachineBasicBlock* Dest = I->BB;
2516 for (i = 0; i < count; ++i)
2517 if (Dest == CasesBits[i].BB)
2521 assert((count < 3) && "Too much destinations to test!");
2522 CasesBits.push_back(CaseBits(0, Dest, 0, 0/*Weight*/));
2526 const APInt& lowValue = cast<ConstantInt>(I->Low)->getValue();
2527 const APInt& highValue = cast<ConstantInt>(I->High)->getValue();
2529 uint64_t lo = (lowValue - lowBound).getZExtValue();
2530 uint64_t hi = (highValue - lowBound).getZExtValue();
2531 CasesBits[i].ExtraWeight += I->ExtraWeight;
2533 for (uint64_t j = lo; j <= hi; j++) {
2534 CasesBits[i].Mask |= 1ULL << j;
2535 CasesBits[i].Bits++;
2539 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp());
2543 // Figure out which block is immediately after the current one.
2544 MachineFunction::iterator BBI = CR.CaseBB;
2547 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2549 DEBUG(dbgs() << "Cases:\n");
2550 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) {
2551 DEBUG(dbgs() << "Mask: " << CasesBits[i].Mask
2552 << ", Bits: " << CasesBits[i].Bits
2553 << ", BB: " << CasesBits[i].BB << '\n');
2555 MachineBasicBlock *CaseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2556 CurMF->insert(BBI, CaseBB);
2557 BTC.push_back(BitTestCase(CasesBits[i].Mask,
2559 CasesBits[i].BB, CasesBits[i].ExtraWeight));
2561 // Put SV in a virtual register to make it available from the new blocks.
2562 ExportFromCurrentBlock(SV);
2565 BitTestBlock BTB(lowBound, cmpRange, SV,
2566 -1U, MVT::Other, (CR.CaseBB == SwitchBB),
2567 CR.CaseBB, Default, BTC);
2569 if (CR.CaseBB == SwitchBB)
2570 visitBitTestHeader(BTB, SwitchBB);
2572 BitTestCases.push_back(BTB);
2577 /// Clusterify - Transform simple list of Cases into list of CaseRange's
2578 size_t SelectionDAGBuilder::Clusterify(CaseVector& Cases,
2579 const SwitchInst& SI) {
2582 BranchProbabilityInfo *BPI = FuncInfo.BPI;
2583 // Start with "simple" cases
2584 for (SwitchInst::ConstCaseIt i = SI.case_begin(), e = SI.case_end();
2586 const BasicBlock *SuccBB = i.getCaseSuccessor();
2587 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SuccBB];
2589 uint32_t ExtraWeight =
2590 BPI ? BPI->getEdgeWeight(SI.getParent(), i.getSuccessorIndex()) : 0;
2592 Cases.push_back(Case(i.getCaseValue(), i.getCaseValue(),
2593 SMBB, ExtraWeight));
2595 std::sort(Cases.begin(), Cases.end(), CaseCmp());
2597 // Merge case into clusters
2598 if (Cases.size() >= 2)
2599 // Must recompute end() each iteration because it may be
2600 // invalidated by erase if we hold on to it
2601 for (CaseItr I = Cases.begin(), J = llvm::next(Cases.begin());
2602 J != Cases.end(); ) {
2603 const APInt& nextValue = cast<ConstantInt>(J->Low)->getValue();
2604 const APInt& currentValue = cast<ConstantInt>(I->High)->getValue();
2605 MachineBasicBlock* nextBB = J->BB;
2606 MachineBasicBlock* currentBB = I->BB;
2608 // If the two neighboring cases go to the same destination, merge them
2609 // into a single case.
2610 if ((nextValue - currentValue == 1) && (currentBB == nextBB)) {
2612 I->ExtraWeight += J->ExtraWeight;
2619 for (CaseItr I=Cases.begin(), E=Cases.end(); I!=E; ++I, ++numCmps) {
2620 if (I->Low != I->High)
2621 // A range counts double, since it requires two compares.
2628 void SelectionDAGBuilder::UpdateSplitBlock(MachineBasicBlock *First,
2629 MachineBasicBlock *Last) {
2631 for (unsigned i = 0, e = JTCases.size(); i != e; ++i)
2632 if (JTCases[i].first.HeaderBB == First)
2633 JTCases[i].first.HeaderBB = Last;
2635 // Update BitTestCases.
2636 for (unsigned i = 0, e = BitTestCases.size(); i != e; ++i)
2637 if (BitTestCases[i].Parent == First)
2638 BitTestCases[i].Parent = Last;
2641 void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) {
2642 MachineBasicBlock *SwitchMBB = FuncInfo.MBB;
2644 // Figure out which block is immediately after the current one.
2645 MachineBasicBlock *NextBlock = 0;
2646 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()];
2648 // If there is only the default destination, branch to it if it is not the
2649 // next basic block. Otherwise, just fall through.
2650 if (!SI.getNumCases()) {
2651 // Update machine-CFG edges.
2653 // If this is not a fall-through branch, emit the branch.
2654 SwitchMBB->addSuccessor(Default);
2655 if (Default != NextBlock)
2656 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(),
2657 MVT::Other, getControlRoot(),
2658 DAG.getBasicBlock(Default)));
2663 // If there are any non-default case statements, create a vector of Cases
2664 // representing each one, and sort the vector so that we can efficiently
2665 // create a binary search tree from them.
2667 size_t numCmps = Clusterify(Cases, SI);
2668 DEBUG(dbgs() << "Clusterify finished. Total clusters: " << Cases.size()
2669 << ". Total compares: " << numCmps << '\n');
2672 // Get the Value to be switched on and default basic blocks, which will be
2673 // inserted into CaseBlock records, representing basic blocks in the binary
2675 const Value *SV = SI.getCondition();
2677 // Push the initial CaseRec onto the worklist
2678 CaseRecVector WorkList;
2679 WorkList.push_back(CaseRec(SwitchMBB,0,0,
2680 CaseRange(Cases.begin(),Cases.end())));
2682 while (!WorkList.empty()) {
2683 // Grab a record representing a case range to process off the worklist
2684 CaseRec CR = WorkList.back();
2685 WorkList.pop_back();
2687 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2690 // If the range has few cases (two or less) emit a series of specific
2692 if (handleSmallSwitchRange(CR, WorkList, SV, Default, SwitchMBB))
2695 // If the switch has more than N blocks, and is at least 40% dense, and the
2696 // target supports indirect branches, then emit a jump table rather than
2697 // lowering the switch to a binary tree of conditional branches.
2698 // N defaults to 4 and is controlled via TLS.getMinimumJumpTableEntries().
2699 if (handleJTSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2702 // Emit binary tree. We need to pick a pivot, and push left and right ranges
2703 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call.
2704 handleBTSplitSwitchCase(CR, WorkList, SV, Default, SwitchMBB);
2708 void SelectionDAGBuilder::visitIndirectBr(const IndirectBrInst &I) {
2709 MachineBasicBlock *IndirectBrMBB = FuncInfo.MBB;
2711 // Update machine-CFG edges with unique successors.
2712 SmallSet<BasicBlock*, 32> Done;
2713 for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i) {
2714 BasicBlock *BB = I.getSuccessor(i);
2715 bool Inserted = Done.insert(BB);
2719 MachineBasicBlock *Succ = FuncInfo.MBBMap[BB];
2720 addSuccessorWithWeight(IndirectBrMBB, Succ);
2723 DAG.setRoot(DAG.getNode(ISD::BRIND, getCurSDLoc(),
2724 MVT::Other, getControlRoot(),
2725 getValue(I.getAddress())));
2728 void SelectionDAGBuilder::visitFSub(const User &I) {
2729 // -0.0 - X --> fneg
2730 Type *Ty = I.getType();
2731 if (isa<Constant>(I.getOperand(0)) &&
2732 I.getOperand(0) == ConstantFP::getZeroValueForNegation(Ty)) {
2733 SDValue Op2 = getValue(I.getOperand(1));
2734 setValue(&I, DAG.getNode(ISD::FNEG, getCurSDLoc(),
2735 Op2.getValueType(), Op2));
2739 visitBinary(I, ISD::FSUB);
2742 void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) {
2743 SDValue Op1 = getValue(I.getOperand(0));
2744 SDValue Op2 = getValue(I.getOperand(1));
2745 setValue(&I, DAG.getNode(OpCode, getCurSDLoc(),
2746 Op1.getValueType(), Op1, Op2));
2749 void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) {
2750 SDValue Op1 = getValue(I.getOperand(0));
2751 SDValue Op2 = getValue(I.getOperand(1));
2753 EVT ShiftTy = TM.getTargetLowering()->getShiftAmountTy(Op2.getValueType());
2755 // Coerce the shift amount to the right type if we can.
2756 if (!I.getType()->isVectorTy() && Op2.getValueType() != ShiftTy) {
2757 unsigned ShiftSize = ShiftTy.getSizeInBits();
2758 unsigned Op2Size = Op2.getValueType().getSizeInBits();
2759 SDLoc DL = getCurSDLoc();
2761 // If the operand is smaller than the shift count type, promote it.
2762 if (ShiftSize > Op2Size)
2763 Op2 = DAG.getNode(ISD::ZERO_EXTEND, DL, ShiftTy, Op2);
2765 // If the operand is larger than the shift count type but the shift
2766 // count type has enough bits to represent any shift value, truncate
2767 // it now. This is a common case and it exposes the truncate to
2768 // optimization early.
2769 else if (ShiftSize >= Log2_32_Ceil(Op2.getValueType().getSizeInBits()))
2770 Op2 = DAG.getNode(ISD::TRUNCATE, DL, ShiftTy, Op2);
2771 // Otherwise we'll need to temporarily settle for some other convenient
2772 // type. Type legalization will make adjustments once the shiftee is split.
2774 Op2 = DAG.getZExtOrTrunc(Op2, DL, MVT::i32);
2777 setValue(&I, DAG.getNode(Opcode, getCurSDLoc(),
2778 Op1.getValueType(), Op1, Op2));
2781 void SelectionDAGBuilder::visitSDiv(const User &I) {
2782 SDValue Op1 = getValue(I.getOperand(0));
2783 SDValue Op2 = getValue(I.getOperand(1));
2785 // Turn exact SDivs into multiplications.
2786 // FIXME: This should be in DAGCombiner, but it doesn't have access to the
2788 if (isa<BinaryOperator>(&I) && cast<BinaryOperator>(&I)->isExact() &&
2789 !isa<ConstantSDNode>(Op1) &&
2790 isa<ConstantSDNode>(Op2) && !cast<ConstantSDNode>(Op2)->isNullValue())
2791 setValue(&I, TM.getTargetLowering()->BuildExactSDIV(Op1, Op2,
2792 getCurSDLoc(), DAG));
2794 setValue(&I, DAG.getNode(ISD::SDIV, getCurSDLoc(), Op1.getValueType(),
2798 void SelectionDAGBuilder::visitICmp(const User &I) {
2799 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2800 if (const ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2801 predicate = IC->getPredicate();
2802 else if (const ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2803 predicate = ICmpInst::Predicate(IC->getPredicate());
2804 SDValue Op1 = getValue(I.getOperand(0));
2805 SDValue Op2 = getValue(I.getOperand(1));
2806 ISD::CondCode Opcode = getICmpCondCode(predicate);
2808 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2809 setValue(&I, DAG.getSetCC(getCurSDLoc(), DestVT, Op1, Op2, Opcode));
2812 void SelectionDAGBuilder::visitFCmp(const User &I) {
2813 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2814 if (const FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2815 predicate = FC->getPredicate();
2816 else if (const ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2817 predicate = FCmpInst::Predicate(FC->getPredicate());
2818 SDValue Op1 = getValue(I.getOperand(0));
2819 SDValue Op2 = getValue(I.getOperand(1));
2820 ISD::CondCode Condition = getFCmpCondCode(predicate);
2821 if (TM.Options.NoNaNsFPMath)
2822 Condition = getFCmpCodeWithoutNaN(Condition);
2823 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2824 setValue(&I, DAG.getSetCC(getCurSDLoc(), DestVT, Op1, Op2, Condition));
2827 void SelectionDAGBuilder::visitSelect(const User &I) {
2828 SmallVector<EVT, 4> ValueVTs;
2829 ComputeValueVTs(*TM.getTargetLowering(), I.getType(), ValueVTs);
2830 unsigned NumValues = ValueVTs.size();
2831 if (NumValues == 0) return;
2833 SmallVector<SDValue, 4> Values(NumValues);
2834 SDValue Cond = getValue(I.getOperand(0));
2835 SDValue TrueVal = getValue(I.getOperand(1));
2836 SDValue FalseVal = getValue(I.getOperand(2));
2837 ISD::NodeType OpCode = Cond.getValueType().isVector() ?
2838 ISD::VSELECT : ISD::SELECT;
2840 for (unsigned i = 0; i != NumValues; ++i)
2841 Values[i] = DAG.getNode(OpCode, getCurSDLoc(),
2842 TrueVal.getNode()->getValueType(TrueVal.getResNo()+i),
2844 SDValue(TrueVal.getNode(),
2845 TrueVal.getResNo() + i),
2846 SDValue(FalseVal.getNode(),
2847 FalseVal.getResNo() + i));
2849 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
2850 DAG.getVTList(&ValueVTs[0], NumValues),
2851 &Values[0], NumValues));
2854 void SelectionDAGBuilder::visitTrunc(const User &I) {
2855 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2856 SDValue N = getValue(I.getOperand(0));
2857 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2858 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), DestVT, N));
2861 void SelectionDAGBuilder::visitZExt(const User &I) {
2862 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2863 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2864 SDValue N = getValue(I.getOperand(0));
2865 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2866 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurSDLoc(), DestVT, N));
2869 void SelectionDAGBuilder::visitSExt(const User &I) {
2870 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2871 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2872 SDValue N = getValue(I.getOperand(0));
2873 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2874 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurSDLoc(), DestVT, N));
2877 void SelectionDAGBuilder::visitFPTrunc(const User &I) {
2878 // FPTrunc is never a no-op cast, no need to check
2879 SDValue N = getValue(I.getOperand(0));
2880 const TargetLowering *TLI = TM.getTargetLowering();
2881 EVT DestVT = TLI->getValueType(I.getType());
2882 setValue(&I, DAG.getNode(ISD::FP_ROUND, getCurSDLoc(),
2884 DAG.getTargetConstant(0, TLI->getPointerTy())));
2887 void SelectionDAGBuilder::visitFPExt(const User &I) {
2888 // FPExt is never a no-op cast, no need to check
2889 SDValue N = getValue(I.getOperand(0));
2890 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2891 setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurSDLoc(), DestVT, N));
2894 void SelectionDAGBuilder::visitFPToUI(const User &I) {
2895 // FPToUI is never a no-op cast, no need to check
2896 SDValue N = getValue(I.getOperand(0));
2897 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2898 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurSDLoc(), DestVT, N));
2901 void SelectionDAGBuilder::visitFPToSI(const User &I) {
2902 // FPToSI is never a no-op cast, no need to check
2903 SDValue N = getValue(I.getOperand(0));
2904 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2905 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurSDLoc(), DestVT, N));
2908 void SelectionDAGBuilder::visitUIToFP(const User &I) {
2909 // UIToFP is never a no-op cast, no need to check
2910 SDValue N = getValue(I.getOperand(0));
2911 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2912 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurSDLoc(), DestVT, N));
2915 void SelectionDAGBuilder::visitSIToFP(const User &I) {
2916 // SIToFP is never a no-op cast, no need to check
2917 SDValue N = getValue(I.getOperand(0));
2918 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2919 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurSDLoc(), DestVT, N));
2922 void SelectionDAGBuilder::visitPtrToInt(const User &I) {
2923 // What to do depends on the size of the integer and the size of the pointer.
2924 // We can either truncate, zero extend, or no-op, accordingly.
2925 SDValue N = getValue(I.getOperand(0));
2926 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2927 setValue(&I, DAG.getZExtOrTrunc(N, getCurSDLoc(), DestVT));
2930 void SelectionDAGBuilder::visitIntToPtr(const User &I) {
2931 // What to do depends on the size of the integer and the size of the pointer.
2932 // We can either truncate, zero extend, or no-op, accordingly.
2933 SDValue N = getValue(I.getOperand(0));
2934 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2935 setValue(&I, DAG.getZExtOrTrunc(N, getCurSDLoc(), DestVT));
2938 void SelectionDAGBuilder::visitBitCast(const User &I) {
2939 SDValue N = getValue(I.getOperand(0));
2940 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2942 // BitCast assures us that source and destination are the same size so this is
2943 // either a BITCAST or a no-op.
2944 if (DestVT != N.getValueType())
2945 setValue(&I, DAG.getNode(ISD::BITCAST, getCurSDLoc(),
2946 DestVT, N)); // convert types.
2948 setValue(&I, N); // noop cast.
2951 void SelectionDAGBuilder::visitAddrSpaceCast(const User &I) {
2952 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2953 const Value *SV = I.getOperand(0);
2954 SDValue N = getValue(SV);
2955 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2957 unsigned SrcAS = SV->getType()->getPointerAddressSpace();
2958 unsigned DestAS = I.getType()->getPointerAddressSpace();
2960 if (!TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
2961 N = DAG.getAddrSpaceCast(getCurSDLoc(), DestVT, N, SrcAS, DestAS);
2966 void SelectionDAGBuilder::visitInsertElement(const User &I) {
2967 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2968 SDValue InVec = getValue(I.getOperand(0));
2969 SDValue InVal = getValue(I.getOperand(1));
2970 SDValue InIdx = DAG.getSExtOrTrunc(getValue(I.getOperand(2)),
2971 getCurSDLoc(), TLI.getVectorIdxTy());
2972 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurSDLoc(),
2973 TM.getTargetLowering()->getValueType(I.getType()),
2974 InVec, InVal, InIdx));
2977 void SelectionDAGBuilder::visitExtractElement(const User &I) {
2978 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2979 SDValue InVec = getValue(I.getOperand(0));
2980 SDValue InIdx = DAG.getSExtOrTrunc(getValue(I.getOperand(1)),
2981 getCurSDLoc(), TLI.getVectorIdxTy());
2982 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurSDLoc(),
2983 TM.getTargetLowering()->getValueType(I.getType()),
2987 // Utility for visitShuffleVector - Return true if every element in Mask,
2988 // beginning from position Pos and ending in Pos+Size, falls within the
2989 // specified sequential range [L, L+Pos). or is undef.
2990 static bool isSequentialInRange(const SmallVectorImpl<int> &Mask,
2991 unsigned Pos, unsigned Size, int Low) {
2992 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
2993 if (Mask[i] >= 0 && Mask[i] != Low)
2998 void SelectionDAGBuilder::visitShuffleVector(const User &I) {
2999 SDValue Src1 = getValue(I.getOperand(0));
3000 SDValue Src2 = getValue(I.getOperand(1));
3002 SmallVector<int, 8> Mask;
3003 ShuffleVectorInst::getShuffleMask(cast<Constant>(I.getOperand(2)), Mask);
3004 unsigned MaskNumElts = Mask.size();
3006 const TargetLowering *TLI = TM.getTargetLowering();
3007 EVT VT = TLI->getValueType(I.getType());
3008 EVT SrcVT = Src1.getValueType();
3009 unsigned SrcNumElts = SrcVT.getVectorNumElements();
3011 if (SrcNumElts == MaskNumElts) {
3012 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
3017 // Normalize the shuffle vector since mask and vector length don't match.
3018 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) {
3019 // Mask is longer than the source vectors and is a multiple of the source
3020 // vectors. We can use concatenate vector to make the mask and vectors
3022 if (SrcNumElts*2 == MaskNumElts) {
3023 // First check for Src1 in low and Src2 in high
3024 if (isSequentialInRange(Mask, 0, SrcNumElts, 0) &&
3025 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, SrcNumElts)) {
3026 // The shuffle is concatenating two vectors together.
3027 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurSDLoc(),
3031 // Then check for Src2 in low and Src1 in high
3032 if (isSequentialInRange(Mask, 0, SrcNumElts, SrcNumElts) &&
3033 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, 0)) {
3034 // The shuffle is concatenating two vectors together.
3035 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurSDLoc(),
3041 // Pad both vectors with undefs to make them the same length as the mask.
3042 unsigned NumConcat = MaskNumElts / SrcNumElts;
3043 bool Src1U = Src1.getOpcode() == ISD::UNDEF;
3044 bool Src2U = Src2.getOpcode() == ISD::UNDEF;
3045 SDValue UndefVal = DAG.getUNDEF(SrcVT);
3047 SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal);
3048 SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal);
3052 Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
3054 &MOps1[0], NumConcat);
3055 Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
3057 &MOps2[0], NumConcat);
3059 // Readjust mask for new input vector length.
3060 SmallVector<int, 8> MappedOps;
3061 for (unsigned i = 0; i != MaskNumElts; ++i) {
3063 if (Idx >= (int)SrcNumElts)
3064 Idx -= SrcNumElts - MaskNumElts;
3065 MappedOps.push_back(Idx);
3068 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
3073 if (SrcNumElts > MaskNumElts) {
3074 // Analyze the access pattern of the vector to see if we can extract
3075 // two subvectors and do the shuffle. The analysis is done by calculating
3076 // the range of elements the mask access on both vectors.
3077 int MinRange[2] = { static_cast<int>(SrcNumElts),
3078 static_cast<int>(SrcNumElts)};
3079 int MaxRange[2] = {-1, -1};
3081 for (unsigned i = 0; i != MaskNumElts; ++i) {
3087 if (Idx >= (int)SrcNumElts) {
3091 if (Idx > MaxRange[Input])
3092 MaxRange[Input] = Idx;
3093 if (Idx < MinRange[Input])
3094 MinRange[Input] = Idx;
3097 // Check if the access is smaller than the vector size and can we find
3098 // a reasonable extract index.
3099 int RangeUse[2] = { -1, -1 }; // 0 = Unused, 1 = Extract, -1 = Can not
3101 int StartIdx[2]; // StartIdx to extract from
3102 for (unsigned Input = 0; Input < 2; ++Input) {
3103 if (MinRange[Input] >= (int)SrcNumElts && MaxRange[Input] < 0) {
3104 RangeUse[Input] = 0; // Unused
3105 StartIdx[Input] = 0;
3109 // Find a good start index that is a multiple of the mask length. Then
3110 // see if the rest of the elements are in range.
3111 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts;
3112 if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts &&
3113 StartIdx[Input] + MaskNumElts <= SrcNumElts)
3114 RangeUse[Input] = 1; // Extract from a multiple of the mask length.
3117 if (RangeUse[0] == 0 && RangeUse[1] == 0) {
3118 setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used.
3121 if (RangeUse[0] >= 0 && RangeUse[1] >= 0) {
3122 // Extract appropriate subvector and generate a vector shuffle
3123 for (unsigned Input = 0; Input < 2; ++Input) {
3124 SDValue &Src = Input == 0 ? Src1 : Src2;
3125 if (RangeUse[Input] == 0)
3126 Src = DAG.getUNDEF(VT);
3128 Src = DAG.getNode(ISD::EXTRACT_SUBVECTOR, getCurSDLoc(), VT,
3129 Src, DAG.getConstant(StartIdx[Input],
3130 TLI->getVectorIdxTy()));
3133 // Calculate new mask.
3134 SmallVector<int, 8> MappedOps;
3135 for (unsigned i = 0; i != MaskNumElts; ++i) {
3138 if (Idx < (int)SrcNumElts)
3141 Idx -= SrcNumElts + StartIdx[1] - MaskNumElts;
3143 MappedOps.push_back(Idx);
3146 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
3152 // We can't use either concat vectors or extract subvectors so fall back to
3153 // replacing the shuffle with extract and build vector.
3154 // to insert and build vector.
3155 EVT EltVT = VT.getVectorElementType();
3156 EVT IdxVT = TLI->getVectorIdxTy();
3157 SmallVector<SDValue,8> Ops;
3158 for (unsigned i = 0; i != MaskNumElts; ++i) {
3163 Res = DAG.getUNDEF(EltVT);
3165 SDValue &Src = Idx < (int)SrcNumElts ? Src1 : Src2;
3166 if (Idx >= (int)SrcNumElts) Idx -= SrcNumElts;
3168 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurSDLoc(),
3169 EltVT, Src, DAG.getConstant(Idx, IdxVT));
3175 setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(),
3176 VT, &Ops[0], Ops.size()));
3179 void SelectionDAGBuilder::visitInsertValue(const InsertValueInst &I) {
3180 const Value *Op0 = I.getOperand(0);
3181 const Value *Op1 = I.getOperand(1);
3182 Type *AggTy = I.getType();
3183 Type *ValTy = Op1->getType();
3184 bool IntoUndef = isa<UndefValue>(Op0);
3185 bool FromUndef = isa<UndefValue>(Op1);
3187 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
3189 const TargetLowering *TLI = TM.getTargetLowering();
3190 SmallVector<EVT, 4> AggValueVTs;
3191 ComputeValueVTs(*TLI, AggTy, AggValueVTs);
3192 SmallVector<EVT, 4> ValValueVTs;
3193 ComputeValueVTs(*TLI, ValTy, ValValueVTs);
3195 unsigned NumAggValues = AggValueVTs.size();
3196 unsigned NumValValues = ValValueVTs.size();
3197 SmallVector<SDValue, 4> Values(NumAggValues);
3199 SDValue Agg = getValue(Op0);
3201 // Copy the beginning value(s) from the original aggregate.
3202 for (; i != LinearIndex; ++i)
3203 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
3204 SDValue(Agg.getNode(), Agg.getResNo() + i);
3205 // Copy values from the inserted value(s).
3207 SDValue Val = getValue(Op1);
3208 for (; i != LinearIndex + NumValValues; ++i)
3209 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) :
3210 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex);
3212 // Copy remaining value(s) from the original aggregate.
3213 for (; i != NumAggValues; ++i)
3214 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
3215 SDValue(Agg.getNode(), Agg.getResNo() + i);
3217 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
3218 DAG.getVTList(&AggValueVTs[0], NumAggValues),
3219 &Values[0], NumAggValues));
3222 void SelectionDAGBuilder::visitExtractValue(const ExtractValueInst &I) {
3223 const Value *Op0 = I.getOperand(0);
3224 Type *AggTy = Op0->getType();
3225 Type *ValTy = I.getType();
3226 bool OutOfUndef = isa<UndefValue>(Op0);
3228 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
3230 const TargetLowering *TLI = TM.getTargetLowering();
3231 SmallVector<EVT, 4> ValValueVTs;
3232 ComputeValueVTs(*TLI, ValTy, ValValueVTs);
3234 unsigned NumValValues = ValValueVTs.size();
3236 // Ignore a extractvalue that produces an empty object
3237 if (!NumValValues) {
3238 setValue(&I, DAG.getUNDEF(MVT(MVT::Other)));
3242 SmallVector<SDValue, 4> Values(NumValValues);
3244 SDValue Agg = getValue(Op0);
3245 // Copy out the selected value(s).
3246 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i)
3247 Values[i - LinearIndex] =
3249 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) :
3250 SDValue(Agg.getNode(), Agg.getResNo() + i);
3252 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
3253 DAG.getVTList(&ValValueVTs[0], NumValValues),
3254 &Values[0], NumValValues));
3257 void SelectionDAGBuilder::visitGetElementPtr(const User &I) {
3258 Value *Op0 = I.getOperand(0);
3259 // Note that the pointer operand may be a vector of pointers. Take the scalar
3260 // element which holds a pointer.
3261 Type *Ty = Op0->getType()->getScalarType();
3262 unsigned AS = Ty->getPointerAddressSpace();
3263 SDValue N = getValue(Op0);
3265 for (GetElementPtrInst::const_op_iterator OI = I.op_begin()+1, E = I.op_end();
3267 const Value *Idx = *OI;
3268 if (StructType *StTy = dyn_cast<StructType>(Ty)) {
3269 unsigned Field = cast<Constant>(Idx)->getUniqueInteger().getZExtValue();
3272 uint64_t Offset = TD->getStructLayout(StTy)->getElementOffset(Field);
3273 N = DAG.getNode(ISD::ADD, getCurSDLoc(), N.getValueType(), N,
3274 DAG.getConstant(Offset, N.getValueType()));
3277 Ty = StTy->getElementType(Field);
3279 Ty = cast<SequentialType>(Ty)->getElementType();
3281 // If this is a constant subscript, handle it quickly.
3282 const TargetLowering *TLI = TM.getTargetLowering();
3283 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
3284 if (CI->isZero()) continue;
3286 TD->getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
3288 EVT PTy = TLI->getPointerTy(AS);
3289 unsigned PtrBits = PTy.getSizeInBits();
3291 OffsVal = DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), PTy,
3292 DAG.getConstant(Offs, MVT::i64));
3294 OffsVal = DAG.getConstant(Offs, PTy);
3296 N = DAG.getNode(ISD::ADD, getCurSDLoc(), N.getValueType(), N,
3301 // N = N + Idx * ElementSize;
3302 APInt ElementSize = APInt(TLI->getPointerSizeInBits(AS),
3303 TD->getTypeAllocSize(Ty));
3304 SDValue IdxN = getValue(Idx);
3306 // If the index is smaller or larger than intptr_t, truncate or extend
3308 IdxN = DAG.getSExtOrTrunc(IdxN, getCurSDLoc(), N.getValueType());
3310 // If this is a multiply by a power of two, turn it into a shl
3311 // immediately. This is a very common case.
3312 if (ElementSize != 1) {
3313 if (ElementSize.isPowerOf2()) {
3314 unsigned Amt = ElementSize.logBase2();
3315 IdxN = DAG.getNode(ISD::SHL, getCurSDLoc(),
3316 N.getValueType(), IdxN,
3317 DAG.getConstant(Amt, IdxN.getValueType()));
3319 SDValue Scale = DAG.getConstant(ElementSize, IdxN.getValueType());
3320 IdxN = DAG.getNode(ISD::MUL, getCurSDLoc(),
3321 N.getValueType(), IdxN, Scale);
3325 N = DAG.getNode(ISD::ADD, getCurSDLoc(),
3326 N.getValueType(), N, IdxN);
3333 void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) {
3334 // If this is a fixed sized alloca in the entry block of the function,
3335 // allocate it statically on the stack.
3336 if (FuncInfo.StaticAllocaMap.count(&I))
3337 return; // getValue will auto-populate this.
3339 Type *Ty = I.getAllocatedType();
3340 const TargetLowering *TLI = TM.getTargetLowering();
3341 uint64_t TySize = TLI->getDataLayout()->getTypeAllocSize(Ty);
3343 std::max((unsigned)TLI->getDataLayout()->getPrefTypeAlignment(Ty),
3346 SDValue AllocSize = getValue(I.getArraySize());
3348 EVT IntPtr = TLI->getPointerTy();
3349 if (AllocSize.getValueType() != IntPtr)
3350 AllocSize = DAG.getZExtOrTrunc(AllocSize, getCurSDLoc(), IntPtr);
3352 AllocSize = DAG.getNode(ISD::MUL, getCurSDLoc(), IntPtr,
3354 DAG.getConstant(TySize, IntPtr));
3356 // Handle alignment. If the requested alignment is less than or equal to
3357 // the stack alignment, ignore it. If the size is greater than or equal to
3358 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
3359 unsigned StackAlign = TM.getFrameLowering()->getStackAlignment();
3360 if (Align <= StackAlign)
3363 // Round the size of the allocation up to the stack alignment size
3364 // by add SA-1 to the size.
3365 AllocSize = DAG.getNode(ISD::ADD, getCurSDLoc(),
3366 AllocSize.getValueType(), AllocSize,
3367 DAG.getIntPtrConstant(StackAlign-1));
3369 // Mask out the low bits for alignment purposes.
3370 AllocSize = DAG.getNode(ISD::AND, getCurSDLoc(),
3371 AllocSize.getValueType(), AllocSize,
3372 DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1)));
3374 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) };
3375 SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other);
3376 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, getCurSDLoc(),
3379 DAG.setRoot(DSA.getValue(1));
3381 // Inform the Frame Information that we have just allocated a variable-sized
3383 FuncInfo.MF->getFrameInfo()->CreateVariableSizedObject(Align ? Align : 1);
3386 void SelectionDAGBuilder::visitLoad(const LoadInst &I) {
3388 return visitAtomicLoad(I);
3390 const Value *SV = I.getOperand(0);
3391 SDValue Ptr = getValue(SV);
3393 Type *Ty = I.getType();
3395 bool isVolatile = I.isVolatile();
3396 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
3397 bool isInvariant = I.getMetadata("invariant.load") != 0;
3398 unsigned Alignment = I.getAlignment();
3399 const MDNode *TBAAInfo = I.getMetadata(LLVMContext::MD_tbaa);
3400 const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range);
3402 SmallVector<EVT, 4> ValueVTs;
3403 SmallVector<uint64_t, 4> Offsets;
3404 ComputeValueVTs(*TM.getTargetLowering(), Ty, ValueVTs, &Offsets);
3405 unsigned NumValues = ValueVTs.size();
3410 bool ConstantMemory = false;
3411 if (isVolatile || NumValues > MaxParallelChains)
3412 // Serialize volatile loads with other side effects.
3414 else if (AA->pointsToConstantMemory(
3415 AliasAnalysis::Location(SV, AA->getTypeStoreSize(Ty), TBAAInfo))) {
3416 // Do not serialize (non-volatile) loads of constant memory with anything.
3417 Root = DAG.getEntryNode();
3418 ConstantMemory = true;
3420 // Do not serialize non-volatile loads against each other.
3421 Root = DAG.getRoot();
3424 const TargetLowering *TLI = TM.getTargetLowering();
3426 Root = TLI->prepareVolatileOrAtomicLoad(Root, getCurSDLoc(), DAG);
3428 SmallVector<SDValue, 4> Values(NumValues);
3429 SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains),
3431 EVT PtrVT = Ptr.getValueType();
3432 unsigned ChainI = 0;
3433 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3434 // Serializing loads here may result in excessive register pressure, and
3435 // TokenFactor places arbitrary choke points on the scheduler. SD scheduling
3436 // could recover a bit by hoisting nodes upward in the chain by recognizing
3437 // they are side-effect free or do not alias. The optimizer should really
3438 // avoid this case by converting large object/array copies to llvm.memcpy
3439 // (MaxParallelChains should always remain as failsafe).
3440 if (ChainI == MaxParallelChains) {
3441 assert(PendingLoads.empty() && "PendingLoads must be serialized first");
3442 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(),
3443 MVT::Other, &Chains[0], ChainI);
3447 SDValue A = DAG.getNode(ISD::ADD, getCurSDLoc(),
3449 DAG.getConstant(Offsets[i], PtrVT));
3450 SDValue L = DAG.getLoad(ValueVTs[i], getCurSDLoc(), Root,
3451 A, MachinePointerInfo(SV, Offsets[i]), isVolatile,
3452 isNonTemporal, isInvariant, Alignment, TBAAInfo,
3456 Chains[ChainI] = L.getValue(1);
3459 if (!ConstantMemory) {
3460 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(),
3461 MVT::Other, &Chains[0], ChainI);
3465 PendingLoads.push_back(Chain);
3468 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
3469 DAG.getVTList(&ValueVTs[0], NumValues),
3470 &Values[0], NumValues));
3473 void SelectionDAGBuilder::visitStore(const StoreInst &I) {
3475 return visitAtomicStore(I);
3477 const Value *SrcV = I.getOperand(0);
3478 const Value *PtrV = I.getOperand(1);
3480 SmallVector<EVT, 4> ValueVTs;
3481 SmallVector<uint64_t, 4> Offsets;
3482 ComputeValueVTs(*TM.getTargetLowering(), SrcV->getType(), ValueVTs, &Offsets);
3483 unsigned NumValues = ValueVTs.size();
3487 // Get the lowered operands. Note that we do this after
3488 // checking if NumResults is zero, because with zero results
3489 // the operands won't have values in the map.
3490 SDValue Src = getValue(SrcV);
3491 SDValue Ptr = getValue(PtrV);
3493 SDValue Root = getRoot();
3494 SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains),
3496 EVT PtrVT = Ptr.getValueType();
3497 bool isVolatile = I.isVolatile();
3498 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
3499 unsigned Alignment = I.getAlignment();
3500 const MDNode *TBAAInfo = I.getMetadata(LLVMContext::MD_tbaa);
3502 unsigned ChainI = 0;
3503 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3504 // See visitLoad comments.
3505 if (ChainI == MaxParallelChains) {
3506 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(),
3507 MVT::Other, &Chains[0], ChainI);
3511 SDValue Add = DAG.getNode(ISD::ADD, getCurSDLoc(), PtrVT, Ptr,
3512 DAG.getConstant(Offsets[i], PtrVT));
3513 SDValue St = DAG.getStore(Root, getCurSDLoc(),
3514 SDValue(Src.getNode(), Src.getResNo() + i),
3515 Add, MachinePointerInfo(PtrV, Offsets[i]),
3516 isVolatile, isNonTemporal, Alignment, TBAAInfo);
3517 Chains[ChainI] = St;
3520 SDValue StoreNode = DAG.getNode(ISD::TokenFactor, getCurSDLoc(),
3521 MVT::Other, &Chains[0], ChainI);
3522 DAG.setRoot(StoreNode);
3525 static SDValue InsertFenceForAtomic(SDValue Chain, AtomicOrdering Order,
3526 SynchronizationScope Scope,
3527 bool Before, SDLoc dl,
3529 const TargetLowering &TLI) {
3530 // Fence, if necessary
3532 if (Order == AcquireRelease || Order == SequentiallyConsistent)
3534 else if (Order == Acquire || Order == Monotonic)
3537 if (Order == AcquireRelease)
3539 else if (Order == Release || Order == Monotonic)
3544 Ops[1] = DAG.getConstant(Order, TLI.getPointerTy());
3545 Ops[2] = DAG.getConstant(Scope, TLI.getPointerTy());
3546 return DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops, 3);
3549 void SelectionDAGBuilder::visitAtomicCmpXchg(const AtomicCmpXchgInst &I) {
3550 SDLoc dl = getCurSDLoc();
3551 AtomicOrdering Order = I.getOrdering();
3552 SynchronizationScope Scope = I.getSynchScope();
3554 SDValue InChain = getRoot();
3556 const TargetLowering *TLI = TM.getTargetLowering();
3557 if (TLI->getInsertFencesForAtomic())
3558 InChain = InsertFenceForAtomic(InChain, Order, Scope, true, dl,
3562 DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, dl,
3563 getValue(I.getCompareOperand()).getSimpleValueType(),
3565 getValue(I.getPointerOperand()),
3566 getValue(I.getCompareOperand()),
3567 getValue(I.getNewValOperand()),
3568 MachinePointerInfo(I.getPointerOperand()), 0 /* Alignment */,
3569 TLI->getInsertFencesForAtomic() ? Monotonic : Order,
3572 SDValue OutChain = L.getValue(1);
3574 if (TLI->getInsertFencesForAtomic())
3575 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3579 DAG.setRoot(OutChain);
3582 void SelectionDAGBuilder::visitAtomicRMW(const AtomicRMWInst &I) {
3583 SDLoc dl = getCurSDLoc();
3585 switch (I.getOperation()) {
3586 default: llvm_unreachable("Unknown atomicrmw operation");
3587 case AtomicRMWInst::Xchg: NT = ISD::ATOMIC_SWAP; break;
3588 case AtomicRMWInst::Add: NT = ISD::ATOMIC_LOAD_ADD; break;
3589 case AtomicRMWInst::Sub: NT = ISD::ATOMIC_LOAD_SUB; break;
3590 case AtomicRMWInst::And: NT = ISD::ATOMIC_LOAD_AND; break;
3591 case AtomicRMWInst::Nand: NT = ISD::ATOMIC_LOAD_NAND; break;
3592 case AtomicRMWInst::Or: NT = ISD::ATOMIC_LOAD_OR; break;
3593 case AtomicRMWInst::Xor: NT = ISD::ATOMIC_LOAD_XOR; break;
3594 case AtomicRMWInst::Max: NT = ISD::ATOMIC_LOAD_MAX; break;
3595 case AtomicRMWInst::Min: NT = ISD::ATOMIC_LOAD_MIN; break;
3596 case AtomicRMWInst::UMax: NT = ISD::ATOMIC_LOAD_UMAX; break;
3597 case AtomicRMWInst::UMin: NT = ISD::ATOMIC_LOAD_UMIN; break;
3599 AtomicOrdering Order = I.getOrdering();
3600 SynchronizationScope Scope = I.getSynchScope();
3602 SDValue InChain = getRoot();
3604 const TargetLowering *TLI = TM.getTargetLowering();
3605 if (TLI->getInsertFencesForAtomic())
3606 InChain = InsertFenceForAtomic(InChain, Order, Scope, true, dl,
3610 DAG.getAtomic(NT, dl,
3611 getValue(I.getValOperand()).getSimpleValueType(),
3613 getValue(I.getPointerOperand()),
3614 getValue(I.getValOperand()),
3615 I.getPointerOperand(), 0 /* Alignment */,
3616 TLI->getInsertFencesForAtomic() ? Monotonic : Order,
3619 SDValue OutChain = L.getValue(1);
3621 if (TLI->getInsertFencesForAtomic())
3622 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3626 DAG.setRoot(OutChain);
3629 void SelectionDAGBuilder::visitFence(const FenceInst &I) {
3630 SDLoc dl = getCurSDLoc();
3631 const TargetLowering *TLI = TM.getTargetLowering();
3634 Ops[1] = DAG.getConstant(I.getOrdering(), TLI->getPointerTy());
3635 Ops[2] = DAG.getConstant(I.getSynchScope(), TLI->getPointerTy());
3636 DAG.setRoot(DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops, 3));
3639 void SelectionDAGBuilder::visitAtomicLoad(const LoadInst &I) {
3640 SDLoc dl = getCurSDLoc();
3641 AtomicOrdering Order = I.getOrdering();
3642 SynchronizationScope Scope = I.getSynchScope();
3644 SDValue InChain = getRoot();
3646 const TargetLowering *TLI = TM.getTargetLowering();
3647 EVT VT = TLI->getValueType(I.getType());
3649 if (I.getAlignment() < VT.getSizeInBits() / 8)
3650 report_fatal_error("Cannot generate unaligned atomic load");
3652 InChain = TLI->prepareVolatileOrAtomicLoad(InChain, dl, DAG);
3654 DAG.getAtomic(ISD::ATOMIC_LOAD, dl, VT, VT, InChain,
3655 getValue(I.getPointerOperand()),
3656 I.getPointerOperand(), I.getAlignment(),
3657 TLI->getInsertFencesForAtomic() ? Monotonic : Order,
3660 SDValue OutChain = L.getValue(1);
3662 if (TLI->getInsertFencesForAtomic())
3663 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3667 DAG.setRoot(OutChain);
3670 void SelectionDAGBuilder::visitAtomicStore(const StoreInst &I) {
3671 SDLoc dl = getCurSDLoc();
3673 AtomicOrdering Order = I.getOrdering();
3674 SynchronizationScope Scope = I.getSynchScope();
3676 SDValue InChain = getRoot();
3678 const TargetLowering *TLI = TM.getTargetLowering();
3679 EVT VT = TLI->getValueType(I.getValueOperand()->getType());
3681 if (I.getAlignment() < VT.getSizeInBits() / 8)
3682 report_fatal_error("Cannot generate unaligned atomic store");
3684 if (TLI->getInsertFencesForAtomic())
3685 InChain = InsertFenceForAtomic(InChain, Order, Scope, true, dl,
3689 DAG.getAtomic(ISD::ATOMIC_STORE, dl, VT,
3691 getValue(I.getPointerOperand()),
3692 getValue(I.getValueOperand()),
3693 I.getPointerOperand(), I.getAlignment(),
3694 TLI->getInsertFencesForAtomic() ? Monotonic : Order,
3697 if (TLI->getInsertFencesForAtomic())
3698 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3701 DAG.setRoot(OutChain);
3704 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
3706 void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I,
3707 unsigned Intrinsic) {
3708 bool HasChain = !I.doesNotAccessMemory();
3709 bool OnlyLoad = HasChain && I.onlyReadsMemory();
3711 // Build the operand list.
3712 SmallVector<SDValue, 8> Ops;
3713 if (HasChain) { // If this intrinsic has side-effects, chainify it.
3715 // We don't need to serialize loads against other loads.
3716 Ops.push_back(DAG.getRoot());
3718 Ops.push_back(getRoot());
3722 // Info is set by getTgtMemInstrinsic
3723 TargetLowering::IntrinsicInfo Info;
3724 const TargetLowering *TLI = TM.getTargetLowering();
3725 bool IsTgtIntrinsic = TLI->getTgtMemIntrinsic(Info, I, Intrinsic);
3727 // Add the intrinsic ID as an integer operand if it's not a target intrinsic.
3728 if (!IsTgtIntrinsic || Info.opc == ISD::INTRINSIC_VOID ||
3729 Info.opc == ISD::INTRINSIC_W_CHAIN)
3730 Ops.push_back(DAG.getTargetConstant(Intrinsic, TLI->getPointerTy()));
3732 // Add all operands of the call to the operand list.
3733 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) {
3734 SDValue Op = getValue(I.getArgOperand(i));
3738 SmallVector<EVT, 4> ValueVTs;
3739 ComputeValueVTs(*TLI, I.getType(), ValueVTs);
3742 ValueVTs.push_back(MVT::Other);
3744 SDVTList VTs = DAG.getVTList(ValueVTs.data(), ValueVTs.size());
3748 if (IsTgtIntrinsic) {
3749 // This is target intrinsic that touches memory
3750 Result = DAG.getMemIntrinsicNode(Info.opc, getCurSDLoc(),
3751 VTs, &Ops[0], Ops.size(),
3753 MachinePointerInfo(Info.ptrVal, Info.offset),
3754 Info.align, Info.vol,
3755 Info.readMem, Info.writeMem);
3756 } else if (!HasChain) {
3757 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurSDLoc(),
3758 VTs, &Ops[0], Ops.size());
3759 } else if (!I.getType()->isVoidTy()) {
3760 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurSDLoc(),
3761 VTs, &Ops[0], Ops.size());
3763 Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurSDLoc(),
3764 VTs, &Ops[0], Ops.size());
3768 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1);
3770 PendingLoads.push_back(Chain);
3775 if (!I.getType()->isVoidTy()) {
3776 if (VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
3777 EVT VT = TLI->getValueType(PTy);
3778 Result = DAG.getNode(ISD::BITCAST, getCurSDLoc(), VT, Result);
3781 setValue(&I, Result);
3785 /// GetSignificand - Get the significand and build it into a floating-point
3786 /// number with exponent of 1:
3788 /// Op = (Op & 0x007fffff) | 0x3f800000;
3790 /// where Op is the hexadecimal representation of floating point value.
3792 GetSignificand(SelectionDAG &DAG, SDValue Op, SDLoc dl) {
3793 SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3794 DAG.getConstant(0x007fffff, MVT::i32));
3795 SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1,
3796 DAG.getConstant(0x3f800000, MVT::i32));
3797 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, t2);
3800 /// GetExponent - Get the exponent:
3802 /// (float)(int)(((Op & 0x7f800000) >> 23) - 127);
3804 /// where Op is the hexadecimal representation of floating point value.
3806 GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI,
3808 SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3809 DAG.getConstant(0x7f800000, MVT::i32));
3810 SDValue t1 = DAG.getNode(ISD::SRL, dl, MVT::i32, t0,
3811 DAG.getConstant(23, TLI.getPointerTy()));
3812 SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1,
3813 DAG.getConstant(127, MVT::i32));
3814 return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2);
3817 /// getF32Constant - Get 32-bit floating point constant.
3819 getF32Constant(SelectionDAG &DAG, unsigned Flt) {
3820 return DAG.getConstantFP(APFloat(APFloat::IEEEsingle, APInt(32, Flt)),
3824 /// expandExp - Lower an exp intrinsic. Handles the special sequences for
3825 /// limited-precision mode.
3826 static SDValue expandExp(SDLoc dl, SDValue Op, SelectionDAG &DAG,
3827 const TargetLowering &TLI) {
3828 if (Op.getValueType() == MVT::f32 &&
3829 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3831 // Put the exponent in the right bit position for later addition to the
3834 // #define LOG2OFe 1.4426950f
3835 // IntegerPartOfX = ((int32_t)(X * LOG2OFe));
3836 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3837 getF32Constant(DAG, 0x3fb8aa3b));
3838 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3840 // FractionalPartOfX = (X * LOG2OFe) - (float)IntegerPartOfX;
3841 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3842 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3844 // IntegerPartOfX <<= 23;
3845 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3846 DAG.getConstant(23, TLI.getPointerTy()));
3848 SDValue TwoToFracPartOfX;
3849 if (LimitFloatPrecision <= 6) {
3850 // For floating-point precision of 6:
3852 // TwoToFractionalPartOfX =
3854 // (0.735607626f + 0.252464424f * x) * x;
3856 // error 0.0144103317, which is 6 bits
3857 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3858 getF32Constant(DAG, 0x3e814304));
3859 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3860 getF32Constant(DAG, 0x3f3c50c8));
3861 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3862 TwoToFracPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3863 getF32Constant(DAG, 0x3f7f5e7e));
3864 } else if (LimitFloatPrecision <= 12) {
3865 // For floating-point precision of 12:
3867 // TwoToFractionalPartOfX =
3870 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3872 // 0.000107046256 error, which is 13 to 14 bits
3873 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3874 getF32Constant(DAG, 0x3da235e3));
3875 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3876 getF32Constant(DAG, 0x3e65b8f3));
3877 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3878 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3879 getF32Constant(DAG, 0x3f324b07));
3880 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3881 TwoToFracPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3882 getF32Constant(DAG, 0x3f7ff8fd));
3883 } else { // LimitFloatPrecision <= 18
3884 // For floating-point precision of 18:
3886 // TwoToFractionalPartOfX =
3890 // (0.554906021e-1f +
3891 // (0.961591928e-2f +
3892 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3894 // error 2.47208000*10^(-7), which is better than 18 bits
3895 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3896 getF32Constant(DAG, 0x3924b03e));
3897 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3898 getF32Constant(DAG, 0x3ab24b87));
3899 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3900 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3901 getF32Constant(DAG, 0x3c1d8c17));
3902 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3903 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3904 getF32Constant(DAG, 0x3d634a1d));
3905 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3906 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3907 getF32Constant(DAG, 0x3e75fe14));
3908 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3909 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3910 getF32Constant(DAG, 0x3f317234));
3911 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3912 TwoToFracPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3913 getF32Constant(DAG, 0x3f800000));
3916 // Add the exponent into the result in integer domain.
3917 SDValue t13 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, TwoToFracPartOfX);
3918 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3919 DAG.getNode(ISD::ADD, dl, MVT::i32,
3920 t13, IntegerPartOfX));
3923 // No special expansion.
3924 return DAG.getNode(ISD::FEXP, dl, Op.getValueType(), Op);
3927 /// expandLog - Lower a log intrinsic. Handles the special sequences for
3928 /// limited-precision mode.
3929 static SDValue expandLog(SDLoc dl, SDValue Op, SelectionDAG &DAG,
3930 const TargetLowering &TLI) {
3931 if (Op.getValueType() == MVT::f32 &&
3932 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3933 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3935 // Scale the exponent by log(2) [0.69314718f].
3936 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3937 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3938 getF32Constant(DAG, 0x3f317218));
3940 // Get the significand and build it into a floating-point number with
3942 SDValue X = GetSignificand(DAG, Op1, dl);
3944 SDValue LogOfMantissa;
3945 if (LimitFloatPrecision <= 6) {
3946 // For floating-point precision of 6:
3950 // (1.4034025f - 0.23903021f * x) * x;
3952 // error 0.0034276066, which is better than 8 bits
3953 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3954 getF32Constant(DAG, 0xbe74c456));
3955 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3956 getF32Constant(DAG, 0x3fb3a2b1));
3957 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3958 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3959 getF32Constant(DAG, 0x3f949a29));
3960 } else if (LimitFloatPrecision <= 12) {
3961 // For floating-point precision of 12:
3967 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x;
3969 // error 0.000061011436, which is 14 bits
3970 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3971 getF32Constant(DAG, 0xbd67b6d6));
3972 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3973 getF32Constant(DAG, 0x3ee4f4b8));
3974 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3975 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3976 getF32Constant(DAG, 0x3fbc278b));
3977 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3978 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3979 getF32Constant(DAG, 0x40348e95));
3980 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3981 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3982 getF32Constant(DAG, 0x3fdef31a));
3983 } else { // LimitFloatPrecision <= 18
3984 // For floating-point precision of 18:
3992 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x;
3994 // error 0.0000023660568, which is better than 18 bits
3995 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3996 getF32Constant(DAG, 0xbc91e5ac));
3997 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3998 getF32Constant(DAG, 0x3e4350aa));
3999 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4000 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4001 getF32Constant(DAG, 0x3f60d3e3));
4002 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4003 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4004 getF32Constant(DAG, 0x4011cdf0));
4005 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4006 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
4007 getF32Constant(DAG, 0x406cfd1c));
4008 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4009 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4010 getF32Constant(DAG, 0x408797cb));
4011 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4012 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
4013 getF32Constant(DAG, 0x4006dcab));
4016 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, LogOfMantissa);
4019 // No special expansion.
4020 return DAG.getNode(ISD::FLOG, dl, Op.getValueType(), Op);
4023 /// expandLog2 - Lower a log2 intrinsic. Handles the special sequences for
4024 /// limited-precision mode.
4025 static SDValue expandLog2(SDLoc dl, SDValue Op, SelectionDAG &DAG,
4026 const TargetLowering &TLI) {
4027 if (Op.getValueType() == MVT::f32 &&
4028 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4029 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
4031 // Get the exponent.
4032 SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl);
4034 // Get the significand and build it into a floating-point number with
4036 SDValue X = GetSignificand(DAG, Op1, dl);
4038 // Different possible minimax approximations of significand in
4039 // floating-point for various degrees of accuracy over [1,2].
4040 SDValue Log2ofMantissa;
4041 if (LimitFloatPrecision <= 6) {
4042 // For floating-point precision of 6:
4044 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x;
4046 // error 0.0049451742, which is more than 7 bits
4047 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4048 getF32Constant(DAG, 0xbeb08fe0));
4049 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4050 getF32Constant(DAG, 0x40019463));
4051 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4052 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4053 getF32Constant(DAG, 0x3fd6633d));
4054 } else if (LimitFloatPrecision <= 12) {
4055 // For floating-point precision of 12:
4061 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x;
4063 // error 0.0000876136000, which is better than 13 bits
4064 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4065 getF32Constant(DAG, 0xbda7262e));
4066 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4067 getF32Constant(DAG, 0x3f25280b));
4068 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4069 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4070 getF32Constant(DAG, 0x4007b923));
4071 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4072 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4073 getF32Constant(DAG, 0x40823e2f));
4074 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4075 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
4076 getF32Constant(DAG, 0x4020d29c));
4077 } else { // LimitFloatPrecision <= 18
4078 // For floating-point precision of 18:
4087 // 0.25691327e-1f * x) * x) * x) * x) * x) * x;
4089 // error 0.0000018516, which is better than 18 bits
4090 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4091 getF32Constant(DAG, 0xbcd2769e));
4092 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4093 getF32Constant(DAG, 0x3e8ce0b9));
4094 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4095 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4096 getF32Constant(DAG, 0x3fa22ae7));
4097 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4098 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4099 getF32Constant(DAG, 0x40525723));
4100 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4101 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
4102 getF32Constant(DAG, 0x40aaf200));
4103 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4104 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4105 getF32Constant(DAG, 0x40c39dad));
4106 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4107 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
4108 getF32Constant(DAG, 0x4042902c));
4111 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, Log2ofMantissa);
4114 // No special expansion.
4115 return DAG.getNode(ISD::FLOG2, dl, Op.getValueType(), Op);
4118 /// expandLog10 - Lower a log10 intrinsic. Handles the special sequences for
4119 /// limited-precision mode.
4120 static SDValue expandLog10(SDLoc dl, SDValue Op, SelectionDAG &DAG,
4121 const TargetLowering &TLI) {
4122 if (Op.getValueType() == MVT::f32 &&
4123 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4124 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
4126 // Scale the exponent by log10(2) [0.30102999f].
4127 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
4128 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
4129 getF32Constant(DAG, 0x3e9a209a));
4131 // Get the significand and build it into a floating-point number with
4133 SDValue X = GetSignificand(DAG, Op1, dl);
4135 SDValue Log10ofMantissa;
4136 if (LimitFloatPrecision <= 6) {
4137 // For floating-point precision of 6:
4139 // Log10ofMantissa =
4141 // (0.60948995f - 0.10380950f * x) * x;
4143 // error 0.0014886165, which is 6 bits
4144 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4145 getF32Constant(DAG, 0xbdd49a13));
4146 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4147 getF32Constant(DAG, 0x3f1c0789));
4148 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4149 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4150 getF32Constant(DAG, 0x3f011300));
4151 } else if (LimitFloatPrecision <= 12) {
4152 // For floating-point precision of 12:
4154 // Log10ofMantissa =
4157 // (-0.31664806f + 0.47637168e-1f * x) * x) * x;
4159 // error 0.00019228036, which is better than 12 bits
4160 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4161 getF32Constant(DAG, 0x3d431f31));
4162 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
4163 getF32Constant(DAG, 0x3ea21fb2));
4164 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4165 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4166 getF32Constant(DAG, 0x3f6ae232));
4167 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4168 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
4169 getF32Constant(DAG, 0x3f25f7c3));
4170 } else { // LimitFloatPrecision <= 18
4171 // For floating-point precision of 18:
4173 // Log10ofMantissa =
4178 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x;
4180 // error 0.0000037995730, which is better than 18 bits
4181 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4182 getF32Constant(DAG, 0x3c5d51ce));
4183 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
4184 getF32Constant(DAG, 0x3e00685a));
4185 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4186 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4187 getF32Constant(DAG, 0x3efb6798));
4188 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4189 SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
4190 getF32Constant(DAG, 0x3f88d192));
4191 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4192 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4193 getF32Constant(DAG, 0x3fc4316c));
4194 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4195 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8,
4196 getF32Constant(DAG, 0x3f57ce70));
4199 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, Log10ofMantissa);
4202 // No special expansion.
4203 return DAG.getNode(ISD::FLOG10, dl, Op.getValueType(), Op);
4206 /// expandExp2 - Lower an exp2 intrinsic. Handles the special sequences for
4207 /// limited-precision mode.
4208 static SDValue expandExp2(SDLoc dl, SDValue Op, SelectionDAG &DAG,
4209 const TargetLowering &TLI) {
4210 if (Op.getValueType() == MVT::f32 &&
4211 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4212 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Op);
4214 // FractionalPartOfX = x - (float)IntegerPartOfX;
4215 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
4216 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, Op, t1);
4218 // IntegerPartOfX <<= 23;
4219 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
4220 DAG.getConstant(23, TLI.getPointerTy()));
4222 SDValue TwoToFractionalPartOfX;
4223 if (LimitFloatPrecision <= 6) {
4224 // For floating-point precision of 6:
4226 // TwoToFractionalPartOfX =
4228 // (0.735607626f + 0.252464424f * x) * x;
4230 // error 0.0144103317, which is 6 bits
4231 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4232 getF32Constant(DAG, 0x3e814304));
4233 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4234 getF32Constant(DAG, 0x3f3c50c8));
4235 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4236 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4237 getF32Constant(DAG, 0x3f7f5e7e));
4238 } else if (LimitFloatPrecision <= 12) {
4239 // For floating-point precision of 12:
4241 // TwoToFractionalPartOfX =
4244 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
4246 // error 0.000107046256, which is 13 to 14 bits
4247 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4248 getF32Constant(DAG, 0x3da235e3));
4249 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4250 getF32Constant(DAG, 0x3e65b8f3));
4251 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4252 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4253 getF32Constant(DAG, 0x3f324b07));
4254 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4255 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4256 getF32Constant(DAG, 0x3f7ff8fd));
4257 } else { // LimitFloatPrecision <= 18
4258 // For floating-point precision of 18:
4260 // TwoToFractionalPartOfX =
4264 // (0.554906021e-1f +
4265 // (0.961591928e-2f +
4266 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
4267 // error 2.47208000*10^(-7), which is better than 18 bits
4268 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4269 getF32Constant(DAG, 0x3924b03e));
4270 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4271 getF32Constant(DAG, 0x3ab24b87));
4272 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4273 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4274 getF32Constant(DAG, 0x3c1d8c17));
4275 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4276 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4277 getF32Constant(DAG, 0x3d634a1d));
4278 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4279 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4280 getF32Constant(DAG, 0x3e75fe14));
4281 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4282 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
4283 getF32Constant(DAG, 0x3f317234));
4284 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
4285 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
4286 getF32Constant(DAG, 0x3f800000));
4289 // Add the exponent into the result in integer domain.
4290 SDValue t13 = DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4291 TwoToFractionalPartOfX);
4292 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4293 DAG.getNode(ISD::ADD, dl, MVT::i32,
4294 t13, IntegerPartOfX));
4297 // No special expansion.
4298 return DAG.getNode(ISD::FEXP2, dl, Op.getValueType(), Op);
4301 /// visitPow - Lower a pow intrinsic. Handles the special sequences for
4302 /// limited-precision mode with x == 10.0f.
4303 static SDValue expandPow(SDLoc dl, SDValue LHS, SDValue RHS,
4304 SelectionDAG &DAG, const TargetLowering &TLI) {
4305 bool IsExp10 = false;
4306 if (LHS.getValueType() == MVT::f32 && RHS.getValueType() == MVT::f32 &&
4307 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4308 if (ConstantFPSDNode *LHSC = dyn_cast<ConstantFPSDNode>(LHS)) {
4310 IsExp10 = LHSC->isExactlyValue(Ten);
4315 // Put the exponent in the right bit position for later addition to the
4318 // #define LOG2OF10 3.3219281f
4319 // IntegerPartOfX = (int32_t)(x * LOG2OF10);
4320 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, RHS,
4321 getF32Constant(DAG, 0x40549a78));
4322 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
4324 // FractionalPartOfX = x - (float)IntegerPartOfX;
4325 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
4326 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
4328 // IntegerPartOfX <<= 23;
4329 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
4330 DAG.getConstant(23, TLI.getPointerTy()));
4332 SDValue TwoToFractionalPartOfX;
4333 if (LimitFloatPrecision <= 6) {
4334 // For floating-point precision of 6:
4336 // twoToFractionalPartOfX =
4338 // (0.735607626f + 0.252464424f * x) * x;
4340 // error 0.0144103317, which is 6 bits
4341 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4342 getF32Constant(DAG, 0x3e814304));
4343 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4344 getF32Constant(DAG, 0x3f3c50c8));
4345 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4346 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4347 getF32Constant(DAG, 0x3f7f5e7e));
4348 } else if (LimitFloatPrecision <= 12) {
4349 // For floating-point precision of 12:
4351 // TwoToFractionalPartOfX =
4354 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
4356 // error 0.000107046256, which is 13 to 14 bits
4357 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4358 getF32Constant(DAG, 0x3da235e3));
4359 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4360 getF32Constant(DAG, 0x3e65b8f3));
4361 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4362 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4363 getF32Constant(DAG, 0x3f324b07));
4364 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4365 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4366 getF32Constant(DAG, 0x3f7ff8fd));
4367 } else { // LimitFloatPrecision <= 18
4368 // For floating-point precision of 18:
4370 // TwoToFractionalPartOfX =
4374 // (0.554906021e-1f +
4375 // (0.961591928e-2f +
4376 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
4377 // error 2.47208000*10^(-7), which is better than 18 bits
4378 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4379 getF32Constant(DAG, 0x3924b03e));
4380 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4381 getF32Constant(DAG, 0x3ab24b87));
4382 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4383 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4384 getF32Constant(DAG, 0x3c1d8c17));
4385 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4386 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4387 getF32Constant(DAG, 0x3d634a1d));
4388 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4389 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4390 getF32Constant(DAG, 0x3e75fe14));
4391 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4392 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
4393 getF32Constant(DAG, 0x3f317234));
4394 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
4395 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
4396 getF32Constant(DAG, 0x3f800000));
4399 SDValue t13 = DAG.getNode(ISD::BITCAST, dl,MVT::i32,TwoToFractionalPartOfX);
4400 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4401 DAG.getNode(ISD::ADD, dl, MVT::i32,
4402 t13, IntegerPartOfX));
4405 // No special expansion.
4406 return DAG.getNode(ISD::FPOW, dl, LHS.getValueType(), LHS, RHS);
4410 /// ExpandPowI - Expand a llvm.powi intrinsic.
4411 static SDValue ExpandPowI(SDLoc DL, SDValue LHS, SDValue RHS,
4412 SelectionDAG &DAG) {
4413 // If RHS is a constant, we can expand this out to a multiplication tree,
4414 // otherwise we end up lowering to a call to __powidf2 (for example). When
4415 // optimizing for size, we only want to do this if the expansion would produce
4416 // a small number of multiplies, otherwise we do the full expansion.
4417 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
4418 // Get the exponent as a positive value.
4419 unsigned Val = RHSC->getSExtValue();
4420 if ((int)Val < 0) Val = -Val;
4422 // powi(x, 0) -> 1.0
4424 return DAG.getConstantFP(1.0, LHS.getValueType());
4426 const Function *F = DAG.getMachineFunction().getFunction();
4427 if (!F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
4428 Attribute::OptimizeForSize) ||
4429 // If optimizing for size, don't insert too many multiplies. This
4430 // inserts up to 5 multiplies.
4431 CountPopulation_32(Val)+Log2_32(Val) < 7) {
4432 // We use the simple binary decomposition method to generate the multiply
4433 // sequence. There are more optimal ways to do this (for example,
4434 // powi(x,15) generates one more multiply than it should), but this has
4435 // the benefit of being both really simple and much better than a libcall.
4436 SDValue Res; // Logically starts equal to 1.0
4437 SDValue CurSquare = LHS;
4441 Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare);
4443 Res = CurSquare; // 1.0*CurSquare.
4446 CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(),
4447 CurSquare, CurSquare);
4451 // If the original was negative, invert the result, producing 1/(x*x*x).
4452 if (RHSC->getSExtValue() < 0)
4453 Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(),
4454 DAG.getConstantFP(1.0, LHS.getValueType()), Res);
4459 // Otherwise, expand to a libcall.
4460 return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS);
4463 // getTruncatedArgReg - Find underlying register used for an truncated
4465 static unsigned getTruncatedArgReg(const SDValue &N) {
4466 if (N.getOpcode() != ISD::TRUNCATE)
4469 const SDValue &Ext = N.getOperand(0);
4470 if (Ext.getOpcode() == ISD::AssertZext ||
4471 Ext.getOpcode() == ISD::AssertSext) {
4472 const SDValue &CFR = Ext.getOperand(0);
4473 if (CFR.getOpcode() == ISD::CopyFromReg)
4474 return cast<RegisterSDNode>(CFR.getOperand(1))->getReg();
4475 if (CFR.getOpcode() == ISD::TRUNCATE)
4476 return getTruncatedArgReg(CFR);
4481 /// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a function
4482 /// argument, create the corresponding DBG_VALUE machine instruction for it now.
4483 /// At the end of instruction selection, they will be inserted to the entry BB.
4485 SelectionDAGBuilder::EmitFuncArgumentDbgValue(const Value *V, MDNode *Variable,
4488 const Argument *Arg = dyn_cast<Argument>(V);
4492 MachineFunction &MF = DAG.getMachineFunction();
4493 const TargetInstrInfo *TII = DAG.getTarget().getInstrInfo();
4495 // Ignore inlined function arguments here.
4496 DIVariable DV(Variable);
4497 if (DV.isInlinedFnArgument(MF.getFunction()))
4500 Optional<MachineOperand> Op;
4501 // Some arguments' frame index is recorded during argument lowering.
4502 if (int FI = FuncInfo.getArgumentFrameIndex(Arg))
4503 Op = MachineOperand::CreateFI(FI);
4505 if (!Op && N.getNode()) {
4507 if (N.getOpcode() == ISD::CopyFromReg)
4508 Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg();
4510 Reg = getTruncatedArgReg(N);
4511 if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) {
4512 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4513 unsigned PR = RegInfo.getLiveInPhysReg(Reg);
4518 Op = MachineOperand::CreateReg(Reg, false);
4522 // Check if ValueMap has reg number.
4523 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
4524 if (VMI != FuncInfo.ValueMap.end())
4525 Op = MachineOperand::CreateReg(VMI->second, false);
4528 if (!Op && N.getNode())
4529 // Check if frame index is available.
4530 if (LoadSDNode *LNode = dyn_cast<LoadSDNode>(N.getNode()))
4531 if (FrameIndexSDNode *FINode =
4532 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode()))
4533 Op = MachineOperand::CreateFI(FINode->getIndex());
4538 // FIXME: This does not handle register-indirect values at offset 0.
4539 bool IsIndirect = Offset != 0;
4541 FuncInfo.ArgDbgValues.push_back(BuildMI(MF, getCurDebugLoc(),
4542 TII->get(TargetOpcode::DBG_VALUE),
4544 Op->getReg(), Offset, Variable));
4546 FuncInfo.ArgDbgValues.push_back(
4547 BuildMI(MF, getCurDebugLoc(), TII->get(TargetOpcode::DBG_VALUE))
4548 .addOperand(*Op).addImm(Offset).addMetadata(Variable));
4553 // VisualStudio defines setjmp as _setjmp
4554 #if defined(_MSC_VER) && defined(setjmp) && \
4555 !defined(setjmp_undefined_for_msvc)
4556 # pragma push_macro("setjmp")
4558 # define setjmp_undefined_for_msvc
4561 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
4562 /// we want to emit this as a call to a named external function, return the name
4563 /// otherwise lower it and return null.
4565 SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) {
4566 const TargetLowering *TLI = TM.getTargetLowering();
4567 SDLoc sdl = getCurSDLoc();
4568 DebugLoc dl = getCurDebugLoc();
4571 switch (Intrinsic) {
4573 // By default, turn this into a target intrinsic node.
4574 visitTargetIntrinsic(I, Intrinsic);
4576 case Intrinsic::vastart: visitVAStart(I); return 0;
4577 case Intrinsic::vaend: visitVAEnd(I); return 0;
4578 case Intrinsic::vacopy: visitVACopy(I); return 0;
4579 case Intrinsic::returnaddress:
4580 setValue(&I, DAG.getNode(ISD::RETURNADDR, sdl, TLI->getPointerTy(),
4581 getValue(I.getArgOperand(0))));
4583 case Intrinsic::frameaddress:
4584 setValue(&I, DAG.getNode(ISD::FRAMEADDR, sdl, TLI->getPointerTy(),
4585 getValue(I.getArgOperand(0))));
4587 case Intrinsic::setjmp:
4588 return &"_setjmp"[!TLI->usesUnderscoreSetJmp()];
4589 case Intrinsic::longjmp:
4590 return &"_longjmp"[!TLI->usesUnderscoreLongJmp()];
4591 case Intrinsic::memcpy: {
4592 // Assert for address < 256 since we support only user defined address
4594 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4596 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4598 "Unknown address space");
4599 SDValue Op1 = getValue(I.getArgOperand(0));
4600 SDValue Op2 = getValue(I.getArgOperand(1));
4601 SDValue Op3 = getValue(I.getArgOperand(2));
4602 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4604 Align = 1; // @llvm.memcpy defines 0 and 1 to both mean no alignment.
4605 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4606 DAG.setRoot(DAG.getMemcpy(getRoot(), sdl, Op1, Op2, Op3, Align, isVol, false,
4607 MachinePointerInfo(I.getArgOperand(0)),
4608 MachinePointerInfo(I.getArgOperand(1))));
4611 case Intrinsic::memset: {
4612 // Assert for address < 256 since we support only user defined address
4614 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4616 "Unknown address space");
4617 SDValue Op1 = getValue(I.getArgOperand(0));
4618 SDValue Op2 = getValue(I.getArgOperand(1));
4619 SDValue Op3 = getValue(I.getArgOperand(2));
4620 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4622 Align = 1; // @llvm.memset defines 0 and 1 to both mean no alignment.
4623 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4624 DAG.setRoot(DAG.getMemset(getRoot(), sdl, Op1, Op2, Op3, Align, isVol,
4625 MachinePointerInfo(I.getArgOperand(0))));
4628 case Intrinsic::memmove: {
4629 // Assert for address < 256 since we support only user defined address
4631 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4633 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4635 "Unknown address space");
4636 SDValue Op1 = getValue(I.getArgOperand(0));
4637 SDValue Op2 = getValue(I.getArgOperand(1));
4638 SDValue Op3 = getValue(I.getArgOperand(2));
4639 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4641 Align = 1; // @llvm.memmove defines 0 and 1 to both mean no alignment.
4642 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4643 DAG.setRoot(DAG.getMemmove(getRoot(), sdl, Op1, Op2, Op3, Align, isVol,
4644 MachinePointerInfo(I.getArgOperand(0)),
4645 MachinePointerInfo(I.getArgOperand(1))));
4648 case Intrinsic::dbg_declare: {
4649 const DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
4650 MDNode *Variable = DI.getVariable();
4651 const Value *Address = DI.getAddress();
4652 DIVariable DIVar(Variable);
4653 assert((!DIVar || DIVar.isVariable()) &&
4654 "Variable in DbgDeclareInst should be either null or a DIVariable.");
4655 if (!Address || !DIVar) {
4656 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4660 // Check if address has undef value.
4661 if (isa<UndefValue>(Address) ||
4662 (Address->use_empty() && !isa<Argument>(Address))) {
4663 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4667 SDValue &N = NodeMap[Address];
4668 if (!N.getNode() && isa<Argument>(Address))
4669 // Check unused arguments map.
4670 N = UnusedArgNodeMap[Address];
4673 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
4674 Address = BCI->getOperand(0);
4675 // Parameters are handled specially.
4677 (DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable ||
4678 isa<Argument>(Address));
4680 const AllocaInst *AI = dyn_cast<AllocaInst>(Address);
4682 if (isParameter && !AI) {
4683 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode());
4685 // Byval parameter. We have a frame index at this point.
4686 SDV = DAG.getDbgValue(Variable, FINode->getIndex(),
4687 0, dl, SDNodeOrder);
4689 // Address is an argument, so try to emit its dbg value using
4690 // virtual register info from the FuncInfo.ValueMap.
4691 EmitFuncArgumentDbgValue(Address, Variable, 0, N);
4695 SDV = DAG.getDbgValue(Variable, N.getNode(), N.getResNo(),
4696 0, dl, SDNodeOrder);
4698 // Can't do anything with other non-AI cases yet.
4699 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4700 DEBUG(dbgs() << "non-AllocaInst issue for Address: \n\t");
4701 DEBUG(Address->dump());
4704 DAG.AddDbgValue(SDV, N.getNode(), isParameter);
4706 // If Address is an argument then try to emit its dbg value using
4707 // virtual register info from the FuncInfo.ValueMap.
4708 if (!EmitFuncArgumentDbgValue(Address, Variable, 0, N)) {
4709 // If variable is pinned by a alloca in dominating bb then
4710 // use StaticAllocaMap.
4711 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Address)) {
4712 if (AI->getParent() != DI.getParent()) {
4713 DenseMap<const AllocaInst*, int>::iterator SI =
4714 FuncInfo.StaticAllocaMap.find(AI);
4715 if (SI != FuncInfo.StaticAllocaMap.end()) {
4716 SDV = DAG.getDbgValue(Variable, SI->second,
4717 0, dl, SDNodeOrder);
4718 DAG.AddDbgValue(SDV, 0, false);
4723 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4728 case Intrinsic::dbg_value: {
4729 const DbgValueInst &DI = cast<DbgValueInst>(I);
4730 DIVariable DIVar(DI.getVariable());
4731 assert((!DIVar || DIVar.isVariable()) &&
4732 "Variable in DbgValueInst should be either null or a DIVariable.");
4736 MDNode *Variable = DI.getVariable();
4737 uint64_t Offset = DI.getOffset();
4738 const Value *V = DI.getValue();
4743 if (isa<ConstantInt>(V) || isa<ConstantFP>(V) || isa<UndefValue>(V)) {
4744 SDV = DAG.getDbgValue(Variable, V, Offset, dl, SDNodeOrder);
4745 DAG.AddDbgValue(SDV, 0, false);
4747 // Do not use getValue() in here; we don't want to generate code at
4748 // this point if it hasn't been done yet.
4749 SDValue N = NodeMap[V];
4750 if (!N.getNode() && isa<Argument>(V))
4751 // Check unused arguments map.
4752 N = UnusedArgNodeMap[V];
4754 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, N)) {
4755 SDV = DAG.getDbgValue(Variable, N.getNode(),
4756 N.getResNo(), Offset, dl, SDNodeOrder);
4757 DAG.AddDbgValue(SDV, N.getNode(), false);
4759 } else if (!V->use_empty() ) {
4760 // Do not call getValue(V) yet, as we don't want to generate code.
4761 // Remember it for later.
4762 DanglingDebugInfo DDI(&DI, dl, SDNodeOrder);
4763 DanglingDebugInfoMap[V] = DDI;
4765 // We may expand this to cover more cases. One case where we have no
4766 // data available is an unreferenced parameter.
4767 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4771 // Build a debug info table entry.
4772 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
4773 V = BCI->getOperand(0);
4774 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
4775 // Don't handle byval struct arguments or VLAs, for example.
4777 DEBUG(dbgs() << "Dropping debug location info for:\n " << DI << "\n");
4778 DEBUG(dbgs() << " Last seen at:\n " << *V << "\n");
4781 DenseMap<const AllocaInst*, int>::iterator SI =
4782 FuncInfo.StaticAllocaMap.find(AI);
4783 if (SI == FuncInfo.StaticAllocaMap.end())
4785 int FI = SI->second;
4787 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4788 if (!DI.getDebugLoc().isUnknown() && MMI.hasDebugInfo())
4789 MMI.setVariableDbgInfo(Variable, FI, DI.getDebugLoc());
4793 case Intrinsic::eh_typeid_for: {
4794 // Find the type id for the given typeinfo.
4795 GlobalVariable *GV = ExtractTypeInfo(I.getArgOperand(0));
4796 unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(GV);
4797 Res = DAG.getConstant(TypeID, MVT::i32);
4802 case Intrinsic::eh_return_i32:
4803 case Intrinsic::eh_return_i64:
4804 DAG.getMachineFunction().getMMI().setCallsEHReturn(true);
4805 DAG.setRoot(DAG.getNode(ISD::EH_RETURN, sdl,
4808 getValue(I.getArgOperand(0)),
4809 getValue(I.getArgOperand(1))));
4811 case Intrinsic::eh_unwind_init:
4812 DAG.getMachineFunction().getMMI().setCallsUnwindInit(true);
4814 case Intrinsic::eh_dwarf_cfa: {
4815 SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getArgOperand(0)), sdl,
4816 TLI->getPointerTy());
4817 SDValue Offset = DAG.getNode(ISD::ADD, sdl,
4818 CfaArg.getValueType(),
4819 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, sdl,
4820 CfaArg.getValueType()),
4822 SDValue FA = DAG.getNode(ISD::FRAMEADDR, sdl,
4823 TLI->getPointerTy(),
4824 DAG.getConstant(0, TLI->getPointerTy()));
4825 setValue(&I, DAG.getNode(ISD::ADD, sdl, FA.getValueType(),
4829 case Intrinsic::eh_sjlj_callsite: {
4830 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4831 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(0));
4832 assert(CI && "Non-constant call site value in eh.sjlj.callsite!");
4833 assert(MMI.getCurrentCallSite() == 0 && "Overlapping call sites!");
4835 MMI.setCurrentCallSite(CI->getZExtValue());
4838 case Intrinsic::eh_sjlj_functioncontext: {
4839 // Get and store the index of the function context.
4840 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4842 cast<AllocaInst>(I.getArgOperand(0)->stripPointerCasts());
4843 int FI = FuncInfo.StaticAllocaMap[FnCtx];
4844 MFI->setFunctionContextIndex(FI);
4847 case Intrinsic::eh_sjlj_setjmp: {
4850 Ops[1] = getValue(I.getArgOperand(0));
4851 SDValue Op = DAG.getNode(ISD::EH_SJLJ_SETJMP, sdl,
4852 DAG.getVTList(MVT::i32, MVT::Other),
4854 setValue(&I, Op.getValue(0));
4855 DAG.setRoot(Op.getValue(1));
4858 case Intrinsic::eh_sjlj_longjmp: {
4859 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, sdl, MVT::Other,
4860 getRoot(), getValue(I.getArgOperand(0))));
4864 case Intrinsic::x86_mmx_pslli_w:
4865 case Intrinsic::x86_mmx_pslli_d:
4866 case Intrinsic::x86_mmx_pslli_q:
4867 case Intrinsic::x86_mmx_psrli_w:
4868 case Intrinsic::x86_mmx_psrli_d:
4869 case Intrinsic::x86_mmx_psrli_q:
4870 case Intrinsic::x86_mmx_psrai_w:
4871 case Intrinsic::x86_mmx_psrai_d: {
4872 SDValue ShAmt = getValue(I.getArgOperand(1));
4873 if (isa<ConstantSDNode>(ShAmt)) {
4874 visitTargetIntrinsic(I, Intrinsic);
4877 unsigned NewIntrinsic = 0;
4878 EVT ShAmtVT = MVT::v2i32;
4879 switch (Intrinsic) {
4880 case Intrinsic::x86_mmx_pslli_w:
4881 NewIntrinsic = Intrinsic::x86_mmx_psll_w;
4883 case Intrinsic::x86_mmx_pslli_d:
4884 NewIntrinsic = Intrinsic::x86_mmx_psll_d;
4886 case Intrinsic::x86_mmx_pslli_q:
4887 NewIntrinsic = Intrinsic::x86_mmx_psll_q;
4889 case Intrinsic::x86_mmx_psrli_w:
4890 NewIntrinsic = Intrinsic::x86_mmx_psrl_w;
4892 case Intrinsic::x86_mmx_psrli_d:
4893 NewIntrinsic = Intrinsic::x86_mmx_psrl_d;
4895 case Intrinsic::x86_mmx_psrli_q:
4896 NewIntrinsic = Intrinsic::x86_mmx_psrl_q;
4898 case Intrinsic::x86_mmx_psrai_w:
4899 NewIntrinsic = Intrinsic::x86_mmx_psra_w;
4901 case Intrinsic::x86_mmx_psrai_d:
4902 NewIntrinsic = Intrinsic::x86_mmx_psra_d;
4904 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4907 // The vector shift intrinsics with scalars uses 32b shift amounts but
4908 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
4910 // We must do this early because v2i32 is not a legal type.
4913 ShOps[1] = DAG.getConstant(0, MVT::i32);
4914 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, sdl, ShAmtVT, &ShOps[0], 2);
4915 EVT DestVT = TLI->getValueType(I.getType());
4916 ShAmt = DAG.getNode(ISD::BITCAST, sdl, DestVT, ShAmt);
4917 Res = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, sdl, DestVT,
4918 DAG.getConstant(NewIntrinsic, MVT::i32),
4919 getValue(I.getArgOperand(0)), ShAmt);
4923 case Intrinsic::x86_avx_vinsertf128_pd_256:
4924 case Intrinsic::x86_avx_vinsertf128_ps_256:
4925 case Intrinsic::x86_avx_vinsertf128_si_256:
4926 case Intrinsic::x86_avx2_vinserti128: {
4927 EVT DestVT = TLI->getValueType(I.getType());
4928 EVT ElVT = TLI->getValueType(I.getArgOperand(1)->getType());
4929 uint64_t Idx = (cast<ConstantInt>(I.getArgOperand(2))->getZExtValue() & 1) *
4930 ElVT.getVectorNumElements();
4931 Res = DAG.getNode(ISD::INSERT_SUBVECTOR, sdl, DestVT,
4932 getValue(I.getArgOperand(0)),
4933 getValue(I.getArgOperand(1)),
4934 DAG.getConstant(Idx, TLI->getVectorIdxTy()));
4938 case Intrinsic::x86_avx_vextractf128_pd_256:
4939 case Intrinsic::x86_avx_vextractf128_ps_256:
4940 case Intrinsic::x86_avx_vextractf128_si_256:
4941 case Intrinsic::x86_avx2_vextracti128: {
4942 EVT DestVT = TLI->getValueType(I.getType());
4943 uint64_t Idx = (cast<ConstantInt>(I.getArgOperand(1))->getZExtValue() & 1) *
4944 DestVT.getVectorNumElements();
4945 Res = DAG.getNode(ISD::EXTRACT_SUBVECTOR, sdl, DestVT,
4946 getValue(I.getArgOperand(0)),
4947 DAG.getConstant(Idx, TLI->getVectorIdxTy()));
4951 case Intrinsic::convertff:
4952 case Intrinsic::convertfsi:
4953 case Intrinsic::convertfui:
4954 case Intrinsic::convertsif:
4955 case Intrinsic::convertuif:
4956 case Intrinsic::convertss:
4957 case Intrinsic::convertsu:
4958 case Intrinsic::convertus:
4959 case Intrinsic::convertuu: {
4960 ISD::CvtCode Code = ISD::CVT_INVALID;
4961 switch (Intrinsic) {
4962 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4963 case Intrinsic::convertff: Code = ISD::CVT_FF; break;
4964 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break;
4965 case Intrinsic::convertfui: Code = ISD::CVT_FU; break;
4966 case Intrinsic::convertsif: Code = ISD::CVT_SF; break;
4967 case Intrinsic::convertuif: Code = ISD::CVT_UF; break;
4968 case Intrinsic::convertss: Code = ISD::CVT_SS; break;
4969 case Intrinsic::convertsu: Code = ISD::CVT_SU; break;
4970 case Intrinsic::convertus: Code = ISD::CVT_US; break;
4971 case Intrinsic::convertuu: Code = ISD::CVT_UU; break;
4973 EVT DestVT = TLI->getValueType(I.getType());
4974 const Value *Op1 = I.getArgOperand(0);
4975 Res = DAG.getConvertRndSat(DestVT, sdl, getValue(Op1),
4976 DAG.getValueType(DestVT),
4977 DAG.getValueType(getValue(Op1).getValueType()),
4978 getValue(I.getArgOperand(1)),
4979 getValue(I.getArgOperand(2)),
4984 case Intrinsic::powi:
4985 setValue(&I, ExpandPowI(sdl, getValue(I.getArgOperand(0)),
4986 getValue(I.getArgOperand(1)), DAG));
4988 case Intrinsic::log:
4989 setValue(&I, expandLog(sdl, getValue(I.getArgOperand(0)), DAG, *TLI));
4991 case Intrinsic::log2:
4992 setValue(&I, expandLog2(sdl, getValue(I.getArgOperand(0)), DAG, *TLI));
4994 case Intrinsic::log10:
4995 setValue(&I, expandLog10(sdl, getValue(I.getArgOperand(0)), DAG, *TLI));
4997 case Intrinsic::exp:
4998 setValue(&I, expandExp(sdl, getValue(I.getArgOperand(0)), DAG, *TLI));
5000 case Intrinsic::exp2:
5001 setValue(&I, expandExp2(sdl, getValue(I.getArgOperand(0)), DAG, *TLI));
5003 case Intrinsic::pow:
5004 setValue(&I, expandPow(sdl, getValue(I.getArgOperand(0)),
5005 getValue(I.getArgOperand(1)), DAG, *TLI));
5007 case Intrinsic::sqrt:
5008 case Intrinsic::fabs:
5009 case Intrinsic::sin:
5010 case Intrinsic::cos:
5011 case Intrinsic::floor:
5012 case Intrinsic::ceil:
5013 case Intrinsic::trunc:
5014 case Intrinsic::rint:
5015 case Intrinsic::nearbyint:
5016 case Intrinsic::round: {
5018 switch (Intrinsic) {
5019 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
5020 case Intrinsic::sqrt: Opcode = ISD::FSQRT; break;
5021 case Intrinsic::fabs: Opcode = ISD::FABS; break;
5022 case Intrinsic::sin: Opcode = ISD::FSIN; break;
5023 case Intrinsic::cos: Opcode = ISD::FCOS; break;
5024 case Intrinsic::floor: Opcode = ISD::FFLOOR; break;
5025 case Intrinsic::ceil: Opcode = ISD::FCEIL; break;
5026 case Intrinsic::trunc: Opcode = ISD::FTRUNC; break;
5027 case Intrinsic::rint: Opcode = ISD::FRINT; break;
5028 case Intrinsic::nearbyint: Opcode = ISD::FNEARBYINT; break;
5029 case Intrinsic::round: Opcode = ISD::FROUND; break;
5032 setValue(&I, DAG.getNode(Opcode, sdl,
5033 getValue(I.getArgOperand(0)).getValueType(),
5034 getValue(I.getArgOperand(0))));
5037 case Intrinsic::copysign:
5038 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, sdl,
5039 getValue(I.getArgOperand(0)).getValueType(),
5040 getValue(I.getArgOperand(0)),
5041 getValue(I.getArgOperand(1))));
5043 case Intrinsic::fma:
5044 setValue(&I, DAG.getNode(ISD::FMA, sdl,
5045 getValue(I.getArgOperand(0)).getValueType(),
5046 getValue(I.getArgOperand(0)),
5047 getValue(I.getArgOperand(1)),
5048 getValue(I.getArgOperand(2))));
5050 case Intrinsic::fmuladd: {
5051 EVT VT = TLI->getValueType(I.getType());
5052 if (TM.Options.AllowFPOpFusion != FPOpFusion::Strict &&
5053 TLI->isFMAFasterThanFMulAndFAdd(VT)) {
5054 setValue(&I, DAG.getNode(ISD::FMA, sdl,
5055 getValue(I.getArgOperand(0)).getValueType(),
5056 getValue(I.getArgOperand(0)),
5057 getValue(I.getArgOperand(1)),
5058 getValue(I.getArgOperand(2))));
5060 SDValue Mul = DAG.getNode(ISD::FMUL, sdl,
5061 getValue(I.getArgOperand(0)).getValueType(),
5062 getValue(I.getArgOperand(0)),
5063 getValue(I.getArgOperand(1)));
5064 SDValue Add = DAG.getNode(ISD::FADD, sdl,
5065 getValue(I.getArgOperand(0)).getValueType(),
5067 getValue(I.getArgOperand(2)));
5072 case Intrinsic::convert_to_fp16:
5073 setValue(&I, DAG.getNode(ISD::FP32_TO_FP16, sdl,
5074 MVT::i16, getValue(I.getArgOperand(0))));
5076 case Intrinsic::convert_from_fp16:
5077 setValue(&I, DAG.getNode(ISD::FP16_TO_FP32, sdl,
5078 MVT::f32, getValue(I.getArgOperand(0))));
5080 case Intrinsic::pcmarker: {
5081 SDValue Tmp = getValue(I.getArgOperand(0));
5082 DAG.setRoot(DAG.getNode(ISD::PCMARKER, sdl, MVT::Other, getRoot(), Tmp));
5085 case Intrinsic::readcyclecounter: {
5086 SDValue Op = getRoot();
5087 Res = DAG.getNode(ISD::READCYCLECOUNTER, sdl,
5088 DAG.getVTList(MVT::i64, MVT::Other),
5091 DAG.setRoot(Res.getValue(1));
5094 case Intrinsic::bswap:
5095 setValue(&I, DAG.getNode(ISD::BSWAP, sdl,
5096 getValue(I.getArgOperand(0)).getValueType(),
5097 getValue(I.getArgOperand(0))));
5099 case Intrinsic::cttz: {
5100 SDValue Arg = getValue(I.getArgOperand(0));
5101 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1));
5102 EVT Ty = Arg.getValueType();
5103 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTTZ : ISD::CTTZ_ZERO_UNDEF,
5107 case Intrinsic::ctlz: {
5108 SDValue Arg = getValue(I.getArgOperand(0));
5109 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1));
5110 EVT Ty = Arg.getValueType();
5111 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTLZ : ISD::CTLZ_ZERO_UNDEF,
5115 case Intrinsic::ctpop: {
5116 SDValue Arg = getValue(I.getArgOperand(0));
5117 EVT Ty = Arg.getValueType();
5118 setValue(&I, DAG.getNode(ISD::CTPOP, sdl, Ty, Arg));
5121 case Intrinsic::stacksave: {
5122 SDValue Op = getRoot();
5123 Res = DAG.getNode(ISD::STACKSAVE, sdl,
5124 DAG.getVTList(TLI->getPointerTy(), MVT::Other), &Op, 1);
5126 DAG.setRoot(Res.getValue(1));
5129 case Intrinsic::stackrestore: {
5130 Res = getValue(I.getArgOperand(0));
5131 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, sdl, MVT::Other, getRoot(), Res));
5134 case Intrinsic::stackprotector: {
5135 // Emit code into the DAG to store the stack guard onto the stack.
5136 MachineFunction &MF = DAG.getMachineFunction();
5137 MachineFrameInfo *MFI = MF.getFrameInfo();
5138 EVT PtrTy = TLI->getPointerTy();
5140 SDValue Src = getValue(I.getArgOperand(0)); // The guard's value.
5141 AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
5143 int FI = FuncInfo.StaticAllocaMap[Slot];
5144 MFI->setStackProtectorIndex(FI);
5146 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
5148 // Store the stack protector onto the stack.
5149 Res = DAG.getStore(getRoot(), sdl, Src, FIN,
5150 MachinePointerInfo::getFixedStack(FI),
5156 case Intrinsic::objectsize: {
5157 // If we don't know by now, we're never going to know.
5158 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1));
5160 assert(CI && "Non-constant type in __builtin_object_size?");
5162 SDValue Arg = getValue(I.getCalledValue());
5163 EVT Ty = Arg.getValueType();
5166 Res = DAG.getConstant(-1ULL, Ty);
5168 Res = DAG.getConstant(0, Ty);
5173 case Intrinsic::annotation:
5174 case Intrinsic::ptr_annotation:
5175 // Drop the intrinsic, but forward the value
5176 setValue(&I, getValue(I.getOperand(0)));
5178 case Intrinsic::var_annotation:
5179 // Discard annotate attributes
5182 case Intrinsic::init_trampoline: {
5183 const Function *F = cast<Function>(I.getArgOperand(1)->stripPointerCasts());
5187 Ops[1] = getValue(I.getArgOperand(0));
5188 Ops[2] = getValue(I.getArgOperand(1));
5189 Ops[3] = getValue(I.getArgOperand(2));
5190 Ops[4] = DAG.getSrcValue(I.getArgOperand(0));
5191 Ops[5] = DAG.getSrcValue(F);
5193 Res = DAG.getNode(ISD::INIT_TRAMPOLINE, sdl, MVT::Other, Ops, 6);
5198 case Intrinsic::adjust_trampoline: {
5199 setValue(&I, DAG.getNode(ISD::ADJUST_TRAMPOLINE, sdl,
5200 TLI->getPointerTy(),
5201 getValue(I.getArgOperand(0))));
5204 case Intrinsic::gcroot:
5206 const Value *Alloca = I.getArgOperand(0)->stripPointerCasts();
5207 const Constant *TypeMap = cast<Constant>(I.getArgOperand(1));
5209 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode());
5210 GFI->addStackRoot(FI->getIndex(), TypeMap);
5213 case Intrinsic::gcread:
5214 case Intrinsic::gcwrite:
5215 llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!");
5216 case Intrinsic::flt_rounds:
5217 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, sdl, MVT::i32));
5220 case Intrinsic::expect: {
5221 // Just replace __builtin_expect(exp, c) with EXP.
5222 setValue(&I, getValue(I.getArgOperand(0)));
5226 case Intrinsic::debugtrap:
5227 case Intrinsic::trap: {
5228 StringRef TrapFuncName = TM.Options.getTrapFunctionName();
5229 if (TrapFuncName.empty()) {
5230 ISD::NodeType Op = (Intrinsic == Intrinsic::trap) ?
5231 ISD::TRAP : ISD::DEBUGTRAP;
5232 DAG.setRoot(DAG.getNode(Op, sdl,MVT::Other, getRoot()));
5235 TargetLowering::ArgListTy Args;
5237 CallLoweringInfo CLI(getRoot(), I.getType(),
5238 false, false, false, false, 0, CallingConv::C,
5239 /*isTailCall=*/false,
5240 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
5241 DAG.getExternalSymbol(TrapFuncName.data(),
5242 TLI->getPointerTy()),
5244 std::pair<SDValue, SDValue> Result = TLI->LowerCallTo(CLI);
5245 DAG.setRoot(Result.second);
5249 case Intrinsic::uadd_with_overflow:
5250 case Intrinsic::sadd_with_overflow:
5251 case Intrinsic::usub_with_overflow:
5252 case Intrinsic::ssub_with_overflow:
5253 case Intrinsic::umul_with_overflow:
5254 case Intrinsic::smul_with_overflow: {
5256 switch (Intrinsic) {
5257 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
5258 case Intrinsic::uadd_with_overflow: Op = ISD::UADDO; break;
5259 case Intrinsic::sadd_with_overflow: Op = ISD::SADDO; break;
5260 case Intrinsic::usub_with_overflow: Op = ISD::USUBO; break;
5261 case Intrinsic::ssub_with_overflow: Op = ISD::SSUBO; break;
5262 case Intrinsic::umul_with_overflow: Op = ISD::UMULO; break;
5263 case Intrinsic::smul_with_overflow: Op = ISD::SMULO; break;
5265 SDValue Op1 = getValue(I.getArgOperand(0));
5266 SDValue Op2 = getValue(I.getArgOperand(1));
5268 SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1);
5269 setValue(&I, DAG.getNode(Op, sdl, VTs, Op1, Op2));
5272 case Intrinsic::prefetch: {
5274 unsigned rw = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
5276 Ops[1] = getValue(I.getArgOperand(0));
5277 Ops[2] = getValue(I.getArgOperand(1));
5278 Ops[3] = getValue(I.getArgOperand(2));
5279 Ops[4] = getValue(I.getArgOperand(3));
5280 DAG.setRoot(DAG.getMemIntrinsicNode(ISD::PREFETCH, sdl,
5281 DAG.getVTList(MVT::Other),
5283 EVT::getIntegerVT(*Context, 8),
5284 MachinePointerInfo(I.getArgOperand(0)),
5286 false, /* volatile */
5288 rw==1)); /* write */
5291 case Intrinsic::lifetime_start:
5292 case Intrinsic::lifetime_end: {
5293 bool IsStart = (Intrinsic == Intrinsic::lifetime_start);
5294 // Stack coloring is not enabled in O0, discard region information.
5295 if (TM.getOptLevel() == CodeGenOpt::None)
5298 SmallVector<Value *, 4> Allocas;
5299 GetUnderlyingObjects(I.getArgOperand(1), Allocas, TD);
5301 for (SmallVectorImpl<Value*>::iterator Object = Allocas.begin(),
5302 E = Allocas.end(); Object != E; ++Object) {
5303 AllocaInst *LifetimeObject = dyn_cast_or_null<AllocaInst>(*Object);
5305 // Could not find an Alloca.
5306 if (!LifetimeObject)
5309 int FI = FuncInfo.StaticAllocaMap[LifetimeObject];
5313 Ops[1] = DAG.getFrameIndex(FI, TLI->getPointerTy(), true);
5314 unsigned Opcode = (IsStart ? ISD::LIFETIME_START : ISD::LIFETIME_END);
5316 Res = DAG.getNode(Opcode, sdl, MVT::Other, Ops, 2);
5321 case Intrinsic::invariant_start:
5322 // Discard region information.
5323 setValue(&I, DAG.getUNDEF(TLI->getPointerTy()));
5325 case Intrinsic::invariant_end:
5326 // Discard region information.
5328 case Intrinsic::stackprotectorcheck: {
5329 // Do not actually emit anything for this basic block. Instead we initialize
5330 // the stack protector descriptor and export the guard variable so we can
5331 // access it in FinishBasicBlock.
5332 const BasicBlock *BB = I.getParent();
5333 SPDescriptor.initialize(BB, FuncInfo.MBBMap[BB], I);
5334 ExportFromCurrentBlock(SPDescriptor.getGuard());
5336 // Flush our exports since we are going to process a terminator.
5337 (void)getControlRoot();
5340 case Intrinsic::donothing:
5343 case Intrinsic::experimental_stackmap: {
5347 case Intrinsic::experimental_patchpoint_void:
5348 case Intrinsic::experimental_patchpoint_i64: {
5355 void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee,
5357 MachineBasicBlock *LandingPad) {
5358 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
5359 FunctionType *FTy = cast<FunctionType>(PT->getElementType());
5360 Type *RetTy = FTy->getReturnType();
5361 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5362 MCSymbol *BeginLabel = 0;
5364 TargetLowering::ArgListTy Args;
5365 TargetLowering::ArgListEntry Entry;
5366 Args.reserve(CS.arg_size());
5368 // Check whether the function can return without sret-demotion.
5369 SmallVector<ISD::OutputArg, 4> Outs;
5370 const TargetLowering *TLI = TM.getTargetLowering();
5371 GetReturnInfo(RetTy, CS.getAttributes(), Outs, *TLI);
5373 bool CanLowerReturn = TLI->CanLowerReturn(CS.getCallingConv(),
5374 DAG.getMachineFunction(),
5375 FTy->isVarArg(), Outs,
5378 SDValue DemoteStackSlot;
5379 int DemoteStackIdx = -100;
5381 if (!CanLowerReturn) {
5382 uint64_t TySize = TLI->getDataLayout()->getTypeAllocSize(
5383 FTy->getReturnType());
5384 unsigned Align = TLI->getDataLayout()->getPrefTypeAlignment(
5385 FTy->getReturnType());
5386 MachineFunction &MF = DAG.getMachineFunction();
5387 DemoteStackIdx = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
5388 Type *StackSlotPtrType = PointerType::getUnqual(FTy->getReturnType());
5390 DemoteStackSlot = DAG.getFrameIndex(DemoteStackIdx, TLI->getPointerTy());
5391 Entry.Node = DemoteStackSlot;
5392 Entry.Ty = StackSlotPtrType;
5393 Entry.isSExt = false;
5394 Entry.isZExt = false;
5395 Entry.isInReg = false;
5396 Entry.isSRet = true;
5397 Entry.isNest = false;
5398 Entry.isByVal = false;
5399 Entry.isReturned = false;
5400 Entry.Alignment = Align;
5401 Args.push_back(Entry);
5402 RetTy = Type::getVoidTy(FTy->getContext());
5405 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
5407 const Value *V = *i;
5410 if (V->getType()->isEmptyTy())
5413 SDValue ArgNode = getValue(V);
5414 Entry.Node = ArgNode; Entry.Ty = V->getType();
5416 // Skip the first return-type Attribute to get to params.
5417 Entry.setAttributes(&CS, i - CS.arg_begin() + 1);
5418 Args.push_back(Entry);
5422 // Insert a label before the invoke call to mark the try range. This can be
5423 // used to detect deletion of the invoke via the MachineModuleInfo.
5424 BeginLabel = MMI.getContext().CreateTempSymbol();
5426 // For SjLj, keep track of which landing pads go with which invokes
5427 // so as to maintain the ordering of pads in the LSDA.
5428 unsigned CallSiteIndex = MMI.getCurrentCallSite();
5429 if (CallSiteIndex) {
5430 MMI.setCallSiteBeginLabel(BeginLabel, CallSiteIndex);
5431 LPadToCallSiteMap[LandingPad].push_back(CallSiteIndex);
5433 // Now that the call site is handled, stop tracking it.
5434 MMI.setCurrentCallSite(0);
5437 // Both PendingLoads and PendingExports must be flushed here;
5438 // this call might not return.
5440 DAG.setRoot(DAG.getEHLabel(getCurSDLoc(), getControlRoot(), BeginLabel));
5443 // Check if target-independent constraints permit a tail call here.
5444 // Target-dependent constraints are checked within TLI->LowerCallTo.
5445 if (isTailCall && !isInTailCallPosition(CS, *TLI))
5449 CallLoweringInfo CLI(getRoot(), RetTy, FTy, isTailCall, Callee, Args, DAG,
5451 std::pair<SDValue,SDValue> Result = TLI->LowerCallTo(CLI);
5452 assert((isTailCall || Result.second.getNode()) &&
5453 "Non-null chain expected with non-tail call!");
5454 assert((Result.second.getNode() || !Result.first.getNode()) &&
5455 "Null value expected with tail call!");
5456 if (Result.first.getNode()) {
5457 setValue(CS.getInstruction(), Result.first);
5458 } else if (!CanLowerReturn && Result.second.getNode()) {
5459 // The instruction result is the result of loading from the
5460 // hidden sret parameter.
5461 SmallVector<EVT, 1> PVTs;
5462 Type *PtrRetTy = PointerType::getUnqual(FTy->getReturnType());
5464 ComputeValueVTs(*TLI, PtrRetTy, PVTs);
5465 assert(PVTs.size() == 1 && "Pointers should fit in one register");
5466 EVT PtrVT = PVTs[0];
5468 SmallVector<EVT, 4> RetTys;
5469 SmallVector<uint64_t, 4> Offsets;
5470 RetTy = FTy->getReturnType();
5471 ComputeValueVTs(*TLI, RetTy, RetTys, &Offsets);
5473 unsigned NumValues = RetTys.size();
5474 SmallVector<SDValue, 4> Values(NumValues);
5475 SmallVector<SDValue, 4> Chains(NumValues);
5477 for (unsigned i = 0; i < NumValues; ++i) {
5478 SDValue Add = DAG.getNode(ISD::ADD, getCurSDLoc(), PtrVT,
5480 DAG.getConstant(Offsets[i], PtrVT));
5481 SDValue L = DAG.getLoad(RetTys[i], getCurSDLoc(), Result.second, Add,
5482 MachinePointerInfo::getFixedStack(DemoteStackIdx, Offsets[i]),
5483 false, false, false, 1);
5485 Chains[i] = L.getValue(1);
5488 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(),
5489 MVT::Other, &Chains[0], NumValues);
5490 PendingLoads.push_back(Chain);
5492 setValue(CS.getInstruction(),
5493 DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
5494 DAG.getVTList(&RetTys[0], RetTys.size()),
5495 &Values[0], Values.size()));
5498 if (!Result.second.getNode()) {
5499 // As a special case, a null chain means that a tail call has been emitted
5500 // and the DAG root is already updated.
5503 // Since there's no actual continuation from this block, nothing can be
5504 // relying on us setting vregs for them.
5505 PendingExports.clear();
5507 DAG.setRoot(Result.second);
5511 // Insert a label at the end of the invoke call to mark the try range. This
5512 // can be used to detect deletion of the invoke via the MachineModuleInfo.
5513 MCSymbol *EndLabel = MMI.getContext().CreateTempSymbol();
5514 DAG.setRoot(DAG.getEHLabel(getCurSDLoc(), getRoot(), EndLabel));
5516 // Inform MachineModuleInfo of range.
5517 MMI.addInvoke(LandingPad, BeginLabel, EndLabel);
5521 /// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the
5522 /// value is equal or not-equal to zero.
5523 static bool IsOnlyUsedInZeroEqualityComparison(const Value *V) {
5524 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end();
5526 if (const ICmpInst *IC = dyn_cast<ICmpInst>(*UI))
5527 if (IC->isEquality())
5528 if (const Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
5529 if (C->isNullValue())
5531 // Unknown instruction.
5537 static SDValue getMemCmpLoad(const Value *PtrVal, MVT LoadVT,
5539 SelectionDAGBuilder &Builder) {
5541 // Check to see if this load can be trivially constant folded, e.g. if the
5542 // input is from a string literal.
5543 if (const Constant *LoadInput = dyn_cast<Constant>(PtrVal)) {
5544 // Cast pointer to the type we really want to load.
5545 LoadInput = ConstantExpr::getBitCast(const_cast<Constant *>(LoadInput),
5546 PointerType::getUnqual(LoadTy));
5548 if (const Constant *LoadCst =
5549 ConstantFoldLoadFromConstPtr(const_cast<Constant *>(LoadInput),
5551 return Builder.getValue(LoadCst);
5554 // Otherwise, we have to emit the load. If the pointer is to unfoldable but
5555 // still constant memory, the input chain can be the entry node.
5557 bool ConstantMemory = false;
5559 // Do not serialize (non-volatile) loads of constant memory with anything.
5560 if (Builder.AA->pointsToConstantMemory(PtrVal)) {
5561 Root = Builder.DAG.getEntryNode();
5562 ConstantMemory = true;
5564 // Do not serialize non-volatile loads against each other.
5565 Root = Builder.DAG.getRoot();
5568 SDValue Ptr = Builder.getValue(PtrVal);
5569 SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurSDLoc(), Root,
5570 Ptr, MachinePointerInfo(PtrVal),
5572 false /*nontemporal*/,
5573 false /*isinvariant*/, 1 /* align=1 */);
5575 if (!ConstantMemory)
5576 Builder.PendingLoads.push_back(LoadVal.getValue(1));
5580 /// processIntegerCallValue - Record the value for an instruction that
5581 /// produces an integer result, converting the type where necessary.
5582 void SelectionDAGBuilder::processIntegerCallValue(const Instruction &I,
5585 EVT VT = TM.getTargetLowering()->getValueType(I.getType(), true);
5587 Value = DAG.getSExtOrTrunc(Value, getCurSDLoc(), VT);
5589 Value = DAG.getZExtOrTrunc(Value, getCurSDLoc(), VT);
5590 setValue(&I, Value);
5593 /// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form.
5594 /// If so, return true and lower it, otherwise return false and it will be
5595 /// lowered like a normal call.
5596 bool SelectionDAGBuilder::visitMemCmpCall(const CallInst &I) {
5597 // Verify that the prototype makes sense. int memcmp(void*,void*,size_t)
5598 if (I.getNumArgOperands() != 3)
5601 const Value *LHS = I.getArgOperand(0), *RHS = I.getArgOperand(1);
5602 if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() ||
5603 !I.getArgOperand(2)->getType()->isIntegerTy() ||
5604 !I.getType()->isIntegerTy())
5607 const Value *Size = I.getArgOperand(2);
5608 const ConstantInt *CSize = dyn_cast<ConstantInt>(Size);
5609 if (CSize && CSize->getZExtValue() == 0) {
5610 EVT CallVT = TM.getTargetLowering()->getValueType(I.getType(), true);
5611 setValue(&I, DAG.getConstant(0, CallVT));
5615 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5616 std::pair<SDValue, SDValue> Res =
5617 TSI.EmitTargetCodeForMemcmp(DAG, getCurSDLoc(), DAG.getRoot(),
5618 getValue(LHS), getValue(RHS), getValue(Size),
5619 MachinePointerInfo(LHS),
5620 MachinePointerInfo(RHS));
5621 if (Res.first.getNode()) {
5622 processIntegerCallValue(I, Res.first, true);
5623 PendingLoads.push_back(Res.second);
5627 // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0
5628 // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0
5629 if (CSize && IsOnlyUsedInZeroEqualityComparison(&I)) {
5630 bool ActuallyDoIt = true;
5633 switch (CSize->getZExtValue()) {
5635 LoadVT = MVT::Other;
5637 ActuallyDoIt = false;
5641 LoadTy = Type::getInt16Ty(CSize->getContext());
5645 LoadTy = Type::getInt32Ty(CSize->getContext());
5649 LoadTy = Type::getInt64Ty(CSize->getContext());
5653 LoadVT = MVT::v4i32;
5654 LoadTy = Type::getInt32Ty(CSize->getContext());
5655 LoadTy = VectorType::get(LoadTy, 4);
5660 // This turns into unaligned loads. We only do this if the target natively
5661 // supports the MVT we'll be loading or if it is small enough (<= 4) that
5662 // we'll only produce a small number of byte loads.
5664 // Require that we can find a legal MVT, and only do this if the target
5665 // supports unaligned loads of that type. Expanding into byte loads would
5667 const TargetLowering *TLI = TM.getTargetLowering();
5668 if (ActuallyDoIt && CSize->getZExtValue() > 4) {
5669 // TODO: Handle 5 byte compare as 4-byte + 1 byte.
5670 // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads.
5671 if (!TLI->isTypeLegal(LoadVT) ||!TLI->allowsUnalignedMemoryAccesses(LoadVT))
5672 ActuallyDoIt = false;
5676 SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this);
5677 SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this);
5679 SDValue Res = DAG.getSetCC(getCurSDLoc(), MVT::i1, LHSVal, RHSVal,
5681 processIntegerCallValue(I, Res, false);
5690 /// visitMemChrCall -- See if we can lower a memchr call into an optimized
5691 /// form. If so, return true and lower it, otherwise return false and it
5692 /// will be lowered like a normal call.
5693 bool SelectionDAGBuilder::visitMemChrCall(const CallInst &I) {
5694 // Verify that the prototype makes sense. void *memchr(void *, int, size_t)
5695 if (I.getNumArgOperands() != 3)
5698 const Value *Src = I.getArgOperand(0);
5699 const Value *Char = I.getArgOperand(1);
5700 const Value *Length = I.getArgOperand(2);
5701 if (!Src->getType()->isPointerTy() ||
5702 !Char->getType()->isIntegerTy() ||
5703 !Length->getType()->isIntegerTy() ||
5704 !I.getType()->isPointerTy())
5707 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5708 std::pair<SDValue, SDValue> Res =
5709 TSI.EmitTargetCodeForMemchr(DAG, getCurSDLoc(), DAG.getRoot(),
5710 getValue(Src), getValue(Char), getValue(Length),
5711 MachinePointerInfo(Src));
5712 if (Res.first.getNode()) {
5713 setValue(&I, Res.first);
5714 PendingLoads.push_back(Res.second);
5721 /// visitStrCpyCall -- See if we can lower a strcpy or stpcpy call into an
5722 /// optimized form. If so, return true and lower it, otherwise return false
5723 /// and it will be lowered like a normal call.
5724 bool SelectionDAGBuilder::visitStrCpyCall(const CallInst &I, bool isStpcpy) {
5725 // Verify that the prototype makes sense. char *strcpy(char *, char *)
5726 if (I.getNumArgOperands() != 2)
5729 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5730 if (!Arg0->getType()->isPointerTy() ||
5731 !Arg1->getType()->isPointerTy() ||
5732 !I.getType()->isPointerTy())
5735 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5736 std::pair<SDValue, SDValue> Res =
5737 TSI.EmitTargetCodeForStrcpy(DAG, getCurSDLoc(), getRoot(),
5738 getValue(Arg0), getValue(Arg1),
5739 MachinePointerInfo(Arg0),
5740 MachinePointerInfo(Arg1), isStpcpy);
5741 if (Res.first.getNode()) {
5742 setValue(&I, Res.first);
5743 DAG.setRoot(Res.second);
5750 /// visitStrCmpCall - See if we can lower a call to strcmp in an optimized form.
5751 /// If so, return true and lower it, otherwise return false and it will be
5752 /// lowered like a normal call.
5753 bool SelectionDAGBuilder::visitStrCmpCall(const CallInst &I) {
5754 // Verify that the prototype makes sense. int strcmp(void*,void*)
5755 if (I.getNumArgOperands() != 2)
5758 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5759 if (!Arg0->getType()->isPointerTy() ||
5760 !Arg1->getType()->isPointerTy() ||
5761 !I.getType()->isIntegerTy())
5764 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5765 std::pair<SDValue, SDValue> Res =
5766 TSI.EmitTargetCodeForStrcmp(DAG, getCurSDLoc(), DAG.getRoot(),
5767 getValue(Arg0), getValue(Arg1),
5768 MachinePointerInfo(Arg0),
5769 MachinePointerInfo(Arg1));
5770 if (Res.first.getNode()) {
5771 processIntegerCallValue(I, Res.first, true);
5772 PendingLoads.push_back(Res.second);
5779 /// visitStrLenCall -- See if we can lower a strlen call into an optimized
5780 /// form. If so, return true and lower it, otherwise return false and it
5781 /// will be lowered like a normal call.
5782 bool SelectionDAGBuilder::visitStrLenCall(const CallInst &I) {
5783 // Verify that the prototype makes sense. size_t strlen(char *)
5784 if (I.getNumArgOperands() != 1)
5787 const Value *Arg0 = I.getArgOperand(0);
5788 if (!Arg0->getType()->isPointerTy() || !I.getType()->isIntegerTy())
5791 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5792 std::pair<SDValue, SDValue> Res =
5793 TSI.EmitTargetCodeForStrlen(DAG, getCurSDLoc(), DAG.getRoot(),
5794 getValue(Arg0), MachinePointerInfo(Arg0));
5795 if (Res.first.getNode()) {
5796 processIntegerCallValue(I, Res.first, false);
5797 PendingLoads.push_back(Res.second);
5804 /// visitStrNLenCall -- See if we can lower a strnlen call into an optimized
5805 /// form. If so, return true and lower it, otherwise return false and it
5806 /// will be lowered like a normal call.
5807 bool SelectionDAGBuilder::visitStrNLenCall(const CallInst &I) {
5808 // Verify that the prototype makes sense. size_t strnlen(char *, size_t)
5809 if (I.getNumArgOperands() != 2)
5812 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5813 if (!Arg0->getType()->isPointerTy() ||
5814 !Arg1->getType()->isIntegerTy() ||
5815 !I.getType()->isIntegerTy())
5818 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5819 std::pair<SDValue, SDValue> Res =
5820 TSI.EmitTargetCodeForStrnlen(DAG, getCurSDLoc(), DAG.getRoot(),
5821 getValue(Arg0), getValue(Arg1),
5822 MachinePointerInfo(Arg0));
5823 if (Res.first.getNode()) {
5824 processIntegerCallValue(I, Res.first, false);
5825 PendingLoads.push_back(Res.second);
5832 /// visitUnaryFloatCall - If a call instruction is a unary floating-point
5833 /// operation (as expected), translate it to an SDNode with the specified opcode
5834 /// and return true.
5835 bool SelectionDAGBuilder::visitUnaryFloatCall(const CallInst &I,
5837 // Sanity check that it really is a unary floating-point call.
5838 if (I.getNumArgOperands() != 1 ||
5839 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
5840 I.getType() != I.getArgOperand(0)->getType() ||
5841 !I.onlyReadsMemory())
5844 SDValue Tmp = getValue(I.getArgOperand(0));
5845 setValue(&I, DAG.getNode(Opcode, getCurSDLoc(), Tmp.getValueType(), Tmp));
5849 void SelectionDAGBuilder::visitCall(const CallInst &I) {
5850 // Handle inline assembly differently.
5851 if (isa<InlineAsm>(I.getCalledValue())) {
5856 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5857 ComputeUsesVAFloatArgument(I, &MMI);
5859 const char *RenameFn = 0;
5860 if (Function *F = I.getCalledFunction()) {
5861 if (F->isDeclaration()) {
5862 if (const TargetIntrinsicInfo *II = TM.getIntrinsicInfo()) {
5863 if (unsigned IID = II->getIntrinsicID(F)) {
5864 RenameFn = visitIntrinsicCall(I, IID);
5869 if (unsigned IID = F->getIntrinsicID()) {
5870 RenameFn = visitIntrinsicCall(I, IID);
5876 // Check for well-known libc/libm calls. If the function is internal, it
5877 // can't be a library call.
5879 if (!F->hasLocalLinkage() && F->hasName() &&
5880 LibInfo->getLibFunc(F->getName(), Func) &&
5881 LibInfo->hasOptimizedCodeGen(Func)) {
5884 case LibFunc::copysign:
5885 case LibFunc::copysignf:
5886 case LibFunc::copysignl:
5887 if (I.getNumArgOperands() == 2 && // Basic sanity checks.
5888 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5889 I.getType() == I.getArgOperand(0)->getType() &&
5890 I.getType() == I.getArgOperand(1)->getType() &&
5891 I.onlyReadsMemory()) {
5892 SDValue LHS = getValue(I.getArgOperand(0));
5893 SDValue RHS = getValue(I.getArgOperand(1));
5894 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurSDLoc(),
5895 LHS.getValueType(), LHS, RHS));
5900 case LibFunc::fabsf:
5901 case LibFunc::fabsl:
5902 if (visitUnaryFloatCall(I, ISD::FABS))
5908 if (visitUnaryFloatCall(I, ISD::FSIN))
5914 if (visitUnaryFloatCall(I, ISD::FCOS))
5918 case LibFunc::sqrtf:
5919 case LibFunc::sqrtl:
5920 case LibFunc::sqrt_finite:
5921 case LibFunc::sqrtf_finite:
5922 case LibFunc::sqrtl_finite:
5923 if (visitUnaryFloatCall(I, ISD::FSQRT))
5926 case LibFunc::floor:
5927 case LibFunc::floorf:
5928 case LibFunc::floorl:
5929 if (visitUnaryFloatCall(I, ISD::FFLOOR))
5932 case LibFunc::nearbyint:
5933 case LibFunc::nearbyintf:
5934 case LibFunc::nearbyintl:
5935 if (visitUnaryFloatCall(I, ISD::FNEARBYINT))
5939 case LibFunc::ceilf:
5940 case LibFunc::ceill:
5941 if (visitUnaryFloatCall(I, ISD::FCEIL))
5945 case LibFunc::rintf:
5946 case LibFunc::rintl:
5947 if (visitUnaryFloatCall(I, ISD::FRINT))
5950 case LibFunc::round:
5951 case LibFunc::roundf:
5952 case LibFunc::roundl:
5953 if (visitUnaryFloatCall(I, ISD::FROUND))
5956 case LibFunc::trunc:
5957 case LibFunc::truncf:
5958 case LibFunc::truncl:
5959 if (visitUnaryFloatCall(I, ISD::FTRUNC))
5963 case LibFunc::log2f:
5964 case LibFunc::log2l:
5965 if (visitUnaryFloatCall(I, ISD::FLOG2))
5969 case LibFunc::exp2f:
5970 case LibFunc::exp2l:
5971 if (visitUnaryFloatCall(I, ISD::FEXP2))
5974 case LibFunc::memcmp:
5975 if (visitMemCmpCall(I))
5978 case LibFunc::memchr:
5979 if (visitMemChrCall(I))
5982 case LibFunc::strcpy:
5983 if (visitStrCpyCall(I, false))
5986 case LibFunc::stpcpy:
5987 if (visitStrCpyCall(I, true))
5990 case LibFunc::strcmp:
5991 if (visitStrCmpCall(I))
5994 case LibFunc::strlen:
5995 if (visitStrLenCall(I))
5998 case LibFunc::strnlen:
5999 if (visitStrNLenCall(I))
6008 Callee = getValue(I.getCalledValue());
6010 Callee = DAG.getExternalSymbol(RenameFn,
6011 TM.getTargetLowering()->getPointerTy());
6013 // Check if we can potentially perform a tail call. More detailed checking is
6014 // be done within LowerCallTo, after more information about the call is known.
6015 LowerCallTo(&I, Callee, I.isTailCall());
6020 /// AsmOperandInfo - This contains information for each constraint that we are
6022 class SDISelAsmOperandInfo : public TargetLowering::AsmOperandInfo {
6024 /// CallOperand - If this is the result output operand or a clobber
6025 /// this is null, otherwise it is the incoming operand to the CallInst.
6026 /// This gets modified as the asm is processed.
6027 SDValue CallOperand;
6029 /// AssignedRegs - If this is a register or register class operand, this
6030 /// contains the set of register corresponding to the operand.
6031 RegsForValue AssignedRegs;
6033 explicit SDISelAsmOperandInfo(const TargetLowering::AsmOperandInfo &info)
6034 : TargetLowering::AsmOperandInfo(info), CallOperand(0,0) {
6037 /// getCallOperandValEVT - Return the EVT of the Value* that this operand
6038 /// corresponds to. If there is no Value* for this operand, it returns
6040 EVT getCallOperandValEVT(LLVMContext &Context,
6041 const TargetLowering &TLI,
6042 const DataLayout *TD) const {
6043 if (CallOperandVal == 0) return MVT::Other;
6045 if (isa<BasicBlock>(CallOperandVal))
6046 return TLI.getPointerTy();
6048 llvm::Type *OpTy = CallOperandVal->getType();
6050 // FIXME: code duplicated from TargetLowering::ParseConstraints().
6051 // If this is an indirect operand, the operand is a pointer to the
6054 llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
6056 report_fatal_error("Indirect operand for inline asm not a pointer!");
6057 OpTy = PtrTy->getElementType();
6060 // Look for vector wrapped in a struct. e.g. { <16 x i8> }.
6061 if (StructType *STy = dyn_cast<StructType>(OpTy))
6062 if (STy->getNumElements() == 1)
6063 OpTy = STy->getElementType(0);
6065 // If OpTy is not a single value, it may be a struct/union that we
6066 // can tile with integers.
6067 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
6068 unsigned BitSize = TD->getTypeSizeInBits(OpTy);
6077 OpTy = IntegerType::get(Context, BitSize);
6082 return TLI.getValueType(OpTy, true);
6086 typedef SmallVector<SDISelAsmOperandInfo,16> SDISelAsmOperandInfoVector;
6088 } // end anonymous namespace
6090 /// GetRegistersForValue - Assign registers (virtual or physical) for the
6091 /// specified operand. We prefer to assign virtual registers, to allow the
6092 /// register allocator to handle the assignment process. However, if the asm
6093 /// uses features that we can't model on machineinstrs, we have SDISel do the
6094 /// allocation. This produces generally horrible, but correct, code.
6096 /// OpInfo describes the operand.
6098 static void GetRegistersForValue(SelectionDAG &DAG,
6099 const TargetLowering &TLI,
6101 SDISelAsmOperandInfo &OpInfo) {
6102 LLVMContext &Context = *DAG.getContext();
6104 MachineFunction &MF = DAG.getMachineFunction();
6105 SmallVector<unsigned, 4> Regs;
6107 // If this is a constraint for a single physreg, or a constraint for a
6108 // register class, find it.
6109 std::pair<unsigned, const TargetRegisterClass*> PhysReg =
6110 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
6111 OpInfo.ConstraintVT);
6113 unsigned NumRegs = 1;
6114 if (OpInfo.ConstraintVT != MVT::Other) {
6115 // If this is a FP input in an integer register (or visa versa) insert a bit
6116 // cast of the input value. More generally, handle any case where the input
6117 // value disagrees with the register class we plan to stick this in.
6118 if (OpInfo.Type == InlineAsm::isInput &&
6119 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) {
6120 // Try to convert to the first EVT that the reg class contains. If the
6121 // types are identical size, use a bitcast to convert (e.g. two differing
6123 MVT RegVT = *PhysReg.second->vt_begin();
6124 if (RegVT.getSizeInBits() == OpInfo.ConstraintVT.getSizeInBits()) {
6125 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
6126 RegVT, OpInfo.CallOperand);
6127 OpInfo.ConstraintVT = RegVT;
6128 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) {
6129 // If the input is a FP value and we want it in FP registers, do a
6130 // bitcast to the corresponding integer type. This turns an f64 value
6131 // into i64, which can be passed with two i32 values on a 32-bit
6133 RegVT = MVT::getIntegerVT(OpInfo.ConstraintVT.getSizeInBits());
6134 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
6135 RegVT, OpInfo.CallOperand);
6136 OpInfo.ConstraintVT = RegVT;
6140 NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT);
6144 EVT ValueVT = OpInfo.ConstraintVT;
6146 // If this is a constraint for a specific physical register, like {r17},
6148 if (unsigned AssignedReg = PhysReg.first) {
6149 const TargetRegisterClass *RC = PhysReg.second;
6150 if (OpInfo.ConstraintVT == MVT::Other)
6151 ValueVT = *RC->vt_begin();
6153 // Get the actual register value type. This is important, because the user
6154 // may have asked for (e.g.) the AX register in i32 type. We need to
6155 // remember that AX is actually i16 to get the right extension.
6156 RegVT = *RC->vt_begin();
6158 // This is a explicit reference to a physical register.
6159 Regs.push_back(AssignedReg);
6161 // If this is an expanded reference, add the rest of the regs to Regs.
6163 TargetRegisterClass::iterator I = RC->begin();
6164 for (; *I != AssignedReg; ++I)
6165 assert(I != RC->end() && "Didn't find reg!");
6167 // Already added the first reg.
6169 for (; NumRegs; --NumRegs, ++I) {
6170 assert(I != RC->end() && "Ran out of registers to allocate!");
6175 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
6179 // Otherwise, if this was a reference to an LLVM register class, create vregs
6180 // for this reference.
6181 if (const TargetRegisterClass *RC = PhysReg.second) {
6182 RegVT = *RC->vt_begin();
6183 if (OpInfo.ConstraintVT == MVT::Other)
6186 // Create the appropriate number of virtual registers.
6187 MachineRegisterInfo &RegInfo = MF.getRegInfo();
6188 for (; NumRegs; --NumRegs)
6189 Regs.push_back(RegInfo.createVirtualRegister(RC));
6191 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
6195 // Otherwise, we couldn't allocate enough registers for this.
6198 /// visitInlineAsm - Handle a call to an InlineAsm object.
6200 void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) {
6201 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
6203 /// ConstraintOperands - Information about all of the constraints.
6204 SDISelAsmOperandInfoVector ConstraintOperands;
6206 const TargetLowering *TLI = TM.getTargetLowering();
6207 TargetLowering::AsmOperandInfoVector
6208 TargetConstraints = TLI->ParseConstraints(CS);
6210 bool hasMemory = false;
6212 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
6213 unsigned ResNo = 0; // ResNo - The result number of the next output.
6214 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
6215 ConstraintOperands.push_back(SDISelAsmOperandInfo(TargetConstraints[i]));
6216 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
6218 MVT OpVT = MVT::Other;
6220 // Compute the value type for each operand.
6221 switch (OpInfo.Type) {
6222 case InlineAsm::isOutput:
6223 // Indirect outputs just consume an argument.
6224 if (OpInfo.isIndirect) {
6225 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
6229 // The return value of the call is this value. As such, there is no
6230 // corresponding argument.
6231 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
6232 if (StructType *STy = dyn_cast<StructType>(CS.getType())) {
6233 OpVT = TLI->getSimpleValueType(STy->getElementType(ResNo));
6235 assert(ResNo == 0 && "Asm only has one result!");
6236 OpVT = TLI->getSimpleValueType(CS.getType());
6240 case InlineAsm::isInput:
6241 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
6243 case InlineAsm::isClobber:
6248 // If this is an input or an indirect output, process the call argument.
6249 // BasicBlocks are labels, currently appearing only in asm's.
6250 if (OpInfo.CallOperandVal) {
6251 if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) {
6252 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]);
6254 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
6257 OpVT = OpInfo.getCallOperandValEVT(*DAG.getContext(), *TLI, TD).
6261 OpInfo.ConstraintVT = OpVT;
6263 // Indirect operand accesses access memory.
6264 if (OpInfo.isIndirect)
6267 for (unsigned j = 0, ee = OpInfo.Codes.size(); j != ee; ++j) {
6268 TargetLowering::ConstraintType
6269 CType = TLI->getConstraintType(OpInfo.Codes[j]);
6270 if (CType == TargetLowering::C_Memory) {
6278 SDValue Chain, Flag;
6280 // We won't need to flush pending loads if this asm doesn't touch
6281 // memory and is nonvolatile.
6282 if (hasMemory || IA->hasSideEffects())
6285 Chain = DAG.getRoot();
6287 // Second pass over the constraints: compute which constraint option to use
6288 // and assign registers to constraints that want a specific physreg.
6289 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6290 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6292 // If this is an output operand with a matching input operand, look up the
6293 // matching input. If their types mismatch, e.g. one is an integer, the
6294 // other is floating point, or their sizes are different, flag it as an
6296 if (OpInfo.hasMatchingInput()) {
6297 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
6299 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
6300 std::pair<unsigned, const TargetRegisterClass*> MatchRC =
6301 TLI->getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
6302 OpInfo.ConstraintVT);
6303 std::pair<unsigned, const TargetRegisterClass*> InputRC =
6304 TLI->getRegForInlineAsmConstraint(Input.ConstraintCode,
6305 Input.ConstraintVT);
6306 if ((OpInfo.ConstraintVT.isInteger() !=
6307 Input.ConstraintVT.isInteger()) ||
6308 (MatchRC.second != InputRC.second)) {
6309 report_fatal_error("Unsupported asm: input constraint"
6310 " with a matching output constraint of"
6311 " incompatible type!");
6313 Input.ConstraintVT = OpInfo.ConstraintVT;
6317 // Compute the constraint code and ConstraintType to use.
6318 TLI->ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, &DAG);
6320 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
6321 OpInfo.Type == InlineAsm::isClobber)
6324 // If this is a memory input, and if the operand is not indirect, do what we
6325 // need to to provide an address for the memory input.
6326 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
6327 !OpInfo.isIndirect) {
6328 assert((OpInfo.isMultipleAlternative ||
6329 (OpInfo.Type == InlineAsm::isInput)) &&
6330 "Can only indirectify direct input operands!");
6332 // Memory operands really want the address of the value. If we don't have
6333 // an indirect input, put it in the constpool if we can, otherwise spill
6334 // it to a stack slot.
6335 // TODO: This isn't quite right. We need to handle these according to
6336 // the addressing mode that the constraint wants. Also, this may take
6337 // an additional register for the computation and we don't want that
6340 // If the operand is a float, integer, or vector constant, spill to a
6341 // constant pool entry to get its address.
6342 const Value *OpVal = OpInfo.CallOperandVal;
6343 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
6344 isa<ConstantVector>(OpVal) || isa<ConstantDataVector>(OpVal)) {
6345 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal),
6346 TLI->getPointerTy());
6348 // Otherwise, create a stack slot and emit a store to it before the
6350 Type *Ty = OpVal->getType();
6351 uint64_t TySize = TLI->getDataLayout()->getTypeAllocSize(Ty);
6352 unsigned Align = TLI->getDataLayout()->getPrefTypeAlignment(Ty);
6353 MachineFunction &MF = DAG.getMachineFunction();
6354 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
6355 SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI->getPointerTy());
6356 Chain = DAG.getStore(Chain, getCurSDLoc(),
6357 OpInfo.CallOperand, StackSlot,
6358 MachinePointerInfo::getFixedStack(SSFI),
6360 OpInfo.CallOperand = StackSlot;
6363 // There is no longer a Value* corresponding to this operand.
6364 OpInfo.CallOperandVal = 0;
6366 // It is now an indirect operand.
6367 OpInfo.isIndirect = true;
6370 // If this constraint is for a specific register, allocate it before
6372 if (OpInfo.ConstraintType == TargetLowering::C_Register)
6373 GetRegistersForValue(DAG, *TLI, getCurSDLoc(), OpInfo);
6376 // Second pass - Loop over all of the operands, assigning virtual or physregs
6377 // to register class operands.
6378 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6379 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6381 // C_Register operands have already been allocated, Other/Memory don't need
6383 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
6384 GetRegistersForValue(DAG, *TLI, getCurSDLoc(), OpInfo);
6387 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
6388 std::vector<SDValue> AsmNodeOperands;
6389 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain
6390 AsmNodeOperands.push_back(
6391 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(),
6392 TLI->getPointerTy()));
6394 // If we have a !srcloc metadata node associated with it, we want to attach
6395 // this to the ultimately generated inline asm machineinstr. To do this, we
6396 // pass in the third operand as this (potentially null) inline asm MDNode.
6397 const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc");
6398 AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc));
6400 // Remember the HasSideEffect, AlignStack, AsmDialect, MayLoad and MayStore
6401 // bits as operand 3.
6402 unsigned ExtraInfo = 0;
6403 if (IA->hasSideEffects())
6404 ExtraInfo |= InlineAsm::Extra_HasSideEffects;
6405 if (IA->isAlignStack())
6406 ExtraInfo |= InlineAsm::Extra_IsAlignStack;
6407 // Set the asm dialect.
6408 ExtraInfo |= IA->getDialect() * InlineAsm::Extra_AsmDialect;
6410 // Determine if this InlineAsm MayLoad or MayStore based on the constraints.
6411 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
6412 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
6414 // Compute the constraint code and ConstraintType to use.
6415 TLI->ComputeConstraintToUse(OpInfo, SDValue());
6417 // Ideally, we would only check against memory constraints. However, the
6418 // meaning of an other constraint can be target-specific and we can't easily
6419 // reason about it. Therefore, be conservative and set MayLoad/MayStore
6420 // for other constriants as well.
6421 if (OpInfo.ConstraintType == TargetLowering::C_Memory ||
6422 OpInfo.ConstraintType == TargetLowering::C_Other) {
6423 if (OpInfo.Type == InlineAsm::isInput)
6424 ExtraInfo |= InlineAsm::Extra_MayLoad;
6425 else if (OpInfo.Type == InlineAsm::isOutput)
6426 ExtraInfo |= InlineAsm::Extra_MayStore;
6427 else if (OpInfo.Type == InlineAsm::isClobber)
6428 ExtraInfo |= (InlineAsm::Extra_MayLoad | InlineAsm::Extra_MayStore);
6432 AsmNodeOperands.push_back(DAG.getTargetConstant(ExtraInfo,
6433 TLI->getPointerTy()));
6435 // Loop over all of the inputs, copying the operand values into the
6436 // appropriate registers and processing the output regs.
6437 RegsForValue RetValRegs;
6439 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
6440 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
6442 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6443 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6445 switch (OpInfo.Type) {
6446 case InlineAsm::isOutput: {
6447 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
6448 OpInfo.ConstraintType != TargetLowering::C_Register) {
6449 // Memory output, or 'other' output (e.g. 'X' constraint).
6450 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
6452 // Add information to the INLINEASM node to know about this output.
6453 unsigned OpFlags = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
6454 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlags,
6455 TLI->getPointerTy()));
6456 AsmNodeOperands.push_back(OpInfo.CallOperand);
6460 // Otherwise, this is a register or register class output.
6462 // Copy the output from the appropriate register. Find a register that
6464 if (OpInfo.AssignedRegs.Regs.empty()) {
6465 LLVMContext &Ctx = *DAG.getContext();
6466 Ctx.emitError(CS.getInstruction(),
6467 "couldn't allocate output register for constraint '" +
6468 Twine(OpInfo.ConstraintCode) + "'");
6472 // If this is an indirect operand, store through the pointer after the
6474 if (OpInfo.isIndirect) {
6475 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
6476 OpInfo.CallOperandVal));
6478 // This is the result value of the call.
6479 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
6480 // Concatenate this output onto the outputs list.
6481 RetValRegs.append(OpInfo.AssignedRegs);
6484 // Add information to the INLINEASM node to know that this register is
6487 .AddInlineAsmOperands(OpInfo.isEarlyClobber
6488 ? InlineAsm::Kind_RegDefEarlyClobber
6489 : InlineAsm::Kind_RegDef,
6490 false, 0, DAG, AsmNodeOperands);
6493 case InlineAsm::isInput: {
6494 SDValue InOperandVal = OpInfo.CallOperand;
6496 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint?
6497 // If this is required to match an output register we have already set,
6498 // just use its register.
6499 unsigned OperandNo = OpInfo.getMatchedOperand();
6501 // Scan until we find the definition we already emitted of this operand.
6502 // When we find it, create a RegsForValue operand.
6503 unsigned CurOp = InlineAsm::Op_FirstOperand;
6504 for (; OperandNo; --OperandNo) {
6505 // Advance to the next operand.
6507 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6508 assert((InlineAsm::isRegDefKind(OpFlag) ||
6509 InlineAsm::isRegDefEarlyClobberKind(OpFlag) ||
6510 InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?");
6511 CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1;
6515 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6516 if (InlineAsm::isRegDefKind(OpFlag) ||
6517 InlineAsm::isRegDefEarlyClobberKind(OpFlag)) {
6518 // Add (OpFlag&0xffff)>>3 registers to MatchedRegs.
6519 if (OpInfo.isIndirect) {
6520 // This happens on gcc/testsuite/gcc.dg/pr8788-1.c
6521 LLVMContext &Ctx = *DAG.getContext();
6522 Ctx.emitError(CS.getInstruction(), "inline asm not supported yet:"
6523 " don't know how to handle tied "
6524 "indirect register inputs");
6528 RegsForValue MatchedRegs;
6529 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
6530 MVT RegVT = AsmNodeOperands[CurOp+1].getSimpleValueType();
6531 MatchedRegs.RegVTs.push_back(RegVT);
6532 MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo();
6533 for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag);
6535 if (const TargetRegisterClass *RC = TLI->getRegClassFor(RegVT))
6536 MatchedRegs.Regs.push_back(RegInfo.createVirtualRegister(RC));
6538 LLVMContext &Ctx = *DAG.getContext();
6539 Ctx.emitError(CS.getInstruction(),
6540 "inline asm error: This value"
6541 " type register class is not natively supported!");
6545 // Use the produced MatchedRegs object to
6546 MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurSDLoc(),
6547 Chain, &Flag, CS.getInstruction());
6548 MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse,
6549 true, OpInfo.getMatchedOperand(),
6550 DAG, AsmNodeOperands);
6554 assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!");
6555 assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 &&
6556 "Unexpected number of operands");
6557 // Add information to the INLINEASM node to know about this input.
6558 // See InlineAsm.h isUseOperandTiedToDef.
6559 OpFlag = InlineAsm::getFlagWordForMatchingOp(OpFlag,
6560 OpInfo.getMatchedOperand());
6561 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlag,
6562 TLI->getPointerTy()));
6563 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]);
6567 // Treat indirect 'X' constraint as memory.
6568 if (OpInfo.ConstraintType == TargetLowering::C_Other &&
6570 OpInfo.ConstraintType = TargetLowering::C_Memory;
6572 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
6573 std::vector<SDValue> Ops;
6574 TLI->LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode,
6577 LLVMContext &Ctx = *DAG.getContext();
6578 Ctx.emitError(CS.getInstruction(),
6579 "invalid operand for inline asm constraint '" +
6580 Twine(OpInfo.ConstraintCode) + "'");
6584 // Add information to the INLINEASM node to know about this input.
6585 unsigned ResOpType =
6586 InlineAsm::getFlagWord(InlineAsm::Kind_Imm, Ops.size());
6587 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
6588 TLI->getPointerTy()));
6589 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
6593 if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
6594 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
6595 assert(InOperandVal.getValueType() == TLI->getPointerTy() &&
6596 "Memory operands expect pointer values");
6598 // Add information to the INLINEASM node to know about this input.
6599 unsigned ResOpType = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
6600 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
6601 TLI->getPointerTy()));
6602 AsmNodeOperands.push_back(InOperandVal);
6606 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
6607 OpInfo.ConstraintType == TargetLowering::C_Register) &&
6608 "Unknown constraint type!");
6610 // TODO: Support this.
6611 if (OpInfo.isIndirect) {
6612 LLVMContext &Ctx = *DAG.getContext();
6613 Ctx.emitError(CS.getInstruction(),
6614 "Don't know how to handle indirect register inputs yet "
6615 "for constraint '" +
6616 Twine(OpInfo.ConstraintCode) + "'");
6620 // Copy the input into the appropriate registers.
6621 if (OpInfo.AssignedRegs.Regs.empty()) {
6622 LLVMContext &Ctx = *DAG.getContext();
6623 Ctx.emitError(CS.getInstruction(),
6624 "couldn't allocate input reg for constraint '" +
6625 Twine(OpInfo.ConstraintCode) + "'");
6629 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, getCurSDLoc(),
6630 Chain, &Flag, CS.getInstruction());
6632 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, false, 0,
6633 DAG, AsmNodeOperands);
6636 case InlineAsm::isClobber: {
6637 // Add the clobbered value to the operand list, so that the register
6638 // allocator is aware that the physreg got clobbered.
6639 if (!OpInfo.AssignedRegs.Regs.empty())
6640 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_Clobber,
6648 // Finish up input operands. Set the input chain and add the flag last.
6649 AsmNodeOperands[InlineAsm::Op_InputChain] = Chain;
6650 if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
6652 Chain = DAG.getNode(ISD::INLINEASM, getCurSDLoc(),
6653 DAG.getVTList(MVT::Other, MVT::Glue),
6654 &AsmNodeOperands[0], AsmNodeOperands.size());
6655 Flag = Chain.getValue(1);
6657 // If this asm returns a register value, copy the result from that register
6658 // and set it as the value of the call.
6659 if (!RetValRegs.Regs.empty()) {
6660 SDValue Val = RetValRegs.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(),
6661 Chain, &Flag, CS.getInstruction());
6663 // FIXME: Why don't we do this for inline asms with MRVs?
6664 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) {
6665 EVT ResultType = TLI->getValueType(CS.getType());
6667 // If any of the results of the inline asm is a vector, it may have the
6668 // wrong width/num elts. This can happen for register classes that can
6669 // contain multiple different value types. The preg or vreg allocated may
6670 // not have the same VT as was expected. Convert it to the right type
6671 // with bit_convert.
6672 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) {
6673 Val = DAG.getNode(ISD::BITCAST, getCurSDLoc(),
6676 } else if (ResultType != Val.getValueType() &&
6677 ResultType.isInteger() && Val.getValueType().isInteger()) {
6678 // If a result value was tied to an input value, the computed result may
6679 // have a wider width than the expected result. Extract the relevant
6681 Val = DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), ResultType, Val);
6684 assert(ResultType == Val.getValueType() && "Asm result value mismatch!");
6687 setValue(CS.getInstruction(), Val);
6688 // Don't need to use this as a chain in this case.
6689 if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty())
6693 std::vector<std::pair<SDValue, const Value *> > StoresToEmit;
6695 // Process indirect outputs, first output all of the flagged copies out of
6697 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
6698 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
6699 const Value *Ptr = IndirectStoresToEmit[i].second;
6700 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(),
6702 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
6705 // Emit the non-flagged stores from the physregs.
6706 SmallVector<SDValue, 8> OutChains;
6707 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) {
6708 SDValue Val = DAG.getStore(Chain, getCurSDLoc(),
6709 StoresToEmit[i].first,
6710 getValue(StoresToEmit[i].second),
6711 MachinePointerInfo(StoresToEmit[i].second),
6713 OutChains.push_back(Val);
6716 if (!OutChains.empty())
6717 Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
6718 &OutChains[0], OutChains.size());
6723 void SelectionDAGBuilder::visitVAStart(const CallInst &I) {
6724 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurSDLoc(),
6725 MVT::Other, getRoot(),
6726 getValue(I.getArgOperand(0)),
6727 DAG.getSrcValue(I.getArgOperand(0))));
6730 void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) {
6731 const TargetLowering *TLI = TM.getTargetLowering();
6732 const DataLayout &TD = *TLI->getDataLayout();
6733 SDValue V = DAG.getVAArg(TLI->getValueType(I.getType()), getCurSDLoc(),
6734 getRoot(), getValue(I.getOperand(0)),
6735 DAG.getSrcValue(I.getOperand(0)),
6736 TD.getABITypeAlignment(I.getType()));
6738 DAG.setRoot(V.getValue(1));
6741 void SelectionDAGBuilder::visitVAEnd(const CallInst &I) {
6742 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurSDLoc(),
6743 MVT::Other, getRoot(),
6744 getValue(I.getArgOperand(0)),
6745 DAG.getSrcValue(I.getArgOperand(0))));
6748 void SelectionDAGBuilder::visitVACopy(const CallInst &I) {
6749 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurSDLoc(),
6750 MVT::Other, getRoot(),
6751 getValue(I.getArgOperand(0)),
6752 getValue(I.getArgOperand(1)),
6753 DAG.getSrcValue(I.getArgOperand(0)),
6754 DAG.getSrcValue(I.getArgOperand(1))));
6757 /// \brief Lower an argument list according to the target calling convention.
6759 /// \return A tuple of <return-value, token-chain>
6761 /// This is a helper for lowering intrinsics that follow a target calling
6762 /// convention or require stack pointer adjustment. Only a subset of the
6763 /// intrinsic's operands need to participate in the calling convention.
6764 std::pair<SDValue, SDValue>
6765 SelectionDAGBuilder::LowerCallOperands(const CallInst &CI, unsigned ArgIdx,
6766 unsigned NumArgs, SDValue Callee,
6768 TargetLowering::ArgListTy Args;
6769 Args.reserve(NumArgs);
6771 // Populate the argument list.
6772 // Attributes for args start at offset 1, after the return attribute.
6773 ImmutableCallSite CS(&CI);
6774 for (unsigned ArgI = ArgIdx, ArgE = ArgIdx + NumArgs, AttrI = ArgIdx + 1;
6775 ArgI != ArgE; ++ArgI) {
6776 const Value *V = CI.getOperand(ArgI);
6778 assert(!V->getType()->isEmptyTy() && "Empty type passed to intrinsic.");
6780 TargetLowering::ArgListEntry Entry;
6781 Entry.Node = getValue(V);
6782 Entry.Ty = V->getType();
6783 Entry.setAttributes(&CS, AttrI);
6784 Args.push_back(Entry);
6787 Type *retTy = useVoidTy ? Type::getVoidTy(*DAG.getContext()) : CI.getType();
6788 TargetLowering::CallLoweringInfo CLI(getRoot(), retTy, /*retSExt*/ false,
6789 /*retZExt*/ false, /*isVarArg*/ false, /*isInReg*/ false, NumArgs,
6790 CI.getCallingConv(), /*isTailCall*/ false, /*doesNotReturn*/ false,
6791 /*isReturnValueUsed*/ CI.use_empty(), Callee, Args, DAG, getCurSDLoc());
6793 const TargetLowering *TLI = TM.getTargetLowering();
6794 return TLI->LowerCallTo(CLI);
6797 /// \brief Add a stack map intrinsic call's live variable operands to a stackmap
6798 /// or patchpoint target node's operand list.
6800 /// Constants are converted to TargetConstants purely as an optimization to
6801 /// avoid constant materialization and register allocation.
6803 /// FrameIndex operands are converted to TargetFrameIndex so that ISEL does not
6804 /// generate addess computation nodes, and so ExpandISelPseudo can convert the
6805 /// TargetFrameIndex into a DirectMemRefOp StackMap location. This avoids
6806 /// address materialization and register allocation, but may also be required
6807 /// for correctness. If a StackMap (or PatchPoint) intrinsic directly uses an
6808 /// alloca in the entry block, then the runtime may assume that the alloca's
6809 /// StackMap location can be read immediately after compilation and that the
6810 /// location is valid at any point during execution (this is similar to the
6811 /// assumption made by the llvm.gcroot intrinsic). If the alloca's location were
6812 /// only available in a register, then the runtime would need to trap when
6813 /// execution reaches the StackMap in order to read the alloca's location.
6814 static void addStackMapLiveVars(const CallInst &CI, unsigned StartIdx,
6815 SmallVectorImpl<SDValue> &Ops,
6816 SelectionDAGBuilder &Builder) {
6817 for (unsigned i = StartIdx, e = CI.getNumArgOperands(); i != e; ++i) {
6818 SDValue OpVal = Builder.getValue(CI.getArgOperand(i));
6819 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(OpVal)) {
6821 Builder.DAG.getTargetConstant(StackMaps::ConstantOp, MVT::i64));
6823 Builder.DAG.getTargetConstant(C->getSExtValue(), MVT::i64));
6824 } else if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(OpVal)) {
6825 const TargetLowering &TLI = Builder.DAG.getTargetLoweringInfo();
6827 Builder.DAG.getTargetFrameIndex(FI->getIndex(), TLI.getPointerTy()));
6829 Ops.push_back(OpVal);
6833 /// \brief Lower llvm.experimental.stackmap directly to its target opcode.
6834 void SelectionDAGBuilder::visitStackmap(const CallInst &CI) {
6835 // void @llvm.experimental.stackmap(i32 <id>, i32 <numShadowBytes>,
6836 // [live variables...])
6838 assert(CI.getType()->isVoidTy() && "Stackmap cannot return a value.");
6840 SDValue Callee = getValue(CI.getCalledValue());
6842 // Lower into a call sequence with no args and no return value.
6843 std::pair<SDValue, SDValue> Result = LowerCallOperands(CI, 0, 0, Callee);
6844 // Set the root to the target-lowered call chain.
6845 SDValue Chain = Result.second;
6848 /// Get a call instruction from the call sequence chain.
6849 /// Tail calls are not allowed.
6850 SDNode *CallEnd = Chain.getNode();
6851 assert(CallEnd->getOpcode() == ISD::CALLSEQ_END &&
6852 "Expected a callseq node.");
6853 SDNode *Call = CallEnd->getOperand(0).getNode();
6854 bool hasGlue = Call->getGluedNode();
6856 // Replace the target specific call node with the stackmap intrinsic.
6857 SmallVector<SDValue, 8> Ops;
6859 // Add the <id> and <numShadowBytes> constants.
6860 for (unsigned i = 0; i < 2; ++i) {
6861 SDValue tmp = getValue(CI.getOperand(i));
6862 Ops.push_back(DAG.getTargetConstant(
6863 cast<ConstantSDNode>(tmp)->getZExtValue(), MVT::i32));
6865 // Push live variables for the stack map.
6866 addStackMapLiveVars(CI, 2, Ops, *this);
6868 // Push the chain (this is originally the first operand of the call, but
6869 // becomes now the last or second to last operand).
6870 Ops.push_back(*(Call->op_begin()));
6872 // Push the glue flag (last operand).
6874 Ops.push_back(*(Call->op_end()-1));
6876 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6878 // Replace the target specific call node with a STACKMAP node.
6879 MachineSDNode *MN = DAG.getMachineNode(TargetOpcode::STACKMAP, getCurSDLoc(),
6882 // StackMap generates no value, so nothing goes in the NodeMap.
6884 // Fixup the consumers of the intrinsic. The chain and glue may be used in the
6886 DAG.ReplaceAllUsesWith(Call, MN);
6888 DAG.DeleteNode(Call);
6890 // Inform the Frame Information that we have a stackmap in this function.
6891 FuncInfo.MF->getFrameInfo()->setHasStackMap();
6894 /// \brief Lower llvm.experimental.patchpoint directly to its target opcode.
6895 void SelectionDAGBuilder::visitPatchpoint(const CallInst &CI) {
6896 // void|i64 @llvm.experimental.patchpoint.void|i64(i64 <id>,
6901 // [live variables...])
6903 CallingConv::ID CC = CI.getCallingConv();
6904 bool isAnyRegCC = CC == CallingConv::AnyReg;
6905 bool hasDef = !CI.getType()->isVoidTy();
6906 SDValue Callee = getValue(CI.getOperand(2)); // <target>
6908 // Get the real number of arguments participating in the call <numArgs>
6909 SDValue NArgVal = getValue(CI.getArgOperand(PatchPointOpers::NArgPos));
6910 unsigned NumArgs = cast<ConstantSDNode>(NArgVal)->getZExtValue();
6912 // Skip the four meta args: <id>, <numNopBytes>, <target>, <numArgs>
6913 // Intrinsics include all meta-operands up to but not including CC.
6914 unsigned NumMetaOpers = PatchPointOpers::CCPos;
6915 assert(CI.getNumArgOperands() >= NumMetaOpers + NumArgs &&
6916 "Not enough arguments provided to the patchpoint intrinsic");
6918 // For AnyRegCC the arguments are lowered later on manually.
6919 unsigned NumCallArgs = isAnyRegCC ? 0 : NumArgs;
6920 std::pair<SDValue, SDValue> Result =
6921 LowerCallOperands(CI, NumMetaOpers, NumCallArgs, Callee, isAnyRegCC);
6923 // Set the root to the target-lowered call chain.
6924 SDValue Chain = Result.second;
6927 SDNode *CallEnd = Chain.getNode();
6928 if (hasDef && (CallEnd->getOpcode() == ISD::CopyFromReg))
6929 CallEnd = CallEnd->getOperand(0).getNode();
6931 /// Get a call instruction from the call sequence chain.
6932 /// Tail calls are not allowed.
6933 assert(CallEnd->getOpcode() == ISD::CALLSEQ_END &&
6934 "Expected a callseq node.");
6935 SDNode *Call = CallEnd->getOperand(0).getNode();
6936 bool hasGlue = Call->getGluedNode();
6938 // Replace the target specific call node with the patchable intrinsic.
6939 SmallVector<SDValue, 8> Ops;
6941 // Add the <id> and <numBytes> constants.
6942 SDValue IDVal = getValue(CI.getOperand(PatchPointOpers::IDPos));
6943 Ops.push_back(DAG.getTargetConstant(
6944 cast<ConstantSDNode>(IDVal)->getZExtValue(), MVT::i64));
6945 SDValue NBytesVal = getValue(CI.getOperand(PatchPointOpers::NBytesPos));
6946 Ops.push_back(DAG.getTargetConstant(
6947 cast<ConstantSDNode>(NBytesVal)->getZExtValue(), MVT::i32));
6949 // Assume that the Callee is a constant address.
6950 // FIXME: handle function symbols in the future.
6952 DAG.getIntPtrConstant(cast<ConstantSDNode>(Callee)->getZExtValue(),
6953 /*isTarget=*/true));
6955 // Adjust <numArgs> to account for any arguments that have been passed on the
6957 // Call Node: Chain, Target, {Args}, RegMask, [Glue]
6958 unsigned NumCallRegArgs = Call->getNumOperands() - (hasGlue ? 4 : 3);
6959 NumCallRegArgs = isAnyRegCC ? NumArgs : NumCallRegArgs;
6960 Ops.push_back(DAG.getTargetConstant(NumCallRegArgs, MVT::i32));
6962 // Add the calling convention
6963 Ops.push_back(DAG.getTargetConstant((unsigned)CC, MVT::i32));
6965 // Add the arguments we omitted previously. The register allocator should
6966 // place these in any free register.
6968 for (unsigned i = NumMetaOpers, e = NumMetaOpers + NumArgs; i != e; ++i)
6969 Ops.push_back(getValue(CI.getArgOperand(i)));
6971 // Push the arguments from the call instruction up to the register mask.
6972 SDNode::op_iterator e = hasGlue ? Call->op_end()-2 : Call->op_end()-1;
6973 for (SDNode::op_iterator i = Call->op_begin()+2; i != e; ++i)
6976 // Push live variables for the stack map.
6977 addStackMapLiveVars(CI, NumMetaOpers + NumArgs, Ops, *this);
6979 // Push the register mask info.
6981 Ops.push_back(*(Call->op_end()-2));
6983 Ops.push_back(*(Call->op_end()-1));
6985 // Push the chain (this is originally the first operand of the call, but
6986 // becomes now the last or second to last operand).
6987 Ops.push_back(*(Call->op_begin()));
6989 // Push the glue flag (last operand).
6991 Ops.push_back(*(Call->op_end()-1));
6994 if (isAnyRegCC && hasDef) {
6995 // Create the return types based on the intrinsic definition
6996 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6997 SmallVector<EVT, 3> ValueVTs;
6998 ComputeValueVTs(TLI, CI.getType(), ValueVTs);
6999 assert(ValueVTs.size() == 1 && "Expected only one return value type.");
7001 // There is always a chain and a glue type at the end
7002 ValueVTs.push_back(MVT::Other);
7003 ValueVTs.push_back(MVT::Glue);
7004 NodeTys = DAG.getVTList(ValueVTs.data(), ValueVTs.size());
7006 NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
7008 // Replace the target specific call node with a PATCHPOINT node.
7009 MachineSDNode *MN = DAG.getMachineNode(TargetOpcode::PATCHPOINT,
7010 getCurSDLoc(), NodeTys, Ops);
7012 // Update the NodeMap.
7015 setValue(&CI, SDValue(MN, 0));
7017 setValue(&CI, Result.first);
7020 // Fixup the consumers of the intrinsic. The chain and glue may be used in the
7021 // call sequence. Furthermore the location of the chain and glue can change
7022 // when the AnyReg calling convention is used and the intrinsic returns a
7024 if (isAnyRegCC && hasDef) {
7025 SDValue From[] = {SDValue(Call, 0), SDValue(Call, 1)};
7026 SDValue To[] = {SDValue(MN, 1), SDValue(MN, 2)};
7027 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
7029 DAG.ReplaceAllUsesWith(Call, MN);
7030 DAG.DeleteNode(Call);
7032 // Inform the Frame Information that we have a patchpoint in this function.
7033 FuncInfo.MF->getFrameInfo()->setHasPatchPoint();
7036 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
7037 /// implementation, which just calls LowerCall.
7038 /// FIXME: When all targets are
7039 /// migrated to using LowerCall, this hook should be integrated into SDISel.
7040 std::pair<SDValue, SDValue>
7041 TargetLowering::LowerCallTo(TargetLowering::CallLoweringInfo &CLI) const {
7042 // Handle the incoming return values from the call.
7044 SmallVector<EVT, 4> RetTys;
7045 ComputeValueVTs(*this, CLI.RetTy, RetTys);
7046 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
7048 MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
7049 unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
7050 for (unsigned i = 0; i != NumRegs; ++i) {
7051 ISD::InputArg MyFlags;
7052 MyFlags.VT = RegisterVT;
7054 MyFlags.Used = CLI.IsReturnValueUsed;
7056 MyFlags.Flags.setSExt();
7058 MyFlags.Flags.setZExt();
7060 MyFlags.Flags.setInReg();
7061 CLI.Ins.push_back(MyFlags);
7065 // Handle all of the outgoing arguments.
7067 CLI.OutVals.clear();
7068 ArgListTy &Args = CLI.Args;
7069 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
7070 SmallVector<EVT, 4> ValueVTs;
7071 ComputeValueVTs(*this, Args[i].Ty, ValueVTs);
7072 for (unsigned Value = 0, NumValues = ValueVTs.size();
7073 Value != NumValues; ++Value) {
7074 EVT VT = ValueVTs[Value];
7075 Type *ArgTy = VT.getTypeForEVT(CLI.RetTy->getContext());
7076 SDValue Op = SDValue(Args[i].Node.getNode(),
7077 Args[i].Node.getResNo() + Value);
7078 ISD::ArgFlagsTy Flags;
7079 unsigned OriginalAlignment =
7080 getDataLayout()->getABITypeAlignment(ArgTy);
7086 if (Args[i].isInReg)
7090 if (Args[i].isByVal) {
7092 PointerType *Ty = cast<PointerType>(Args[i].Ty);
7093 Type *ElementTy = Ty->getElementType();
7094 Flags.setByValSize(getDataLayout()->getTypeAllocSize(ElementTy));
7095 // For ByVal, alignment should come from FE. BE will guess if this
7096 // info is not there but there are cases it cannot get right.
7097 unsigned FrameAlign;
7098 if (Args[i].Alignment)
7099 FrameAlign = Args[i].Alignment;
7101 FrameAlign = getByValTypeAlignment(ElementTy);
7102 Flags.setByValAlign(FrameAlign);
7106 Flags.setOrigAlign(OriginalAlignment);
7108 MVT PartVT = getRegisterType(CLI.RetTy->getContext(), VT);
7109 unsigned NumParts = getNumRegisters(CLI.RetTy->getContext(), VT);
7110 SmallVector<SDValue, 4> Parts(NumParts);
7111 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
7114 ExtendKind = ISD::SIGN_EXTEND;
7115 else if (Args[i].isZExt)
7116 ExtendKind = ISD::ZERO_EXTEND;
7118 // Conservatively only handle 'returned' on non-vectors for now
7119 if (Args[i].isReturned && !Op.getValueType().isVector()) {
7120 assert(CLI.RetTy == Args[i].Ty && RetTys.size() == NumValues &&
7121 "unexpected use of 'returned'");
7122 // Before passing 'returned' to the target lowering code, ensure that
7123 // either the register MVT and the actual EVT are the same size or that
7124 // the return value and argument are extended in the same way; in these
7125 // cases it's safe to pass the argument register value unchanged as the
7126 // return register value (although it's at the target's option whether
7128 // TODO: allow code generation to take advantage of partially preserved
7129 // registers rather than clobbering the entire register when the
7130 // parameter extension method is not compatible with the return
7132 if ((NumParts * PartVT.getSizeInBits() == VT.getSizeInBits()) ||
7133 (ExtendKind != ISD::ANY_EXTEND &&
7134 CLI.RetSExt == Args[i].isSExt && CLI.RetZExt == Args[i].isZExt))
7135 Flags.setReturned();
7138 getCopyToParts(CLI.DAG, CLI.DL, Op, &Parts[0], NumParts,
7139 PartVT, CLI.CS ? CLI.CS->getInstruction() : 0, ExtendKind);
7141 for (unsigned j = 0; j != NumParts; ++j) {
7142 // if it isn't first piece, alignment must be 1
7143 ISD::OutputArg MyFlags(Flags, Parts[j].getValueType(), VT,
7144 i < CLI.NumFixedArgs,
7145 i, j*Parts[j].getValueType().getStoreSize());
7146 if (NumParts > 1 && j == 0)
7147 MyFlags.Flags.setSplit();
7149 MyFlags.Flags.setOrigAlign(1);
7151 CLI.Outs.push_back(MyFlags);
7152 CLI.OutVals.push_back(Parts[j]);
7157 SmallVector<SDValue, 4> InVals;
7158 CLI.Chain = LowerCall(CLI, InVals);
7160 // Verify that the target's LowerCall behaved as expected.
7161 assert(CLI.Chain.getNode() && CLI.Chain.getValueType() == MVT::Other &&
7162 "LowerCall didn't return a valid chain!");
7163 assert((!CLI.IsTailCall || InVals.empty()) &&
7164 "LowerCall emitted a return value for a tail call!");
7165 assert((CLI.IsTailCall || InVals.size() == CLI.Ins.size()) &&
7166 "LowerCall didn't emit the correct number of values!");
7168 // For a tail call, the return value is merely live-out and there aren't
7169 // any nodes in the DAG representing it. Return a special value to
7170 // indicate that a tail call has been emitted and no more Instructions
7171 // should be processed in the current block.
7172 if (CLI.IsTailCall) {
7173 CLI.DAG.setRoot(CLI.Chain);
7174 return std::make_pair(SDValue(), SDValue());
7177 DEBUG(for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++i) {
7178 assert(InVals[i].getNode() &&
7179 "LowerCall emitted a null value!");
7180 assert(EVT(CLI.Ins[i].VT) == InVals[i].getValueType() &&
7181 "LowerCall emitted a value with the wrong type!");
7184 // Collect the legal value parts into potentially illegal values
7185 // that correspond to the original function's return values.
7186 ISD::NodeType AssertOp = ISD::DELETED_NODE;
7188 AssertOp = ISD::AssertSext;
7189 else if (CLI.RetZExt)
7190 AssertOp = ISD::AssertZext;
7191 SmallVector<SDValue, 4> ReturnValues;
7192 unsigned CurReg = 0;
7193 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
7195 MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
7196 unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
7198 ReturnValues.push_back(getCopyFromParts(CLI.DAG, CLI.DL, &InVals[CurReg],
7199 NumRegs, RegisterVT, VT, NULL,
7204 // For a function returning void, there is no return value. We can't create
7205 // such a node, so we just return a null return value in that case. In
7206 // that case, nothing will actually look at the value.
7207 if (ReturnValues.empty())
7208 return std::make_pair(SDValue(), CLI.Chain);
7210 SDValue Res = CLI.DAG.getNode(ISD::MERGE_VALUES, CLI.DL,
7211 CLI.DAG.getVTList(&RetTys[0], RetTys.size()),
7212 &ReturnValues[0], ReturnValues.size());
7213 return std::make_pair(Res, CLI.Chain);
7216 void TargetLowering::LowerOperationWrapper(SDNode *N,
7217 SmallVectorImpl<SDValue> &Results,
7218 SelectionDAG &DAG) const {
7219 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
7221 Results.push_back(Res);
7224 SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
7225 llvm_unreachable("LowerOperation not implemented for this target!");
7229 SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) {
7230 SDValue Op = getNonRegisterValue(V);
7231 assert((Op.getOpcode() != ISD::CopyFromReg ||
7232 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
7233 "Copy from a reg to the same reg!");
7234 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg");
7236 const TargetLowering *TLI = TM.getTargetLowering();
7237 RegsForValue RFV(V->getContext(), *TLI, Reg, V->getType());
7238 SDValue Chain = DAG.getEntryNode();
7239 RFV.getCopyToRegs(Op, DAG, getCurSDLoc(), Chain, 0, V);
7240 PendingExports.push_back(Chain);
7243 #include "llvm/CodeGen/SelectionDAGISel.h"
7245 /// isOnlyUsedInEntryBlock - If the specified argument is only used in the
7246 /// entry block, return true. This includes arguments used by switches, since
7247 /// the switch may expand into multiple basic blocks.
7248 static bool isOnlyUsedInEntryBlock(const Argument *A, bool FastISel) {
7249 // With FastISel active, we may be splitting blocks, so force creation
7250 // of virtual registers for all non-dead arguments.
7252 return A->use_empty();
7254 const BasicBlock *Entry = A->getParent()->begin();
7255 for (Value::const_use_iterator UI = A->use_begin(), E = A->use_end();
7257 const User *U = *UI;
7258 if (cast<Instruction>(U)->getParent() != Entry || isa<SwitchInst>(U))
7259 return false; // Use not in entry block.
7264 void SelectionDAGISel::LowerArguments(const Function &F) {
7265 SelectionDAG &DAG = SDB->DAG;
7266 SDLoc dl = SDB->getCurSDLoc();
7267 const TargetLowering *TLI = getTargetLowering();
7268 const DataLayout *TD = TLI->getDataLayout();
7269 SmallVector<ISD::InputArg, 16> Ins;
7271 if (!FuncInfo->CanLowerReturn) {
7272 // Put in an sret pointer parameter before all the other parameters.
7273 SmallVector<EVT, 1> ValueVTs;
7274 ComputeValueVTs(*getTargetLowering(),
7275 PointerType::getUnqual(F.getReturnType()), ValueVTs);
7277 // NOTE: Assuming that a pointer will never break down to more than one VT
7279 ISD::ArgFlagsTy Flags;
7281 MVT RegisterVT = TLI->getRegisterType(*DAG.getContext(), ValueVTs[0]);
7282 ISD::InputArg RetArg(Flags, RegisterVT, ValueVTs[0], true, 0, 0);
7283 Ins.push_back(RetArg);
7286 // Set up the incoming argument description vector.
7288 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end();
7289 I != E; ++I, ++Idx) {
7290 SmallVector<EVT, 4> ValueVTs;
7291 ComputeValueVTs(*TLI, I->getType(), ValueVTs);
7292 bool isArgValueUsed = !I->use_empty();
7293 unsigned PartBase = 0;
7294 for (unsigned Value = 0, NumValues = ValueVTs.size();
7295 Value != NumValues; ++Value) {
7296 EVT VT = ValueVTs[Value];
7297 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
7298 ISD::ArgFlagsTy Flags;
7299 unsigned OriginalAlignment =
7300 TD->getABITypeAlignment(ArgTy);
7302 if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt))
7304 if (F.getAttributes().hasAttribute(Idx, Attribute::SExt))
7306 if (F.getAttributes().hasAttribute(Idx, Attribute::InReg))
7308 if (F.getAttributes().hasAttribute(Idx, Attribute::StructRet))
7310 if (F.getAttributes().hasAttribute(Idx, Attribute::ByVal)) {
7312 PointerType *Ty = cast<PointerType>(I->getType());
7313 Type *ElementTy = Ty->getElementType();
7314 Flags.setByValSize(TD->getTypeAllocSize(ElementTy));
7315 // For ByVal, alignment should be passed from FE. BE will guess if
7316 // this info is not there but there are cases it cannot get right.
7317 unsigned FrameAlign;
7318 if (F.getParamAlignment(Idx))
7319 FrameAlign = F.getParamAlignment(Idx);
7321 FrameAlign = TLI->getByValTypeAlignment(ElementTy);
7322 Flags.setByValAlign(FrameAlign);
7324 if (F.getAttributes().hasAttribute(Idx, Attribute::Nest))
7326 Flags.setOrigAlign(OriginalAlignment);
7328 MVT RegisterVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7329 unsigned NumRegs = TLI->getNumRegisters(*CurDAG->getContext(), VT);
7330 for (unsigned i = 0; i != NumRegs; ++i) {
7331 ISD::InputArg MyFlags(Flags, RegisterVT, VT, isArgValueUsed,
7332 Idx-1, PartBase+i*RegisterVT.getStoreSize());
7333 if (NumRegs > 1 && i == 0)
7334 MyFlags.Flags.setSplit();
7335 // if it isn't first piece, alignment must be 1
7337 MyFlags.Flags.setOrigAlign(1);
7338 Ins.push_back(MyFlags);
7340 PartBase += VT.getStoreSize();
7344 // Call the target to set up the argument values.
7345 SmallVector<SDValue, 8> InVals;
7346 SDValue NewRoot = TLI->LowerFormalArguments(DAG.getRoot(), F.getCallingConv(),
7350 // Verify that the target's LowerFormalArguments behaved as expected.
7351 assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other &&
7352 "LowerFormalArguments didn't return a valid chain!");
7353 assert(InVals.size() == Ins.size() &&
7354 "LowerFormalArguments didn't emit the correct number of values!");
7356 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
7357 assert(InVals[i].getNode() &&
7358 "LowerFormalArguments emitted a null value!");
7359 assert(EVT(Ins[i].VT) == InVals[i].getValueType() &&
7360 "LowerFormalArguments emitted a value with the wrong type!");
7364 // Update the DAG with the new chain value resulting from argument lowering.
7365 DAG.setRoot(NewRoot);
7367 // Set up the argument values.
7370 if (!FuncInfo->CanLowerReturn) {
7371 // Create a virtual register for the sret pointer, and put in a copy
7372 // from the sret argument into it.
7373 SmallVector<EVT, 1> ValueVTs;
7374 ComputeValueVTs(*TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
7375 MVT VT = ValueVTs[0].getSimpleVT();
7376 MVT RegVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7377 ISD::NodeType AssertOp = ISD::DELETED_NODE;
7378 SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1,
7379 RegVT, VT, NULL, AssertOp);
7381 MachineFunction& MF = SDB->DAG.getMachineFunction();
7382 MachineRegisterInfo& RegInfo = MF.getRegInfo();
7383 unsigned SRetReg = RegInfo.createVirtualRegister(TLI->getRegClassFor(RegVT));
7384 FuncInfo->DemoteRegister = SRetReg;
7385 NewRoot = SDB->DAG.getCopyToReg(NewRoot, SDB->getCurSDLoc(),
7387 DAG.setRoot(NewRoot);
7389 // i indexes lowered arguments. Bump it past the hidden sret argument.
7390 // Idx indexes LLVM arguments. Don't touch it.
7394 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
7396 SmallVector<SDValue, 4> ArgValues;
7397 SmallVector<EVT, 4> ValueVTs;
7398 ComputeValueVTs(*TLI, I->getType(), ValueVTs);
7399 unsigned NumValues = ValueVTs.size();
7401 // If this argument is unused then remember its value. It is used to generate
7402 // debugging information.
7403 if (I->use_empty() && NumValues) {
7404 SDB->setUnusedArgValue(I, InVals[i]);
7406 // Also remember any frame index for use in FastISel.
7407 if (FrameIndexSDNode *FI =
7408 dyn_cast<FrameIndexSDNode>(InVals[i].getNode()))
7409 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
7412 for (unsigned Val = 0; Val != NumValues; ++Val) {
7413 EVT VT = ValueVTs[Val];
7414 MVT PartVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7415 unsigned NumParts = TLI->getNumRegisters(*CurDAG->getContext(), VT);
7417 if (!I->use_empty()) {
7418 ISD::NodeType AssertOp = ISD::DELETED_NODE;
7419 if (F.getAttributes().hasAttribute(Idx, Attribute::SExt))
7420 AssertOp = ISD::AssertSext;
7421 else if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt))
7422 AssertOp = ISD::AssertZext;
7424 ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i],
7425 NumParts, PartVT, VT,
7432 // We don't need to do anything else for unused arguments.
7433 if (ArgValues.empty())
7436 // Note down frame index.
7437 if (FrameIndexSDNode *FI =
7438 dyn_cast<FrameIndexSDNode>(ArgValues[0].getNode()))
7439 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
7441 SDValue Res = DAG.getMergeValues(&ArgValues[0], NumValues,
7442 SDB->getCurSDLoc());
7444 SDB->setValue(I, Res);
7445 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::BUILD_PAIR) {
7446 if (LoadSDNode *LNode =
7447 dyn_cast<LoadSDNode>(Res.getOperand(0).getNode()))
7448 if (FrameIndexSDNode *FI =
7449 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode()))
7450 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
7453 // If this argument is live outside of the entry block, insert a copy from
7454 // wherever we got it to the vreg that other BB's will reference it as.
7455 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::CopyFromReg) {
7456 // If we can, though, try to skip creating an unnecessary vreg.
7457 // FIXME: This isn't very clean... it would be nice to make this more
7458 // general. It's also subtly incompatible with the hacks FastISel
7460 unsigned Reg = cast<RegisterSDNode>(Res.getOperand(1))->getReg();
7461 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
7462 FuncInfo->ValueMap[I] = Reg;
7466 if (!isOnlyUsedInEntryBlock(I, TM.Options.EnableFastISel)) {
7467 FuncInfo->InitializeRegForValue(I);
7468 SDB->CopyToExportRegsIfNeeded(I);
7472 assert(i == InVals.size() && "Argument register count mismatch!");
7474 // Finally, if the target has anything special to do, allow it to do so.
7475 // FIXME: this should insert code into the DAG!
7476 EmitFunctionEntryCode();
7479 /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
7480 /// ensure constants are generated when needed. Remember the virtual registers
7481 /// that need to be added to the Machine PHI nodes as input. We cannot just
7482 /// directly add them, because expansion might result in multiple MBB's for one
7483 /// BB. As such, the start of the BB might correspond to a different MBB than
7487 SelectionDAGBuilder::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
7488 const TerminatorInst *TI = LLVMBB->getTerminator();
7490 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
7492 // Check successor nodes' PHI nodes that expect a constant to be available
7494 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
7495 const BasicBlock *SuccBB = TI->getSuccessor(succ);
7496 if (!isa<PHINode>(SuccBB->begin())) continue;
7497 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
7499 // If this terminator has multiple identical successors (common for
7500 // switches), only handle each succ once.
7501 if (!SuccsHandled.insert(SuccMBB)) continue;
7503 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
7505 // At this point we know that there is a 1-1 correspondence between LLVM PHI
7506 // nodes and Machine PHI nodes, but the incoming operands have not been
7508 for (BasicBlock::const_iterator I = SuccBB->begin();
7509 const PHINode *PN = dyn_cast<PHINode>(I); ++I) {
7510 // Ignore dead phi's.
7511 if (PN->use_empty()) continue;
7514 if (PN->getType()->isEmptyTy())
7518 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
7520 if (const Constant *C = dyn_cast<Constant>(PHIOp)) {
7521 unsigned &RegOut = ConstantsOut[C];
7523 RegOut = FuncInfo.CreateRegs(C->getType());
7524 CopyValueToVirtualRegister(C, RegOut);
7528 DenseMap<const Value *, unsigned>::iterator I =
7529 FuncInfo.ValueMap.find(PHIOp);
7530 if (I != FuncInfo.ValueMap.end())
7533 assert(isa<AllocaInst>(PHIOp) &&
7534 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
7535 "Didn't codegen value into a register!??");
7536 Reg = FuncInfo.CreateRegs(PHIOp->getType());
7537 CopyValueToVirtualRegister(PHIOp, Reg);
7541 // Remember that this register needs to added to the machine PHI node as
7542 // the input for this MBB.
7543 SmallVector<EVT, 4> ValueVTs;
7544 const TargetLowering *TLI = TM.getTargetLowering();
7545 ComputeValueVTs(*TLI, PN->getType(), ValueVTs);
7546 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
7547 EVT VT = ValueVTs[vti];
7548 unsigned NumRegisters = TLI->getNumRegisters(*DAG.getContext(), VT);
7549 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
7550 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
7551 Reg += NumRegisters;
7556 ConstantsOut.clear();
7559 /// Add a successor MBB to ParentMBB< creating a new MachineBB for BB if SuccMBB
7562 SelectionDAGBuilder::StackProtectorDescriptor::
7563 AddSuccessorMBB(const BasicBlock *BB,
7564 MachineBasicBlock *ParentMBB,
7565 MachineBasicBlock *SuccMBB) {
7566 // If SuccBB has not been created yet, create it.
7568 MachineFunction *MF = ParentMBB->getParent();
7569 MachineFunction::iterator BBI = ParentMBB;
7570 SuccMBB = MF->CreateMachineBasicBlock(BB);
7571 MF->insert(++BBI, SuccMBB);
7573 // Add it as a successor of ParentMBB.
7574 ParentMBB->addSuccessor(SuccMBB);