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 #include "SelectionDAGBuilder.h"
15 #include "SDNodeDbgValue.h"
16 #include "llvm/ADT/BitVector.h"
17 #include "llvm/ADT/Optional.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/Analysis/AliasAnalysis.h"
20 #include "llvm/Analysis/BranchProbabilityInfo.h"
21 #include "llvm/Analysis/ConstantFolding.h"
22 #include "llvm/Analysis/ValueTracking.h"
23 #include "llvm/CodeGen/Analysis.h"
24 #include "llvm/CodeGen/FastISel.h"
25 #include "llvm/CodeGen/FunctionLoweringInfo.h"
26 #include "llvm/CodeGen/GCMetadata.h"
27 #include "llvm/CodeGen/GCStrategy.h"
28 #include "llvm/CodeGen/MachineFrameInfo.h"
29 #include "llvm/CodeGen/MachineFunction.h"
30 #include "llvm/CodeGen/MachineInstrBuilder.h"
31 #include "llvm/CodeGen/MachineJumpTableInfo.h"
32 #include "llvm/CodeGen/MachineModuleInfo.h"
33 #include "llvm/CodeGen/MachineRegisterInfo.h"
34 #include "llvm/CodeGen/SelectionDAG.h"
35 #include "llvm/CodeGen/StackMaps.h"
36 #include "llvm/IR/CallingConv.h"
37 #include "llvm/IR/Constants.h"
38 #include "llvm/IR/DataLayout.h"
39 #include "llvm/IR/DebugInfo.h"
40 #include "llvm/IR/DerivedTypes.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/GlobalVariable.h"
43 #include "llvm/IR/InlineAsm.h"
44 #include "llvm/IR/Instructions.h"
45 #include "llvm/IR/IntrinsicInst.h"
46 #include "llvm/IR/Intrinsics.h"
47 #include "llvm/IR/LLVMContext.h"
48 #include "llvm/IR/Module.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/raw_ostream.h"
54 #include "llvm/Target/TargetFrameLowering.h"
55 #include "llvm/Target/TargetInstrInfo.h"
56 #include "llvm/Target/TargetIntrinsicInfo.h"
57 #include "llvm/Target/TargetLibraryInfo.h"
58 #include "llvm/Target/TargetLowering.h"
59 #include "llvm/Target/TargetOptions.h"
60 #include "llvm/Target/TargetSelectionDAGInfo.h"
64 #define DEBUG_TYPE "isel"
66 /// LimitFloatPrecision - Generate low-precision inline sequences for
67 /// some float libcalls (6, 8 or 12 bits).
68 static unsigned LimitFloatPrecision;
70 static cl::opt<unsigned, true>
71 LimitFPPrecision("limit-float-precision",
72 cl::desc("Generate low-precision inline sequences "
73 "for some float libcalls"),
74 cl::location(LimitFloatPrecision),
77 // Limit the width of DAG chains. This is important in general to prevent
78 // prevent DAG-based analysis from blowing up. For example, alias analysis and
79 // load clustering may not complete in reasonable time. It is difficult to
80 // recognize and avoid this situation within each individual analysis, and
81 // future analyses are likely to have the same behavior. Limiting DAG width is
82 // the safe approach, and will be especially important with global DAGs.
84 // MaxParallelChains default is arbitrarily high to avoid affecting
85 // optimization, but could be lowered to improve compile time. Any ld-ld-st-st
86 // sequence over this should have been converted to llvm.memcpy by the
87 // frontend. It easy to induce this behavior with .ll code such as:
88 // %buffer = alloca [4096 x i8]
89 // %data = load [4096 x i8]* %argPtr
90 // store [4096 x i8] %data, [4096 x i8]* %buffer
91 static const unsigned MaxParallelChains = 64;
93 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, SDLoc DL,
94 const SDValue *Parts, unsigned NumParts,
95 MVT PartVT, EVT ValueVT, const Value *V);
97 /// getCopyFromParts - Create a value that contains the specified legal parts
98 /// combined into the value they represent. If the parts combine to a type
99 /// larger then ValueVT then AssertOp can be used to specify whether the extra
100 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
101 /// (ISD::AssertSext).
102 static SDValue getCopyFromParts(SelectionDAG &DAG, SDLoc DL,
103 const SDValue *Parts,
104 unsigned NumParts, MVT PartVT, EVT ValueVT,
106 ISD::NodeType AssertOp = ISD::DELETED_NODE) {
107 if (ValueVT.isVector())
108 return getCopyFromPartsVector(DAG, DL, Parts, NumParts,
111 assert(NumParts > 0 && "No parts to assemble!");
112 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
113 SDValue Val = Parts[0];
116 // Assemble the value from multiple parts.
117 if (ValueVT.isInteger()) {
118 unsigned PartBits = PartVT.getSizeInBits();
119 unsigned ValueBits = ValueVT.getSizeInBits();
121 // Assemble the power of 2 part.
122 unsigned RoundParts = NumParts & (NumParts - 1) ?
123 1 << Log2_32(NumParts) : NumParts;
124 unsigned RoundBits = PartBits * RoundParts;
125 EVT RoundVT = RoundBits == ValueBits ?
126 ValueVT : EVT::getIntegerVT(*DAG.getContext(), RoundBits);
129 EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2);
131 if (RoundParts > 2) {
132 Lo = getCopyFromParts(DAG, DL, Parts, RoundParts / 2,
134 Hi = getCopyFromParts(DAG, DL, Parts + RoundParts / 2,
135 RoundParts / 2, PartVT, HalfVT, V);
137 Lo = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[0]);
138 Hi = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[1]);
141 if (TLI.isBigEndian())
144 Val = DAG.getNode(ISD::BUILD_PAIR, DL, RoundVT, Lo, Hi);
146 if (RoundParts < NumParts) {
147 // Assemble the trailing non-power-of-2 part.
148 unsigned OddParts = NumParts - RoundParts;
149 EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits);
150 Hi = getCopyFromParts(DAG, DL,
151 Parts + RoundParts, OddParts, PartVT, OddVT, V);
153 // Combine the round and odd parts.
155 if (TLI.isBigEndian())
157 EVT TotalVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
158 Hi = DAG.getNode(ISD::ANY_EXTEND, DL, TotalVT, Hi);
159 Hi = DAG.getNode(ISD::SHL, DL, TotalVT, Hi,
160 DAG.getConstant(Lo.getValueType().getSizeInBits(),
161 TLI.getPointerTy()));
162 Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, TotalVT, Lo);
163 Val = DAG.getNode(ISD::OR, DL, TotalVT, Lo, Hi);
165 } else if (PartVT.isFloatingPoint()) {
166 // FP split into multiple FP parts (for ppcf128)
167 assert(ValueVT == EVT(MVT::ppcf128) && PartVT == MVT::f64 &&
170 Lo = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[0]);
171 Hi = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[1]);
172 if (TLI.hasBigEndianPartOrdering(ValueVT))
174 Val = DAG.getNode(ISD::BUILD_PAIR, DL, ValueVT, Lo, Hi);
176 // FP split into integer parts (soft fp)
177 assert(ValueVT.isFloatingPoint() && PartVT.isInteger() &&
178 !PartVT.isVector() && "Unexpected split");
179 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits());
180 Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT, V);
184 // There is now one part, held in Val. Correct it to match ValueVT.
185 EVT PartEVT = Val.getValueType();
187 if (PartEVT == ValueVT)
190 if (PartEVT.isInteger() && ValueVT.isInteger()) {
191 if (ValueVT.bitsLT(PartEVT)) {
192 // For a truncate, see if we have any information to
193 // indicate whether the truncated bits will always be
194 // zero or sign-extension.
195 if (AssertOp != ISD::DELETED_NODE)
196 Val = DAG.getNode(AssertOp, DL, PartEVT, Val,
197 DAG.getValueType(ValueVT));
198 return DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
200 return DAG.getNode(ISD::ANY_EXTEND, DL, ValueVT, Val);
203 if (PartEVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
204 // FP_ROUND's are always exact here.
205 if (ValueVT.bitsLT(Val.getValueType()))
206 return DAG.getNode(ISD::FP_ROUND, DL, ValueVT, Val,
207 DAG.getTargetConstant(1, TLI.getPointerTy()));
209 return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val);
212 if (PartEVT.getSizeInBits() == ValueVT.getSizeInBits())
213 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
215 llvm_unreachable("Unknown mismatch!");
218 static void diagnosePossiblyInvalidConstraint(LLVMContext &Ctx, const Value *V,
219 const Twine &ErrMsg) {
220 const Instruction *I = dyn_cast_or_null<Instruction>(V);
222 return Ctx.emitError(ErrMsg);
224 const char *AsmError = ", possible invalid constraint for vector type";
225 if (const CallInst *CI = dyn_cast<CallInst>(I))
226 if (isa<InlineAsm>(CI->getCalledValue()))
227 return Ctx.emitError(I, ErrMsg + AsmError);
229 return Ctx.emitError(I, ErrMsg);
232 /// getCopyFromPartsVector - Create a value that contains the specified legal
233 /// parts combined into the value they represent. If the parts combine to a
234 /// type larger then ValueVT then AssertOp can be used to specify whether the
235 /// extra bits are known to be zero (ISD::AssertZext) or sign extended from
236 /// ValueVT (ISD::AssertSext).
237 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, SDLoc DL,
238 const SDValue *Parts, unsigned NumParts,
239 MVT PartVT, EVT ValueVT, const Value *V) {
240 assert(ValueVT.isVector() && "Not a vector value");
241 assert(NumParts > 0 && "No parts to assemble!");
242 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
243 SDValue Val = Parts[0];
245 // Handle a multi-element vector.
249 unsigned NumIntermediates;
251 TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT,
252 NumIntermediates, RegisterVT);
253 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
254 NumParts = NumRegs; // Silence a compiler warning.
255 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
256 assert(RegisterVT == Parts[0].getSimpleValueType() &&
257 "Part type doesn't match part!");
259 // Assemble the parts into intermediate operands.
260 SmallVector<SDValue, 8> Ops(NumIntermediates);
261 if (NumIntermediates == NumParts) {
262 // If the register was not expanded, truncate or copy the value,
264 for (unsigned i = 0; i != NumParts; ++i)
265 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i], 1,
266 PartVT, IntermediateVT, V);
267 } else if (NumParts > 0) {
268 // If the intermediate type was expanded, build the intermediate
269 // operands from the parts.
270 assert(NumParts % NumIntermediates == 0 &&
271 "Must expand into a divisible number of parts!");
272 unsigned Factor = NumParts / NumIntermediates;
273 for (unsigned i = 0; i != NumIntermediates; ++i)
274 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor,
275 PartVT, IntermediateVT, V);
278 // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the
279 // intermediate operands.
280 Val = DAG.getNode(IntermediateVT.isVector() ? ISD::CONCAT_VECTORS
285 // There is now one part, held in Val. Correct it to match ValueVT.
286 EVT PartEVT = Val.getValueType();
288 if (PartEVT == ValueVT)
291 if (PartEVT.isVector()) {
292 // If the element type of the source/dest vectors are the same, but the
293 // parts vector has more elements than the value vector, then we have a
294 // vector widening case (e.g. <2 x float> -> <4 x float>). Extract the
296 if (PartEVT.getVectorElementType() == ValueVT.getVectorElementType()) {
297 assert(PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements() &&
298 "Cannot narrow, it would be a lossy transformation");
299 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val,
300 DAG.getConstant(0, TLI.getVectorIdxTy()));
303 // Vector/Vector bitcast.
304 if (ValueVT.getSizeInBits() == PartEVT.getSizeInBits())
305 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
307 assert(PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements() &&
308 "Cannot handle this kind of promotion");
309 // Promoted vector extract
310 bool Smaller = ValueVT.bitsLE(PartEVT);
311 return DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
316 // Trivial bitcast if the types are the same size and the destination
317 // vector type is legal.
318 if (PartEVT.getSizeInBits() == ValueVT.getSizeInBits() &&
319 TLI.isTypeLegal(ValueVT))
320 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
322 // Handle cases such as i8 -> <1 x i1>
323 if (ValueVT.getVectorNumElements() != 1) {
324 diagnosePossiblyInvalidConstraint(*DAG.getContext(), V,
325 "non-trivial scalar-to-vector conversion");
326 return DAG.getUNDEF(ValueVT);
329 if (ValueVT.getVectorNumElements() == 1 &&
330 ValueVT.getVectorElementType() != PartEVT) {
331 bool Smaller = ValueVT.bitsLE(PartEVT);
332 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
333 DL, ValueVT.getScalarType(), Val);
336 return DAG.getNode(ISD::BUILD_VECTOR, DL, ValueVT, Val);
339 static void getCopyToPartsVector(SelectionDAG &DAG, SDLoc dl,
340 SDValue Val, SDValue *Parts, unsigned NumParts,
341 MVT PartVT, const Value *V);
343 /// getCopyToParts - Create a series of nodes that contain the specified value
344 /// split into legal parts. If the parts contain more bits than Val, then, for
345 /// integers, ExtendKind can be used to specify how to generate the extra bits.
346 static void getCopyToParts(SelectionDAG &DAG, SDLoc DL,
347 SDValue Val, SDValue *Parts, unsigned NumParts,
348 MVT PartVT, const Value *V,
349 ISD::NodeType ExtendKind = ISD::ANY_EXTEND) {
350 EVT ValueVT = Val.getValueType();
352 // Handle the vector case separately.
353 if (ValueVT.isVector())
354 return getCopyToPartsVector(DAG, DL, Val, Parts, NumParts, PartVT, V);
356 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
357 unsigned PartBits = PartVT.getSizeInBits();
358 unsigned OrigNumParts = NumParts;
359 assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!");
364 assert(!ValueVT.isVector() && "Vector case handled elsewhere");
365 EVT PartEVT = PartVT;
366 if (PartEVT == ValueVT) {
367 assert(NumParts == 1 && "No-op copy with multiple parts!");
372 if (NumParts * PartBits > ValueVT.getSizeInBits()) {
373 // If the parts cover more bits than the value has, promote the value.
374 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
375 assert(NumParts == 1 && "Do not know what to promote to!");
376 Val = DAG.getNode(ISD::FP_EXTEND, DL, PartVT, Val);
378 assert((PartVT.isInteger() || PartVT == MVT::x86mmx) &&
379 ValueVT.isInteger() &&
380 "Unknown mismatch!");
381 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
382 Val = DAG.getNode(ExtendKind, DL, ValueVT, Val);
383 if (PartVT == MVT::x86mmx)
384 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
386 } else if (PartBits == ValueVT.getSizeInBits()) {
387 // Different types of the same size.
388 assert(NumParts == 1 && PartEVT != ValueVT);
389 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
390 } else if (NumParts * PartBits < ValueVT.getSizeInBits()) {
391 // If the parts cover less bits than value has, truncate the value.
392 assert((PartVT.isInteger() || PartVT == MVT::x86mmx) &&
393 ValueVT.isInteger() &&
394 "Unknown mismatch!");
395 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
396 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
397 if (PartVT == MVT::x86mmx)
398 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
401 // The value may have changed - recompute ValueVT.
402 ValueVT = Val.getValueType();
403 assert(NumParts * PartBits == ValueVT.getSizeInBits() &&
404 "Failed to tile the value with PartVT!");
407 if (PartEVT != ValueVT)
408 diagnosePossiblyInvalidConstraint(*DAG.getContext(), V,
409 "scalar-to-vector conversion failed");
415 // Expand the value into multiple parts.
416 if (NumParts & (NumParts - 1)) {
417 // The number of parts is not a power of 2. Split off and copy the tail.
418 assert(PartVT.isInteger() && ValueVT.isInteger() &&
419 "Do not know what to expand to!");
420 unsigned RoundParts = 1 << Log2_32(NumParts);
421 unsigned RoundBits = RoundParts * PartBits;
422 unsigned OddParts = NumParts - RoundParts;
423 SDValue OddVal = DAG.getNode(ISD::SRL, DL, ValueVT, Val,
424 DAG.getIntPtrConstant(RoundBits));
425 getCopyToParts(DAG, DL, OddVal, Parts + RoundParts, OddParts, PartVT, V);
427 if (TLI.isBigEndian())
428 // The odd parts were reversed by getCopyToParts - unreverse them.
429 std::reverse(Parts + RoundParts, Parts + NumParts);
431 NumParts = RoundParts;
432 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
433 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
436 // The number of parts is a power of 2. Repeatedly bisect the value using
438 Parts[0] = DAG.getNode(ISD::BITCAST, DL,
439 EVT::getIntegerVT(*DAG.getContext(),
440 ValueVT.getSizeInBits()),
443 for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) {
444 for (unsigned i = 0; i < NumParts; i += StepSize) {
445 unsigned ThisBits = StepSize * PartBits / 2;
446 EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits);
447 SDValue &Part0 = Parts[i];
448 SDValue &Part1 = Parts[i+StepSize/2];
450 Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
451 ThisVT, Part0, DAG.getIntPtrConstant(1));
452 Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
453 ThisVT, Part0, DAG.getIntPtrConstant(0));
455 if (ThisBits == PartBits && ThisVT != PartVT) {
456 Part0 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part0);
457 Part1 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part1);
462 if (TLI.isBigEndian())
463 std::reverse(Parts, Parts + OrigNumParts);
467 /// getCopyToPartsVector - Create a series of nodes that contain the specified
468 /// value split into legal parts.
469 static void getCopyToPartsVector(SelectionDAG &DAG, SDLoc DL,
470 SDValue Val, SDValue *Parts, unsigned NumParts,
471 MVT PartVT, const Value *V) {
472 EVT ValueVT = Val.getValueType();
473 assert(ValueVT.isVector() && "Not a vector");
474 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
477 EVT PartEVT = PartVT;
478 if (PartEVT == ValueVT) {
480 } else if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) {
481 // Bitconvert vector->vector case.
482 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
483 } else if (PartVT.isVector() &&
484 PartEVT.getVectorElementType() == ValueVT.getVectorElementType() &&
485 PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements()) {
486 EVT ElementVT = PartVT.getVectorElementType();
487 // Vector widening case, e.g. <2 x float> -> <4 x float>. Shuffle in
489 SmallVector<SDValue, 16> Ops;
490 for (unsigned i = 0, e = ValueVT.getVectorNumElements(); i != e; ++i)
491 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
492 ElementVT, Val, DAG.getConstant(i,
493 TLI.getVectorIdxTy())));
495 for (unsigned i = ValueVT.getVectorNumElements(),
496 e = PartVT.getVectorNumElements(); i != e; ++i)
497 Ops.push_back(DAG.getUNDEF(ElementVT));
499 Val = DAG.getNode(ISD::BUILD_VECTOR, DL, PartVT, Ops);
501 // FIXME: Use CONCAT for 2x -> 4x.
503 //SDValue UndefElts = DAG.getUNDEF(VectorTy);
504 //Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts);
505 } else if (PartVT.isVector() &&
506 PartEVT.getVectorElementType().bitsGE(
507 ValueVT.getVectorElementType()) &&
508 PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements()) {
510 // Promoted vector extract
511 bool Smaller = PartEVT.bitsLE(ValueVT);
512 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
515 // Vector -> scalar conversion.
516 assert(ValueVT.getVectorNumElements() == 1 &&
517 "Only trivial vector-to-scalar conversions should get here!");
518 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
519 PartVT, Val, DAG.getConstant(0, TLI.getVectorIdxTy()));
521 bool Smaller = ValueVT.bitsLE(PartVT);
522 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
530 // Handle a multi-element vector.
533 unsigned NumIntermediates;
534 unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT,
536 NumIntermediates, RegisterVT);
537 unsigned NumElements = ValueVT.getVectorNumElements();
539 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
540 NumParts = NumRegs; // Silence a compiler warning.
541 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
543 // Split the vector into intermediate operands.
544 SmallVector<SDValue, 8> Ops(NumIntermediates);
545 for (unsigned i = 0; i != NumIntermediates; ++i) {
546 if (IntermediateVT.isVector())
547 Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL,
549 DAG.getConstant(i * (NumElements / NumIntermediates),
550 TLI.getVectorIdxTy()));
552 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
554 DAG.getConstant(i, TLI.getVectorIdxTy()));
557 // Split the intermediate operands into legal parts.
558 if (NumParts == NumIntermediates) {
559 // If the register was not expanded, promote or copy the value,
561 for (unsigned i = 0; i != NumParts; ++i)
562 getCopyToParts(DAG, DL, Ops[i], &Parts[i], 1, PartVT, V);
563 } else if (NumParts > 0) {
564 // If the intermediate type was expanded, split each the value into
566 assert(NumParts % NumIntermediates == 0 &&
567 "Must expand into a divisible number of parts!");
568 unsigned Factor = NumParts / NumIntermediates;
569 for (unsigned i = 0; i != NumIntermediates; ++i)
570 getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT, V);
575 /// RegsForValue - This struct represents the registers (physical or virtual)
576 /// that a particular set of values is assigned, and the type information
577 /// about the value. The most common situation is to represent one value at a
578 /// time, but struct or array values are handled element-wise as multiple
579 /// values. The splitting of aggregates is performed recursively, so that we
580 /// never have aggregate-typed registers. The values at this point do not
581 /// necessarily have legal types, so each value may require one or more
582 /// registers of some legal type.
584 struct RegsForValue {
585 /// ValueVTs - The value types of the values, which may not be legal, and
586 /// may need be promoted or synthesized from one or more registers.
588 SmallVector<EVT, 4> ValueVTs;
590 /// RegVTs - The value types of the registers. This is the same size as
591 /// ValueVTs and it records, for each value, what the type of the assigned
592 /// register or registers are. (Individual values are never synthesized
593 /// from more than one type of register.)
595 /// With virtual registers, the contents of RegVTs is redundant with TLI's
596 /// getRegisterType member function, however when with physical registers
597 /// it is necessary to have a separate record of the types.
599 SmallVector<MVT, 4> RegVTs;
601 /// Regs - This list holds the registers assigned to the values.
602 /// Each legal or promoted value requires one register, and each
603 /// expanded value requires multiple registers.
605 SmallVector<unsigned, 4> Regs;
609 RegsForValue(const SmallVector<unsigned, 4> ®s,
610 MVT regvt, EVT valuevt)
611 : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {}
613 RegsForValue(LLVMContext &Context, const TargetLowering &tli,
614 unsigned Reg, Type *Ty) {
615 ComputeValueVTs(tli, Ty, ValueVTs);
617 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
618 EVT ValueVT = ValueVTs[Value];
619 unsigned NumRegs = tli.getNumRegisters(Context, ValueVT);
620 MVT RegisterVT = tli.getRegisterType(Context, ValueVT);
621 for (unsigned i = 0; i != NumRegs; ++i)
622 Regs.push_back(Reg + i);
623 RegVTs.push_back(RegisterVT);
628 /// append - Add the specified values to this one.
629 void append(const RegsForValue &RHS) {
630 ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end());
631 RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end());
632 Regs.append(RHS.Regs.begin(), RHS.Regs.end());
635 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
636 /// this value and returns the result as a ValueVTs value. This uses
637 /// Chain/Flag as the input and updates them for the output Chain/Flag.
638 /// If the Flag pointer is NULL, no flag is used.
639 SDValue getCopyFromRegs(SelectionDAG &DAG, FunctionLoweringInfo &FuncInfo,
641 SDValue &Chain, SDValue *Flag,
642 const Value *V = nullptr) const;
644 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
645 /// specified value into the registers specified by this object. This uses
646 /// Chain/Flag as the input and updates them for the output Chain/Flag.
647 /// If the Flag pointer is NULL, no flag is used.
648 void getCopyToRegs(SDValue Val, SelectionDAG &DAG, SDLoc dl,
649 SDValue &Chain, SDValue *Flag, const Value *V) const;
651 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
652 /// operand list. This adds the code marker, matching input operand index
653 /// (if applicable), and includes the number of values added into it.
654 void AddInlineAsmOperands(unsigned Kind,
655 bool HasMatching, unsigned MatchingIdx,
657 std::vector<SDValue> &Ops) const;
661 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
662 /// this value and returns the result as a ValueVT value. This uses
663 /// Chain/Flag as the input and updates them for the output Chain/Flag.
664 /// If the Flag pointer is NULL, no flag is used.
665 SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
666 FunctionLoweringInfo &FuncInfo,
668 SDValue &Chain, SDValue *Flag,
669 const Value *V) const {
670 // A Value with type {} or [0 x %t] needs no registers.
671 if (ValueVTs.empty())
674 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
676 // Assemble the legal parts into the final values.
677 SmallVector<SDValue, 4> Values(ValueVTs.size());
678 SmallVector<SDValue, 8> Parts;
679 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
680 // Copy the legal parts from the registers.
681 EVT ValueVT = ValueVTs[Value];
682 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
683 MVT RegisterVT = RegVTs[Value];
685 Parts.resize(NumRegs);
686 for (unsigned i = 0; i != NumRegs; ++i) {
689 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT);
691 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag);
692 *Flag = P.getValue(2);
695 Chain = P.getValue(1);
698 // If the source register was virtual and if we know something about it,
699 // add an assert node.
700 if (!TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) ||
701 !RegisterVT.isInteger() || RegisterVT.isVector())
704 const FunctionLoweringInfo::LiveOutInfo *LOI =
705 FuncInfo.GetLiveOutRegInfo(Regs[Part+i]);
709 unsigned RegSize = RegisterVT.getSizeInBits();
710 unsigned NumSignBits = LOI->NumSignBits;
711 unsigned NumZeroBits = LOI->KnownZero.countLeadingOnes();
713 if (NumZeroBits == RegSize) {
714 // The current value is a zero.
715 // Explicitly express that as it would be easier for
716 // optimizations to kick in.
717 Parts[i] = DAG.getConstant(0, RegisterVT);
721 // FIXME: We capture more information than the dag can represent. For
722 // now, just use the tightest assertzext/assertsext possible.
724 EVT FromVT(MVT::Other);
725 if (NumSignBits == RegSize)
726 isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1
727 else if (NumZeroBits >= RegSize-1)
728 isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1
729 else if (NumSignBits > RegSize-8)
730 isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8
731 else if (NumZeroBits >= RegSize-8)
732 isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8
733 else if (NumSignBits > RegSize-16)
734 isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16
735 else if (NumZeroBits >= RegSize-16)
736 isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16
737 else if (NumSignBits > RegSize-32)
738 isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32
739 else if (NumZeroBits >= RegSize-32)
740 isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32
744 // Add an assertion node.
745 assert(FromVT != MVT::Other);
746 Parts[i] = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl,
747 RegisterVT, P, DAG.getValueType(FromVT));
750 Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(),
751 NumRegs, RegisterVT, ValueVT, V);
756 return DAG.getNode(ISD::MERGE_VALUES, dl, DAG.getVTList(ValueVTs), Values);
759 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
760 /// specified value into the registers specified by this object. This uses
761 /// Chain/Flag as the input and updates them for the output Chain/Flag.
762 /// If the Flag pointer is NULL, no flag is used.
763 void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, SDLoc dl,
764 SDValue &Chain, SDValue *Flag,
765 const Value *V) const {
766 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
768 // Get the list of the values's legal parts.
769 unsigned NumRegs = Regs.size();
770 SmallVector<SDValue, 8> Parts(NumRegs);
771 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
772 EVT ValueVT = ValueVTs[Value];
773 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
774 MVT RegisterVT = RegVTs[Value];
775 ISD::NodeType ExtendKind =
776 TLI.isZExtFree(Val, RegisterVT)? ISD::ZERO_EXTEND: ISD::ANY_EXTEND;
778 getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value),
779 &Parts[Part], NumParts, RegisterVT, V, ExtendKind);
783 // Copy the parts into the registers.
784 SmallVector<SDValue, 8> Chains(NumRegs);
785 for (unsigned i = 0; i != NumRegs; ++i) {
788 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]);
790 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag);
791 *Flag = Part.getValue(1);
794 Chains[i] = Part.getValue(0);
797 if (NumRegs == 1 || Flag)
798 // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is
799 // flagged to it. That is the CopyToReg nodes and the user are considered
800 // a single scheduling unit. If we create a TokenFactor and return it as
801 // chain, then the TokenFactor is both a predecessor (operand) of the
802 // user as well as a successor (the TF operands are flagged to the user).
803 // c1, f1 = CopyToReg
804 // c2, f2 = CopyToReg
805 // c3 = TokenFactor c1, c2
808 Chain = Chains[NumRegs-1];
810 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
813 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
814 /// operand list. This adds the code marker and includes the number of
815 /// values added into it.
816 void RegsForValue::AddInlineAsmOperands(unsigned Code, bool HasMatching,
817 unsigned MatchingIdx,
819 std::vector<SDValue> &Ops) const {
820 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
822 unsigned Flag = InlineAsm::getFlagWord(Code, Regs.size());
824 Flag = InlineAsm::getFlagWordForMatchingOp(Flag, MatchingIdx);
825 else if (!Regs.empty() &&
826 TargetRegisterInfo::isVirtualRegister(Regs.front())) {
827 // Put the register class of the virtual registers in the flag word. That
828 // way, later passes can recompute register class constraints for inline
829 // assembly as well as normal instructions.
830 // Don't do this for tied operands that can use the regclass information
832 const MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
833 const TargetRegisterClass *RC = MRI.getRegClass(Regs.front());
834 Flag = InlineAsm::getFlagWordForRegClass(Flag, RC->getID());
837 SDValue Res = DAG.getTargetConstant(Flag, MVT::i32);
840 unsigned SP = TLI.getStackPointerRegisterToSaveRestore();
841 for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) {
842 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVTs[Value]);
843 MVT RegisterVT = RegVTs[Value];
844 for (unsigned i = 0; i != NumRegs; ++i) {
845 assert(Reg < Regs.size() && "Mismatch in # registers expected");
846 unsigned TheReg = Regs[Reg++];
847 Ops.push_back(DAG.getRegister(TheReg, RegisterVT));
849 if (TheReg == SP && Code == InlineAsm::Kind_Clobber) {
850 // If we clobbered the stack pointer, MFI should know about it.
851 assert(DAG.getMachineFunction().getFrameInfo()->
852 hasInlineAsmWithSPAdjust());
858 void SelectionDAGBuilder::init(GCFunctionInfo *gfi, AliasAnalysis &aa,
859 const TargetLibraryInfo *li) {
863 DL = DAG.getTarget().getDataLayout();
864 Context = DAG.getContext();
865 LPadToCallSiteMap.clear();
868 /// clear - Clear out the current SelectionDAG and the associated
869 /// state and prepare this SelectionDAGBuilder object to be used
870 /// for a new block. This doesn't clear out information about
871 /// additional blocks that are needed to complete switch lowering
872 /// or PHI node updating; that information is cleared out as it is
874 void SelectionDAGBuilder::clear() {
876 UnusedArgNodeMap.clear();
877 PendingLoads.clear();
878 PendingExports.clear();
881 SDNodeOrder = LowestSDNodeOrder;
884 /// clearDanglingDebugInfo - Clear the dangling debug information
885 /// map. This function is separated from the clear so that debug
886 /// information that is dangling in a basic block can be properly
887 /// resolved in a different basic block. This allows the
888 /// SelectionDAG to resolve dangling debug information attached
890 void SelectionDAGBuilder::clearDanglingDebugInfo() {
891 DanglingDebugInfoMap.clear();
894 /// getRoot - Return the current virtual root of the Selection DAG,
895 /// flushing any PendingLoad items. This must be done before emitting
896 /// a store or any other node that may need to be ordered after any
897 /// prior load instructions.
899 SDValue SelectionDAGBuilder::getRoot() {
900 if (PendingLoads.empty())
901 return DAG.getRoot();
903 if (PendingLoads.size() == 1) {
904 SDValue Root = PendingLoads[0];
906 PendingLoads.clear();
910 // Otherwise, we have to make a token factor node.
911 SDValue Root = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
913 PendingLoads.clear();
918 /// getControlRoot - Similar to getRoot, but instead of flushing all the
919 /// PendingLoad items, flush all the PendingExports items. It is necessary
920 /// to do this before emitting a terminator instruction.
922 SDValue SelectionDAGBuilder::getControlRoot() {
923 SDValue Root = DAG.getRoot();
925 if (PendingExports.empty())
928 // Turn all of the CopyToReg chains into one factored node.
929 if (Root.getOpcode() != ISD::EntryToken) {
930 unsigned i = 0, e = PendingExports.size();
931 for (; i != e; ++i) {
932 assert(PendingExports[i].getNode()->getNumOperands() > 1);
933 if (PendingExports[i].getNode()->getOperand(0) == Root)
934 break; // Don't add the root if we already indirectly depend on it.
938 PendingExports.push_back(Root);
941 Root = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
943 PendingExports.clear();
948 void SelectionDAGBuilder::visit(const Instruction &I) {
949 // Set up outgoing PHI node register values before emitting the terminator.
950 if (isa<TerminatorInst>(&I))
951 HandlePHINodesInSuccessorBlocks(I.getParent());
957 visit(I.getOpcode(), I);
959 if (!isa<TerminatorInst>(&I) && !HasTailCall)
960 CopyToExportRegsIfNeeded(&I);
965 void SelectionDAGBuilder::visitPHI(const PHINode &) {
966 llvm_unreachable("SelectionDAGBuilder shouldn't visit PHI nodes!");
969 void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) {
970 // Note: this doesn't use InstVisitor, because it has to work with
971 // ConstantExpr's in addition to instructions.
973 default: llvm_unreachable("Unknown instruction type encountered!");
974 // Build the switch statement using the Instruction.def file.
975 #define HANDLE_INST(NUM, OPCODE, CLASS) \
976 case Instruction::OPCODE: visit##OPCODE((const CLASS&)I); break;
977 #include "llvm/IR/Instruction.def"
981 // resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V,
982 // generate the debug data structures now that we've seen its definition.
983 void SelectionDAGBuilder::resolveDanglingDebugInfo(const Value *V,
985 DanglingDebugInfo &DDI = DanglingDebugInfoMap[V];
987 const DbgValueInst *DI = DDI.getDI();
988 DebugLoc dl = DDI.getdl();
989 unsigned DbgSDNodeOrder = DDI.getSDNodeOrder();
990 MDNode *Variable = DI->getVariable();
991 uint64_t Offset = DI->getOffset();
992 // A dbg.value for an alloca is always indirect.
993 bool IsIndirect = isa<AllocaInst>(V) || Offset != 0;
996 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, IsIndirect, Val)) {
997 SDV = DAG.getDbgValue(Variable, Val.getNode(),
998 Val.getResNo(), IsIndirect,
999 Offset, dl, DbgSDNodeOrder);
1000 DAG.AddDbgValue(SDV, Val.getNode(), false);
1003 DEBUG(dbgs() << "Dropping debug info for " << *DI << "\n");
1004 DanglingDebugInfoMap[V] = DanglingDebugInfo();
1008 /// getValue - Return an SDValue for the given Value.
1009 SDValue SelectionDAGBuilder::getValue(const Value *V) {
1010 // If we already have an SDValue for this value, use it. It's important
1011 // to do this first, so that we don't create a CopyFromReg if we already
1012 // have a regular SDValue.
1013 SDValue &N = NodeMap[V];
1014 if (N.getNode()) return N;
1016 // If there's a virtual register allocated and initialized for this
1018 DenseMap<const Value *, unsigned>::iterator It = FuncInfo.ValueMap.find(V);
1019 if (It != FuncInfo.ValueMap.end()) {
1020 unsigned InReg = It->second;
1021 RegsForValue RFV(*DAG.getContext(), *TM.getTargetLowering(),
1022 InReg, V->getType());
1023 SDValue Chain = DAG.getEntryNode();
1024 N = RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, nullptr, V);
1025 resolveDanglingDebugInfo(V, N);
1029 // Otherwise create a new SDValue and remember it.
1030 SDValue Val = getValueImpl(V);
1032 resolveDanglingDebugInfo(V, Val);
1036 /// getNonRegisterValue - Return an SDValue for the given Value, but
1037 /// don't look in FuncInfo.ValueMap for a virtual register.
1038 SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) {
1039 // If we already have an SDValue for this value, use it.
1040 SDValue &N = NodeMap[V];
1041 if (N.getNode()) return N;
1043 // Otherwise create a new SDValue and remember it.
1044 SDValue Val = getValueImpl(V);
1046 resolveDanglingDebugInfo(V, Val);
1050 /// getValueImpl - Helper function for getValue and getNonRegisterValue.
1051 /// Create an SDValue for the given value.
1052 SDValue SelectionDAGBuilder::getValueImpl(const Value *V) {
1053 const TargetLowering *TLI = TM.getTargetLowering();
1055 if (const Constant *C = dyn_cast<Constant>(V)) {
1056 EVT VT = TLI->getValueType(V->getType(), true);
1058 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C))
1059 return DAG.getConstant(*CI, VT);
1061 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
1062 return DAG.getGlobalAddress(GV, getCurSDLoc(), VT);
1064 if (isa<ConstantPointerNull>(C)) {
1065 unsigned AS = V->getType()->getPointerAddressSpace();
1066 return DAG.getConstant(0, TLI->getPointerTy(AS));
1069 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
1070 return DAG.getConstantFP(*CFP, VT);
1072 if (isa<UndefValue>(C) && !V->getType()->isAggregateType())
1073 return DAG.getUNDEF(VT);
1075 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
1076 visit(CE->getOpcode(), *CE);
1077 SDValue N1 = NodeMap[V];
1078 assert(N1.getNode() && "visit didn't populate the NodeMap!");
1082 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) {
1083 SmallVector<SDValue, 4> Constants;
1084 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end();
1086 SDNode *Val = getValue(*OI).getNode();
1087 // If the operand is an empty aggregate, there are no values.
1089 // Add each leaf value from the operand to the Constants list
1090 // to form a flattened list of all the values.
1091 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
1092 Constants.push_back(SDValue(Val, i));
1095 return DAG.getMergeValues(Constants, getCurSDLoc());
1098 if (const ConstantDataSequential *CDS =
1099 dyn_cast<ConstantDataSequential>(C)) {
1100 SmallVector<SDValue, 4> Ops;
1101 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1102 SDNode *Val = getValue(CDS->getElementAsConstant(i)).getNode();
1103 // Add each leaf value from the operand to the Constants list
1104 // to form a flattened list of all the values.
1105 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
1106 Ops.push_back(SDValue(Val, i));
1109 if (isa<ArrayType>(CDS->getType()))
1110 return DAG.getMergeValues(Ops, getCurSDLoc());
1111 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(),
1115 if (C->getType()->isStructTy() || C->getType()->isArrayTy()) {
1116 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
1117 "Unknown struct or array constant!");
1119 SmallVector<EVT, 4> ValueVTs;
1120 ComputeValueVTs(*TLI, C->getType(), ValueVTs);
1121 unsigned NumElts = ValueVTs.size();
1123 return SDValue(); // empty struct
1124 SmallVector<SDValue, 4> Constants(NumElts);
1125 for (unsigned i = 0; i != NumElts; ++i) {
1126 EVT EltVT = ValueVTs[i];
1127 if (isa<UndefValue>(C))
1128 Constants[i] = DAG.getUNDEF(EltVT);
1129 else if (EltVT.isFloatingPoint())
1130 Constants[i] = DAG.getConstantFP(0, EltVT);
1132 Constants[i] = DAG.getConstant(0, EltVT);
1135 return DAG.getMergeValues(Constants, getCurSDLoc());
1138 if (const BlockAddress *BA = dyn_cast<BlockAddress>(C))
1139 return DAG.getBlockAddress(BA, VT);
1141 VectorType *VecTy = cast<VectorType>(V->getType());
1142 unsigned NumElements = VecTy->getNumElements();
1144 // Now that we know the number and type of the elements, get that number of
1145 // elements into the Ops array based on what kind of constant it is.
1146 SmallVector<SDValue, 16> Ops;
1147 if (const ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
1148 for (unsigned i = 0; i != NumElements; ++i)
1149 Ops.push_back(getValue(CV->getOperand(i)));
1151 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!");
1152 EVT EltVT = TLI->getValueType(VecTy->getElementType());
1155 if (EltVT.isFloatingPoint())
1156 Op = DAG.getConstantFP(0, EltVT);
1158 Op = DAG.getConstant(0, EltVT);
1159 Ops.assign(NumElements, Op);
1162 // Create a BUILD_VECTOR node.
1163 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(), VT, Ops);
1166 // If this is a static alloca, generate it as the frameindex instead of
1168 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
1169 DenseMap<const AllocaInst*, int>::iterator SI =
1170 FuncInfo.StaticAllocaMap.find(AI);
1171 if (SI != FuncInfo.StaticAllocaMap.end())
1172 return DAG.getFrameIndex(SI->second, TLI->getPointerTy());
1175 // If this is an instruction which fast-isel has deferred, select it now.
1176 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
1177 unsigned InReg = FuncInfo.InitializeRegForValue(Inst);
1178 RegsForValue RFV(*DAG.getContext(), *TLI, InReg, Inst->getType());
1179 SDValue Chain = DAG.getEntryNode();
1180 return RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, nullptr, V);
1183 llvm_unreachable("Can't get register for value!");
1186 void SelectionDAGBuilder::visitRet(const ReturnInst &I) {
1187 const TargetLowering *TLI = TM.getTargetLowering();
1188 SDValue Chain = getControlRoot();
1189 SmallVector<ISD::OutputArg, 8> Outs;
1190 SmallVector<SDValue, 8> OutVals;
1192 if (!FuncInfo.CanLowerReturn) {
1193 unsigned DemoteReg = FuncInfo.DemoteRegister;
1194 const Function *F = I.getParent()->getParent();
1196 // Emit a store of the return value through the virtual register.
1197 // Leave Outs empty so that LowerReturn won't try to load return
1198 // registers the usual way.
1199 SmallVector<EVT, 1> PtrValueVTs;
1200 ComputeValueVTs(*TLI, PointerType::getUnqual(F->getReturnType()),
1203 SDValue RetPtr = DAG.getRegister(DemoteReg, PtrValueVTs[0]);
1204 SDValue RetOp = getValue(I.getOperand(0));
1206 SmallVector<EVT, 4> ValueVTs;
1207 SmallVector<uint64_t, 4> Offsets;
1208 ComputeValueVTs(*TLI, I.getOperand(0)->getType(), ValueVTs, &Offsets);
1209 unsigned NumValues = ValueVTs.size();
1211 SmallVector<SDValue, 4> Chains(NumValues);
1212 for (unsigned i = 0; i != NumValues; ++i) {
1213 SDValue Add = DAG.getNode(ISD::ADD, getCurSDLoc(),
1214 RetPtr.getValueType(), RetPtr,
1215 DAG.getIntPtrConstant(Offsets[i]));
1217 DAG.getStore(Chain, getCurSDLoc(),
1218 SDValue(RetOp.getNode(), RetOp.getResNo() + i),
1219 // FIXME: better loc info would be nice.
1220 Add, MachinePointerInfo(), false, false, 0);
1223 Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(),
1224 MVT::Other, Chains);
1225 } else if (I.getNumOperands() != 0) {
1226 SmallVector<EVT, 4> ValueVTs;
1227 ComputeValueVTs(*TLI, I.getOperand(0)->getType(), ValueVTs);
1228 unsigned NumValues = ValueVTs.size();
1230 SDValue RetOp = getValue(I.getOperand(0));
1231 for (unsigned j = 0, f = NumValues; j != f; ++j) {
1232 EVT VT = ValueVTs[j];
1234 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1236 const Function *F = I.getParent()->getParent();
1237 if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1239 ExtendKind = ISD::SIGN_EXTEND;
1240 else if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1242 ExtendKind = ISD::ZERO_EXTEND;
1244 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger())
1245 VT = TLI->getTypeForExtArgOrReturn(VT.getSimpleVT(), ExtendKind);
1247 unsigned NumParts = TLI->getNumRegisters(*DAG.getContext(), VT);
1248 MVT PartVT = TLI->getRegisterType(*DAG.getContext(), VT);
1249 SmallVector<SDValue, 4> Parts(NumParts);
1250 getCopyToParts(DAG, getCurSDLoc(),
1251 SDValue(RetOp.getNode(), RetOp.getResNo() + j),
1252 &Parts[0], NumParts, PartVT, &I, ExtendKind);
1254 // 'inreg' on function refers to return value
1255 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1256 if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1260 // Propagate extension type if any
1261 if (ExtendKind == ISD::SIGN_EXTEND)
1263 else if (ExtendKind == ISD::ZERO_EXTEND)
1266 for (unsigned i = 0; i < NumParts; ++i) {
1267 Outs.push_back(ISD::OutputArg(Flags, Parts[i].getValueType(),
1268 VT, /*isfixed=*/true, 0, 0));
1269 OutVals.push_back(Parts[i]);
1275 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1276 CallingConv::ID CallConv =
1277 DAG.getMachineFunction().getFunction()->getCallingConv();
1278 Chain = TM.getTargetLowering()->LowerReturn(Chain, CallConv, isVarArg,
1279 Outs, OutVals, getCurSDLoc(),
1282 // Verify that the target's LowerReturn behaved as expected.
1283 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
1284 "LowerReturn didn't return a valid chain!");
1286 // Update the DAG with the new chain value resulting from return lowering.
1290 /// CopyToExportRegsIfNeeded - If the given value has virtual registers
1291 /// created for it, emit nodes to copy the value into the virtual
1293 void SelectionDAGBuilder::CopyToExportRegsIfNeeded(const Value *V) {
1295 if (V->getType()->isEmptyTy())
1298 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
1299 if (VMI != FuncInfo.ValueMap.end()) {
1300 assert(!V->use_empty() && "Unused value assigned virtual registers!");
1301 CopyValueToVirtualRegister(V, VMI->second);
1305 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
1306 /// the current basic block, add it to ValueMap now so that we'll get a
1308 void SelectionDAGBuilder::ExportFromCurrentBlock(const Value *V) {
1309 // No need to export constants.
1310 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
1312 // Already exported?
1313 if (FuncInfo.isExportedInst(V)) return;
1315 unsigned Reg = FuncInfo.InitializeRegForValue(V);
1316 CopyValueToVirtualRegister(V, Reg);
1319 bool SelectionDAGBuilder::isExportableFromCurrentBlock(const Value *V,
1320 const BasicBlock *FromBB) {
1321 // The operands of the setcc have to be in this block. We don't know
1322 // how to export them from some other block.
1323 if (const Instruction *VI = dyn_cast<Instruction>(V)) {
1324 // Can export from current BB.
1325 if (VI->getParent() == FromBB)
1328 // Is already exported, noop.
1329 return FuncInfo.isExportedInst(V);
1332 // If this is an argument, we can export it if the BB is the entry block or
1333 // if it is already exported.
1334 if (isa<Argument>(V)) {
1335 if (FromBB == &FromBB->getParent()->getEntryBlock())
1338 // Otherwise, can only export this if it is already exported.
1339 return FuncInfo.isExportedInst(V);
1342 // Otherwise, constants can always be exported.
1346 /// Return branch probability calculated by BranchProbabilityInfo for IR blocks.
1347 uint32_t SelectionDAGBuilder::getEdgeWeight(const MachineBasicBlock *Src,
1348 const MachineBasicBlock *Dst) const {
1349 BranchProbabilityInfo *BPI = FuncInfo.BPI;
1352 const BasicBlock *SrcBB = Src->getBasicBlock();
1353 const BasicBlock *DstBB = Dst->getBasicBlock();
1354 return BPI->getEdgeWeight(SrcBB, DstBB);
1357 void SelectionDAGBuilder::
1358 addSuccessorWithWeight(MachineBasicBlock *Src, MachineBasicBlock *Dst,
1359 uint32_t Weight /* = 0 */) {
1361 Weight = getEdgeWeight(Src, Dst);
1362 Src->addSuccessor(Dst, Weight);
1366 static bool InBlock(const Value *V, const BasicBlock *BB) {
1367 if (const Instruction *I = dyn_cast<Instruction>(V))
1368 return I->getParent() == BB;
1372 /// EmitBranchForMergedCondition - Helper method for FindMergedConditions.
1373 /// This function emits a branch and is used at the leaves of an OR or an
1374 /// AND operator tree.
1377 SelectionDAGBuilder::EmitBranchForMergedCondition(const Value *Cond,
1378 MachineBasicBlock *TBB,
1379 MachineBasicBlock *FBB,
1380 MachineBasicBlock *CurBB,
1381 MachineBasicBlock *SwitchBB,
1384 const BasicBlock *BB = CurBB->getBasicBlock();
1386 // If the leaf of the tree is a comparison, merge the condition into
1388 if (const CmpInst *BOp = dyn_cast<CmpInst>(Cond)) {
1389 // The operands of the cmp have to be in this block. We don't know
1390 // how to export them from some other block. If this is the first block
1391 // of the sequence, no exporting is needed.
1392 if (CurBB == SwitchBB ||
1393 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
1394 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) {
1395 ISD::CondCode Condition;
1396 if (const ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
1397 Condition = getICmpCondCode(IC->getPredicate());
1398 } else if (const FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) {
1399 Condition = getFCmpCondCode(FC->getPredicate());
1400 if (TM.Options.NoNaNsFPMath)
1401 Condition = getFCmpCodeWithoutNaN(Condition);
1403 Condition = ISD::SETEQ; // silence warning.
1404 llvm_unreachable("Unknown compare instruction");
1407 CaseBlock CB(Condition, BOp->getOperand(0), BOp->getOperand(1), nullptr,
1408 TBB, FBB, CurBB, TWeight, FWeight);
1409 SwitchCases.push_back(CB);
1414 // Create a CaseBlock record representing this branch.
1415 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()),
1416 nullptr, TBB, FBB, CurBB, TWeight, FWeight);
1417 SwitchCases.push_back(CB);
1420 /// Scale down both weights to fit into uint32_t.
1421 static void ScaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
1422 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
1423 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
1424 NewTrue = NewTrue / Scale;
1425 NewFalse = NewFalse / Scale;
1428 /// FindMergedConditions - If Cond is an expression like
1429 void SelectionDAGBuilder::FindMergedConditions(const Value *Cond,
1430 MachineBasicBlock *TBB,
1431 MachineBasicBlock *FBB,
1432 MachineBasicBlock *CurBB,
1433 MachineBasicBlock *SwitchBB,
1434 unsigned Opc, uint32_t TWeight,
1436 // If this node is not part of the or/and tree, emit it as a branch.
1437 const Instruction *BOp = dyn_cast<Instruction>(Cond);
1438 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
1439 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
1440 BOp->getParent() != CurBB->getBasicBlock() ||
1441 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
1442 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
1443 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB,
1448 // Create TmpBB after CurBB.
1449 MachineFunction::iterator BBI = CurBB;
1450 MachineFunction &MF = DAG.getMachineFunction();
1451 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock());
1452 CurBB->getParent()->insert(++BBI, TmpBB);
1454 if (Opc == Instruction::Or) {
1455 // Codegen X | Y as:
1464 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
1465 // The requirement is that
1466 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
1467 // = TrueProb for orignal BB.
1468 // Assuming the orignal weights are A and B, one choice is to set BB1's
1469 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
1471 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
1472 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
1473 // TmpBB, but the math is more complicated.
1475 uint64_t NewTrueWeight = TWeight;
1476 uint64_t NewFalseWeight = (uint64_t)TWeight + 2 * (uint64_t)FWeight;
1477 ScaleWeights(NewTrueWeight, NewFalseWeight);
1478 // Emit the LHS condition.
1479 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, SwitchBB, Opc,
1480 NewTrueWeight, NewFalseWeight);
1482 NewTrueWeight = TWeight;
1483 NewFalseWeight = 2 * (uint64_t)FWeight;
1484 ScaleWeights(NewTrueWeight, NewFalseWeight);
1485 // Emit the RHS condition into TmpBB.
1486 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc,
1487 NewTrueWeight, NewFalseWeight);
1489 assert(Opc == Instruction::And && "Unknown merge op!");
1490 // Codegen X & Y as:
1498 // This requires creation of TmpBB after CurBB.
1500 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
1501 // The requirement is that
1502 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
1503 // = FalseProb for orignal BB.
1504 // Assuming the orignal weights are A and B, one choice is to set BB1's
1505 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
1507 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
1509 uint64_t NewTrueWeight = 2 * (uint64_t)TWeight + (uint64_t)FWeight;
1510 uint64_t NewFalseWeight = FWeight;
1511 ScaleWeights(NewTrueWeight, NewFalseWeight);
1512 // Emit the LHS condition.
1513 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, SwitchBB, Opc,
1514 NewTrueWeight, NewFalseWeight);
1516 NewTrueWeight = 2 * (uint64_t)TWeight;
1517 NewFalseWeight = FWeight;
1518 ScaleWeights(NewTrueWeight, NewFalseWeight);
1519 // Emit the RHS condition into TmpBB.
1520 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc,
1521 NewTrueWeight, NewFalseWeight);
1525 /// If the set of cases should be emitted as a series of branches, return true.
1526 /// If we should emit this as a bunch of and/or'd together conditions, return
1529 SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases) {
1530 if (Cases.size() != 2) return true;
1532 // If this is two comparisons of the same values or'd or and'd together, they
1533 // will get folded into a single comparison, so don't emit two blocks.
1534 if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
1535 Cases[0].CmpRHS == Cases[1].CmpRHS) ||
1536 (Cases[0].CmpRHS == Cases[1].CmpLHS &&
1537 Cases[0].CmpLHS == Cases[1].CmpRHS)) {
1541 // Handle: (X != null) | (Y != null) --> (X|Y) != 0
1542 // Handle: (X == null) & (Y == null) --> (X|Y) == 0
1543 if (Cases[0].CmpRHS == Cases[1].CmpRHS &&
1544 Cases[0].CC == Cases[1].CC &&
1545 isa<Constant>(Cases[0].CmpRHS) &&
1546 cast<Constant>(Cases[0].CmpRHS)->isNullValue()) {
1547 if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB)
1549 if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB)
1556 void SelectionDAGBuilder::visitBr(const BranchInst &I) {
1557 MachineBasicBlock *BrMBB = FuncInfo.MBB;
1559 // Update machine-CFG edges.
1560 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
1562 // Figure out which block is immediately after the current one.
1563 MachineBasicBlock *NextBlock = nullptr;
1564 MachineFunction::iterator BBI = BrMBB;
1565 if (++BBI != FuncInfo.MF->end())
1568 if (I.isUnconditional()) {
1569 // Update machine-CFG edges.
1570 BrMBB->addSuccessor(Succ0MBB);
1572 // If this is not a fall-through branch or optimizations are switched off,
1574 if (Succ0MBB != NextBlock || TM.getOptLevel() == CodeGenOpt::None)
1575 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(),
1576 MVT::Other, getControlRoot(),
1577 DAG.getBasicBlock(Succ0MBB)));
1582 // If this condition is one of the special cases we handle, do special stuff
1584 const Value *CondVal = I.getCondition();
1585 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
1587 // If this is a series of conditions that are or'd or and'd together, emit
1588 // this as a sequence of branches instead of setcc's with and/or operations.
1589 // As long as jumps are not expensive, this should improve performance.
1590 // For example, instead of something like:
1603 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
1604 if (!TM.getTargetLowering()->isJumpExpensive() &&
1606 (BOp->getOpcode() == Instruction::And ||
1607 BOp->getOpcode() == Instruction::Or)) {
1608 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB,
1609 BOp->getOpcode(), getEdgeWeight(BrMBB, Succ0MBB),
1610 getEdgeWeight(BrMBB, Succ1MBB));
1611 // If the compares in later blocks need to use values not currently
1612 // exported from this block, export them now. This block should always
1613 // be the first entry.
1614 assert(SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!");
1616 // Allow some cases to be rejected.
1617 if (ShouldEmitAsBranches(SwitchCases)) {
1618 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1619 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1620 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1623 // Emit the branch for this block.
1624 visitSwitchCase(SwitchCases[0], BrMBB);
1625 SwitchCases.erase(SwitchCases.begin());
1629 // Okay, we decided not to do this, remove any inserted MBB's and clear
1631 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
1632 FuncInfo.MF->erase(SwitchCases[i].ThisBB);
1634 SwitchCases.clear();
1638 // Create a CaseBlock record representing this branch.
1639 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()),
1640 nullptr, Succ0MBB, Succ1MBB, BrMBB);
1642 // Use visitSwitchCase to actually insert the fast branch sequence for this
1644 visitSwitchCase(CB, BrMBB);
1647 /// visitSwitchCase - Emits the necessary code to represent a single node in
1648 /// the binary search tree resulting from lowering a switch instruction.
1649 void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB,
1650 MachineBasicBlock *SwitchBB) {
1652 SDValue CondLHS = getValue(CB.CmpLHS);
1653 SDLoc dl = getCurSDLoc();
1655 // Build the setcc now.
1657 // Fold "(X == true)" to X and "(X == false)" to !X to
1658 // handle common cases produced by branch lowering.
1659 if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) &&
1660 CB.CC == ISD::SETEQ)
1662 else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) &&
1663 CB.CC == ISD::SETEQ) {
1664 SDValue True = DAG.getConstant(1, CondLHS.getValueType());
1665 Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True);
1667 Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1669 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now");
1671 const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue();
1672 const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue();
1674 SDValue CmpOp = getValue(CB.CmpMHS);
1675 EVT VT = CmpOp.getValueType();
1677 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) {
1678 Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, VT),
1681 SDValue SUB = DAG.getNode(ISD::SUB, dl,
1682 VT, CmpOp, DAG.getConstant(Low, VT));
1683 Cond = DAG.getSetCC(dl, MVT::i1, SUB,
1684 DAG.getConstant(High-Low, VT), ISD::SETULE);
1688 // Update successor info
1689 addSuccessorWithWeight(SwitchBB, CB.TrueBB, CB.TrueWeight);
1690 // TrueBB and FalseBB are always different unless the incoming IR is
1691 // degenerate. This only happens when running llc on weird IR.
1692 if (CB.TrueBB != CB.FalseBB)
1693 addSuccessorWithWeight(SwitchBB, CB.FalseBB, CB.FalseWeight);
1695 // Set NextBlock to be the MBB immediately after the current one, if any.
1696 // This is used to avoid emitting unnecessary branches to the next block.
1697 MachineBasicBlock *NextBlock = nullptr;
1698 MachineFunction::iterator BBI = SwitchBB;
1699 if (++BBI != FuncInfo.MF->end())
1702 // If the lhs block is the next block, invert the condition so that we can
1703 // fall through to the lhs instead of the rhs block.
1704 if (CB.TrueBB == NextBlock) {
1705 std::swap(CB.TrueBB, CB.FalseBB);
1706 SDValue True = DAG.getConstant(1, Cond.getValueType());
1707 Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True);
1710 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1711 MVT::Other, getControlRoot(), Cond,
1712 DAG.getBasicBlock(CB.TrueBB));
1714 // Insert the false branch. Do this even if it's a fall through branch,
1715 // this makes it easier to do DAG optimizations which require inverting
1716 // the branch condition.
1717 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1718 DAG.getBasicBlock(CB.FalseBB));
1720 DAG.setRoot(BrCond);
1723 /// visitJumpTable - Emit JumpTable node in the current MBB
1724 void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) {
1725 // Emit the code for the jump table
1726 assert(JT.Reg != -1U && "Should lower JT Header first!");
1727 EVT PTy = TM.getTargetLowering()->getPointerTy();
1728 SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurSDLoc(),
1730 SDValue Table = DAG.getJumpTable(JT.JTI, PTy);
1731 SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurSDLoc(),
1732 MVT::Other, Index.getValue(1),
1734 DAG.setRoot(BrJumpTable);
1737 /// visitJumpTableHeader - This function emits necessary code to produce index
1738 /// in the JumpTable from switch case.
1739 void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT,
1740 JumpTableHeader &JTH,
1741 MachineBasicBlock *SwitchBB) {
1742 // Subtract the lowest switch case value from the value being switched on and
1743 // conditional branch to default mbb if the result is greater than the
1744 // difference between smallest and largest cases.
1745 SDValue SwitchOp = getValue(JTH.SValue);
1746 EVT VT = SwitchOp.getValueType();
1747 SDValue Sub = DAG.getNode(ISD::SUB, getCurSDLoc(), VT, SwitchOp,
1748 DAG.getConstant(JTH.First, VT));
1750 // The SDNode we just created, which holds the value being switched on minus
1751 // the smallest case value, needs to be copied to a virtual register so it
1752 // can be used as an index into the jump table in a subsequent basic block.
1753 // This value may be smaller or larger than the target's pointer type, and
1754 // therefore require extension or truncating.
1755 const TargetLowering *TLI = TM.getTargetLowering();
1756 SwitchOp = DAG.getZExtOrTrunc(Sub, getCurSDLoc(), TLI->getPointerTy());
1758 unsigned JumpTableReg = FuncInfo.CreateReg(TLI->getPointerTy());
1759 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurSDLoc(),
1760 JumpTableReg, SwitchOp);
1761 JT.Reg = JumpTableReg;
1763 // Emit the range check for the jump table, and branch to the default block
1764 // for the switch statement if the value being switched on exceeds the largest
1765 // case in the switch.
1766 SDValue CMP = DAG.getSetCC(getCurSDLoc(),
1767 TLI->getSetCCResultType(*DAG.getContext(),
1768 Sub.getValueType()),
1770 DAG.getConstant(JTH.Last - JTH.First,VT),
1773 // Set NextBlock to be the MBB immediately after the current one, if any.
1774 // This is used to avoid emitting unnecessary branches to the next block.
1775 MachineBasicBlock *NextBlock = nullptr;
1776 MachineFunction::iterator BBI = SwitchBB;
1778 if (++BBI != FuncInfo.MF->end())
1781 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurSDLoc(),
1782 MVT::Other, CopyTo, CMP,
1783 DAG.getBasicBlock(JT.Default));
1785 if (JT.MBB != NextBlock)
1786 BrCond = DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, BrCond,
1787 DAG.getBasicBlock(JT.MBB));
1789 DAG.setRoot(BrCond);
1792 /// Codegen a new tail for a stack protector check ParentMBB which has had its
1793 /// tail spliced into a stack protector check success bb.
1795 /// For a high level explanation of how this fits into the stack protector
1796 /// generation see the comment on the declaration of class
1797 /// StackProtectorDescriptor.
1798 void SelectionDAGBuilder::visitSPDescriptorParent(StackProtectorDescriptor &SPD,
1799 MachineBasicBlock *ParentBB) {
1801 // First create the loads to the guard/stack slot for the comparison.
1802 const TargetLowering *TLI = TM.getTargetLowering();
1803 EVT PtrTy = TLI->getPointerTy();
1805 MachineFrameInfo *MFI = ParentBB->getParent()->getFrameInfo();
1806 int FI = MFI->getStackProtectorIndex();
1808 const Value *IRGuard = SPD.getGuard();
1809 SDValue GuardPtr = getValue(IRGuard);
1810 SDValue StackSlotPtr = DAG.getFrameIndex(FI, PtrTy);
1813 TLI->getDataLayout()->getPrefTypeAlignment(IRGuard->getType());
1817 // If useLoadStackGuardNode returns true, retrieve the guard value from
1818 // the virtual register holding the value. Otherwise, emit a volatile load
1819 // to retrieve the stack guard value.
1820 if (TLI->useLoadStackGuardNode())
1821 Guard = DAG.getCopyFromReg(DAG.getEntryNode(), getCurSDLoc(),
1822 SPD.getGuardReg(), PtrTy);
1824 Guard = DAG.getLoad(PtrTy, getCurSDLoc(), DAG.getEntryNode(),
1825 GuardPtr, MachinePointerInfo(IRGuard, 0),
1826 true, false, false, Align);
1828 SDValue StackSlot = DAG.getLoad(PtrTy, getCurSDLoc(), DAG.getEntryNode(),
1830 MachinePointerInfo::getFixedStack(FI),
1831 true, false, false, Align);
1833 // Perform the comparison via a subtract/getsetcc.
1834 EVT VT = Guard.getValueType();
1835 SDValue Sub = DAG.getNode(ISD::SUB, getCurSDLoc(), VT, Guard, StackSlot);
1837 SDValue Cmp = DAG.getSetCC(getCurSDLoc(),
1838 TLI->getSetCCResultType(*DAG.getContext(),
1839 Sub.getValueType()),
1840 Sub, DAG.getConstant(0, VT),
1843 // If the sub is not 0, then we know the guard/stackslot do not equal, so
1844 // branch to failure MBB.
1845 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurSDLoc(),
1846 MVT::Other, StackSlot.getOperand(0),
1847 Cmp, DAG.getBasicBlock(SPD.getFailureMBB()));
1848 // Otherwise branch to success MBB.
1849 SDValue Br = DAG.getNode(ISD::BR, getCurSDLoc(),
1851 DAG.getBasicBlock(SPD.getSuccessMBB()));
1856 /// Codegen the failure basic block for a stack protector check.
1858 /// A failure stack protector machine basic block consists simply of a call to
1859 /// __stack_chk_fail().
1861 /// For a high level explanation of how this fits into the stack protector
1862 /// generation see the comment on the declaration of class
1863 /// StackProtectorDescriptor.
1865 SelectionDAGBuilder::visitSPDescriptorFailure(StackProtectorDescriptor &SPD) {
1866 const TargetLowering *TLI = TM.getTargetLowering();
1867 SDValue Chain = TLI->makeLibCall(DAG, RTLIB::STACKPROTECTOR_CHECK_FAIL,
1868 MVT::isVoid, nullptr, 0, false,
1869 getCurSDLoc(), false, false).second;
1873 /// visitBitTestHeader - This function emits necessary code to produce value
1874 /// suitable for "bit tests"
1875 void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B,
1876 MachineBasicBlock *SwitchBB) {
1877 // Subtract the minimum value
1878 SDValue SwitchOp = getValue(B.SValue);
1879 EVT VT = SwitchOp.getValueType();
1880 SDValue Sub = DAG.getNode(ISD::SUB, getCurSDLoc(), VT, SwitchOp,
1881 DAG.getConstant(B.First, VT));
1884 const TargetLowering *TLI = TM.getTargetLowering();
1885 SDValue RangeCmp = DAG.getSetCC(getCurSDLoc(),
1886 TLI->getSetCCResultType(*DAG.getContext(),
1887 Sub.getValueType()),
1888 Sub, DAG.getConstant(B.Range, VT),
1891 // Determine the type of the test operands.
1892 bool UsePtrType = false;
1893 if (!TLI->isTypeLegal(VT))
1896 for (unsigned i = 0, e = B.Cases.size(); i != e; ++i)
1897 if (!isUIntN(VT.getSizeInBits(), B.Cases[i].Mask)) {
1898 // Switch table case range are encoded into series of masks.
1899 // Just use pointer type, it's guaranteed to fit.
1905 VT = TLI->getPointerTy();
1906 Sub = DAG.getZExtOrTrunc(Sub, getCurSDLoc(), VT);
1909 B.RegVT = VT.getSimpleVT();
1910 B.Reg = FuncInfo.CreateReg(B.RegVT);
1911 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurSDLoc(),
1914 // Set NextBlock to be the MBB immediately after the current one, if any.
1915 // This is used to avoid emitting unnecessary branches to the next block.
1916 MachineBasicBlock *NextBlock = nullptr;
1917 MachineFunction::iterator BBI = SwitchBB;
1918 if (++BBI != FuncInfo.MF->end())
1921 MachineBasicBlock* MBB = B.Cases[0].ThisBB;
1923 addSuccessorWithWeight(SwitchBB, B.Default);
1924 addSuccessorWithWeight(SwitchBB, MBB);
1926 SDValue BrRange = DAG.getNode(ISD::BRCOND, getCurSDLoc(),
1927 MVT::Other, CopyTo, RangeCmp,
1928 DAG.getBasicBlock(B.Default));
1930 if (MBB != NextBlock)
1931 BrRange = DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, CopyTo,
1932 DAG.getBasicBlock(MBB));
1934 DAG.setRoot(BrRange);
1937 /// visitBitTestCase - this function produces one "bit test"
1938 void SelectionDAGBuilder::visitBitTestCase(BitTestBlock &BB,
1939 MachineBasicBlock* NextMBB,
1940 uint32_t BranchWeightToNext,
1943 MachineBasicBlock *SwitchBB) {
1945 SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurSDLoc(),
1948 unsigned PopCount = CountPopulation_64(B.Mask);
1949 const TargetLowering *TLI = TM.getTargetLowering();
1950 if (PopCount == 1) {
1951 // Testing for a single bit; just compare the shift count with what it
1952 // would need to be to shift a 1 bit in that position.
1953 Cmp = DAG.getSetCC(getCurSDLoc(),
1954 TLI->getSetCCResultType(*DAG.getContext(), VT),
1956 DAG.getConstant(countTrailingZeros(B.Mask), VT),
1958 } else if (PopCount == BB.Range) {
1959 // There is only one zero bit in the range, test for it directly.
1960 Cmp = DAG.getSetCC(getCurSDLoc(),
1961 TLI->getSetCCResultType(*DAG.getContext(), VT),
1963 DAG.getConstant(CountTrailingOnes_64(B.Mask), VT),
1966 // Make desired shift
1967 SDValue SwitchVal = DAG.getNode(ISD::SHL, getCurSDLoc(), VT,
1968 DAG.getConstant(1, VT), ShiftOp);
1970 // Emit bit tests and jumps
1971 SDValue AndOp = DAG.getNode(ISD::AND, getCurSDLoc(),
1972 VT, SwitchVal, DAG.getConstant(B.Mask, VT));
1973 Cmp = DAG.getSetCC(getCurSDLoc(),
1974 TLI->getSetCCResultType(*DAG.getContext(), VT),
1975 AndOp, DAG.getConstant(0, VT),
1979 // The branch weight from SwitchBB to B.TargetBB is B.ExtraWeight.
1980 addSuccessorWithWeight(SwitchBB, B.TargetBB, B.ExtraWeight);
1981 // The branch weight from SwitchBB to NextMBB is BranchWeightToNext.
1982 addSuccessorWithWeight(SwitchBB, NextMBB, BranchWeightToNext);
1984 SDValue BrAnd = DAG.getNode(ISD::BRCOND, getCurSDLoc(),
1985 MVT::Other, getControlRoot(),
1986 Cmp, DAG.getBasicBlock(B.TargetBB));
1988 // Set NextBlock to be the MBB immediately after the current one, if any.
1989 // This is used to avoid emitting unnecessary branches to the next block.
1990 MachineBasicBlock *NextBlock = nullptr;
1991 MachineFunction::iterator BBI = SwitchBB;
1992 if (++BBI != FuncInfo.MF->end())
1995 if (NextMBB != NextBlock)
1996 BrAnd = DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, BrAnd,
1997 DAG.getBasicBlock(NextMBB));
2002 void SelectionDAGBuilder::visitInvoke(const InvokeInst &I) {
2003 MachineBasicBlock *InvokeMBB = FuncInfo.MBB;
2005 // Retrieve successors.
2006 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
2007 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)];
2009 const Value *Callee(I.getCalledValue());
2010 const Function *Fn = dyn_cast<Function>(Callee);
2011 if (isa<InlineAsm>(Callee))
2013 else if (Fn && Fn->isIntrinsic()) {
2014 assert(Fn->getIntrinsicID() == Intrinsic::donothing);
2015 // Ignore invokes to @llvm.donothing: jump directly to the next BB.
2017 LowerCallTo(&I, getValue(Callee), false, LandingPad);
2019 // If the value of the invoke is used outside of its defining block, make it
2020 // available as a virtual register.
2021 CopyToExportRegsIfNeeded(&I);
2023 // Update successor info
2024 addSuccessorWithWeight(InvokeMBB, Return);
2025 addSuccessorWithWeight(InvokeMBB, LandingPad);
2027 // Drop into normal successor.
2028 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(),
2029 MVT::Other, getControlRoot(),
2030 DAG.getBasicBlock(Return)));
2033 void SelectionDAGBuilder::visitResume(const ResumeInst &RI) {
2034 llvm_unreachable("SelectionDAGBuilder shouldn't visit resume instructions!");
2037 void SelectionDAGBuilder::visitLandingPad(const LandingPadInst &LP) {
2038 assert(FuncInfo.MBB->isLandingPad() &&
2039 "Call to landingpad not in landing pad!");
2041 MachineBasicBlock *MBB = FuncInfo.MBB;
2042 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
2043 AddLandingPadInfo(LP, MMI, MBB);
2045 // If there aren't registers to copy the values into (e.g., during SjLj
2046 // exceptions), then don't bother to create these DAG nodes.
2047 const TargetLowering *TLI = TM.getTargetLowering();
2048 if (TLI->getExceptionPointerRegister() == 0 &&
2049 TLI->getExceptionSelectorRegister() == 0)
2052 SmallVector<EVT, 2> ValueVTs;
2053 ComputeValueVTs(*TLI, LP.getType(), ValueVTs);
2054 assert(ValueVTs.size() == 2 && "Only two-valued landingpads are supported");
2056 // Get the two live-in registers as SDValues. The physregs have already been
2057 // copied into virtual registers.
2059 Ops[0] = DAG.getZExtOrTrunc(
2060 DAG.getCopyFromReg(DAG.getEntryNode(), getCurSDLoc(),
2061 FuncInfo.ExceptionPointerVirtReg, TLI->getPointerTy()),
2062 getCurSDLoc(), ValueVTs[0]);
2063 Ops[1] = DAG.getZExtOrTrunc(
2064 DAG.getCopyFromReg(DAG.getEntryNode(), getCurSDLoc(),
2065 FuncInfo.ExceptionSelectorVirtReg, TLI->getPointerTy()),
2066 getCurSDLoc(), ValueVTs[1]);
2069 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
2070 DAG.getVTList(ValueVTs), Ops);
2074 /// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for
2075 /// small case ranges).
2076 bool SelectionDAGBuilder::handleSmallSwitchRange(CaseRec& CR,
2077 CaseRecVector& WorkList,
2079 MachineBasicBlock *Default,
2080 MachineBasicBlock *SwitchBB) {
2081 // Size is the number of Cases represented by this range.
2082 size_t Size = CR.Range.second - CR.Range.first;
2086 // Get the MachineFunction which holds the current MBB. This is used when
2087 // inserting any additional MBBs necessary to represent the switch.
2088 MachineFunction *CurMF = FuncInfo.MF;
2090 // Figure out which block is immediately after the current one.
2091 MachineBasicBlock *NextBlock = nullptr;
2092 MachineFunction::iterator BBI = CR.CaseBB;
2094 if (++BBI != FuncInfo.MF->end())
2097 BranchProbabilityInfo *BPI = FuncInfo.BPI;
2098 // If any two of the cases has the same destination, and if one value
2099 // is the same as the other, but has one bit unset that the other has set,
2100 // use bit manipulation to do two compares at once. For example:
2101 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
2102 // TODO: This could be extended to merge any 2 cases in switches with 3 cases.
2103 // TODO: Handle cases where CR.CaseBB != SwitchBB.
2104 if (Size == 2 && CR.CaseBB == SwitchBB) {
2105 Case &Small = *CR.Range.first;
2106 Case &Big = *(CR.Range.second-1);
2108 if (Small.Low == Small.High && Big.Low == Big.High && Small.BB == Big.BB) {
2109 const APInt& SmallValue = cast<ConstantInt>(Small.Low)->getValue();
2110 const APInt& BigValue = cast<ConstantInt>(Big.Low)->getValue();
2112 // Check that there is only one bit different.
2113 if (BigValue.countPopulation() == SmallValue.countPopulation() + 1 &&
2114 (SmallValue | BigValue) == BigValue) {
2115 // Isolate the common bit.
2116 APInt CommonBit = BigValue & ~SmallValue;
2117 assert((SmallValue | CommonBit) == BigValue &&
2118 CommonBit.countPopulation() == 1 && "Not a common bit?");
2120 SDValue CondLHS = getValue(SV);
2121 EVT VT = CondLHS.getValueType();
2122 SDLoc DL = getCurSDLoc();
2124 SDValue Or = DAG.getNode(ISD::OR, DL, VT, CondLHS,
2125 DAG.getConstant(CommonBit, VT));
2126 SDValue Cond = DAG.getSetCC(DL, MVT::i1,
2127 Or, DAG.getConstant(BigValue, VT),
2130 // Update successor info.
2131 // Both Small and Big will jump to Small.BB, so we sum up the weights.
2132 addSuccessorWithWeight(SwitchBB, Small.BB,
2133 Small.ExtraWeight + Big.ExtraWeight);
2134 addSuccessorWithWeight(SwitchBB, Default,
2135 // The default destination is the first successor in IR.
2136 BPI ? BPI->getEdgeWeight(SwitchBB->getBasicBlock(), (unsigned)0) : 0);
2138 // Insert the true branch.
2139 SDValue BrCond = DAG.getNode(ISD::BRCOND, DL, MVT::Other,
2140 getControlRoot(), Cond,
2141 DAG.getBasicBlock(Small.BB));
2143 // Insert the false branch.
2144 BrCond = DAG.getNode(ISD::BR, DL, MVT::Other, BrCond,
2145 DAG.getBasicBlock(Default));
2147 DAG.setRoot(BrCond);
2153 // Order cases by weight so the most likely case will be checked first.
2154 uint32_t UnhandledWeights = 0;
2156 for (CaseItr I = CR.Range.first, IE = CR.Range.second; I != IE; ++I) {
2157 uint32_t IWeight = I->ExtraWeight;
2158 UnhandledWeights += IWeight;
2159 for (CaseItr J = CR.Range.first; J < I; ++J) {
2160 uint32_t JWeight = J->ExtraWeight;
2161 if (IWeight > JWeight)
2166 // Rearrange the case blocks so that the last one falls through if possible.
2167 Case &BackCase = *(CR.Range.second-1);
2169 NextBlock && Default != NextBlock && BackCase.BB != NextBlock) {
2170 // The last case block won't fall through into 'NextBlock' if we emit the
2171 // branches in this order. See if rearranging a case value would help.
2172 // We start at the bottom as it's the case with the least weight.
2173 for (Case *I = &*(CR.Range.second-2), *E = &*CR.Range.first-1; I != E; --I)
2174 if (I->BB == NextBlock) {
2175 std::swap(*I, BackCase);
2180 // Create a CaseBlock record representing a conditional branch to
2181 // the Case's target mbb if the value being switched on SV is equal
2183 MachineBasicBlock *CurBlock = CR.CaseBB;
2184 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
2185 MachineBasicBlock *FallThrough;
2187 FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock());
2188 CurMF->insert(BBI, FallThrough);
2190 // Put SV in a virtual register to make it available from the new blocks.
2191 ExportFromCurrentBlock(SV);
2193 // If the last case doesn't match, go to the default block.
2194 FallThrough = Default;
2197 const Value *RHS, *LHS, *MHS;
2199 if (I->High == I->Low) {
2200 // This is just small small case range :) containing exactly 1 case
2202 LHS = SV; RHS = I->High; MHS = nullptr;
2205 LHS = I->Low; MHS = SV; RHS = I->High;
2208 // The false weight should be sum of all un-handled cases.
2209 UnhandledWeights -= I->ExtraWeight;
2210 CaseBlock CB(CC, LHS, RHS, MHS, /* truebb */ I->BB, /* falsebb */ FallThrough,
2212 /* trueweight */ I->ExtraWeight,
2213 /* falseweight */ UnhandledWeights);
2215 // If emitting the first comparison, just call visitSwitchCase to emit the
2216 // code into the current block. Otherwise, push the CaseBlock onto the
2217 // vector to be later processed by SDISel, and insert the node's MBB
2218 // before the next MBB.
2219 if (CurBlock == SwitchBB)
2220 visitSwitchCase(CB, SwitchBB);
2222 SwitchCases.push_back(CB);
2224 CurBlock = FallThrough;
2230 static inline bool areJTsAllowed(const TargetLowering &TLI) {
2231 return TLI.supportJumpTables() &&
2232 (TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
2233 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other));
2236 static APInt ComputeRange(const APInt &First, const APInt &Last) {
2237 uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1;
2238 APInt LastExt = Last.sext(BitWidth), FirstExt = First.sext(BitWidth);
2239 return (LastExt - FirstExt + 1ULL);
2242 /// handleJTSwitchCase - Emit jumptable for current switch case range
2243 bool SelectionDAGBuilder::handleJTSwitchCase(CaseRec &CR,
2244 CaseRecVector &WorkList,
2246 MachineBasicBlock *Default,
2247 MachineBasicBlock *SwitchBB) {
2248 Case& FrontCase = *CR.Range.first;
2249 Case& BackCase = *(CR.Range.second-1);
2251 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
2252 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
2254 APInt TSize(First.getBitWidth(), 0);
2255 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I)
2258 const TargetLowering *TLI = TM.getTargetLowering();
2259 if (!areJTsAllowed(*TLI) || TSize.ult(TLI->getMinimumJumpTableEntries()))
2262 APInt Range = ComputeRange(First, Last);
2263 // The density is TSize / Range. Require at least 40%.
2264 // It should not be possible for IntTSize to saturate for sane code, but make
2265 // sure we handle Range saturation correctly.
2266 uint64_t IntRange = Range.getLimitedValue(UINT64_MAX/10);
2267 uint64_t IntTSize = TSize.getLimitedValue(UINT64_MAX/10);
2268 if (IntTSize * 10 < IntRange * 4)
2271 DEBUG(dbgs() << "Lowering jump table\n"
2272 << "First entry: " << First << ". Last entry: " << Last << '\n'
2273 << "Range: " << Range << ". Size: " << TSize << ".\n\n");
2275 // Get the MachineFunction which holds the current MBB. This is used when
2276 // inserting any additional MBBs necessary to represent the switch.
2277 MachineFunction *CurMF = FuncInfo.MF;
2279 // Figure out which block is immediately after the current one.
2280 MachineFunction::iterator BBI = CR.CaseBB;
2283 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2285 // Create a new basic block to hold the code for loading the address
2286 // of the jump table, and jumping to it. Update successor information;
2287 // we will either branch to the default case for the switch, or the jump
2289 MachineBasicBlock *JumpTableBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2290 CurMF->insert(BBI, JumpTableBB);
2292 addSuccessorWithWeight(CR.CaseBB, Default);
2293 addSuccessorWithWeight(CR.CaseBB, JumpTableBB);
2295 // Build a vector of destination BBs, corresponding to each target
2296 // of the jump table. If the value of the jump table slot corresponds to
2297 // a case statement, push the case's BB onto the vector, otherwise, push
2299 std::vector<MachineBasicBlock*> DestBBs;
2301 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) {
2302 const APInt &Low = cast<ConstantInt>(I->Low)->getValue();
2303 const APInt &High = cast<ConstantInt>(I->High)->getValue();
2305 if (Low.sle(TEI) && TEI.sle(High)) {
2306 DestBBs.push_back(I->BB);
2310 DestBBs.push_back(Default);
2314 // Calculate weight for each unique destination in CR.
2315 DenseMap<MachineBasicBlock*, uint32_t> DestWeights;
2317 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
2318 DenseMap<MachineBasicBlock*, uint32_t>::iterator Itr =
2319 DestWeights.find(I->BB);
2320 if (Itr != DestWeights.end())
2321 Itr->second += I->ExtraWeight;
2323 DestWeights[I->BB] = I->ExtraWeight;
2326 // Update successor info. Add one edge to each unique successor.
2327 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs());
2328 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(),
2329 E = DestBBs.end(); I != E; ++I) {
2330 if (!SuccsHandled[(*I)->getNumber()]) {
2331 SuccsHandled[(*I)->getNumber()] = true;
2332 DenseMap<MachineBasicBlock*, uint32_t>::iterator Itr =
2333 DestWeights.find(*I);
2334 addSuccessorWithWeight(JumpTableBB, *I,
2335 Itr != DestWeights.end() ? Itr->second : 0);
2339 // Create a jump table index for this jump table.
2340 unsigned JTEncoding = TLI->getJumpTableEncoding();
2341 unsigned JTI = CurMF->getOrCreateJumpTableInfo(JTEncoding)
2342 ->createJumpTableIndex(DestBBs);
2344 // Set the jump table information so that we can codegen it as a second
2345 // MachineBasicBlock
2346 JumpTable JT(-1U, JTI, JumpTableBB, Default);
2347 JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == SwitchBB));
2348 if (CR.CaseBB == SwitchBB)
2349 visitJumpTableHeader(JT, JTH, SwitchBB);
2351 JTCases.push_back(JumpTableBlock(JTH, JT));
2355 /// handleBTSplitSwitchCase - emit comparison and split binary search tree into
2357 bool SelectionDAGBuilder::handleBTSplitSwitchCase(CaseRec& CR,
2358 CaseRecVector& WorkList,
2360 MachineBasicBlock* Default,
2361 MachineBasicBlock* SwitchBB) {
2362 // Get the MachineFunction which holds the current MBB. This is used when
2363 // inserting any additional MBBs necessary to represent the switch.
2364 MachineFunction *CurMF = FuncInfo.MF;
2366 // Figure out which block is immediately after the current one.
2367 MachineFunction::iterator BBI = CR.CaseBB;
2370 Case& FrontCase = *CR.Range.first;
2371 Case& BackCase = *(CR.Range.second-1);
2372 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2374 // Size is the number of Cases represented by this range.
2375 unsigned Size = CR.Range.second - CR.Range.first;
2377 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
2378 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
2380 CaseItr Pivot = CR.Range.first + Size/2;
2382 // Select optimal pivot, maximizing sum density of LHS and RHS. This will
2383 // (heuristically) allow us to emit JumpTable's later.
2384 APInt TSize(First.getBitWidth(), 0);
2385 for (CaseItr I = CR.Range.first, E = CR.Range.second;
2389 APInt LSize = FrontCase.size();
2390 APInt RSize = TSize-LSize;
2391 DEBUG(dbgs() << "Selecting best pivot: \n"
2392 << "First: " << First << ", Last: " << Last <<'\n'
2393 << "LSize: " << LSize << ", RSize: " << RSize << '\n');
2394 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second;
2396 const APInt &LEnd = cast<ConstantInt>(I->High)->getValue();
2397 const APInt &RBegin = cast<ConstantInt>(J->Low)->getValue();
2398 APInt Range = ComputeRange(LEnd, RBegin);
2399 assert((Range - 2ULL).isNonNegative() &&
2400 "Invalid case distance");
2401 // Use volatile double here to avoid excess precision issues on some hosts,
2402 // e.g. that use 80-bit X87 registers.
2403 volatile double LDensity =
2404 (double)LSize.roundToDouble() /
2405 (LEnd - First + 1ULL).roundToDouble();
2406 volatile double RDensity =
2407 (double)RSize.roundToDouble() /
2408 (Last - RBegin + 1ULL).roundToDouble();
2409 volatile double Metric = Range.logBase2()*(LDensity+RDensity);
2410 // Should always split in some non-trivial place
2411 DEBUG(dbgs() <<"=>Step\n"
2412 << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n'
2413 << "LDensity: " << LDensity
2414 << ", RDensity: " << RDensity << '\n'
2415 << "Metric: " << Metric << '\n');
2416 if (FMetric < Metric) {
2419 DEBUG(dbgs() << "Current metric set to: " << FMetric << '\n');
2426 const TargetLowering *TLI = TM.getTargetLowering();
2427 if (areJTsAllowed(*TLI)) {
2428 // If our case is dense we *really* should handle it earlier!
2429 assert((FMetric > 0) && "Should handle dense range earlier!");
2431 Pivot = CR.Range.first + Size/2;
2434 CaseRange LHSR(CR.Range.first, Pivot);
2435 CaseRange RHSR(Pivot, CR.Range.second);
2436 const Constant *C = Pivot->Low;
2437 MachineBasicBlock *FalseBB = nullptr, *TrueBB = nullptr;
2439 // We know that we branch to the LHS if the Value being switched on is
2440 // less than the Pivot value, C. We use this to optimize our binary
2441 // tree a bit, by recognizing that if SV is greater than or equal to the
2442 // LHS's Case Value, and that Case Value is exactly one less than the
2443 // Pivot's Value, then we can branch directly to the LHS's Target,
2444 // rather than creating a leaf node for it.
2445 if ((LHSR.second - LHSR.first) == 1 &&
2446 LHSR.first->High == CR.GE &&
2447 cast<ConstantInt>(C)->getValue() ==
2448 (cast<ConstantInt>(CR.GE)->getValue() + 1LL)) {
2449 TrueBB = LHSR.first->BB;
2451 TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2452 CurMF->insert(BBI, TrueBB);
2453 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR));
2455 // Put SV in a virtual register to make it available from the new blocks.
2456 ExportFromCurrentBlock(SV);
2459 // Similar to the optimization above, if the Value being switched on is
2460 // known to be less than the Constant CR.LT, and the current Case Value
2461 // is CR.LT - 1, then we can branch directly to the target block for
2462 // the current Case Value, rather than emitting a RHS leaf node for it.
2463 if ((RHSR.second - RHSR.first) == 1 && CR.LT &&
2464 cast<ConstantInt>(RHSR.first->Low)->getValue() ==
2465 (cast<ConstantInt>(CR.LT)->getValue() - 1LL)) {
2466 FalseBB = RHSR.first->BB;
2468 FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2469 CurMF->insert(BBI, FalseBB);
2470 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR));
2472 // Put SV in a virtual register to make it available from the new blocks.
2473 ExportFromCurrentBlock(SV);
2476 // Create a CaseBlock record representing a conditional branch to
2477 // the LHS node if the value being switched on SV is less than C.
2478 // Otherwise, branch to LHS.
2479 CaseBlock CB(ISD::SETLT, SV, C, nullptr, TrueBB, FalseBB, CR.CaseBB);
2481 if (CR.CaseBB == SwitchBB)
2482 visitSwitchCase(CB, SwitchBB);
2484 SwitchCases.push_back(CB);
2489 /// handleBitTestsSwitchCase - if current case range has few destination and
2490 /// range span less, than machine word bitwidth, encode case range into series
2491 /// of masks and emit bit tests with these masks.
2492 bool SelectionDAGBuilder::handleBitTestsSwitchCase(CaseRec& CR,
2493 CaseRecVector& WorkList,
2495 MachineBasicBlock* Default,
2496 MachineBasicBlock* SwitchBB) {
2497 const TargetLowering *TLI = TM.getTargetLowering();
2498 EVT PTy = TLI->getPointerTy();
2499 unsigned IntPtrBits = PTy.getSizeInBits();
2501 Case& FrontCase = *CR.Range.first;
2502 Case& BackCase = *(CR.Range.second-1);
2504 // Get the MachineFunction which holds the current MBB. This is used when
2505 // inserting any additional MBBs necessary to represent the switch.
2506 MachineFunction *CurMF = FuncInfo.MF;
2508 // If target does not have legal shift left, do not emit bit tests at all.
2509 if (!TLI->isOperationLegal(ISD::SHL, PTy))
2513 for (CaseItr I = CR.Range.first, E = CR.Range.second;
2515 // Single case counts one, case range - two.
2516 numCmps += (I->Low == I->High ? 1 : 2);
2519 // Count unique destinations
2520 SmallSet<MachineBasicBlock*, 4> Dests;
2521 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2522 Dests.insert(I->BB);
2523 if (Dests.size() > 3)
2524 // Don't bother the code below, if there are too much unique destinations
2527 DEBUG(dbgs() << "Total number of unique destinations: "
2528 << Dests.size() << '\n'
2529 << "Total number of comparisons: " << numCmps << '\n');
2531 // Compute span of values.
2532 const APInt& minValue = cast<ConstantInt>(FrontCase.Low)->getValue();
2533 const APInt& maxValue = cast<ConstantInt>(BackCase.High)->getValue();
2534 APInt cmpRange = maxValue - minValue;
2536 DEBUG(dbgs() << "Compare range: " << cmpRange << '\n'
2537 << "Low bound: " << minValue << '\n'
2538 << "High bound: " << maxValue << '\n');
2540 if (cmpRange.uge(IntPtrBits) ||
2541 (!(Dests.size() == 1 && numCmps >= 3) &&
2542 !(Dests.size() == 2 && numCmps >= 5) &&
2543 !(Dests.size() >= 3 && numCmps >= 6)))
2546 DEBUG(dbgs() << "Emitting bit tests\n");
2547 APInt lowBound = APInt::getNullValue(cmpRange.getBitWidth());
2549 // Optimize the case where all the case values fit in a
2550 // word without having to subtract minValue. In this case,
2551 // we can optimize away the subtraction.
2552 if (minValue.isNonNegative() && maxValue.slt(IntPtrBits)) {
2553 cmpRange = maxValue;
2555 lowBound = minValue;
2558 CaseBitsVector CasesBits;
2559 unsigned i, count = 0;
2561 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2562 MachineBasicBlock* Dest = I->BB;
2563 for (i = 0; i < count; ++i)
2564 if (Dest == CasesBits[i].BB)
2568 assert((count < 3) && "Too much destinations to test!");
2569 CasesBits.push_back(CaseBits(0, Dest, 0, 0/*Weight*/));
2573 const APInt& lowValue = cast<ConstantInt>(I->Low)->getValue();
2574 const APInt& highValue = cast<ConstantInt>(I->High)->getValue();
2576 uint64_t lo = (lowValue - lowBound).getZExtValue();
2577 uint64_t hi = (highValue - lowBound).getZExtValue();
2578 CasesBits[i].ExtraWeight += I->ExtraWeight;
2580 for (uint64_t j = lo; j <= hi; j++) {
2581 CasesBits[i].Mask |= 1ULL << j;
2582 CasesBits[i].Bits++;
2586 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp());
2590 // Figure out which block is immediately after the current one.
2591 MachineFunction::iterator BBI = CR.CaseBB;
2594 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2596 DEBUG(dbgs() << "Cases:\n");
2597 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) {
2598 DEBUG(dbgs() << "Mask: " << CasesBits[i].Mask
2599 << ", Bits: " << CasesBits[i].Bits
2600 << ", BB: " << CasesBits[i].BB << '\n');
2602 MachineBasicBlock *CaseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2603 CurMF->insert(BBI, CaseBB);
2604 BTC.push_back(BitTestCase(CasesBits[i].Mask,
2606 CasesBits[i].BB, CasesBits[i].ExtraWeight));
2608 // Put SV in a virtual register to make it available from the new blocks.
2609 ExportFromCurrentBlock(SV);
2612 BitTestBlock BTB(lowBound, cmpRange, SV,
2613 -1U, MVT::Other, (CR.CaseBB == SwitchBB),
2614 CR.CaseBB, Default, BTC);
2616 if (CR.CaseBB == SwitchBB)
2617 visitBitTestHeader(BTB, SwitchBB);
2619 BitTestCases.push_back(BTB);
2624 /// Clusterify - Transform simple list of Cases into list of CaseRange's
2625 size_t SelectionDAGBuilder::Clusterify(CaseVector& Cases,
2626 const SwitchInst& SI) {
2629 BranchProbabilityInfo *BPI = FuncInfo.BPI;
2630 // Start with "simple" cases
2631 for (SwitchInst::ConstCaseIt i = SI.case_begin(), e = SI.case_end();
2633 const BasicBlock *SuccBB = i.getCaseSuccessor();
2634 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SuccBB];
2636 uint32_t ExtraWeight =
2637 BPI ? BPI->getEdgeWeight(SI.getParent(), i.getSuccessorIndex()) : 0;
2639 Cases.push_back(Case(i.getCaseValue(), i.getCaseValue(),
2640 SMBB, ExtraWeight));
2642 std::sort(Cases.begin(), Cases.end(), CaseCmp());
2644 // Merge case into clusters
2645 if (Cases.size() >= 2)
2646 // Must recompute end() each iteration because it may be
2647 // invalidated by erase if we hold on to it
2648 for (CaseItr I = Cases.begin(), J = std::next(Cases.begin());
2649 J != Cases.end(); ) {
2650 const APInt& nextValue = cast<ConstantInt>(J->Low)->getValue();
2651 const APInt& currentValue = cast<ConstantInt>(I->High)->getValue();
2652 MachineBasicBlock* nextBB = J->BB;
2653 MachineBasicBlock* currentBB = I->BB;
2655 // If the two neighboring cases go to the same destination, merge them
2656 // into a single case.
2657 if ((nextValue - currentValue == 1) && (currentBB == nextBB)) {
2659 I->ExtraWeight += J->ExtraWeight;
2666 for (CaseItr I=Cases.begin(), E=Cases.end(); I!=E; ++I, ++numCmps) {
2667 if (I->Low != I->High)
2668 // A range counts double, since it requires two compares.
2675 void SelectionDAGBuilder::UpdateSplitBlock(MachineBasicBlock *First,
2676 MachineBasicBlock *Last) {
2678 for (unsigned i = 0, e = JTCases.size(); i != e; ++i)
2679 if (JTCases[i].first.HeaderBB == First)
2680 JTCases[i].first.HeaderBB = Last;
2682 // Update BitTestCases.
2683 for (unsigned i = 0, e = BitTestCases.size(); i != e; ++i)
2684 if (BitTestCases[i].Parent == First)
2685 BitTestCases[i].Parent = Last;
2688 void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) {
2689 MachineBasicBlock *SwitchMBB = FuncInfo.MBB;
2691 // Figure out which block is immediately after the current one.
2692 MachineBasicBlock *NextBlock = nullptr;
2693 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()];
2695 // If there is only the default destination, branch to it if it is not the
2696 // next basic block. Otherwise, just fall through.
2697 if (!SI.getNumCases()) {
2698 // Update machine-CFG edges.
2700 // If this is not a fall-through branch, emit the branch.
2701 SwitchMBB->addSuccessor(Default);
2702 if (Default != NextBlock)
2703 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(),
2704 MVT::Other, getControlRoot(),
2705 DAG.getBasicBlock(Default)));
2710 // If there are any non-default case statements, create a vector of Cases
2711 // representing each one, and sort the vector so that we can efficiently
2712 // create a binary search tree from them.
2714 size_t numCmps = Clusterify(Cases, SI);
2715 DEBUG(dbgs() << "Clusterify finished. Total clusters: " << Cases.size()
2716 << ". Total compares: " << numCmps << '\n');
2719 // Get the Value to be switched on and default basic blocks, which will be
2720 // inserted into CaseBlock records, representing basic blocks in the binary
2722 const Value *SV = SI.getCondition();
2724 // Push the initial CaseRec onto the worklist
2725 CaseRecVector WorkList;
2726 WorkList.push_back(CaseRec(SwitchMBB,nullptr,nullptr,
2727 CaseRange(Cases.begin(),Cases.end())));
2729 while (!WorkList.empty()) {
2730 // Grab a record representing a case range to process off the worklist
2731 CaseRec CR = WorkList.back();
2732 WorkList.pop_back();
2734 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2737 // If the range has few cases (two or less) emit a series of specific
2739 if (handleSmallSwitchRange(CR, WorkList, SV, Default, SwitchMBB))
2742 // If the switch has more than N blocks, and is at least 40% dense, and the
2743 // target supports indirect branches, then emit a jump table rather than
2744 // lowering the switch to a binary tree of conditional branches.
2745 // N defaults to 4 and is controlled via TLS.getMinimumJumpTableEntries().
2746 if (handleJTSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2749 // Emit binary tree. We need to pick a pivot, and push left and right ranges
2750 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call.
2751 handleBTSplitSwitchCase(CR, WorkList, SV, Default, SwitchMBB);
2755 void SelectionDAGBuilder::visitIndirectBr(const IndirectBrInst &I) {
2756 MachineBasicBlock *IndirectBrMBB = FuncInfo.MBB;
2758 // Update machine-CFG edges with unique successors.
2759 SmallSet<BasicBlock*, 32> Done;
2760 for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i) {
2761 BasicBlock *BB = I.getSuccessor(i);
2762 bool Inserted = Done.insert(BB);
2766 MachineBasicBlock *Succ = FuncInfo.MBBMap[BB];
2767 addSuccessorWithWeight(IndirectBrMBB, Succ);
2770 DAG.setRoot(DAG.getNode(ISD::BRIND, getCurSDLoc(),
2771 MVT::Other, getControlRoot(),
2772 getValue(I.getAddress())));
2775 void SelectionDAGBuilder::visitUnreachable(const UnreachableInst &I) {
2776 if (DAG.getTarget().Options.TrapUnreachable)
2777 DAG.setRoot(DAG.getNode(ISD::TRAP, getCurSDLoc(), MVT::Other, DAG.getRoot()));
2780 void SelectionDAGBuilder::visitFSub(const User &I) {
2781 // -0.0 - X --> fneg
2782 Type *Ty = I.getType();
2783 if (isa<Constant>(I.getOperand(0)) &&
2784 I.getOperand(0) == ConstantFP::getZeroValueForNegation(Ty)) {
2785 SDValue Op2 = getValue(I.getOperand(1));
2786 setValue(&I, DAG.getNode(ISD::FNEG, getCurSDLoc(),
2787 Op2.getValueType(), Op2));
2791 visitBinary(I, ISD::FSUB);
2794 void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) {
2795 SDValue Op1 = getValue(I.getOperand(0));
2796 SDValue Op2 = getValue(I.getOperand(1));
2801 if (const OverflowingBinaryOperator *OFBinOp =
2802 dyn_cast<const OverflowingBinaryOperator>(&I)) {
2803 nuw = OFBinOp->hasNoUnsignedWrap();
2804 nsw = OFBinOp->hasNoSignedWrap();
2806 if (const PossiblyExactOperator *ExactOp =
2807 dyn_cast<const PossiblyExactOperator>(&I))
2808 exact = ExactOp->isExact();
2810 SDValue BinNodeValue = DAG.getNode(OpCode, getCurSDLoc(), Op1.getValueType(),
2811 Op1, Op2, nuw, nsw, exact);
2812 setValue(&I, BinNodeValue);
2815 void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) {
2816 SDValue Op1 = getValue(I.getOperand(0));
2817 SDValue Op2 = getValue(I.getOperand(1));
2819 EVT ShiftTy = TM.getTargetLowering()->getShiftAmountTy(Op2.getValueType());
2821 // Coerce the shift amount to the right type if we can.
2822 if (!I.getType()->isVectorTy() && Op2.getValueType() != ShiftTy) {
2823 unsigned ShiftSize = ShiftTy.getSizeInBits();
2824 unsigned Op2Size = Op2.getValueType().getSizeInBits();
2825 SDLoc DL = getCurSDLoc();
2827 // If the operand is smaller than the shift count type, promote it.
2828 if (ShiftSize > Op2Size)
2829 Op2 = DAG.getNode(ISD::ZERO_EXTEND, DL, ShiftTy, Op2);
2831 // If the operand is larger than the shift count type but the shift
2832 // count type has enough bits to represent any shift value, truncate
2833 // it now. This is a common case and it exposes the truncate to
2834 // optimization early.
2835 else if (ShiftSize >= Log2_32_Ceil(Op2.getValueType().getSizeInBits()))
2836 Op2 = DAG.getNode(ISD::TRUNCATE, DL, ShiftTy, Op2);
2837 // Otherwise we'll need to temporarily settle for some other convenient
2838 // type. Type legalization will make adjustments once the shiftee is split.
2840 Op2 = DAG.getZExtOrTrunc(Op2, DL, MVT::i32);
2847 if (Opcode == ISD::SRL || Opcode == ISD::SRA || Opcode == ISD::SHL) {
2849 if (const OverflowingBinaryOperator *OFBinOp =
2850 dyn_cast<const OverflowingBinaryOperator>(&I)) {
2851 nuw = OFBinOp->hasNoUnsignedWrap();
2852 nsw = OFBinOp->hasNoSignedWrap();
2854 if (const PossiblyExactOperator *ExactOp =
2855 dyn_cast<const PossiblyExactOperator>(&I))
2856 exact = ExactOp->isExact();
2859 SDValue Res = DAG.getNode(Opcode, getCurSDLoc(), Op1.getValueType(), Op1, Op2,
2864 void SelectionDAGBuilder::visitSDiv(const User &I) {
2865 SDValue Op1 = getValue(I.getOperand(0));
2866 SDValue Op2 = getValue(I.getOperand(1));
2868 // Turn exact SDivs into multiplications.
2869 // FIXME: This should be in DAGCombiner, but it doesn't have access to the
2871 if (isa<BinaryOperator>(&I) && cast<BinaryOperator>(&I)->isExact() &&
2872 !isa<ConstantSDNode>(Op1) &&
2873 isa<ConstantSDNode>(Op2) && !cast<ConstantSDNode>(Op2)->isNullValue())
2874 setValue(&I, TM.getTargetLowering()->BuildExactSDIV(Op1, Op2,
2875 getCurSDLoc(), DAG));
2877 setValue(&I, DAG.getNode(ISD::SDIV, getCurSDLoc(), Op1.getValueType(),
2881 void SelectionDAGBuilder::visitICmp(const User &I) {
2882 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2883 if (const ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2884 predicate = IC->getPredicate();
2885 else if (const ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2886 predicate = ICmpInst::Predicate(IC->getPredicate());
2887 SDValue Op1 = getValue(I.getOperand(0));
2888 SDValue Op2 = getValue(I.getOperand(1));
2889 ISD::CondCode Opcode = getICmpCondCode(predicate);
2891 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2892 setValue(&I, DAG.getSetCC(getCurSDLoc(), DestVT, Op1, Op2, Opcode));
2895 void SelectionDAGBuilder::visitFCmp(const User &I) {
2896 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2897 if (const FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2898 predicate = FC->getPredicate();
2899 else if (const ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2900 predicate = FCmpInst::Predicate(FC->getPredicate());
2901 SDValue Op1 = getValue(I.getOperand(0));
2902 SDValue Op2 = getValue(I.getOperand(1));
2903 ISD::CondCode Condition = getFCmpCondCode(predicate);
2904 if (TM.Options.NoNaNsFPMath)
2905 Condition = getFCmpCodeWithoutNaN(Condition);
2906 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2907 setValue(&I, DAG.getSetCC(getCurSDLoc(), DestVT, Op1, Op2, Condition));
2910 void SelectionDAGBuilder::visitSelect(const User &I) {
2911 SmallVector<EVT, 4> ValueVTs;
2912 ComputeValueVTs(*TM.getTargetLowering(), I.getType(), ValueVTs);
2913 unsigned NumValues = ValueVTs.size();
2914 if (NumValues == 0) return;
2916 SmallVector<SDValue, 4> Values(NumValues);
2917 SDValue Cond = getValue(I.getOperand(0));
2918 SDValue TrueVal = getValue(I.getOperand(1));
2919 SDValue FalseVal = getValue(I.getOperand(2));
2920 ISD::NodeType OpCode = Cond.getValueType().isVector() ?
2921 ISD::VSELECT : ISD::SELECT;
2923 for (unsigned i = 0; i != NumValues; ++i)
2924 Values[i] = DAG.getNode(OpCode, getCurSDLoc(),
2925 TrueVal.getNode()->getValueType(TrueVal.getResNo()+i),
2927 SDValue(TrueVal.getNode(),
2928 TrueVal.getResNo() + i),
2929 SDValue(FalseVal.getNode(),
2930 FalseVal.getResNo() + i));
2932 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
2933 DAG.getVTList(ValueVTs), Values));
2936 void SelectionDAGBuilder::visitTrunc(const User &I) {
2937 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2938 SDValue N = getValue(I.getOperand(0));
2939 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2940 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), DestVT, N));
2943 void SelectionDAGBuilder::visitZExt(const User &I) {
2944 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2945 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2946 SDValue N = getValue(I.getOperand(0));
2947 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2948 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurSDLoc(), DestVT, N));
2951 void SelectionDAGBuilder::visitSExt(const User &I) {
2952 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2953 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2954 SDValue N = getValue(I.getOperand(0));
2955 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2956 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurSDLoc(), DestVT, N));
2959 void SelectionDAGBuilder::visitFPTrunc(const User &I) {
2960 // FPTrunc is never a no-op cast, no need to check
2961 SDValue N = getValue(I.getOperand(0));
2962 const TargetLowering *TLI = TM.getTargetLowering();
2963 EVT DestVT = TLI->getValueType(I.getType());
2964 setValue(&I, DAG.getNode(ISD::FP_ROUND, getCurSDLoc(),
2966 DAG.getTargetConstant(0, TLI->getPointerTy())));
2969 void SelectionDAGBuilder::visitFPExt(const User &I) {
2970 // FPExt is never a no-op cast, no need to check
2971 SDValue N = getValue(I.getOperand(0));
2972 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2973 setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurSDLoc(), DestVT, N));
2976 void SelectionDAGBuilder::visitFPToUI(const User &I) {
2977 // FPToUI is never a no-op cast, no need to check
2978 SDValue N = getValue(I.getOperand(0));
2979 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2980 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurSDLoc(), DestVT, N));
2983 void SelectionDAGBuilder::visitFPToSI(const User &I) {
2984 // FPToSI is never a no-op cast, no need to check
2985 SDValue N = getValue(I.getOperand(0));
2986 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2987 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurSDLoc(), DestVT, N));
2990 void SelectionDAGBuilder::visitUIToFP(const User &I) {
2991 // UIToFP is never a no-op cast, no need to check
2992 SDValue N = getValue(I.getOperand(0));
2993 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2994 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurSDLoc(), DestVT, N));
2997 void SelectionDAGBuilder::visitSIToFP(const User &I) {
2998 // SIToFP is never a no-op cast, no need to check
2999 SDValue N = getValue(I.getOperand(0));
3000 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
3001 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurSDLoc(), DestVT, N));
3004 void SelectionDAGBuilder::visitPtrToInt(const User &I) {
3005 // What to do depends on the size of the integer and the size of the pointer.
3006 // We can either truncate, zero extend, or no-op, accordingly.
3007 SDValue N = getValue(I.getOperand(0));
3008 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
3009 setValue(&I, DAG.getZExtOrTrunc(N, getCurSDLoc(), DestVT));
3012 void SelectionDAGBuilder::visitIntToPtr(const User &I) {
3013 // What to do depends on the size of the integer and the size of the pointer.
3014 // We can either truncate, zero extend, or no-op, accordingly.
3015 SDValue N = getValue(I.getOperand(0));
3016 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
3017 setValue(&I, DAG.getZExtOrTrunc(N, getCurSDLoc(), DestVT));
3020 void SelectionDAGBuilder::visitBitCast(const User &I) {
3021 SDValue N = getValue(I.getOperand(0));
3022 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
3024 // BitCast assures us that source and destination are the same size so this is
3025 // either a BITCAST or a no-op.
3026 if (DestVT != N.getValueType())
3027 setValue(&I, DAG.getNode(ISD::BITCAST, getCurSDLoc(),
3028 DestVT, N)); // convert types.
3029 // Check if the original LLVM IR Operand was a ConstantInt, because getValue()
3030 // might fold any kind of constant expression to an integer constant and that
3031 // is not what we are looking for. Only regcognize a bitcast of a genuine
3032 // constant integer as an opaque constant.
3033 else if(ConstantInt *C = dyn_cast<ConstantInt>(I.getOperand(0)))
3034 setValue(&I, DAG.getConstant(C->getValue(), DestVT, /*isTarget=*/false,
3037 setValue(&I, N); // noop cast.
3040 void SelectionDAGBuilder::visitAddrSpaceCast(const User &I) {
3041 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3042 const Value *SV = I.getOperand(0);
3043 SDValue N = getValue(SV);
3044 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
3046 unsigned SrcAS = SV->getType()->getPointerAddressSpace();
3047 unsigned DestAS = I.getType()->getPointerAddressSpace();
3049 if (!TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
3050 N = DAG.getAddrSpaceCast(getCurSDLoc(), DestVT, N, SrcAS, DestAS);
3055 void SelectionDAGBuilder::visitInsertElement(const User &I) {
3056 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3057 SDValue InVec = getValue(I.getOperand(0));
3058 SDValue InVal = getValue(I.getOperand(1));
3059 SDValue InIdx = DAG.getSExtOrTrunc(getValue(I.getOperand(2)),
3060 getCurSDLoc(), TLI.getVectorIdxTy());
3061 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurSDLoc(),
3062 TM.getTargetLowering()->getValueType(I.getType()),
3063 InVec, InVal, InIdx));
3066 void SelectionDAGBuilder::visitExtractElement(const User &I) {
3067 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3068 SDValue InVec = getValue(I.getOperand(0));
3069 SDValue InIdx = DAG.getSExtOrTrunc(getValue(I.getOperand(1)),
3070 getCurSDLoc(), TLI.getVectorIdxTy());
3071 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurSDLoc(),
3072 TM.getTargetLowering()->getValueType(I.getType()),
3076 // Utility for visitShuffleVector - Return true if every element in Mask,
3077 // beginning from position Pos and ending in Pos+Size, falls within the
3078 // specified sequential range [L, L+Pos). or is undef.
3079 static bool isSequentialInRange(const SmallVectorImpl<int> &Mask,
3080 unsigned Pos, unsigned Size, int Low) {
3081 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
3082 if (Mask[i] >= 0 && Mask[i] != Low)
3087 void SelectionDAGBuilder::visitShuffleVector(const User &I) {
3088 SDValue Src1 = getValue(I.getOperand(0));
3089 SDValue Src2 = getValue(I.getOperand(1));
3091 SmallVector<int, 8> Mask;
3092 ShuffleVectorInst::getShuffleMask(cast<Constant>(I.getOperand(2)), Mask);
3093 unsigned MaskNumElts = Mask.size();
3095 const TargetLowering *TLI = TM.getTargetLowering();
3096 EVT VT = TLI->getValueType(I.getType());
3097 EVT SrcVT = Src1.getValueType();
3098 unsigned SrcNumElts = SrcVT.getVectorNumElements();
3100 if (SrcNumElts == MaskNumElts) {
3101 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
3106 // Normalize the shuffle vector since mask and vector length don't match.
3107 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) {
3108 // Mask is longer than the source vectors and is a multiple of the source
3109 // vectors. We can use concatenate vector to make the mask and vectors
3111 if (SrcNumElts*2 == MaskNumElts) {
3112 // First check for Src1 in low and Src2 in high
3113 if (isSequentialInRange(Mask, 0, SrcNumElts, 0) &&
3114 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, SrcNumElts)) {
3115 // The shuffle is concatenating two vectors together.
3116 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurSDLoc(),
3120 // Then check for Src2 in low and Src1 in high
3121 if (isSequentialInRange(Mask, 0, SrcNumElts, SrcNumElts) &&
3122 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, 0)) {
3123 // The shuffle is concatenating two vectors together.
3124 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurSDLoc(),
3130 // Pad both vectors with undefs to make them the same length as the mask.
3131 unsigned NumConcat = MaskNumElts / SrcNumElts;
3132 bool Src1U = Src1.getOpcode() == ISD::UNDEF;
3133 bool Src2U = Src2.getOpcode() == ISD::UNDEF;
3134 SDValue UndefVal = DAG.getUNDEF(SrcVT);
3136 SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal);
3137 SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal);
3141 Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
3142 getCurSDLoc(), VT, MOps1);
3143 Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
3144 getCurSDLoc(), VT, MOps2);
3146 // Readjust mask for new input vector length.
3147 SmallVector<int, 8> MappedOps;
3148 for (unsigned i = 0; i != MaskNumElts; ++i) {
3150 if (Idx >= (int)SrcNumElts)
3151 Idx -= SrcNumElts - MaskNumElts;
3152 MappedOps.push_back(Idx);
3155 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
3160 if (SrcNumElts > MaskNumElts) {
3161 // Analyze the access pattern of the vector to see if we can extract
3162 // two subvectors and do the shuffle. The analysis is done by calculating
3163 // the range of elements the mask access on both vectors.
3164 int MinRange[2] = { static_cast<int>(SrcNumElts),
3165 static_cast<int>(SrcNumElts)};
3166 int MaxRange[2] = {-1, -1};
3168 for (unsigned i = 0; i != MaskNumElts; ++i) {
3174 if (Idx >= (int)SrcNumElts) {
3178 if (Idx > MaxRange[Input])
3179 MaxRange[Input] = Idx;
3180 if (Idx < MinRange[Input])
3181 MinRange[Input] = Idx;
3184 // Check if the access is smaller than the vector size and can we find
3185 // a reasonable extract index.
3186 int RangeUse[2] = { -1, -1 }; // 0 = Unused, 1 = Extract, -1 = Can not
3188 int StartIdx[2]; // StartIdx to extract from
3189 for (unsigned Input = 0; Input < 2; ++Input) {
3190 if (MinRange[Input] >= (int)SrcNumElts && MaxRange[Input] < 0) {
3191 RangeUse[Input] = 0; // Unused
3192 StartIdx[Input] = 0;
3196 // Find a good start index that is a multiple of the mask length. Then
3197 // see if the rest of the elements are in range.
3198 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts;
3199 if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts &&
3200 StartIdx[Input] + MaskNumElts <= SrcNumElts)
3201 RangeUse[Input] = 1; // Extract from a multiple of the mask length.
3204 if (RangeUse[0] == 0 && RangeUse[1] == 0) {
3205 setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used.
3208 if (RangeUse[0] >= 0 && RangeUse[1] >= 0) {
3209 // Extract appropriate subvector and generate a vector shuffle
3210 for (unsigned Input = 0; Input < 2; ++Input) {
3211 SDValue &Src = Input == 0 ? Src1 : Src2;
3212 if (RangeUse[Input] == 0)
3213 Src = DAG.getUNDEF(VT);
3215 Src = DAG.getNode(ISD::EXTRACT_SUBVECTOR, getCurSDLoc(), VT,
3216 Src, DAG.getConstant(StartIdx[Input],
3217 TLI->getVectorIdxTy()));
3220 // Calculate new mask.
3221 SmallVector<int, 8> MappedOps;
3222 for (unsigned i = 0; i != MaskNumElts; ++i) {
3225 if (Idx < (int)SrcNumElts)
3228 Idx -= SrcNumElts + StartIdx[1] - MaskNumElts;
3230 MappedOps.push_back(Idx);
3233 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
3239 // We can't use either concat vectors or extract subvectors so fall back to
3240 // replacing the shuffle with extract and build vector.
3241 // to insert and build vector.
3242 EVT EltVT = VT.getVectorElementType();
3243 EVT IdxVT = TLI->getVectorIdxTy();
3244 SmallVector<SDValue,8> Ops;
3245 for (unsigned i = 0; i != MaskNumElts; ++i) {
3250 Res = DAG.getUNDEF(EltVT);
3252 SDValue &Src = Idx < (int)SrcNumElts ? Src1 : Src2;
3253 if (Idx >= (int)SrcNumElts) Idx -= SrcNumElts;
3255 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurSDLoc(),
3256 EltVT, Src, DAG.getConstant(Idx, IdxVT));
3262 setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(), VT, Ops));
3265 void SelectionDAGBuilder::visitInsertValue(const InsertValueInst &I) {
3266 const Value *Op0 = I.getOperand(0);
3267 const Value *Op1 = I.getOperand(1);
3268 Type *AggTy = I.getType();
3269 Type *ValTy = Op1->getType();
3270 bool IntoUndef = isa<UndefValue>(Op0);
3271 bool FromUndef = isa<UndefValue>(Op1);
3273 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
3275 const TargetLowering *TLI = TM.getTargetLowering();
3276 SmallVector<EVT, 4> AggValueVTs;
3277 ComputeValueVTs(*TLI, AggTy, AggValueVTs);
3278 SmallVector<EVT, 4> ValValueVTs;
3279 ComputeValueVTs(*TLI, ValTy, ValValueVTs);
3281 unsigned NumAggValues = AggValueVTs.size();
3282 unsigned NumValValues = ValValueVTs.size();
3283 SmallVector<SDValue, 4> Values(NumAggValues);
3285 SDValue Agg = getValue(Op0);
3287 // Copy the beginning value(s) from the original aggregate.
3288 for (; i != LinearIndex; ++i)
3289 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
3290 SDValue(Agg.getNode(), Agg.getResNo() + i);
3291 // Copy values from the inserted value(s).
3293 SDValue Val = getValue(Op1);
3294 for (; i != LinearIndex + NumValValues; ++i)
3295 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) :
3296 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex);
3298 // Copy remaining value(s) from the original aggregate.
3299 for (; i != NumAggValues; ++i)
3300 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
3301 SDValue(Agg.getNode(), Agg.getResNo() + i);
3303 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
3304 DAG.getVTList(AggValueVTs), Values));
3307 void SelectionDAGBuilder::visitExtractValue(const ExtractValueInst &I) {
3308 const Value *Op0 = I.getOperand(0);
3309 Type *AggTy = Op0->getType();
3310 Type *ValTy = I.getType();
3311 bool OutOfUndef = isa<UndefValue>(Op0);
3313 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
3315 const TargetLowering *TLI = TM.getTargetLowering();
3316 SmallVector<EVT, 4> ValValueVTs;
3317 ComputeValueVTs(*TLI, ValTy, ValValueVTs);
3319 unsigned NumValValues = ValValueVTs.size();
3321 // Ignore a extractvalue that produces an empty object
3322 if (!NumValValues) {
3323 setValue(&I, DAG.getUNDEF(MVT(MVT::Other)));
3327 SmallVector<SDValue, 4> Values(NumValValues);
3329 SDValue Agg = getValue(Op0);
3330 // Copy out the selected value(s).
3331 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i)
3332 Values[i - LinearIndex] =
3334 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) :
3335 SDValue(Agg.getNode(), Agg.getResNo() + i);
3337 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
3338 DAG.getVTList(ValValueVTs), Values));
3341 void SelectionDAGBuilder::visitGetElementPtr(const User &I) {
3342 Value *Op0 = I.getOperand(0);
3343 // Note that the pointer operand may be a vector of pointers. Take the scalar
3344 // element which holds a pointer.
3345 Type *Ty = Op0->getType()->getScalarType();
3346 unsigned AS = Ty->getPointerAddressSpace();
3347 SDValue N = getValue(Op0);
3349 for (GetElementPtrInst::const_op_iterator OI = I.op_begin()+1, E = I.op_end();
3351 const Value *Idx = *OI;
3352 if (StructType *StTy = dyn_cast<StructType>(Ty)) {
3353 unsigned Field = cast<Constant>(Idx)->getUniqueInteger().getZExtValue();
3356 uint64_t Offset = DL->getStructLayout(StTy)->getElementOffset(Field);
3357 N = DAG.getNode(ISD::ADD, getCurSDLoc(), N.getValueType(), N,
3358 DAG.getConstant(Offset, N.getValueType()));
3361 Ty = StTy->getElementType(Field);
3363 Ty = cast<SequentialType>(Ty)->getElementType();
3365 // If this is a constant subscript, handle it quickly.
3366 const TargetLowering *TLI = TM.getTargetLowering();
3367 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
3368 if (CI->isZero()) continue;
3370 DL->getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
3372 EVT PTy = TLI->getPointerTy(AS);
3373 unsigned PtrBits = PTy.getSizeInBits();
3375 OffsVal = DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), PTy,
3376 DAG.getConstant(Offs, MVT::i64));
3378 OffsVal = DAG.getConstant(Offs, PTy);
3380 N = DAG.getNode(ISD::ADD, getCurSDLoc(), N.getValueType(), N,
3385 // N = N + Idx * ElementSize;
3386 APInt ElementSize = APInt(TLI->getPointerSizeInBits(AS),
3387 DL->getTypeAllocSize(Ty));
3388 SDValue IdxN = getValue(Idx);
3390 // If the index is smaller or larger than intptr_t, truncate or extend
3392 IdxN = DAG.getSExtOrTrunc(IdxN, getCurSDLoc(), N.getValueType());
3394 // If this is a multiply by a power of two, turn it into a shl
3395 // immediately. This is a very common case.
3396 if (ElementSize != 1) {
3397 if (ElementSize.isPowerOf2()) {
3398 unsigned Amt = ElementSize.logBase2();
3399 IdxN = DAG.getNode(ISD::SHL, getCurSDLoc(),
3400 N.getValueType(), IdxN,
3401 DAG.getConstant(Amt, IdxN.getValueType()));
3403 SDValue Scale = DAG.getConstant(ElementSize, IdxN.getValueType());
3404 IdxN = DAG.getNode(ISD::MUL, getCurSDLoc(),
3405 N.getValueType(), IdxN, Scale);
3409 N = DAG.getNode(ISD::ADD, getCurSDLoc(),
3410 N.getValueType(), N, IdxN);
3417 void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) {
3418 // If this is a fixed sized alloca in the entry block of the function,
3419 // allocate it statically on the stack.
3420 if (FuncInfo.StaticAllocaMap.count(&I))
3421 return; // getValue will auto-populate this.
3423 Type *Ty = I.getAllocatedType();
3424 const TargetLowering *TLI = TM.getTargetLowering();
3425 uint64_t TySize = TLI->getDataLayout()->getTypeAllocSize(Ty);
3427 std::max((unsigned)TLI->getDataLayout()->getPrefTypeAlignment(Ty),
3430 SDValue AllocSize = getValue(I.getArraySize());
3432 EVT IntPtr = TLI->getPointerTy();
3433 if (AllocSize.getValueType() != IntPtr)
3434 AllocSize = DAG.getZExtOrTrunc(AllocSize, getCurSDLoc(), IntPtr);
3436 AllocSize = DAG.getNode(ISD::MUL, getCurSDLoc(), IntPtr,
3438 DAG.getConstant(TySize, IntPtr));
3440 // Handle alignment. If the requested alignment is less than or equal to
3441 // the stack alignment, ignore it. If the size is greater than or equal to
3442 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
3443 unsigned StackAlign = TM.getFrameLowering()->getStackAlignment();
3444 if (Align <= StackAlign)
3447 // Round the size of the allocation up to the stack alignment size
3448 // by add SA-1 to the size.
3449 AllocSize = DAG.getNode(ISD::ADD, getCurSDLoc(),
3450 AllocSize.getValueType(), AllocSize,
3451 DAG.getIntPtrConstant(StackAlign-1));
3453 // Mask out the low bits for alignment purposes.
3454 AllocSize = DAG.getNode(ISD::AND, getCurSDLoc(),
3455 AllocSize.getValueType(), AllocSize,
3456 DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1)));
3458 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) };
3459 SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other);
3460 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, getCurSDLoc(), VTs, Ops);
3462 DAG.setRoot(DSA.getValue(1));
3464 assert(FuncInfo.MF->getFrameInfo()->hasVarSizedObjects());
3467 void SelectionDAGBuilder::visitLoad(const LoadInst &I) {
3469 return visitAtomicLoad(I);
3471 const Value *SV = I.getOperand(0);
3472 SDValue Ptr = getValue(SV);
3474 Type *Ty = I.getType();
3476 bool isVolatile = I.isVolatile();
3477 bool isNonTemporal = I.getMetadata("nontemporal") != nullptr;
3478 bool isInvariant = I.getMetadata("invariant.load") != nullptr;
3479 unsigned Alignment = I.getAlignment();
3482 I.getAAMetadata(AAInfo);
3483 const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range);
3485 SmallVector<EVT, 4> ValueVTs;
3486 SmallVector<uint64_t, 4> Offsets;
3487 ComputeValueVTs(*TM.getTargetLowering(), Ty, ValueVTs, &Offsets);
3488 unsigned NumValues = ValueVTs.size();
3493 bool ConstantMemory = false;
3494 if (isVolatile || NumValues > MaxParallelChains)
3495 // Serialize volatile loads with other side effects.
3497 else if (AA->pointsToConstantMemory(
3498 AliasAnalysis::Location(SV, AA->getTypeStoreSize(Ty), AAInfo))) {
3499 // Do not serialize (non-volatile) loads of constant memory with anything.
3500 Root = DAG.getEntryNode();
3501 ConstantMemory = true;
3503 // Do not serialize non-volatile loads against each other.
3504 Root = DAG.getRoot();
3507 const TargetLowering *TLI = TM.getTargetLowering();
3509 Root = TLI->prepareVolatileOrAtomicLoad(Root, getCurSDLoc(), DAG);
3511 SmallVector<SDValue, 4> Values(NumValues);
3512 SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains),
3514 EVT PtrVT = Ptr.getValueType();
3515 unsigned ChainI = 0;
3516 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3517 // Serializing loads here may result in excessive register pressure, and
3518 // TokenFactor places arbitrary choke points on the scheduler. SD scheduling
3519 // could recover a bit by hoisting nodes upward in the chain by recognizing
3520 // they are side-effect free or do not alias. The optimizer should really
3521 // avoid this case by converting large object/array copies to llvm.memcpy
3522 // (MaxParallelChains should always remain as failsafe).
3523 if (ChainI == MaxParallelChains) {
3524 assert(PendingLoads.empty() && "PendingLoads must be serialized first");
3525 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
3526 makeArrayRef(Chains.data(), ChainI));
3530 SDValue A = DAG.getNode(ISD::ADD, getCurSDLoc(),
3532 DAG.getConstant(Offsets[i], PtrVT));
3533 SDValue L = DAG.getLoad(ValueVTs[i], getCurSDLoc(), Root,
3534 A, MachinePointerInfo(SV, Offsets[i]), isVolatile,
3535 isNonTemporal, isInvariant, Alignment, AAInfo,
3539 Chains[ChainI] = L.getValue(1);
3542 if (!ConstantMemory) {
3543 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
3544 makeArrayRef(Chains.data(), ChainI));
3548 PendingLoads.push_back(Chain);
3551 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
3552 DAG.getVTList(ValueVTs), Values));
3555 void SelectionDAGBuilder::visitStore(const StoreInst &I) {
3557 return visitAtomicStore(I);
3559 const Value *SrcV = I.getOperand(0);
3560 const Value *PtrV = I.getOperand(1);
3562 SmallVector<EVT, 4> ValueVTs;
3563 SmallVector<uint64_t, 4> Offsets;
3564 ComputeValueVTs(*TM.getTargetLowering(), SrcV->getType(), ValueVTs, &Offsets);
3565 unsigned NumValues = ValueVTs.size();
3569 // Get the lowered operands. Note that we do this after
3570 // checking if NumResults is zero, because with zero results
3571 // the operands won't have values in the map.
3572 SDValue Src = getValue(SrcV);
3573 SDValue Ptr = getValue(PtrV);
3575 SDValue Root = getRoot();
3576 SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains),
3578 EVT PtrVT = Ptr.getValueType();
3579 bool isVolatile = I.isVolatile();
3580 bool isNonTemporal = I.getMetadata("nontemporal") != nullptr;
3581 unsigned Alignment = I.getAlignment();
3584 I.getAAMetadata(AAInfo);
3586 unsigned ChainI = 0;
3587 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3588 // See visitLoad comments.
3589 if (ChainI == MaxParallelChains) {
3590 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
3591 makeArrayRef(Chains.data(), ChainI));
3595 SDValue Add = DAG.getNode(ISD::ADD, getCurSDLoc(), PtrVT, Ptr,
3596 DAG.getConstant(Offsets[i], PtrVT));
3597 SDValue St = DAG.getStore(Root, getCurSDLoc(),
3598 SDValue(Src.getNode(), Src.getResNo() + i),
3599 Add, MachinePointerInfo(PtrV, Offsets[i]),
3600 isVolatile, isNonTemporal, Alignment, AAInfo);
3601 Chains[ChainI] = St;
3604 SDValue StoreNode = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
3605 makeArrayRef(Chains.data(), ChainI));
3606 DAG.setRoot(StoreNode);
3609 static SDValue InsertFenceForAtomic(SDValue Chain, AtomicOrdering Order,
3610 SynchronizationScope Scope,
3611 bool Before, SDLoc dl,
3613 const TargetLowering &TLI) {
3614 // Fence, if necessary
3616 if (Order == AcquireRelease || Order == SequentiallyConsistent)
3618 else if (Order == Acquire || Order == Monotonic || Order == Unordered)
3621 if (Order == AcquireRelease)
3623 else if (Order == Release || Order == Monotonic || Order == Unordered)
3628 Ops[1] = DAG.getConstant(Order, TLI.getPointerTy());
3629 Ops[2] = DAG.getConstant(Scope, TLI.getPointerTy());
3630 return DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops);
3633 void SelectionDAGBuilder::visitAtomicCmpXchg(const AtomicCmpXchgInst &I) {
3634 SDLoc dl = getCurSDLoc();
3635 AtomicOrdering SuccessOrder = I.getSuccessOrdering();
3636 AtomicOrdering FailureOrder = I.getFailureOrdering();
3637 SynchronizationScope Scope = I.getSynchScope();
3639 SDValue InChain = getRoot();
3641 const TargetLowering *TLI = TM.getTargetLowering();
3642 if (TLI->getInsertFencesForAtomic())
3643 InChain = InsertFenceForAtomic(InChain, SuccessOrder, Scope, true, dl,
3646 MVT MemVT = getValue(I.getCompareOperand()).getSimpleValueType();
3647 SDVTList VTs = DAG.getVTList(MemVT, MVT::i1, MVT::Other);
3648 SDValue L = DAG.getAtomicCmpSwap(
3649 ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, dl, MemVT, VTs, InChain,
3650 getValue(I.getPointerOperand()), getValue(I.getCompareOperand()),
3651 getValue(I.getNewValOperand()), MachinePointerInfo(I.getPointerOperand()),
3653 TLI->getInsertFencesForAtomic() ? Monotonic : SuccessOrder,
3654 TLI->getInsertFencesForAtomic() ? Monotonic : FailureOrder, Scope);
3656 SDValue OutChain = L.getValue(2);
3658 if (TLI->getInsertFencesForAtomic())
3659 OutChain = InsertFenceForAtomic(OutChain, SuccessOrder, Scope, false, dl,
3663 DAG.setRoot(OutChain);
3666 void SelectionDAGBuilder::visitAtomicRMW(const AtomicRMWInst &I) {
3667 SDLoc dl = getCurSDLoc();
3669 switch (I.getOperation()) {
3670 default: llvm_unreachable("Unknown atomicrmw operation");
3671 case AtomicRMWInst::Xchg: NT = ISD::ATOMIC_SWAP; break;
3672 case AtomicRMWInst::Add: NT = ISD::ATOMIC_LOAD_ADD; break;
3673 case AtomicRMWInst::Sub: NT = ISD::ATOMIC_LOAD_SUB; break;
3674 case AtomicRMWInst::And: NT = ISD::ATOMIC_LOAD_AND; break;
3675 case AtomicRMWInst::Nand: NT = ISD::ATOMIC_LOAD_NAND; break;
3676 case AtomicRMWInst::Or: NT = ISD::ATOMIC_LOAD_OR; break;
3677 case AtomicRMWInst::Xor: NT = ISD::ATOMIC_LOAD_XOR; break;
3678 case AtomicRMWInst::Max: NT = ISD::ATOMIC_LOAD_MAX; break;
3679 case AtomicRMWInst::Min: NT = ISD::ATOMIC_LOAD_MIN; break;
3680 case AtomicRMWInst::UMax: NT = ISD::ATOMIC_LOAD_UMAX; break;
3681 case AtomicRMWInst::UMin: NT = ISD::ATOMIC_LOAD_UMIN; break;
3683 AtomicOrdering Order = I.getOrdering();
3684 SynchronizationScope Scope = I.getSynchScope();
3686 SDValue InChain = getRoot();
3688 const TargetLowering *TLI = TM.getTargetLowering();
3689 if (TLI->getInsertFencesForAtomic())
3690 InChain = InsertFenceForAtomic(InChain, Order, Scope, true, dl,
3694 DAG.getAtomic(NT, dl,
3695 getValue(I.getValOperand()).getSimpleValueType(),
3697 getValue(I.getPointerOperand()),
3698 getValue(I.getValOperand()),
3699 I.getPointerOperand(), 0 /* Alignment */,
3700 TLI->getInsertFencesForAtomic() ? Monotonic : Order,
3703 SDValue OutChain = L.getValue(1);
3705 if (TLI->getInsertFencesForAtomic())
3706 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3710 DAG.setRoot(OutChain);
3713 void SelectionDAGBuilder::visitFence(const FenceInst &I) {
3714 SDLoc dl = getCurSDLoc();
3715 const TargetLowering *TLI = TM.getTargetLowering();
3718 Ops[1] = DAG.getConstant(I.getOrdering(), TLI->getPointerTy());
3719 Ops[2] = DAG.getConstant(I.getSynchScope(), TLI->getPointerTy());
3720 DAG.setRoot(DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops));
3723 void SelectionDAGBuilder::visitAtomicLoad(const LoadInst &I) {
3724 SDLoc dl = getCurSDLoc();
3725 AtomicOrdering Order = I.getOrdering();
3726 SynchronizationScope Scope = I.getSynchScope();
3728 SDValue InChain = getRoot();
3730 const TargetLowering *TLI = TM.getTargetLowering();
3731 EVT VT = TLI->getValueType(I.getType());
3733 if (I.getAlignment() < VT.getSizeInBits() / 8)
3734 report_fatal_error("Cannot generate unaligned atomic load");
3736 MachineMemOperand *MMO =
3737 DAG.getMachineFunction().
3738 getMachineMemOperand(MachinePointerInfo(I.getPointerOperand()),
3739 MachineMemOperand::MOVolatile |
3740 MachineMemOperand::MOLoad,
3742 I.getAlignment() ? I.getAlignment() :
3743 DAG.getEVTAlignment(VT));
3745 InChain = TLI->prepareVolatileOrAtomicLoad(InChain, dl, DAG);
3747 DAG.getAtomic(ISD::ATOMIC_LOAD, dl, VT, VT, InChain,
3748 getValue(I.getPointerOperand()), MMO,
3749 TLI->getInsertFencesForAtomic() ? Monotonic : Order,
3752 SDValue OutChain = L.getValue(1);
3754 if (TLI->getInsertFencesForAtomic())
3755 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3759 DAG.setRoot(OutChain);
3762 void SelectionDAGBuilder::visitAtomicStore(const StoreInst &I) {
3763 SDLoc dl = getCurSDLoc();
3765 AtomicOrdering Order = I.getOrdering();
3766 SynchronizationScope Scope = I.getSynchScope();
3768 SDValue InChain = getRoot();
3770 const TargetLowering *TLI = TM.getTargetLowering();
3771 EVT VT = TLI->getValueType(I.getValueOperand()->getType());
3773 if (I.getAlignment() < VT.getSizeInBits() / 8)
3774 report_fatal_error("Cannot generate unaligned atomic store");
3776 if (TLI->getInsertFencesForAtomic())
3777 InChain = InsertFenceForAtomic(InChain, Order, Scope, true, dl,
3781 DAG.getAtomic(ISD::ATOMIC_STORE, dl, VT,
3783 getValue(I.getPointerOperand()),
3784 getValue(I.getValueOperand()),
3785 I.getPointerOperand(), I.getAlignment(),
3786 TLI->getInsertFencesForAtomic() ? Monotonic : Order,
3789 if (TLI->getInsertFencesForAtomic())
3790 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3793 DAG.setRoot(OutChain);
3796 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
3798 void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I,
3799 unsigned Intrinsic) {
3800 bool HasChain = !I.doesNotAccessMemory();
3801 bool OnlyLoad = HasChain && I.onlyReadsMemory();
3803 // Build the operand list.
3804 SmallVector<SDValue, 8> Ops;
3805 if (HasChain) { // If this intrinsic has side-effects, chainify it.
3807 // We don't need to serialize loads against other loads.
3808 Ops.push_back(DAG.getRoot());
3810 Ops.push_back(getRoot());
3814 // Info is set by getTgtMemInstrinsic
3815 TargetLowering::IntrinsicInfo Info;
3816 const TargetLowering *TLI = TM.getTargetLowering();
3817 bool IsTgtIntrinsic = TLI->getTgtMemIntrinsic(Info, I, Intrinsic);
3819 // Add the intrinsic ID as an integer operand if it's not a target intrinsic.
3820 if (!IsTgtIntrinsic || Info.opc == ISD::INTRINSIC_VOID ||
3821 Info.opc == ISD::INTRINSIC_W_CHAIN)
3822 Ops.push_back(DAG.getTargetConstant(Intrinsic, TLI->getPointerTy()));
3824 // Add all operands of the call to the operand list.
3825 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) {
3826 SDValue Op = getValue(I.getArgOperand(i));
3830 SmallVector<EVT, 4> ValueVTs;
3831 ComputeValueVTs(*TLI, I.getType(), ValueVTs);
3834 ValueVTs.push_back(MVT::Other);
3836 SDVTList VTs = DAG.getVTList(ValueVTs);
3840 if (IsTgtIntrinsic) {
3841 // This is target intrinsic that touches memory
3842 Result = DAG.getMemIntrinsicNode(Info.opc, getCurSDLoc(),
3843 VTs, Ops, Info.memVT,
3844 MachinePointerInfo(Info.ptrVal, Info.offset),
3845 Info.align, Info.vol,
3846 Info.readMem, Info.writeMem);
3847 } else if (!HasChain) {
3848 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurSDLoc(), VTs, Ops);
3849 } else if (!I.getType()->isVoidTy()) {
3850 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurSDLoc(), VTs, Ops);
3852 Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurSDLoc(), VTs, Ops);
3856 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1);
3858 PendingLoads.push_back(Chain);
3863 if (!I.getType()->isVoidTy()) {
3864 if (VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
3865 EVT VT = TLI->getValueType(PTy);
3866 Result = DAG.getNode(ISD::BITCAST, getCurSDLoc(), VT, Result);
3869 setValue(&I, Result);
3873 /// GetSignificand - Get the significand and build it into a floating-point
3874 /// number with exponent of 1:
3876 /// Op = (Op & 0x007fffff) | 0x3f800000;
3878 /// where Op is the hexadecimal representation of floating point value.
3880 GetSignificand(SelectionDAG &DAG, SDValue Op, SDLoc dl) {
3881 SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3882 DAG.getConstant(0x007fffff, MVT::i32));
3883 SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1,
3884 DAG.getConstant(0x3f800000, MVT::i32));
3885 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, t2);
3888 /// GetExponent - Get the exponent:
3890 /// (float)(int)(((Op & 0x7f800000) >> 23) - 127);
3892 /// where Op is the hexadecimal representation of floating point value.
3894 GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI,
3896 SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3897 DAG.getConstant(0x7f800000, MVT::i32));
3898 SDValue t1 = DAG.getNode(ISD::SRL, dl, MVT::i32, t0,
3899 DAG.getConstant(23, TLI.getPointerTy()));
3900 SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1,
3901 DAG.getConstant(127, MVT::i32));
3902 return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2);
3905 /// getF32Constant - Get 32-bit floating point constant.
3907 getF32Constant(SelectionDAG &DAG, unsigned Flt) {
3908 return DAG.getConstantFP(APFloat(APFloat::IEEEsingle, APInt(32, Flt)),
3912 /// expandExp - Lower an exp intrinsic. Handles the special sequences for
3913 /// limited-precision mode.
3914 static SDValue expandExp(SDLoc dl, SDValue Op, SelectionDAG &DAG,
3915 const TargetLowering &TLI) {
3916 if (Op.getValueType() == MVT::f32 &&
3917 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3919 // Put the exponent in the right bit position for later addition to the
3922 // #define LOG2OFe 1.4426950f
3923 // IntegerPartOfX = ((int32_t)(X * LOG2OFe));
3924 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3925 getF32Constant(DAG, 0x3fb8aa3b));
3926 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3928 // FractionalPartOfX = (X * LOG2OFe) - (float)IntegerPartOfX;
3929 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3930 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3932 // IntegerPartOfX <<= 23;
3933 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3934 DAG.getConstant(23, TLI.getPointerTy()));
3936 SDValue TwoToFracPartOfX;
3937 if (LimitFloatPrecision <= 6) {
3938 // For floating-point precision of 6:
3940 // TwoToFractionalPartOfX =
3942 // (0.735607626f + 0.252464424f * x) * x;
3944 // error 0.0144103317, which is 6 bits
3945 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3946 getF32Constant(DAG, 0x3e814304));
3947 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3948 getF32Constant(DAG, 0x3f3c50c8));
3949 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3950 TwoToFracPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3951 getF32Constant(DAG, 0x3f7f5e7e));
3952 } else if (LimitFloatPrecision <= 12) {
3953 // For floating-point precision of 12:
3955 // TwoToFractionalPartOfX =
3958 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3960 // 0.000107046256 error, which is 13 to 14 bits
3961 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3962 getF32Constant(DAG, 0x3da235e3));
3963 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3964 getF32Constant(DAG, 0x3e65b8f3));
3965 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3966 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3967 getF32Constant(DAG, 0x3f324b07));
3968 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3969 TwoToFracPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3970 getF32Constant(DAG, 0x3f7ff8fd));
3971 } else { // LimitFloatPrecision <= 18
3972 // For floating-point precision of 18:
3974 // TwoToFractionalPartOfX =
3978 // (0.554906021e-1f +
3979 // (0.961591928e-2f +
3980 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3982 // error 2.47208000*10^(-7), which is better than 18 bits
3983 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3984 getF32Constant(DAG, 0x3924b03e));
3985 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3986 getF32Constant(DAG, 0x3ab24b87));
3987 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3988 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3989 getF32Constant(DAG, 0x3c1d8c17));
3990 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3991 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3992 getF32Constant(DAG, 0x3d634a1d));
3993 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3994 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3995 getF32Constant(DAG, 0x3e75fe14));
3996 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3997 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3998 getF32Constant(DAG, 0x3f317234));
3999 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
4000 TwoToFracPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
4001 getF32Constant(DAG, 0x3f800000));
4004 // Add the exponent into the result in integer domain.
4005 SDValue t13 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, TwoToFracPartOfX);
4006 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4007 DAG.getNode(ISD::ADD, dl, MVT::i32,
4008 t13, IntegerPartOfX));
4011 // No special expansion.
4012 return DAG.getNode(ISD::FEXP, dl, Op.getValueType(), Op);
4015 /// expandLog - Lower a log intrinsic. Handles the special sequences for
4016 /// limited-precision mode.
4017 static SDValue expandLog(SDLoc dl, SDValue Op, SelectionDAG &DAG,
4018 const TargetLowering &TLI) {
4019 if (Op.getValueType() == MVT::f32 &&
4020 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4021 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
4023 // Scale the exponent by log(2) [0.69314718f].
4024 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
4025 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
4026 getF32Constant(DAG, 0x3f317218));
4028 // Get the significand and build it into a floating-point number with
4030 SDValue X = GetSignificand(DAG, Op1, dl);
4032 SDValue LogOfMantissa;
4033 if (LimitFloatPrecision <= 6) {
4034 // For floating-point precision of 6:
4038 // (1.4034025f - 0.23903021f * x) * x;
4040 // error 0.0034276066, which is better than 8 bits
4041 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4042 getF32Constant(DAG, 0xbe74c456));
4043 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4044 getF32Constant(DAG, 0x3fb3a2b1));
4045 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4046 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4047 getF32Constant(DAG, 0x3f949a29));
4048 } else if (LimitFloatPrecision <= 12) {
4049 // For floating-point precision of 12:
4055 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x;
4057 // error 0.000061011436, which is 14 bits
4058 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4059 getF32Constant(DAG, 0xbd67b6d6));
4060 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4061 getF32Constant(DAG, 0x3ee4f4b8));
4062 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4063 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4064 getF32Constant(DAG, 0x3fbc278b));
4065 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4066 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4067 getF32Constant(DAG, 0x40348e95));
4068 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4069 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
4070 getF32Constant(DAG, 0x3fdef31a));
4071 } else { // LimitFloatPrecision <= 18
4072 // For floating-point precision of 18:
4080 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x;
4082 // error 0.0000023660568, which is better than 18 bits
4083 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4084 getF32Constant(DAG, 0xbc91e5ac));
4085 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4086 getF32Constant(DAG, 0x3e4350aa));
4087 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4088 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4089 getF32Constant(DAG, 0x3f60d3e3));
4090 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4091 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4092 getF32Constant(DAG, 0x4011cdf0));
4093 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4094 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
4095 getF32Constant(DAG, 0x406cfd1c));
4096 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4097 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4098 getF32Constant(DAG, 0x408797cb));
4099 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4100 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
4101 getF32Constant(DAG, 0x4006dcab));
4104 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, LogOfMantissa);
4107 // No special expansion.
4108 return DAG.getNode(ISD::FLOG, dl, Op.getValueType(), Op);
4111 /// expandLog2 - Lower a log2 intrinsic. Handles the special sequences for
4112 /// limited-precision mode.
4113 static SDValue expandLog2(SDLoc dl, SDValue Op, SelectionDAG &DAG,
4114 const TargetLowering &TLI) {
4115 if (Op.getValueType() == MVT::f32 &&
4116 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4117 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
4119 // Get the exponent.
4120 SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl);
4122 // Get the significand and build it into a floating-point number with
4124 SDValue X = GetSignificand(DAG, Op1, dl);
4126 // Different possible minimax approximations of significand in
4127 // floating-point for various degrees of accuracy over [1,2].
4128 SDValue Log2ofMantissa;
4129 if (LimitFloatPrecision <= 6) {
4130 // For floating-point precision of 6:
4132 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x;
4134 // error 0.0049451742, which is more than 7 bits
4135 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4136 getF32Constant(DAG, 0xbeb08fe0));
4137 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4138 getF32Constant(DAG, 0x40019463));
4139 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4140 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4141 getF32Constant(DAG, 0x3fd6633d));
4142 } else if (LimitFloatPrecision <= 12) {
4143 // For floating-point precision of 12:
4149 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x;
4151 // error 0.0000876136000, which is better than 13 bits
4152 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4153 getF32Constant(DAG, 0xbda7262e));
4154 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4155 getF32Constant(DAG, 0x3f25280b));
4156 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4157 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4158 getF32Constant(DAG, 0x4007b923));
4159 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4160 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4161 getF32Constant(DAG, 0x40823e2f));
4162 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4163 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
4164 getF32Constant(DAG, 0x4020d29c));
4165 } else { // LimitFloatPrecision <= 18
4166 // For floating-point precision of 18:
4175 // 0.25691327e-1f * x) * x) * x) * x) * x) * x;
4177 // error 0.0000018516, which is better than 18 bits
4178 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4179 getF32Constant(DAG, 0xbcd2769e));
4180 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4181 getF32Constant(DAG, 0x3e8ce0b9));
4182 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4183 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4184 getF32Constant(DAG, 0x3fa22ae7));
4185 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4186 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4187 getF32Constant(DAG, 0x40525723));
4188 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4189 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
4190 getF32Constant(DAG, 0x40aaf200));
4191 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4192 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4193 getF32Constant(DAG, 0x40c39dad));
4194 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4195 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
4196 getF32Constant(DAG, 0x4042902c));
4199 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, Log2ofMantissa);
4202 // No special expansion.
4203 return DAG.getNode(ISD::FLOG2, dl, Op.getValueType(), Op);
4206 /// expandLog10 - Lower a log10 intrinsic. Handles the special sequences for
4207 /// limited-precision mode.
4208 static SDValue expandLog10(SDLoc dl, SDValue Op, SelectionDAG &DAG,
4209 const TargetLowering &TLI) {
4210 if (Op.getValueType() == MVT::f32 &&
4211 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4212 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
4214 // Scale the exponent by log10(2) [0.30102999f].
4215 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
4216 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
4217 getF32Constant(DAG, 0x3e9a209a));
4219 // Get the significand and build it into a floating-point number with
4221 SDValue X = GetSignificand(DAG, Op1, dl);
4223 SDValue Log10ofMantissa;
4224 if (LimitFloatPrecision <= 6) {
4225 // For floating-point precision of 6:
4227 // Log10ofMantissa =
4229 // (0.60948995f - 0.10380950f * x) * x;
4231 // error 0.0014886165, which is 6 bits
4232 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4233 getF32Constant(DAG, 0xbdd49a13));
4234 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4235 getF32Constant(DAG, 0x3f1c0789));
4236 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4237 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4238 getF32Constant(DAG, 0x3f011300));
4239 } else if (LimitFloatPrecision <= 12) {
4240 // For floating-point precision of 12:
4242 // Log10ofMantissa =
4245 // (-0.31664806f + 0.47637168e-1f * x) * x) * x;
4247 // error 0.00019228036, which is better than 12 bits
4248 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4249 getF32Constant(DAG, 0x3d431f31));
4250 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
4251 getF32Constant(DAG, 0x3ea21fb2));
4252 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4253 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4254 getF32Constant(DAG, 0x3f6ae232));
4255 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4256 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
4257 getF32Constant(DAG, 0x3f25f7c3));
4258 } else { // LimitFloatPrecision <= 18
4259 // For floating-point precision of 18:
4261 // Log10ofMantissa =
4266 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x;
4268 // error 0.0000037995730, which is better than 18 bits
4269 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4270 getF32Constant(DAG, 0x3c5d51ce));
4271 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
4272 getF32Constant(DAG, 0x3e00685a));
4273 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4274 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4275 getF32Constant(DAG, 0x3efb6798));
4276 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4277 SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
4278 getF32Constant(DAG, 0x3f88d192));
4279 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4280 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4281 getF32Constant(DAG, 0x3fc4316c));
4282 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4283 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8,
4284 getF32Constant(DAG, 0x3f57ce70));
4287 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, Log10ofMantissa);
4290 // No special expansion.
4291 return DAG.getNode(ISD::FLOG10, dl, Op.getValueType(), Op);
4294 /// expandExp2 - Lower an exp2 intrinsic. Handles the special sequences for
4295 /// limited-precision mode.
4296 static SDValue expandExp2(SDLoc dl, SDValue Op, SelectionDAG &DAG,
4297 const TargetLowering &TLI) {
4298 if (Op.getValueType() == MVT::f32 &&
4299 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4300 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Op);
4302 // FractionalPartOfX = x - (float)IntegerPartOfX;
4303 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
4304 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, Op, t1);
4306 // IntegerPartOfX <<= 23;
4307 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
4308 DAG.getConstant(23, TLI.getPointerTy()));
4310 SDValue TwoToFractionalPartOfX;
4311 if (LimitFloatPrecision <= 6) {
4312 // For floating-point precision of 6:
4314 // TwoToFractionalPartOfX =
4316 // (0.735607626f + 0.252464424f * x) * x;
4318 // error 0.0144103317, which is 6 bits
4319 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4320 getF32Constant(DAG, 0x3e814304));
4321 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4322 getF32Constant(DAG, 0x3f3c50c8));
4323 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4324 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4325 getF32Constant(DAG, 0x3f7f5e7e));
4326 } else if (LimitFloatPrecision <= 12) {
4327 // For floating-point precision of 12:
4329 // TwoToFractionalPartOfX =
4332 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
4334 // error 0.000107046256, which is 13 to 14 bits
4335 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4336 getF32Constant(DAG, 0x3da235e3));
4337 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4338 getF32Constant(DAG, 0x3e65b8f3));
4339 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4340 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4341 getF32Constant(DAG, 0x3f324b07));
4342 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4343 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4344 getF32Constant(DAG, 0x3f7ff8fd));
4345 } else { // LimitFloatPrecision <= 18
4346 // For floating-point precision of 18:
4348 // TwoToFractionalPartOfX =
4352 // (0.554906021e-1f +
4353 // (0.961591928e-2f +
4354 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
4355 // error 2.47208000*10^(-7), which is better than 18 bits
4356 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4357 getF32Constant(DAG, 0x3924b03e));
4358 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4359 getF32Constant(DAG, 0x3ab24b87));
4360 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4361 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4362 getF32Constant(DAG, 0x3c1d8c17));
4363 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4364 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4365 getF32Constant(DAG, 0x3d634a1d));
4366 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4367 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4368 getF32Constant(DAG, 0x3e75fe14));
4369 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4370 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
4371 getF32Constant(DAG, 0x3f317234));
4372 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
4373 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
4374 getF32Constant(DAG, 0x3f800000));
4377 // Add the exponent into the result in integer domain.
4378 SDValue t13 = DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4379 TwoToFractionalPartOfX);
4380 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4381 DAG.getNode(ISD::ADD, dl, MVT::i32,
4382 t13, IntegerPartOfX));
4385 // No special expansion.
4386 return DAG.getNode(ISD::FEXP2, dl, Op.getValueType(), Op);
4389 /// visitPow - Lower a pow intrinsic. Handles the special sequences for
4390 /// limited-precision mode with x == 10.0f.
4391 static SDValue expandPow(SDLoc dl, SDValue LHS, SDValue RHS,
4392 SelectionDAG &DAG, const TargetLowering &TLI) {
4393 bool IsExp10 = false;
4394 if (LHS.getValueType() == MVT::f32 && RHS.getValueType() == MVT::f32 &&
4395 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4396 if (ConstantFPSDNode *LHSC = dyn_cast<ConstantFPSDNode>(LHS)) {
4398 IsExp10 = LHSC->isExactlyValue(Ten);
4403 // Put the exponent in the right bit position for later addition to the
4406 // #define LOG2OF10 3.3219281f
4407 // IntegerPartOfX = (int32_t)(x * LOG2OF10);
4408 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, RHS,
4409 getF32Constant(DAG, 0x40549a78));
4410 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
4412 // FractionalPartOfX = x - (float)IntegerPartOfX;
4413 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
4414 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
4416 // IntegerPartOfX <<= 23;
4417 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
4418 DAG.getConstant(23, TLI.getPointerTy()));
4420 SDValue TwoToFractionalPartOfX;
4421 if (LimitFloatPrecision <= 6) {
4422 // For floating-point precision of 6:
4424 // twoToFractionalPartOfX =
4426 // (0.735607626f + 0.252464424f * x) * x;
4428 // error 0.0144103317, which is 6 bits
4429 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4430 getF32Constant(DAG, 0x3e814304));
4431 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4432 getF32Constant(DAG, 0x3f3c50c8));
4433 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4434 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4435 getF32Constant(DAG, 0x3f7f5e7e));
4436 } else if (LimitFloatPrecision <= 12) {
4437 // For floating-point precision of 12:
4439 // TwoToFractionalPartOfX =
4442 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
4444 // error 0.000107046256, which is 13 to 14 bits
4445 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4446 getF32Constant(DAG, 0x3da235e3));
4447 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4448 getF32Constant(DAG, 0x3e65b8f3));
4449 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4450 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4451 getF32Constant(DAG, 0x3f324b07));
4452 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4453 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4454 getF32Constant(DAG, 0x3f7ff8fd));
4455 } else { // LimitFloatPrecision <= 18
4456 // For floating-point precision of 18:
4458 // TwoToFractionalPartOfX =
4462 // (0.554906021e-1f +
4463 // (0.961591928e-2f +
4464 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
4465 // error 2.47208000*10^(-7), which is better than 18 bits
4466 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4467 getF32Constant(DAG, 0x3924b03e));
4468 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4469 getF32Constant(DAG, 0x3ab24b87));
4470 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4471 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4472 getF32Constant(DAG, 0x3c1d8c17));
4473 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4474 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4475 getF32Constant(DAG, 0x3d634a1d));
4476 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4477 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4478 getF32Constant(DAG, 0x3e75fe14));
4479 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4480 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
4481 getF32Constant(DAG, 0x3f317234));
4482 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
4483 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
4484 getF32Constant(DAG, 0x3f800000));
4487 SDValue t13 = DAG.getNode(ISD::BITCAST, dl,MVT::i32,TwoToFractionalPartOfX);
4488 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4489 DAG.getNode(ISD::ADD, dl, MVT::i32,
4490 t13, IntegerPartOfX));
4493 // No special expansion.
4494 return DAG.getNode(ISD::FPOW, dl, LHS.getValueType(), LHS, RHS);
4498 /// ExpandPowI - Expand a llvm.powi intrinsic.
4499 static SDValue ExpandPowI(SDLoc DL, SDValue LHS, SDValue RHS,
4500 SelectionDAG &DAG) {
4501 // If RHS is a constant, we can expand this out to a multiplication tree,
4502 // otherwise we end up lowering to a call to __powidf2 (for example). When
4503 // optimizing for size, we only want to do this if the expansion would produce
4504 // a small number of multiplies, otherwise we do the full expansion.
4505 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
4506 // Get the exponent as a positive value.
4507 unsigned Val = RHSC->getSExtValue();
4508 if ((int)Val < 0) Val = -Val;
4510 // powi(x, 0) -> 1.0
4512 return DAG.getConstantFP(1.0, LHS.getValueType());
4514 const Function *F = DAG.getMachineFunction().getFunction();
4515 if (!F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
4516 Attribute::OptimizeForSize) ||
4517 // If optimizing for size, don't insert too many multiplies. This
4518 // inserts up to 5 multiplies.
4519 CountPopulation_32(Val)+Log2_32(Val) < 7) {
4520 // We use the simple binary decomposition method to generate the multiply
4521 // sequence. There are more optimal ways to do this (for example,
4522 // powi(x,15) generates one more multiply than it should), but this has
4523 // the benefit of being both really simple and much better than a libcall.
4524 SDValue Res; // Logically starts equal to 1.0
4525 SDValue CurSquare = LHS;
4529 Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare);
4531 Res = CurSquare; // 1.0*CurSquare.
4534 CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(),
4535 CurSquare, CurSquare);
4539 // If the original was negative, invert the result, producing 1/(x*x*x).
4540 if (RHSC->getSExtValue() < 0)
4541 Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(),
4542 DAG.getConstantFP(1.0, LHS.getValueType()), Res);
4547 // Otherwise, expand to a libcall.
4548 return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS);
4551 // getTruncatedArgReg - Find underlying register used for an truncated
4553 static unsigned getTruncatedArgReg(const SDValue &N) {
4554 if (N.getOpcode() != ISD::TRUNCATE)
4557 const SDValue &Ext = N.getOperand(0);
4558 if (Ext.getOpcode() == ISD::AssertZext ||
4559 Ext.getOpcode() == ISD::AssertSext) {
4560 const SDValue &CFR = Ext.getOperand(0);
4561 if (CFR.getOpcode() == ISD::CopyFromReg)
4562 return cast<RegisterSDNode>(CFR.getOperand(1))->getReg();
4563 if (CFR.getOpcode() == ISD::TRUNCATE)
4564 return getTruncatedArgReg(CFR);
4569 /// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a function
4570 /// argument, create the corresponding DBG_VALUE machine instruction for it now.
4571 /// At the end of instruction selection, they will be inserted to the entry BB.
4573 SelectionDAGBuilder::EmitFuncArgumentDbgValue(const Value *V, MDNode *Variable,
4574 int64_t Offset, bool IsIndirect,
4576 const Argument *Arg = dyn_cast<Argument>(V);
4580 MachineFunction &MF = DAG.getMachineFunction();
4581 const TargetInstrInfo *TII = DAG.getTarget().getInstrInfo();
4583 // Ignore inlined function arguments here.
4584 DIVariable DV(Variable);
4585 if (DV.isInlinedFnArgument(MF.getFunction()))
4588 Optional<MachineOperand> Op;
4589 // Some arguments' frame index is recorded during argument lowering.
4590 if (int FI = FuncInfo.getArgumentFrameIndex(Arg))
4591 Op = MachineOperand::CreateFI(FI);
4593 if (!Op && N.getNode()) {
4595 if (N.getOpcode() == ISD::CopyFromReg)
4596 Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg();
4598 Reg = getTruncatedArgReg(N);
4599 if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) {
4600 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4601 unsigned PR = RegInfo.getLiveInPhysReg(Reg);
4606 Op = MachineOperand::CreateReg(Reg, false);
4610 // Check if ValueMap has reg number.
4611 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
4612 if (VMI != FuncInfo.ValueMap.end())
4613 Op = MachineOperand::CreateReg(VMI->second, false);
4616 if (!Op && N.getNode())
4617 // Check if frame index is available.
4618 if (LoadSDNode *LNode = dyn_cast<LoadSDNode>(N.getNode()))
4619 if (FrameIndexSDNode *FINode =
4620 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode()))
4621 Op = MachineOperand::CreateFI(FINode->getIndex());
4627 FuncInfo.ArgDbgValues.push_back(BuildMI(MF, getCurDebugLoc(),
4628 TII->get(TargetOpcode::DBG_VALUE),
4630 Op->getReg(), Offset, Variable));
4632 FuncInfo.ArgDbgValues.push_back(
4633 BuildMI(MF, getCurDebugLoc(), TII->get(TargetOpcode::DBG_VALUE))
4634 .addOperand(*Op).addImm(Offset).addMetadata(Variable));
4639 // VisualStudio defines setjmp as _setjmp
4640 #if defined(_MSC_VER) && defined(setjmp) && \
4641 !defined(setjmp_undefined_for_msvc)
4642 # pragma push_macro("setjmp")
4644 # define setjmp_undefined_for_msvc
4647 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
4648 /// we want to emit this as a call to a named external function, return the name
4649 /// otherwise lower it and return null.
4651 SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) {
4652 const TargetLowering *TLI = TM.getTargetLowering();
4653 SDLoc sdl = getCurSDLoc();
4654 DebugLoc dl = getCurDebugLoc();
4657 switch (Intrinsic) {
4659 // By default, turn this into a target intrinsic node.
4660 visitTargetIntrinsic(I, Intrinsic);
4662 case Intrinsic::vastart: visitVAStart(I); return nullptr;
4663 case Intrinsic::vaend: visitVAEnd(I); return nullptr;
4664 case Intrinsic::vacopy: visitVACopy(I); return nullptr;
4665 case Intrinsic::returnaddress:
4666 setValue(&I, DAG.getNode(ISD::RETURNADDR, sdl, TLI->getPointerTy(),
4667 getValue(I.getArgOperand(0))));
4669 case Intrinsic::frameaddress:
4670 setValue(&I, DAG.getNode(ISD::FRAMEADDR, sdl, TLI->getPointerTy(),
4671 getValue(I.getArgOperand(0))));
4673 case Intrinsic::read_register: {
4674 Value *Reg = I.getArgOperand(0);
4675 SDValue RegName = DAG.getMDNode(cast<MDNode>(Reg));
4676 EVT VT = TM.getTargetLowering()->getValueType(I.getType());
4677 setValue(&I, DAG.getNode(ISD::READ_REGISTER, sdl, VT, RegName));
4680 case Intrinsic::write_register: {
4681 Value *Reg = I.getArgOperand(0);
4682 Value *RegValue = I.getArgOperand(1);
4683 SDValue Chain = getValue(RegValue).getOperand(0);
4684 SDValue RegName = DAG.getMDNode(cast<MDNode>(Reg));
4685 DAG.setRoot(DAG.getNode(ISD::WRITE_REGISTER, sdl, MVT::Other, Chain,
4686 RegName, getValue(RegValue)));
4689 case Intrinsic::setjmp:
4690 return &"_setjmp"[!TLI->usesUnderscoreSetJmp()];
4691 case Intrinsic::longjmp:
4692 return &"_longjmp"[!TLI->usesUnderscoreLongJmp()];
4693 case Intrinsic::memcpy: {
4694 // Assert for address < 256 since we support only user defined address
4696 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4698 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4700 "Unknown address space");
4701 SDValue Op1 = getValue(I.getArgOperand(0));
4702 SDValue Op2 = getValue(I.getArgOperand(1));
4703 SDValue Op3 = getValue(I.getArgOperand(2));
4704 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4706 Align = 1; // @llvm.memcpy defines 0 and 1 to both mean no alignment.
4707 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4708 DAG.setRoot(DAG.getMemcpy(getRoot(), sdl, Op1, Op2, Op3, Align, isVol, false,
4709 MachinePointerInfo(I.getArgOperand(0)),
4710 MachinePointerInfo(I.getArgOperand(1))));
4713 case Intrinsic::memset: {
4714 // Assert for address < 256 since we support only user defined address
4716 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4718 "Unknown address space");
4719 SDValue Op1 = getValue(I.getArgOperand(0));
4720 SDValue Op2 = getValue(I.getArgOperand(1));
4721 SDValue Op3 = getValue(I.getArgOperand(2));
4722 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4724 Align = 1; // @llvm.memset defines 0 and 1 to both mean no alignment.
4725 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4726 DAG.setRoot(DAG.getMemset(getRoot(), sdl, Op1, Op2, Op3, Align, isVol,
4727 MachinePointerInfo(I.getArgOperand(0))));
4730 case Intrinsic::memmove: {
4731 // Assert for address < 256 since we support only user defined address
4733 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4735 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4737 "Unknown address space");
4738 SDValue Op1 = getValue(I.getArgOperand(0));
4739 SDValue Op2 = getValue(I.getArgOperand(1));
4740 SDValue Op3 = getValue(I.getArgOperand(2));
4741 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4743 Align = 1; // @llvm.memmove defines 0 and 1 to both mean no alignment.
4744 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4745 DAG.setRoot(DAG.getMemmove(getRoot(), sdl, Op1, Op2, Op3, Align, isVol,
4746 MachinePointerInfo(I.getArgOperand(0)),
4747 MachinePointerInfo(I.getArgOperand(1))));
4750 case Intrinsic::dbg_declare: {
4751 const DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
4752 MDNode *Variable = DI.getVariable();
4753 const Value *Address = DI.getAddress();
4754 DIVariable DIVar(Variable);
4755 assert((!DIVar || DIVar.isVariable()) &&
4756 "Variable in DbgDeclareInst should be either null or a DIVariable.");
4757 if (!Address || !DIVar) {
4758 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4762 // Check if address has undef value.
4763 if (isa<UndefValue>(Address) ||
4764 (Address->use_empty() && !isa<Argument>(Address))) {
4765 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4769 SDValue &N = NodeMap[Address];
4770 if (!N.getNode() && isa<Argument>(Address))
4771 // Check unused arguments map.
4772 N = UnusedArgNodeMap[Address];
4775 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
4776 Address = BCI->getOperand(0);
4777 // Parameters are handled specially.
4779 (DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable ||
4780 isa<Argument>(Address));
4782 const AllocaInst *AI = dyn_cast<AllocaInst>(Address);
4784 if (isParameter && !AI) {
4785 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode());
4787 // Byval parameter. We have a frame index at this point.
4788 SDV = DAG.getFrameIndexDbgValue(Variable, FINode->getIndex(),
4789 0, dl, SDNodeOrder);
4791 // Address is an argument, so try to emit its dbg value using
4792 // virtual register info from the FuncInfo.ValueMap.
4793 EmitFuncArgumentDbgValue(Address, Variable, 0, false, N);
4797 SDV = DAG.getDbgValue(Variable, N.getNode(), N.getResNo(),
4798 true, 0, dl, SDNodeOrder);
4800 // Can't do anything with other non-AI cases yet.
4801 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4802 DEBUG(dbgs() << "non-AllocaInst issue for Address: \n\t");
4803 DEBUG(Address->dump());
4806 DAG.AddDbgValue(SDV, N.getNode(), isParameter);
4808 // If Address is an argument then try to emit its dbg value using
4809 // virtual register info from the FuncInfo.ValueMap.
4810 if (!EmitFuncArgumentDbgValue(Address, Variable, 0, false, N)) {
4811 // If variable is pinned by a alloca in dominating bb then
4812 // use StaticAllocaMap.
4813 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Address)) {
4814 if (AI->getParent() != DI.getParent()) {
4815 DenseMap<const AllocaInst*, int>::iterator SI =
4816 FuncInfo.StaticAllocaMap.find(AI);
4817 if (SI != FuncInfo.StaticAllocaMap.end()) {
4818 SDV = DAG.getFrameIndexDbgValue(Variable, SI->second,
4819 0, dl, SDNodeOrder);
4820 DAG.AddDbgValue(SDV, nullptr, false);
4825 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4830 case Intrinsic::dbg_value: {
4831 const DbgValueInst &DI = cast<DbgValueInst>(I);
4832 DIVariable DIVar(DI.getVariable());
4833 assert((!DIVar || DIVar.isVariable()) &&
4834 "Variable in DbgValueInst should be either null or a DIVariable.");
4838 MDNode *Variable = DI.getVariable();
4839 uint64_t Offset = DI.getOffset();
4840 const Value *V = DI.getValue();
4845 if (isa<ConstantInt>(V) || isa<ConstantFP>(V) || isa<UndefValue>(V)) {
4846 SDV = DAG.getConstantDbgValue(Variable, V, Offset, dl, SDNodeOrder);
4847 DAG.AddDbgValue(SDV, nullptr, false);
4849 // Do not use getValue() in here; we don't want to generate code at
4850 // this point if it hasn't been done yet.
4851 SDValue N = NodeMap[V];
4852 if (!N.getNode() && isa<Argument>(V))
4853 // Check unused arguments map.
4854 N = UnusedArgNodeMap[V];
4856 // A dbg.value for an alloca is always indirect.
4857 bool IsIndirect = isa<AllocaInst>(V) || Offset != 0;
4858 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, IsIndirect, N)) {
4859 SDV = DAG.getDbgValue(Variable, N.getNode(),
4860 N.getResNo(), IsIndirect,
4861 Offset, dl, SDNodeOrder);
4862 DAG.AddDbgValue(SDV, N.getNode(), false);
4864 } else if (!V->use_empty() ) {
4865 // Do not call getValue(V) yet, as we don't want to generate code.
4866 // Remember it for later.
4867 DanglingDebugInfo DDI(&DI, dl, SDNodeOrder);
4868 DanglingDebugInfoMap[V] = DDI;
4870 // We may expand this to cover more cases. One case where we have no
4871 // data available is an unreferenced parameter.
4872 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4876 // Build a debug info table entry.
4877 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
4878 V = BCI->getOperand(0);
4879 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
4880 // Don't handle byval struct arguments or VLAs, for example.
4882 DEBUG(dbgs() << "Dropping debug location info for:\n " << DI << "\n");
4883 DEBUG(dbgs() << " Last seen at:\n " << *V << "\n");
4886 DenseMap<const AllocaInst*, int>::iterator SI =
4887 FuncInfo.StaticAllocaMap.find(AI);
4888 if (SI == FuncInfo.StaticAllocaMap.end())
4889 return nullptr; // VLAs.
4893 case Intrinsic::eh_typeid_for: {
4894 // Find the type id for the given typeinfo.
4895 GlobalVariable *GV = ExtractTypeInfo(I.getArgOperand(0));
4896 unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(GV);
4897 Res = DAG.getConstant(TypeID, MVT::i32);
4902 case Intrinsic::eh_return_i32:
4903 case Intrinsic::eh_return_i64:
4904 DAG.getMachineFunction().getMMI().setCallsEHReturn(true);
4905 DAG.setRoot(DAG.getNode(ISD::EH_RETURN, sdl,
4908 getValue(I.getArgOperand(0)),
4909 getValue(I.getArgOperand(1))));
4911 case Intrinsic::eh_unwind_init:
4912 DAG.getMachineFunction().getMMI().setCallsUnwindInit(true);
4914 case Intrinsic::eh_dwarf_cfa: {
4915 SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getArgOperand(0)), sdl,
4916 TLI->getPointerTy());
4917 SDValue Offset = DAG.getNode(ISD::ADD, sdl,
4918 CfaArg.getValueType(),
4919 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, sdl,
4920 CfaArg.getValueType()),
4922 SDValue FA = DAG.getNode(ISD::FRAMEADDR, sdl,
4923 TLI->getPointerTy(),
4924 DAG.getConstant(0, TLI->getPointerTy()));
4925 setValue(&I, DAG.getNode(ISD::ADD, sdl, FA.getValueType(),
4929 case Intrinsic::eh_sjlj_callsite: {
4930 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4931 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(0));
4932 assert(CI && "Non-constant call site value in eh.sjlj.callsite!");
4933 assert(MMI.getCurrentCallSite() == 0 && "Overlapping call sites!");
4935 MMI.setCurrentCallSite(CI->getZExtValue());
4938 case Intrinsic::eh_sjlj_functioncontext: {
4939 // Get and store the index of the function context.
4940 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4942 cast<AllocaInst>(I.getArgOperand(0)->stripPointerCasts());
4943 int FI = FuncInfo.StaticAllocaMap[FnCtx];
4944 MFI->setFunctionContextIndex(FI);
4947 case Intrinsic::eh_sjlj_setjmp: {
4950 Ops[1] = getValue(I.getArgOperand(0));
4951 SDValue Op = DAG.getNode(ISD::EH_SJLJ_SETJMP, sdl,
4952 DAG.getVTList(MVT::i32, MVT::Other), Ops);
4953 setValue(&I, Op.getValue(0));
4954 DAG.setRoot(Op.getValue(1));
4957 case Intrinsic::eh_sjlj_longjmp: {
4958 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, sdl, MVT::Other,
4959 getRoot(), getValue(I.getArgOperand(0))));
4963 case Intrinsic::x86_mmx_pslli_w:
4964 case Intrinsic::x86_mmx_pslli_d:
4965 case Intrinsic::x86_mmx_pslli_q:
4966 case Intrinsic::x86_mmx_psrli_w:
4967 case Intrinsic::x86_mmx_psrli_d:
4968 case Intrinsic::x86_mmx_psrli_q:
4969 case Intrinsic::x86_mmx_psrai_w:
4970 case Intrinsic::x86_mmx_psrai_d: {
4971 SDValue ShAmt = getValue(I.getArgOperand(1));
4972 if (isa<ConstantSDNode>(ShAmt)) {
4973 visitTargetIntrinsic(I, Intrinsic);
4976 unsigned NewIntrinsic = 0;
4977 EVT ShAmtVT = MVT::v2i32;
4978 switch (Intrinsic) {
4979 case Intrinsic::x86_mmx_pslli_w:
4980 NewIntrinsic = Intrinsic::x86_mmx_psll_w;
4982 case Intrinsic::x86_mmx_pslli_d:
4983 NewIntrinsic = Intrinsic::x86_mmx_psll_d;
4985 case Intrinsic::x86_mmx_pslli_q:
4986 NewIntrinsic = Intrinsic::x86_mmx_psll_q;
4988 case Intrinsic::x86_mmx_psrli_w:
4989 NewIntrinsic = Intrinsic::x86_mmx_psrl_w;
4991 case Intrinsic::x86_mmx_psrli_d:
4992 NewIntrinsic = Intrinsic::x86_mmx_psrl_d;
4994 case Intrinsic::x86_mmx_psrli_q:
4995 NewIntrinsic = Intrinsic::x86_mmx_psrl_q;
4997 case Intrinsic::x86_mmx_psrai_w:
4998 NewIntrinsic = Intrinsic::x86_mmx_psra_w;
5000 case Intrinsic::x86_mmx_psrai_d:
5001 NewIntrinsic = Intrinsic::x86_mmx_psra_d;
5003 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
5006 // The vector shift intrinsics with scalars uses 32b shift amounts but
5007 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
5009 // We must do this early because v2i32 is not a legal type.
5012 ShOps[1] = DAG.getConstant(0, MVT::i32);
5013 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, sdl, ShAmtVT, ShOps);
5014 EVT DestVT = TLI->getValueType(I.getType());
5015 ShAmt = DAG.getNode(ISD::BITCAST, sdl, DestVT, ShAmt);
5016 Res = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, sdl, DestVT,
5017 DAG.getConstant(NewIntrinsic, MVT::i32),
5018 getValue(I.getArgOperand(0)), ShAmt);
5022 case Intrinsic::x86_avx_vinsertf128_pd_256:
5023 case Intrinsic::x86_avx_vinsertf128_ps_256:
5024 case Intrinsic::x86_avx_vinsertf128_si_256:
5025 case Intrinsic::x86_avx2_vinserti128: {
5026 EVT DestVT = TLI->getValueType(I.getType());
5027 EVT ElVT = TLI->getValueType(I.getArgOperand(1)->getType());
5028 uint64_t Idx = (cast<ConstantInt>(I.getArgOperand(2))->getZExtValue() & 1) *
5029 ElVT.getVectorNumElements();
5030 Res = DAG.getNode(ISD::INSERT_SUBVECTOR, sdl, DestVT,
5031 getValue(I.getArgOperand(0)),
5032 getValue(I.getArgOperand(1)),
5033 DAG.getConstant(Idx, TLI->getVectorIdxTy()));
5037 case Intrinsic::x86_avx_vextractf128_pd_256:
5038 case Intrinsic::x86_avx_vextractf128_ps_256:
5039 case Intrinsic::x86_avx_vextractf128_si_256:
5040 case Intrinsic::x86_avx2_vextracti128: {
5041 EVT DestVT = TLI->getValueType(I.getType());
5042 uint64_t Idx = (cast<ConstantInt>(I.getArgOperand(1))->getZExtValue() & 1) *
5043 DestVT.getVectorNumElements();
5044 Res = DAG.getNode(ISD::EXTRACT_SUBVECTOR, sdl, DestVT,
5045 getValue(I.getArgOperand(0)),
5046 DAG.getConstant(Idx, TLI->getVectorIdxTy()));
5050 case Intrinsic::convertff:
5051 case Intrinsic::convertfsi:
5052 case Intrinsic::convertfui:
5053 case Intrinsic::convertsif:
5054 case Intrinsic::convertuif:
5055 case Intrinsic::convertss:
5056 case Intrinsic::convertsu:
5057 case Intrinsic::convertus:
5058 case Intrinsic::convertuu: {
5059 ISD::CvtCode Code = ISD::CVT_INVALID;
5060 switch (Intrinsic) {
5061 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
5062 case Intrinsic::convertff: Code = ISD::CVT_FF; break;
5063 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break;
5064 case Intrinsic::convertfui: Code = ISD::CVT_FU; break;
5065 case Intrinsic::convertsif: Code = ISD::CVT_SF; break;
5066 case Intrinsic::convertuif: Code = ISD::CVT_UF; break;
5067 case Intrinsic::convertss: Code = ISD::CVT_SS; break;
5068 case Intrinsic::convertsu: Code = ISD::CVT_SU; break;
5069 case Intrinsic::convertus: Code = ISD::CVT_US; break;
5070 case Intrinsic::convertuu: Code = ISD::CVT_UU; break;
5072 EVT DestVT = TLI->getValueType(I.getType());
5073 const Value *Op1 = I.getArgOperand(0);
5074 Res = DAG.getConvertRndSat(DestVT, sdl, getValue(Op1),
5075 DAG.getValueType(DestVT),
5076 DAG.getValueType(getValue(Op1).getValueType()),
5077 getValue(I.getArgOperand(1)),
5078 getValue(I.getArgOperand(2)),
5083 case Intrinsic::powi:
5084 setValue(&I, ExpandPowI(sdl, getValue(I.getArgOperand(0)),
5085 getValue(I.getArgOperand(1)), DAG));
5087 case Intrinsic::log:
5088 setValue(&I, expandLog(sdl, getValue(I.getArgOperand(0)), DAG, *TLI));
5090 case Intrinsic::log2:
5091 setValue(&I, expandLog2(sdl, getValue(I.getArgOperand(0)), DAG, *TLI));
5093 case Intrinsic::log10:
5094 setValue(&I, expandLog10(sdl, getValue(I.getArgOperand(0)), DAG, *TLI));
5096 case Intrinsic::exp:
5097 setValue(&I, expandExp(sdl, getValue(I.getArgOperand(0)), DAG, *TLI));
5099 case Intrinsic::exp2:
5100 setValue(&I, expandExp2(sdl, getValue(I.getArgOperand(0)), DAG, *TLI));
5102 case Intrinsic::pow:
5103 setValue(&I, expandPow(sdl, getValue(I.getArgOperand(0)),
5104 getValue(I.getArgOperand(1)), DAG, *TLI));
5106 case Intrinsic::sqrt:
5107 case Intrinsic::fabs:
5108 case Intrinsic::sin:
5109 case Intrinsic::cos:
5110 case Intrinsic::floor:
5111 case Intrinsic::ceil:
5112 case Intrinsic::trunc:
5113 case Intrinsic::rint:
5114 case Intrinsic::nearbyint:
5115 case Intrinsic::round: {
5117 switch (Intrinsic) {
5118 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
5119 case Intrinsic::sqrt: Opcode = ISD::FSQRT; break;
5120 case Intrinsic::fabs: Opcode = ISD::FABS; break;
5121 case Intrinsic::sin: Opcode = ISD::FSIN; break;
5122 case Intrinsic::cos: Opcode = ISD::FCOS; break;
5123 case Intrinsic::floor: Opcode = ISD::FFLOOR; break;
5124 case Intrinsic::ceil: Opcode = ISD::FCEIL; break;
5125 case Intrinsic::trunc: Opcode = ISD::FTRUNC; break;
5126 case Intrinsic::rint: Opcode = ISD::FRINT; break;
5127 case Intrinsic::nearbyint: Opcode = ISD::FNEARBYINT; break;
5128 case Intrinsic::round: Opcode = ISD::FROUND; break;
5131 setValue(&I, DAG.getNode(Opcode, sdl,
5132 getValue(I.getArgOperand(0)).getValueType(),
5133 getValue(I.getArgOperand(0))));
5136 case Intrinsic::copysign:
5137 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, sdl,
5138 getValue(I.getArgOperand(0)).getValueType(),
5139 getValue(I.getArgOperand(0)),
5140 getValue(I.getArgOperand(1))));
5142 case Intrinsic::fma:
5143 setValue(&I, DAG.getNode(ISD::FMA, sdl,
5144 getValue(I.getArgOperand(0)).getValueType(),
5145 getValue(I.getArgOperand(0)),
5146 getValue(I.getArgOperand(1)),
5147 getValue(I.getArgOperand(2))));
5149 case Intrinsic::fmuladd: {
5150 EVT VT = TLI->getValueType(I.getType());
5151 if (TM.Options.AllowFPOpFusion != FPOpFusion::Strict &&
5152 TLI->isFMAFasterThanFMulAndFAdd(VT)) {
5153 setValue(&I, DAG.getNode(ISD::FMA, sdl,
5154 getValue(I.getArgOperand(0)).getValueType(),
5155 getValue(I.getArgOperand(0)),
5156 getValue(I.getArgOperand(1)),
5157 getValue(I.getArgOperand(2))));
5159 SDValue Mul = DAG.getNode(ISD::FMUL, sdl,
5160 getValue(I.getArgOperand(0)).getValueType(),
5161 getValue(I.getArgOperand(0)),
5162 getValue(I.getArgOperand(1)));
5163 SDValue Add = DAG.getNode(ISD::FADD, sdl,
5164 getValue(I.getArgOperand(0)).getValueType(),
5166 getValue(I.getArgOperand(2)));
5171 case Intrinsic::convert_to_fp16:
5172 setValue(&I, DAG.getNode(ISD::BITCAST, sdl, MVT::i16,
5173 DAG.getNode(ISD::FP_ROUND, sdl, MVT::f16,
5174 getValue(I.getArgOperand(0)),
5175 DAG.getTargetConstant(0, MVT::i32))));
5177 case Intrinsic::convert_from_fp16:
5179 DAG.getNode(ISD::FP_EXTEND, sdl, TLI->getValueType(I.getType()),
5180 DAG.getNode(ISD::BITCAST, sdl, MVT::f16,
5181 getValue(I.getArgOperand(0)))));
5183 case Intrinsic::pcmarker: {
5184 SDValue Tmp = getValue(I.getArgOperand(0));
5185 DAG.setRoot(DAG.getNode(ISD::PCMARKER, sdl, MVT::Other, getRoot(), Tmp));
5188 case Intrinsic::readcyclecounter: {
5189 SDValue Op = getRoot();
5190 Res = DAG.getNode(ISD::READCYCLECOUNTER, sdl,
5191 DAG.getVTList(MVT::i64, MVT::Other), Op);
5193 DAG.setRoot(Res.getValue(1));
5196 case Intrinsic::bswap:
5197 setValue(&I, DAG.getNode(ISD::BSWAP, sdl,
5198 getValue(I.getArgOperand(0)).getValueType(),
5199 getValue(I.getArgOperand(0))));
5201 case Intrinsic::cttz: {
5202 SDValue Arg = getValue(I.getArgOperand(0));
5203 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1));
5204 EVT Ty = Arg.getValueType();
5205 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTTZ : ISD::CTTZ_ZERO_UNDEF,
5209 case Intrinsic::ctlz: {
5210 SDValue Arg = getValue(I.getArgOperand(0));
5211 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1));
5212 EVT Ty = Arg.getValueType();
5213 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTLZ : ISD::CTLZ_ZERO_UNDEF,
5217 case Intrinsic::ctpop: {
5218 SDValue Arg = getValue(I.getArgOperand(0));
5219 EVT Ty = Arg.getValueType();
5220 setValue(&I, DAG.getNode(ISD::CTPOP, sdl, Ty, Arg));
5223 case Intrinsic::stacksave: {
5224 SDValue Op = getRoot();
5225 Res = DAG.getNode(ISD::STACKSAVE, sdl,
5226 DAG.getVTList(TLI->getPointerTy(), MVT::Other), Op);
5228 DAG.setRoot(Res.getValue(1));
5231 case Intrinsic::stackrestore: {
5232 Res = getValue(I.getArgOperand(0));
5233 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, sdl, MVT::Other, getRoot(), Res));
5236 case Intrinsic::stackprotector: {
5237 // Emit code into the DAG to store the stack guard onto the stack.
5238 MachineFunction &MF = DAG.getMachineFunction();
5239 MachineFrameInfo *MFI = MF.getFrameInfo();
5240 EVT PtrTy = TLI->getPointerTy();
5241 SDValue Src, Chain = getRoot();
5243 if (TLI->useLoadStackGuardNode()) {
5244 // Emit a LOAD_STACK_GUARD node.
5245 MachineSDNode *Node = DAG.getMachineNode(TargetOpcode::LOAD_STACK_GUARD,
5247 LoadInst *LI = cast<LoadInst>(I.getArgOperand(0));
5248 MachinePointerInfo MPInfo(LI->getPointerOperand());
5249 MachineInstr::mmo_iterator MemRefs = MF.allocateMemRefsArray(1);
5250 unsigned Flags = MachineMemOperand::MOLoad |
5251 MachineMemOperand::MOInvariant;
5252 *MemRefs = MF.getMachineMemOperand(MPInfo, Flags,
5253 PtrTy.getSizeInBits() / 8,
5254 DAG.getEVTAlignment(PtrTy));
5255 Node->setMemRefs(MemRefs, MemRefs + 1);
5257 // Copy the guard value to a virtual register so that it can be
5258 // retrieved in the epilogue.
5259 Src = SDValue(Node, 0);
5260 const TargetRegisterClass *RC =
5261 TLI->getRegClassFor(Src.getSimpleValueType());
5262 unsigned Reg = MF.getRegInfo().createVirtualRegister(RC);
5264 SPDescriptor.setGuardReg(Reg);
5265 Chain = DAG.getCopyToReg(Chain, sdl, Reg, Src);
5267 Src = getValue(I.getArgOperand(0)); // The guard's value.
5270 AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
5272 int FI = FuncInfo.StaticAllocaMap[Slot];
5273 MFI->setStackProtectorIndex(FI);
5275 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
5277 // Store the stack protector onto the stack.
5278 Res = DAG.getStore(Chain, sdl, Src, FIN,
5279 MachinePointerInfo::getFixedStack(FI),
5285 case Intrinsic::objectsize: {
5286 // If we don't know by now, we're never going to know.
5287 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1));
5289 assert(CI && "Non-constant type in __builtin_object_size?");
5291 SDValue Arg = getValue(I.getCalledValue());
5292 EVT Ty = Arg.getValueType();
5295 Res = DAG.getConstant(-1ULL, Ty);
5297 Res = DAG.getConstant(0, Ty);
5302 case Intrinsic::annotation:
5303 case Intrinsic::ptr_annotation:
5304 // Drop the intrinsic, but forward the value
5305 setValue(&I, getValue(I.getOperand(0)));
5307 case Intrinsic::assume:
5308 case Intrinsic::var_annotation:
5309 // Discard annotate attributes and assumptions
5312 case Intrinsic::init_trampoline: {
5313 const Function *F = cast<Function>(I.getArgOperand(1)->stripPointerCasts());
5317 Ops[1] = getValue(I.getArgOperand(0));
5318 Ops[2] = getValue(I.getArgOperand(1));
5319 Ops[3] = getValue(I.getArgOperand(2));
5320 Ops[4] = DAG.getSrcValue(I.getArgOperand(0));
5321 Ops[5] = DAG.getSrcValue(F);
5323 Res = DAG.getNode(ISD::INIT_TRAMPOLINE, sdl, MVT::Other, Ops);
5328 case Intrinsic::adjust_trampoline: {
5329 setValue(&I, DAG.getNode(ISD::ADJUST_TRAMPOLINE, sdl,
5330 TLI->getPointerTy(),
5331 getValue(I.getArgOperand(0))));
5334 case Intrinsic::gcroot:
5336 const Value *Alloca = I.getArgOperand(0)->stripPointerCasts();
5337 const Constant *TypeMap = cast<Constant>(I.getArgOperand(1));
5339 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode());
5340 GFI->addStackRoot(FI->getIndex(), TypeMap);
5343 case Intrinsic::gcread:
5344 case Intrinsic::gcwrite:
5345 llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!");
5346 case Intrinsic::flt_rounds:
5347 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, sdl, MVT::i32));
5350 case Intrinsic::expect: {
5351 // Just replace __builtin_expect(exp, c) with EXP.
5352 setValue(&I, getValue(I.getArgOperand(0)));
5356 case Intrinsic::debugtrap:
5357 case Intrinsic::trap: {
5358 StringRef TrapFuncName = TM.Options.getTrapFunctionName();
5359 if (TrapFuncName.empty()) {
5360 ISD::NodeType Op = (Intrinsic == Intrinsic::trap) ?
5361 ISD::TRAP : ISD::DEBUGTRAP;
5362 DAG.setRoot(DAG.getNode(Op, sdl,MVT::Other, getRoot()));
5365 TargetLowering::ArgListTy Args;
5367 TargetLowering::CallLoweringInfo CLI(DAG);
5368 CLI.setDebugLoc(sdl).setChain(getRoot())
5369 .setCallee(CallingConv::C, I.getType(),
5370 DAG.getExternalSymbol(TrapFuncName.data(), TLI->getPointerTy()),
5371 std::move(Args), 0);
5373 std::pair<SDValue, SDValue> Result = TLI->LowerCallTo(CLI);
5374 DAG.setRoot(Result.second);
5378 case Intrinsic::uadd_with_overflow:
5379 case Intrinsic::sadd_with_overflow:
5380 case Intrinsic::usub_with_overflow:
5381 case Intrinsic::ssub_with_overflow:
5382 case Intrinsic::umul_with_overflow:
5383 case Intrinsic::smul_with_overflow: {
5385 switch (Intrinsic) {
5386 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
5387 case Intrinsic::uadd_with_overflow: Op = ISD::UADDO; break;
5388 case Intrinsic::sadd_with_overflow: Op = ISD::SADDO; break;
5389 case Intrinsic::usub_with_overflow: Op = ISD::USUBO; break;
5390 case Intrinsic::ssub_with_overflow: Op = ISD::SSUBO; break;
5391 case Intrinsic::umul_with_overflow: Op = ISD::UMULO; break;
5392 case Intrinsic::smul_with_overflow: Op = ISD::SMULO; break;
5394 SDValue Op1 = getValue(I.getArgOperand(0));
5395 SDValue Op2 = getValue(I.getArgOperand(1));
5397 SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1);
5398 setValue(&I, DAG.getNode(Op, sdl, VTs, Op1, Op2));
5401 case Intrinsic::prefetch: {
5403 unsigned rw = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
5405 Ops[1] = getValue(I.getArgOperand(0));
5406 Ops[2] = getValue(I.getArgOperand(1));
5407 Ops[3] = getValue(I.getArgOperand(2));
5408 Ops[4] = getValue(I.getArgOperand(3));
5409 DAG.setRoot(DAG.getMemIntrinsicNode(ISD::PREFETCH, sdl,
5410 DAG.getVTList(MVT::Other), Ops,
5411 EVT::getIntegerVT(*Context, 8),
5412 MachinePointerInfo(I.getArgOperand(0)),
5414 false, /* volatile */
5416 rw==1)); /* write */
5419 case Intrinsic::lifetime_start:
5420 case Intrinsic::lifetime_end: {
5421 bool IsStart = (Intrinsic == Intrinsic::lifetime_start);
5422 // Stack coloring is not enabled in O0, discard region information.
5423 if (TM.getOptLevel() == CodeGenOpt::None)
5426 SmallVector<Value *, 4> Allocas;
5427 GetUnderlyingObjects(I.getArgOperand(1), Allocas, DL);
5429 for (SmallVectorImpl<Value*>::iterator Object = Allocas.begin(),
5430 E = Allocas.end(); Object != E; ++Object) {
5431 AllocaInst *LifetimeObject = dyn_cast_or_null<AllocaInst>(*Object);
5433 // Could not find an Alloca.
5434 if (!LifetimeObject)
5437 int FI = FuncInfo.StaticAllocaMap[LifetimeObject];
5441 Ops[1] = DAG.getFrameIndex(FI, TLI->getPointerTy(), true);
5442 unsigned Opcode = (IsStart ? ISD::LIFETIME_START : ISD::LIFETIME_END);
5444 Res = DAG.getNode(Opcode, sdl, MVT::Other, Ops);
5449 case Intrinsic::invariant_start:
5450 // Discard region information.
5451 setValue(&I, DAG.getUNDEF(TLI->getPointerTy()));
5453 case Intrinsic::invariant_end:
5454 // Discard region information.
5456 case Intrinsic::stackprotectorcheck: {
5457 // Do not actually emit anything for this basic block. Instead we initialize
5458 // the stack protector descriptor and export the guard variable so we can
5459 // access it in FinishBasicBlock.
5460 const BasicBlock *BB = I.getParent();
5461 SPDescriptor.initialize(BB, FuncInfo.MBBMap[BB], I);
5462 ExportFromCurrentBlock(SPDescriptor.getGuard());
5464 // Flush our exports since we are going to process a terminator.
5465 (void)getControlRoot();
5468 case Intrinsic::clear_cache:
5469 return TLI->getClearCacheBuiltinName();
5470 case Intrinsic::donothing:
5473 case Intrinsic::experimental_stackmap: {
5477 case Intrinsic::experimental_patchpoint_void:
5478 case Intrinsic::experimental_patchpoint_i64: {
5485 void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee,
5487 MachineBasicBlock *LandingPad) {
5488 const TargetLowering *TLI = TM.getTargetLowering();
5489 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
5490 FunctionType *FTy = cast<FunctionType>(PT->getElementType());
5491 Type *RetTy = FTy->getReturnType();
5492 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5493 MCSymbol *BeginLabel = nullptr;
5495 TargetLowering::ArgListTy Args;
5496 TargetLowering::ArgListEntry Entry;
5497 Args.reserve(CS.arg_size());
5499 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
5501 const Value *V = *i;
5504 if (V->getType()->isEmptyTy())
5507 SDValue ArgNode = getValue(V);
5508 Entry.Node = ArgNode; Entry.Ty = V->getType();
5510 // Skip the first return-type Attribute to get to params.
5511 Entry.setAttributes(&CS, i - CS.arg_begin() + 1);
5512 Args.push_back(Entry);
5516 // Insert a label before the invoke call to mark the try range. This can be
5517 // used to detect deletion of the invoke via the MachineModuleInfo.
5518 BeginLabel = MMI.getContext().CreateTempSymbol();
5520 // For SjLj, keep track of which landing pads go with which invokes
5521 // so as to maintain the ordering of pads in the LSDA.
5522 unsigned CallSiteIndex = MMI.getCurrentCallSite();
5523 if (CallSiteIndex) {
5524 MMI.setCallSiteBeginLabel(BeginLabel, CallSiteIndex);
5525 LPadToCallSiteMap[LandingPad].push_back(CallSiteIndex);
5527 // Now that the call site is handled, stop tracking it.
5528 MMI.setCurrentCallSite(0);
5531 // Both PendingLoads and PendingExports must be flushed here;
5532 // this call might not return.
5534 DAG.setRoot(DAG.getEHLabel(getCurSDLoc(), getControlRoot(), BeginLabel));
5537 // Check if target-independent constraints permit a tail call here.
5538 // Target-dependent constraints are checked within TLI->LowerCallTo.
5539 if (isTailCall && !isInTailCallPosition(CS, DAG.getTarget()))
5542 TargetLowering::CallLoweringInfo CLI(DAG);
5543 CLI.setDebugLoc(getCurSDLoc()).setChain(getRoot())
5544 .setCallee(RetTy, FTy, Callee, std::move(Args), CS).setTailCall(isTailCall);
5546 std::pair<SDValue,SDValue> Result = TLI->LowerCallTo(CLI);
5547 assert((isTailCall || Result.second.getNode()) &&
5548 "Non-null chain expected with non-tail call!");
5549 assert((Result.second.getNode() || !Result.first.getNode()) &&
5550 "Null value expected with tail call!");
5551 if (Result.first.getNode())
5552 setValue(CS.getInstruction(), Result.first);
5554 if (!Result.second.getNode()) {
5555 // As a special case, a null chain means that a tail call has been emitted
5556 // and the DAG root is already updated.
5559 // Since there's no actual continuation from this block, nothing can be
5560 // relying on us setting vregs for them.
5561 PendingExports.clear();
5563 DAG.setRoot(Result.second);
5567 // Insert a label at the end of the invoke call to mark the try range. This
5568 // can be used to detect deletion of the invoke via the MachineModuleInfo.
5569 MCSymbol *EndLabel = MMI.getContext().CreateTempSymbol();
5570 DAG.setRoot(DAG.getEHLabel(getCurSDLoc(), getRoot(), EndLabel));
5572 // Inform MachineModuleInfo of range.
5573 MMI.addInvoke(LandingPad, BeginLabel, EndLabel);
5577 /// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the
5578 /// value is equal or not-equal to zero.
5579 static bool IsOnlyUsedInZeroEqualityComparison(const Value *V) {
5580 for (const User *U : V->users()) {
5581 if (const ICmpInst *IC = dyn_cast<ICmpInst>(U))
5582 if (IC->isEquality())
5583 if (const Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
5584 if (C->isNullValue())
5586 // Unknown instruction.
5592 static SDValue getMemCmpLoad(const Value *PtrVal, MVT LoadVT,
5594 SelectionDAGBuilder &Builder) {
5596 // Check to see if this load can be trivially constant folded, e.g. if the
5597 // input is from a string literal.
5598 if (const Constant *LoadInput = dyn_cast<Constant>(PtrVal)) {
5599 // Cast pointer to the type we really want to load.
5600 LoadInput = ConstantExpr::getBitCast(const_cast<Constant *>(LoadInput),
5601 PointerType::getUnqual(LoadTy));
5603 if (const Constant *LoadCst =
5604 ConstantFoldLoadFromConstPtr(const_cast<Constant *>(LoadInput),
5606 return Builder.getValue(LoadCst);
5609 // Otherwise, we have to emit the load. If the pointer is to unfoldable but
5610 // still constant memory, the input chain can be the entry node.
5612 bool ConstantMemory = false;
5614 // Do not serialize (non-volatile) loads of constant memory with anything.
5615 if (Builder.AA->pointsToConstantMemory(PtrVal)) {
5616 Root = Builder.DAG.getEntryNode();
5617 ConstantMemory = true;
5619 // Do not serialize non-volatile loads against each other.
5620 Root = Builder.DAG.getRoot();
5623 SDValue Ptr = Builder.getValue(PtrVal);
5624 SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurSDLoc(), Root,
5625 Ptr, MachinePointerInfo(PtrVal),
5627 false /*nontemporal*/,
5628 false /*isinvariant*/, 1 /* align=1 */);
5630 if (!ConstantMemory)
5631 Builder.PendingLoads.push_back(LoadVal.getValue(1));
5635 /// processIntegerCallValue - Record the value for an instruction that
5636 /// produces an integer result, converting the type where necessary.
5637 void SelectionDAGBuilder::processIntegerCallValue(const Instruction &I,
5640 EVT VT = TM.getTargetLowering()->getValueType(I.getType(), true);
5642 Value = DAG.getSExtOrTrunc(Value, getCurSDLoc(), VT);
5644 Value = DAG.getZExtOrTrunc(Value, getCurSDLoc(), VT);
5645 setValue(&I, Value);
5648 /// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form.
5649 /// If so, return true and lower it, otherwise return false and it will be
5650 /// lowered like a normal call.
5651 bool SelectionDAGBuilder::visitMemCmpCall(const CallInst &I) {
5652 // Verify that the prototype makes sense. int memcmp(void*,void*,size_t)
5653 if (I.getNumArgOperands() != 3)
5656 const Value *LHS = I.getArgOperand(0), *RHS = I.getArgOperand(1);
5657 if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() ||
5658 !I.getArgOperand(2)->getType()->isIntegerTy() ||
5659 !I.getType()->isIntegerTy())
5662 const Value *Size = I.getArgOperand(2);
5663 const ConstantInt *CSize = dyn_cast<ConstantInt>(Size);
5664 if (CSize && CSize->getZExtValue() == 0) {
5665 EVT CallVT = TM.getTargetLowering()->getValueType(I.getType(), true);
5666 setValue(&I, DAG.getConstant(0, CallVT));
5670 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5671 std::pair<SDValue, SDValue> Res =
5672 TSI.EmitTargetCodeForMemcmp(DAG, getCurSDLoc(), DAG.getRoot(),
5673 getValue(LHS), getValue(RHS), getValue(Size),
5674 MachinePointerInfo(LHS),
5675 MachinePointerInfo(RHS));
5676 if (Res.first.getNode()) {
5677 processIntegerCallValue(I, Res.first, true);
5678 PendingLoads.push_back(Res.second);
5682 // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0
5683 // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0
5684 if (CSize && IsOnlyUsedInZeroEqualityComparison(&I)) {
5685 bool ActuallyDoIt = true;
5688 switch (CSize->getZExtValue()) {
5690 LoadVT = MVT::Other;
5692 ActuallyDoIt = false;
5696 LoadTy = Type::getInt16Ty(CSize->getContext());
5700 LoadTy = Type::getInt32Ty(CSize->getContext());
5704 LoadTy = Type::getInt64Ty(CSize->getContext());
5708 LoadVT = MVT::v4i32;
5709 LoadTy = Type::getInt32Ty(CSize->getContext());
5710 LoadTy = VectorType::get(LoadTy, 4);
5715 // This turns into unaligned loads. We only do this if the target natively
5716 // supports the MVT we'll be loading or if it is small enough (<= 4) that
5717 // we'll only produce a small number of byte loads.
5719 // Require that we can find a legal MVT, and only do this if the target
5720 // supports unaligned loads of that type. Expanding into byte loads would
5722 const TargetLowering *TLI = TM.getTargetLowering();
5723 if (ActuallyDoIt && CSize->getZExtValue() > 4) {
5724 unsigned DstAS = LHS->getType()->getPointerAddressSpace();
5725 unsigned SrcAS = RHS->getType()->getPointerAddressSpace();
5726 // TODO: Handle 5 byte compare as 4-byte + 1 byte.
5727 // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads.
5728 // TODO: Check alignment of src and dest ptrs.
5729 if (!TLI->isTypeLegal(LoadVT) ||
5730 !TLI->allowsMisalignedMemoryAccesses(LoadVT, SrcAS) ||
5731 !TLI->allowsMisalignedMemoryAccesses(LoadVT, DstAS))
5732 ActuallyDoIt = false;
5736 SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this);
5737 SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this);
5739 SDValue Res = DAG.getSetCC(getCurSDLoc(), MVT::i1, LHSVal, RHSVal,
5741 processIntegerCallValue(I, Res, false);
5750 /// visitMemChrCall -- See if we can lower a memchr call into an optimized
5751 /// form. If so, return true and lower it, otherwise return false and it
5752 /// will be lowered like a normal call.
5753 bool SelectionDAGBuilder::visitMemChrCall(const CallInst &I) {
5754 // Verify that the prototype makes sense. void *memchr(void *, int, size_t)
5755 if (I.getNumArgOperands() != 3)
5758 const Value *Src = I.getArgOperand(0);
5759 const Value *Char = I.getArgOperand(1);
5760 const Value *Length = I.getArgOperand(2);
5761 if (!Src->getType()->isPointerTy() ||
5762 !Char->getType()->isIntegerTy() ||
5763 !Length->getType()->isIntegerTy() ||
5764 !I.getType()->isPointerTy())
5767 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5768 std::pair<SDValue, SDValue> Res =
5769 TSI.EmitTargetCodeForMemchr(DAG, getCurSDLoc(), DAG.getRoot(),
5770 getValue(Src), getValue(Char), getValue(Length),
5771 MachinePointerInfo(Src));
5772 if (Res.first.getNode()) {
5773 setValue(&I, Res.first);
5774 PendingLoads.push_back(Res.second);
5781 /// visitStrCpyCall -- See if we can lower a strcpy or stpcpy call into an
5782 /// optimized form. If so, return true and lower it, otherwise return false
5783 /// and it will be lowered like a normal call.
5784 bool SelectionDAGBuilder::visitStrCpyCall(const CallInst &I, bool isStpcpy) {
5785 // Verify that the prototype makes sense. char *strcpy(char *, char *)
5786 if (I.getNumArgOperands() != 2)
5789 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5790 if (!Arg0->getType()->isPointerTy() ||
5791 !Arg1->getType()->isPointerTy() ||
5792 !I.getType()->isPointerTy())
5795 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5796 std::pair<SDValue, SDValue> Res =
5797 TSI.EmitTargetCodeForStrcpy(DAG, getCurSDLoc(), getRoot(),
5798 getValue(Arg0), getValue(Arg1),
5799 MachinePointerInfo(Arg0),
5800 MachinePointerInfo(Arg1), isStpcpy);
5801 if (Res.first.getNode()) {
5802 setValue(&I, Res.first);
5803 DAG.setRoot(Res.second);
5810 /// visitStrCmpCall - See if we can lower a call to strcmp in an optimized form.
5811 /// If so, return true and lower it, otherwise return false and it will be
5812 /// lowered like a normal call.
5813 bool SelectionDAGBuilder::visitStrCmpCall(const CallInst &I) {
5814 // Verify that the prototype makes sense. int strcmp(void*,void*)
5815 if (I.getNumArgOperands() != 2)
5818 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5819 if (!Arg0->getType()->isPointerTy() ||
5820 !Arg1->getType()->isPointerTy() ||
5821 !I.getType()->isIntegerTy())
5824 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5825 std::pair<SDValue, SDValue> Res =
5826 TSI.EmitTargetCodeForStrcmp(DAG, getCurSDLoc(), DAG.getRoot(),
5827 getValue(Arg0), getValue(Arg1),
5828 MachinePointerInfo(Arg0),
5829 MachinePointerInfo(Arg1));
5830 if (Res.first.getNode()) {
5831 processIntegerCallValue(I, Res.first, true);
5832 PendingLoads.push_back(Res.second);
5839 /// visitStrLenCall -- See if we can lower a strlen call into an optimized
5840 /// form. If so, return true and lower it, otherwise return false and it
5841 /// will be lowered like a normal call.
5842 bool SelectionDAGBuilder::visitStrLenCall(const CallInst &I) {
5843 // Verify that the prototype makes sense. size_t strlen(char *)
5844 if (I.getNumArgOperands() != 1)
5847 const Value *Arg0 = I.getArgOperand(0);
5848 if (!Arg0->getType()->isPointerTy() || !I.getType()->isIntegerTy())
5851 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5852 std::pair<SDValue, SDValue> Res =
5853 TSI.EmitTargetCodeForStrlen(DAG, getCurSDLoc(), DAG.getRoot(),
5854 getValue(Arg0), MachinePointerInfo(Arg0));
5855 if (Res.first.getNode()) {
5856 processIntegerCallValue(I, Res.first, false);
5857 PendingLoads.push_back(Res.second);
5864 /// visitStrNLenCall -- See if we can lower a strnlen call into an optimized
5865 /// form. If so, return true and lower it, otherwise return false and it
5866 /// will be lowered like a normal call.
5867 bool SelectionDAGBuilder::visitStrNLenCall(const CallInst &I) {
5868 // Verify that the prototype makes sense. size_t strnlen(char *, size_t)
5869 if (I.getNumArgOperands() != 2)
5872 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5873 if (!Arg0->getType()->isPointerTy() ||
5874 !Arg1->getType()->isIntegerTy() ||
5875 !I.getType()->isIntegerTy())
5878 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5879 std::pair<SDValue, SDValue> Res =
5880 TSI.EmitTargetCodeForStrnlen(DAG, getCurSDLoc(), DAG.getRoot(),
5881 getValue(Arg0), getValue(Arg1),
5882 MachinePointerInfo(Arg0));
5883 if (Res.first.getNode()) {
5884 processIntegerCallValue(I, Res.first, false);
5885 PendingLoads.push_back(Res.second);
5892 /// visitUnaryFloatCall - If a call instruction is a unary floating-point
5893 /// operation (as expected), translate it to an SDNode with the specified opcode
5894 /// and return true.
5895 bool SelectionDAGBuilder::visitUnaryFloatCall(const CallInst &I,
5897 // Sanity check that it really is a unary floating-point call.
5898 if (I.getNumArgOperands() != 1 ||
5899 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
5900 I.getType() != I.getArgOperand(0)->getType() ||
5901 !I.onlyReadsMemory())
5904 SDValue Tmp = getValue(I.getArgOperand(0));
5905 setValue(&I, DAG.getNode(Opcode, getCurSDLoc(), Tmp.getValueType(), Tmp));
5909 void SelectionDAGBuilder::visitCall(const CallInst &I) {
5910 // Handle inline assembly differently.
5911 if (isa<InlineAsm>(I.getCalledValue())) {
5916 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5917 ComputeUsesVAFloatArgument(I, &MMI);
5919 const char *RenameFn = nullptr;
5920 if (Function *F = I.getCalledFunction()) {
5921 if (F->isDeclaration()) {
5922 if (const TargetIntrinsicInfo *II = TM.getIntrinsicInfo()) {
5923 if (unsigned IID = II->getIntrinsicID(F)) {
5924 RenameFn = visitIntrinsicCall(I, IID);
5929 if (unsigned IID = F->getIntrinsicID()) {
5930 RenameFn = visitIntrinsicCall(I, IID);
5936 // Check for well-known libc/libm calls. If the function is internal, it
5937 // can't be a library call.
5939 if (!F->hasLocalLinkage() && F->hasName() &&
5940 LibInfo->getLibFunc(F->getName(), Func) &&
5941 LibInfo->hasOptimizedCodeGen(Func)) {
5944 case LibFunc::copysign:
5945 case LibFunc::copysignf:
5946 case LibFunc::copysignl:
5947 if (I.getNumArgOperands() == 2 && // Basic sanity checks.
5948 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5949 I.getType() == I.getArgOperand(0)->getType() &&
5950 I.getType() == I.getArgOperand(1)->getType() &&
5951 I.onlyReadsMemory()) {
5952 SDValue LHS = getValue(I.getArgOperand(0));
5953 SDValue RHS = getValue(I.getArgOperand(1));
5954 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurSDLoc(),
5955 LHS.getValueType(), LHS, RHS));
5960 case LibFunc::fabsf:
5961 case LibFunc::fabsl:
5962 if (visitUnaryFloatCall(I, ISD::FABS))
5968 if (visitUnaryFloatCall(I, ISD::FSIN))
5974 if (visitUnaryFloatCall(I, ISD::FCOS))
5978 case LibFunc::sqrtf:
5979 case LibFunc::sqrtl:
5980 case LibFunc::sqrt_finite:
5981 case LibFunc::sqrtf_finite:
5982 case LibFunc::sqrtl_finite:
5983 if (visitUnaryFloatCall(I, ISD::FSQRT))
5986 case LibFunc::floor:
5987 case LibFunc::floorf:
5988 case LibFunc::floorl:
5989 if (visitUnaryFloatCall(I, ISD::FFLOOR))
5992 case LibFunc::nearbyint:
5993 case LibFunc::nearbyintf:
5994 case LibFunc::nearbyintl:
5995 if (visitUnaryFloatCall(I, ISD::FNEARBYINT))
5999 case LibFunc::ceilf:
6000 case LibFunc::ceill:
6001 if (visitUnaryFloatCall(I, ISD::FCEIL))
6005 case LibFunc::rintf:
6006 case LibFunc::rintl:
6007 if (visitUnaryFloatCall(I, ISD::FRINT))
6010 case LibFunc::round:
6011 case LibFunc::roundf:
6012 case LibFunc::roundl:
6013 if (visitUnaryFloatCall(I, ISD::FROUND))
6016 case LibFunc::trunc:
6017 case LibFunc::truncf:
6018 case LibFunc::truncl:
6019 if (visitUnaryFloatCall(I, ISD::FTRUNC))
6023 case LibFunc::log2f:
6024 case LibFunc::log2l:
6025 if (visitUnaryFloatCall(I, ISD::FLOG2))
6029 case LibFunc::exp2f:
6030 case LibFunc::exp2l:
6031 if (visitUnaryFloatCall(I, ISD::FEXP2))
6034 case LibFunc::memcmp:
6035 if (visitMemCmpCall(I))
6038 case LibFunc::memchr:
6039 if (visitMemChrCall(I))
6042 case LibFunc::strcpy:
6043 if (visitStrCpyCall(I, false))
6046 case LibFunc::stpcpy:
6047 if (visitStrCpyCall(I, true))
6050 case LibFunc::strcmp:
6051 if (visitStrCmpCall(I))
6054 case LibFunc::strlen:
6055 if (visitStrLenCall(I))
6058 case LibFunc::strnlen:
6059 if (visitStrNLenCall(I))
6068 Callee = getValue(I.getCalledValue());
6070 Callee = DAG.getExternalSymbol(RenameFn,
6071 TM.getTargetLowering()->getPointerTy());
6073 // Check if we can potentially perform a tail call. More detailed checking is
6074 // be done within LowerCallTo, after more information about the call is known.
6075 LowerCallTo(&I, Callee, I.isTailCall());
6080 /// AsmOperandInfo - This contains information for each constraint that we are
6082 class SDISelAsmOperandInfo : public TargetLowering::AsmOperandInfo {
6084 /// CallOperand - If this is the result output operand or a clobber
6085 /// this is null, otherwise it is the incoming operand to the CallInst.
6086 /// This gets modified as the asm is processed.
6087 SDValue CallOperand;
6089 /// AssignedRegs - If this is a register or register class operand, this
6090 /// contains the set of register corresponding to the operand.
6091 RegsForValue AssignedRegs;
6093 explicit SDISelAsmOperandInfo(const TargetLowering::AsmOperandInfo &info)
6094 : TargetLowering::AsmOperandInfo(info), CallOperand(nullptr,0) {
6097 /// getCallOperandValEVT - Return the EVT of the Value* that this operand
6098 /// corresponds to. If there is no Value* for this operand, it returns
6100 EVT getCallOperandValEVT(LLVMContext &Context,
6101 const TargetLowering &TLI,
6102 const DataLayout *DL) const {
6103 if (!CallOperandVal) return MVT::Other;
6105 if (isa<BasicBlock>(CallOperandVal))
6106 return TLI.getPointerTy();
6108 llvm::Type *OpTy = CallOperandVal->getType();
6110 // FIXME: code duplicated from TargetLowering::ParseConstraints().
6111 // If this is an indirect operand, the operand is a pointer to the
6114 llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
6116 report_fatal_error("Indirect operand for inline asm not a pointer!");
6117 OpTy = PtrTy->getElementType();
6120 // Look for vector wrapped in a struct. e.g. { <16 x i8> }.
6121 if (StructType *STy = dyn_cast<StructType>(OpTy))
6122 if (STy->getNumElements() == 1)
6123 OpTy = STy->getElementType(0);
6125 // If OpTy is not a single value, it may be a struct/union that we
6126 // can tile with integers.
6127 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
6128 unsigned BitSize = DL->getTypeSizeInBits(OpTy);
6137 OpTy = IntegerType::get(Context, BitSize);
6142 return TLI.getValueType(OpTy, true);
6146 typedef SmallVector<SDISelAsmOperandInfo,16> SDISelAsmOperandInfoVector;
6148 } // end anonymous namespace
6150 /// GetRegistersForValue - Assign registers (virtual or physical) for the
6151 /// specified operand. We prefer to assign virtual registers, to allow the
6152 /// register allocator to handle the assignment process. However, if the asm
6153 /// uses features that we can't model on machineinstrs, we have SDISel do the
6154 /// allocation. This produces generally horrible, but correct, code.
6156 /// OpInfo describes the operand.
6158 static void GetRegistersForValue(SelectionDAG &DAG,
6159 const TargetLowering &TLI,
6161 SDISelAsmOperandInfo &OpInfo) {
6162 LLVMContext &Context = *DAG.getContext();
6164 MachineFunction &MF = DAG.getMachineFunction();
6165 SmallVector<unsigned, 4> Regs;
6167 // If this is a constraint for a single physreg, or a constraint for a
6168 // register class, find it.
6169 std::pair<unsigned, const TargetRegisterClass*> PhysReg =
6170 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
6171 OpInfo.ConstraintVT);
6173 unsigned NumRegs = 1;
6174 if (OpInfo.ConstraintVT != MVT::Other) {
6175 // If this is a FP input in an integer register (or visa versa) insert a bit
6176 // cast of the input value. More generally, handle any case where the input
6177 // value disagrees with the register class we plan to stick this in.
6178 if (OpInfo.Type == InlineAsm::isInput &&
6179 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) {
6180 // Try to convert to the first EVT that the reg class contains. If the
6181 // types are identical size, use a bitcast to convert (e.g. two differing
6183 MVT RegVT = *PhysReg.second->vt_begin();
6184 if (RegVT.getSizeInBits() == OpInfo.CallOperand.getValueSizeInBits()) {
6185 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
6186 RegVT, OpInfo.CallOperand);
6187 OpInfo.ConstraintVT = RegVT;
6188 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) {
6189 // If the input is a FP value and we want it in FP registers, do a
6190 // bitcast to the corresponding integer type. This turns an f64 value
6191 // into i64, which can be passed with two i32 values on a 32-bit
6193 RegVT = MVT::getIntegerVT(OpInfo.ConstraintVT.getSizeInBits());
6194 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
6195 RegVT, OpInfo.CallOperand);
6196 OpInfo.ConstraintVT = RegVT;
6200 NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT);
6204 EVT ValueVT = OpInfo.ConstraintVT;
6206 // If this is a constraint for a specific physical register, like {r17},
6208 if (unsigned AssignedReg = PhysReg.first) {
6209 const TargetRegisterClass *RC = PhysReg.second;
6210 if (OpInfo.ConstraintVT == MVT::Other)
6211 ValueVT = *RC->vt_begin();
6213 // Get the actual register value type. This is important, because the user
6214 // may have asked for (e.g.) the AX register in i32 type. We need to
6215 // remember that AX is actually i16 to get the right extension.
6216 RegVT = *RC->vt_begin();
6218 // This is a explicit reference to a physical register.
6219 Regs.push_back(AssignedReg);
6221 // If this is an expanded reference, add the rest of the regs to Regs.
6223 TargetRegisterClass::iterator I = RC->begin();
6224 for (; *I != AssignedReg; ++I)
6225 assert(I != RC->end() && "Didn't find reg!");
6227 // Already added the first reg.
6229 for (; NumRegs; --NumRegs, ++I) {
6230 assert(I != RC->end() && "Ran out of registers to allocate!");
6235 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
6239 // Otherwise, if this was a reference to an LLVM register class, create vregs
6240 // for this reference.
6241 if (const TargetRegisterClass *RC = PhysReg.second) {
6242 RegVT = *RC->vt_begin();
6243 if (OpInfo.ConstraintVT == MVT::Other)
6246 // Create the appropriate number of virtual registers.
6247 MachineRegisterInfo &RegInfo = MF.getRegInfo();
6248 for (; NumRegs; --NumRegs)
6249 Regs.push_back(RegInfo.createVirtualRegister(RC));
6251 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
6255 // Otherwise, we couldn't allocate enough registers for this.
6258 /// visitInlineAsm - Handle a call to an InlineAsm object.
6260 void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) {
6261 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
6263 /// ConstraintOperands - Information about all of the constraints.
6264 SDISelAsmOperandInfoVector ConstraintOperands;
6266 const TargetLowering *TLI = TM.getTargetLowering();
6267 TargetLowering::AsmOperandInfoVector
6268 TargetConstraints = TLI->ParseConstraints(CS);
6270 bool hasMemory = false;
6272 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
6273 unsigned ResNo = 0; // ResNo - The result number of the next output.
6274 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
6275 ConstraintOperands.push_back(SDISelAsmOperandInfo(TargetConstraints[i]));
6276 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
6278 MVT OpVT = MVT::Other;
6280 // Compute the value type for each operand.
6281 switch (OpInfo.Type) {
6282 case InlineAsm::isOutput:
6283 // Indirect outputs just consume an argument.
6284 if (OpInfo.isIndirect) {
6285 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
6289 // The return value of the call is this value. As such, there is no
6290 // corresponding argument.
6291 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
6292 if (StructType *STy = dyn_cast<StructType>(CS.getType())) {
6293 OpVT = TLI->getSimpleValueType(STy->getElementType(ResNo));
6295 assert(ResNo == 0 && "Asm only has one result!");
6296 OpVT = TLI->getSimpleValueType(CS.getType());
6300 case InlineAsm::isInput:
6301 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
6303 case InlineAsm::isClobber:
6308 // If this is an input or an indirect output, process the call argument.
6309 // BasicBlocks are labels, currently appearing only in asm's.
6310 if (OpInfo.CallOperandVal) {
6311 if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) {
6312 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]);
6314 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
6317 OpVT = OpInfo.getCallOperandValEVT(*DAG.getContext(), *TLI, DL).
6321 OpInfo.ConstraintVT = OpVT;
6323 // Indirect operand accesses access memory.
6324 if (OpInfo.isIndirect)
6327 for (unsigned j = 0, ee = OpInfo.Codes.size(); j != ee; ++j) {
6328 TargetLowering::ConstraintType
6329 CType = TLI->getConstraintType(OpInfo.Codes[j]);
6330 if (CType == TargetLowering::C_Memory) {
6338 SDValue Chain, Flag;
6340 // We won't need to flush pending loads if this asm doesn't touch
6341 // memory and is nonvolatile.
6342 if (hasMemory || IA->hasSideEffects())
6345 Chain = DAG.getRoot();
6347 // Second pass over the constraints: compute which constraint option to use
6348 // and assign registers to constraints that want a specific physreg.
6349 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6350 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6352 // If this is an output operand with a matching input operand, look up the
6353 // matching input. If their types mismatch, e.g. one is an integer, the
6354 // other is floating point, or their sizes are different, flag it as an
6356 if (OpInfo.hasMatchingInput()) {
6357 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
6359 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
6360 std::pair<unsigned, const TargetRegisterClass*> MatchRC =
6361 TLI->getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
6362 OpInfo.ConstraintVT);
6363 std::pair<unsigned, const TargetRegisterClass*> InputRC =
6364 TLI->getRegForInlineAsmConstraint(Input.ConstraintCode,
6365 Input.ConstraintVT);
6366 if ((OpInfo.ConstraintVT.isInteger() !=
6367 Input.ConstraintVT.isInteger()) ||
6368 (MatchRC.second != InputRC.second)) {
6369 report_fatal_error("Unsupported asm: input constraint"
6370 " with a matching output constraint of"
6371 " incompatible type!");
6373 Input.ConstraintVT = OpInfo.ConstraintVT;
6377 // Compute the constraint code and ConstraintType to use.
6378 TLI->ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, &DAG);
6380 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
6381 OpInfo.Type == InlineAsm::isClobber)
6384 // If this is a memory input, and if the operand is not indirect, do what we
6385 // need to to provide an address for the memory input.
6386 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
6387 !OpInfo.isIndirect) {
6388 assert((OpInfo.isMultipleAlternative ||
6389 (OpInfo.Type == InlineAsm::isInput)) &&
6390 "Can only indirectify direct input operands!");
6392 // Memory operands really want the address of the value. If we don't have
6393 // an indirect input, put it in the constpool if we can, otherwise spill
6394 // it to a stack slot.
6395 // TODO: This isn't quite right. We need to handle these according to
6396 // the addressing mode that the constraint wants. Also, this may take
6397 // an additional register for the computation and we don't want that
6400 // If the operand is a float, integer, or vector constant, spill to a
6401 // constant pool entry to get its address.
6402 const Value *OpVal = OpInfo.CallOperandVal;
6403 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
6404 isa<ConstantVector>(OpVal) || isa<ConstantDataVector>(OpVal)) {
6405 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal),
6406 TLI->getPointerTy());
6408 // Otherwise, create a stack slot and emit a store to it before the
6410 Type *Ty = OpVal->getType();
6411 uint64_t TySize = TLI->getDataLayout()->getTypeAllocSize(Ty);
6412 unsigned Align = TLI->getDataLayout()->getPrefTypeAlignment(Ty);
6413 MachineFunction &MF = DAG.getMachineFunction();
6414 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
6415 SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI->getPointerTy());
6416 Chain = DAG.getStore(Chain, getCurSDLoc(),
6417 OpInfo.CallOperand, StackSlot,
6418 MachinePointerInfo::getFixedStack(SSFI),
6420 OpInfo.CallOperand = StackSlot;
6423 // There is no longer a Value* corresponding to this operand.
6424 OpInfo.CallOperandVal = nullptr;
6426 // It is now an indirect operand.
6427 OpInfo.isIndirect = true;
6430 // If this constraint is for a specific register, allocate it before
6432 if (OpInfo.ConstraintType == TargetLowering::C_Register)
6433 GetRegistersForValue(DAG, *TLI, getCurSDLoc(), OpInfo);
6436 // Second pass - Loop over all of the operands, assigning virtual or physregs
6437 // to register class operands.
6438 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6439 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6441 // C_Register operands have already been allocated, Other/Memory don't need
6443 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
6444 GetRegistersForValue(DAG, *TLI, getCurSDLoc(), OpInfo);
6447 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
6448 std::vector<SDValue> AsmNodeOperands;
6449 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain
6450 AsmNodeOperands.push_back(
6451 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(),
6452 TLI->getPointerTy()));
6454 // If we have a !srcloc metadata node associated with it, we want to attach
6455 // this to the ultimately generated inline asm machineinstr. To do this, we
6456 // pass in the third operand as this (potentially null) inline asm MDNode.
6457 const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc");
6458 AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc));
6460 // Remember the HasSideEffect, AlignStack, AsmDialect, MayLoad and MayStore
6461 // bits as operand 3.
6462 unsigned ExtraInfo = 0;
6463 if (IA->hasSideEffects())
6464 ExtraInfo |= InlineAsm::Extra_HasSideEffects;
6465 if (IA->isAlignStack())
6466 ExtraInfo |= InlineAsm::Extra_IsAlignStack;
6467 // Set the asm dialect.
6468 ExtraInfo |= IA->getDialect() * InlineAsm::Extra_AsmDialect;
6470 // Determine if this InlineAsm MayLoad or MayStore based on the constraints.
6471 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
6472 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
6474 // Compute the constraint code and ConstraintType to use.
6475 TLI->ComputeConstraintToUse(OpInfo, SDValue());
6477 // Ideally, we would only check against memory constraints. However, the
6478 // meaning of an other constraint can be target-specific and we can't easily
6479 // reason about it. Therefore, be conservative and set MayLoad/MayStore
6480 // for other constriants as well.
6481 if (OpInfo.ConstraintType == TargetLowering::C_Memory ||
6482 OpInfo.ConstraintType == TargetLowering::C_Other) {
6483 if (OpInfo.Type == InlineAsm::isInput)
6484 ExtraInfo |= InlineAsm::Extra_MayLoad;
6485 else if (OpInfo.Type == InlineAsm::isOutput)
6486 ExtraInfo |= InlineAsm::Extra_MayStore;
6487 else if (OpInfo.Type == InlineAsm::isClobber)
6488 ExtraInfo |= (InlineAsm::Extra_MayLoad | InlineAsm::Extra_MayStore);
6492 AsmNodeOperands.push_back(DAG.getTargetConstant(ExtraInfo,
6493 TLI->getPointerTy()));
6495 // Loop over all of the inputs, copying the operand values into the
6496 // appropriate registers and processing the output regs.
6497 RegsForValue RetValRegs;
6499 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
6500 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
6502 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6503 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6505 switch (OpInfo.Type) {
6506 case InlineAsm::isOutput: {
6507 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
6508 OpInfo.ConstraintType != TargetLowering::C_Register) {
6509 // Memory output, or 'other' output (e.g. 'X' constraint).
6510 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
6512 // Add information to the INLINEASM node to know about this output.
6513 unsigned OpFlags = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
6514 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlags,
6515 TLI->getPointerTy()));
6516 AsmNodeOperands.push_back(OpInfo.CallOperand);
6520 // Otherwise, this is a register or register class output.
6522 // Copy the output from the appropriate register. Find a register that
6524 if (OpInfo.AssignedRegs.Regs.empty()) {
6525 LLVMContext &Ctx = *DAG.getContext();
6526 Ctx.emitError(CS.getInstruction(),
6527 "couldn't allocate output register for constraint '" +
6528 Twine(OpInfo.ConstraintCode) + "'");
6532 // If this is an indirect operand, store through the pointer after the
6534 if (OpInfo.isIndirect) {
6535 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
6536 OpInfo.CallOperandVal));
6538 // This is the result value of the call.
6539 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
6540 // Concatenate this output onto the outputs list.
6541 RetValRegs.append(OpInfo.AssignedRegs);
6544 // Add information to the INLINEASM node to know that this register is
6547 .AddInlineAsmOperands(OpInfo.isEarlyClobber
6548 ? InlineAsm::Kind_RegDefEarlyClobber
6549 : InlineAsm::Kind_RegDef,
6550 false, 0, DAG, AsmNodeOperands);
6553 case InlineAsm::isInput: {
6554 SDValue InOperandVal = OpInfo.CallOperand;
6556 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint?
6557 // If this is required to match an output register we have already set,
6558 // just use its register.
6559 unsigned OperandNo = OpInfo.getMatchedOperand();
6561 // Scan until we find the definition we already emitted of this operand.
6562 // When we find it, create a RegsForValue operand.
6563 unsigned CurOp = InlineAsm::Op_FirstOperand;
6564 for (; OperandNo; --OperandNo) {
6565 // Advance to the next operand.
6567 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6568 assert((InlineAsm::isRegDefKind(OpFlag) ||
6569 InlineAsm::isRegDefEarlyClobberKind(OpFlag) ||
6570 InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?");
6571 CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1;
6575 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6576 if (InlineAsm::isRegDefKind(OpFlag) ||
6577 InlineAsm::isRegDefEarlyClobberKind(OpFlag)) {
6578 // Add (OpFlag&0xffff)>>3 registers to MatchedRegs.
6579 if (OpInfo.isIndirect) {
6580 // This happens on gcc/testsuite/gcc.dg/pr8788-1.c
6581 LLVMContext &Ctx = *DAG.getContext();
6582 Ctx.emitError(CS.getInstruction(), "inline asm not supported yet:"
6583 " don't know how to handle tied "
6584 "indirect register inputs");
6588 RegsForValue MatchedRegs;
6589 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
6590 MVT RegVT = AsmNodeOperands[CurOp+1].getSimpleValueType();
6591 MatchedRegs.RegVTs.push_back(RegVT);
6592 MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo();
6593 for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag);
6595 if (const TargetRegisterClass *RC = TLI->getRegClassFor(RegVT))
6596 MatchedRegs.Regs.push_back(RegInfo.createVirtualRegister(RC));
6598 LLVMContext &Ctx = *DAG.getContext();
6599 Ctx.emitError(CS.getInstruction(),
6600 "inline asm error: This value"
6601 " type register class is not natively supported!");
6605 // Use the produced MatchedRegs object to
6606 MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurSDLoc(),
6607 Chain, &Flag, CS.getInstruction());
6608 MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse,
6609 true, OpInfo.getMatchedOperand(),
6610 DAG, AsmNodeOperands);
6614 assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!");
6615 assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 &&
6616 "Unexpected number of operands");
6617 // Add information to the INLINEASM node to know about this input.
6618 // See InlineAsm.h isUseOperandTiedToDef.
6619 OpFlag = InlineAsm::getFlagWordForMatchingOp(OpFlag,
6620 OpInfo.getMatchedOperand());
6621 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlag,
6622 TLI->getPointerTy()));
6623 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]);
6627 // Treat indirect 'X' constraint as memory.
6628 if (OpInfo.ConstraintType == TargetLowering::C_Other &&
6630 OpInfo.ConstraintType = TargetLowering::C_Memory;
6632 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
6633 std::vector<SDValue> Ops;
6634 TLI->LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode,
6637 LLVMContext &Ctx = *DAG.getContext();
6638 Ctx.emitError(CS.getInstruction(),
6639 "invalid operand for inline asm constraint '" +
6640 Twine(OpInfo.ConstraintCode) + "'");
6644 // Add information to the INLINEASM node to know about this input.
6645 unsigned ResOpType =
6646 InlineAsm::getFlagWord(InlineAsm::Kind_Imm, Ops.size());
6647 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
6648 TLI->getPointerTy()));
6649 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
6653 if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
6654 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
6655 assert(InOperandVal.getValueType() == TLI->getPointerTy() &&
6656 "Memory operands expect pointer values");
6658 // Add information to the INLINEASM node to know about this input.
6659 unsigned ResOpType = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
6660 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
6661 TLI->getPointerTy()));
6662 AsmNodeOperands.push_back(InOperandVal);
6666 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
6667 OpInfo.ConstraintType == TargetLowering::C_Register) &&
6668 "Unknown constraint type!");
6670 // TODO: Support this.
6671 if (OpInfo.isIndirect) {
6672 LLVMContext &Ctx = *DAG.getContext();
6673 Ctx.emitError(CS.getInstruction(),
6674 "Don't know how to handle indirect register inputs yet "
6675 "for constraint '" +
6676 Twine(OpInfo.ConstraintCode) + "'");
6680 // Copy the input into the appropriate registers.
6681 if (OpInfo.AssignedRegs.Regs.empty()) {
6682 LLVMContext &Ctx = *DAG.getContext();
6683 Ctx.emitError(CS.getInstruction(),
6684 "couldn't allocate input reg for constraint '" +
6685 Twine(OpInfo.ConstraintCode) + "'");
6689 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, getCurSDLoc(),
6690 Chain, &Flag, CS.getInstruction());
6692 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, false, 0,
6693 DAG, AsmNodeOperands);
6696 case InlineAsm::isClobber: {
6697 // Add the clobbered value to the operand list, so that the register
6698 // allocator is aware that the physreg got clobbered.
6699 if (!OpInfo.AssignedRegs.Regs.empty())
6700 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_Clobber,
6708 // Finish up input operands. Set the input chain and add the flag last.
6709 AsmNodeOperands[InlineAsm::Op_InputChain] = Chain;
6710 if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
6712 Chain = DAG.getNode(ISD::INLINEASM, getCurSDLoc(),
6713 DAG.getVTList(MVT::Other, MVT::Glue), AsmNodeOperands);
6714 Flag = Chain.getValue(1);
6716 // If this asm returns a register value, copy the result from that register
6717 // and set it as the value of the call.
6718 if (!RetValRegs.Regs.empty()) {
6719 SDValue Val = RetValRegs.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(),
6720 Chain, &Flag, CS.getInstruction());
6722 // FIXME: Why don't we do this for inline asms with MRVs?
6723 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) {
6724 EVT ResultType = TLI->getValueType(CS.getType());
6726 // If any of the results of the inline asm is a vector, it may have the
6727 // wrong width/num elts. This can happen for register classes that can
6728 // contain multiple different value types. The preg or vreg allocated may
6729 // not have the same VT as was expected. Convert it to the right type
6730 // with bit_convert.
6731 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) {
6732 Val = DAG.getNode(ISD::BITCAST, getCurSDLoc(),
6735 } else if (ResultType != Val.getValueType() &&
6736 ResultType.isInteger() && Val.getValueType().isInteger()) {
6737 // If a result value was tied to an input value, the computed result may
6738 // have a wider width than the expected result. Extract the relevant
6740 Val = DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), ResultType, Val);
6743 assert(ResultType == Val.getValueType() && "Asm result value mismatch!");
6746 setValue(CS.getInstruction(), Val);
6747 // Don't need to use this as a chain in this case.
6748 if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty())
6752 std::vector<std::pair<SDValue, const Value *> > StoresToEmit;
6754 // Process indirect outputs, first output all of the flagged copies out of
6756 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
6757 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
6758 const Value *Ptr = IndirectStoresToEmit[i].second;
6759 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(),
6761 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
6764 // Emit the non-flagged stores from the physregs.
6765 SmallVector<SDValue, 8> OutChains;
6766 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) {
6767 SDValue Val = DAG.getStore(Chain, getCurSDLoc(),
6768 StoresToEmit[i].first,
6769 getValue(StoresToEmit[i].second),
6770 MachinePointerInfo(StoresToEmit[i].second),
6772 OutChains.push_back(Val);
6775 if (!OutChains.empty())
6776 Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other, OutChains);
6781 void SelectionDAGBuilder::visitVAStart(const CallInst &I) {
6782 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurSDLoc(),
6783 MVT::Other, getRoot(),
6784 getValue(I.getArgOperand(0)),
6785 DAG.getSrcValue(I.getArgOperand(0))));
6788 void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) {
6789 const TargetLowering *TLI = TM.getTargetLowering();
6790 const DataLayout &DL = *TLI->getDataLayout();
6791 SDValue V = DAG.getVAArg(TLI->getValueType(I.getType()), getCurSDLoc(),
6792 getRoot(), getValue(I.getOperand(0)),
6793 DAG.getSrcValue(I.getOperand(0)),
6794 DL.getABITypeAlignment(I.getType()));
6796 DAG.setRoot(V.getValue(1));
6799 void SelectionDAGBuilder::visitVAEnd(const CallInst &I) {
6800 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurSDLoc(),
6801 MVT::Other, getRoot(),
6802 getValue(I.getArgOperand(0)),
6803 DAG.getSrcValue(I.getArgOperand(0))));
6806 void SelectionDAGBuilder::visitVACopy(const CallInst &I) {
6807 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurSDLoc(),
6808 MVT::Other, getRoot(),
6809 getValue(I.getArgOperand(0)),
6810 getValue(I.getArgOperand(1)),
6811 DAG.getSrcValue(I.getArgOperand(0)),
6812 DAG.getSrcValue(I.getArgOperand(1))));
6815 /// \brief Lower an argument list according to the target calling convention.
6817 /// \return A tuple of <return-value, token-chain>
6819 /// This is a helper for lowering intrinsics that follow a target calling
6820 /// convention or require stack pointer adjustment. Only a subset of the
6821 /// intrinsic's operands need to participate in the calling convention.
6822 std::pair<SDValue, SDValue>
6823 SelectionDAGBuilder::LowerCallOperands(const CallInst &CI, unsigned ArgIdx,
6824 unsigned NumArgs, SDValue Callee,
6826 TargetLowering::ArgListTy Args;
6827 Args.reserve(NumArgs);
6829 // Populate the argument list.
6830 // Attributes for args start at offset 1, after the return attribute.
6831 ImmutableCallSite CS(&CI);
6832 for (unsigned ArgI = ArgIdx, ArgE = ArgIdx + NumArgs, AttrI = ArgIdx + 1;
6833 ArgI != ArgE; ++ArgI) {
6834 const Value *V = CI.getOperand(ArgI);
6836 assert(!V->getType()->isEmptyTy() && "Empty type passed to intrinsic.");
6838 TargetLowering::ArgListEntry Entry;
6839 Entry.Node = getValue(V);
6840 Entry.Ty = V->getType();
6841 Entry.setAttributes(&CS, AttrI);
6842 Args.push_back(Entry);
6845 Type *retTy = useVoidTy ? Type::getVoidTy(*DAG.getContext()) : CI.getType();
6846 TargetLowering::CallLoweringInfo CLI(DAG);
6847 CLI.setDebugLoc(getCurSDLoc()).setChain(getRoot())
6848 .setCallee(CI.getCallingConv(), retTy, Callee, std::move(Args), NumArgs)
6849 .setDiscardResult(!CI.use_empty());
6851 const TargetLowering *TLI = TM.getTargetLowering();
6852 return TLI->LowerCallTo(CLI);
6855 /// \brief Add a stack map intrinsic call's live variable operands to a stackmap
6856 /// or patchpoint target node's operand list.
6858 /// Constants are converted to TargetConstants purely as an optimization to
6859 /// avoid constant materialization and register allocation.
6861 /// FrameIndex operands are converted to TargetFrameIndex so that ISEL does not
6862 /// generate addess computation nodes, and so ExpandISelPseudo can convert the
6863 /// TargetFrameIndex into a DirectMemRefOp StackMap location. This avoids
6864 /// address materialization and register allocation, but may also be required
6865 /// for correctness. If a StackMap (or PatchPoint) intrinsic directly uses an
6866 /// alloca in the entry block, then the runtime may assume that the alloca's
6867 /// StackMap location can be read immediately after compilation and that the
6868 /// location is valid at any point during execution (this is similar to the
6869 /// assumption made by the llvm.gcroot intrinsic). If the alloca's location were
6870 /// only available in a register, then the runtime would need to trap when
6871 /// execution reaches the StackMap in order to read the alloca's location.
6872 static void addStackMapLiveVars(const CallInst &CI, unsigned StartIdx,
6873 SmallVectorImpl<SDValue> &Ops,
6874 SelectionDAGBuilder &Builder) {
6875 for (unsigned i = StartIdx, e = CI.getNumArgOperands(); i != e; ++i) {
6876 SDValue OpVal = Builder.getValue(CI.getArgOperand(i));
6877 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(OpVal)) {
6879 Builder.DAG.getTargetConstant(StackMaps::ConstantOp, MVT::i64));
6881 Builder.DAG.getTargetConstant(C->getSExtValue(), MVT::i64));
6882 } else if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(OpVal)) {
6883 const TargetLowering &TLI = Builder.DAG.getTargetLoweringInfo();
6885 Builder.DAG.getTargetFrameIndex(FI->getIndex(), TLI.getPointerTy()));
6887 Ops.push_back(OpVal);
6891 /// \brief Lower llvm.experimental.stackmap directly to its target opcode.
6892 void SelectionDAGBuilder::visitStackmap(const CallInst &CI) {
6893 // void @llvm.experimental.stackmap(i32 <id>, i32 <numShadowBytes>,
6894 // [live variables...])
6896 assert(CI.getType()->isVoidTy() && "Stackmap cannot return a value.");
6898 SDValue Chain, InFlag, Callee, NullPtr;
6899 SmallVector<SDValue, 32> Ops;
6901 SDLoc DL = getCurSDLoc();
6902 Callee = getValue(CI.getCalledValue());
6903 NullPtr = DAG.getIntPtrConstant(0, true);
6905 // The stackmap intrinsic only records the live variables (the arguemnts
6906 // passed to it) and emits NOPS (if requested). Unlike the patchpoint
6907 // intrinsic, this won't be lowered to a function call. This means we don't
6908 // have to worry about calling conventions and target specific lowering code.
6909 // Instead we perform the call lowering right here.
6911 // chain, flag = CALLSEQ_START(chain, 0)
6912 // chain, flag = STACKMAP(id, nbytes, ..., chain, flag)
6913 // chain, flag = CALLSEQ_END(chain, 0, 0, flag)
6915 Chain = DAG.getCALLSEQ_START(getRoot(), NullPtr, DL);
6916 InFlag = Chain.getValue(1);
6918 // Add the <id> and <numBytes> constants.
6919 SDValue IDVal = getValue(CI.getOperand(PatchPointOpers::IDPos));
6920 Ops.push_back(DAG.getTargetConstant(
6921 cast<ConstantSDNode>(IDVal)->getZExtValue(), MVT::i64));
6922 SDValue NBytesVal = getValue(CI.getOperand(PatchPointOpers::NBytesPos));
6923 Ops.push_back(DAG.getTargetConstant(
6924 cast<ConstantSDNode>(NBytesVal)->getZExtValue(), MVT::i32));
6926 // Push live variables for the stack map.
6927 addStackMapLiveVars(CI, 2, Ops, *this);
6929 // We are not pushing any register mask info here on the operands list,
6930 // because the stackmap doesn't clobber anything.
6932 // Push the chain and the glue flag.
6933 Ops.push_back(Chain);
6934 Ops.push_back(InFlag);
6936 // Create the STACKMAP node.
6937 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6938 SDNode *SM = DAG.getMachineNode(TargetOpcode::STACKMAP, DL, NodeTys, Ops);
6939 Chain = SDValue(SM, 0);
6940 InFlag = Chain.getValue(1);
6942 Chain = DAG.getCALLSEQ_END(Chain, NullPtr, NullPtr, InFlag, DL);
6944 // Stackmaps don't generate values, so nothing goes into the NodeMap.
6946 // Set the root to the target-lowered call chain.
6949 // Inform the Frame Information that we have a stackmap in this function.
6950 FuncInfo.MF->getFrameInfo()->setHasStackMap();
6953 /// \brief Lower llvm.experimental.patchpoint directly to its target opcode.
6954 void SelectionDAGBuilder::visitPatchpoint(const CallInst &CI) {
6955 // void|i64 @llvm.experimental.patchpoint.void|i64(i64 <id>,
6960 // [live variables...])
6962 CallingConv::ID CC = CI.getCallingConv();
6963 bool isAnyRegCC = CC == CallingConv::AnyReg;
6964 bool hasDef = !CI.getType()->isVoidTy();
6965 SDValue Callee = getValue(CI.getOperand(2)); // <target>
6967 // Get the real number of arguments participating in the call <numArgs>
6968 SDValue NArgVal = getValue(CI.getArgOperand(PatchPointOpers::NArgPos));
6969 unsigned NumArgs = cast<ConstantSDNode>(NArgVal)->getZExtValue();
6971 // Skip the four meta args: <id>, <numNopBytes>, <target>, <numArgs>
6972 // Intrinsics include all meta-operands up to but not including CC.
6973 unsigned NumMetaOpers = PatchPointOpers::CCPos;
6974 assert(CI.getNumArgOperands() >= NumMetaOpers + NumArgs &&
6975 "Not enough arguments provided to the patchpoint intrinsic");
6977 // For AnyRegCC the arguments are lowered later on manually.
6978 unsigned NumCallArgs = isAnyRegCC ? 0 : NumArgs;
6979 std::pair<SDValue, SDValue> Result =
6980 LowerCallOperands(CI, NumMetaOpers, NumCallArgs, Callee, isAnyRegCC);
6982 // Set the root to the target-lowered call chain.
6983 SDValue Chain = Result.second;
6986 SDNode *CallEnd = Chain.getNode();
6987 if (hasDef && (CallEnd->getOpcode() == ISD::CopyFromReg))
6988 CallEnd = CallEnd->getOperand(0).getNode();
6990 /// Get a call instruction from the call sequence chain.
6991 /// Tail calls are not allowed.
6992 assert(CallEnd->getOpcode() == ISD::CALLSEQ_END &&
6993 "Expected a callseq node.");
6994 SDNode *Call = CallEnd->getOperand(0).getNode();
6995 bool hasGlue = Call->getGluedNode();
6997 // Replace the target specific call node with the patchable intrinsic.
6998 SmallVector<SDValue, 8> Ops;
7000 // Add the <id> and <numBytes> constants.
7001 SDValue IDVal = getValue(CI.getOperand(PatchPointOpers::IDPos));
7002 Ops.push_back(DAG.getTargetConstant(
7003 cast<ConstantSDNode>(IDVal)->getZExtValue(), MVT::i64));
7004 SDValue NBytesVal = getValue(CI.getOperand(PatchPointOpers::NBytesPos));
7005 Ops.push_back(DAG.getTargetConstant(
7006 cast<ConstantSDNode>(NBytesVal)->getZExtValue(), MVT::i32));
7008 // Assume that the Callee is a constant address.
7009 // FIXME: handle function symbols in the future.
7011 DAG.getIntPtrConstant(cast<ConstantSDNode>(Callee)->getZExtValue(),
7012 /*isTarget=*/true));
7014 // Adjust <numArgs> to account for any arguments that have been passed on the
7016 // Call Node: Chain, Target, {Args}, RegMask, [Glue]
7017 unsigned NumCallRegArgs = Call->getNumOperands() - (hasGlue ? 4 : 3);
7018 NumCallRegArgs = isAnyRegCC ? NumArgs : NumCallRegArgs;
7019 Ops.push_back(DAG.getTargetConstant(NumCallRegArgs, MVT::i32));
7021 // Add the calling convention
7022 Ops.push_back(DAG.getTargetConstant((unsigned)CC, MVT::i32));
7024 // Add the arguments we omitted previously. The register allocator should
7025 // place these in any free register.
7027 for (unsigned i = NumMetaOpers, e = NumMetaOpers + NumArgs; i != e; ++i)
7028 Ops.push_back(getValue(CI.getArgOperand(i)));
7030 // Push the arguments from the call instruction up to the register mask.
7031 SDNode::op_iterator e = hasGlue ? Call->op_end()-2 : Call->op_end()-1;
7032 for (SDNode::op_iterator i = Call->op_begin()+2; i != e; ++i)
7035 // Push live variables for the stack map.
7036 addStackMapLiveVars(CI, NumMetaOpers + NumArgs, Ops, *this);
7038 // Push the register mask info.
7040 Ops.push_back(*(Call->op_end()-2));
7042 Ops.push_back(*(Call->op_end()-1));
7044 // Push the chain (this is originally the first operand of the call, but
7045 // becomes now the last or second to last operand).
7046 Ops.push_back(*(Call->op_begin()));
7048 // Push the glue flag (last operand).
7050 Ops.push_back(*(Call->op_end()-1));
7053 if (isAnyRegCC && hasDef) {
7054 // Create the return types based on the intrinsic definition
7055 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7056 SmallVector<EVT, 3> ValueVTs;
7057 ComputeValueVTs(TLI, CI.getType(), ValueVTs);
7058 assert(ValueVTs.size() == 1 && "Expected only one return value type.");
7060 // There is always a chain and a glue type at the end
7061 ValueVTs.push_back(MVT::Other);
7062 ValueVTs.push_back(MVT::Glue);
7063 NodeTys = DAG.getVTList(ValueVTs);
7065 NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
7067 // Replace the target specific call node with a PATCHPOINT node.
7068 MachineSDNode *MN = DAG.getMachineNode(TargetOpcode::PATCHPOINT,
7069 getCurSDLoc(), NodeTys, Ops);
7071 // Update the NodeMap.
7074 setValue(&CI, SDValue(MN, 0));
7076 setValue(&CI, Result.first);
7079 // Fixup the consumers of the intrinsic. The chain and glue may be used in the
7080 // call sequence. Furthermore the location of the chain and glue can change
7081 // when the AnyReg calling convention is used and the intrinsic returns a
7083 if (isAnyRegCC && hasDef) {
7084 SDValue From[] = {SDValue(Call, 0), SDValue(Call, 1)};
7085 SDValue To[] = {SDValue(MN, 1), SDValue(MN, 2)};
7086 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
7088 DAG.ReplaceAllUsesWith(Call, MN);
7089 DAG.DeleteNode(Call);
7091 // Inform the Frame Information that we have a patchpoint in this function.
7092 FuncInfo.MF->getFrameInfo()->setHasPatchPoint();
7095 /// Returns an AttributeSet representing the attributes applied to the return
7096 /// value of the given call.
7097 static AttributeSet getReturnAttrs(TargetLowering::CallLoweringInfo &CLI) {
7098 SmallVector<Attribute::AttrKind, 2> Attrs;
7100 Attrs.push_back(Attribute::SExt);
7102 Attrs.push_back(Attribute::ZExt);
7104 Attrs.push_back(Attribute::InReg);
7106 return AttributeSet::get(CLI.RetTy->getContext(), AttributeSet::ReturnIndex,
7110 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
7111 /// implementation, which just calls LowerCall.
7112 /// FIXME: When all targets are
7113 /// migrated to using LowerCall, this hook should be integrated into SDISel.
7114 std::pair<SDValue, SDValue>
7115 TargetLowering::LowerCallTo(TargetLowering::CallLoweringInfo &CLI) const {
7116 // Handle the incoming return values from the call.
7118 Type *OrigRetTy = CLI.RetTy;
7119 SmallVector<EVT, 4> RetTys;
7120 SmallVector<uint64_t, 4> Offsets;
7121 ComputeValueVTs(*this, CLI.RetTy, RetTys, &Offsets);
7123 SmallVector<ISD::OutputArg, 4> Outs;
7124 GetReturnInfo(CLI.RetTy, getReturnAttrs(CLI), Outs, *this);
7126 bool CanLowerReturn =
7127 this->CanLowerReturn(CLI.CallConv, CLI.DAG.getMachineFunction(),
7128 CLI.IsVarArg, Outs, CLI.RetTy->getContext());
7130 SDValue DemoteStackSlot;
7131 int DemoteStackIdx = -100;
7132 if (!CanLowerReturn) {
7133 // FIXME: equivalent assert?
7134 // assert(!CS.hasInAllocaArgument() &&
7135 // "sret demotion is incompatible with inalloca");
7136 uint64_t TySize = getDataLayout()->getTypeAllocSize(CLI.RetTy);
7137 unsigned Align = getDataLayout()->getPrefTypeAlignment(CLI.RetTy);
7138 MachineFunction &MF = CLI.DAG.getMachineFunction();
7139 DemoteStackIdx = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
7140 Type *StackSlotPtrType = PointerType::getUnqual(CLI.RetTy);
7142 DemoteStackSlot = CLI.DAG.getFrameIndex(DemoteStackIdx, getPointerTy());
7144 Entry.Node = DemoteStackSlot;
7145 Entry.Ty = StackSlotPtrType;
7146 Entry.isSExt = false;
7147 Entry.isZExt = false;
7148 Entry.isInReg = false;
7149 Entry.isSRet = true;
7150 Entry.isNest = false;
7151 Entry.isByVal = false;
7152 Entry.isReturned = false;
7153 Entry.Alignment = Align;
7154 CLI.getArgs().insert(CLI.getArgs().begin(), Entry);
7155 CLI.RetTy = Type::getVoidTy(CLI.RetTy->getContext());
7157 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
7159 MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
7160 unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
7161 for (unsigned i = 0; i != NumRegs; ++i) {
7162 ISD::InputArg MyFlags;
7163 MyFlags.VT = RegisterVT;
7165 MyFlags.Used = CLI.IsReturnValueUsed;
7167 MyFlags.Flags.setSExt();
7169 MyFlags.Flags.setZExt();
7171 MyFlags.Flags.setInReg();
7172 CLI.Ins.push_back(MyFlags);
7177 // Handle all of the outgoing arguments.
7179 CLI.OutVals.clear();
7180 ArgListTy &Args = CLI.getArgs();
7181 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
7182 SmallVector<EVT, 4> ValueVTs;
7183 ComputeValueVTs(*this, Args[i].Ty, ValueVTs);
7184 Type *FinalType = Args[i].Ty;
7185 if (Args[i].isByVal)
7186 FinalType = cast<PointerType>(Args[i].Ty)->getElementType();
7187 bool NeedsRegBlock = functionArgumentNeedsConsecutiveRegisters(
7188 FinalType, CLI.CallConv, CLI.IsVarArg);
7189 for (unsigned Value = 0, NumValues = ValueVTs.size(); Value != NumValues;
7191 EVT VT = ValueVTs[Value];
7192 Type *ArgTy = VT.getTypeForEVT(CLI.RetTy->getContext());
7193 SDValue Op = SDValue(Args[i].Node.getNode(),
7194 Args[i].Node.getResNo() + Value);
7195 ISD::ArgFlagsTy Flags;
7196 unsigned OriginalAlignment = getDataLayout()->getABITypeAlignment(ArgTy);
7202 if (Args[i].isInReg)
7206 if (Args[i].isByVal)
7208 if (Args[i].isInAlloca) {
7209 Flags.setInAlloca();
7210 // Set the byval flag for CCAssignFn callbacks that don't know about
7211 // inalloca. This way we can know how many bytes we should've allocated
7212 // and how many bytes a callee cleanup function will pop. If we port
7213 // inalloca to more targets, we'll have to add custom inalloca handling
7214 // in the various CC lowering callbacks.
7217 if (Args[i].isByVal || Args[i].isInAlloca) {
7218 PointerType *Ty = cast<PointerType>(Args[i].Ty);
7219 Type *ElementTy = Ty->getElementType();
7220 Flags.setByValSize(getDataLayout()->getTypeAllocSize(ElementTy));
7221 // For ByVal, alignment should come from FE. BE will guess if this
7222 // info is not there but there are cases it cannot get right.
7223 unsigned FrameAlign;
7224 if (Args[i].Alignment)
7225 FrameAlign = Args[i].Alignment;
7227 FrameAlign = getByValTypeAlignment(ElementTy);
7228 Flags.setByValAlign(FrameAlign);
7233 Flags.setInConsecutiveRegs();
7234 Flags.setOrigAlign(OriginalAlignment);
7236 MVT PartVT = getRegisterType(CLI.RetTy->getContext(), VT);
7237 unsigned NumParts = getNumRegisters(CLI.RetTy->getContext(), VT);
7238 SmallVector<SDValue, 4> Parts(NumParts);
7239 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
7242 ExtendKind = ISD::SIGN_EXTEND;
7243 else if (Args[i].isZExt)
7244 ExtendKind = ISD::ZERO_EXTEND;
7246 // Conservatively only handle 'returned' on non-vectors for now
7247 if (Args[i].isReturned && !Op.getValueType().isVector()) {
7248 assert(CLI.RetTy == Args[i].Ty && RetTys.size() == NumValues &&
7249 "unexpected use of 'returned'");
7250 // Before passing 'returned' to the target lowering code, ensure that
7251 // either the register MVT and the actual EVT are the same size or that
7252 // the return value and argument are extended in the same way; in these
7253 // cases it's safe to pass the argument register value unchanged as the
7254 // return register value (although it's at the target's option whether
7256 // TODO: allow code generation to take advantage of partially preserved
7257 // registers rather than clobbering the entire register when the
7258 // parameter extension method is not compatible with the return
7260 if ((NumParts * PartVT.getSizeInBits() == VT.getSizeInBits()) ||
7261 (ExtendKind != ISD::ANY_EXTEND &&
7262 CLI.RetSExt == Args[i].isSExt && CLI.RetZExt == Args[i].isZExt))
7263 Flags.setReturned();
7266 getCopyToParts(CLI.DAG, CLI.DL, Op, &Parts[0], NumParts, PartVT,
7267 CLI.CS ? CLI.CS->getInstruction() : nullptr, ExtendKind);
7269 for (unsigned j = 0; j != NumParts; ++j) {
7270 // if it isn't first piece, alignment must be 1
7271 ISD::OutputArg MyFlags(Flags, Parts[j].getValueType(), VT,
7272 i < CLI.NumFixedArgs,
7273 i, j*Parts[j].getValueType().getStoreSize());
7274 if (NumParts > 1 && j == 0)
7275 MyFlags.Flags.setSplit();
7277 MyFlags.Flags.setOrigAlign(1);
7279 // Only mark the end at the last register of the last value.
7280 if (NeedsRegBlock && Value == NumValues - 1 && j == NumParts - 1)
7281 MyFlags.Flags.setInConsecutiveRegsLast();
7283 CLI.Outs.push_back(MyFlags);
7284 CLI.OutVals.push_back(Parts[j]);
7289 SmallVector<SDValue, 4> InVals;
7290 CLI.Chain = LowerCall(CLI, InVals);
7292 // Verify that the target's LowerCall behaved as expected.
7293 assert(CLI.Chain.getNode() && CLI.Chain.getValueType() == MVT::Other &&
7294 "LowerCall didn't return a valid chain!");
7295 assert((!CLI.IsTailCall || InVals.empty()) &&
7296 "LowerCall emitted a return value for a tail call!");
7297 assert((CLI.IsTailCall || InVals.size() == CLI.Ins.size()) &&
7298 "LowerCall didn't emit the correct number of values!");
7300 // For a tail call, the return value is merely live-out and there aren't
7301 // any nodes in the DAG representing it. Return a special value to
7302 // indicate that a tail call has been emitted and no more Instructions
7303 // should be processed in the current block.
7304 if (CLI.IsTailCall) {
7305 CLI.DAG.setRoot(CLI.Chain);
7306 return std::make_pair(SDValue(), SDValue());
7309 DEBUG(for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++i) {
7310 assert(InVals[i].getNode() &&
7311 "LowerCall emitted a null value!");
7312 assert(EVT(CLI.Ins[i].VT) == InVals[i].getValueType() &&
7313 "LowerCall emitted a value with the wrong type!");
7316 SmallVector<SDValue, 4> ReturnValues;
7317 if (!CanLowerReturn) {
7318 // The instruction result is the result of loading from the
7319 // hidden sret parameter.
7320 SmallVector<EVT, 1> PVTs;
7321 Type *PtrRetTy = PointerType::getUnqual(OrigRetTy);
7323 ComputeValueVTs(*this, PtrRetTy, PVTs);
7324 assert(PVTs.size() == 1 && "Pointers should fit in one register");
7325 EVT PtrVT = PVTs[0];
7327 unsigned NumValues = RetTys.size();
7328 ReturnValues.resize(NumValues);
7329 SmallVector<SDValue, 4> Chains(NumValues);
7331 for (unsigned i = 0; i < NumValues; ++i) {
7332 SDValue Add = CLI.DAG.getNode(ISD::ADD, CLI.DL, PtrVT, DemoteStackSlot,
7333 CLI.DAG.getConstant(Offsets[i], PtrVT));
7334 SDValue L = CLI.DAG.getLoad(
7335 RetTys[i], CLI.DL, CLI.Chain, Add,
7336 MachinePointerInfo::getFixedStack(DemoteStackIdx, Offsets[i]), false,
7338 ReturnValues[i] = L;
7339 Chains[i] = L.getValue(1);
7342 CLI.Chain = CLI.DAG.getNode(ISD::TokenFactor, CLI.DL, MVT::Other, Chains);
7344 // Collect the legal value parts into potentially illegal values
7345 // that correspond to the original function's return values.
7346 ISD::NodeType AssertOp = ISD::DELETED_NODE;
7348 AssertOp = ISD::AssertSext;
7349 else if (CLI.RetZExt)
7350 AssertOp = ISD::AssertZext;
7351 unsigned CurReg = 0;
7352 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
7354 MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
7355 unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
7357 ReturnValues.push_back(getCopyFromParts(CLI.DAG, CLI.DL, &InVals[CurReg],
7358 NumRegs, RegisterVT, VT, nullptr,
7363 // For a function returning void, there is no return value. We can't create
7364 // such a node, so we just return a null return value in that case. In
7365 // that case, nothing will actually look at the value.
7366 if (ReturnValues.empty())
7367 return std::make_pair(SDValue(), CLI.Chain);
7370 SDValue Res = CLI.DAG.getNode(ISD::MERGE_VALUES, CLI.DL,
7371 CLI.DAG.getVTList(RetTys), ReturnValues);
7372 return std::make_pair(Res, CLI.Chain);
7375 void TargetLowering::LowerOperationWrapper(SDNode *N,
7376 SmallVectorImpl<SDValue> &Results,
7377 SelectionDAG &DAG) const {
7378 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
7380 Results.push_back(Res);
7383 SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
7384 llvm_unreachable("LowerOperation not implemented for this target!");
7388 SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) {
7389 SDValue Op = getNonRegisterValue(V);
7390 assert((Op.getOpcode() != ISD::CopyFromReg ||
7391 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
7392 "Copy from a reg to the same reg!");
7393 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg");
7395 const TargetLowering *TLI = TM.getTargetLowering();
7396 RegsForValue RFV(V->getContext(), *TLI, Reg, V->getType());
7397 SDValue Chain = DAG.getEntryNode();
7398 RFV.getCopyToRegs(Op, DAG, getCurSDLoc(), Chain, nullptr, V);
7399 PendingExports.push_back(Chain);
7402 #include "llvm/CodeGen/SelectionDAGISel.h"
7404 /// isOnlyUsedInEntryBlock - If the specified argument is only used in the
7405 /// entry block, return true. This includes arguments used by switches, since
7406 /// the switch may expand into multiple basic blocks.
7407 static bool isOnlyUsedInEntryBlock(const Argument *A, bool FastISel) {
7408 // With FastISel active, we may be splitting blocks, so force creation
7409 // of virtual registers for all non-dead arguments.
7411 return A->use_empty();
7413 const BasicBlock *Entry = A->getParent()->begin();
7414 for (const User *U : A->users())
7415 if (cast<Instruction>(U)->getParent() != Entry || isa<SwitchInst>(U))
7416 return false; // Use not in entry block.
7421 void SelectionDAGISel::LowerArguments(const Function &F) {
7422 SelectionDAG &DAG = SDB->DAG;
7423 SDLoc dl = SDB->getCurSDLoc();
7424 const TargetLowering *TLI = getTargetLowering();
7425 const DataLayout *DL = TLI->getDataLayout();
7426 SmallVector<ISD::InputArg, 16> Ins;
7428 if (!FuncInfo->CanLowerReturn) {
7429 // Put in an sret pointer parameter before all the other parameters.
7430 SmallVector<EVT, 1> ValueVTs;
7431 ComputeValueVTs(*getTargetLowering(),
7432 PointerType::getUnqual(F.getReturnType()), ValueVTs);
7434 // NOTE: Assuming that a pointer will never break down to more than one VT
7436 ISD::ArgFlagsTy Flags;
7438 MVT RegisterVT = TLI->getRegisterType(*DAG.getContext(), ValueVTs[0]);
7439 ISD::InputArg RetArg(Flags, RegisterVT, ValueVTs[0], true, 0, 0);
7440 Ins.push_back(RetArg);
7443 // Set up the incoming argument description vector.
7445 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end();
7446 I != E; ++I, ++Idx) {
7447 SmallVector<EVT, 4> ValueVTs;
7448 ComputeValueVTs(*TLI, I->getType(), ValueVTs);
7449 bool isArgValueUsed = !I->use_empty();
7450 unsigned PartBase = 0;
7451 Type *FinalType = I->getType();
7452 if (F.getAttributes().hasAttribute(Idx, Attribute::ByVal))
7453 FinalType = cast<PointerType>(FinalType)->getElementType();
7454 bool NeedsRegBlock = TLI->functionArgumentNeedsConsecutiveRegisters(
7455 FinalType, F.getCallingConv(), F.isVarArg());
7456 for (unsigned Value = 0, NumValues = ValueVTs.size();
7457 Value != NumValues; ++Value) {
7458 EVT VT = ValueVTs[Value];
7459 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
7460 ISD::ArgFlagsTy Flags;
7461 unsigned OriginalAlignment = DL->getABITypeAlignment(ArgTy);
7463 if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt))
7465 if (F.getAttributes().hasAttribute(Idx, Attribute::SExt))
7467 if (F.getAttributes().hasAttribute(Idx, Attribute::InReg))
7469 if (F.getAttributes().hasAttribute(Idx, Attribute::StructRet))
7471 if (F.getAttributes().hasAttribute(Idx, Attribute::ByVal))
7473 if (F.getAttributes().hasAttribute(Idx, Attribute::InAlloca)) {
7474 Flags.setInAlloca();
7475 // Set the byval flag for CCAssignFn callbacks that don't know about
7476 // inalloca. This way we can know how many bytes we should've allocated
7477 // and how many bytes a callee cleanup function will pop. If we port
7478 // inalloca to more targets, we'll have to add custom inalloca handling
7479 // in the various CC lowering callbacks.
7482 if (Flags.isByVal() || Flags.isInAlloca()) {
7483 PointerType *Ty = cast<PointerType>(I->getType());
7484 Type *ElementTy = Ty->getElementType();
7485 Flags.setByValSize(DL->getTypeAllocSize(ElementTy));
7486 // For ByVal, alignment should be passed from FE. BE will guess if
7487 // this info is not there but there are cases it cannot get right.
7488 unsigned FrameAlign;
7489 if (F.getParamAlignment(Idx))
7490 FrameAlign = F.getParamAlignment(Idx);
7492 FrameAlign = TLI->getByValTypeAlignment(ElementTy);
7493 Flags.setByValAlign(FrameAlign);
7495 if (F.getAttributes().hasAttribute(Idx, Attribute::Nest))
7498 Flags.setInConsecutiveRegs();
7499 Flags.setOrigAlign(OriginalAlignment);
7501 MVT RegisterVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7502 unsigned NumRegs = TLI->getNumRegisters(*CurDAG->getContext(), VT);
7503 for (unsigned i = 0; i != NumRegs; ++i) {
7504 ISD::InputArg MyFlags(Flags, RegisterVT, VT, isArgValueUsed,
7505 Idx-1, PartBase+i*RegisterVT.getStoreSize());
7506 if (NumRegs > 1 && i == 0)
7507 MyFlags.Flags.setSplit();
7508 // if it isn't first piece, alignment must be 1
7510 MyFlags.Flags.setOrigAlign(1);
7512 // Only mark the end at the last register of the last value.
7513 if (NeedsRegBlock && Value == NumValues - 1 && i == NumRegs - 1)
7514 MyFlags.Flags.setInConsecutiveRegsLast();
7516 Ins.push_back(MyFlags);
7518 PartBase += VT.getStoreSize();
7522 // Call the target to set up the argument values.
7523 SmallVector<SDValue, 8> InVals;
7524 SDValue NewRoot = TLI->LowerFormalArguments(DAG.getRoot(), F.getCallingConv(),
7528 // Verify that the target's LowerFormalArguments behaved as expected.
7529 assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other &&
7530 "LowerFormalArguments didn't return a valid chain!");
7531 assert(InVals.size() == Ins.size() &&
7532 "LowerFormalArguments didn't emit the correct number of values!");
7534 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
7535 assert(InVals[i].getNode() &&
7536 "LowerFormalArguments emitted a null value!");
7537 assert(EVT(Ins[i].VT) == InVals[i].getValueType() &&
7538 "LowerFormalArguments emitted a value with the wrong type!");
7542 // Update the DAG with the new chain value resulting from argument lowering.
7543 DAG.setRoot(NewRoot);
7545 // Set up the argument values.
7548 if (!FuncInfo->CanLowerReturn) {
7549 // Create a virtual register for the sret pointer, and put in a copy
7550 // from the sret argument into it.
7551 SmallVector<EVT, 1> ValueVTs;
7552 ComputeValueVTs(*TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
7553 MVT VT = ValueVTs[0].getSimpleVT();
7554 MVT RegVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7555 ISD::NodeType AssertOp = ISD::DELETED_NODE;
7556 SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1,
7557 RegVT, VT, nullptr, AssertOp);
7559 MachineFunction& MF = SDB->DAG.getMachineFunction();
7560 MachineRegisterInfo& RegInfo = MF.getRegInfo();
7561 unsigned SRetReg = RegInfo.createVirtualRegister(TLI->getRegClassFor(RegVT));
7562 FuncInfo->DemoteRegister = SRetReg;
7563 NewRoot = SDB->DAG.getCopyToReg(NewRoot, SDB->getCurSDLoc(),
7565 DAG.setRoot(NewRoot);
7567 // i indexes lowered arguments. Bump it past the hidden sret argument.
7568 // Idx indexes LLVM arguments. Don't touch it.
7572 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
7574 SmallVector<SDValue, 4> ArgValues;
7575 SmallVector<EVT, 4> ValueVTs;
7576 ComputeValueVTs(*TLI, I->getType(), ValueVTs);
7577 unsigned NumValues = ValueVTs.size();
7579 // If this argument is unused then remember its value. It is used to generate
7580 // debugging information.
7581 if (I->use_empty() && NumValues) {
7582 SDB->setUnusedArgValue(I, InVals[i]);
7584 // Also remember any frame index for use in FastISel.
7585 if (FrameIndexSDNode *FI =
7586 dyn_cast<FrameIndexSDNode>(InVals[i].getNode()))
7587 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
7590 for (unsigned Val = 0; Val != NumValues; ++Val) {
7591 EVT VT = ValueVTs[Val];
7592 MVT PartVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7593 unsigned NumParts = TLI->getNumRegisters(*CurDAG->getContext(), VT);
7595 if (!I->use_empty()) {
7596 ISD::NodeType AssertOp = ISD::DELETED_NODE;
7597 if (F.getAttributes().hasAttribute(Idx, Attribute::SExt))
7598 AssertOp = ISD::AssertSext;
7599 else if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt))
7600 AssertOp = ISD::AssertZext;
7602 ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i],
7603 NumParts, PartVT, VT,
7604 nullptr, AssertOp));
7610 // We don't need to do anything else for unused arguments.
7611 if (ArgValues.empty())
7614 // Note down frame index.
7615 if (FrameIndexSDNode *FI =
7616 dyn_cast<FrameIndexSDNode>(ArgValues[0].getNode()))
7617 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
7619 SDValue Res = DAG.getMergeValues(makeArrayRef(ArgValues.data(), NumValues),
7620 SDB->getCurSDLoc());
7622 SDB->setValue(I, Res);
7623 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::BUILD_PAIR) {
7624 if (LoadSDNode *LNode =
7625 dyn_cast<LoadSDNode>(Res.getOperand(0).getNode()))
7626 if (FrameIndexSDNode *FI =
7627 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode()))
7628 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
7631 // If this argument is live outside of the entry block, insert a copy from
7632 // wherever we got it to the vreg that other BB's will reference it as.
7633 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::CopyFromReg) {
7634 // If we can, though, try to skip creating an unnecessary vreg.
7635 // FIXME: This isn't very clean... it would be nice to make this more
7636 // general. It's also subtly incompatible with the hacks FastISel
7638 unsigned Reg = cast<RegisterSDNode>(Res.getOperand(1))->getReg();
7639 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
7640 FuncInfo->ValueMap[I] = Reg;
7644 if (!isOnlyUsedInEntryBlock(I, TM.Options.EnableFastISel)) {
7645 FuncInfo->InitializeRegForValue(I);
7646 SDB->CopyToExportRegsIfNeeded(I);
7650 assert(i == InVals.size() && "Argument register count mismatch!");
7652 // Finally, if the target has anything special to do, allow it to do so.
7653 // FIXME: this should insert code into the DAG!
7654 EmitFunctionEntryCode();
7657 /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
7658 /// ensure constants are generated when needed. Remember the virtual registers
7659 /// that need to be added to the Machine PHI nodes as input. We cannot just
7660 /// directly add them, because expansion might result in multiple MBB's for one
7661 /// BB. As such, the start of the BB might correspond to a different MBB than
7665 SelectionDAGBuilder::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
7666 const TerminatorInst *TI = LLVMBB->getTerminator();
7668 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
7670 // Check successor nodes' PHI nodes that expect a constant to be available
7672 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
7673 const BasicBlock *SuccBB = TI->getSuccessor(succ);
7674 if (!isa<PHINode>(SuccBB->begin())) continue;
7675 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
7677 // If this terminator has multiple identical successors (common for
7678 // switches), only handle each succ once.
7679 if (!SuccsHandled.insert(SuccMBB)) continue;
7681 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
7683 // At this point we know that there is a 1-1 correspondence between LLVM PHI
7684 // nodes and Machine PHI nodes, but the incoming operands have not been
7686 for (BasicBlock::const_iterator I = SuccBB->begin();
7687 const PHINode *PN = dyn_cast<PHINode>(I); ++I) {
7688 // Ignore dead phi's.
7689 if (PN->use_empty()) continue;
7692 if (PN->getType()->isEmptyTy())
7696 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
7698 if (const Constant *C = dyn_cast<Constant>(PHIOp)) {
7699 unsigned &RegOut = ConstantsOut[C];
7701 RegOut = FuncInfo.CreateRegs(C->getType());
7702 CopyValueToVirtualRegister(C, RegOut);
7706 DenseMap<const Value *, unsigned>::iterator I =
7707 FuncInfo.ValueMap.find(PHIOp);
7708 if (I != FuncInfo.ValueMap.end())
7711 assert(isa<AllocaInst>(PHIOp) &&
7712 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
7713 "Didn't codegen value into a register!??");
7714 Reg = FuncInfo.CreateRegs(PHIOp->getType());
7715 CopyValueToVirtualRegister(PHIOp, Reg);
7719 // Remember that this register needs to added to the machine PHI node as
7720 // the input for this MBB.
7721 SmallVector<EVT, 4> ValueVTs;
7722 const TargetLowering *TLI = TM.getTargetLowering();
7723 ComputeValueVTs(*TLI, PN->getType(), ValueVTs);
7724 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
7725 EVT VT = ValueVTs[vti];
7726 unsigned NumRegisters = TLI->getNumRegisters(*DAG.getContext(), VT);
7727 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
7728 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
7729 Reg += NumRegisters;
7734 ConstantsOut.clear();
7737 /// Add a successor MBB to ParentMBB< creating a new MachineBB for BB if SuccMBB
7740 SelectionDAGBuilder::StackProtectorDescriptor::
7741 AddSuccessorMBB(const BasicBlock *BB,
7742 MachineBasicBlock *ParentMBB,
7743 MachineBasicBlock *SuccMBB) {
7744 // If SuccBB has not been created yet, create it.
7746 MachineFunction *MF = ParentMBB->getParent();
7747 MachineFunction::iterator BBI = ParentMBB;
7748 SuccMBB = MF->CreateMachineBasicBlock(BB);
7749 MF->insert(++BBI, SuccMBB);
7751 // Add it as a successor of ParentMBB.
7752 ParentMBB->addSuccessor(SuccMBB);