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/ADT/Statistic.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/Analysis/BranchProbabilityInfo.h"
22 #include "llvm/Analysis/ConstantFolding.h"
23 #include "llvm/Analysis/TargetLibraryInfo.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/CodeGen/FastISel.h"
26 #include "llvm/CodeGen/FunctionLoweringInfo.h"
27 #include "llvm/CodeGen/GCMetadata.h"
28 #include "llvm/CodeGen/GCStrategy.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineJumpTableInfo.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/CodeGen/SelectionDAG.h"
36 #include "llvm/CodeGen/StackMaps.h"
37 #include "llvm/CodeGen/WinEHFuncInfo.h"
38 #include "llvm/IR/CallingConv.h"
39 #include "llvm/IR/Constants.h"
40 #include "llvm/IR/DataLayout.h"
41 #include "llvm/IR/DebugInfo.h"
42 #include "llvm/IR/DerivedTypes.h"
43 #include "llvm/IR/Function.h"
44 #include "llvm/IR/GlobalVariable.h"
45 #include "llvm/IR/InlineAsm.h"
46 #include "llvm/IR/Instructions.h"
47 #include "llvm/IR/IntrinsicInst.h"
48 #include "llvm/IR/Intrinsics.h"
49 #include "llvm/IR/LLVMContext.h"
50 #include "llvm/IR/Module.h"
51 #include "llvm/IR/Statepoint.h"
52 #include "llvm/MC/MCSymbol.h"
53 #include "llvm/Support/CommandLine.h"
54 #include "llvm/Support/Debug.h"
55 #include "llvm/Support/ErrorHandling.h"
56 #include "llvm/Support/MathExtras.h"
57 #include "llvm/Support/raw_ostream.h"
58 #include "llvm/Target/TargetFrameLowering.h"
59 #include "llvm/Target/TargetInstrInfo.h"
60 #include "llvm/Target/TargetIntrinsicInfo.h"
61 #include "llvm/Target/TargetLowering.h"
62 #include "llvm/Target/TargetOptions.h"
63 #include "llvm/Target/TargetSelectionDAGInfo.h"
64 #include "llvm/Target/TargetSubtargetInfo.h"
68 #define DEBUG_TYPE "isel"
70 /// LimitFloatPrecision - Generate low-precision inline sequences for
71 /// some float libcalls (6, 8 or 12 bits).
72 static unsigned LimitFloatPrecision;
74 static cl::opt<unsigned, true>
75 LimitFPPrecision("limit-float-precision",
76 cl::desc("Generate low-precision inline sequences "
77 "for some float libcalls"),
78 cl::location(LimitFloatPrecision),
82 EnableFMFInDAG("enable-fmf-dag", cl::init(false), cl::Hidden,
83 cl::desc("Enable fast-math-flags for DAG nodes"));
85 // Limit the width of DAG chains. This is important in general to prevent
86 // DAG-based analysis from blowing up. For example, alias analysis and
87 // load clustering may not complete in reasonable time. It is difficult to
88 // recognize and avoid this situation within each individual analysis, and
89 // future analyses are likely to have the same behavior. Limiting DAG width is
90 // the safe approach and will be especially important with global DAGs.
92 // MaxParallelChains default is arbitrarily high to avoid affecting
93 // optimization, but could be lowered to improve compile time. Any ld-ld-st-st
94 // sequence over this should have been converted to llvm.memcpy by the
95 // frontend. It easy to induce this behavior with .ll code such as:
96 // %buffer = alloca [4096 x i8]
97 // %data = load [4096 x i8]* %argPtr
98 // store [4096 x i8] %data, [4096 x i8]* %buffer
99 static const unsigned MaxParallelChains = 64;
101 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, SDLoc DL,
102 const SDValue *Parts, unsigned NumParts,
103 MVT PartVT, EVT ValueVT, const Value *V);
105 /// getCopyFromParts - Create a value that contains the specified legal parts
106 /// combined into the value they represent. If the parts combine to a type
107 /// larger then ValueVT then AssertOp can be used to specify whether the extra
108 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
109 /// (ISD::AssertSext).
110 static SDValue getCopyFromParts(SelectionDAG &DAG, SDLoc DL,
111 const SDValue *Parts,
112 unsigned NumParts, MVT PartVT, EVT ValueVT,
114 ISD::NodeType AssertOp = ISD::DELETED_NODE) {
115 if (ValueVT.isVector())
116 return getCopyFromPartsVector(DAG, DL, Parts, NumParts,
119 assert(NumParts > 0 && "No parts to assemble!");
120 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
121 SDValue Val = Parts[0];
124 // Assemble the value from multiple parts.
125 if (ValueVT.isInteger()) {
126 unsigned PartBits = PartVT.getSizeInBits();
127 unsigned ValueBits = ValueVT.getSizeInBits();
129 // Assemble the power of 2 part.
130 unsigned RoundParts = NumParts & (NumParts - 1) ?
131 1 << Log2_32(NumParts) : NumParts;
132 unsigned RoundBits = PartBits * RoundParts;
133 EVT RoundVT = RoundBits == ValueBits ?
134 ValueVT : EVT::getIntegerVT(*DAG.getContext(), RoundBits);
137 EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2);
139 if (RoundParts > 2) {
140 Lo = getCopyFromParts(DAG, DL, Parts, RoundParts / 2,
142 Hi = getCopyFromParts(DAG, DL, Parts + RoundParts / 2,
143 RoundParts / 2, PartVT, HalfVT, V);
145 Lo = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[0]);
146 Hi = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[1]);
149 if (TLI.isBigEndian())
152 Val = DAG.getNode(ISD::BUILD_PAIR, DL, RoundVT, Lo, Hi);
154 if (RoundParts < NumParts) {
155 // Assemble the trailing non-power-of-2 part.
156 unsigned OddParts = NumParts - RoundParts;
157 EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits);
158 Hi = getCopyFromParts(DAG, DL,
159 Parts + RoundParts, OddParts, PartVT, OddVT, V);
161 // Combine the round and odd parts.
163 if (TLI.isBigEndian())
165 EVT TotalVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
166 Hi = DAG.getNode(ISD::ANY_EXTEND, DL, TotalVT, Hi);
167 Hi = DAG.getNode(ISD::SHL, DL, TotalVT, Hi,
168 DAG.getConstant(Lo.getValueType().getSizeInBits(), DL,
169 TLI.getPointerTy()));
170 Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, TotalVT, Lo);
171 Val = DAG.getNode(ISD::OR, DL, TotalVT, Lo, Hi);
173 } else if (PartVT.isFloatingPoint()) {
174 // FP split into multiple FP parts (for ppcf128)
175 assert(ValueVT == EVT(MVT::ppcf128) && PartVT == MVT::f64 &&
178 Lo = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[0]);
179 Hi = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[1]);
180 if (TLI.hasBigEndianPartOrdering(ValueVT))
182 Val = DAG.getNode(ISD::BUILD_PAIR, DL, ValueVT, Lo, Hi);
184 // FP split into integer parts (soft fp)
185 assert(ValueVT.isFloatingPoint() && PartVT.isInteger() &&
186 !PartVT.isVector() && "Unexpected split");
187 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits());
188 Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT, V);
192 // There is now one part, held in Val. Correct it to match ValueVT.
193 EVT PartEVT = Val.getValueType();
195 if (PartEVT == ValueVT)
198 if (PartEVT.isInteger() && ValueVT.isInteger()) {
199 if (ValueVT.bitsLT(PartEVT)) {
200 // For a truncate, see if we have any information to
201 // indicate whether the truncated bits will always be
202 // zero or sign-extension.
203 if (AssertOp != ISD::DELETED_NODE)
204 Val = DAG.getNode(AssertOp, DL, PartEVT, Val,
205 DAG.getValueType(ValueVT));
206 return DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
208 return DAG.getNode(ISD::ANY_EXTEND, DL, ValueVT, Val);
211 if (PartEVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
212 // FP_ROUND's are always exact here.
213 if (ValueVT.bitsLT(Val.getValueType()))
214 return DAG.getNode(ISD::FP_ROUND, DL, ValueVT, Val,
215 DAG.getTargetConstant(1, DL, TLI.getPointerTy()));
217 return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val);
220 if (PartEVT.getSizeInBits() == ValueVT.getSizeInBits())
221 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
223 llvm_unreachable("Unknown mismatch!");
226 static void diagnosePossiblyInvalidConstraint(LLVMContext &Ctx, const Value *V,
227 const Twine &ErrMsg) {
228 const Instruction *I = dyn_cast_or_null<Instruction>(V);
230 return Ctx.emitError(ErrMsg);
232 const char *AsmError = ", possible invalid constraint for vector type";
233 if (const CallInst *CI = dyn_cast<CallInst>(I))
234 if (isa<InlineAsm>(CI->getCalledValue()))
235 return Ctx.emitError(I, ErrMsg + AsmError);
237 return Ctx.emitError(I, ErrMsg);
240 /// getCopyFromPartsVector - Create a value that contains the specified legal
241 /// parts combined into the value they represent. If the parts combine to a
242 /// type larger then ValueVT then AssertOp can be used to specify whether the
243 /// extra bits are known to be zero (ISD::AssertZext) or sign extended from
244 /// ValueVT (ISD::AssertSext).
245 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, SDLoc DL,
246 const SDValue *Parts, unsigned NumParts,
247 MVT PartVT, EVT ValueVT, const Value *V) {
248 assert(ValueVT.isVector() && "Not a vector value");
249 assert(NumParts > 0 && "No parts to assemble!");
250 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
251 SDValue Val = Parts[0];
253 // Handle a multi-element vector.
257 unsigned NumIntermediates;
259 TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT,
260 NumIntermediates, RegisterVT);
261 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
262 NumParts = NumRegs; // Silence a compiler warning.
263 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
264 assert(RegisterVT.getSizeInBits() ==
265 Parts[0].getSimpleValueType().getSizeInBits() &&
266 "Part type sizes don't match!");
268 // Assemble the parts into intermediate operands.
269 SmallVector<SDValue, 8> Ops(NumIntermediates);
270 if (NumIntermediates == NumParts) {
271 // If the register was not expanded, truncate or copy the value,
273 for (unsigned i = 0; i != NumParts; ++i)
274 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i], 1,
275 PartVT, IntermediateVT, V);
276 } else if (NumParts > 0) {
277 // If the intermediate type was expanded, build the intermediate
278 // operands from the parts.
279 assert(NumParts % NumIntermediates == 0 &&
280 "Must expand into a divisible number of parts!");
281 unsigned Factor = NumParts / NumIntermediates;
282 for (unsigned i = 0; i != NumIntermediates; ++i)
283 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor,
284 PartVT, IntermediateVT, V);
287 // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the
288 // intermediate operands.
289 Val = DAG.getNode(IntermediateVT.isVector() ? ISD::CONCAT_VECTORS
294 // There is now one part, held in Val. Correct it to match ValueVT.
295 EVT PartEVT = Val.getValueType();
297 if (PartEVT == ValueVT)
300 if (PartEVT.isVector()) {
301 // If the element type of the source/dest vectors are the same, but the
302 // parts vector has more elements than the value vector, then we have a
303 // vector widening case (e.g. <2 x float> -> <4 x float>). Extract the
305 if (PartEVT.getVectorElementType() == ValueVT.getVectorElementType()) {
306 assert(PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements() &&
307 "Cannot narrow, it would be a lossy transformation");
308 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val,
309 DAG.getConstant(0, DL, TLI.getVectorIdxTy()));
312 // Vector/Vector bitcast.
313 if (ValueVT.getSizeInBits() == PartEVT.getSizeInBits())
314 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
316 assert(PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements() &&
317 "Cannot handle this kind of promotion");
318 // Promoted vector extract
319 bool Smaller = ValueVT.bitsLE(PartEVT);
320 return DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
325 // Trivial bitcast if the types are the same size and the destination
326 // vector type is legal.
327 if (PartEVT.getSizeInBits() == ValueVT.getSizeInBits() &&
328 TLI.isTypeLegal(ValueVT))
329 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
331 // Handle cases such as i8 -> <1 x i1>
332 if (ValueVT.getVectorNumElements() != 1) {
333 diagnosePossiblyInvalidConstraint(*DAG.getContext(), V,
334 "non-trivial scalar-to-vector conversion");
335 return DAG.getUNDEF(ValueVT);
338 if (ValueVT.getVectorNumElements() == 1 &&
339 ValueVT.getVectorElementType() != PartEVT) {
340 bool Smaller = ValueVT.bitsLE(PartEVT);
341 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
342 DL, ValueVT.getScalarType(), Val);
345 return DAG.getNode(ISD::BUILD_VECTOR, DL, ValueVT, Val);
348 static void getCopyToPartsVector(SelectionDAG &DAG, SDLoc dl,
349 SDValue Val, SDValue *Parts, unsigned NumParts,
350 MVT PartVT, const Value *V);
352 /// getCopyToParts - Create a series of nodes that contain the specified value
353 /// split into legal parts. If the parts contain more bits than Val, then, for
354 /// integers, ExtendKind can be used to specify how to generate the extra bits.
355 static void getCopyToParts(SelectionDAG &DAG, SDLoc DL,
356 SDValue Val, SDValue *Parts, unsigned NumParts,
357 MVT PartVT, const Value *V,
358 ISD::NodeType ExtendKind = ISD::ANY_EXTEND) {
359 EVT ValueVT = Val.getValueType();
361 // Handle the vector case separately.
362 if (ValueVT.isVector())
363 return getCopyToPartsVector(DAG, DL, Val, Parts, NumParts, PartVT, V);
365 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
366 unsigned PartBits = PartVT.getSizeInBits();
367 unsigned OrigNumParts = NumParts;
368 assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!");
373 assert(!ValueVT.isVector() && "Vector case handled elsewhere");
374 EVT PartEVT = PartVT;
375 if (PartEVT == ValueVT) {
376 assert(NumParts == 1 && "No-op copy with multiple parts!");
381 if (NumParts * PartBits > ValueVT.getSizeInBits()) {
382 // If the parts cover more bits than the value has, promote the value.
383 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
384 assert(NumParts == 1 && "Do not know what to promote to!");
385 Val = DAG.getNode(ISD::FP_EXTEND, DL, PartVT, Val);
387 assert((PartVT.isInteger() || PartVT == MVT::x86mmx) &&
388 ValueVT.isInteger() &&
389 "Unknown mismatch!");
390 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
391 Val = DAG.getNode(ExtendKind, DL, ValueVT, Val);
392 if (PartVT == MVT::x86mmx)
393 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
395 } else if (PartBits == ValueVT.getSizeInBits()) {
396 // Different types of the same size.
397 assert(NumParts == 1 && PartEVT != ValueVT);
398 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
399 } else if (NumParts * PartBits < ValueVT.getSizeInBits()) {
400 // If the parts cover less bits than value has, truncate the value.
401 assert((PartVT.isInteger() || PartVT == MVT::x86mmx) &&
402 ValueVT.isInteger() &&
403 "Unknown mismatch!");
404 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
405 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
406 if (PartVT == MVT::x86mmx)
407 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
410 // The value may have changed - recompute ValueVT.
411 ValueVT = Val.getValueType();
412 assert(NumParts * PartBits == ValueVT.getSizeInBits() &&
413 "Failed to tile the value with PartVT!");
416 if (PartEVT != ValueVT)
417 diagnosePossiblyInvalidConstraint(*DAG.getContext(), V,
418 "scalar-to-vector conversion failed");
424 // Expand the value into multiple parts.
425 if (NumParts & (NumParts - 1)) {
426 // The number of parts is not a power of 2. Split off and copy the tail.
427 assert(PartVT.isInteger() && ValueVT.isInteger() &&
428 "Do not know what to expand to!");
429 unsigned RoundParts = 1 << Log2_32(NumParts);
430 unsigned RoundBits = RoundParts * PartBits;
431 unsigned OddParts = NumParts - RoundParts;
432 SDValue OddVal = DAG.getNode(ISD::SRL, DL, ValueVT, Val,
433 DAG.getIntPtrConstant(RoundBits, DL));
434 getCopyToParts(DAG, DL, OddVal, Parts + RoundParts, OddParts, PartVT, V);
436 if (TLI.isBigEndian())
437 // The odd parts were reversed by getCopyToParts - unreverse them.
438 std::reverse(Parts + RoundParts, Parts + NumParts);
440 NumParts = RoundParts;
441 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
442 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
445 // The number of parts is a power of 2. Repeatedly bisect the value using
447 Parts[0] = DAG.getNode(ISD::BITCAST, DL,
448 EVT::getIntegerVT(*DAG.getContext(),
449 ValueVT.getSizeInBits()),
452 for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) {
453 for (unsigned i = 0; i < NumParts; i += StepSize) {
454 unsigned ThisBits = StepSize * PartBits / 2;
455 EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits);
456 SDValue &Part0 = Parts[i];
457 SDValue &Part1 = Parts[i+StepSize/2];
459 Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
460 ThisVT, Part0, DAG.getIntPtrConstant(1, DL));
461 Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
462 ThisVT, Part0, DAG.getIntPtrConstant(0, DL));
464 if (ThisBits == PartBits && ThisVT != PartVT) {
465 Part0 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part0);
466 Part1 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part1);
471 if (TLI.isBigEndian())
472 std::reverse(Parts, Parts + OrigNumParts);
476 /// getCopyToPartsVector - Create a series of nodes that contain the specified
477 /// value split into legal parts.
478 static void getCopyToPartsVector(SelectionDAG &DAG, SDLoc DL,
479 SDValue Val, SDValue *Parts, unsigned NumParts,
480 MVT PartVT, const Value *V) {
481 EVT ValueVT = Val.getValueType();
482 assert(ValueVT.isVector() && "Not a vector");
483 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
486 EVT PartEVT = PartVT;
487 if (PartEVT == ValueVT) {
489 } else if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) {
490 // Bitconvert vector->vector case.
491 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
492 } else if (PartVT.isVector() &&
493 PartEVT.getVectorElementType() == ValueVT.getVectorElementType() &&
494 PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements()) {
495 EVT ElementVT = PartVT.getVectorElementType();
496 // Vector widening case, e.g. <2 x float> -> <4 x float>. Shuffle in
498 SmallVector<SDValue, 16> Ops;
499 for (unsigned i = 0, e = ValueVT.getVectorNumElements(); i != e; ++i)
500 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
501 ElementVT, Val, DAG.getConstant(i, DL,
502 TLI.getVectorIdxTy())));
504 for (unsigned i = ValueVT.getVectorNumElements(),
505 e = PartVT.getVectorNumElements(); i != e; ++i)
506 Ops.push_back(DAG.getUNDEF(ElementVT));
508 Val = DAG.getNode(ISD::BUILD_VECTOR, DL, PartVT, Ops);
510 // FIXME: Use CONCAT for 2x -> 4x.
512 //SDValue UndefElts = DAG.getUNDEF(VectorTy);
513 //Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts);
514 } else if (PartVT.isVector() &&
515 PartEVT.getVectorElementType().bitsGE(
516 ValueVT.getVectorElementType()) &&
517 PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements()) {
519 // Promoted vector extract
520 bool Smaller = PartEVT.bitsLE(ValueVT);
521 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
524 // Vector -> scalar conversion.
525 assert(ValueVT.getVectorNumElements() == 1 &&
526 "Only trivial vector-to-scalar conversions should get here!");
527 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
529 DAG.getConstant(0, DL, TLI.getVectorIdxTy()));
531 bool Smaller = ValueVT.bitsLE(PartVT);
532 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
540 // Handle a multi-element vector.
543 unsigned NumIntermediates;
544 unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT,
546 NumIntermediates, RegisterVT);
547 unsigned NumElements = ValueVT.getVectorNumElements();
549 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
550 NumParts = NumRegs; // Silence a compiler warning.
551 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
553 // Split the vector into intermediate operands.
554 SmallVector<SDValue, 8> Ops(NumIntermediates);
555 for (unsigned i = 0; i != NumIntermediates; ++i) {
556 if (IntermediateVT.isVector())
557 Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL,
559 DAG.getConstant(i * (NumElements / NumIntermediates), DL,
560 TLI.getVectorIdxTy()));
562 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
564 DAG.getConstant(i, DL, TLI.getVectorIdxTy()));
567 // Split the intermediate operands into legal parts.
568 if (NumParts == NumIntermediates) {
569 // If the register was not expanded, promote or copy the value,
571 for (unsigned i = 0; i != NumParts; ++i)
572 getCopyToParts(DAG, DL, Ops[i], &Parts[i], 1, PartVT, V);
573 } else if (NumParts > 0) {
574 // If the intermediate type was expanded, split each the value into
576 assert(NumIntermediates != 0 && "division by zero");
577 assert(NumParts % NumIntermediates == 0 &&
578 "Must expand into a divisible number of parts!");
579 unsigned Factor = NumParts / NumIntermediates;
580 for (unsigned i = 0; i != NumIntermediates; ++i)
581 getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT, V);
585 RegsForValue::RegsForValue() {}
587 RegsForValue::RegsForValue(const SmallVector<unsigned, 4> ®s, MVT regvt,
589 : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {}
591 RegsForValue::RegsForValue(LLVMContext &Context, const TargetLowering &tli,
592 unsigned Reg, Type *Ty) {
593 ComputeValueVTs(tli, Ty, ValueVTs);
595 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
596 EVT ValueVT = ValueVTs[Value];
597 unsigned NumRegs = tli.getNumRegisters(Context, ValueVT);
598 MVT RegisterVT = tli.getRegisterType(Context, ValueVT);
599 for (unsigned i = 0; i != NumRegs; ++i)
600 Regs.push_back(Reg + i);
601 RegVTs.push_back(RegisterVT);
606 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
607 /// this value and returns the result as a ValueVT value. This uses
608 /// Chain/Flag as the input and updates them for the output Chain/Flag.
609 /// If the Flag pointer is NULL, no flag is used.
610 SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
611 FunctionLoweringInfo &FuncInfo,
613 SDValue &Chain, SDValue *Flag,
614 const Value *V) const {
615 // A Value with type {} or [0 x %t] needs no registers.
616 if (ValueVTs.empty())
619 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
621 // Assemble the legal parts into the final values.
622 SmallVector<SDValue, 4> Values(ValueVTs.size());
623 SmallVector<SDValue, 8> Parts;
624 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
625 // Copy the legal parts from the registers.
626 EVT ValueVT = ValueVTs[Value];
627 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
628 MVT RegisterVT = RegVTs[Value];
630 Parts.resize(NumRegs);
631 for (unsigned i = 0; i != NumRegs; ++i) {
634 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT);
636 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag);
637 *Flag = P.getValue(2);
640 Chain = P.getValue(1);
643 // If the source register was virtual and if we know something about it,
644 // add an assert node.
645 if (!TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) ||
646 !RegisterVT.isInteger() || RegisterVT.isVector())
649 const FunctionLoweringInfo::LiveOutInfo *LOI =
650 FuncInfo.GetLiveOutRegInfo(Regs[Part+i]);
654 unsigned RegSize = RegisterVT.getSizeInBits();
655 unsigned NumSignBits = LOI->NumSignBits;
656 unsigned NumZeroBits = LOI->KnownZero.countLeadingOnes();
658 if (NumZeroBits == RegSize) {
659 // The current value is a zero.
660 // Explicitly express that as it would be easier for
661 // optimizations to kick in.
662 Parts[i] = DAG.getConstant(0, dl, RegisterVT);
666 // FIXME: We capture more information than the dag can represent. For
667 // now, just use the tightest assertzext/assertsext possible.
669 EVT FromVT(MVT::Other);
670 if (NumSignBits == RegSize)
671 isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1
672 else if (NumZeroBits >= RegSize-1)
673 isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1
674 else if (NumSignBits > RegSize-8)
675 isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8
676 else if (NumZeroBits >= RegSize-8)
677 isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8
678 else if (NumSignBits > RegSize-16)
679 isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16
680 else if (NumZeroBits >= RegSize-16)
681 isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16
682 else if (NumSignBits > RegSize-32)
683 isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32
684 else if (NumZeroBits >= RegSize-32)
685 isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32
689 // Add an assertion node.
690 assert(FromVT != MVT::Other);
691 Parts[i] = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl,
692 RegisterVT, P, DAG.getValueType(FromVT));
695 Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(),
696 NumRegs, RegisterVT, ValueVT, V);
701 return DAG.getNode(ISD::MERGE_VALUES, dl, DAG.getVTList(ValueVTs), Values);
704 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
705 /// specified value into the registers specified by this object. This uses
706 /// Chain/Flag as the input and updates them for the output Chain/Flag.
707 /// If the Flag pointer is NULL, no flag is used.
708 void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, SDLoc dl,
709 SDValue &Chain, SDValue *Flag, const Value *V,
710 ISD::NodeType PreferredExtendType) const {
711 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
712 ISD::NodeType ExtendKind = PreferredExtendType;
714 // Get the list of the values's legal parts.
715 unsigned NumRegs = Regs.size();
716 SmallVector<SDValue, 8> Parts(NumRegs);
717 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
718 EVT ValueVT = ValueVTs[Value];
719 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
720 MVT RegisterVT = RegVTs[Value];
722 if (ExtendKind == ISD::ANY_EXTEND && TLI.isZExtFree(Val, RegisterVT))
723 ExtendKind = ISD::ZERO_EXTEND;
725 getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value),
726 &Parts[Part], NumParts, RegisterVT, V, ExtendKind);
730 // Copy the parts into the registers.
731 SmallVector<SDValue, 8> Chains(NumRegs);
732 for (unsigned i = 0; i != NumRegs; ++i) {
735 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]);
737 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag);
738 *Flag = Part.getValue(1);
741 Chains[i] = Part.getValue(0);
744 if (NumRegs == 1 || Flag)
745 // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is
746 // flagged to it. That is the CopyToReg nodes and the user are considered
747 // a single scheduling unit. If we create a TokenFactor and return it as
748 // chain, then the TokenFactor is both a predecessor (operand) of the
749 // user as well as a successor (the TF operands are flagged to the user).
750 // c1, f1 = CopyToReg
751 // c2, f2 = CopyToReg
752 // c3 = TokenFactor c1, c2
755 Chain = Chains[NumRegs-1];
757 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
760 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
761 /// operand list. This adds the code marker and includes the number of
762 /// values added into it.
763 void RegsForValue::AddInlineAsmOperands(unsigned Code, bool HasMatching,
764 unsigned MatchingIdx, SDLoc dl,
766 std::vector<SDValue> &Ops) const {
767 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
769 unsigned Flag = InlineAsm::getFlagWord(Code, Regs.size());
771 Flag = InlineAsm::getFlagWordForMatchingOp(Flag, MatchingIdx);
772 else if (!Regs.empty() &&
773 TargetRegisterInfo::isVirtualRegister(Regs.front())) {
774 // Put the register class of the virtual registers in the flag word. That
775 // way, later passes can recompute register class constraints for inline
776 // assembly as well as normal instructions.
777 // Don't do this for tied operands that can use the regclass information
779 const MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
780 const TargetRegisterClass *RC = MRI.getRegClass(Regs.front());
781 Flag = InlineAsm::getFlagWordForRegClass(Flag, RC->getID());
784 SDValue Res = DAG.getTargetConstant(Flag, dl, MVT::i32);
787 unsigned SP = TLI.getStackPointerRegisterToSaveRestore();
788 for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) {
789 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVTs[Value]);
790 MVT RegisterVT = RegVTs[Value];
791 for (unsigned i = 0; i != NumRegs; ++i) {
792 assert(Reg < Regs.size() && "Mismatch in # registers expected");
793 unsigned TheReg = Regs[Reg++];
794 Ops.push_back(DAG.getRegister(TheReg, RegisterVT));
796 if (TheReg == SP && Code == InlineAsm::Kind_Clobber) {
797 // If we clobbered the stack pointer, MFI should know about it.
798 assert(DAG.getMachineFunction().getFrameInfo()->
799 hasInlineAsmWithSPAdjust());
805 void SelectionDAGBuilder::init(GCFunctionInfo *gfi, AliasAnalysis &aa,
806 const TargetLibraryInfo *li) {
810 DL = DAG.getTarget().getDataLayout();
811 Context = DAG.getContext();
812 LPadToCallSiteMap.clear();
815 /// clear - Clear out the current SelectionDAG and the associated
816 /// state and prepare this SelectionDAGBuilder object to be used
817 /// for a new block. This doesn't clear out information about
818 /// additional blocks that are needed to complete switch lowering
819 /// or PHI node updating; that information is cleared out as it is
821 void SelectionDAGBuilder::clear() {
823 UnusedArgNodeMap.clear();
824 PendingLoads.clear();
825 PendingExports.clear();
828 SDNodeOrder = LowestSDNodeOrder;
829 StatepointLowering.clear();
832 /// clearDanglingDebugInfo - Clear the dangling debug information
833 /// map. This function is separated from the clear so that debug
834 /// information that is dangling in a basic block can be properly
835 /// resolved in a different basic block. This allows the
836 /// SelectionDAG to resolve dangling debug information attached
838 void SelectionDAGBuilder::clearDanglingDebugInfo() {
839 DanglingDebugInfoMap.clear();
842 /// getRoot - Return the current virtual root of the Selection DAG,
843 /// flushing any PendingLoad items. This must be done before emitting
844 /// a store or any other node that may need to be ordered after any
845 /// prior load instructions.
847 SDValue SelectionDAGBuilder::getRoot() {
848 if (PendingLoads.empty())
849 return DAG.getRoot();
851 if (PendingLoads.size() == 1) {
852 SDValue Root = PendingLoads[0];
854 PendingLoads.clear();
858 // Otherwise, we have to make a token factor node.
859 SDValue Root = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
861 PendingLoads.clear();
866 /// getControlRoot - Similar to getRoot, but instead of flushing all the
867 /// PendingLoad items, flush all the PendingExports items. It is necessary
868 /// to do this before emitting a terminator instruction.
870 SDValue SelectionDAGBuilder::getControlRoot() {
871 SDValue Root = DAG.getRoot();
873 if (PendingExports.empty())
876 // Turn all of the CopyToReg chains into one factored node.
877 if (Root.getOpcode() != ISD::EntryToken) {
878 unsigned i = 0, e = PendingExports.size();
879 for (; i != e; ++i) {
880 assert(PendingExports[i].getNode()->getNumOperands() > 1);
881 if (PendingExports[i].getNode()->getOperand(0) == Root)
882 break; // Don't add the root if we already indirectly depend on it.
886 PendingExports.push_back(Root);
889 Root = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
891 PendingExports.clear();
896 void SelectionDAGBuilder::visit(const Instruction &I) {
897 // Set up outgoing PHI node register values before emitting the terminator.
898 if (isa<TerminatorInst>(&I))
899 HandlePHINodesInSuccessorBlocks(I.getParent());
905 visit(I.getOpcode(), I);
907 if (!isa<TerminatorInst>(&I) && !HasTailCall)
908 CopyToExportRegsIfNeeded(&I);
913 void SelectionDAGBuilder::visitPHI(const PHINode &) {
914 llvm_unreachable("SelectionDAGBuilder shouldn't visit PHI nodes!");
917 void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) {
918 // Note: this doesn't use InstVisitor, because it has to work with
919 // ConstantExpr's in addition to instructions.
921 default: llvm_unreachable("Unknown instruction type encountered!");
922 // Build the switch statement using the Instruction.def file.
923 #define HANDLE_INST(NUM, OPCODE, CLASS) \
924 case Instruction::OPCODE: visit##OPCODE((const CLASS&)I); break;
925 #include "llvm/IR/Instruction.def"
929 // resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V,
930 // generate the debug data structures now that we've seen its definition.
931 void SelectionDAGBuilder::resolveDanglingDebugInfo(const Value *V,
933 DanglingDebugInfo &DDI = DanglingDebugInfoMap[V];
935 const DbgValueInst *DI = DDI.getDI();
936 DebugLoc dl = DDI.getdl();
937 unsigned DbgSDNodeOrder = DDI.getSDNodeOrder();
938 DILocalVariable *Variable = DI->getVariable();
939 DIExpression *Expr = DI->getExpression();
940 assert(Variable->isValidLocationForIntrinsic(dl) &&
941 "Expected inlined-at fields to agree");
942 uint64_t Offset = DI->getOffset();
943 // A dbg.value for an alloca is always indirect.
944 bool IsIndirect = isa<AllocaInst>(V) || Offset != 0;
947 if (!EmitFuncArgumentDbgValue(V, Variable, Expr, dl, Offset, IsIndirect,
949 SDV = DAG.getDbgValue(Variable, Expr, Val.getNode(), Val.getResNo(),
950 IsIndirect, Offset, dl, DbgSDNodeOrder);
951 DAG.AddDbgValue(SDV, Val.getNode(), false);
954 DEBUG(dbgs() << "Dropping debug info for " << *DI << "\n");
955 DanglingDebugInfoMap[V] = DanglingDebugInfo();
959 /// getCopyFromRegs - If there was virtual register allocated for the value V
960 /// emit CopyFromReg of the specified type Ty. Return empty SDValue() otherwise.
961 SDValue SelectionDAGBuilder::getCopyFromRegs(const Value *V, Type *Ty) {
962 DenseMap<const Value *, unsigned>::iterator It = FuncInfo.ValueMap.find(V);
965 if (It != FuncInfo.ValueMap.end()) {
966 unsigned InReg = It->second;
967 RegsForValue RFV(*DAG.getContext(), DAG.getTargetLoweringInfo(), InReg,
969 SDValue Chain = DAG.getEntryNode();
970 Result = RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, nullptr, V);
971 resolveDanglingDebugInfo(V, Result);
977 /// getValue - Return an SDValue for the given Value.
978 SDValue SelectionDAGBuilder::getValue(const Value *V) {
979 // If we already have an SDValue for this value, use it. It's important
980 // to do this first, so that we don't create a CopyFromReg if we already
981 // have a regular SDValue.
982 SDValue &N = NodeMap[V];
983 if (N.getNode()) return N;
985 // If there's a virtual register allocated and initialized for this
987 SDValue copyFromReg = getCopyFromRegs(V, V->getType());
988 if (copyFromReg.getNode()) {
992 // Otherwise create a new SDValue and remember it.
993 SDValue Val = getValueImpl(V);
995 resolveDanglingDebugInfo(V, Val);
999 // Return true if SDValue exists for the given Value
1000 bool SelectionDAGBuilder::findValue(const Value *V) const {
1001 return (NodeMap.find(V) != NodeMap.end()) ||
1002 (FuncInfo.ValueMap.find(V) != FuncInfo.ValueMap.end());
1005 /// getNonRegisterValue - Return an SDValue for the given Value, but
1006 /// don't look in FuncInfo.ValueMap for a virtual register.
1007 SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) {
1008 // If we already have an SDValue for this value, use it.
1009 SDValue &N = NodeMap[V];
1011 if (isa<ConstantSDNode>(N) || isa<ConstantFPSDNode>(N)) {
1012 // Remove the debug location from the node as the node is about to be used
1013 // in a location which may differ from the original debug location. This
1014 // is relevant to Constant and ConstantFP nodes because they can appear
1015 // as constant expressions inside PHI nodes.
1016 N->setDebugLoc(DebugLoc());
1021 // Otherwise create a new SDValue and remember it.
1022 SDValue Val = getValueImpl(V);
1024 resolveDanglingDebugInfo(V, Val);
1028 /// getValueImpl - Helper function for getValue and getNonRegisterValue.
1029 /// Create an SDValue for the given value.
1030 SDValue SelectionDAGBuilder::getValueImpl(const Value *V) {
1031 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1033 if (const Constant *C = dyn_cast<Constant>(V)) {
1034 EVT VT = TLI.getValueType(V->getType(), true);
1036 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C))
1037 return DAG.getConstant(*CI, getCurSDLoc(), VT);
1039 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
1040 return DAG.getGlobalAddress(GV, getCurSDLoc(), VT);
1042 if (isa<ConstantPointerNull>(C)) {
1043 unsigned AS = V->getType()->getPointerAddressSpace();
1044 return DAG.getConstant(0, getCurSDLoc(), TLI.getPointerTy(AS));
1047 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
1048 return DAG.getConstantFP(*CFP, getCurSDLoc(), VT);
1050 if (isa<UndefValue>(C) && !V->getType()->isAggregateType())
1051 return DAG.getUNDEF(VT);
1053 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
1054 visit(CE->getOpcode(), *CE);
1055 SDValue N1 = NodeMap[V];
1056 assert(N1.getNode() && "visit didn't populate the NodeMap!");
1060 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) {
1061 SmallVector<SDValue, 4> Constants;
1062 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end();
1064 SDNode *Val = getValue(*OI).getNode();
1065 // If the operand is an empty aggregate, there are no values.
1067 // Add each leaf value from the operand to the Constants list
1068 // to form a flattened list of all the values.
1069 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
1070 Constants.push_back(SDValue(Val, i));
1073 return DAG.getMergeValues(Constants, getCurSDLoc());
1076 if (const ConstantDataSequential *CDS =
1077 dyn_cast<ConstantDataSequential>(C)) {
1078 SmallVector<SDValue, 4> Ops;
1079 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1080 SDNode *Val = getValue(CDS->getElementAsConstant(i)).getNode();
1081 // Add each leaf value from the operand to the Constants list
1082 // to form a flattened list of all the values.
1083 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
1084 Ops.push_back(SDValue(Val, i));
1087 if (isa<ArrayType>(CDS->getType()))
1088 return DAG.getMergeValues(Ops, getCurSDLoc());
1089 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(),
1093 if (C->getType()->isStructTy() || C->getType()->isArrayTy()) {
1094 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
1095 "Unknown struct or array constant!");
1097 SmallVector<EVT, 4> ValueVTs;
1098 ComputeValueVTs(TLI, C->getType(), ValueVTs);
1099 unsigned NumElts = ValueVTs.size();
1101 return SDValue(); // empty struct
1102 SmallVector<SDValue, 4> Constants(NumElts);
1103 for (unsigned i = 0; i != NumElts; ++i) {
1104 EVT EltVT = ValueVTs[i];
1105 if (isa<UndefValue>(C))
1106 Constants[i] = DAG.getUNDEF(EltVT);
1107 else if (EltVT.isFloatingPoint())
1108 Constants[i] = DAG.getConstantFP(0, getCurSDLoc(), EltVT);
1110 Constants[i] = DAG.getConstant(0, getCurSDLoc(), EltVT);
1113 return DAG.getMergeValues(Constants, getCurSDLoc());
1116 if (const BlockAddress *BA = dyn_cast<BlockAddress>(C))
1117 return DAG.getBlockAddress(BA, VT);
1119 VectorType *VecTy = cast<VectorType>(V->getType());
1120 unsigned NumElements = VecTy->getNumElements();
1122 // Now that we know the number and type of the elements, get that number of
1123 // elements into the Ops array based on what kind of constant it is.
1124 SmallVector<SDValue, 16> Ops;
1125 if (const ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
1126 for (unsigned i = 0; i != NumElements; ++i)
1127 Ops.push_back(getValue(CV->getOperand(i)));
1129 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!");
1130 EVT EltVT = TLI.getValueType(VecTy->getElementType());
1133 if (EltVT.isFloatingPoint())
1134 Op = DAG.getConstantFP(0, getCurSDLoc(), EltVT);
1136 Op = DAG.getConstant(0, getCurSDLoc(), EltVT);
1137 Ops.assign(NumElements, Op);
1140 // Create a BUILD_VECTOR node.
1141 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(), VT, Ops);
1144 // If this is a static alloca, generate it as the frameindex instead of
1146 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
1147 DenseMap<const AllocaInst*, int>::iterator SI =
1148 FuncInfo.StaticAllocaMap.find(AI);
1149 if (SI != FuncInfo.StaticAllocaMap.end())
1150 return DAG.getFrameIndex(SI->second, TLI.getPointerTy());
1153 // If this is an instruction which fast-isel has deferred, select it now.
1154 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
1155 unsigned InReg = FuncInfo.InitializeRegForValue(Inst);
1156 RegsForValue RFV(*DAG.getContext(), TLI, InReg, Inst->getType());
1157 SDValue Chain = DAG.getEntryNode();
1158 return RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, nullptr, V);
1161 llvm_unreachable("Can't get register for value!");
1164 void SelectionDAGBuilder::visitRet(const ReturnInst &I) {
1165 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1166 SDValue Chain = getControlRoot();
1167 SmallVector<ISD::OutputArg, 8> Outs;
1168 SmallVector<SDValue, 8> OutVals;
1170 if (!FuncInfo.CanLowerReturn) {
1171 unsigned DemoteReg = FuncInfo.DemoteRegister;
1172 const Function *F = I.getParent()->getParent();
1174 // Emit a store of the return value through the virtual register.
1175 // Leave Outs empty so that LowerReturn won't try to load return
1176 // registers the usual way.
1177 SmallVector<EVT, 1> PtrValueVTs;
1178 ComputeValueVTs(TLI, PointerType::getUnqual(F->getReturnType()),
1181 SDValue RetPtr = DAG.getRegister(DemoteReg, PtrValueVTs[0]);
1182 SDValue RetOp = getValue(I.getOperand(0));
1184 SmallVector<EVT, 4> ValueVTs;
1185 SmallVector<uint64_t, 4> Offsets;
1186 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs, &Offsets);
1187 unsigned NumValues = ValueVTs.size();
1189 SmallVector<SDValue, 4> Chains(NumValues);
1190 for (unsigned i = 0; i != NumValues; ++i) {
1191 SDValue Add = DAG.getNode(ISD::ADD, getCurSDLoc(),
1192 RetPtr.getValueType(), RetPtr,
1193 DAG.getIntPtrConstant(Offsets[i],
1196 DAG.getStore(Chain, getCurSDLoc(),
1197 SDValue(RetOp.getNode(), RetOp.getResNo() + i),
1198 // FIXME: better loc info would be nice.
1199 Add, MachinePointerInfo(), false, false, 0);
1202 Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(),
1203 MVT::Other, Chains);
1204 } else if (I.getNumOperands() != 0) {
1205 SmallVector<EVT, 4> ValueVTs;
1206 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs);
1207 unsigned NumValues = ValueVTs.size();
1209 SDValue RetOp = getValue(I.getOperand(0));
1211 const Function *F = I.getParent()->getParent();
1213 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1214 if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1216 ExtendKind = ISD::SIGN_EXTEND;
1217 else if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1219 ExtendKind = ISD::ZERO_EXTEND;
1221 LLVMContext &Context = F->getContext();
1222 bool RetInReg = F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1225 for (unsigned j = 0; j != NumValues; ++j) {
1226 EVT VT = ValueVTs[j];
1228 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger())
1229 VT = TLI.getTypeForExtArgOrReturn(Context, VT, ExtendKind);
1231 unsigned NumParts = TLI.getNumRegisters(Context, VT);
1232 MVT PartVT = TLI.getRegisterType(Context, VT);
1233 SmallVector<SDValue, 4> Parts(NumParts);
1234 getCopyToParts(DAG, getCurSDLoc(),
1235 SDValue(RetOp.getNode(), RetOp.getResNo() + j),
1236 &Parts[0], NumParts, PartVT, &I, ExtendKind);
1238 // 'inreg' on function refers to return value
1239 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1243 // Propagate extension type if any
1244 if (ExtendKind == ISD::SIGN_EXTEND)
1246 else if (ExtendKind == ISD::ZERO_EXTEND)
1249 for (unsigned i = 0; i < NumParts; ++i) {
1250 Outs.push_back(ISD::OutputArg(Flags, Parts[i].getValueType(),
1251 VT, /*isfixed=*/true, 0, 0));
1252 OutVals.push_back(Parts[i]);
1258 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1259 CallingConv::ID CallConv =
1260 DAG.getMachineFunction().getFunction()->getCallingConv();
1261 Chain = DAG.getTargetLoweringInfo().LowerReturn(
1262 Chain, CallConv, isVarArg, Outs, OutVals, getCurSDLoc(), DAG);
1264 // Verify that the target's LowerReturn behaved as expected.
1265 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
1266 "LowerReturn didn't return a valid chain!");
1268 // Update the DAG with the new chain value resulting from return lowering.
1272 /// CopyToExportRegsIfNeeded - If the given value has virtual registers
1273 /// created for it, emit nodes to copy the value into the virtual
1275 void SelectionDAGBuilder::CopyToExportRegsIfNeeded(const Value *V) {
1277 if (V->getType()->isEmptyTy())
1280 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
1281 if (VMI != FuncInfo.ValueMap.end()) {
1282 assert(!V->use_empty() && "Unused value assigned virtual registers!");
1283 CopyValueToVirtualRegister(V, VMI->second);
1287 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
1288 /// the current basic block, add it to ValueMap now so that we'll get a
1290 void SelectionDAGBuilder::ExportFromCurrentBlock(const Value *V) {
1291 // No need to export constants.
1292 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
1294 // Already exported?
1295 if (FuncInfo.isExportedInst(V)) return;
1297 unsigned Reg = FuncInfo.InitializeRegForValue(V);
1298 CopyValueToVirtualRegister(V, Reg);
1301 bool SelectionDAGBuilder::isExportableFromCurrentBlock(const Value *V,
1302 const BasicBlock *FromBB) {
1303 // The operands of the setcc have to be in this block. We don't know
1304 // how to export them from some other block.
1305 if (const Instruction *VI = dyn_cast<Instruction>(V)) {
1306 // Can export from current BB.
1307 if (VI->getParent() == FromBB)
1310 // Is already exported, noop.
1311 return FuncInfo.isExportedInst(V);
1314 // If this is an argument, we can export it if the BB is the entry block or
1315 // if it is already exported.
1316 if (isa<Argument>(V)) {
1317 if (FromBB == &FromBB->getParent()->getEntryBlock())
1320 // Otherwise, can only export this if it is already exported.
1321 return FuncInfo.isExportedInst(V);
1324 // Otherwise, constants can always be exported.
1328 /// Return branch probability calculated by BranchProbabilityInfo for IR blocks.
1329 uint32_t SelectionDAGBuilder::getEdgeWeight(const MachineBasicBlock *Src,
1330 const MachineBasicBlock *Dst) const {
1331 BranchProbabilityInfo *BPI = FuncInfo.BPI;
1334 const BasicBlock *SrcBB = Src->getBasicBlock();
1335 const BasicBlock *DstBB = Dst->getBasicBlock();
1336 return BPI->getEdgeWeight(SrcBB, DstBB);
1339 void SelectionDAGBuilder::
1340 addSuccessorWithWeight(MachineBasicBlock *Src, MachineBasicBlock *Dst,
1341 uint32_t Weight /* = 0 */) {
1343 Weight = getEdgeWeight(Src, Dst);
1344 Src->addSuccessor(Dst, Weight);
1348 static bool InBlock(const Value *V, const BasicBlock *BB) {
1349 if (const Instruction *I = dyn_cast<Instruction>(V))
1350 return I->getParent() == BB;
1354 /// EmitBranchForMergedCondition - Helper method for FindMergedConditions.
1355 /// This function emits a branch and is used at the leaves of an OR or an
1356 /// AND operator tree.
1359 SelectionDAGBuilder::EmitBranchForMergedCondition(const Value *Cond,
1360 MachineBasicBlock *TBB,
1361 MachineBasicBlock *FBB,
1362 MachineBasicBlock *CurBB,
1363 MachineBasicBlock *SwitchBB,
1366 const BasicBlock *BB = CurBB->getBasicBlock();
1368 // If the leaf of the tree is a comparison, merge the condition into
1370 if (const CmpInst *BOp = dyn_cast<CmpInst>(Cond)) {
1371 // The operands of the cmp have to be in this block. We don't know
1372 // how to export them from some other block. If this is the first block
1373 // of the sequence, no exporting is needed.
1374 if (CurBB == SwitchBB ||
1375 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
1376 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) {
1377 ISD::CondCode Condition;
1378 if (const ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
1379 Condition = getICmpCondCode(IC->getPredicate());
1380 } else if (const FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) {
1381 Condition = getFCmpCondCode(FC->getPredicate());
1382 if (TM.Options.NoNaNsFPMath)
1383 Condition = getFCmpCodeWithoutNaN(Condition);
1385 (void)Condition; // silence warning.
1386 llvm_unreachable("Unknown compare instruction");
1389 CaseBlock CB(Condition, BOp->getOperand(0), BOp->getOperand(1), nullptr,
1390 TBB, FBB, CurBB, TWeight, FWeight);
1391 SwitchCases.push_back(CB);
1396 // Create a CaseBlock record representing this branch.
1397 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()),
1398 nullptr, TBB, FBB, CurBB, TWeight, FWeight);
1399 SwitchCases.push_back(CB);
1402 /// Scale down both weights to fit into uint32_t.
1403 static void ScaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
1404 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
1405 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
1406 NewTrue = NewTrue / Scale;
1407 NewFalse = NewFalse / Scale;
1410 /// FindMergedConditions - If Cond is an expression like
1411 void SelectionDAGBuilder::FindMergedConditions(const Value *Cond,
1412 MachineBasicBlock *TBB,
1413 MachineBasicBlock *FBB,
1414 MachineBasicBlock *CurBB,
1415 MachineBasicBlock *SwitchBB,
1416 unsigned Opc, uint32_t TWeight,
1418 // If this node is not part of the or/and tree, emit it as a branch.
1419 const Instruction *BOp = dyn_cast<Instruction>(Cond);
1420 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
1421 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
1422 BOp->getParent() != CurBB->getBasicBlock() ||
1423 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
1424 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
1425 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB,
1430 // Create TmpBB after CurBB.
1431 MachineFunction::iterator BBI = CurBB;
1432 MachineFunction &MF = DAG.getMachineFunction();
1433 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock());
1434 CurBB->getParent()->insert(++BBI, TmpBB);
1436 if (Opc == Instruction::Or) {
1437 // Codegen X | Y as:
1446 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
1447 // The requirement is that
1448 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
1449 // = TrueProb for original BB.
1450 // Assuming the original weights are A and B, one choice is to set BB1's
1451 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
1453 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
1454 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
1455 // TmpBB, but the math is more complicated.
1457 uint64_t NewTrueWeight = TWeight;
1458 uint64_t NewFalseWeight = (uint64_t)TWeight + 2 * (uint64_t)FWeight;
1459 ScaleWeights(NewTrueWeight, NewFalseWeight);
1460 // Emit the LHS condition.
1461 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, SwitchBB, Opc,
1462 NewTrueWeight, NewFalseWeight);
1464 NewTrueWeight = TWeight;
1465 NewFalseWeight = 2 * (uint64_t)FWeight;
1466 ScaleWeights(NewTrueWeight, NewFalseWeight);
1467 // Emit the RHS condition into TmpBB.
1468 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc,
1469 NewTrueWeight, NewFalseWeight);
1471 assert(Opc == Instruction::And && "Unknown merge op!");
1472 // Codegen X & Y as:
1480 // This requires creation of TmpBB after CurBB.
1482 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
1483 // The requirement is that
1484 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
1485 // = FalseProb for original BB.
1486 // Assuming the original weights are A and B, one choice is to set BB1's
1487 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
1489 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
1491 uint64_t NewTrueWeight = 2 * (uint64_t)TWeight + (uint64_t)FWeight;
1492 uint64_t NewFalseWeight = FWeight;
1493 ScaleWeights(NewTrueWeight, NewFalseWeight);
1494 // Emit the LHS condition.
1495 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, SwitchBB, Opc,
1496 NewTrueWeight, NewFalseWeight);
1498 NewTrueWeight = 2 * (uint64_t)TWeight;
1499 NewFalseWeight = FWeight;
1500 ScaleWeights(NewTrueWeight, NewFalseWeight);
1501 // Emit the RHS condition into TmpBB.
1502 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc,
1503 NewTrueWeight, NewFalseWeight);
1507 /// If the set of cases should be emitted as a series of branches, return true.
1508 /// If we should emit this as a bunch of and/or'd together conditions, return
1511 SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases) {
1512 if (Cases.size() != 2) return true;
1514 // If this is two comparisons of the same values or'd or and'd together, they
1515 // will get folded into a single comparison, so don't emit two blocks.
1516 if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
1517 Cases[0].CmpRHS == Cases[1].CmpRHS) ||
1518 (Cases[0].CmpRHS == Cases[1].CmpLHS &&
1519 Cases[0].CmpLHS == Cases[1].CmpRHS)) {
1523 // Handle: (X != null) | (Y != null) --> (X|Y) != 0
1524 // Handle: (X == null) & (Y == null) --> (X|Y) == 0
1525 if (Cases[0].CmpRHS == Cases[1].CmpRHS &&
1526 Cases[0].CC == Cases[1].CC &&
1527 isa<Constant>(Cases[0].CmpRHS) &&
1528 cast<Constant>(Cases[0].CmpRHS)->isNullValue()) {
1529 if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB)
1531 if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB)
1538 void SelectionDAGBuilder::visitBr(const BranchInst &I) {
1539 MachineBasicBlock *BrMBB = FuncInfo.MBB;
1541 // Update machine-CFG edges.
1542 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
1544 if (I.isUnconditional()) {
1545 // Update machine-CFG edges.
1546 BrMBB->addSuccessor(Succ0MBB);
1548 // If this is not a fall-through branch or optimizations are switched off,
1550 if (Succ0MBB != NextBlock(BrMBB) || TM.getOptLevel() == CodeGenOpt::None)
1551 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(),
1552 MVT::Other, getControlRoot(),
1553 DAG.getBasicBlock(Succ0MBB)));
1558 // If this condition is one of the special cases we handle, do special stuff
1560 const Value *CondVal = I.getCondition();
1561 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
1563 // If this is a series of conditions that are or'd or and'd together, emit
1564 // this as a sequence of branches instead of setcc's with and/or operations.
1565 // As long as jumps are not expensive, this should improve performance.
1566 // For example, instead of something like:
1579 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
1580 if (!DAG.getTargetLoweringInfo().isJumpExpensive() &&
1581 BOp->hasOneUse() && (BOp->getOpcode() == Instruction::And ||
1582 BOp->getOpcode() == Instruction::Or)) {
1583 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB,
1584 BOp->getOpcode(), getEdgeWeight(BrMBB, Succ0MBB),
1585 getEdgeWeight(BrMBB, Succ1MBB));
1586 // If the compares in later blocks need to use values not currently
1587 // exported from this block, export them now. This block should always
1588 // be the first entry.
1589 assert(SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!");
1591 // Allow some cases to be rejected.
1592 if (ShouldEmitAsBranches(SwitchCases)) {
1593 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1594 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1595 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1598 // Emit the branch for this block.
1599 visitSwitchCase(SwitchCases[0], BrMBB);
1600 SwitchCases.erase(SwitchCases.begin());
1604 // Okay, we decided not to do this, remove any inserted MBB's and clear
1606 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
1607 FuncInfo.MF->erase(SwitchCases[i].ThisBB);
1609 SwitchCases.clear();
1613 // Create a CaseBlock record representing this branch.
1614 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()),
1615 nullptr, Succ0MBB, Succ1MBB, BrMBB);
1617 // Use visitSwitchCase to actually insert the fast branch sequence for this
1619 visitSwitchCase(CB, BrMBB);
1622 /// visitSwitchCase - Emits the necessary code to represent a single node in
1623 /// the binary search tree resulting from lowering a switch instruction.
1624 void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB,
1625 MachineBasicBlock *SwitchBB) {
1627 SDValue CondLHS = getValue(CB.CmpLHS);
1628 SDLoc dl = getCurSDLoc();
1630 // Build the setcc now.
1632 // Fold "(X == true)" to X and "(X == false)" to !X to
1633 // handle common cases produced by branch lowering.
1634 if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) &&
1635 CB.CC == ISD::SETEQ)
1637 else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) &&
1638 CB.CC == ISD::SETEQ) {
1639 SDValue True = DAG.getConstant(1, dl, CondLHS.getValueType());
1640 Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True);
1642 Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1644 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now");
1646 const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue();
1647 const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue();
1649 SDValue CmpOp = getValue(CB.CmpMHS);
1650 EVT VT = CmpOp.getValueType();
1652 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) {
1653 Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, dl, VT),
1656 SDValue SUB = DAG.getNode(ISD::SUB, dl,
1657 VT, CmpOp, DAG.getConstant(Low, dl, VT));
1658 Cond = DAG.getSetCC(dl, MVT::i1, SUB,
1659 DAG.getConstant(High-Low, dl, VT), ISD::SETULE);
1663 // Update successor info
1664 addSuccessorWithWeight(SwitchBB, CB.TrueBB, CB.TrueWeight);
1665 // TrueBB and FalseBB are always different unless the incoming IR is
1666 // degenerate. This only happens when running llc on weird IR.
1667 if (CB.TrueBB != CB.FalseBB)
1668 addSuccessorWithWeight(SwitchBB, CB.FalseBB, CB.FalseWeight);
1670 // If the lhs block is the next block, invert the condition so that we can
1671 // fall through to the lhs instead of the rhs block.
1672 if (CB.TrueBB == NextBlock(SwitchBB)) {
1673 std::swap(CB.TrueBB, CB.FalseBB);
1674 SDValue True = DAG.getConstant(1, dl, Cond.getValueType());
1675 Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True);
1678 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1679 MVT::Other, getControlRoot(), Cond,
1680 DAG.getBasicBlock(CB.TrueBB));
1682 // Insert the false branch. Do this even if it's a fall through branch,
1683 // this makes it easier to do DAG optimizations which require inverting
1684 // the branch condition.
1685 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1686 DAG.getBasicBlock(CB.FalseBB));
1688 DAG.setRoot(BrCond);
1691 /// visitJumpTable - Emit JumpTable node in the current MBB
1692 void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) {
1693 // Emit the code for the jump table
1694 assert(JT.Reg != -1U && "Should lower JT Header first!");
1695 EVT PTy = DAG.getTargetLoweringInfo().getPointerTy();
1696 SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurSDLoc(),
1698 SDValue Table = DAG.getJumpTable(JT.JTI, PTy);
1699 SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurSDLoc(),
1700 MVT::Other, Index.getValue(1),
1702 DAG.setRoot(BrJumpTable);
1705 /// visitJumpTableHeader - This function emits necessary code to produce index
1706 /// in the JumpTable from switch case.
1707 void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT,
1708 JumpTableHeader &JTH,
1709 MachineBasicBlock *SwitchBB) {
1710 SDLoc dl = getCurSDLoc();
1712 // Subtract the lowest switch case value from the value being switched on and
1713 // conditional branch to default mbb if the result is greater than the
1714 // difference between smallest and largest cases.
1715 SDValue SwitchOp = getValue(JTH.SValue);
1716 EVT VT = SwitchOp.getValueType();
1717 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, SwitchOp,
1718 DAG.getConstant(JTH.First, dl, VT));
1720 // The SDNode we just created, which holds the value being switched on minus
1721 // the smallest case value, needs to be copied to a virtual register so it
1722 // can be used as an index into the jump table in a subsequent basic block.
1723 // This value may be smaller or larger than the target's pointer type, and
1724 // therefore require extension or truncating.
1725 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1726 SwitchOp = DAG.getZExtOrTrunc(Sub, dl, TLI.getPointerTy());
1728 unsigned JumpTableReg = FuncInfo.CreateReg(TLI.getPointerTy());
1729 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), dl,
1730 JumpTableReg, SwitchOp);
1731 JT.Reg = JumpTableReg;
1733 // Emit the range check for the jump table, and branch to the default block
1734 // for the switch statement if the value being switched on exceeds the largest
1735 // case in the switch.
1737 DAG.getSetCC(dl, TLI.getSetCCResultType(*DAG.getContext(),
1738 Sub.getValueType()),
1739 Sub, DAG.getConstant(JTH.Last - JTH.First, dl, VT),
1742 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1743 MVT::Other, CopyTo, CMP,
1744 DAG.getBasicBlock(JT.Default));
1746 // Avoid emitting unnecessary branches to the next block.
1747 if (JT.MBB != NextBlock(SwitchBB))
1748 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1749 DAG.getBasicBlock(JT.MBB));
1751 DAG.setRoot(BrCond);
1754 /// Codegen a new tail for a stack protector check ParentMBB which has had its
1755 /// tail spliced into a stack protector check success bb.
1757 /// For a high level explanation of how this fits into the stack protector
1758 /// generation see the comment on the declaration of class
1759 /// StackProtectorDescriptor.
1760 void SelectionDAGBuilder::visitSPDescriptorParent(StackProtectorDescriptor &SPD,
1761 MachineBasicBlock *ParentBB) {
1763 // First create the loads to the guard/stack slot for the comparison.
1764 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1765 EVT PtrTy = TLI.getPointerTy();
1767 MachineFrameInfo *MFI = ParentBB->getParent()->getFrameInfo();
1768 int FI = MFI->getStackProtectorIndex();
1770 const Value *IRGuard = SPD.getGuard();
1771 SDValue GuardPtr = getValue(IRGuard);
1772 SDValue StackSlotPtr = DAG.getFrameIndex(FI, PtrTy);
1775 TLI.getDataLayout()->getPrefTypeAlignment(IRGuard->getType());
1778 SDLoc dl = getCurSDLoc();
1780 // If GuardReg is set and useLoadStackGuardNode returns true, retrieve the
1781 // guard value from the virtual register holding the value. Otherwise, emit a
1782 // volatile load to retrieve the stack guard value.
1783 unsigned GuardReg = SPD.getGuardReg();
1785 if (GuardReg && TLI.useLoadStackGuardNode())
1786 Guard = DAG.getCopyFromReg(DAG.getEntryNode(), dl, GuardReg,
1789 Guard = DAG.getLoad(PtrTy, dl, DAG.getEntryNode(),
1790 GuardPtr, MachinePointerInfo(IRGuard, 0),
1791 true, false, false, Align);
1793 SDValue StackSlot = DAG.getLoad(PtrTy, dl, DAG.getEntryNode(),
1795 MachinePointerInfo::getFixedStack(FI),
1796 true, false, false, Align);
1798 // Perform the comparison via a subtract/getsetcc.
1799 EVT VT = Guard.getValueType();
1800 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, Guard, StackSlot);
1803 DAG.getSetCC(dl, TLI.getSetCCResultType(*DAG.getContext(),
1804 Sub.getValueType()),
1805 Sub, DAG.getConstant(0, dl, VT), ISD::SETNE);
1807 // If the sub is not 0, then we know the guard/stackslot do not equal, so
1808 // branch to failure MBB.
1809 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1810 MVT::Other, StackSlot.getOperand(0),
1811 Cmp, DAG.getBasicBlock(SPD.getFailureMBB()));
1812 // Otherwise branch to success MBB.
1813 SDValue Br = DAG.getNode(ISD::BR, dl,
1815 DAG.getBasicBlock(SPD.getSuccessMBB()));
1820 /// Codegen the failure basic block for a stack protector check.
1822 /// A failure stack protector machine basic block consists simply of a call to
1823 /// __stack_chk_fail().
1825 /// For a high level explanation of how this fits into the stack protector
1826 /// generation see the comment on the declaration of class
1827 /// StackProtectorDescriptor.
1829 SelectionDAGBuilder::visitSPDescriptorFailure(StackProtectorDescriptor &SPD) {
1830 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1832 TLI.makeLibCall(DAG, RTLIB::STACKPROTECTOR_CHECK_FAIL, MVT::isVoid,
1833 nullptr, 0, false, getCurSDLoc(), false, false).second;
1837 /// visitBitTestHeader - This function emits necessary code to produce value
1838 /// suitable for "bit tests"
1839 void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B,
1840 MachineBasicBlock *SwitchBB) {
1841 SDLoc dl = getCurSDLoc();
1843 // Subtract the minimum value
1844 SDValue SwitchOp = getValue(B.SValue);
1845 EVT VT = SwitchOp.getValueType();
1846 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, SwitchOp,
1847 DAG.getConstant(B.First, dl, VT));
1850 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1852 DAG.getSetCC(dl, TLI.getSetCCResultType(*DAG.getContext(),
1853 Sub.getValueType()),
1854 Sub, DAG.getConstant(B.Range, dl, VT), ISD::SETUGT);
1856 // Determine the type of the test operands.
1857 bool UsePtrType = false;
1858 if (!TLI.isTypeLegal(VT))
1861 for (unsigned i = 0, e = B.Cases.size(); i != e; ++i)
1862 if (!isUIntN(VT.getSizeInBits(), B.Cases[i].Mask)) {
1863 // Switch table case range are encoded into series of masks.
1864 // Just use pointer type, it's guaranteed to fit.
1870 VT = TLI.getPointerTy();
1871 Sub = DAG.getZExtOrTrunc(Sub, dl, VT);
1874 B.RegVT = VT.getSimpleVT();
1875 B.Reg = FuncInfo.CreateReg(B.RegVT);
1876 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), dl, B.Reg, Sub);
1878 MachineBasicBlock* MBB = B.Cases[0].ThisBB;
1880 addSuccessorWithWeight(SwitchBB, B.Default);
1881 addSuccessorWithWeight(SwitchBB, MBB);
1883 SDValue BrRange = DAG.getNode(ISD::BRCOND, dl,
1884 MVT::Other, CopyTo, RangeCmp,
1885 DAG.getBasicBlock(B.Default));
1887 // Avoid emitting unnecessary branches to the next block.
1888 if (MBB != NextBlock(SwitchBB))
1889 BrRange = DAG.getNode(ISD::BR, dl, MVT::Other, BrRange,
1890 DAG.getBasicBlock(MBB));
1892 DAG.setRoot(BrRange);
1895 /// visitBitTestCase - this function produces one "bit test"
1896 void SelectionDAGBuilder::visitBitTestCase(BitTestBlock &BB,
1897 MachineBasicBlock* NextMBB,
1898 uint32_t BranchWeightToNext,
1901 MachineBasicBlock *SwitchBB) {
1902 SDLoc dl = getCurSDLoc();
1904 SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), dl, Reg, VT);
1906 unsigned PopCount = countPopulation(B.Mask);
1907 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1908 if (PopCount == 1) {
1909 // Testing for a single bit; just compare the shift count with what it
1910 // would need to be to shift a 1 bit in that position.
1912 dl, TLI.getSetCCResultType(*DAG.getContext(), VT), ShiftOp,
1913 DAG.getConstant(countTrailingZeros(B.Mask), dl, VT), ISD::SETEQ);
1914 } else if (PopCount == BB.Range) {
1915 // There is only one zero bit in the range, test for it directly.
1917 dl, TLI.getSetCCResultType(*DAG.getContext(), VT), ShiftOp,
1918 DAG.getConstant(countTrailingOnes(B.Mask), dl, VT), ISD::SETNE);
1920 // Make desired shift
1921 SDValue SwitchVal = DAG.getNode(ISD::SHL, dl, VT,
1922 DAG.getConstant(1, dl, VT), ShiftOp);
1924 // Emit bit tests and jumps
1925 SDValue AndOp = DAG.getNode(ISD::AND, dl,
1926 VT, SwitchVal, DAG.getConstant(B.Mask, dl, VT));
1927 Cmp = DAG.getSetCC(dl, TLI.getSetCCResultType(*DAG.getContext(), VT), AndOp,
1928 DAG.getConstant(0, dl, VT), ISD::SETNE);
1931 // The branch weight from SwitchBB to B.TargetBB is B.ExtraWeight.
1932 addSuccessorWithWeight(SwitchBB, B.TargetBB, B.ExtraWeight);
1933 // The branch weight from SwitchBB to NextMBB is BranchWeightToNext.
1934 addSuccessorWithWeight(SwitchBB, NextMBB, BranchWeightToNext);
1936 SDValue BrAnd = DAG.getNode(ISD::BRCOND, dl,
1937 MVT::Other, getControlRoot(),
1938 Cmp, DAG.getBasicBlock(B.TargetBB));
1940 // Avoid emitting unnecessary branches to the next block.
1941 if (NextMBB != NextBlock(SwitchBB))
1942 BrAnd = DAG.getNode(ISD::BR, dl, MVT::Other, BrAnd,
1943 DAG.getBasicBlock(NextMBB));
1948 void SelectionDAGBuilder::visitInvoke(const InvokeInst &I) {
1949 MachineBasicBlock *InvokeMBB = FuncInfo.MBB;
1951 // Retrieve successors.
1952 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
1953 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)];
1955 const Value *Callee(I.getCalledValue());
1956 const Function *Fn = dyn_cast<Function>(Callee);
1957 if (isa<InlineAsm>(Callee))
1959 else if (Fn && Fn->isIntrinsic()) {
1960 switch (Fn->getIntrinsicID()) {
1962 llvm_unreachable("Cannot invoke this intrinsic");
1963 case Intrinsic::donothing:
1964 // Ignore invokes to @llvm.donothing: jump directly to the next BB.
1966 case Intrinsic::experimental_patchpoint_void:
1967 case Intrinsic::experimental_patchpoint_i64:
1968 visitPatchpoint(&I, LandingPad);
1970 case Intrinsic::experimental_gc_statepoint:
1971 LowerStatepoint(ImmutableStatepoint(&I), LandingPad);
1975 LowerCallTo(&I, getValue(Callee), false, LandingPad);
1977 // If the value of the invoke is used outside of its defining block, make it
1978 // available as a virtual register.
1979 // We already took care of the exported value for the statepoint instruction
1980 // during call to the LowerStatepoint.
1981 if (!isStatepoint(I)) {
1982 CopyToExportRegsIfNeeded(&I);
1985 // Update successor info
1986 addSuccessorWithWeight(InvokeMBB, Return);
1987 addSuccessorWithWeight(InvokeMBB, LandingPad);
1989 // Drop into normal successor.
1990 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(),
1991 MVT::Other, getControlRoot(),
1992 DAG.getBasicBlock(Return)));
1995 void SelectionDAGBuilder::visitResume(const ResumeInst &RI) {
1996 llvm_unreachable("SelectionDAGBuilder shouldn't visit resume instructions!");
1999 void SelectionDAGBuilder::visitLandingPad(const LandingPadInst &LP) {
2000 assert(FuncInfo.MBB->isLandingPad() &&
2001 "Call to landingpad not in landing pad!");
2003 MachineBasicBlock *MBB = FuncInfo.MBB;
2004 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
2005 AddLandingPadInfo(LP, MMI, MBB);
2007 // If there aren't registers to copy the values into (e.g., during SjLj
2008 // exceptions), then don't bother to create these DAG nodes.
2009 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2010 if (TLI.getExceptionPointerRegister() == 0 &&
2011 TLI.getExceptionSelectorRegister() == 0)
2014 SmallVector<EVT, 2> ValueVTs;
2015 SDLoc dl = getCurSDLoc();
2016 ComputeValueVTs(TLI, LP.getType(), ValueVTs);
2017 assert(ValueVTs.size() == 2 && "Only two-valued landingpads are supported");
2019 // Get the two live-in registers as SDValues. The physregs have already been
2020 // copied into virtual registers.
2022 if (FuncInfo.ExceptionPointerVirtReg) {
2023 Ops[0] = DAG.getZExtOrTrunc(
2024 DAG.getCopyFromReg(DAG.getEntryNode(), dl,
2025 FuncInfo.ExceptionPointerVirtReg, TLI.getPointerTy()),
2028 Ops[0] = DAG.getConstant(0, dl, TLI.getPointerTy());
2030 Ops[1] = DAG.getZExtOrTrunc(
2031 DAG.getCopyFromReg(DAG.getEntryNode(), dl,
2032 FuncInfo.ExceptionSelectorVirtReg, TLI.getPointerTy()),
2036 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, dl,
2037 DAG.getVTList(ValueVTs), Ops);
2042 SelectionDAGBuilder::visitLandingPadClauseBB(GlobalValue *ClauseGV,
2043 MachineBasicBlock *LPadBB) {
2044 SDValue Chain = getControlRoot();
2045 SDLoc dl = getCurSDLoc();
2047 // Get the typeid that we will dispatch on later.
2048 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2049 const TargetRegisterClass *RC = TLI.getRegClassFor(TLI.getPointerTy());
2050 unsigned VReg = FuncInfo.MF->getRegInfo().createVirtualRegister(RC);
2051 unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(ClauseGV);
2052 SDValue Sel = DAG.getConstant(TypeID, dl, TLI.getPointerTy());
2053 Chain = DAG.getCopyToReg(Chain, dl, VReg, Sel);
2055 // Branch to the main landing pad block.
2056 MachineBasicBlock *ClauseMBB = FuncInfo.MBB;
2057 ClauseMBB->addSuccessor(LPadBB);
2058 DAG.setRoot(DAG.getNode(ISD::BR, dl, MVT::Other, Chain,
2059 DAG.getBasicBlock(LPadBB)));
2063 void SelectionDAGBuilder::sortAndRangeify(CaseClusterVector &Clusters) {
2065 for (const CaseCluster &CC : Clusters)
2066 assert(CC.Low == CC.High && "Input clusters must be single-case");
2069 std::sort(Clusters.begin(), Clusters.end(),
2070 [](const CaseCluster &a, const CaseCluster &b) {
2071 return a.Low->getValue().slt(b.Low->getValue());
2074 // Merge adjacent clusters with the same destination.
2075 const unsigned N = Clusters.size();
2076 unsigned DstIndex = 0;
2077 for (unsigned SrcIndex = 0; SrcIndex < N; ++SrcIndex) {
2078 CaseCluster &CC = Clusters[SrcIndex];
2079 const ConstantInt *CaseVal = CC.Low;
2080 MachineBasicBlock *Succ = CC.MBB;
2082 if (DstIndex != 0 && Clusters[DstIndex - 1].MBB == Succ &&
2083 (CaseVal->getValue() - Clusters[DstIndex - 1].High->getValue()) == 1) {
2084 // If this case has the same successor and is a neighbour, merge it into
2085 // the previous cluster.
2086 Clusters[DstIndex - 1].High = CaseVal;
2087 Clusters[DstIndex - 1].Weight += CC.Weight;
2088 assert(Clusters[DstIndex - 1].Weight >= CC.Weight && "Weight overflow!");
2090 std::memmove(&Clusters[DstIndex++], &Clusters[SrcIndex],
2091 sizeof(Clusters[SrcIndex]));
2094 Clusters.resize(DstIndex);
2097 void SelectionDAGBuilder::UpdateSplitBlock(MachineBasicBlock *First,
2098 MachineBasicBlock *Last) {
2100 for (unsigned i = 0, e = JTCases.size(); i != e; ++i)
2101 if (JTCases[i].first.HeaderBB == First)
2102 JTCases[i].first.HeaderBB = Last;
2104 // Update BitTestCases.
2105 for (unsigned i = 0, e = BitTestCases.size(); i != e; ++i)
2106 if (BitTestCases[i].Parent == First)
2107 BitTestCases[i].Parent = Last;
2110 void SelectionDAGBuilder::visitIndirectBr(const IndirectBrInst &I) {
2111 MachineBasicBlock *IndirectBrMBB = FuncInfo.MBB;
2113 // Update machine-CFG edges with unique successors.
2114 SmallSet<BasicBlock*, 32> Done;
2115 for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i) {
2116 BasicBlock *BB = I.getSuccessor(i);
2117 bool Inserted = Done.insert(BB).second;
2121 MachineBasicBlock *Succ = FuncInfo.MBBMap[BB];
2122 addSuccessorWithWeight(IndirectBrMBB, Succ);
2125 DAG.setRoot(DAG.getNode(ISD::BRIND, getCurSDLoc(),
2126 MVT::Other, getControlRoot(),
2127 getValue(I.getAddress())));
2130 void SelectionDAGBuilder::visitUnreachable(const UnreachableInst &I) {
2131 if (DAG.getTarget().Options.TrapUnreachable)
2132 DAG.setRoot(DAG.getNode(ISD::TRAP, getCurSDLoc(), MVT::Other, DAG.getRoot()));
2135 void SelectionDAGBuilder::visitFSub(const User &I) {
2136 // -0.0 - X --> fneg
2137 Type *Ty = I.getType();
2138 if (isa<Constant>(I.getOperand(0)) &&
2139 I.getOperand(0) == ConstantFP::getZeroValueForNegation(Ty)) {
2140 SDValue Op2 = getValue(I.getOperand(1));
2141 setValue(&I, DAG.getNode(ISD::FNEG, getCurSDLoc(),
2142 Op2.getValueType(), Op2));
2146 visitBinary(I, ISD::FSUB);
2149 void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) {
2150 SDValue Op1 = getValue(I.getOperand(0));
2151 SDValue Op2 = getValue(I.getOperand(1));
2158 if (const OverflowingBinaryOperator *OFBinOp =
2159 dyn_cast<const OverflowingBinaryOperator>(&I)) {
2160 nuw = OFBinOp->hasNoUnsignedWrap();
2161 nsw = OFBinOp->hasNoSignedWrap();
2163 if (const PossiblyExactOperator *ExactOp =
2164 dyn_cast<const PossiblyExactOperator>(&I))
2165 exact = ExactOp->isExact();
2166 if (const FPMathOperator *FPOp = dyn_cast<const FPMathOperator>(&I))
2167 FMF = FPOp->getFastMathFlags();
2170 Flags.setExact(exact);
2171 Flags.setNoSignedWrap(nsw);
2172 Flags.setNoUnsignedWrap(nuw);
2173 if (EnableFMFInDAG) {
2174 Flags.setAllowReciprocal(FMF.allowReciprocal());
2175 Flags.setNoInfs(FMF.noInfs());
2176 Flags.setNoNaNs(FMF.noNaNs());
2177 Flags.setNoSignedZeros(FMF.noSignedZeros());
2178 Flags.setUnsafeAlgebra(FMF.unsafeAlgebra());
2180 SDValue BinNodeValue = DAG.getNode(OpCode, getCurSDLoc(), Op1.getValueType(),
2182 setValue(&I, BinNodeValue);
2185 void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) {
2186 SDValue Op1 = getValue(I.getOperand(0));
2187 SDValue Op2 = getValue(I.getOperand(1));
2190 DAG.getTargetLoweringInfo().getShiftAmountTy(Op2.getValueType());
2192 // Coerce the shift amount to the right type if we can.
2193 if (!I.getType()->isVectorTy() && Op2.getValueType() != ShiftTy) {
2194 unsigned ShiftSize = ShiftTy.getSizeInBits();
2195 unsigned Op2Size = Op2.getValueType().getSizeInBits();
2196 SDLoc DL = getCurSDLoc();
2198 // If the operand is smaller than the shift count type, promote it.
2199 if (ShiftSize > Op2Size)
2200 Op2 = DAG.getNode(ISD::ZERO_EXTEND, DL, ShiftTy, Op2);
2202 // If the operand is larger than the shift count type but the shift
2203 // count type has enough bits to represent any shift value, truncate
2204 // it now. This is a common case and it exposes the truncate to
2205 // optimization early.
2206 else if (ShiftSize >= Log2_32_Ceil(Op2.getValueType().getSizeInBits()))
2207 Op2 = DAG.getNode(ISD::TRUNCATE, DL, ShiftTy, Op2);
2208 // Otherwise we'll need to temporarily settle for some other convenient
2209 // type. Type legalization will make adjustments once the shiftee is split.
2211 Op2 = DAG.getZExtOrTrunc(Op2, DL, MVT::i32);
2218 if (Opcode == ISD::SRL || Opcode == ISD::SRA || Opcode == ISD::SHL) {
2220 if (const OverflowingBinaryOperator *OFBinOp =
2221 dyn_cast<const OverflowingBinaryOperator>(&I)) {
2222 nuw = OFBinOp->hasNoUnsignedWrap();
2223 nsw = OFBinOp->hasNoSignedWrap();
2225 if (const PossiblyExactOperator *ExactOp =
2226 dyn_cast<const PossiblyExactOperator>(&I))
2227 exact = ExactOp->isExact();
2230 Flags.setExact(exact);
2231 Flags.setNoSignedWrap(nsw);
2232 Flags.setNoUnsignedWrap(nuw);
2233 SDValue Res = DAG.getNode(Opcode, getCurSDLoc(), Op1.getValueType(), Op1, Op2,
2238 void SelectionDAGBuilder::visitSDiv(const User &I) {
2239 SDValue Op1 = getValue(I.getOperand(0));
2240 SDValue Op2 = getValue(I.getOperand(1));
2243 Flags.setExact(isa<PossiblyExactOperator>(&I) &&
2244 cast<PossiblyExactOperator>(&I)->isExact());
2245 setValue(&I, DAG.getNode(ISD::SDIV, getCurSDLoc(), Op1.getValueType(), Op1,
2249 void SelectionDAGBuilder::visitICmp(const User &I) {
2250 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2251 if (const ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2252 predicate = IC->getPredicate();
2253 else if (const ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2254 predicate = ICmpInst::Predicate(IC->getPredicate());
2255 SDValue Op1 = getValue(I.getOperand(0));
2256 SDValue Op2 = getValue(I.getOperand(1));
2257 ISD::CondCode Opcode = getICmpCondCode(predicate);
2259 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType());
2260 setValue(&I, DAG.getSetCC(getCurSDLoc(), DestVT, Op1, Op2, Opcode));
2263 void SelectionDAGBuilder::visitFCmp(const User &I) {
2264 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2265 if (const FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2266 predicate = FC->getPredicate();
2267 else if (const ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2268 predicate = FCmpInst::Predicate(FC->getPredicate());
2269 SDValue Op1 = getValue(I.getOperand(0));
2270 SDValue Op2 = getValue(I.getOperand(1));
2271 ISD::CondCode Condition = getFCmpCondCode(predicate);
2272 if (TM.Options.NoNaNsFPMath)
2273 Condition = getFCmpCodeWithoutNaN(Condition);
2274 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType());
2275 setValue(&I, DAG.getSetCC(getCurSDLoc(), DestVT, Op1, Op2, Condition));
2278 void SelectionDAGBuilder::visitSelect(const User &I) {
2279 SmallVector<EVT, 4> ValueVTs;
2280 ComputeValueVTs(DAG.getTargetLoweringInfo(), I.getType(), ValueVTs);
2281 unsigned NumValues = ValueVTs.size();
2282 if (NumValues == 0) return;
2284 SmallVector<SDValue, 4> Values(NumValues);
2285 SDValue Cond = getValue(I.getOperand(0));
2286 SDValue LHSVal = getValue(I.getOperand(1));
2287 SDValue RHSVal = getValue(I.getOperand(2));
2288 auto BaseOps = {Cond};
2289 ISD::NodeType OpCode = Cond.getValueType().isVector() ?
2290 ISD::VSELECT : ISD::SELECT;
2292 // Min/max matching is only viable if all output VTs are the same.
2293 if (std::equal(ValueVTs.begin(), ValueVTs.end(), ValueVTs.begin())) {
2295 SelectPatternFlavor SPF = matchSelectPattern(const_cast<User*>(&I), LHS, RHS);
2296 ISD::NodeType Opc = ISD::DELETED_NODE;
2298 case SPF_UMAX: Opc = ISD::UMAX; break;
2299 case SPF_UMIN: Opc = ISD::UMIN; break;
2300 case SPF_SMAX: Opc = ISD::SMAX; break;
2301 case SPF_SMIN: Opc = ISD::SMIN; break;
2305 EVT VT = ValueVTs[0];
2306 LLVMContext &Ctx = *DAG.getContext();
2307 auto &TLI = DAG.getTargetLoweringInfo();
2308 while (TLI.getTypeAction(Ctx, VT) == TargetLoweringBase::TypeSplitVector)
2309 VT = TLI.getTypeToTransformTo(Ctx, VT);
2311 if (Opc != ISD::DELETED_NODE && TLI.isOperationLegalOrCustom(Opc, VT) &&
2312 // If the underlying comparison instruction is used by any other instruction,
2313 // the consumed instructions won't be destroyed, so it is not profitable
2314 // to convert to a min/max.
2315 cast<SelectInst>(&I)->getCondition()->hasOneUse()) {
2317 LHSVal = getValue(LHS);
2318 RHSVal = getValue(RHS);
2323 for (unsigned i = 0; i != NumValues; ++i) {
2324 SmallVector<SDValue, 3> Ops(BaseOps.begin(), BaseOps.end());
2325 Ops.push_back(SDValue(LHSVal.getNode(), LHSVal.getResNo() + i));
2326 Ops.push_back(SDValue(RHSVal.getNode(), RHSVal.getResNo() + i));
2327 Values[i] = DAG.getNode(OpCode, getCurSDLoc(),
2328 LHSVal.getNode()->getValueType(LHSVal.getResNo()+i),
2332 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
2333 DAG.getVTList(ValueVTs), Values));
2336 void SelectionDAGBuilder::visitTrunc(const User &I) {
2337 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2338 SDValue N = getValue(I.getOperand(0));
2339 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType());
2340 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), DestVT, N));
2343 void SelectionDAGBuilder::visitZExt(const User &I) {
2344 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2345 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2346 SDValue N = getValue(I.getOperand(0));
2347 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType());
2348 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurSDLoc(), DestVT, N));
2351 void SelectionDAGBuilder::visitSExt(const User &I) {
2352 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2353 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2354 SDValue N = getValue(I.getOperand(0));
2355 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType());
2356 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurSDLoc(), DestVT, N));
2359 void SelectionDAGBuilder::visitFPTrunc(const User &I) {
2360 // FPTrunc is never a no-op cast, no need to check
2361 SDValue N = getValue(I.getOperand(0));
2362 SDLoc dl = getCurSDLoc();
2363 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2364 EVT DestVT = TLI.getValueType(I.getType());
2365 setValue(&I, DAG.getNode(ISD::FP_ROUND, dl, DestVT, N,
2366 DAG.getTargetConstant(0, dl, TLI.getPointerTy())));
2369 void SelectionDAGBuilder::visitFPExt(const User &I) {
2370 // FPExt is never a no-op cast, no need to check
2371 SDValue N = getValue(I.getOperand(0));
2372 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType());
2373 setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurSDLoc(), DestVT, N));
2376 void SelectionDAGBuilder::visitFPToUI(const User &I) {
2377 // FPToUI is never a no-op cast, no need to check
2378 SDValue N = getValue(I.getOperand(0));
2379 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType());
2380 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurSDLoc(), DestVT, N));
2383 void SelectionDAGBuilder::visitFPToSI(const User &I) {
2384 // FPToSI is never a no-op cast, no need to check
2385 SDValue N = getValue(I.getOperand(0));
2386 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType());
2387 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurSDLoc(), DestVT, N));
2390 void SelectionDAGBuilder::visitUIToFP(const User &I) {
2391 // UIToFP is never a no-op cast, no need to check
2392 SDValue N = getValue(I.getOperand(0));
2393 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType());
2394 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurSDLoc(), DestVT, N));
2397 void SelectionDAGBuilder::visitSIToFP(const User &I) {
2398 // SIToFP is never a no-op cast, no need to check
2399 SDValue N = getValue(I.getOperand(0));
2400 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType());
2401 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurSDLoc(), DestVT, N));
2404 void SelectionDAGBuilder::visitPtrToInt(const User &I) {
2405 // What to do depends on the size of the integer and the size of the pointer.
2406 // We can either truncate, zero extend, or no-op, accordingly.
2407 SDValue N = getValue(I.getOperand(0));
2408 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType());
2409 setValue(&I, DAG.getZExtOrTrunc(N, getCurSDLoc(), DestVT));
2412 void SelectionDAGBuilder::visitIntToPtr(const User &I) {
2413 // What to do depends on the size of the integer and the size of the pointer.
2414 // We can either truncate, zero extend, or no-op, accordingly.
2415 SDValue N = getValue(I.getOperand(0));
2416 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType());
2417 setValue(&I, DAG.getZExtOrTrunc(N, getCurSDLoc(), DestVT));
2420 void SelectionDAGBuilder::visitBitCast(const User &I) {
2421 SDValue N = getValue(I.getOperand(0));
2422 SDLoc dl = getCurSDLoc();
2423 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(I.getType());
2425 // BitCast assures us that source and destination are the same size so this is
2426 // either a BITCAST or a no-op.
2427 if (DestVT != N.getValueType())
2428 setValue(&I, DAG.getNode(ISD::BITCAST, dl,
2429 DestVT, N)); // convert types.
2430 // Check if the original LLVM IR Operand was a ConstantInt, because getValue()
2431 // might fold any kind of constant expression to an integer constant and that
2432 // is not what we are looking for. Only regcognize a bitcast of a genuine
2433 // constant integer as an opaque constant.
2434 else if(ConstantInt *C = dyn_cast<ConstantInt>(I.getOperand(0)))
2435 setValue(&I, DAG.getConstant(C->getValue(), dl, DestVT, /*isTarget=*/false,
2438 setValue(&I, N); // noop cast.
2441 void SelectionDAGBuilder::visitAddrSpaceCast(const User &I) {
2442 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2443 const Value *SV = I.getOperand(0);
2444 SDValue N = getValue(SV);
2445 EVT DestVT = TLI.getValueType(I.getType());
2447 unsigned SrcAS = SV->getType()->getPointerAddressSpace();
2448 unsigned DestAS = I.getType()->getPointerAddressSpace();
2450 if (!TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
2451 N = DAG.getAddrSpaceCast(getCurSDLoc(), DestVT, N, SrcAS, DestAS);
2456 void SelectionDAGBuilder::visitInsertElement(const User &I) {
2457 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2458 SDValue InVec = getValue(I.getOperand(0));
2459 SDValue InVal = getValue(I.getOperand(1));
2460 SDValue InIdx = DAG.getSExtOrTrunc(getValue(I.getOperand(2)),
2461 getCurSDLoc(), TLI.getVectorIdxTy());
2462 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurSDLoc(),
2463 TLI.getValueType(I.getType()), InVec, InVal, InIdx));
2466 void SelectionDAGBuilder::visitExtractElement(const User &I) {
2467 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2468 SDValue InVec = getValue(I.getOperand(0));
2469 SDValue InIdx = DAG.getSExtOrTrunc(getValue(I.getOperand(1)),
2470 getCurSDLoc(), TLI.getVectorIdxTy());
2471 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurSDLoc(),
2472 TLI.getValueType(I.getType()), InVec, InIdx));
2475 // Utility for visitShuffleVector - Return true if every element in Mask,
2476 // beginning from position Pos and ending in Pos+Size, falls within the
2477 // specified sequential range [L, L+Pos). or is undef.
2478 static bool isSequentialInRange(const SmallVectorImpl<int> &Mask,
2479 unsigned Pos, unsigned Size, int Low) {
2480 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
2481 if (Mask[i] >= 0 && Mask[i] != Low)
2486 void SelectionDAGBuilder::visitShuffleVector(const User &I) {
2487 SDValue Src1 = getValue(I.getOperand(0));
2488 SDValue Src2 = getValue(I.getOperand(1));
2490 SmallVector<int, 8> Mask;
2491 ShuffleVectorInst::getShuffleMask(cast<Constant>(I.getOperand(2)), Mask);
2492 unsigned MaskNumElts = Mask.size();
2494 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2495 EVT VT = TLI.getValueType(I.getType());
2496 EVT SrcVT = Src1.getValueType();
2497 unsigned SrcNumElts = SrcVT.getVectorNumElements();
2499 if (SrcNumElts == MaskNumElts) {
2500 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
2505 // Normalize the shuffle vector since mask and vector length don't match.
2506 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) {
2507 // Mask is longer than the source vectors and is a multiple of the source
2508 // vectors. We can use concatenate vector to make the mask and vectors
2510 if (SrcNumElts*2 == MaskNumElts) {
2511 // First check for Src1 in low and Src2 in high
2512 if (isSequentialInRange(Mask, 0, SrcNumElts, 0) &&
2513 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, SrcNumElts)) {
2514 // The shuffle is concatenating two vectors together.
2515 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurSDLoc(),
2519 // Then check for Src2 in low and Src1 in high
2520 if (isSequentialInRange(Mask, 0, SrcNumElts, SrcNumElts) &&
2521 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, 0)) {
2522 // The shuffle is concatenating two vectors together.
2523 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurSDLoc(),
2529 // Pad both vectors with undefs to make them the same length as the mask.
2530 unsigned NumConcat = MaskNumElts / SrcNumElts;
2531 bool Src1U = Src1.getOpcode() == ISD::UNDEF;
2532 bool Src2U = Src2.getOpcode() == ISD::UNDEF;
2533 SDValue UndefVal = DAG.getUNDEF(SrcVT);
2535 SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal);
2536 SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal);
2540 Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2541 getCurSDLoc(), VT, MOps1);
2542 Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2543 getCurSDLoc(), VT, MOps2);
2545 // Readjust mask for new input vector length.
2546 SmallVector<int, 8> MappedOps;
2547 for (unsigned i = 0; i != MaskNumElts; ++i) {
2549 if (Idx >= (int)SrcNumElts)
2550 Idx -= SrcNumElts - MaskNumElts;
2551 MappedOps.push_back(Idx);
2554 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
2559 if (SrcNumElts > MaskNumElts) {
2560 // Analyze the access pattern of the vector to see if we can extract
2561 // two subvectors and do the shuffle. The analysis is done by calculating
2562 // the range of elements the mask access on both vectors.
2563 int MinRange[2] = { static_cast<int>(SrcNumElts),
2564 static_cast<int>(SrcNumElts)};
2565 int MaxRange[2] = {-1, -1};
2567 for (unsigned i = 0; i != MaskNumElts; ++i) {
2573 if (Idx >= (int)SrcNumElts) {
2577 if (Idx > MaxRange[Input])
2578 MaxRange[Input] = Idx;
2579 if (Idx < MinRange[Input])
2580 MinRange[Input] = Idx;
2583 // Check if the access is smaller than the vector size and can we find
2584 // a reasonable extract index.
2585 int RangeUse[2] = { -1, -1 }; // 0 = Unused, 1 = Extract, -1 = Can not
2587 int StartIdx[2]; // StartIdx to extract from
2588 for (unsigned Input = 0; Input < 2; ++Input) {
2589 if (MinRange[Input] >= (int)SrcNumElts && MaxRange[Input] < 0) {
2590 RangeUse[Input] = 0; // Unused
2591 StartIdx[Input] = 0;
2595 // Find a good start index that is a multiple of the mask length. Then
2596 // see if the rest of the elements are in range.
2597 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts;
2598 if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts &&
2599 StartIdx[Input] + MaskNumElts <= SrcNumElts)
2600 RangeUse[Input] = 1; // Extract from a multiple of the mask length.
2603 if (RangeUse[0] == 0 && RangeUse[1] == 0) {
2604 setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used.
2607 if (RangeUse[0] >= 0 && RangeUse[1] >= 0) {
2608 // Extract appropriate subvector and generate a vector shuffle
2609 for (unsigned Input = 0; Input < 2; ++Input) {
2610 SDValue &Src = Input == 0 ? Src1 : Src2;
2611 if (RangeUse[Input] == 0)
2612 Src = DAG.getUNDEF(VT);
2614 SDLoc dl = getCurSDLoc();
2616 ISD::EXTRACT_SUBVECTOR, dl, VT, Src,
2617 DAG.getConstant(StartIdx[Input], dl, TLI.getVectorIdxTy()));
2621 // Calculate new mask.
2622 SmallVector<int, 8> MappedOps;
2623 for (unsigned i = 0; i != MaskNumElts; ++i) {
2626 if (Idx < (int)SrcNumElts)
2629 Idx -= SrcNumElts + StartIdx[1] - MaskNumElts;
2631 MappedOps.push_back(Idx);
2634 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
2640 // We can't use either concat vectors or extract subvectors so fall back to
2641 // replacing the shuffle with extract and build vector.
2642 // to insert and build vector.
2643 EVT EltVT = VT.getVectorElementType();
2644 EVT IdxVT = TLI.getVectorIdxTy();
2645 SDLoc dl = getCurSDLoc();
2646 SmallVector<SDValue,8> Ops;
2647 for (unsigned i = 0; i != MaskNumElts; ++i) {
2652 Res = DAG.getUNDEF(EltVT);
2654 SDValue &Src = Idx < (int)SrcNumElts ? Src1 : Src2;
2655 if (Idx >= (int)SrcNumElts) Idx -= SrcNumElts;
2657 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
2658 EltVT, Src, DAG.getConstant(Idx, dl, IdxVT));
2664 setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops));
2667 void SelectionDAGBuilder::visitInsertValue(const InsertValueInst &I) {
2668 const Value *Op0 = I.getOperand(0);
2669 const Value *Op1 = I.getOperand(1);
2670 Type *AggTy = I.getType();
2671 Type *ValTy = Op1->getType();
2672 bool IntoUndef = isa<UndefValue>(Op0);
2673 bool FromUndef = isa<UndefValue>(Op1);
2675 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
2677 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2678 SmallVector<EVT, 4> AggValueVTs;
2679 ComputeValueVTs(TLI, AggTy, AggValueVTs);
2680 SmallVector<EVT, 4> ValValueVTs;
2681 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2683 unsigned NumAggValues = AggValueVTs.size();
2684 unsigned NumValValues = ValValueVTs.size();
2685 SmallVector<SDValue, 4> Values(NumAggValues);
2687 // Ignore an insertvalue that produces an empty object
2688 if (!NumAggValues) {
2689 setValue(&I, DAG.getUNDEF(MVT(MVT::Other)));
2693 SDValue Agg = getValue(Op0);
2695 // Copy the beginning value(s) from the original aggregate.
2696 for (; i != LinearIndex; ++i)
2697 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2698 SDValue(Agg.getNode(), Agg.getResNo() + i);
2699 // Copy values from the inserted value(s).
2701 SDValue Val = getValue(Op1);
2702 for (; i != LinearIndex + NumValValues; ++i)
2703 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2704 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex);
2706 // Copy remaining value(s) from the original aggregate.
2707 for (; i != NumAggValues; ++i)
2708 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2709 SDValue(Agg.getNode(), Agg.getResNo() + i);
2711 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
2712 DAG.getVTList(AggValueVTs), Values));
2715 void SelectionDAGBuilder::visitExtractValue(const ExtractValueInst &I) {
2716 const Value *Op0 = I.getOperand(0);
2717 Type *AggTy = Op0->getType();
2718 Type *ValTy = I.getType();
2719 bool OutOfUndef = isa<UndefValue>(Op0);
2721 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
2723 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2724 SmallVector<EVT, 4> ValValueVTs;
2725 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2727 unsigned NumValValues = ValValueVTs.size();
2729 // Ignore a extractvalue that produces an empty object
2730 if (!NumValValues) {
2731 setValue(&I, DAG.getUNDEF(MVT(MVT::Other)));
2735 SmallVector<SDValue, 4> Values(NumValValues);
2737 SDValue Agg = getValue(Op0);
2738 // Copy out the selected value(s).
2739 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i)
2740 Values[i - LinearIndex] =
2742 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) :
2743 SDValue(Agg.getNode(), Agg.getResNo() + i);
2745 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
2746 DAG.getVTList(ValValueVTs), Values));
2749 void SelectionDAGBuilder::visitGetElementPtr(const User &I) {
2750 Value *Op0 = I.getOperand(0);
2751 // Note that the pointer operand may be a vector of pointers. Take the scalar
2752 // element which holds a pointer.
2753 Type *Ty = Op0->getType()->getScalarType();
2754 unsigned AS = Ty->getPointerAddressSpace();
2755 SDValue N = getValue(Op0);
2756 SDLoc dl = getCurSDLoc();
2758 for (GetElementPtrInst::const_op_iterator OI = I.op_begin()+1, E = I.op_end();
2760 const Value *Idx = *OI;
2761 if (StructType *StTy = dyn_cast<StructType>(Ty)) {
2762 unsigned Field = cast<Constant>(Idx)->getUniqueInteger().getZExtValue();
2765 uint64_t Offset = DL->getStructLayout(StTy)->getElementOffset(Field);
2766 N = DAG.getNode(ISD::ADD, dl, N.getValueType(), N,
2767 DAG.getConstant(Offset, dl, N.getValueType()));
2770 Ty = StTy->getElementType(Field);
2772 Ty = cast<SequentialType>(Ty)->getElementType();
2773 MVT PtrTy = DAG.getTargetLoweringInfo().getPointerTy(AS);
2774 unsigned PtrSize = PtrTy.getSizeInBits();
2775 APInt ElementSize(PtrSize, DL->getTypeAllocSize(Ty));
2777 // If this is a constant subscript, handle it quickly.
2778 if (const auto *CI = dyn_cast<ConstantInt>(Idx)) {
2781 APInt Offs = ElementSize * CI->getValue().sextOrTrunc(PtrSize);
2782 SDValue OffsVal = DAG.getConstant(Offs, dl, PtrTy);
2783 N = DAG.getNode(ISD::ADD, dl, N.getValueType(), N, OffsVal);
2787 // N = N + Idx * ElementSize;
2788 SDValue IdxN = getValue(Idx);
2790 // If the index is smaller or larger than intptr_t, truncate or extend
2792 IdxN = DAG.getSExtOrTrunc(IdxN, dl, N.getValueType());
2794 // If this is a multiply by a power of two, turn it into a shl
2795 // immediately. This is a very common case.
2796 if (ElementSize != 1) {
2797 if (ElementSize.isPowerOf2()) {
2798 unsigned Amt = ElementSize.logBase2();
2799 IdxN = DAG.getNode(ISD::SHL, dl,
2800 N.getValueType(), IdxN,
2801 DAG.getConstant(Amt, dl, IdxN.getValueType()));
2803 SDValue Scale = DAG.getConstant(ElementSize, dl, IdxN.getValueType());
2804 IdxN = DAG.getNode(ISD::MUL, dl,
2805 N.getValueType(), IdxN, Scale);
2809 N = DAG.getNode(ISD::ADD, dl,
2810 N.getValueType(), N, IdxN);
2817 void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) {
2818 // If this is a fixed sized alloca in the entry block of the function,
2819 // allocate it statically on the stack.
2820 if (FuncInfo.StaticAllocaMap.count(&I))
2821 return; // getValue will auto-populate this.
2823 SDLoc dl = getCurSDLoc();
2824 Type *Ty = I.getAllocatedType();
2825 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2826 uint64_t TySize = TLI.getDataLayout()->getTypeAllocSize(Ty);
2828 std::max((unsigned)TLI.getDataLayout()->getPrefTypeAlignment(Ty),
2831 SDValue AllocSize = getValue(I.getArraySize());
2833 EVT IntPtr = TLI.getPointerTy();
2834 if (AllocSize.getValueType() != IntPtr)
2835 AllocSize = DAG.getZExtOrTrunc(AllocSize, dl, IntPtr);
2837 AllocSize = DAG.getNode(ISD::MUL, dl, IntPtr,
2839 DAG.getConstant(TySize, dl, IntPtr));
2841 // Handle alignment. If the requested alignment is less than or equal to
2842 // the stack alignment, ignore it. If the size is greater than or equal to
2843 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
2844 unsigned StackAlign =
2845 DAG.getSubtarget().getFrameLowering()->getStackAlignment();
2846 if (Align <= StackAlign)
2849 // Round the size of the allocation up to the stack alignment size
2850 // by add SA-1 to the size.
2851 AllocSize = DAG.getNode(ISD::ADD, dl,
2852 AllocSize.getValueType(), AllocSize,
2853 DAG.getIntPtrConstant(StackAlign - 1, dl));
2855 // Mask out the low bits for alignment purposes.
2856 AllocSize = DAG.getNode(ISD::AND, dl,
2857 AllocSize.getValueType(), AllocSize,
2858 DAG.getIntPtrConstant(~(uint64_t)(StackAlign - 1),
2861 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align, dl) };
2862 SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other);
2863 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, dl, VTs, Ops);
2865 DAG.setRoot(DSA.getValue(1));
2867 assert(FuncInfo.MF->getFrameInfo()->hasVarSizedObjects());
2870 void SelectionDAGBuilder::visitLoad(const LoadInst &I) {
2872 return visitAtomicLoad(I);
2874 const Value *SV = I.getOperand(0);
2875 SDValue Ptr = getValue(SV);
2877 Type *Ty = I.getType();
2879 bool isVolatile = I.isVolatile();
2880 bool isNonTemporal = I.getMetadata(LLVMContext::MD_nontemporal) != nullptr;
2882 // The IR notion of invariant_load only guarantees that all *non-faulting*
2883 // invariant loads result in the same value. The MI notion of invariant load
2884 // guarantees that the load can be legally moved to any location within its
2885 // containing function. The MI notion of invariant_load is stronger than the
2886 // IR notion of invariant_load -- an MI invariant_load is an IR invariant_load
2887 // with a guarantee that the location being loaded from is dereferenceable
2888 // throughout the function's lifetime.
2890 bool isInvariant = I.getMetadata(LLVMContext::MD_invariant_load) != nullptr &&
2891 isDereferenceablePointer(SV, *DAG.getTarget().getDataLayout());
2892 unsigned Alignment = I.getAlignment();
2895 I.getAAMetadata(AAInfo);
2896 const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range);
2898 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2899 SmallVector<EVT, 4> ValueVTs;
2900 SmallVector<uint64_t, 4> Offsets;
2901 ComputeValueVTs(TLI, Ty, ValueVTs, &Offsets);
2902 unsigned NumValues = ValueVTs.size();
2907 bool ConstantMemory = false;
2908 if (isVolatile || NumValues > MaxParallelChains)
2909 // Serialize volatile loads with other side effects.
2911 else if (AA->pointsToConstantMemory(
2912 MemoryLocation(SV, AA->getTypeStoreSize(Ty), AAInfo))) {
2913 // Do not serialize (non-volatile) loads of constant memory with anything.
2914 Root = DAG.getEntryNode();
2915 ConstantMemory = true;
2917 // Do not serialize non-volatile loads against each other.
2918 Root = DAG.getRoot();
2921 SDLoc dl = getCurSDLoc();
2924 Root = TLI.prepareVolatileOrAtomicLoad(Root, dl, DAG);
2926 SmallVector<SDValue, 4> Values(NumValues);
2927 SmallVector<SDValue, 4> Chains(std::min(MaxParallelChains, NumValues));
2928 EVT PtrVT = Ptr.getValueType();
2929 unsigned ChainI = 0;
2930 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
2931 // Serializing loads here may result in excessive register pressure, and
2932 // TokenFactor places arbitrary choke points on the scheduler. SD scheduling
2933 // could recover a bit by hoisting nodes upward in the chain by recognizing
2934 // they are side-effect free or do not alias. The optimizer should really
2935 // avoid this case by converting large object/array copies to llvm.memcpy
2936 // (MaxParallelChains should always remain as failsafe).
2937 if (ChainI == MaxParallelChains) {
2938 assert(PendingLoads.empty() && "PendingLoads must be serialized first");
2939 SDValue Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2940 makeArrayRef(Chains.data(), ChainI));
2944 SDValue A = DAG.getNode(ISD::ADD, dl,
2946 DAG.getConstant(Offsets[i], dl, PtrVT));
2947 SDValue L = DAG.getLoad(ValueVTs[i], dl, Root,
2948 A, MachinePointerInfo(SV, Offsets[i]), isVolatile,
2949 isNonTemporal, isInvariant, Alignment, AAInfo,
2953 Chains[ChainI] = L.getValue(1);
2956 if (!ConstantMemory) {
2957 SDValue Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2958 makeArrayRef(Chains.data(), ChainI));
2962 PendingLoads.push_back(Chain);
2965 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, dl,
2966 DAG.getVTList(ValueVTs), Values));
2969 void SelectionDAGBuilder::visitStore(const StoreInst &I) {
2971 return visitAtomicStore(I);
2973 const Value *SrcV = I.getOperand(0);
2974 const Value *PtrV = I.getOperand(1);
2976 SmallVector<EVT, 4> ValueVTs;
2977 SmallVector<uint64_t, 4> Offsets;
2978 ComputeValueVTs(DAG.getTargetLoweringInfo(), SrcV->getType(),
2979 ValueVTs, &Offsets);
2980 unsigned NumValues = ValueVTs.size();
2984 // Get the lowered operands. Note that we do this after
2985 // checking if NumResults is zero, because with zero results
2986 // the operands won't have values in the map.
2987 SDValue Src = getValue(SrcV);
2988 SDValue Ptr = getValue(PtrV);
2990 SDValue Root = getRoot();
2991 SmallVector<SDValue, 4> Chains(std::min(MaxParallelChains, NumValues));
2992 EVT PtrVT = Ptr.getValueType();
2993 bool isVolatile = I.isVolatile();
2994 bool isNonTemporal = I.getMetadata(LLVMContext::MD_nontemporal) != nullptr;
2995 unsigned Alignment = I.getAlignment();
2996 SDLoc dl = getCurSDLoc();
2999 I.getAAMetadata(AAInfo);
3001 unsigned ChainI = 0;
3002 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3003 // See visitLoad comments.
3004 if (ChainI == MaxParallelChains) {
3005 SDValue Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3006 makeArrayRef(Chains.data(), ChainI));
3010 SDValue Add = DAG.getNode(ISD::ADD, dl, PtrVT, Ptr,
3011 DAG.getConstant(Offsets[i], dl, PtrVT));
3012 SDValue St = DAG.getStore(Root, dl,
3013 SDValue(Src.getNode(), Src.getResNo() + i),
3014 Add, MachinePointerInfo(PtrV, Offsets[i]),
3015 isVolatile, isNonTemporal, Alignment, AAInfo);
3016 Chains[ChainI] = St;
3019 SDValue StoreNode = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3020 makeArrayRef(Chains.data(), ChainI));
3021 DAG.setRoot(StoreNode);
3024 void SelectionDAGBuilder::visitMaskedStore(const CallInst &I) {
3025 SDLoc sdl = getCurSDLoc();
3027 // llvm.masked.store.*(Src0, Ptr, alignemt, Mask)
3028 Value *PtrOperand = I.getArgOperand(1);
3029 SDValue Ptr = getValue(PtrOperand);
3030 SDValue Src0 = getValue(I.getArgOperand(0));
3031 SDValue Mask = getValue(I.getArgOperand(3));
3032 EVT VT = Src0.getValueType();
3033 unsigned Alignment = (cast<ConstantInt>(I.getArgOperand(2)))->getZExtValue();
3035 Alignment = DAG.getEVTAlignment(VT);
3038 I.getAAMetadata(AAInfo);
3040 MachineMemOperand *MMO =
3041 DAG.getMachineFunction().
3042 getMachineMemOperand(MachinePointerInfo(PtrOperand),
3043 MachineMemOperand::MOStore, VT.getStoreSize(),
3045 SDValue StoreNode = DAG.getMaskedStore(getRoot(), sdl, Src0, Ptr, Mask, VT,
3047 DAG.setRoot(StoreNode);
3048 setValue(&I, StoreNode);
3051 // Gather/scatter receive a vector of pointers.
3052 // This vector of pointers may be represented as a base pointer + vector of
3053 // indices, it depends on GEP and instruction preceeding GEP
3054 // that calculates indices
3055 static bool getUniformBase(Value *& Ptr, SDValue& Base, SDValue& Index,
3056 SelectionDAGBuilder* SDB) {
3058 assert (Ptr->getType()->isVectorTy() && "Uexpected pointer type");
3059 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
3060 if (!Gep || Gep->getNumOperands() > 2)
3062 ShuffleVectorInst *ShuffleInst =
3063 dyn_cast<ShuffleVectorInst>(Gep->getPointerOperand());
3064 if (!ShuffleInst || !ShuffleInst->getMask()->isNullValue() ||
3065 cast<Instruction>(ShuffleInst->getOperand(0))->getOpcode() !=
3066 Instruction::InsertElement)
3069 Ptr = cast<InsertElementInst>(ShuffleInst->getOperand(0))->getOperand(1);
3071 SelectionDAG& DAG = SDB->DAG;
3072 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3073 // Check is the Ptr is inside current basic block
3074 // If not, look for the shuffle instruction
3075 if (SDB->findValue(Ptr))
3076 Base = SDB->getValue(Ptr);
3077 else if (SDB->findValue(ShuffleInst)) {
3078 SDValue ShuffleNode = SDB->getValue(ShuffleInst);
3079 SDLoc sdl = ShuffleNode;
3080 Base = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, sdl,
3081 ShuffleNode.getValueType().getScalarType(), ShuffleNode,
3082 DAG.getConstant(0, sdl, TLI.getVectorIdxTy()));
3083 SDB->setValue(Ptr, Base);
3088 Value *IndexVal = Gep->getOperand(1);
3089 if (SDB->findValue(IndexVal)) {
3090 Index = SDB->getValue(IndexVal);
3092 if (SExtInst* Sext = dyn_cast<SExtInst>(IndexVal)) {
3093 IndexVal = Sext->getOperand(0);
3094 if (SDB->findValue(IndexVal))
3095 Index = SDB->getValue(IndexVal);
3102 void SelectionDAGBuilder::visitMaskedScatter(const CallInst &I) {
3103 SDLoc sdl = getCurSDLoc();
3105 // llvm.masked.scatter.*(Src0, Ptrs, alignemt, Mask)
3106 Value *Ptr = I.getArgOperand(1);
3107 SDValue Src0 = getValue(I.getArgOperand(0));
3108 SDValue Mask = getValue(I.getArgOperand(3));
3109 EVT VT = Src0.getValueType();
3110 unsigned Alignment = (cast<ConstantInt>(I.getArgOperand(2)))->getZExtValue();
3112 Alignment = DAG.getEVTAlignment(VT);
3113 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3116 I.getAAMetadata(AAInfo);
3120 Value *BasePtr = Ptr;
3121 bool UniformBase = getUniformBase(BasePtr, Base, Index, this);
3123 Value *MemOpBasePtr = UniformBase ? BasePtr : nullptr;
3124 MachineMemOperand *MMO = DAG.getMachineFunction().
3125 getMachineMemOperand(MachinePointerInfo(MemOpBasePtr),
3126 MachineMemOperand::MOStore, VT.getStoreSize(),
3129 Base = DAG.getTargetConstant(0, sdl, TLI.getPointerTy());
3130 Index = getValue(Ptr);
3132 SDValue Ops[] = { getRoot(), Src0, Mask, Base, Index };
3133 SDValue Scatter = DAG.getMaskedScatter(DAG.getVTList(MVT::Other), VT, sdl,
3135 DAG.setRoot(Scatter);
3136 setValue(&I, Scatter);
3139 void SelectionDAGBuilder::visitMaskedLoad(const CallInst &I) {
3140 SDLoc sdl = getCurSDLoc();
3142 // @llvm.masked.load.*(Ptr, alignment, Mask, Src0)
3143 Value *PtrOperand = I.getArgOperand(0);
3144 SDValue Ptr = getValue(PtrOperand);
3145 SDValue Src0 = getValue(I.getArgOperand(3));
3146 SDValue Mask = getValue(I.getArgOperand(2));
3148 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3149 EVT VT = TLI.getValueType(I.getType());
3150 unsigned Alignment = (cast<ConstantInt>(I.getArgOperand(1)))->getZExtValue();
3152 Alignment = DAG.getEVTAlignment(VT);
3155 I.getAAMetadata(AAInfo);
3156 const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range);
3158 SDValue InChain = DAG.getRoot();
3159 if (AA->pointsToConstantMemory(MemoryLocation(
3160 PtrOperand, AA->getTypeStoreSize(I.getType()), AAInfo))) {
3161 // Do not serialize (non-volatile) loads of constant memory with anything.
3162 InChain = DAG.getEntryNode();
3165 MachineMemOperand *MMO =
3166 DAG.getMachineFunction().
3167 getMachineMemOperand(MachinePointerInfo(PtrOperand),
3168 MachineMemOperand::MOLoad, VT.getStoreSize(),
3169 Alignment, AAInfo, Ranges);
3171 SDValue Load = DAG.getMaskedLoad(VT, sdl, InChain, Ptr, Mask, Src0, VT, MMO,
3173 SDValue OutChain = Load.getValue(1);
3174 DAG.setRoot(OutChain);
3178 void SelectionDAGBuilder::visitMaskedGather(const CallInst &I) {
3179 SDLoc sdl = getCurSDLoc();
3181 // @llvm.masked.gather.*(Ptrs, alignment, Mask, Src0)
3182 Value *Ptr = I.getArgOperand(0);
3183 SDValue Src0 = getValue(I.getArgOperand(3));
3184 SDValue Mask = getValue(I.getArgOperand(2));
3186 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3187 EVT VT = TLI.getValueType(I.getType());
3188 unsigned Alignment = (cast<ConstantInt>(I.getArgOperand(1)))->getZExtValue();
3190 Alignment = DAG.getEVTAlignment(VT);
3193 I.getAAMetadata(AAInfo);
3194 const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range);
3196 SDValue Root = DAG.getRoot();
3199 Value *BasePtr = Ptr;
3200 bool UniformBase = getUniformBase(BasePtr, Base, Index, this);
3201 bool ConstantMemory = false;
3203 AA->pointsToConstantMemory(
3204 MemoryLocation(BasePtr, AA->getTypeStoreSize(I.getType()), AAInfo))) {
3205 // Do not serialize (non-volatile) loads of constant memory with anything.
3206 Root = DAG.getEntryNode();
3207 ConstantMemory = true;
3210 MachineMemOperand *MMO =
3211 DAG.getMachineFunction().
3212 getMachineMemOperand(MachinePointerInfo(UniformBase ? BasePtr : nullptr),
3213 MachineMemOperand::MOLoad, VT.getStoreSize(),
3214 Alignment, AAInfo, Ranges);
3217 Base = DAG.getTargetConstant(0, sdl, TLI.getPointerTy());
3218 Index = getValue(Ptr);
3220 SDValue Ops[] = { Root, Src0, Mask, Base, Index };
3221 SDValue Gather = DAG.getMaskedGather(DAG.getVTList(VT, MVT::Other), VT, sdl,
3224 SDValue OutChain = Gather.getValue(1);
3225 if (!ConstantMemory)
3226 PendingLoads.push_back(OutChain);
3227 setValue(&I, Gather);
3230 void SelectionDAGBuilder::visitAtomicCmpXchg(const AtomicCmpXchgInst &I) {
3231 SDLoc dl = getCurSDLoc();
3232 AtomicOrdering SuccessOrder = I.getSuccessOrdering();
3233 AtomicOrdering FailureOrder = I.getFailureOrdering();
3234 SynchronizationScope Scope = I.getSynchScope();
3236 SDValue InChain = getRoot();
3238 MVT MemVT = getValue(I.getCompareOperand()).getSimpleValueType();
3239 SDVTList VTs = DAG.getVTList(MemVT, MVT::i1, MVT::Other);
3240 SDValue L = DAG.getAtomicCmpSwap(
3241 ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, dl, MemVT, VTs, InChain,
3242 getValue(I.getPointerOperand()), getValue(I.getCompareOperand()),
3243 getValue(I.getNewValOperand()), MachinePointerInfo(I.getPointerOperand()),
3244 /*Alignment=*/ 0, SuccessOrder, FailureOrder, Scope);
3246 SDValue OutChain = L.getValue(2);
3249 DAG.setRoot(OutChain);
3252 void SelectionDAGBuilder::visitAtomicRMW(const AtomicRMWInst &I) {
3253 SDLoc dl = getCurSDLoc();
3255 switch (I.getOperation()) {
3256 default: llvm_unreachable("Unknown atomicrmw operation");
3257 case AtomicRMWInst::Xchg: NT = ISD::ATOMIC_SWAP; break;
3258 case AtomicRMWInst::Add: NT = ISD::ATOMIC_LOAD_ADD; break;
3259 case AtomicRMWInst::Sub: NT = ISD::ATOMIC_LOAD_SUB; break;
3260 case AtomicRMWInst::And: NT = ISD::ATOMIC_LOAD_AND; break;
3261 case AtomicRMWInst::Nand: NT = ISD::ATOMIC_LOAD_NAND; break;
3262 case AtomicRMWInst::Or: NT = ISD::ATOMIC_LOAD_OR; break;
3263 case AtomicRMWInst::Xor: NT = ISD::ATOMIC_LOAD_XOR; break;
3264 case AtomicRMWInst::Max: NT = ISD::ATOMIC_LOAD_MAX; break;
3265 case AtomicRMWInst::Min: NT = ISD::ATOMIC_LOAD_MIN; break;
3266 case AtomicRMWInst::UMax: NT = ISD::ATOMIC_LOAD_UMAX; break;
3267 case AtomicRMWInst::UMin: NT = ISD::ATOMIC_LOAD_UMIN; break;
3269 AtomicOrdering Order = I.getOrdering();
3270 SynchronizationScope Scope = I.getSynchScope();
3272 SDValue InChain = getRoot();
3275 DAG.getAtomic(NT, dl,
3276 getValue(I.getValOperand()).getSimpleValueType(),
3278 getValue(I.getPointerOperand()),
3279 getValue(I.getValOperand()),
3280 I.getPointerOperand(),
3281 /* Alignment=*/ 0, Order, Scope);
3283 SDValue OutChain = L.getValue(1);
3286 DAG.setRoot(OutChain);
3289 void SelectionDAGBuilder::visitFence(const FenceInst &I) {
3290 SDLoc dl = getCurSDLoc();
3291 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3294 Ops[1] = DAG.getConstant(I.getOrdering(), dl, TLI.getPointerTy());
3295 Ops[2] = DAG.getConstant(I.getSynchScope(), dl, TLI.getPointerTy());
3296 DAG.setRoot(DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops));
3299 void SelectionDAGBuilder::visitAtomicLoad(const LoadInst &I) {
3300 SDLoc dl = getCurSDLoc();
3301 AtomicOrdering Order = I.getOrdering();
3302 SynchronizationScope Scope = I.getSynchScope();
3304 SDValue InChain = getRoot();
3306 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3307 EVT VT = TLI.getValueType(I.getType());
3309 if (I.getAlignment() < VT.getSizeInBits() / 8)
3310 report_fatal_error("Cannot generate unaligned atomic load");
3312 MachineMemOperand *MMO =
3313 DAG.getMachineFunction().
3314 getMachineMemOperand(MachinePointerInfo(I.getPointerOperand()),
3315 MachineMemOperand::MOVolatile |
3316 MachineMemOperand::MOLoad,
3318 I.getAlignment() ? I.getAlignment() :
3319 DAG.getEVTAlignment(VT));
3321 InChain = TLI.prepareVolatileOrAtomicLoad(InChain, dl, DAG);
3323 DAG.getAtomic(ISD::ATOMIC_LOAD, dl, VT, VT, InChain,
3324 getValue(I.getPointerOperand()), MMO,
3327 SDValue OutChain = L.getValue(1);
3330 DAG.setRoot(OutChain);
3333 void SelectionDAGBuilder::visitAtomicStore(const StoreInst &I) {
3334 SDLoc dl = getCurSDLoc();
3336 AtomicOrdering Order = I.getOrdering();
3337 SynchronizationScope Scope = I.getSynchScope();
3339 SDValue InChain = getRoot();
3341 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3342 EVT VT = TLI.getValueType(I.getValueOperand()->getType());
3344 if (I.getAlignment() < VT.getSizeInBits() / 8)
3345 report_fatal_error("Cannot generate unaligned atomic store");
3348 DAG.getAtomic(ISD::ATOMIC_STORE, dl, VT,
3350 getValue(I.getPointerOperand()),
3351 getValue(I.getValueOperand()),
3352 I.getPointerOperand(), I.getAlignment(),
3355 DAG.setRoot(OutChain);
3358 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
3360 void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I,
3361 unsigned Intrinsic) {
3362 bool HasChain = !I.doesNotAccessMemory();
3363 bool OnlyLoad = HasChain && I.onlyReadsMemory();
3365 // Build the operand list.
3366 SmallVector<SDValue, 8> Ops;
3367 if (HasChain) { // If this intrinsic has side-effects, chainify it.
3369 // We don't need to serialize loads against other loads.
3370 Ops.push_back(DAG.getRoot());
3372 Ops.push_back(getRoot());
3376 // Info is set by getTgtMemInstrinsic
3377 TargetLowering::IntrinsicInfo Info;
3378 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3379 bool IsTgtIntrinsic = TLI.getTgtMemIntrinsic(Info, I, Intrinsic);
3381 // Add the intrinsic ID as an integer operand if it's not a target intrinsic.
3382 if (!IsTgtIntrinsic || Info.opc == ISD::INTRINSIC_VOID ||
3383 Info.opc == ISD::INTRINSIC_W_CHAIN)
3384 Ops.push_back(DAG.getTargetConstant(Intrinsic, getCurSDLoc(),
3385 TLI.getPointerTy()));
3387 // Add all operands of the call to the operand list.
3388 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) {
3389 SDValue Op = getValue(I.getArgOperand(i));
3393 SmallVector<EVT, 4> ValueVTs;
3394 ComputeValueVTs(TLI, I.getType(), ValueVTs);
3397 ValueVTs.push_back(MVT::Other);
3399 SDVTList VTs = DAG.getVTList(ValueVTs);
3403 if (IsTgtIntrinsic) {
3404 // This is target intrinsic that touches memory
3405 Result = DAG.getMemIntrinsicNode(Info.opc, getCurSDLoc(),
3406 VTs, Ops, Info.memVT,
3407 MachinePointerInfo(Info.ptrVal, Info.offset),
3408 Info.align, Info.vol,
3409 Info.readMem, Info.writeMem, Info.size);
3410 } else if (!HasChain) {
3411 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurSDLoc(), VTs, Ops);
3412 } else if (!I.getType()->isVoidTy()) {
3413 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurSDLoc(), VTs, Ops);
3415 Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurSDLoc(), VTs, Ops);
3419 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1);
3421 PendingLoads.push_back(Chain);
3426 if (!I.getType()->isVoidTy()) {
3427 if (VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
3428 EVT VT = TLI.getValueType(PTy);
3429 Result = DAG.getNode(ISD::BITCAST, getCurSDLoc(), VT, Result);
3432 setValue(&I, Result);
3436 /// GetSignificand - Get the significand and build it into a floating-point
3437 /// number with exponent of 1:
3439 /// Op = (Op & 0x007fffff) | 0x3f800000;
3441 /// where Op is the hexadecimal representation of floating point value.
3443 GetSignificand(SelectionDAG &DAG, SDValue Op, SDLoc dl) {
3444 SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3445 DAG.getConstant(0x007fffff, dl, MVT::i32));
3446 SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1,
3447 DAG.getConstant(0x3f800000, dl, MVT::i32));
3448 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, t2);
3451 /// GetExponent - Get the exponent:
3453 /// (float)(int)(((Op & 0x7f800000) >> 23) - 127);
3455 /// where Op is the hexadecimal representation of floating point value.
3457 GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI,
3459 SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3460 DAG.getConstant(0x7f800000, dl, MVT::i32));
3461 SDValue t1 = DAG.getNode(ISD::SRL, dl, MVT::i32, t0,
3462 DAG.getConstant(23, dl, TLI.getPointerTy()));
3463 SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1,
3464 DAG.getConstant(127, dl, MVT::i32));
3465 return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2);
3468 /// getF32Constant - Get 32-bit floating point constant.
3470 getF32Constant(SelectionDAG &DAG, unsigned Flt, SDLoc dl) {
3471 return DAG.getConstantFP(APFloat(APFloat::IEEEsingle, APInt(32, Flt)), dl,
3475 static SDValue getLimitedPrecisionExp2(SDValue t0, SDLoc dl,
3476 SelectionDAG &DAG) {
3477 // IntegerPartOfX = ((int32_t)(t0);
3478 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3480 // FractionalPartOfX = t0 - (float)IntegerPartOfX;
3481 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3482 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3484 // IntegerPartOfX <<= 23;
3485 IntegerPartOfX = DAG.getNode(
3486 ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3487 DAG.getConstant(23, dl, DAG.getTargetLoweringInfo().getPointerTy()));
3489 SDValue TwoToFractionalPartOfX;
3490 if (LimitFloatPrecision <= 6) {
3491 // For floating-point precision of 6:
3493 // TwoToFractionalPartOfX =
3495 // (0.735607626f + 0.252464424f * x) * x;
3497 // error 0.0144103317, which is 6 bits
3498 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3499 getF32Constant(DAG, 0x3e814304, dl));
3500 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3501 getF32Constant(DAG, 0x3f3c50c8, dl));
3502 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3503 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3504 getF32Constant(DAG, 0x3f7f5e7e, dl));
3505 } else if (LimitFloatPrecision <= 12) {
3506 // For floating-point precision of 12:
3508 // TwoToFractionalPartOfX =
3511 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3513 // error 0.000107046256, which is 13 to 14 bits
3514 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3515 getF32Constant(DAG, 0x3da235e3, dl));
3516 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3517 getF32Constant(DAG, 0x3e65b8f3, dl));
3518 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3519 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3520 getF32Constant(DAG, 0x3f324b07, dl));
3521 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3522 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3523 getF32Constant(DAG, 0x3f7ff8fd, dl));
3524 } else { // LimitFloatPrecision <= 18
3525 // For floating-point precision of 18:
3527 // TwoToFractionalPartOfX =
3531 // (0.554906021e-1f +
3532 // (0.961591928e-2f +
3533 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3534 // error 2.47208000*10^(-7), which is better than 18 bits
3535 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3536 getF32Constant(DAG, 0x3924b03e, dl));
3537 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3538 getF32Constant(DAG, 0x3ab24b87, dl));
3539 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3540 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3541 getF32Constant(DAG, 0x3c1d8c17, dl));
3542 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3543 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3544 getF32Constant(DAG, 0x3d634a1d, dl));
3545 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3546 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3547 getF32Constant(DAG, 0x3e75fe14, dl));
3548 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3549 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3550 getF32Constant(DAG, 0x3f317234, dl));
3551 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3552 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3553 getF32Constant(DAG, 0x3f800000, dl));
3556 // Add the exponent into the result in integer domain.
3557 SDValue t13 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, TwoToFractionalPartOfX);
3558 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3559 DAG.getNode(ISD::ADD, dl, MVT::i32, t13, IntegerPartOfX));
3562 /// expandExp - Lower an exp intrinsic. Handles the special sequences for
3563 /// limited-precision mode.
3564 static SDValue expandExp(SDLoc dl, SDValue Op, SelectionDAG &DAG,
3565 const TargetLowering &TLI) {
3566 if (Op.getValueType() == MVT::f32 &&
3567 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3569 // Put the exponent in the right bit position for later addition to the
3572 // #define LOG2OFe 1.4426950f
3573 // t0 = Op * LOG2OFe
3574 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3575 getF32Constant(DAG, 0x3fb8aa3b, dl));
3576 return getLimitedPrecisionExp2(t0, dl, DAG);
3579 // No special expansion.
3580 return DAG.getNode(ISD::FEXP, dl, Op.getValueType(), Op);
3583 /// expandLog - Lower a log intrinsic. Handles the special sequences for
3584 /// limited-precision mode.
3585 static SDValue expandLog(SDLoc dl, SDValue Op, SelectionDAG &DAG,
3586 const TargetLowering &TLI) {
3587 if (Op.getValueType() == MVT::f32 &&
3588 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3589 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3591 // Scale the exponent by log(2) [0.69314718f].
3592 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3593 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3594 getF32Constant(DAG, 0x3f317218, dl));
3596 // Get the significand and build it into a floating-point number with
3598 SDValue X = GetSignificand(DAG, Op1, dl);
3600 SDValue LogOfMantissa;
3601 if (LimitFloatPrecision <= 6) {
3602 // For floating-point precision of 6:
3606 // (1.4034025f - 0.23903021f * x) * x;
3608 // error 0.0034276066, which is better than 8 bits
3609 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3610 getF32Constant(DAG, 0xbe74c456, dl));
3611 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3612 getF32Constant(DAG, 0x3fb3a2b1, dl));
3613 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3614 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3615 getF32Constant(DAG, 0x3f949a29, dl));
3616 } else if (LimitFloatPrecision <= 12) {
3617 // For floating-point precision of 12:
3623 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x;
3625 // error 0.000061011436, which is 14 bits
3626 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3627 getF32Constant(DAG, 0xbd67b6d6, dl));
3628 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3629 getF32Constant(DAG, 0x3ee4f4b8, dl));
3630 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3631 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3632 getF32Constant(DAG, 0x3fbc278b, dl));
3633 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3634 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3635 getF32Constant(DAG, 0x40348e95, dl));
3636 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3637 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3638 getF32Constant(DAG, 0x3fdef31a, dl));
3639 } else { // LimitFloatPrecision <= 18
3640 // For floating-point precision of 18:
3648 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x;
3650 // error 0.0000023660568, which is better than 18 bits
3651 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3652 getF32Constant(DAG, 0xbc91e5ac, dl));
3653 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3654 getF32Constant(DAG, 0x3e4350aa, dl));
3655 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3656 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3657 getF32Constant(DAG, 0x3f60d3e3, dl));
3658 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3659 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3660 getF32Constant(DAG, 0x4011cdf0, dl));
3661 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3662 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3663 getF32Constant(DAG, 0x406cfd1c, dl));
3664 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3665 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3666 getF32Constant(DAG, 0x408797cb, dl));
3667 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3668 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3669 getF32Constant(DAG, 0x4006dcab, dl));
3672 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, LogOfMantissa);
3675 // No special expansion.
3676 return DAG.getNode(ISD::FLOG, dl, Op.getValueType(), Op);
3679 /// expandLog2 - Lower a log2 intrinsic. Handles the special sequences for
3680 /// limited-precision mode.
3681 static SDValue expandLog2(SDLoc dl, SDValue Op, SelectionDAG &DAG,
3682 const TargetLowering &TLI) {
3683 if (Op.getValueType() == MVT::f32 &&
3684 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3685 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3687 // Get the exponent.
3688 SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl);
3690 // Get the significand and build it into a floating-point number with
3692 SDValue X = GetSignificand(DAG, Op1, dl);
3694 // Different possible minimax approximations of significand in
3695 // floating-point for various degrees of accuracy over [1,2].
3696 SDValue Log2ofMantissa;
3697 if (LimitFloatPrecision <= 6) {
3698 // For floating-point precision of 6:
3700 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x;
3702 // error 0.0049451742, which is more than 7 bits
3703 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3704 getF32Constant(DAG, 0xbeb08fe0, dl));
3705 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3706 getF32Constant(DAG, 0x40019463, dl));
3707 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3708 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3709 getF32Constant(DAG, 0x3fd6633d, dl));
3710 } else if (LimitFloatPrecision <= 12) {
3711 // For floating-point precision of 12:
3717 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x;
3719 // error 0.0000876136000, which is better than 13 bits
3720 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3721 getF32Constant(DAG, 0xbda7262e, dl));
3722 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3723 getF32Constant(DAG, 0x3f25280b, dl));
3724 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3725 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3726 getF32Constant(DAG, 0x4007b923, dl));
3727 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3728 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3729 getF32Constant(DAG, 0x40823e2f, dl));
3730 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3731 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3732 getF32Constant(DAG, 0x4020d29c, dl));
3733 } else { // LimitFloatPrecision <= 18
3734 // For floating-point precision of 18:
3743 // 0.25691327e-1f * x) * x) * x) * x) * x) * x;
3745 // error 0.0000018516, which is better than 18 bits
3746 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3747 getF32Constant(DAG, 0xbcd2769e, dl));
3748 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3749 getF32Constant(DAG, 0x3e8ce0b9, dl));
3750 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3751 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3752 getF32Constant(DAG, 0x3fa22ae7, dl));
3753 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3754 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3755 getF32Constant(DAG, 0x40525723, dl));
3756 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3757 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3758 getF32Constant(DAG, 0x40aaf200, dl));
3759 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3760 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3761 getF32Constant(DAG, 0x40c39dad, dl));
3762 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3763 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3764 getF32Constant(DAG, 0x4042902c, dl));
3767 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, Log2ofMantissa);
3770 // No special expansion.
3771 return DAG.getNode(ISD::FLOG2, dl, Op.getValueType(), Op);
3774 /// expandLog10 - Lower a log10 intrinsic. Handles the special sequences for
3775 /// limited-precision mode.
3776 static SDValue expandLog10(SDLoc dl, SDValue Op, SelectionDAG &DAG,
3777 const TargetLowering &TLI) {
3778 if (Op.getValueType() == MVT::f32 &&
3779 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3780 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3782 // Scale the exponent by log10(2) [0.30102999f].
3783 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3784 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3785 getF32Constant(DAG, 0x3e9a209a, dl));
3787 // Get the significand and build it into a floating-point number with
3789 SDValue X = GetSignificand(DAG, Op1, dl);
3791 SDValue Log10ofMantissa;
3792 if (LimitFloatPrecision <= 6) {
3793 // For floating-point precision of 6:
3795 // Log10ofMantissa =
3797 // (0.60948995f - 0.10380950f * x) * x;
3799 // error 0.0014886165, which is 6 bits
3800 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3801 getF32Constant(DAG, 0xbdd49a13, dl));
3802 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3803 getF32Constant(DAG, 0x3f1c0789, dl));
3804 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3805 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3806 getF32Constant(DAG, 0x3f011300, dl));
3807 } else if (LimitFloatPrecision <= 12) {
3808 // For floating-point precision of 12:
3810 // Log10ofMantissa =
3813 // (-0.31664806f + 0.47637168e-1f * x) * x) * x;
3815 // error 0.00019228036, which is better than 12 bits
3816 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3817 getF32Constant(DAG, 0x3d431f31, dl));
3818 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
3819 getF32Constant(DAG, 0x3ea21fb2, dl));
3820 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3821 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3822 getF32Constant(DAG, 0x3f6ae232, dl));
3823 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3824 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
3825 getF32Constant(DAG, 0x3f25f7c3, dl));
3826 } else { // LimitFloatPrecision <= 18
3827 // For floating-point precision of 18:
3829 // Log10ofMantissa =
3834 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x;
3836 // error 0.0000037995730, which is better than 18 bits
3837 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3838 getF32Constant(DAG, 0x3c5d51ce, dl));
3839 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
3840 getF32Constant(DAG, 0x3e00685a, dl));
3841 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3842 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3843 getF32Constant(DAG, 0x3efb6798, dl));
3844 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3845 SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
3846 getF32Constant(DAG, 0x3f88d192, dl));
3847 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3848 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3849 getF32Constant(DAG, 0x3fc4316c, dl));
3850 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3851 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8,
3852 getF32Constant(DAG, 0x3f57ce70, dl));
3855 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, Log10ofMantissa);
3858 // No special expansion.
3859 return DAG.getNode(ISD::FLOG10, dl, Op.getValueType(), Op);
3862 /// expandExp2 - Lower an exp2 intrinsic. Handles the special sequences for
3863 /// limited-precision mode.
3864 static SDValue expandExp2(SDLoc dl, SDValue Op, SelectionDAG &DAG,
3865 const TargetLowering &TLI) {
3866 if (Op.getValueType() == MVT::f32 &&
3867 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18)
3868 return getLimitedPrecisionExp2(Op, dl, DAG);
3870 // No special expansion.
3871 return DAG.getNode(ISD::FEXP2, dl, Op.getValueType(), Op);
3874 /// visitPow - Lower a pow intrinsic. Handles the special sequences for
3875 /// limited-precision mode with x == 10.0f.
3876 static SDValue expandPow(SDLoc dl, SDValue LHS, SDValue RHS,
3877 SelectionDAG &DAG, const TargetLowering &TLI) {
3878 bool IsExp10 = false;
3879 if (LHS.getValueType() == MVT::f32 && RHS.getValueType() == MVT::f32 &&
3880 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3881 if (ConstantFPSDNode *LHSC = dyn_cast<ConstantFPSDNode>(LHS)) {
3883 IsExp10 = LHSC->isExactlyValue(Ten);
3888 // Put the exponent in the right bit position for later addition to the
3891 // #define LOG2OF10 3.3219281f
3892 // t0 = Op * LOG2OF10;
3893 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, RHS,
3894 getF32Constant(DAG, 0x40549a78, dl));
3895 return getLimitedPrecisionExp2(t0, dl, DAG);
3898 // No special expansion.
3899 return DAG.getNode(ISD::FPOW, dl, LHS.getValueType(), LHS, RHS);
3903 /// ExpandPowI - Expand a llvm.powi intrinsic.
3904 static SDValue ExpandPowI(SDLoc DL, SDValue LHS, SDValue RHS,
3905 SelectionDAG &DAG) {
3906 // If RHS is a constant, we can expand this out to a multiplication tree,
3907 // otherwise we end up lowering to a call to __powidf2 (for example). When
3908 // optimizing for size, we only want to do this if the expansion would produce
3909 // a small number of multiplies, otherwise we do the full expansion.
3910 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3911 // Get the exponent as a positive value.
3912 unsigned Val = RHSC->getSExtValue();
3913 if ((int)Val < 0) Val = -Val;
3915 // powi(x, 0) -> 1.0
3917 return DAG.getConstantFP(1.0, DL, LHS.getValueType());
3919 const Function *F = DAG.getMachineFunction().getFunction();
3920 if (!F->hasFnAttribute(Attribute::OptimizeForSize) ||
3921 // If optimizing for size, don't insert too many multiplies. This
3922 // inserts up to 5 multiplies.
3923 countPopulation(Val) + Log2_32(Val) < 7) {
3924 // We use the simple binary decomposition method to generate the multiply
3925 // sequence. There are more optimal ways to do this (for example,
3926 // powi(x,15) generates one more multiply than it should), but this has
3927 // the benefit of being both really simple and much better than a libcall.
3928 SDValue Res; // Logically starts equal to 1.0
3929 SDValue CurSquare = LHS;
3933 Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare);
3935 Res = CurSquare; // 1.0*CurSquare.
3938 CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(),
3939 CurSquare, CurSquare);
3943 // If the original was negative, invert the result, producing 1/(x*x*x).
3944 if (RHSC->getSExtValue() < 0)
3945 Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(),
3946 DAG.getConstantFP(1.0, DL, LHS.getValueType()), Res);
3951 // Otherwise, expand to a libcall.
3952 return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS);
3955 // getTruncatedArgReg - Find underlying register used for an truncated
3957 static unsigned getTruncatedArgReg(const SDValue &N) {
3958 if (N.getOpcode() != ISD::TRUNCATE)
3961 const SDValue &Ext = N.getOperand(0);
3962 if (Ext.getOpcode() == ISD::AssertZext ||
3963 Ext.getOpcode() == ISD::AssertSext) {
3964 const SDValue &CFR = Ext.getOperand(0);
3965 if (CFR.getOpcode() == ISD::CopyFromReg)
3966 return cast<RegisterSDNode>(CFR.getOperand(1))->getReg();
3967 if (CFR.getOpcode() == ISD::TRUNCATE)
3968 return getTruncatedArgReg(CFR);
3973 /// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a function
3974 /// argument, create the corresponding DBG_VALUE machine instruction for it now.
3975 /// At the end of instruction selection, they will be inserted to the entry BB.
3976 bool SelectionDAGBuilder::EmitFuncArgumentDbgValue(
3977 const Value *V, DILocalVariable *Variable, DIExpression *Expr,
3978 DILocation *DL, int64_t Offset, bool IsIndirect, const SDValue &N) {
3979 const Argument *Arg = dyn_cast<Argument>(V);
3983 MachineFunction &MF = DAG.getMachineFunction();
3984 const TargetInstrInfo *TII = DAG.getSubtarget().getInstrInfo();
3986 // Ignore inlined function arguments here.
3988 // FIXME: Should we be checking DL->inlinedAt() to determine this?
3989 if (!Variable->getScope()->getSubprogram()->describes(MF.getFunction()))
3992 Optional<MachineOperand> Op;
3993 // Some arguments' frame index is recorded during argument lowering.
3994 if (int FI = FuncInfo.getArgumentFrameIndex(Arg))
3995 Op = MachineOperand::CreateFI(FI);
3997 if (!Op && N.getNode()) {
3999 if (N.getOpcode() == ISD::CopyFromReg)
4000 Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg();
4002 Reg = getTruncatedArgReg(N);
4003 if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) {
4004 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4005 unsigned PR = RegInfo.getLiveInPhysReg(Reg);
4010 Op = MachineOperand::CreateReg(Reg, false);
4014 // Check if ValueMap has reg number.
4015 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
4016 if (VMI != FuncInfo.ValueMap.end())
4017 Op = MachineOperand::CreateReg(VMI->second, false);
4020 if (!Op && N.getNode())
4021 // Check if frame index is available.
4022 if (LoadSDNode *LNode = dyn_cast<LoadSDNode>(N.getNode()))
4023 if (FrameIndexSDNode *FINode =
4024 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode()))
4025 Op = MachineOperand::CreateFI(FINode->getIndex());
4030 assert(Variable->isValidLocationForIntrinsic(DL) &&
4031 "Expected inlined-at fields to agree");
4033 FuncInfo.ArgDbgValues.push_back(
4034 BuildMI(MF, DL, TII->get(TargetOpcode::DBG_VALUE), IsIndirect,
4035 Op->getReg(), Offset, Variable, Expr));
4037 FuncInfo.ArgDbgValues.push_back(
4038 BuildMI(MF, DL, TII->get(TargetOpcode::DBG_VALUE))
4041 .addMetadata(Variable)
4042 .addMetadata(Expr));
4047 // VisualStudio defines setjmp as _setjmp
4048 #if defined(_MSC_VER) && defined(setjmp) && \
4049 !defined(setjmp_undefined_for_msvc)
4050 # pragma push_macro("setjmp")
4052 # define setjmp_undefined_for_msvc
4055 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
4056 /// we want to emit this as a call to a named external function, return the name
4057 /// otherwise lower it and return null.
4059 SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) {
4060 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4061 SDLoc sdl = getCurSDLoc();
4062 DebugLoc dl = getCurDebugLoc();
4065 switch (Intrinsic) {
4067 // By default, turn this into a target intrinsic node.
4068 visitTargetIntrinsic(I, Intrinsic);
4070 case Intrinsic::vastart: visitVAStart(I); return nullptr;
4071 case Intrinsic::vaend: visitVAEnd(I); return nullptr;
4072 case Intrinsic::vacopy: visitVACopy(I); return nullptr;
4073 case Intrinsic::returnaddress:
4074 setValue(&I, DAG.getNode(ISD::RETURNADDR, sdl, TLI.getPointerTy(),
4075 getValue(I.getArgOperand(0))));
4077 case Intrinsic::frameaddress:
4078 setValue(&I, DAG.getNode(ISD::FRAMEADDR, sdl, TLI.getPointerTy(),
4079 getValue(I.getArgOperand(0))));
4081 case Intrinsic::read_register: {
4082 Value *Reg = I.getArgOperand(0);
4083 SDValue Chain = getRoot();
4085 DAG.getMDNode(cast<MDNode>(cast<MetadataAsValue>(Reg)->getMetadata()));
4086 EVT VT = TLI.getValueType(I.getType());
4087 Res = DAG.getNode(ISD::READ_REGISTER, sdl,
4088 DAG.getVTList(VT, MVT::Other), Chain, RegName);
4090 DAG.setRoot(Res.getValue(1));
4093 case Intrinsic::write_register: {
4094 Value *Reg = I.getArgOperand(0);
4095 Value *RegValue = I.getArgOperand(1);
4096 SDValue Chain = getRoot();
4098 DAG.getMDNode(cast<MDNode>(cast<MetadataAsValue>(Reg)->getMetadata()));
4099 DAG.setRoot(DAG.getNode(ISD::WRITE_REGISTER, sdl, MVT::Other, Chain,
4100 RegName, getValue(RegValue)));
4103 case Intrinsic::setjmp:
4104 return &"_setjmp"[!TLI.usesUnderscoreSetJmp()];
4105 case Intrinsic::longjmp:
4106 return &"_longjmp"[!TLI.usesUnderscoreLongJmp()];
4107 case Intrinsic::memcpy: {
4108 // FIXME: this definition of "user defined address space" is x86-specific
4109 // Assert for address < 256 since we support only user defined address
4111 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4113 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4115 "Unknown address space");
4116 SDValue Op1 = getValue(I.getArgOperand(0));
4117 SDValue Op2 = getValue(I.getArgOperand(1));
4118 SDValue Op3 = getValue(I.getArgOperand(2));
4119 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4121 Align = 1; // @llvm.memcpy defines 0 and 1 to both mean no alignment.
4122 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4123 bool isTC = I.isTailCall() && isInTailCallPosition(&I, DAG.getTarget());
4124 SDValue MC = DAG.getMemcpy(getRoot(), sdl, Op1, Op2, Op3, Align, isVol,
4126 MachinePointerInfo(I.getArgOperand(0)),
4127 MachinePointerInfo(I.getArgOperand(1)));
4128 updateDAGForMaybeTailCall(MC);
4131 case Intrinsic::memset: {
4132 // FIXME: this definition of "user defined address space" is x86-specific
4133 // Assert for address < 256 since we support only user defined address
4135 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4137 "Unknown address space");
4138 SDValue Op1 = getValue(I.getArgOperand(0));
4139 SDValue Op2 = getValue(I.getArgOperand(1));
4140 SDValue Op3 = getValue(I.getArgOperand(2));
4141 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4143 Align = 1; // @llvm.memset defines 0 and 1 to both mean no alignment.
4144 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4145 bool isTC = I.isTailCall() && isInTailCallPosition(&I, DAG.getTarget());
4146 SDValue MS = DAG.getMemset(getRoot(), sdl, Op1, Op2, Op3, Align, isVol,
4147 isTC, MachinePointerInfo(I.getArgOperand(0)));
4148 updateDAGForMaybeTailCall(MS);
4151 case Intrinsic::memmove: {
4152 // FIXME: this definition of "user defined address space" is x86-specific
4153 // Assert for address < 256 since we support only user defined address
4155 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4157 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4159 "Unknown address space");
4160 SDValue Op1 = getValue(I.getArgOperand(0));
4161 SDValue Op2 = getValue(I.getArgOperand(1));
4162 SDValue Op3 = getValue(I.getArgOperand(2));
4163 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4165 Align = 1; // @llvm.memmove defines 0 and 1 to both mean no alignment.
4166 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4167 bool isTC = I.isTailCall() && isInTailCallPosition(&I, DAG.getTarget());
4168 SDValue MM = DAG.getMemmove(getRoot(), sdl, Op1, Op2, Op3, Align, isVol,
4169 isTC, MachinePointerInfo(I.getArgOperand(0)),
4170 MachinePointerInfo(I.getArgOperand(1)));
4171 updateDAGForMaybeTailCall(MM);
4174 case Intrinsic::dbg_declare: {
4175 const DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
4176 DILocalVariable *Variable = DI.getVariable();
4177 DIExpression *Expression = DI.getExpression();
4178 const Value *Address = DI.getAddress();
4179 assert(Variable && "Missing variable");
4181 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4185 // Check if address has undef value.
4186 if (isa<UndefValue>(Address) ||
4187 (Address->use_empty() && !isa<Argument>(Address))) {
4188 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4192 SDValue &N = NodeMap[Address];
4193 if (!N.getNode() && isa<Argument>(Address))
4194 // Check unused arguments map.
4195 N = UnusedArgNodeMap[Address];
4198 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
4199 Address = BCI->getOperand(0);
4200 // Parameters are handled specially.
4201 bool isParameter = Variable->getTag() == dwarf::DW_TAG_arg_variable ||
4202 isa<Argument>(Address);
4204 const AllocaInst *AI = dyn_cast<AllocaInst>(Address);
4206 if (isParameter && !AI) {
4207 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode());
4209 // Byval parameter. We have a frame index at this point.
4210 SDV = DAG.getFrameIndexDbgValue(
4211 Variable, Expression, FINode->getIndex(), 0, dl, SDNodeOrder);
4213 // Address is an argument, so try to emit its dbg value using
4214 // virtual register info from the FuncInfo.ValueMap.
4215 EmitFuncArgumentDbgValue(Address, Variable, Expression, dl, 0, false,
4220 SDV = DAG.getDbgValue(Variable, Expression, N.getNode(), N.getResNo(),
4221 true, 0, dl, SDNodeOrder);
4223 // Can't do anything with other non-AI cases yet.
4224 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4225 DEBUG(dbgs() << "non-AllocaInst issue for Address: \n\t");
4226 DEBUG(Address->dump());
4229 DAG.AddDbgValue(SDV, N.getNode(), isParameter);
4231 // If Address is an argument then try to emit its dbg value using
4232 // virtual register info from the FuncInfo.ValueMap.
4233 if (!EmitFuncArgumentDbgValue(Address, Variable, Expression, dl, 0, false,
4235 // If variable is pinned by a alloca in dominating bb then
4236 // use StaticAllocaMap.
4237 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Address)) {
4238 if (AI->getParent() != DI.getParent()) {
4239 DenseMap<const AllocaInst*, int>::iterator SI =
4240 FuncInfo.StaticAllocaMap.find(AI);
4241 if (SI != FuncInfo.StaticAllocaMap.end()) {
4242 SDV = DAG.getFrameIndexDbgValue(Variable, Expression, SI->second,
4243 0, dl, SDNodeOrder);
4244 DAG.AddDbgValue(SDV, nullptr, false);
4249 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4254 case Intrinsic::dbg_value: {
4255 const DbgValueInst &DI = cast<DbgValueInst>(I);
4256 assert(DI.getVariable() && "Missing variable");
4258 DILocalVariable *Variable = DI.getVariable();
4259 DIExpression *Expression = DI.getExpression();
4260 uint64_t Offset = DI.getOffset();
4261 const Value *V = DI.getValue();
4266 if (isa<ConstantInt>(V) || isa<ConstantFP>(V) || isa<UndefValue>(V)) {
4267 SDV = DAG.getConstantDbgValue(Variable, Expression, V, Offset, dl,
4269 DAG.AddDbgValue(SDV, nullptr, false);
4271 // Do not use getValue() in here; we don't want to generate code at
4272 // this point if it hasn't been done yet.
4273 SDValue N = NodeMap[V];
4274 if (!N.getNode() && isa<Argument>(V))
4275 // Check unused arguments map.
4276 N = UnusedArgNodeMap[V];
4278 // A dbg.value for an alloca is always indirect.
4279 bool IsIndirect = isa<AllocaInst>(V) || Offset != 0;
4280 if (!EmitFuncArgumentDbgValue(V, Variable, Expression, dl, Offset,
4282 SDV = DAG.getDbgValue(Variable, Expression, N.getNode(), N.getResNo(),
4283 IsIndirect, Offset, dl, SDNodeOrder);
4284 DAG.AddDbgValue(SDV, N.getNode(), false);
4286 } else if (!V->use_empty() ) {
4287 // Do not call getValue(V) yet, as we don't want to generate code.
4288 // Remember it for later.
4289 DanglingDebugInfo DDI(&DI, dl, SDNodeOrder);
4290 DanglingDebugInfoMap[V] = DDI;
4292 // We may expand this to cover more cases. One case where we have no
4293 // data available is an unreferenced parameter.
4294 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4298 // Build a debug info table entry.
4299 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
4300 V = BCI->getOperand(0);
4301 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
4302 // Don't handle byval struct arguments or VLAs, for example.
4304 DEBUG(dbgs() << "Dropping debug location info for:\n " << DI << "\n");
4305 DEBUG(dbgs() << " Last seen at:\n " << *V << "\n");
4308 DenseMap<const AllocaInst*, int>::iterator SI =
4309 FuncInfo.StaticAllocaMap.find(AI);
4310 if (SI == FuncInfo.StaticAllocaMap.end())
4311 return nullptr; // VLAs.
4315 case Intrinsic::eh_typeid_for: {
4316 // Find the type id for the given typeinfo.
4317 GlobalValue *GV = ExtractTypeInfo(I.getArgOperand(0));
4318 unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(GV);
4319 Res = DAG.getConstant(TypeID, sdl, MVT::i32);
4324 case Intrinsic::eh_return_i32:
4325 case Intrinsic::eh_return_i64:
4326 DAG.getMachineFunction().getMMI().setCallsEHReturn(true);
4327 DAG.setRoot(DAG.getNode(ISD::EH_RETURN, sdl,
4330 getValue(I.getArgOperand(0)),
4331 getValue(I.getArgOperand(1))));
4333 case Intrinsic::eh_unwind_init:
4334 DAG.getMachineFunction().getMMI().setCallsUnwindInit(true);
4336 case Intrinsic::eh_dwarf_cfa: {
4337 SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getArgOperand(0)), sdl,
4338 TLI.getPointerTy());
4339 SDValue Offset = DAG.getNode(ISD::ADD, sdl,
4340 CfaArg.getValueType(),
4341 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, sdl,
4342 CfaArg.getValueType()),
4344 SDValue FA = DAG.getNode(ISD::FRAMEADDR, sdl, TLI.getPointerTy(),
4345 DAG.getConstant(0, sdl, TLI.getPointerTy()));
4346 setValue(&I, DAG.getNode(ISD::ADD, sdl, FA.getValueType(),
4350 case Intrinsic::eh_sjlj_callsite: {
4351 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4352 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(0));
4353 assert(CI && "Non-constant call site value in eh.sjlj.callsite!");
4354 assert(MMI.getCurrentCallSite() == 0 && "Overlapping call sites!");
4356 MMI.setCurrentCallSite(CI->getZExtValue());
4359 case Intrinsic::eh_sjlj_functioncontext: {
4360 // Get and store the index of the function context.
4361 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4363 cast<AllocaInst>(I.getArgOperand(0)->stripPointerCasts());
4364 int FI = FuncInfo.StaticAllocaMap[FnCtx];
4365 MFI->setFunctionContextIndex(FI);
4368 case Intrinsic::eh_sjlj_setjmp: {
4371 Ops[1] = getValue(I.getArgOperand(0));
4372 SDValue Op = DAG.getNode(ISD::EH_SJLJ_SETJMP, sdl,
4373 DAG.getVTList(MVT::i32, MVT::Other), Ops);
4374 setValue(&I, Op.getValue(0));
4375 DAG.setRoot(Op.getValue(1));
4378 case Intrinsic::eh_sjlj_longjmp: {
4379 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, sdl, MVT::Other,
4380 getRoot(), getValue(I.getArgOperand(0))));
4384 case Intrinsic::masked_gather:
4385 visitMaskedGather(I);
4387 case Intrinsic::masked_load:
4390 case Intrinsic::masked_scatter:
4391 visitMaskedScatter(I);
4393 case Intrinsic::masked_store:
4394 visitMaskedStore(I);
4396 case Intrinsic::x86_mmx_pslli_w:
4397 case Intrinsic::x86_mmx_pslli_d:
4398 case Intrinsic::x86_mmx_pslli_q:
4399 case Intrinsic::x86_mmx_psrli_w:
4400 case Intrinsic::x86_mmx_psrli_d:
4401 case Intrinsic::x86_mmx_psrli_q:
4402 case Intrinsic::x86_mmx_psrai_w:
4403 case Intrinsic::x86_mmx_psrai_d: {
4404 SDValue ShAmt = getValue(I.getArgOperand(1));
4405 if (isa<ConstantSDNode>(ShAmt)) {
4406 visitTargetIntrinsic(I, Intrinsic);
4409 unsigned NewIntrinsic = 0;
4410 EVT ShAmtVT = MVT::v2i32;
4411 switch (Intrinsic) {
4412 case Intrinsic::x86_mmx_pslli_w:
4413 NewIntrinsic = Intrinsic::x86_mmx_psll_w;
4415 case Intrinsic::x86_mmx_pslli_d:
4416 NewIntrinsic = Intrinsic::x86_mmx_psll_d;
4418 case Intrinsic::x86_mmx_pslli_q:
4419 NewIntrinsic = Intrinsic::x86_mmx_psll_q;
4421 case Intrinsic::x86_mmx_psrli_w:
4422 NewIntrinsic = Intrinsic::x86_mmx_psrl_w;
4424 case Intrinsic::x86_mmx_psrli_d:
4425 NewIntrinsic = Intrinsic::x86_mmx_psrl_d;
4427 case Intrinsic::x86_mmx_psrli_q:
4428 NewIntrinsic = Intrinsic::x86_mmx_psrl_q;
4430 case Intrinsic::x86_mmx_psrai_w:
4431 NewIntrinsic = Intrinsic::x86_mmx_psra_w;
4433 case Intrinsic::x86_mmx_psrai_d:
4434 NewIntrinsic = Intrinsic::x86_mmx_psra_d;
4436 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4439 // The vector shift intrinsics with scalars uses 32b shift amounts but
4440 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
4442 // We must do this early because v2i32 is not a legal type.
4445 ShOps[1] = DAG.getConstant(0, sdl, MVT::i32);
4446 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, sdl, ShAmtVT, ShOps);
4447 EVT DestVT = TLI.getValueType(I.getType());
4448 ShAmt = DAG.getNode(ISD::BITCAST, sdl, DestVT, ShAmt);
4449 Res = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, sdl, DestVT,
4450 DAG.getConstant(NewIntrinsic, sdl, MVT::i32),
4451 getValue(I.getArgOperand(0)), ShAmt);
4455 case Intrinsic::convertff:
4456 case Intrinsic::convertfsi:
4457 case Intrinsic::convertfui:
4458 case Intrinsic::convertsif:
4459 case Intrinsic::convertuif:
4460 case Intrinsic::convertss:
4461 case Intrinsic::convertsu:
4462 case Intrinsic::convertus:
4463 case Intrinsic::convertuu: {
4464 ISD::CvtCode Code = ISD::CVT_INVALID;
4465 switch (Intrinsic) {
4466 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4467 case Intrinsic::convertff: Code = ISD::CVT_FF; break;
4468 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break;
4469 case Intrinsic::convertfui: Code = ISD::CVT_FU; break;
4470 case Intrinsic::convertsif: Code = ISD::CVT_SF; break;
4471 case Intrinsic::convertuif: Code = ISD::CVT_UF; break;
4472 case Intrinsic::convertss: Code = ISD::CVT_SS; break;
4473 case Intrinsic::convertsu: Code = ISD::CVT_SU; break;
4474 case Intrinsic::convertus: Code = ISD::CVT_US; break;
4475 case Intrinsic::convertuu: Code = ISD::CVT_UU; break;
4477 EVT DestVT = TLI.getValueType(I.getType());
4478 const Value *Op1 = I.getArgOperand(0);
4479 Res = DAG.getConvertRndSat(DestVT, sdl, getValue(Op1),
4480 DAG.getValueType(DestVT),
4481 DAG.getValueType(getValue(Op1).getValueType()),
4482 getValue(I.getArgOperand(1)),
4483 getValue(I.getArgOperand(2)),
4488 case Intrinsic::powi:
4489 setValue(&I, ExpandPowI(sdl, getValue(I.getArgOperand(0)),
4490 getValue(I.getArgOperand(1)), DAG));
4492 case Intrinsic::log:
4493 setValue(&I, expandLog(sdl, getValue(I.getArgOperand(0)), DAG, TLI));
4495 case Intrinsic::log2:
4496 setValue(&I, expandLog2(sdl, getValue(I.getArgOperand(0)), DAG, TLI));
4498 case Intrinsic::log10:
4499 setValue(&I, expandLog10(sdl, getValue(I.getArgOperand(0)), DAG, TLI));
4501 case Intrinsic::exp:
4502 setValue(&I, expandExp(sdl, getValue(I.getArgOperand(0)), DAG, TLI));
4504 case Intrinsic::exp2:
4505 setValue(&I, expandExp2(sdl, getValue(I.getArgOperand(0)), DAG, TLI));
4507 case Intrinsic::pow:
4508 setValue(&I, expandPow(sdl, getValue(I.getArgOperand(0)),
4509 getValue(I.getArgOperand(1)), DAG, TLI));
4511 case Intrinsic::sqrt:
4512 case Intrinsic::fabs:
4513 case Intrinsic::sin:
4514 case Intrinsic::cos:
4515 case Intrinsic::floor:
4516 case Intrinsic::ceil:
4517 case Intrinsic::trunc:
4518 case Intrinsic::rint:
4519 case Intrinsic::nearbyint:
4520 case Intrinsic::round: {
4522 switch (Intrinsic) {
4523 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4524 case Intrinsic::sqrt: Opcode = ISD::FSQRT; break;
4525 case Intrinsic::fabs: Opcode = ISD::FABS; break;
4526 case Intrinsic::sin: Opcode = ISD::FSIN; break;
4527 case Intrinsic::cos: Opcode = ISD::FCOS; break;
4528 case Intrinsic::floor: Opcode = ISD::FFLOOR; break;
4529 case Intrinsic::ceil: Opcode = ISD::FCEIL; break;
4530 case Intrinsic::trunc: Opcode = ISD::FTRUNC; break;
4531 case Intrinsic::rint: Opcode = ISD::FRINT; break;
4532 case Intrinsic::nearbyint: Opcode = ISD::FNEARBYINT; break;
4533 case Intrinsic::round: Opcode = ISD::FROUND; break;
4536 setValue(&I, DAG.getNode(Opcode, sdl,
4537 getValue(I.getArgOperand(0)).getValueType(),
4538 getValue(I.getArgOperand(0))));
4541 case Intrinsic::minnum:
4542 setValue(&I, DAG.getNode(ISD::FMINNUM, sdl,
4543 getValue(I.getArgOperand(0)).getValueType(),
4544 getValue(I.getArgOperand(0)),
4545 getValue(I.getArgOperand(1))));
4547 case Intrinsic::maxnum:
4548 setValue(&I, DAG.getNode(ISD::FMAXNUM, sdl,
4549 getValue(I.getArgOperand(0)).getValueType(),
4550 getValue(I.getArgOperand(0)),
4551 getValue(I.getArgOperand(1))));
4553 case Intrinsic::copysign:
4554 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, sdl,
4555 getValue(I.getArgOperand(0)).getValueType(),
4556 getValue(I.getArgOperand(0)),
4557 getValue(I.getArgOperand(1))));
4559 case Intrinsic::fma:
4560 setValue(&I, DAG.getNode(ISD::FMA, sdl,
4561 getValue(I.getArgOperand(0)).getValueType(),
4562 getValue(I.getArgOperand(0)),
4563 getValue(I.getArgOperand(1)),
4564 getValue(I.getArgOperand(2))));
4566 case Intrinsic::fmuladd: {
4567 EVT VT = TLI.getValueType(I.getType());
4568 if (TM.Options.AllowFPOpFusion != FPOpFusion::Strict &&
4569 TLI.isFMAFasterThanFMulAndFAdd(VT)) {
4570 setValue(&I, DAG.getNode(ISD::FMA, sdl,
4571 getValue(I.getArgOperand(0)).getValueType(),
4572 getValue(I.getArgOperand(0)),
4573 getValue(I.getArgOperand(1)),
4574 getValue(I.getArgOperand(2))));
4576 SDValue Mul = DAG.getNode(ISD::FMUL, sdl,
4577 getValue(I.getArgOperand(0)).getValueType(),
4578 getValue(I.getArgOperand(0)),
4579 getValue(I.getArgOperand(1)));
4580 SDValue Add = DAG.getNode(ISD::FADD, sdl,
4581 getValue(I.getArgOperand(0)).getValueType(),
4583 getValue(I.getArgOperand(2)));
4588 case Intrinsic::convert_to_fp16:
4589 setValue(&I, DAG.getNode(ISD::BITCAST, sdl, MVT::i16,
4590 DAG.getNode(ISD::FP_ROUND, sdl, MVT::f16,
4591 getValue(I.getArgOperand(0)),
4592 DAG.getTargetConstant(0, sdl,
4595 case Intrinsic::convert_from_fp16:
4597 DAG.getNode(ISD::FP_EXTEND, sdl, TLI.getValueType(I.getType()),
4598 DAG.getNode(ISD::BITCAST, sdl, MVT::f16,
4599 getValue(I.getArgOperand(0)))));
4601 case Intrinsic::pcmarker: {
4602 SDValue Tmp = getValue(I.getArgOperand(0));
4603 DAG.setRoot(DAG.getNode(ISD::PCMARKER, sdl, MVT::Other, getRoot(), Tmp));
4606 case Intrinsic::readcyclecounter: {
4607 SDValue Op = getRoot();
4608 Res = DAG.getNode(ISD::READCYCLECOUNTER, sdl,
4609 DAG.getVTList(MVT::i64, MVT::Other), Op);
4611 DAG.setRoot(Res.getValue(1));
4614 case Intrinsic::bswap:
4615 setValue(&I, DAG.getNode(ISD::BSWAP, sdl,
4616 getValue(I.getArgOperand(0)).getValueType(),
4617 getValue(I.getArgOperand(0))));
4619 case Intrinsic::cttz: {
4620 SDValue Arg = getValue(I.getArgOperand(0));
4621 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1));
4622 EVT Ty = Arg.getValueType();
4623 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTTZ : ISD::CTTZ_ZERO_UNDEF,
4627 case Intrinsic::ctlz: {
4628 SDValue Arg = getValue(I.getArgOperand(0));
4629 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1));
4630 EVT Ty = Arg.getValueType();
4631 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTLZ : ISD::CTLZ_ZERO_UNDEF,
4635 case Intrinsic::ctpop: {
4636 SDValue Arg = getValue(I.getArgOperand(0));
4637 EVT Ty = Arg.getValueType();
4638 setValue(&I, DAG.getNode(ISD::CTPOP, sdl, Ty, Arg));
4641 case Intrinsic::stacksave: {
4642 SDValue Op = getRoot();
4643 Res = DAG.getNode(ISD::STACKSAVE, sdl,
4644 DAG.getVTList(TLI.getPointerTy(), MVT::Other), Op);
4646 DAG.setRoot(Res.getValue(1));
4649 case Intrinsic::stackrestore: {
4650 Res = getValue(I.getArgOperand(0));
4651 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, sdl, MVT::Other, getRoot(), Res));
4654 case Intrinsic::stackprotector: {
4655 // Emit code into the DAG to store the stack guard onto the stack.
4656 MachineFunction &MF = DAG.getMachineFunction();
4657 MachineFrameInfo *MFI = MF.getFrameInfo();
4658 EVT PtrTy = TLI.getPointerTy();
4659 SDValue Src, Chain = getRoot();
4660 const Value *Ptr = cast<LoadInst>(I.getArgOperand(0))->getPointerOperand();
4661 const GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr);
4663 // See if Ptr is a bitcast. If it is, look through it and see if we can get
4664 // global variable __stack_chk_guard.
4666 if (const Operator *BC = dyn_cast<Operator>(Ptr))
4667 if (BC->getOpcode() == Instruction::BitCast)
4668 GV = dyn_cast<GlobalVariable>(BC->getOperand(0));
4670 if (GV && TLI.useLoadStackGuardNode()) {
4671 // Emit a LOAD_STACK_GUARD node.
4672 MachineSDNode *Node = DAG.getMachineNode(TargetOpcode::LOAD_STACK_GUARD,
4674 MachinePointerInfo MPInfo(GV);
4675 MachineInstr::mmo_iterator MemRefs = MF.allocateMemRefsArray(1);
4676 unsigned Flags = MachineMemOperand::MOLoad |
4677 MachineMemOperand::MOInvariant;
4678 *MemRefs = MF.getMachineMemOperand(MPInfo, Flags,
4679 PtrTy.getSizeInBits() / 8,
4680 DAG.getEVTAlignment(PtrTy));
4681 Node->setMemRefs(MemRefs, MemRefs + 1);
4683 // Copy the guard value to a virtual register so that it can be
4684 // retrieved in the epilogue.
4685 Src = SDValue(Node, 0);
4686 const TargetRegisterClass *RC =
4687 TLI.getRegClassFor(Src.getSimpleValueType());
4688 unsigned Reg = MF.getRegInfo().createVirtualRegister(RC);
4690 SPDescriptor.setGuardReg(Reg);
4691 Chain = DAG.getCopyToReg(Chain, sdl, Reg, Src);
4693 Src = getValue(I.getArgOperand(0)); // The guard's value.
4696 AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
4698 int FI = FuncInfo.StaticAllocaMap[Slot];
4699 MFI->setStackProtectorIndex(FI);
4701 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
4703 // Store the stack protector onto the stack.
4704 Res = DAG.getStore(Chain, sdl, Src, FIN,
4705 MachinePointerInfo::getFixedStack(FI),
4711 case Intrinsic::objectsize: {
4712 // If we don't know by now, we're never going to know.
4713 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1));
4715 assert(CI && "Non-constant type in __builtin_object_size?");
4717 SDValue Arg = getValue(I.getCalledValue());
4718 EVT Ty = Arg.getValueType();
4721 Res = DAG.getConstant(-1ULL, sdl, Ty);
4723 Res = DAG.getConstant(0, sdl, Ty);
4728 case Intrinsic::annotation:
4729 case Intrinsic::ptr_annotation:
4730 // Drop the intrinsic, but forward the value
4731 setValue(&I, getValue(I.getOperand(0)));
4733 case Intrinsic::assume:
4734 case Intrinsic::var_annotation:
4735 // Discard annotate attributes and assumptions
4738 case Intrinsic::init_trampoline: {
4739 const Function *F = cast<Function>(I.getArgOperand(1)->stripPointerCasts());
4743 Ops[1] = getValue(I.getArgOperand(0));
4744 Ops[2] = getValue(I.getArgOperand(1));
4745 Ops[3] = getValue(I.getArgOperand(2));
4746 Ops[4] = DAG.getSrcValue(I.getArgOperand(0));
4747 Ops[5] = DAG.getSrcValue(F);
4749 Res = DAG.getNode(ISD::INIT_TRAMPOLINE, sdl, MVT::Other, Ops);
4754 case Intrinsic::adjust_trampoline: {
4755 setValue(&I, DAG.getNode(ISD::ADJUST_TRAMPOLINE, sdl,
4757 getValue(I.getArgOperand(0))));
4760 case Intrinsic::gcroot:
4762 const Value *Alloca = I.getArgOperand(0)->stripPointerCasts();
4763 const Constant *TypeMap = cast<Constant>(I.getArgOperand(1));
4765 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode());
4766 GFI->addStackRoot(FI->getIndex(), TypeMap);
4769 case Intrinsic::gcread:
4770 case Intrinsic::gcwrite:
4771 llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!");
4772 case Intrinsic::flt_rounds:
4773 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, sdl, MVT::i32));
4776 case Intrinsic::expect: {
4777 // Just replace __builtin_expect(exp, c) with EXP.
4778 setValue(&I, getValue(I.getArgOperand(0)));
4782 case Intrinsic::debugtrap:
4783 case Intrinsic::trap: {
4784 StringRef TrapFuncName =
4786 .getAttribute(AttributeSet::FunctionIndex, "trap-func-name")
4787 .getValueAsString();
4788 if (TrapFuncName.empty()) {
4789 ISD::NodeType Op = (Intrinsic == Intrinsic::trap) ?
4790 ISD::TRAP : ISD::DEBUGTRAP;
4791 DAG.setRoot(DAG.getNode(Op, sdl,MVT::Other, getRoot()));
4794 TargetLowering::ArgListTy Args;
4796 TargetLowering::CallLoweringInfo CLI(DAG);
4797 CLI.setDebugLoc(sdl).setChain(getRoot())
4798 .setCallee(CallingConv::C, I.getType(),
4799 DAG.getExternalSymbol(TrapFuncName.data(), TLI.getPointerTy()),
4800 std::move(Args), 0);
4802 std::pair<SDValue, SDValue> Result = TLI.LowerCallTo(CLI);
4803 DAG.setRoot(Result.second);
4807 case Intrinsic::uadd_with_overflow:
4808 case Intrinsic::sadd_with_overflow:
4809 case Intrinsic::usub_with_overflow:
4810 case Intrinsic::ssub_with_overflow:
4811 case Intrinsic::umul_with_overflow:
4812 case Intrinsic::smul_with_overflow: {
4814 switch (Intrinsic) {
4815 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4816 case Intrinsic::uadd_with_overflow: Op = ISD::UADDO; break;
4817 case Intrinsic::sadd_with_overflow: Op = ISD::SADDO; break;
4818 case Intrinsic::usub_with_overflow: Op = ISD::USUBO; break;
4819 case Intrinsic::ssub_with_overflow: Op = ISD::SSUBO; break;
4820 case Intrinsic::umul_with_overflow: Op = ISD::UMULO; break;
4821 case Intrinsic::smul_with_overflow: Op = ISD::SMULO; break;
4823 SDValue Op1 = getValue(I.getArgOperand(0));
4824 SDValue Op2 = getValue(I.getArgOperand(1));
4826 SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1);
4827 setValue(&I, DAG.getNode(Op, sdl, VTs, Op1, Op2));
4830 case Intrinsic::prefetch: {
4832 unsigned rw = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
4834 Ops[1] = getValue(I.getArgOperand(0));
4835 Ops[2] = getValue(I.getArgOperand(1));
4836 Ops[3] = getValue(I.getArgOperand(2));
4837 Ops[4] = getValue(I.getArgOperand(3));
4838 DAG.setRoot(DAG.getMemIntrinsicNode(ISD::PREFETCH, sdl,
4839 DAG.getVTList(MVT::Other), Ops,
4840 EVT::getIntegerVT(*Context, 8),
4841 MachinePointerInfo(I.getArgOperand(0)),
4843 false, /* volatile */
4845 rw==1)); /* write */
4848 case Intrinsic::lifetime_start:
4849 case Intrinsic::lifetime_end: {
4850 bool IsStart = (Intrinsic == Intrinsic::lifetime_start);
4851 // Stack coloring is not enabled in O0, discard region information.
4852 if (TM.getOptLevel() == CodeGenOpt::None)
4855 SmallVector<Value *, 4> Allocas;
4856 GetUnderlyingObjects(I.getArgOperand(1), Allocas, *DL);
4858 for (SmallVectorImpl<Value*>::iterator Object = Allocas.begin(),
4859 E = Allocas.end(); Object != E; ++Object) {
4860 AllocaInst *LifetimeObject = dyn_cast_or_null<AllocaInst>(*Object);
4862 // Could not find an Alloca.
4863 if (!LifetimeObject)
4866 // First check that the Alloca is static, otherwise it won't have a
4867 // valid frame index.
4868 auto SI = FuncInfo.StaticAllocaMap.find(LifetimeObject);
4869 if (SI == FuncInfo.StaticAllocaMap.end())
4872 int FI = SI->second;
4876 Ops[1] = DAG.getFrameIndex(FI, TLI.getPointerTy(), true);
4877 unsigned Opcode = (IsStart ? ISD::LIFETIME_START : ISD::LIFETIME_END);
4879 Res = DAG.getNode(Opcode, sdl, MVT::Other, Ops);
4884 case Intrinsic::invariant_start:
4885 // Discard region information.
4886 setValue(&I, DAG.getUNDEF(TLI.getPointerTy()));
4888 case Intrinsic::invariant_end:
4889 // Discard region information.
4891 case Intrinsic::stackprotectorcheck: {
4892 // Do not actually emit anything for this basic block. Instead we initialize
4893 // the stack protector descriptor and export the guard variable so we can
4894 // access it in FinishBasicBlock.
4895 const BasicBlock *BB = I.getParent();
4896 SPDescriptor.initialize(BB, FuncInfo.MBBMap[BB], I);
4897 ExportFromCurrentBlock(SPDescriptor.getGuard());
4899 // Flush our exports since we are going to process a terminator.
4900 (void)getControlRoot();
4903 case Intrinsic::clear_cache:
4904 return TLI.getClearCacheBuiltinName();
4905 case Intrinsic::eh_actions:
4906 setValue(&I, DAG.getUNDEF(TLI.getPointerTy()));
4908 case Intrinsic::donothing:
4911 case Intrinsic::experimental_stackmap: {
4915 case Intrinsic::experimental_patchpoint_void:
4916 case Intrinsic::experimental_patchpoint_i64: {
4917 visitPatchpoint(&I);
4920 case Intrinsic::experimental_gc_statepoint: {
4924 case Intrinsic::experimental_gc_result_int:
4925 case Intrinsic::experimental_gc_result_float:
4926 case Intrinsic::experimental_gc_result_ptr:
4927 case Intrinsic::experimental_gc_result: {
4931 case Intrinsic::experimental_gc_relocate: {
4935 case Intrinsic::instrprof_increment:
4936 llvm_unreachable("instrprof failed to lower an increment");
4938 case Intrinsic::frameescape: {
4939 MachineFunction &MF = DAG.getMachineFunction();
4940 const TargetInstrInfo *TII = DAG.getSubtarget().getInstrInfo();
4942 // Directly emit some FRAME_ALLOC machine instrs. Label assignment emission
4943 // is the same on all targets.
4944 for (unsigned Idx = 0, E = I.getNumArgOperands(); Idx < E; ++Idx) {
4945 Value *Arg = I.getArgOperand(Idx)->stripPointerCasts();
4946 if (isa<ConstantPointerNull>(Arg))
4947 continue; // Skip null pointers. They represent a hole in index space.
4948 AllocaInst *Slot = cast<AllocaInst>(Arg);
4949 assert(FuncInfo.StaticAllocaMap.count(Slot) &&
4950 "can only escape static allocas");
4951 int FI = FuncInfo.StaticAllocaMap[Slot];
4952 MCSymbol *FrameAllocSym =
4953 MF.getMMI().getContext().getOrCreateFrameAllocSymbol(
4954 GlobalValue::getRealLinkageName(MF.getName()), Idx);
4955 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, dl,
4956 TII->get(TargetOpcode::FRAME_ALLOC))
4957 .addSym(FrameAllocSym)
4964 case Intrinsic::framerecover: {
4965 // i8* @llvm.framerecover(i8* %fn, i8* %fp, i32 %idx)
4966 MachineFunction &MF = DAG.getMachineFunction();
4967 MVT PtrVT = TLI.getPointerTy(0);
4969 // Get the symbol that defines the frame offset.
4970 auto *Fn = cast<Function>(I.getArgOperand(0)->stripPointerCasts());
4971 auto *Idx = cast<ConstantInt>(I.getArgOperand(2));
4972 unsigned IdxVal = unsigned(Idx->getLimitedValue(INT_MAX));
4973 MCSymbol *FrameAllocSym =
4974 MF.getMMI().getContext().getOrCreateFrameAllocSymbol(
4975 GlobalValue::getRealLinkageName(Fn->getName()), IdxVal);
4977 // Create a MCSymbol for the label to avoid any target lowering
4978 // that would make this PC relative.
4979 SDValue OffsetSym = DAG.getMCSymbol(FrameAllocSym, PtrVT);
4981 DAG.getNode(ISD::FRAME_ALLOC_RECOVER, sdl, PtrVT, OffsetSym);
4983 // Add the offset to the FP.
4984 Value *FP = I.getArgOperand(1);
4985 SDValue FPVal = getValue(FP);
4986 SDValue Add = DAG.getNode(ISD::ADD, sdl, PtrVT, FPVal, OffsetVal);
4991 case Intrinsic::eh_begincatch:
4992 case Intrinsic::eh_endcatch:
4993 llvm_unreachable("begin/end catch intrinsics not lowered in codegen");
4994 case Intrinsic::eh_exceptioncode: {
4995 unsigned Reg = TLI.getExceptionPointerRegister();
4996 assert(Reg && "cannot get exception code on this platform");
4997 MVT PtrVT = TLI.getPointerTy();
4998 const TargetRegisterClass *PtrRC = TLI.getRegClassFor(PtrVT);
4999 assert(FuncInfo.MBB->isLandingPad() && "eh.exceptioncode in non-lpad");
5000 unsigned VReg = FuncInfo.MBB->addLiveIn(Reg, PtrRC);
5002 DAG.getCopyFromReg(DAG.getEntryNode(), getCurSDLoc(), VReg, PtrVT);
5003 N = DAG.getZExtOrTrunc(N, getCurSDLoc(), MVT::i32);
5010 std::pair<SDValue, SDValue>
5011 SelectionDAGBuilder::lowerInvokable(TargetLowering::CallLoweringInfo &CLI,
5012 MachineBasicBlock *LandingPad) {
5013 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5014 MCSymbol *BeginLabel = nullptr;
5017 // Insert a label before the invoke call to mark the try range. This can be
5018 // used to detect deletion of the invoke via the MachineModuleInfo.
5019 BeginLabel = MMI.getContext().createTempSymbol();
5021 // For SjLj, keep track of which landing pads go with which invokes
5022 // so as to maintain the ordering of pads in the LSDA.
5023 unsigned CallSiteIndex = MMI.getCurrentCallSite();
5024 if (CallSiteIndex) {
5025 MMI.setCallSiteBeginLabel(BeginLabel, CallSiteIndex);
5026 LPadToCallSiteMap[LandingPad].push_back(CallSiteIndex);
5028 // Now that the call site is handled, stop tracking it.
5029 MMI.setCurrentCallSite(0);
5032 // Both PendingLoads and PendingExports must be flushed here;
5033 // this call might not return.
5035 DAG.setRoot(DAG.getEHLabel(getCurSDLoc(), getControlRoot(), BeginLabel));
5037 CLI.setChain(getRoot());
5039 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5040 std::pair<SDValue, SDValue> Result = TLI.LowerCallTo(CLI);
5042 assert((CLI.IsTailCall || Result.second.getNode()) &&
5043 "Non-null chain expected with non-tail call!");
5044 assert((Result.second.getNode() || !Result.first.getNode()) &&
5045 "Null value expected with tail call!");
5047 if (!Result.second.getNode()) {
5048 // As a special case, a null chain means that a tail call has been emitted
5049 // and the DAG root is already updated.
5052 // Since there's no actual continuation from this block, nothing can be
5053 // relying on us setting vregs for them.
5054 PendingExports.clear();
5056 DAG.setRoot(Result.second);
5060 // Insert a label at the end of the invoke call to mark the try range. This
5061 // can be used to detect deletion of the invoke via the MachineModuleInfo.
5062 MCSymbol *EndLabel = MMI.getContext().createTempSymbol();
5063 DAG.setRoot(DAG.getEHLabel(getCurSDLoc(), getRoot(), EndLabel));
5065 // Inform MachineModuleInfo of range.
5066 MMI.addInvoke(LandingPad, BeginLabel, EndLabel);
5072 void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee,
5074 MachineBasicBlock *LandingPad) {
5075 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
5076 FunctionType *FTy = cast<FunctionType>(PT->getElementType());
5077 Type *RetTy = FTy->getReturnType();
5079 TargetLowering::ArgListTy Args;
5080 TargetLowering::ArgListEntry Entry;
5081 Args.reserve(CS.arg_size());
5083 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
5085 const Value *V = *i;
5088 if (V->getType()->isEmptyTy())
5091 SDValue ArgNode = getValue(V);
5092 Entry.Node = ArgNode; Entry.Ty = V->getType();
5094 // Skip the first return-type Attribute to get to params.
5095 Entry.setAttributes(&CS, i - CS.arg_begin() + 1);
5096 Args.push_back(Entry);
5098 // If we have an explicit sret argument that is an Instruction, (i.e., it
5099 // might point to function-local memory), we can't meaningfully tail-call.
5100 if (Entry.isSRet && isa<Instruction>(V))
5104 // Check if target-independent constraints permit a tail call here.
5105 // Target-dependent constraints are checked within TLI->LowerCallTo.
5106 if (isTailCall && !isInTailCallPosition(CS, DAG.getTarget()))
5109 TargetLowering::CallLoweringInfo CLI(DAG);
5110 CLI.setDebugLoc(getCurSDLoc()).setChain(getRoot())
5111 .setCallee(RetTy, FTy, Callee, std::move(Args), CS)
5112 .setTailCall(isTailCall);
5113 std::pair<SDValue,SDValue> Result = lowerInvokable(CLI, LandingPad);
5115 if (Result.first.getNode())
5116 setValue(CS.getInstruction(), Result.first);
5119 /// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the
5120 /// value is equal or not-equal to zero.
5121 static bool IsOnlyUsedInZeroEqualityComparison(const Value *V) {
5122 for (const User *U : V->users()) {
5123 if (const ICmpInst *IC = dyn_cast<ICmpInst>(U))
5124 if (IC->isEquality())
5125 if (const Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
5126 if (C->isNullValue())
5128 // Unknown instruction.
5134 static SDValue getMemCmpLoad(const Value *PtrVal, MVT LoadVT,
5136 SelectionDAGBuilder &Builder) {
5138 // Check to see if this load can be trivially constant folded, e.g. if the
5139 // input is from a string literal.
5140 if (const Constant *LoadInput = dyn_cast<Constant>(PtrVal)) {
5141 // Cast pointer to the type we really want to load.
5142 LoadInput = ConstantExpr::getBitCast(const_cast<Constant *>(LoadInput),
5143 PointerType::getUnqual(LoadTy));
5145 if (const Constant *LoadCst = ConstantFoldLoadFromConstPtr(
5146 const_cast<Constant *>(LoadInput), *Builder.DL))
5147 return Builder.getValue(LoadCst);
5150 // Otherwise, we have to emit the load. If the pointer is to unfoldable but
5151 // still constant memory, the input chain can be the entry node.
5153 bool ConstantMemory = false;
5155 // Do not serialize (non-volatile) loads of constant memory with anything.
5156 if (Builder.AA->pointsToConstantMemory(PtrVal)) {
5157 Root = Builder.DAG.getEntryNode();
5158 ConstantMemory = true;
5160 // Do not serialize non-volatile loads against each other.
5161 Root = Builder.DAG.getRoot();
5164 SDValue Ptr = Builder.getValue(PtrVal);
5165 SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurSDLoc(), Root,
5166 Ptr, MachinePointerInfo(PtrVal),
5168 false /*nontemporal*/,
5169 false /*isinvariant*/, 1 /* align=1 */);
5171 if (!ConstantMemory)
5172 Builder.PendingLoads.push_back(LoadVal.getValue(1));
5176 /// processIntegerCallValue - Record the value for an instruction that
5177 /// produces an integer result, converting the type where necessary.
5178 void SelectionDAGBuilder::processIntegerCallValue(const Instruction &I,
5181 EVT VT = DAG.getTargetLoweringInfo().getValueType(I.getType(), true);
5183 Value = DAG.getSExtOrTrunc(Value, getCurSDLoc(), VT);
5185 Value = DAG.getZExtOrTrunc(Value, getCurSDLoc(), VT);
5186 setValue(&I, Value);
5189 /// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form.
5190 /// If so, return true and lower it, otherwise return false and it will be
5191 /// lowered like a normal call.
5192 bool SelectionDAGBuilder::visitMemCmpCall(const CallInst &I) {
5193 // Verify that the prototype makes sense. int memcmp(void*,void*,size_t)
5194 if (I.getNumArgOperands() != 3)
5197 const Value *LHS = I.getArgOperand(0), *RHS = I.getArgOperand(1);
5198 if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() ||
5199 !I.getArgOperand(2)->getType()->isIntegerTy() ||
5200 !I.getType()->isIntegerTy())
5203 const Value *Size = I.getArgOperand(2);
5204 const ConstantInt *CSize = dyn_cast<ConstantInt>(Size);
5205 if (CSize && CSize->getZExtValue() == 0) {
5206 EVT CallVT = DAG.getTargetLoweringInfo().getValueType(I.getType(), true);
5207 setValue(&I, DAG.getConstant(0, getCurSDLoc(), CallVT));
5211 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5212 std::pair<SDValue, SDValue> Res =
5213 TSI.EmitTargetCodeForMemcmp(DAG, getCurSDLoc(), DAG.getRoot(),
5214 getValue(LHS), getValue(RHS), getValue(Size),
5215 MachinePointerInfo(LHS),
5216 MachinePointerInfo(RHS));
5217 if (Res.first.getNode()) {
5218 processIntegerCallValue(I, Res.first, true);
5219 PendingLoads.push_back(Res.second);
5223 // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0
5224 // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0
5225 if (CSize && IsOnlyUsedInZeroEqualityComparison(&I)) {
5226 bool ActuallyDoIt = true;
5229 switch (CSize->getZExtValue()) {
5231 LoadVT = MVT::Other;
5233 ActuallyDoIt = false;
5237 LoadTy = Type::getInt16Ty(CSize->getContext());
5241 LoadTy = Type::getInt32Ty(CSize->getContext());
5245 LoadTy = Type::getInt64Ty(CSize->getContext());
5249 LoadVT = MVT::v4i32;
5250 LoadTy = Type::getInt32Ty(CSize->getContext());
5251 LoadTy = VectorType::get(LoadTy, 4);
5256 // This turns into unaligned loads. We only do this if the target natively
5257 // supports the MVT we'll be loading or if it is small enough (<= 4) that
5258 // we'll only produce a small number of byte loads.
5260 // Require that we can find a legal MVT, and only do this if the target
5261 // supports unaligned loads of that type. Expanding into byte loads would
5263 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5264 if (ActuallyDoIt && CSize->getZExtValue() > 4) {
5265 unsigned DstAS = LHS->getType()->getPointerAddressSpace();
5266 unsigned SrcAS = RHS->getType()->getPointerAddressSpace();
5267 // TODO: Handle 5 byte compare as 4-byte + 1 byte.
5268 // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads.
5269 // TODO: Check alignment of src and dest ptrs.
5270 if (!TLI.isTypeLegal(LoadVT) ||
5271 !TLI.allowsMisalignedMemoryAccesses(LoadVT, SrcAS) ||
5272 !TLI.allowsMisalignedMemoryAccesses(LoadVT, DstAS))
5273 ActuallyDoIt = false;
5277 SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this);
5278 SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this);
5280 SDValue Res = DAG.getSetCC(getCurSDLoc(), MVT::i1, LHSVal, RHSVal,
5282 processIntegerCallValue(I, Res, false);
5291 /// visitMemChrCall -- See if we can lower a memchr call into an optimized
5292 /// form. If so, return true and lower it, otherwise return false and it
5293 /// will be lowered like a normal call.
5294 bool SelectionDAGBuilder::visitMemChrCall(const CallInst &I) {
5295 // Verify that the prototype makes sense. void *memchr(void *, int, size_t)
5296 if (I.getNumArgOperands() != 3)
5299 const Value *Src = I.getArgOperand(0);
5300 const Value *Char = I.getArgOperand(1);
5301 const Value *Length = I.getArgOperand(2);
5302 if (!Src->getType()->isPointerTy() ||
5303 !Char->getType()->isIntegerTy() ||
5304 !Length->getType()->isIntegerTy() ||
5305 !I.getType()->isPointerTy())
5308 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5309 std::pair<SDValue, SDValue> Res =
5310 TSI.EmitTargetCodeForMemchr(DAG, getCurSDLoc(), DAG.getRoot(),
5311 getValue(Src), getValue(Char), getValue(Length),
5312 MachinePointerInfo(Src));
5313 if (Res.first.getNode()) {
5314 setValue(&I, Res.first);
5315 PendingLoads.push_back(Res.second);
5322 /// visitStrCpyCall -- See if we can lower a strcpy or stpcpy call into an
5323 /// optimized form. If so, return true and lower it, otherwise return false
5324 /// and it will be lowered like a normal call.
5325 bool SelectionDAGBuilder::visitStrCpyCall(const CallInst &I, bool isStpcpy) {
5326 // Verify that the prototype makes sense. char *strcpy(char *, char *)
5327 if (I.getNumArgOperands() != 2)
5330 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5331 if (!Arg0->getType()->isPointerTy() ||
5332 !Arg1->getType()->isPointerTy() ||
5333 !I.getType()->isPointerTy())
5336 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5337 std::pair<SDValue, SDValue> Res =
5338 TSI.EmitTargetCodeForStrcpy(DAG, getCurSDLoc(), getRoot(),
5339 getValue(Arg0), getValue(Arg1),
5340 MachinePointerInfo(Arg0),
5341 MachinePointerInfo(Arg1), isStpcpy);
5342 if (Res.first.getNode()) {
5343 setValue(&I, Res.first);
5344 DAG.setRoot(Res.second);
5351 /// visitStrCmpCall - See if we can lower a call to strcmp in an optimized form.
5352 /// If so, return true and lower it, otherwise return false and it will be
5353 /// lowered like a normal call.
5354 bool SelectionDAGBuilder::visitStrCmpCall(const CallInst &I) {
5355 // Verify that the prototype makes sense. int strcmp(void*,void*)
5356 if (I.getNumArgOperands() != 2)
5359 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5360 if (!Arg0->getType()->isPointerTy() ||
5361 !Arg1->getType()->isPointerTy() ||
5362 !I.getType()->isIntegerTy())
5365 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5366 std::pair<SDValue, SDValue> Res =
5367 TSI.EmitTargetCodeForStrcmp(DAG, getCurSDLoc(), DAG.getRoot(),
5368 getValue(Arg0), getValue(Arg1),
5369 MachinePointerInfo(Arg0),
5370 MachinePointerInfo(Arg1));
5371 if (Res.first.getNode()) {
5372 processIntegerCallValue(I, Res.first, true);
5373 PendingLoads.push_back(Res.second);
5380 /// visitStrLenCall -- See if we can lower a strlen call into an optimized
5381 /// form. If so, return true and lower it, otherwise return false and it
5382 /// will be lowered like a normal call.
5383 bool SelectionDAGBuilder::visitStrLenCall(const CallInst &I) {
5384 // Verify that the prototype makes sense. size_t strlen(char *)
5385 if (I.getNumArgOperands() != 1)
5388 const Value *Arg0 = I.getArgOperand(0);
5389 if (!Arg0->getType()->isPointerTy() || !I.getType()->isIntegerTy())
5392 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5393 std::pair<SDValue, SDValue> Res =
5394 TSI.EmitTargetCodeForStrlen(DAG, getCurSDLoc(), DAG.getRoot(),
5395 getValue(Arg0), MachinePointerInfo(Arg0));
5396 if (Res.first.getNode()) {
5397 processIntegerCallValue(I, Res.first, false);
5398 PendingLoads.push_back(Res.second);
5405 /// visitStrNLenCall -- See if we can lower a strnlen call into an optimized
5406 /// form. If so, return true and lower it, otherwise return false and it
5407 /// will be lowered like a normal call.
5408 bool SelectionDAGBuilder::visitStrNLenCall(const CallInst &I) {
5409 // Verify that the prototype makes sense. size_t strnlen(char *, size_t)
5410 if (I.getNumArgOperands() != 2)
5413 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5414 if (!Arg0->getType()->isPointerTy() ||
5415 !Arg1->getType()->isIntegerTy() ||
5416 !I.getType()->isIntegerTy())
5419 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5420 std::pair<SDValue, SDValue> Res =
5421 TSI.EmitTargetCodeForStrnlen(DAG, getCurSDLoc(), DAG.getRoot(),
5422 getValue(Arg0), getValue(Arg1),
5423 MachinePointerInfo(Arg0));
5424 if (Res.first.getNode()) {
5425 processIntegerCallValue(I, Res.first, false);
5426 PendingLoads.push_back(Res.second);
5433 /// visitUnaryFloatCall - If a call instruction is a unary floating-point
5434 /// operation (as expected), translate it to an SDNode with the specified opcode
5435 /// and return true.
5436 bool SelectionDAGBuilder::visitUnaryFloatCall(const CallInst &I,
5438 // Sanity check that it really is a unary floating-point call.
5439 if (I.getNumArgOperands() != 1 ||
5440 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
5441 I.getType() != I.getArgOperand(0)->getType() ||
5442 !I.onlyReadsMemory())
5445 SDValue Tmp = getValue(I.getArgOperand(0));
5446 setValue(&I, DAG.getNode(Opcode, getCurSDLoc(), Tmp.getValueType(), Tmp));
5450 /// visitBinaryFloatCall - If a call instruction is a binary floating-point
5451 /// operation (as expected), translate it to an SDNode with the specified opcode
5452 /// and return true.
5453 bool SelectionDAGBuilder::visitBinaryFloatCall(const CallInst &I,
5455 // Sanity check that it really is a binary floating-point call.
5456 if (I.getNumArgOperands() != 2 ||
5457 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
5458 I.getType() != I.getArgOperand(0)->getType() ||
5459 I.getType() != I.getArgOperand(1)->getType() ||
5460 !I.onlyReadsMemory())
5463 SDValue Tmp0 = getValue(I.getArgOperand(0));
5464 SDValue Tmp1 = getValue(I.getArgOperand(1));
5465 EVT VT = Tmp0.getValueType();
5466 setValue(&I, DAG.getNode(Opcode, getCurSDLoc(), VT, Tmp0, Tmp1));
5470 void SelectionDAGBuilder::visitCall(const CallInst &I) {
5471 // Handle inline assembly differently.
5472 if (isa<InlineAsm>(I.getCalledValue())) {
5477 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5478 ComputeUsesVAFloatArgument(I, &MMI);
5480 const char *RenameFn = nullptr;
5481 if (Function *F = I.getCalledFunction()) {
5482 if (F->isDeclaration()) {
5483 if (const TargetIntrinsicInfo *II = TM.getIntrinsicInfo()) {
5484 if (unsigned IID = II->getIntrinsicID(F)) {
5485 RenameFn = visitIntrinsicCall(I, IID);
5490 if (Intrinsic::ID IID = F->getIntrinsicID()) {
5491 RenameFn = visitIntrinsicCall(I, IID);
5497 // Check for well-known libc/libm calls. If the function is internal, it
5498 // can't be a library call.
5500 if (!F->hasLocalLinkage() && F->hasName() &&
5501 LibInfo->getLibFunc(F->getName(), Func) &&
5502 LibInfo->hasOptimizedCodeGen(Func)) {
5505 case LibFunc::copysign:
5506 case LibFunc::copysignf:
5507 case LibFunc::copysignl:
5508 if (I.getNumArgOperands() == 2 && // Basic sanity checks.
5509 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5510 I.getType() == I.getArgOperand(0)->getType() &&
5511 I.getType() == I.getArgOperand(1)->getType() &&
5512 I.onlyReadsMemory()) {
5513 SDValue LHS = getValue(I.getArgOperand(0));
5514 SDValue RHS = getValue(I.getArgOperand(1));
5515 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurSDLoc(),
5516 LHS.getValueType(), LHS, RHS));
5521 case LibFunc::fabsf:
5522 case LibFunc::fabsl:
5523 if (visitUnaryFloatCall(I, ISD::FABS))
5527 case LibFunc::fminf:
5528 case LibFunc::fminl:
5529 if (visitBinaryFloatCall(I, ISD::FMINNUM))
5533 case LibFunc::fmaxf:
5534 case LibFunc::fmaxl:
5535 if (visitBinaryFloatCall(I, ISD::FMAXNUM))
5541 if (visitUnaryFloatCall(I, ISD::FSIN))
5547 if (visitUnaryFloatCall(I, ISD::FCOS))
5551 case LibFunc::sqrtf:
5552 case LibFunc::sqrtl:
5553 case LibFunc::sqrt_finite:
5554 case LibFunc::sqrtf_finite:
5555 case LibFunc::sqrtl_finite:
5556 if (visitUnaryFloatCall(I, ISD::FSQRT))
5559 case LibFunc::floor:
5560 case LibFunc::floorf:
5561 case LibFunc::floorl:
5562 if (visitUnaryFloatCall(I, ISD::FFLOOR))
5565 case LibFunc::nearbyint:
5566 case LibFunc::nearbyintf:
5567 case LibFunc::nearbyintl:
5568 if (visitUnaryFloatCall(I, ISD::FNEARBYINT))
5572 case LibFunc::ceilf:
5573 case LibFunc::ceill:
5574 if (visitUnaryFloatCall(I, ISD::FCEIL))
5578 case LibFunc::rintf:
5579 case LibFunc::rintl:
5580 if (visitUnaryFloatCall(I, ISD::FRINT))
5583 case LibFunc::round:
5584 case LibFunc::roundf:
5585 case LibFunc::roundl:
5586 if (visitUnaryFloatCall(I, ISD::FROUND))
5589 case LibFunc::trunc:
5590 case LibFunc::truncf:
5591 case LibFunc::truncl:
5592 if (visitUnaryFloatCall(I, ISD::FTRUNC))
5596 case LibFunc::log2f:
5597 case LibFunc::log2l:
5598 if (visitUnaryFloatCall(I, ISD::FLOG2))
5602 case LibFunc::exp2f:
5603 case LibFunc::exp2l:
5604 if (visitUnaryFloatCall(I, ISD::FEXP2))
5607 case LibFunc::memcmp:
5608 if (visitMemCmpCall(I))
5611 case LibFunc::memchr:
5612 if (visitMemChrCall(I))
5615 case LibFunc::strcpy:
5616 if (visitStrCpyCall(I, false))
5619 case LibFunc::stpcpy:
5620 if (visitStrCpyCall(I, true))
5623 case LibFunc::strcmp:
5624 if (visitStrCmpCall(I))
5627 case LibFunc::strlen:
5628 if (visitStrLenCall(I))
5631 case LibFunc::strnlen:
5632 if (visitStrNLenCall(I))
5641 Callee = getValue(I.getCalledValue());
5643 Callee = DAG.getExternalSymbol(RenameFn,
5644 DAG.getTargetLoweringInfo().getPointerTy());
5646 // Check if we can potentially perform a tail call. More detailed checking is
5647 // be done within LowerCallTo, after more information about the call is known.
5648 LowerCallTo(&I, Callee, I.isTailCall());
5653 /// AsmOperandInfo - This contains information for each constraint that we are
5655 class SDISelAsmOperandInfo : public TargetLowering::AsmOperandInfo {
5657 /// CallOperand - If this is the result output operand or a clobber
5658 /// this is null, otherwise it is the incoming operand to the CallInst.
5659 /// This gets modified as the asm is processed.
5660 SDValue CallOperand;
5662 /// AssignedRegs - If this is a register or register class operand, this
5663 /// contains the set of register corresponding to the operand.
5664 RegsForValue AssignedRegs;
5666 explicit SDISelAsmOperandInfo(const TargetLowering::AsmOperandInfo &info)
5667 : TargetLowering::AsmOperandInfo(info), CallOperand(nullptr,0) {
5670 /// getCallOperandValEVT - Return the EVT of the Value* that this operand
5671 /// corresponds to. If there is no Value* for this operand, it returns
5673 EVT getCallOperandValEVT(LLVMContext &Context,
5674 const TargetLowering &TLI,
5675 const DataLayout *DL) const {
5676 if (!CallOperandVal) return MVT::Other;
5678 if (isa<BasicBlock>(CallOperandVal))
5679 return TLI.getPointerTy();
5681 llvm::Type *OpTy = CallOperandVal->getType();
5683 // FIXME: code duplicated from TargetLowering::ParseConstraints().
5684 // If this is an indirect operand, the operand is a pointer to the
5687 llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
5689 report_fatal_error("Indirect operand for inline asm not a pointer!");
5690 OpTy = PtrTy->getElementType();
5693 // Look for vector wrapped in a struct. e.g. { <16 x i8> }.
5694 if (StructType *STy = dyn_cast<StructType>(OpTy))
5695 if (STy->getNumElements() == 1)
5696 OpTy = STy->getElementType(0);
5698 // If OpTy is not a single value, it may be a struct/union that we
5699 // can tile with integers.
5700 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
5701 unsigned BitSize = DL->getTypeSizeInBits(OpTy);
5710 OpTy = IntegerType::get(Context, BitSize);
5715 return TLI.getValueType(OpTy, true);
5719 typedef SmallVector<SDISelAsmOperandInfo,16> SDISelAsmOperandInfoVector;
5721 } // end anonymous namespace
5723 /// GetRegistersForValue - Assign registers (virtual or physical) for the
5724 /// specified operand. We prefer to assign virtual registers, to allow the
5725 /// register allocator to handle the assignment process. However, if the asm
5726 /// uses features that we can't model on machineinstrs, we have SDISel do the
5727 /// allocation. This produces generally horrible, but correct, code.
5729 /// OpInfo describes the operand.
5731 static void GetRegistersForValue(SelectionDAG &DAG,
5732 const TargetLowering &TLI,
5734 SDISelAsmOperandInfo &OpInfo) {
5735 LLVMContext &Context = *DAG.getContext();
5737 MachineFunction &MF = DAG.getMachineFunction();
5738 SmallVector<unsigned, 4> Regs;
5740 // If this is a constraint for a single physreg, or a constraint for a
5741 // register class, find it.
5742 std::pair<unsigned, const TargetRegisterClass *> PhysReg =
5743 TLI.getRegForInlineAsmConstraint(MF.getSubtarget().getRegisterInfo(),
5744 OpInfo.ConstraintCode,
5745 OpInfo.ConstraintVT);
5747 unsigned NumRegs = 1;
5748 if (OpInfo.ConstraintVT != MVT::Other) {
5749 // If this is a FP input in an integer register (or visa versa) insert a bit
5750 // cast of the input value. More generally, handle any case where the input
5751 // value disagrees with the register class we plan to stick this in.
5752 if (OpInfo.Type == InlineAsm::isInput &&
5753 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) {
5754 // Try to convert to the first EVT that the reg class contains. If the
5755 // types are identical size, use a bitcast to convert (e.g. two differing
5757 MVT RegVT = *PhysReg.second->vt_begin();
5758 if (RegVT.getSizeInBits() == OpInfo.CallOperand.getValueSizeInBits()) {
5759 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
5760 RegVT, OpInfo.CallOperand);
5761 OpInfo.ConstraintVT = RegVT;
5762 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) {
5763 // If the input is a FP value and we want it in FP registers, do a
5764 // bitcast to the corresponding integer type. This turns an f64 value
5765 // into i64, which can be passed with two i32 values on a 32-bit
5767 RegVT = MVT::getIntegerVT(OpInfo.ConstraintVT.getSizeInBits());
5768 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
5769 RegVT, OpInfo.CallOperand);
5770 OpInfo.ConstraintVT = RegVT;
5774 NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT);
5778 EVT ValueVT = OpInfo.ConstraintVT;
5780 // If this is a constraint for a specific physical register, like {r17},
5782 if (unsigned AssignedReg = PhysReg.first) {
5783 const TargetRegisterClass *RC = PhysReg.second;
5784 if (OpInfo.ConstraintVT == MVT::Other)
5785 ValueVT = *RC->vt_begin();
5787 // Get the actual register value type. This is important, because the user
5788 // may have asked for (e.g.) the AX register in i32 type. We need to
5789 // remember that AX is actually i16 to get the right extension.
5790 RegVT = *RC->vt_begin();
5792 // This is a explicit reference to a physical register.
5793 Regs.push_back(AssignedReg);
5795 // If this is an expanded reference, add the rest of the regs to Regs.
5797 TargetRegisterClass::iterator I = RC->begin();
5798 for (; *I != AssignedReg; ++I)
5799 assert(I != RC->end() && "Didn't find reg!");
5801 // Already added the first reg.
5803 for (; NumRegs; --NumRegs, ++I) {
5804 assert(I != RC->end() && "Ran out of registers to allocate!");
5809 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5813 // Otherwise, if this was a reference to an LLVM register class, create vregs
5814 // for this reference.
5815 if (const TargetRegisterClass *RC = PhysReg.second) {
5816 RegVT = *RC->vt_begin();
5817 if (OpInfo.ConstraintVT == MVT::Other)
5820 // Create the appropriate number of virtual registers.
5821 MachineRegisterInfo &RegInfo = MF.getRegInfo();
5822 for (; NumRegs; --NumRegs)
5823 Regs.push_back(RegInfo.createVirtualRegister(RC));
5825 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5829 // Otherwise, we couldn't allocate enough registers for this.
5832 /// visitInlineAsm - Handle a call to an InlineAsm object.
5834 void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) {
5835 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
5837 /// ConstraintOperands - Information about all of the constraints.
5838 SDISelAsmOperandInfoVector ConstraintOperands;
5840 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5841 TargetLowering::AsmOperandInfoVector TargetConstraints =
5842 TLI.ParseConstraints(DAG.getSubtarget().getRegisterInfo(), CS);
5844 bool hasMemory = false;
5846 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
5847 unsigned ResNo = 0; // ResNo - The result number of the next output.
5848 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
5849 ConstraintOperands.push_back(SDISelAsmOperandInfo(TargetConstraints[i]));
5850 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
5852 MVT OpVT = MVT::Other;
5854 // Compute the value type for each operand.
5855 switch (OpInfo.Type) {
5856 case InlineAsm::isOutput:
5857 // Indirect outputs just consume an argument.
5858 if (OpInfo.isIndirect) {
5859 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5863 // The return value of the call is this value. As such, there is no
5864 // corresponding argument.
5865 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
5866 if (StructType *STy = dyn_cast<StructType>(CS.getType())) {
5867 OpVT = TLI.getSimpleValueType(STy->getElementType(ResNo));
5869 assert(ResNo == 0 && "Asm only has one result!");
5870 OpVT = TLI.getSimpleValueType(CS.getType());
5874 case InlineAsm::isInput:
5875 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5877 case InlineAsm::isClobber:
5882 // If this is an input or an indirect output, process the call argument.
5883 // BasicBlocks are labels, currently appearing only in asm's.
5884 if (OpInfo.CallOperandVal) {
5885 if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) {
5886 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]);
5888 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
5892 OpInfo.getCallOperandValEVT(*DAG.getContext(), TLI, DL).getSimpleVT();
5895 OpInfo.ConstraintVT = OpVT;
5897 // Indirect operand accesses access memory.
5898 if (OpInfo.isIndirect)
5901 for (unsigned j = 0, ee = OpInfo.Codes.size(); j != ee; ++j) {
5902 TargetLowering::ConstraintType
5903 CType = TLI.getConstraintType(OpInfo.Codes[j]);
5904 if (CType == TargetLowering::C_Memory) {
5912 SDValue Chain, Flag;
5914 // We won't need to flush pending loads if this asm doesn't touch
5915 // memory and is nonvolatile.
5916 if (hasMemory || IA->hasSideEffects())
5919 Chain = DAG.getRoot();
5921 // Second pass over the constraints: compute which constraint option to use
5922 // and assign registers to constraints that want a specific physreg.
5923 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5924 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5926 // If this is an output operand with a matching input operand, look up the
5927 // matching input. If their types mismatch, e.g. one is an integer, the
5928 // other is floating point, or their sizes are different, flag it as an
5930 if (OpInfo.hasMatchingInput()) {
5931 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
5933 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
5934 const TargetRegisterInfo *TRI = DAG.getSubtarget().getRegisterInfo();
5935 std::pair<unsigned, const TargetRegisterClass *> MatchRC =
5936 TLI.getRegForInlineAsmConstraint(TRI, OpInfo.ConstraintCode,
5937 OpInfo.ConstraintVT);
5938 std::pair<unsigned, const TargetRegisterClass *> InputRC =
5939 TLI.getRegForInlineAsmConstraint(TRI, Input.ConstraintCode,
5940 Input.ConstraintVT);
5941 if ((OpInfo.ConstraintVT.isInteger() !=
5942 Input.ConstraintVT.isInteger()) ||
5943 (MatchRC.second != InputRC.second)) {
5944 report_fatal_error("Unsupported asm: input constraint"
5945 " with a matching output constraint of"
5946 " incompatible type!");
5948 Input.ConstraintVT = OpInfo.ConstraintVT;
5952 // Compute the constraint code and ConstraintType to use.
5953 TLI.ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, &DAG);
5955 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5956 OpInfo.Type == InlineAsm::isClobber)
5959 // If this is a memory input, and if the operand is not indirect, do what we
5960 // need to to provide an address for the memory input.
5961 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5962 !OpInfo.isIndirect) {
5963 assert((OpInfo.isMultipleAlternative ||
5964 (OpInfo.Type == InlineAsm::isInput)) &&
5965 "Can only indirectify direct input operands!");
5967 // Memory operands really want the address of the value. If we don't have
5968 // an indirect input, put it in the constpool if we can, otherwise spill
5969 // it to a stack slot.
5970 // TODO: This isn't quite right. We need to handle these according to
5971 // the addressing mode that the constraint wants. Also, this may take
5972 // an additional register for the computation and we don't want that
5975 // If the operand is a float, integer, or vector constant, spill to a
5976 // constant pool entry to get its address.
5977 const Value *OpVal = OpInfo.CallOperandVal;
5978 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
5979 isa<ConstantVector>(OpVal) || isa<ConstantDataVector>(OpVal)) {
5980 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal),
5981 TLI.getPointerTy());
5983 // Otherwise, create a stack slot and emit a store to it before the
5985 Type *Ty = OpVal->getType();
5986 uint64_t TySize = TLI.getDataLayout()->getTypeAllocSize(Ty);
5987 unsigned Align = TLI.getDataLayout()->getPrefTypeAlignment(Ty);
5988 MachineFunction &MF = DAG.getMachineFunction();
5989 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
5990 SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
5991 Chain = DAG.getStore(Chain, getCurSDLoc(),
5992 OpInfo.CallOperand, StackSlot,
5993 MachinePointerInfo::getFixedStack(SSFI),
5995 OpInfo.CallOperand = StackSlot;
5998 // There is no longer a Value* corresponding to this operand.
5999 OpInfo.CallOperandVal = nullptr;
6001 // It is now an indirect operand.
6002 OpInfo.isIndirect = true;
6005 // If this constraint is for a specific register, allocate it before
6007 if (OpInfo.ConstraintType == TargetLowering::C_Register)
6008 GetRegistersForValue(DAG, TLI, getCurSDLoc(), OpInfo);
6011 // Second pass - Loop over all of the operands, assigning virtual or physregs
6012 // to register class operands.
6013 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6014 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6016 // C_Register operands have already been allocated, Other/Memory don't need
6018 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
6019 GetRegistersForValue(DAG, TLI, getCurSDLoc(), OpInfo);
6022 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
6023 std::vector<SDValue> AsmNodeOperands;
6024 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain
6025 AsmNodeOperands.push_back(
6026 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(),
6027 TLI.getPointerTy()));
6029 // If we have a !srcloc metadata node associated with it, we want to attach
6030 // this to the ultimately generated inline asm machineinstr. To do this, we
6031 // pass in the third operand as this (potentially null) inline asm MDNode.
6032 const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc");
6033 AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc));
6035 // Remember the HasSideEffect, AlignStack, AsmDialect, MayLoad and MayStore
6036 // bits as operand 3.
6037 unsigned ExtraInfo = 0;
6038 if (IA->hasSideEffects())
6039 ExtraInfo |= InlineAsm::Extra_HasSideEffects;
6040 if (IA->isAlignStack())
6041 ExtraInfo |= InlineAsm::Extra_IsAlignStack;
6042 // Set the asm dialect.
6043 ExtraInfo |= IA->getDialect() * InlineAsm::Extra_AsmDialect;
6045 // Determine if this InlineAsm MayLoad or MayStore based on the constraints.
6046 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
6047 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
6049 // Compute the constraint code and ConstraintType to use.
6050 TLI.ComputeConstraintToUse(OpInfo, SDValue());
6052 // Ideally, we would only check against memory constraints. However, the
6053 // meaning of an other constraint can be target-specific and we can't easily
6054 // reason about it. Therefore, be conservative and set MayLoad/MayStore
6055 // for other constriants as well.
6056 if (OpInfo.ConstraintType == TargetLowering::C_Memory ||
6057 OpInfo.ConstraintType == TargetLowering::C_Other) {
6058 if (OpInfo.Type == InlineAsm::isInput)
6059 ExtraInfo |= InlineAsm::Extra_MayLoad;
6060 else if (OpInfo.Type == InlineAsm::isOutput)
6061 ExtraInfo |= InlineAsm::Extra_MayStore;
6062 else if (OpInfo.Type == InlineAsm::isClobber)
6063 ExtraInfo |= (InlineAsm::Extra_MayLoad | InlineAsm::Extra_MayStore);
6067 AsmNodeOperands.push_back(DAG.getTargetConstant(ExtraInfo, getCurSDLoc(),
6068 TLI.getPointerTy()));
6070 // Loop over all of the inputs, copying the operand values into the
6071 // appropriate registers and processing the output regs.
6072 RegsForValue RetValRegs;
6074 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
6075 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
6077 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6078 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6080 switch (OpInfo.Type) {
6081 case InlineAsm::isOutput: {
6082 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
6083 OpInfo.ConstraintType != TargetLowering::C_Register) {
6084 // Memory output, or 'other' output (e.g. 'X' constraint).
6085 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
6087 unsigned ConstraintID =
6088 TLI.getInlineAsmMemConstraint(OpInfo.ConstraintCode);
6089 assert(ConstraintID != InlineAsm::Constraint_Unknown &&
6090 "Failed to convert memory constraint code to constraint id.");
6092 // Add information to the INLINEASM node to know about this output.
6093 unsigned OpFlags = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
6094 OpFlags = InlineAsm::getFlagWordForMem(OpFlags, ConstraintID);
6095 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlags, getCurSDLoc(),
6097 AsmNodeOperands.push_back(OpInfo.CallOperand);
6101 // Otherwise, this is a register or register class output.
6103 // Copy the output from the appropriate register. Find a register that
6105 if (OpInfo.AssignedRegs.Regs.empty()) {
6106 LLVMContext &Ctx = *DAG.getContext();
6107 Ctx.emitError(CS.getInstruction(),
6108 "couldn't allocate output register for constraint '" +
6109 Twine(OpInfo.ConstraintCode) + "'");
6113 // If this is an indirect operand, store through the pointer after the
6115 if (OpInfo.isIndirect) {
6116 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
6117 OpInfo.CallOperandVal));
6119 // This is the result value of the call.
6120 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
6121 // Concatenate this output onto the outputs list.
6122 RetValRegs.append(OpInfo.AssignedRegs);
6125 // Add information to the INLINEASM node to know that this register is
6128 .AddInlineAsmOperands(OpInfo.isEarlyClobber
6129 ? InlineAsm::Kind_RegDefEarlyClobber
6130 : InlineAsm::Kind_RegDef,
6131 false, 0, getCurSDLoc(), DAG, AsmNodeOperands);
6134 case InlineAsm::isInput: {
6135 SDValue InOperandVal = OpInfo.CallOperand;
6137 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint?
6138 // If this is required to match an output register we have already set,
6139 // just use its register.
6140 unsigned OperandNo = OpInfo.getMatchedOperand();
6142 // Scan until we find the definition we already emitted of this operand.
6143 // When we find it, create a RegsForValue operand.
6144 unsigned CurOp = InlineAsm::Op_FirstOperand;
6145 for (; OperandNo; --OperandNo) {
6146 // Advance to the next operand.
6148 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6149 assert((InlineAsm::isRegDefKind(OpFlag) ||
6150 InlineAsm::isRegDefEarlyClobberKind(OpFlag) ||
6151 InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?");
6152 CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1;
6156 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6157 if (InlineAsm::isRegDefKind(OpFlag) ||
6158 InlineAsm::isRegDefEarlyClobberKind(OpFlag)) {
6159 // Add (OpFlag&0xffff)>>3 registers to MatchedRegs.
6160 if (OpInfo.isIndirect) {
6161 // This happens on gcc/testsuite/gcc.dg/pr8788-1.c
6162 LLVMContext &Ctx = *DAG.getContext();
6163 Ctx.emitError(CS.getInstruction(), "inline asm not supported yet:"
6164 " don't know how to handle tied "
6165 "indirect register inputs");
6169 RegsForValue MatchedRegs;
6170 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
6171 MVT RegVT = AsmNodeOperands[CurOp+1].getSimpleValueType();
6172 MatchedRegs.RegVTs.push_back(RegVT);
6173 MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo();
6174 for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag);
6176 if (const TargetRegisterClass *RC = TLI.getRegClassFor(RegVT))
6177 MatchedRegs.Regs.push_back(RegInfo.createVirtualRegister(RC));
6179 LLVMContext &Ctx = *DAG.getContext();
6180 Ctx.emitError(CS.getInstruction(),
6181 "inline asm error: This value"
6182 " type register class is not natively supported!");
6186 SDLoc dl = getCurSDLoc();
6187 // Use the produced MatchedRegs object to
6188 MatchedRegs.getCopyToRegs(InOperandVal, DAG, dl,
6189 Chain, &Flag, CS.getInstruction());
6190 MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse,
6191 true, OpInfo.getMatchedOperand(), dl,
6192 DAG, AsmNodeOperands);
6196 assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!");
6197 assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 &&
6198 "Unexpected number of operands");
6199 // Add information to the INLINEASM node to know about this input.
6200 // See InlineAsm.h isUseOperandTiedToDef.
6201 OpFlag = InlineAsm::convertMemFlagWordToMatchingFlagWord(OpFlag);
6202 OpFlag = InlineAsm::getFlagWordForMatchingOp(OpFlag,
6203 OpInfo.getMatchedOperand());
6204 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlag, getCurSDLoc(),
6205 TLI.getPointerTy()));
6206 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]);
6210 // Treat indirect 'X' constraint as memory.
6211 if (OpInfo.ConstraintType == TargetLowering::C_Other &&
6213 OpInfo.ConstraintType = TargetLowering::C_Memory;
6215 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
6216 std::vector<SDValue> Ops;
6217 TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode,
6220 LLVMContext &Ctx = *DAG.getContext();
6221 Ctx.emitError(CS.getInstruction(),
6222 "invalid operand for inline asm constraint '" +
6223 Twine(OpInfo.ConstraintCode) + "'");
6227 // Add information to the INLINEASM node to know about this input.
6228 unsigned ResOpType =
6229 InlineAsm::getFlagWord(InlineAsm::Kind_Imm, Ops.size());
6230 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
6232 TLI.getPointerTy()));
6233 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
6237 if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
6238 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
6239 assert(InOperandVal.getValueType() == TLI.getPointerTy() &&
6240 "Memory operands expect pointer values");
6242 unsigned ConstraintID =
6243 TLI.getInlineAsmMemConstraint(OpInfo.ConstraintCode);
6244 assert(ConstraintID != InlineAsm::Constraint_Unknown &&
6245 "Failed to convert memory constraint code to constraint id.");
6247 // Add information to the INLINEASM node to know about this input.
6248 unsigned ResOpType = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
6249 ResOpType = InlineAsm::getFlagWordForMem(ResOpType, ConstraintID);
6250 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
6253 AsmNodeOperands.push_back(InOperandVal);
6257 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
6258 OpInfo.ConstraintType == TargetLowering::C_Register) &&
6259 "Unknown constraint type!");
6261 // TODO: Support this.
6262 if (OpInfo.isIndirect) {
6263 LLVMContext &Ctx = *DAG.getContext();
6264 Ctx.emitError(CS.getInstruction(),
6265 "Don't know how to handle indirect register inputs yet "
6266 "for constraint '" +
6267 Twine(OpInfo.ConstraintCode) + "'");
6271 // Copy the input into the appropriate registers.
6272 if (OpInfo.AssignedRegs.Regs.empty()) {
6273 LLVMContext &Ctx = *DAG.getContext();
6274 Ctx.emitError(CS.getInstruction(),
6275 "couldn't allocate input reg for constraint '" +
6276 Twine(OpInfo.ConstraintCode) + "'");
6280 SDLoc dl = getCurSDLoc();
6282 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, dl,
6283 Chain, &Flag, CS.getInstruction());
6285 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, false, 0,
6286 dl, DAG, AsmNodeOperands);
6289 case InlineAsm::isClobber: {
6290 // Add the clobbered value to the operand list, so that the register
6291 // allocator is aware that the physreg got clobbered.
6292 if (!OpInfo.AssignedRegs.Regs.empty())
6293 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_Clobber,
6294 false, 0, getCurSDLoc(), DAG,
6301 // Finish up input operands. Set the input chain and add the flag last.
6302 AsmNodeOperands[InlineAsm::Op_InputChain] = Chain;
6303 if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
6305 Chain = DAG.getNode(ISD::INLINEASM, getCurSDLoc(),
6306 DAG.getVTList(MVT::Other, MVT::Glue), AsmNodeOperands);
6307 Flag = Chain.getValue(1);
6309 // If this asm returns a register value, copy the result from that register
6310 // and set it as the value of the call.
6311 if (!RetValRegs.Regs.empty()) {
6312 SDValue Val = RetValRegs.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(),
6313 Chain, &Flag, CS.getInstruction());
6315 // FIXME: Why don't we do this for inline asms with MRVs?
6316 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) {
6317 EVT ResultType = TLI.getValueType(CS.getType());
6319 // If any of the results of the inline asm is a vector, it may have the
6320 // wrong width/num elts. This can happen for register classes that can
6321 // contain multiple different value types. The preg or vreg allocated may
6322 // not have the same VT as was expected. Convert it to the right type
6323 // with bit_convert.
6324 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) {
6325 Val = DAG.getNode(ISD::BITCAST, getCurSDLoc(),
6328 } else if (ResultType != Val.getValueType() &&
6329 ResultType.isInteger() && Val.getValueType().isInteger()) {
6330 // If a result value was tied to an input value, the computed result may
6331 // have a wider width than the expected result. Extract the relevant
6333 Val = DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), ResultType, Val);
6336 assert(ResultType == Val.getValueType() && "Asm result value mismatch!");
6339 setValue(CS.getInstruction(), Val);
6340 // Don't need to use this as a chain in this case.
6341 if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty())
6345 std::vector<std::pair<SDValue, const Value *> > StoresToEmit;
6347 // Process indirect outputs, first output all of the flagged copies out of
6349 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
6350 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
6351 const Value *Ptr = IndirectStoresToEmit[i].second;
6352 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(),
6354 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
6357 // Emit the non-flagged stores from the physregs.
6358 SmallVector<SDValue, 8> OutChains;
6359 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) {
6360 SDValue Val = DAG.getStore(Chain, getCurSDLoc(),
6361 StoresToEmit[i].first,
6362 getValue(StoresToEmit[i].second),
6363 MachinePointerInfo(StoresToEmit[i].second),
6365 OutChains.push_back(Val);
6368 if (!OutChains.empty())
6369 Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other, OutChains);
6374 void SelectionDAGBuilder::visitVAStart(const CallInst &I) {
6375 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurSDLoc(),
6376 MVT::Other, getRoot(),
6377 getValue(I.getArgOperand(0)),
6378 DAG.getSrcValue(I.getArgOperand(0))));
6381 void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) {
6382 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6383 const DataLayout &DL = *TLI.getDataLayout();
6384 SDValue V = DAG.getVAArg(TLI.getValueType(I.getType()), getCurSDLoc(),
6385 getRoot(), getValue(I.getOperand(0)),
6386 DAG.getSrcValue(I.getOperand(0)),
6387 DL.getABITypeAlignment(I.getType()));
6389 DAG.setRoot(V.getValue(1));
6392 void SelectionDAGBuilder::visitVAEnd(const CallInst &I) {
6393 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurSDLoc(),
6394 MVT::Other, getRoot(),
6395 getValue(I.getArgOperand(0)),
6396 DAG.getSrcValue(I.getArgOperand(0))));
6399 void SelectionDAGBuilder::visitVACopy(const CallInst &I) {
6400 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurSDLoc(),
6401 MVT::Other, getRoot(),
6402 getValue(I.getArgOperand(0)),
6403 getValue(I.getArgOperand(1)),
6404 DAG.getSrcValue(I.getArgOperand(0)),
6405 DAG.getSrcValue(I.getArgOperand(1))));
6408 /// \brief Lower an argument list according to the target calling convention.
6410 /// \return A tuple of <return-value, token-chain>
6412 /// This is a helper for lowering intrinsics that follow a target calling
6413 /// convention or require stack pointer adjustment. Only a subset of the
6414 /// intrinsic's operands need to participate in the calling convention.
6415 std::pair<SDValue, SDValue>
6416 SelectionDAGBuilder::lowerCallOperands(ImmutableCallSite CS, unsigned ArgIdx,
6417 unsigned NumArgs, SDValue Callee,
6419 MachineBasicBlock *LandingPad,
6420 bool IsPatchPoint) {
6421 TargetLowering::ArgListTy Args;
6422 Args.reserve(NumArgs);
6424 // Populate the argument list.
6425 // Attributes for args start at offset 1, after the return attribute.
6426 for (unsigned ArgI = ArgIdx, ArgE = ArgIdx + NumArgs, AttrI = ArgIdx + 1;
6427 ArgI != ArgE; ++ArgI) {
6428 const Value *V = CS->getOperand(ArgI);
6430 assert(!V->getType()->isEmptyTy() && "Empty type passed to intrinsic.");
6432 TargetLowering::ArgListEntry Entry;
6433 Entry.Node = getValue(V);
6434 Entry.Ty = V->getType();
6435 Entry.setAttributes(&CS, AttrI);
6436 Args.push_back(Entry);
6439 TargetLowering::CallLoweringInfo CLI(DAG);
6440 CLI.setDebugLoc(getCurSDLoc()).setChain(getRoot())
6441 .setCallee(CS.getCallingConv(), ReturnTy, Callee, std::move(Args), NumArgs)
6442 .setDiscardResult(CS->use_empty()).setIsPatchPoint(IsPatchPoint);
6444 return lowerInvokable(CLI, LandingPad);
6447 /// \brief Add a stack map intrinsic call's live variable operands to a stackmap
6448 /// or patchpoint target node's operand list.
6450 /// Constants are converted to TargetConstants purely as an optimization to
6451 /// avoid constant materialization and register allocation.
6453 /// FrameIndex operands are converted to TargetFrameIndex so that ISEL does not
6454 /// generate addess computation nodes, and so ExpandISelPseudo can convert the
6455 /// TargetFrameIndex into a DirectMemRefOp StackMap location. This avoids
6456 /// address materialization and register allocation, but may also be required
6457 /// for correctness. If a StackMap (or PatchPoint) intrinsic directly uses an
6458 /// alloca in the entry block, then the runtime may assume that the alloca's
6459 /// StackMap location can be read immediately after compilation and that the
6460 /// location is valid at any point during execution (this is similar to the
6461 /// assumption made by the llvm.gcroot intrinsic). If the alloca's location were
6462 /// only available in a register, then the runtime would need to trap when
6463 /// execution reaches the StackMap in order to read the alloca's location.
6464 static void addStackMapLiveVars(ImmutableCallSite CS, unsigned StartIdx,
6465 SDLoc DL, SmallVectorImpl<SDValue> &Ops,
6466 SelectionDAGBuilder &Builder) {
6467 for (unsigned i = StartIdx, e = CS.arg_size(); i != e; ++i) {
6468 SDValue OpVal = Builder.getValue(CS.getArgument(i));
6469 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(OpVal)) {
6471 Builder.DAG.getTargetConstant(StackMaps::ConstantOp, DL, MVT::i64));
6473 Builder.DAG.getTargetConstant(C->getSExtValue(), DL, MVT::i64));
6474 } else if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(OpVal)) {
6475 const TargetLowering &TLI = Builder.DAG.getTargetLoweringInfo();
6477 Builder.DAG.getTargetFrameIndex(FI->getIndex(), TLI.getPointerTy()));
6479 Ops.push_back(OpVal);
6483 /// \brief Lower llvm.experimental.stackmap directly to its target opcode.
6484 void SelectionDAGBuilder::visitStackmap(const CallInst &CI) {
6485 // void @llvm.experimental.stackmap(i32 <id>, i32 <numShadowBytes>,
6486 // [live variables...])
6488 assert(CI.getType()->isVoidTy() && "Stackmap cannot return a value.");
6490 SDValue Chain, InFlag, Callee, NullPtr;
6491 SmallVector<SDValue, 32> Ops;
6493 SDLoc DL = getCurSDLoc();
6494 Callee = getValue(CI.getCalledValue());
6495 NullPtr = DAG.getIntPtrConstant(0, DL, true);
6497 // The stackmap intrinsic only records the live variables (the arguemnts
6498 // passed to it) and emits NOPS (if requested). Unlike the patchpoint
6499 // intrinsic, this won't be lowered to a function call. This means we don't
6500 // have to worry about calling conventions and target specific lowering code.
6501 // Instead we perform the call lowering right here.
6503 // chain, flag = CALLSEQ_START(chain, 0)
6504 // chain, flag = STACKMAP(id, nbytes, ..., chain, flag)
6505 // chain, flag = CALLSEQ_END(chain, 0, 0, flag)
6507 Chain = DAG.getCALLSEQ_START(getRoot(), NullPtr, DL);
6508 InFlag = Chain.getValue(1);
6510 // Add the <id> and <numBytes> constants.
6511 SDValue IDVal = getValue(CI.getOperand(PatchPointOpers::IDPos));
6512 Ops.push_back(DAG.getTargetConstant(
6513 cast<ConstantSDNode>(IDVal)->getZExtValue(), DL, MVT::i64));
6514 SDValue NBytesVal = getValue(CI.getOperand(PatchPointOpers::NBytesPos));
6515 Ops.push_back(DAG.getTargetConstant(
6516 cast<ConstantSDNode>(NBytesVal)->getZExtValue(), DL,
6519 // Push live variables for the stack map.
6520 addStackMapLiveVars(&CI, 2, DL, Ops, *this);
6522 // We are not pushing any register mask info here on the operands list,
6523 // because the stackmap doesn't clobber anything.
6525 // Push the chain and the glue flag.
6526 Ops.push_back(Chain);
6527 Ops.push_back(InFlag);
6529 // Create the STACKMAP node.
6530 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6531 SDNode *SM = DAG.getMachineNode(TargetOpcode::STACKMAP, DL, NodeTys, Ops);
6532 Chain = SDValue(SM, 0);
6533 InFlag = Chain.getValue(1);
6535 Chain = DAG.getCALLSEQ_END(Chain, NullPtr, NullPtr, InFlag, DL);
6537 // Stackmaps don't generate values, so nothing goes into the NodeMap.
6539 // Set the root to the target-lowered call chain.
6542 // Inform the Frame Information that we have a stackmap in this function.
6543 FuncInfo.MF->getFrameInfo()->setHasStackMap();
6546 /// \brief Lower llvm.experimental.patchpoint directly to its target opcode.
6547 void SelectionDAGBuilder::visitPatchpoint(ImmutableCallSite CS,
6548 MachineBasicBlock *LandingPad) {
6549 // void|i64 @llvm.experimental.patchpoint.void|i64(i64 <id>,
6554 // [live variables...])
6556 CallingConv::ID CC = CS.getCallingConv();
6557 bool IsAnyRegCC = CC == CallingConv::AnyReg;
6558 bool HasDef = !CS->getType()->isVoidTy();
6559 SDLoc dl = getCurSDLoc();
6560 SDValue Callee = getValue(CS->getOperand(PatchPointOpers::TargetPos));
6562 // Handle immediate and symbolic callees.
6563 if (auto* ConstCallee = dyn_cast<ConstantSDNode>(Callee))
6564 Callee = DAG.getIntPtrConstant(ConstCallee->getZExtValue(), dl,
6566 else if (auto* SymbolicCallee = dyn_cast<GlobalAddressSDNode>(Callee))
6567 Callee = DAG.getTargetGlobalAddress(SymbolicCallee->getGlobal(),
6568 SDLoc(SymbolicCallee),
6569 SymbolicCallee->getValueType(0));
6571 // Get the real number of arguments participating in the call <numArgs>
6572 SDValue NArgVal = getValue(CS.getArgument(PatchPointOpers::NArgPos));
6573 unsigned NumArgs = cast<ConstantSDNode>(NArgVal)->getZExtValue();
6575 // Skip the four meta args: <id>, <numNopBytes>, <target>, <numArgs>
6576 // Intrinsics include all meta-operands up to but not including CC.
6577 unsigned NumMetaOpers = PatchPointOpers::CCPos;
6578 assert(CS.arg_size() >= NumMetaOpers + NumArgs &&
6579 "Not enough arguments provided to the patchpoint intrinsic");
6581 // For AnyRegCC the arguments are lowered later on manually.
6582 unsigned NumCallArgs = IsAnyRegCC ? 0 : NumArgs;
6584 IsAnyRegCC ? Type::getVoidTy(*DAG.getContext()) : CS->getType();
6585 std::pair<SDValue, SDValue> Result =
6586 lowerCallOperands(CS, NumMetaOpers, NumCallArgs, Callee, ReturnTy,
6589 SDNode *CallEnd = Result.second.getNode();
6590 if (HasDef && (CallEnd->getOpcode() == ISD::CopyFromReg))
6591 CallEnd = CallEnd->getOperand(0).getNode();
6593 /// Get a call instruction from the call sequence chain.
6594 /// Tail calls are not allowed.
6595 assert(CallEnd->getOpcode() == ISD::CALLSEQ_END &&
6596 "Expected a callseq node.");
6597 SDNode *Call = CallEnd->getOperand(0).getNode();
6598 bool HasGlue = Call->getGluedNode();
6600 // Replace the target specific call node with the patchable intrinsic.
6601 SmallVector<SDValue, 8> Ops;
6603 // Add the <id> and <numBytes> constants.
6604 SDValue IDVal = getValue(CS->getOperand(PatchPointOpers::IDPos));
6605 Ops.push_back(DAG.getTargetConstant(
6606 cast<ConstantSDNode>(IDVal)->getZExtValue(), dl, MVT::i64));
6607 SDValue NBytesVal = getValue(CS->getOperand(PatchPointOpers::NBytesPos));
6608 Ops.push_back(DAG.getTargetConstant(
6609 cast<ConstantSDNode>(NBytesVal)->getZExtValue(), dl,
6613 Ops.push_back(Callee);
6615 // Adjust <numArgs> to account for any arguments that have been passed on the
6617 // Call Node: Chain, Target, {Args}, RegMask, [Glue]
6618 unsigned NumCallRegArgs = Call->getNumOperands() - (HasGlue ? 4 : 3);
6619 NumCallRegArgs = IsAnyRegCC ? NumArgs : NumCallRegArgs;
6620 Ops.push_back(DAG.getTargetConstant(NumCallRegArgs, dl, MVT::i32));
6622 // Add the calling convention
6623 Ops.push_back(DAG.getTargetConstant((unsigned)CC, dl, MVT::i32));
6625 // Add the arguments we omitted previously. The register allocator should
6626 // place these in any free register.
6628 for (unsigned i = NumMetaOpers, e = NumMetaOpers + NumArgs; i != e; ++i)
6629 Ops.push_back(getValue(CS.getArgument(i)));
6631 // Push the arguments from the call instruction up to the register mask.
6632 SDNode::op_iterator e = HasGlue ? Call->op_end()-2 : Call->op_end()-1;
6633 Ops.append(Call->op_begin() + 2, e);
6635 // Push live variables for the stack map.
6636 addStackMapLiveVars(CS, NumMetaOpers + NumArgs, dl, Ops, *this);
6638 // Push the register mask info.
6640 Ops.push_back(*(Call->op_end()-2));
6642 Ops.push_back(*(Call->op_end()-1));
6644 // Push the chain (this is originally the first operand of the call, but
6645 // becomes now the last or second to last operand).
6646 Ops.push_back(*(Call->op_begin()));
6648 // Push the glue flag (last operand).
6650 Ops.push_back(*(Call->op_end()-1));
6653 if (IsAnyRegCC && HasDef) {
6654 // Create the return types based on the intrinsic definition
6655 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6656 SmallVector<EVT, 3> ValueVTs;
6657 ComputeValueVTs(TLI, CS->getType(), ValueVTs);
6658 assert(ValueVTs.size() == 1 && "Expected only one return value type.");
6660 // There is always a chain and a glue type at the end
6661 ValueVTs.push_back(MVT::Other);
6662 ValueVTs.push_back(MVT::Glue);
6663 NodeTys = DAG.getVTList(ValueVTs);
6665 NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6667 // Replace the target specific call node with a PATCHPOINT node.
6668 MachineSDNode *MN = DAG.getMachineNode(TargetOpcode::PATCHPOINT,
6671 // Update the NodeMap.
6674 setValue(CS.getInstruction(), SDValue(MN, 0));
6676 setValue(CS.getInstruction(), Result.first);
6679 // Fixup the consumers of the intrinsic. The chain and glue may be used in the
6680 // call sequence. Furthermore the location of the chain and glue can change
6681 // when the AnyReg calling convention is used and the intrinsic returns a
6683 if (IsAnyRegCC && HasDef) {
6684 SDValue From[] = {SDValue(Call, 0), SDValue(Call, 1)};
6685 SDValue To[] = {SDValue(MN, 1), SDValue(MN, 2)};
6686 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
6688 DAG.ReplaceAllUsesWith(Call, MN);
6689 DAG.DeleteNode(Call);
6691 // Inform the Frame Information that we have a patchpoint in this function.
6692 FuncInfo.MF->getFrameInfo()->setHasPatchPoint();
6695 /// Returns an AttributeSet representing the attributes applied to the return
6696 /// value of the given call.
6697 static AttributeSet getReturnAttrs(TargetLowering::CallLoweringInfo &CLI) {
6698 SmallVector<Attribute::AttrKind, 2> Attrs;
6700 Attrs.push_back(Attribute::SExt);
6702 Attrs.push_back(Attribute::ZExt);
6704 Attrs.push_back(Attribute::InReg);
6706 return AttributeSet::get(CLI.RetTy->getContext(), AttributeSet::ReturnIndex,
6710 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
6711 /// implementation, which just calls LowerCall.
6712 /// FIXME: When all targets are
6713 /// migrated to using LowerCall, this hook should be integrated into SDISel.
6714 std::pair<SDValue, SDValue>
6715 TargetLowering::LowerCallTo(TargetLowering::CallLoweringInfo &CLI) const {
6716 // Handle the incoming return values from the call.
6718 Type *OrigRetTy = CLI.RetTy;
6719 SmallVector<EVT, 4> RetTys;
6720 SmallVector<uint64_t, 4> Offsets;
6721 ComputeValueVTs(*this, CLI.RetTy, RetTys, &Offsets);
6723 SmallVector<ISD::OutputArg, 4> Outs;
6724 GetReturnInfo(CLI.RetTy, getReturnAttrs(CLI), Outs, *this);
6726 bool CanLowerReturn =
6727 this->CanLowerReturn(CLI.CallConv, CLI.DAG.getMachineFunction(),
6728 CLI.IsVarArg, Outs, CLI.RetTy->getContext());
6730 SDValue DemoteStackSlot;
6731 int DemoteStackIdx = -100;
6732 if (!CanLowerReturn) {
6733 // FIXME: equivalent assert?
6734 // assert(!CS.hasInAllocaArgument() &&
6735 // "sret demotion is incompatible with inalloca");
6736 uint64_t TySize = getDataLayout()->getTypeAllocSize(CLI.RetTy);
6737 unsigned Align = getDataLayout()->getPrefTypeAlignment(CLI.RetTy);
6738 MachineFunction &MF = CLI.DAG.getMachineFunction();
6739 DemoteStackIdx = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
6740 Type *StackSlotPtrType = PointerType::getUnqual(CLI.RetTy);
6742 DemoteStackSlot = CLI.DAG.getFrameIndex(DemoteStackIdx, getPointerTy());
6744 Entry.Node = DemoteStackSlot;
6745 Entry.Ty = StackSlotPtrType;
6746 Entry.isSExt = false;
6747 Entry.isZExt = false;
6748 Entry.isInReg = false;
6749 Entry.isSRet = true;
6750 Entry.isNest = false;
6751 Entry.isByVal = false;
6752 Entry.isReturned = false;
6753 Entry.Alignment = Align;
6754 CLI.getArgs().insert(CLI.getArgs().begin(), Entry);
6755 CLI.RetTy = Type::getVoidTy(CLI.RetTy->getContext());
6757 // sret demotion isn't compatible with tail-calls, since the sret argument
6758 // points into the callers stack frame.
6759 CLI.IsTailCall = false;
6761 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
6763 MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
6764 unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
6765 for (unsigned i = 0; i != NumRegs; ++i) {
6766 ISD::InputArg MyFlags;
6767 MyFlags.VT = RegisterVT;
6769 MyFlags.Used = CLI.IsReturnValueUsed;
6771 MyFlags.Flags.setSExt();
6773 MyFlags.Flags.setZExt();
6775 MyFlags.Flags.setInReg();
6776 CLI.Ins.push_back(MyFlags);
6781 // Handle all of the outgoing arguments.
6783 CLI.OutVals.clear();
6784 ArgListTy &Args = CLI.getArgs();
6785 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
6786 SmallVector<EVT, 4> ValueVTs;
6787 ComputeValueVTs(*this, Args[i].Ty, ValueVTs);
6788 Type *FinalType = Args[i].Ty;
6789 if (Args[i].isByVal)
6790 FinalType = cast<PointerType>(Args[i].Ty)->getElementType();
6791 bool NeedsRegBlock = functionArgumentNeedsConsecutiveRegisters(
6792 FinalType, CLI.CallConv, CLI.IsVarArg);
6793 for (unsigned Value = 0, NumValues = ValueVTs.size(); Value != NumValues;
6795 EVT VT = ValueVTs[Value];
6796 Type *ArgTy = VT.getTypeForEVT(CLI.RetTy->getContext());
6797 SDValue Op = SDValue(Args[i].Node.getNode(),
6798 Args[i].Node.getResNo() + Value);
6799 ISD::ArgFlagsTy Flags;
6800 unsigned OriginalAlignment = getDataLayout()->getABITypeAlignment(ArgTy);
6806 if (Args[i].isInReg)
6810 if (Args[i].isByVal)
6812 if (Args[i].isInAlloca) {
6813 Flags.setInAlloca();
6814 // Set the byval flag for CCAssignFn callbacks that don't know about
6815 // inalloca. This way we can know how many bytes we should've allocated
6816 // and how many bytes a callee cleanup function will pop. If we port
6817 // inalloca to more targets, we'll have to add custom inalloca handling
6818 // in the various CC lowering callbacks.
6821 if (Args[i].isByVal || Args[i].isInAlloca) {
6822 PointerType *Ty = cast<PointerType>(Args[i].Ty);
6823 Type *ElementTy = Ty->getElementType();
6824 Flags.setByValSize(getDataLayout()->getTypeAllocSize(ElementTy));
6825 // For ByVal, alignment should come from FE. BE will guess if this
6826 // info is not there but there are cases it cannot get right.
6827 unsigned FrameAlign;
6828 if (Args[i].Alignment)
6829 FrameAlign = Args[i].Alignment;
6831 FrameAlign = getByValTypeAlignment(ElementTy);
6832 Flags.setByValAlign(FrameAlign);
6837 Flags.setInConsecutiveRegs();
6838 Flags.setOrigAlign(OriginalAlignment);
6840 MVT PartVT = getRegisterType(CLI.RetTy->getContext(), VT);
6841 unsigned NumParts = getNumRegisters(CLI.RetTy->getContext(), VT);
6842 SmallVector<SDValue, 4> Parts(NumParts);
6843 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
6846 ExtendKind = ISD::SIGN_EXTEND;
6847 else if (Args[i].isZExt)
6848 ExtendKind = ISD::ZERO_EXTEND;
6850 // Conservatively only handle 'returned' on non-vectors for now
6851 if (Args[i].isReturned && !Op.getValueType().isVector()) {
6852 assert(CLI.RetTy == Args[i].Ty && RetTys.size() == NumValues &&
6853 "unexpected use of 'returned'");
6854 // Before passing 'returned' to the target lowering code, ensure that
6855 // either the register MVT and the actual EVT are the same size or that
6856 // the return value and argument are extended in the same way; in these
6857 // cases it's safe to pass the argument register value unchanged as the
6858 // return register value (although it's at the target's option whether
6860 // TODO: allow code generation to take advantage of partially preserved
6861 // registers rather than clobbering the entire register when the
6862 // parameter extension method is not compatible with the return
6864 if ((NumParts * PartVT.getSizeInBits() == VT.getSizeInBits()) ||
6865 (ExtendKind != ISD::ANY_EXTEND &&
6866 CLI.RetSExt == Args[i].isSExt && CLI.RetZExt == Args[i].isZExt))
6867 Flags.setReturned();
6870 getCopyToParts(CLI.DAG, CLI.DL, Op, &Parts[0], NumParts, PartVT,
6871 CLI.CS ? CLI.CS->getInstruction() : nullptr, ExtendKind);
6873 for (unsigned j = 0; j != NumParts; ++j) {
6874 // if it isn't first piece, alignment must be 1
6875 ISD::OutputArg MyFlags(Flags, Parts[j].getValueType(), VT,
6876 i < CLI.NumFixedArgs,
6877 i, j*Parts[j].getValueType().getStoreSize());
6878 if (NumParts > 1 && j == 0)
6879 MyFlags.Flags.setSplit();
6881 MyFlags.Flags.setOrigAlign(1);
6883 CLI.Outs.push_back(MyFlags);
6884 CLI.OutVals.push_back(Parts[j]);
6887 if (NeedsRegBlock && Value == NumValues - 1)
6888 CLI.Outs[CLI.Outs.size() - 1].Flags.setInConsecutiveRegsLast();
6892 SmallVector<SDValue, 4> InVals;
6893 CLI.Chain = LowerCall(CLI, InVals);
6895 // Verify that the target's LowerCall behaved as expected.
6896 assert(CLI.Chain.getNode() && CLI.Chain.getValueType() == MVT::Other &&
6897 "LowerCall didn't return a valid chain!");
6898 assert((!CLI.IsTailCall || InVals.empty()) &&
6899 "LowerCall emitted a return value for a tail call!");
6900 assert((CLI.IsTailCall || InVals.size() == CLI.Ins.size()) &&
6901 "LowerCall didn't emit the correct number of values!");
6903 // For a tail call, the return value is merely live-out and there aren't
6904 // any nodes in the DAG representing it. Return a special value to
6905 // indicate that a tail call has been emitted and no more Instructions
6906 // should be processed in the current block.
6907 if (CLI.IsTailCall) {
6908 CLI.DAG.setRoot(CLI.Chain);
6909 return std::make_pair(SDValue(), SDValue());
6912 DEBUG(for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++i) {
6913 assert(InVals[i].getNode() &&
6914 "LowerCall emitted a null value!");
6915 assert(EVT(CLI.Ins[i].VT) == InVals[i].getValueType() &&
6916 "LowerCall emitted a value with the wrong type!");
6919 SmallVector<SDValue, 4> ReturnValues;
6920 if (!CanLowerReturn) {
6921 // The instruction result is the result of loading from the
6922 // hidden sret parameter.
6923 SmallVector<EVT, 1> PVTs;
6924 Type *PtrRetTy = PointerType::getUnqual(OrigRetTy);
6926 ComputeValueVTs(*this, PtrRetTy, PVTs);
6927 assert(PVTs.size() == 1 && "Pointers should fit in one register");
6928 EVT PtrVT = PVTs[0];
6930 unsigned NumValues = RetTys.size();
6931 ReturnValues.resize(NumValues);
6932 SmallVector<SDValue, 4> Chains(NumValues);
6934 for (unsigned i = 0; i < NumValues; ++i) {
6935 SDValue Add = CLI.DAG.getNode(ISD::ADD, CLI.DL, PtrVT, DemoteStackSlot,
6936 CLI.DAG.getConstant(Offsets[i], CLI.DL,
6938 SDValue L = CLI.DAG.getLoad(
6939 RetTys[i], CLI.DL, CLI.Chain, Add,
6940 MachinePointerInfo::getFixedStack(DemoteStackIdx, Offsets[i]), false,
6942 ReturnValues[i] = L;
6943 Chains[i] = L.getValue(1);
6946 CLI.Chain = CLI.DAG.getNode(ISD::TokenFactor, CLI.DL, MVT::Other, Chains);
6948 // Collect the legal value parts into potentially illegal values
6949 // that correspond to the original function's return values.
6950 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6952 AssertOp = ISD::AssertSext;
6953 else if (CLI.RetZExt)
6954 AssertOp = ISD::AssertZext;
6955 unsigned CurReg = 0;
6956 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
6958 MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
6959 unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
6961 ReturnValues.push_back(getCopyFromParts(CLI.DAG, CLI.DL, &InVals[CurReg],
6962 NumRegs, RegisterVT, VT, nullptr,
6967 // For a function returning void, there is no return value. We can't create
6968 // such a node, so we just return a null return value in that case. In
6969 // that case, nothing will actually look at the value.
6970 if (ReturnValues.empty())
6971 return std::make_pair(SDValue(), CLI.Chain);
6974 SDValue Res = CLI.DAG.getNode(ISD::MERGE_VALUES, CLI.DL,
6975 CLI.DAG.getVTList(RetTys), ReturnValues);
6976 return std::make_pair(Res, CLI.Chain);
6979 void TargetLowering::LowerOperationWrapper(SDNode *N,
6980 SmallVectorImpl<SDValue> &Results,
6981 SelectionDAG &DAG) const {
6982 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
6984 Results.push_back(Res);
6987 SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6988 llvm_unreachable("LowerOperation not implemented for this target!");
6992 SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) {
6993 SDValue Op = getNonRegisterValue(V);
6994 assert((Op.getOpcode() != ISD::CopyFromReg ||
6995 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
6996 "Copy from a reg to the same reg!");
6997 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg");
6999 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7000 RegsForValue RFV(V->getContext(), TLI, Reg, V->getType());
7001 SDValue Chain = DAG.getEntryNode();
7003 ISD::NodeType ExtendType = (FuncInfo.PreferredExtendType.find(V) ==
7004 FuncInfo.PreferredExtendType.end())
7006 : FuncInfo.PreferredExtendType[V];
7007 RFV.getCopyToRegs(Op, DAG, getCurSDLoc(), Chain, nullptr, V, ExtendType);
7008 PendingExports.push_back(Chain);
7011 #include "llvm/CodeGen/SelectionDAGISel.h"
7013 /// isOnlyUsedInEntryBlock - If the specified argument is only used in the
7014 /// entry block, return true. This includes arguments used by switches, since
7015 /// the switch may expand into multiple basic blocks.
7016 static bool isOnlyUsedInEntryBlock(const Argument *A, bool FastISel) {
7017 // With FastISel active, we may be splitting blocks, so force creation
7018 // of virtual registers for all non-dead arguments.
7020 return A->use_empty();
7022 const BasicBlock *Entry = A->getParent()->begin();
7023 for (const User *U : A->users())
7024 if (cast<Instruction>(U)->getParent() != Entry || isa<SwitchInst>(U))
7025 return false; // Use not in entry block.
7030 void SelectionDAGISel::LowerArguments(const Function &F) {
7031 SelectionDAG &DAG = SDB->DAG;
7032 SDLoc dl = SDB->getCurSDLoc();
7033 const DataLayout *DL = TLI->getDataLayout();
7034 SmallVector<ISD::InputArg, 16> Ins;
7036 if (!FuncInfo->CanLowerReturn) {
7037 // Put in an sret pointer parameter before all the other parameters.
7038 SmallVector<EVT, 1> ValueVTs;
7039 ComputeValueVTs(*TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
7041 // NOTE: Assuming that a pointer will never break down to more than one VT
7043 ISD::ArgFlagsTy Flags;
7045 MVT RegisterVT = TLI->getRegisterType(*DAG.getContext(), ValueVTs[0]);
7046 ISD::InputArg RetArg(Flags, RegisterVT, ValueVTs[0], true,
7047 ISD::InputArg::NoArgIndex, 0);
7048 Ins.push_back(RetArg);
7051 // Set up the incoming argument description vector.
7053 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end();
7054 I != E; ++I, ++Idx) {
7055 SmallVector<EVT, 4> ValueVTs;
7056 ComputeValueVTs(*TLI, I->getType(), ValueVTs);
7057 bool isArgValueUsed = !I->use_empty();
7058 unsigned PartBase = 0;
7059 Type *FinalType = I->getType();
7060 if (F.getAttributes().hasAttribute(Idx, Attribute::ByVal))
7061 FinalType = cast<PointerType>(FinalType)->getElementType();
7062 bool NeedsRegBlock = TLI->functionArgumentNeedsConsecutiveRegisters(
7063 FinalType, F.getCallingConv(), F.isVarArg());
7064 for (unsigned Value = 0, NumValues = ValueVTs.size();
7065 Value != NumValues; ++Value) {
7066 EVT VT = ValueVTs[Value];
7067 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
7068 ISD::ArgFlagsTy Flags;
7069 unsigned OriginalAlignment = DL->getABITypeAlignment(ArgTy);
7071 if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt))
7073 if (F.getAttributes().hasAttribute(Idx, Attribute::SExt))
7075 if (F.getAttributes().hasAttribute(Idx, Attribute::InReg))
7077 if (F.getAttributes().hasAttribute(Idx, Attribute::StructRet))
7079 if (F.getAttributes().hasAttribute(Idx, Attribute::ByVal))
7081 if (F.getAttributes().hasAttribute(Idx, Attribute::InAlloca)) {
7082 Flags.setInAlloca();
7083 // Set the byval flag for CCAssignFn callbacks that don't know about
7084 // inalloca. This way we can know how many bytes we should've allocated
7085 // and how many bytes a callee cleanup function will pop. If we port
7086 // inalloca to more targets, we'll have to add custom inalloca handling
7087 // in the various CC lowering callbacks.
7090 if (Flags.isByVal() || Flags.isInAlloca()) {
7091 PointerType *Ty = cast<PointerType>(I->getType());
7092 Type *ElementTy = Ty->getElementType();
7093 Flags.setByValSize(DL->getTypeAllocSize(ElementTy));
7094 // For ByVal, alignment should be passed from FE. BE will guess if
7095 // this info is not there but there are cases it cannot get right.
7096 unsigned FrameAlign;
7097 if (F.getParamAlignment(Idx))
7098 FrameAlign = F.getParamAlignment(Idx);
7100 FrameAlign = TLI->getByValTypeAlignment(ElementTy);
7101 Flags.setByValAlign(FrameAlign);
7103 if (F.getAttributes().hasAttribute(Idx, Attribute::Nest))
7106 Flags.setInConsecutiveRegs();
7107 Flags.setOrigAlign(OriginalAlignment);
7109 MVT RegisterVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7110 unsigned NumRegs = TLI->getNumRegisters(*CurDAG->getContext(), VT);
7111 for (unsigned i = 0; i != NumRegs; ++i) {
7112 ISD::InputArg MyFlags(Flags, RegisterVT, VT, isArgValueUsed,
7113 Idx-1, PartBase+i*RegisterVT.getStoreSize());
7114 if (NumRegs > 1 && i == 0)
7115 MyFlags.Flags.setSplit();
7116 // if it isn't first piece, alignment must be 1
7118 MyFlags.Flags.setOrigAlign(1);
7119 Ins.push_back(MyFlags);
7121 if (NeedsRegBlock && Value == NumValues - 1)
7122 Ins[Ins.size() - 1].Flags.setInConsecutiveRegsLast();
7123 PartBase += VT.getStoreSize();
7127 // Call the target to set up the argument values.
7128 SmallVector<SDValue, 8> InVals;
7129 SDValue NewRoot = TLI->LowerFormalArguments(
7130 DAG.getRoot(), F.getCallingConv(), F.isVarArg(), Ins, dl, DAG, InVals);
7132 // Verify that the target's LowerFormalArguments behaved as expected.
7133 assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other &&
7134 "LowerFormalArguments didn't return a valid chain!");
7135 assert(InVals.size() == Ins.size() &&
7136 "LowerFormalArguments didn't emit the correct number of values!");
7138 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
7139 assert(InVals[i].getNode() &&
7140 "LowerFormalArguments emitted a null value!");
7141 assert(EVT(Ins[i].VT) == InVals[i].getValueType() &&
7142 "LowerFormalArguments emitted a value with the wrong type!");
7146 // Update the DAG with the new chain value resulting from argument lowering.
7147 DAG.setRoot(NewRoot);
7149 // Set up the argument values.
7152 if (!FuncInfo->CanLowerReturn) {
7153 // Create a virtual register for the sret pointer, and put in a copy
7154 // from the sret argument into it.
7155 SmallVector<EVT, 1> ValueVTs;
7156 ComputeValueVTs(*TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
7157 MVT VT = ValueVTs[0].getSimpleVT();
7158 MVT RegVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7159 ISD::NodeType AssertOp = ISD::DELETED_NODE;
7160 SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1,
7161 RegVT, VT, nullptr, AssertOp);
7163 MachineFunction& MF = SDB->DAG.getMachineFunction();
7164 MachineRegisterInfo& RegInfo = MF.getRegInfo();
7165 unsigned SRetReg = RegInfo.createVirtualRegister(TLI->getRegClassFor(RegVT));
7166 FuncInfo->DemoteRegister = SRetReg;
7168 SDB->DAG.getCopyToReg(NewRoot, SDB->getCurSDLoc(), SRetReg, ArgValue);
7169 DAG.setRoot(NewRoot);
7171 // i indexes lowered arguments. Bump it past the hidden sret argument.
7172 // Idx indexes LLVM arguments. Don't touch it.
7176 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
7178 SmallVector<SDValue, 4> ArgValues;
7179 SmallVector<EVT, 4> ValueVTs;
7180 ComputeValueVTs(*TLI, I->getType(), ValueVTs);
7181 unsigned NumValues = ValueVTs.size();
7183 // If this argument is unused then remember its value. It is used to generate
7184 // debugging information.
7185 if (I->use_empty() && NumValues) {
7186 SDB->setUnusedArgValue(I, InVals[i]);
7188 // Also remember any frame index for use in FastISel.
7189 if (FrameIndexSDNode *FI =
7190 dyn_cast<FrameIndexSDNode>(InVals[i].getNode()))
7191 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
7194 for (unsigned Val = 0; Val != NumValues; ++Val) {
7195 EVT VT = ValueVTs[Val];
7196 MVT PartVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7197 unsigned NumParts = TLI->getNumRegisters(*CurDAG->getContext(), VT);
7199 if (!I->use_empty()) {
7200 ISD::NodeType AssertOp = ISD::DELETED_NODE;
7201 if (F.getAttributes().hasAttribute(Idx, Attribute::SExt))
7202 AssertOp = ISD::AssertSext;
7203 else if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt))
7204 AssertOp = ISD::AssertZext;
7206 ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i],
7207 NumParts, PartVT, VT,
7208 nullptr, AssertOp));
7214 // We don't need to do anything else for unused arguments.
7215 if (ArgValues.empty())
7218 // Note down frame index.
7219 if (FrameIndexSDNode *FI =
7220 dyn_cast<FrameIndexSDNode>(ArgValues[0].getNode()))
7221 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
7223 SDValue Res = DAG.getMergeValues(makeArrayRef(ArgValues.data(), NumValues),
7224 SDB->getCurSDLoc());
7226 SDB->setValue(I, Res);
7227 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::BUILD_PAIR) {
7228 if (LoadSDNode *LNode =
7229 dyn_cast<LoadSDNode>(Res.getOperand(0).getNode()))
7230 if (FrameIndexSDNode *FI =
7231 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode()))
7232 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
7235 // If this argument is live outside of the entry block, insert a copy from
7236 // wherever we got it to the vreg that other BB's will reference it as.
7237 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::CopyFromReg) {
7238 // If we can, though, try to skip creating an unnecessary vreg.
7239 // FIXME: This isn't very clean... it would be nice to make this more
7240 // general. It's also subtly incompatible with the hacks FastISel
7242 unsigned Reg = cast<RegisterSDNode>(Res.getOperand(1))->getReg();
7243 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
7244 FuncInfo->ValueMap[I] = Reg;
7248 if (!isOnlyUsedInEntryBlock(I, TM.Options.EnableFastISel)) {
7249 FuncInfo->InitializeRegForValue(I);
7250 SDB->CopyToExportRegsIfNeeded(I);
7254 assert(i == InVals.size() && "Argument register count mismatch!");
7256 // Finally, if the target has anything special to do, allow it to do so.
7257 EmitFunctionEntryCode();
7260 /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
7261 /// ensure constants are generated when needed. Remember the virtual registers
7262 /// that need to be added to the Machine PHI nodes as input. We cannot just
7263 /// directly add them, because expansion might result in multiple MBB's for one
7264 /// BB. As such, the start of the BB might correspond to a different MBB than
7268 SelectionDAGBuilder::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
7269 const TerminatorInst *TI = LLVMBB->getTerminator();
7271 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
7273 // Check PHI nodes in successors that expect a value to be available from this
7275 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
7276 const BasicBlock *SuccBB = TI->getSuccessor(succ);
7277 if (!isa<PHINode>(SuccBB->begin())) continue;
7278 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
7280 // If this terminator has multiple identical successors (common for
7281 // switches), only handle each succ once.
7282 if (!SuccsHandled.insert(SuccMBB).second)
7285 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
7287 // At this point we know that there is a 1-1 correspondence between LLVM PHI
7288 // nodes and Machine PHI nodes, but the incoming operands have not been
7290 for (BasicBlock::const_iterator I = SuccBB->begin();
7291 const PHINode *PN = dyn_cast<PHINode>(I); ++I) {
7292 // Ignore dead phi's.
7293 if (PN->use_empty()) continue;
7296 if (PN->getType()->isEmptyTy())
7300 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
7302 if (const Constant *C = dyn_cast<Constant>(PHIOp)) {
7303 unsigned &RegOut = ConstantsOut[C];
7305 RegOut = FuncInfo.CreateRegs(C->getType());
7306 CopyValueToVirtualRegister(C, RegOut);
7310 DenseMap<const Value *, unsigned>::iterator I =
7311 FuncInfo.ValueMap.find(PHIOp);
7312 if (I != FuncInfo.ValueMap.end())
7315 assert(isa<AllocaInst>(PHIOp) &&
7316 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
7317 "Didn't codegen value into a register!??");
7318 Reg = FuncInfo.CreateRegs(PHIOp->getType());
7319 CopyValueToVirtualRegister(PHIOp, Reg);
7323 // Remember that this register needs to added to the machine PHI node as
7324 // the input for this MBB.
7325 SmallVector<EVT, 4> ValueVTs;
7326 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7327 ComputeValueVTs(TLI, PN->getType(), ValueVTs);
7328 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
7329 EVT VT = ValueVTs[vti];
7330 unsigned NumRegisters = TLI.getNumRegisters(*DAG.getContext(), VT);
7331 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
7332 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
7333 Reg += NumRegisters;
7338 ConstantsOut.clear();
7341 /// Add a successor MBB to ParentMBB< creating a new MachineBB for BB if SuccMBB
7344 SelectionDAGBuilder::StackProtectorDescriptor::
7345 AddSuccessorMBB(const BasicBlock *BB,
7346 MachineBasicBlock *ParentMBB,
7348 MachineBasicBlock *SuccMBB) {
7349 // If SuccBB has not been created yet, create it.
7351 MachineFunction *MF = ParentMBB->getParent();
7352 MachineFunction::iterator BBI = ParentMBB;
7353 SuccMBB = MF->CreateMachineBasicBlock(BB);
7354 MF->insert(++BBI, SuccMBB);
7356 // Add it as a successor of ParentMBB.
7357 ParentMBB->addSuccessor(
7358 SuccMBB, BranchProbabilityInfo::getBranchWeightStackProtector(IsLikely));
7362 MachineBasicBlock *SelectionDAGBuilder::NextBlock(MachineBasicBlock *MBB) {
7363 MachineFunction::iterator I = MBB;
7364 if (++I == FuncInfo.MF->end())
7369 /// During lowering new call nodes can be created (such as memset, etc.).
7370 /// Those will become new roots of the current DAG, but complications arise
7371 /// when they are tail calls. In such cases, the call lowering will update
7372 /// the root, but the builder still needs to know that a tail call has been
7373 /// lowered in order to avoid generating an additional return.
7374 void SelectionDAGBuilder::updateDAGForMaybeTailCall(SDValue MaybeTC) {
7375 // If the node is null, we do have a tail call.
7376 if (MaybeTC.getNode() != nullptr)
7377 DAG.setRoot(MaybeTC);
7382 bool SelectionDAGBuilder::isDense(const CaseClusterVector &Clusters,
7383 unsigned *TotalCases, unsigned First,
7385 assert(Last >= First);
7386 assert(TotalCases[Last] >= TotalCases[First]);
7388 APInt LowCase = Clusters[First].Low->getValue();
7389 APInt HighCase = Clusters[Last].High->getValue();
7390 assert(LowCase.getBitWidth() == HighCase.getBitWidth());
7392 // FIXME: A range of consecutive cases has 100% density, but only requires one
7393 // comparison to lower. We should discriminate against such consecutive ranges
7396 uint64_t Diff = (HighCase - LowCase).getLimitedValue((UINT64_MAX - 1) / 100);
7397 uint64_t Range = Diff + 1;
7400 TotalCases[Last] - (First == 0 ? 0 : TotalCases[First - 1]);
7402 assert(NumCases < UINT64_MAX / 100);
7403 assert(Range >= NumCases);
7405 return NumCases * 100 >= Range * MinJumpTableDensity;
7408 static inline bool areJTsAllowed(const TargetLowering &TLI) {
7409 return TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
7410 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other);
7413 bool SelectionDAGBuilder::buildJumpTable(CaseClusterVector &Clusters,
7414 unsigned First, unsigned Last,
7415 const SwitchInst *SI,
7416 MachineBasicBlock *DefaultMBB,
7417 CaseCluster &JTCluster) {
7418 assert(First <= Last);
7420 uint32_t Weight = 0;
7421 unsigned NumCmps = 0;
7422 std::vector<MachineBasicBlock*> Table;
7423 DenseMap<MachineBasicBlock*, uint32_t> JTWeights;
7424 for (unsigned I = First; I <= Last; ++I) {
7425 assert(Clusters[I].Kind == CC_Range);
7426 Weight += Clusters[I].Weight;
7427 assert(Weight >= Clusters[I].Weight && "Weight overflow!");
7428 APInt Low = Clusters[I].Low->getValue();
7429 APInt High = Clusters[I].High->getValue();
7430 NumCmps += (Low == High) ? 1 : 2;
7432 // Fill the gap between this and the previous cluster.
7433 APInt PreviousHigh = Clusters[I - 1].High->getValue();
7434 assert(PreviousHigh.slt(Low));
7435 uint64_t Gap = (Low - PreviousHigh).getLimitedValue() - 1;
7436 for (uint64_t J = 0; J < Gap; J++)
7437 Table.push_back(DefaultMBB);
7439 uint64_t ClusterSize = (High - Low).getLimitedValue() + 1;
7440 for (uint64_t J = 0; J < ClusterSize; ++J)
7441 Table.push_back(Clusters[I].MBB);
7442 JTWeights[Clusters[I].MBB] += Clusters[I].Weight;
7445 unsigned NumDests = JTWeights.size();
7446 if (isSuitableForBitTests(NumDests, NumCmps,
7447 Clusters[First].Low->getValue(),
7448 Clusters[Last].High->getValue())) {
7449 // Clusters[First..Last] should be lowered as bit tests instead.
7453 // Create the MBB that will load from and jump through the table.
7454 // Note: We create it here, but it's not inserted into the function yet.
7455 MachineFunction *CurMF = FuncInfo.MF;
7456 MachineBasicBlock *JumpTableMBB =
7457 CurMF->CreateMachineBasicBlock(SI->getParent());
7459 // Add successors. Note: use table order for determinism.
7460 SmallPtrSet<MachineBasicBlock *, 8> Done;
7461 for (MachineBasicBlock *Succ : Table) {
7462 if (Done.count(Succ))
7464 addSuccessorWithWeight(JumpTableMBB, Succ, JTWeights[Succ]);
7468 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7469 unsigned JTI = CurMF->getOrCreateJumpTableInfo(TLI.getJumpTableEncoding())
7470 ->createJumpTableIndex(Table);
7472 // Set up the jump table info.
7473 JumpTable JT(-1U, JTI, JumpTableMBB, nullptr);
7474 JumpTableHeader JTH(Clusters[First].Low->getValue(),
7475 Clusters[Last].High->getValue(), SI->getCondition(),
7477 JTCases.emplace_back(std::move(JTH), std::move(JT));
7479 JTCluster = CaseCluster::jumpTable(Clusters[First].Low, Clusters[Last].High,
7480 JTCases.size() - 1, Weight);
7484 void SelectionDAGBuilder::findJumpTables(CaseClusterVector &Clusters,
7485 const SwitchInst *SI,
7486 MachineBasicBlock *DefaultMBB) {
7488 // Clusters must be non-empty, sorted, and only contain Range clusters.
7489 assert(!Clusters.empty());
7490 for (CaseCluster &C : Clusters)
7491 assert(C.Kind == CC_Range);
7492 for (unsigned i = 1, e = Clusters.size(); i < e; ++i)
7493 assert(Clusters[i - 1].High->getValue().slt(Clusters[i].Low->getValue()));
7496 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7497 if (!areJTsAllowed(TLI))
7500 const int64_t N = Clusters.size();
7501 const unsigned MinJumpTableSize = TLI.getMinimumJumpTableEntries();
7503 // TotalCases[i]: Total nbr of cases in Clusters[0..i].
7504 SmallVector<unsigned, 8> TotalCases(N);
7506 for (unsigned i = 0; i < N; ++i) {
7507 APInt Hi = Clusters[i].High->getValue();
7508 APInt Lo = Clusters[i].Low->getValue();
7509 TotalCases[i] = (Hi - Lo).getLimitedValue() + 1;
7511 TotalCases[i] += TotalCases[i - 1];
7514 if (N >= MinJumpTableSize && isDense(Clusters, &TotalCases[0], 0, N - 1)) {
7515 // Cheap case: the whole range might be suitable for jump table.
7516 CaseCluster JTCluster;
7517 if (buildJumpTable(Clusters, 0, N - 1, SI, DefaultMBB, JTCluster)) {
7518 Clusters[0] = JTCluster;
7524 // The algorithm below is not suitable for -O0.
7525 if (TM.getOptLevel() == CodeGenOpt::None)
7528 // Split Clusters into minimum number of dense partitions. The algorithm uses
7529 // the same idea as Kannan & Proebsting "Correction to 'Producing Good Code
7530 // for the Case Statement'" (1994), but builds the MinPartitions array in
7531 // reverse order to make it easier to reconstruct the partitions in ascending
7532 // order. In the choice between two optimal partitionings, it picks the one
7533 // which yields more jump tables.
7535 // MinPartitions[i] is the minimum nbr of partitions of Clusters[i..N-1].
7536 SmallVector<unsigned, 8> MinPartitions(N);
7537 // LastElement[i] is the last element of the partition starting at i.
7538 SmallVector<unsigned, 8> LastElement(N);
7539 // NumTables[i]: nbr of >= MinJumpTableSize partitions from Clusters[i..N-1].
7540 SmallVector<unsigned, 8> NumTables(N);
7542 // Base case: There is only one way to partition Clusters[N-1].
7543 MinPartitions[N - 1] = 1;
7544 LastElement[N - 1] = N - 1;
7545 assert(MinJumpTableSize > 1);
7546 NumTables[N - 1] = 0;
7548 // Note: loop indexes are signed to avoid underflow.
7549 for (int64_t i = N - 2; i >= 0; i--) {
7550 // Find optimal partitioning of Clusters[i..N-1].
7551 // Baseline: Put Clusters[i] into a partition on its own.
7552 MinPartitions[i] = MinPartitions[i + 1] + 1;
7554 NumTables[i] = NumTables[i + 1];
7556 // Search for a solution that results in fewer partitions.
7557 for (int64_t j = N - 1; j > i; j--) {
7558 // Try building a partition from Clusters[i..j].
7559 if (isDense(Clusters, &TotalCases[0], i, j)) {
7560 unsigned NumPartitions = 1 + (j == N - 1 ? 0 : MinPartitions[j + 1]);
7561 bool IsTable = j - i + 1 >= MinJumpTableSize;
7562 unsigned Tables = IsTable + (j == N - 1 ? 0 : NumTables[j + 1]);
7564 // If this j leads to fewer partitions, or same number of partitions
7565 // with more lookup tables, it is a better partitioning.
7566 if (NumPartitions < MinPartitions[i] ||
7567 (NumPartitions == MinPartitions[i] && Tables > NumTables[i])) {
7568 MinPartitions[i] = NumPartitions;
7570 NumTables[i] = Tables;
7576 // Iterate over the partitions, replacing some with jump tables in-place.
7577 unsigned DstIndex = 0;
7578 for (unsigned First = 0, Last; First < N; First = Last + 1) {
7579 Last = LastElement[First];
7580 assert(Last >= First);
7581 assert(DstIndex <= First);
7582 unsigned NumClusters = Last - First + 1;
7584 CaseCluster JTCluster;
7585 if (NumClusters >= MinJumpTableSize &&
7586 buildJumpTable(Clusters, First, Last, SI, DefaultMBB, JTCluster)) {
7587 Clusters[DstIndex++] = JTCluster;
7589 for (unsigned I = First; I <= Last; ++I)
7590 std::memmove(&Clusters[DstIndex++], &Clusters[I], sizeof(Clusters[I]));
7593 Clusters.resize(DstIndex);
7596 bool SelectionDAGBuilder::rangeFitsInWord(const APInt &Low, const APInt &High) {
7597 // FIXME: Using the pointer type doesn't seem ideal.
7598 uint64_t BW = DAG.getTargetLoweringInfo().getPointerTy().getSizeInBits();
7599 uint64_t Range = (High - Low).getLimitedValue(UINT64_MAX - 1) + 1;
7603 bool SelectionDAGBuilder::isSuitableForBitTests(unsigned NumDests,
7606 const APInt &High) {
7607 // FIXME: I don't think NumCmps is the correct metric: a single case and a
7608 // range of cases both require only one branch to lower. Just looking at the
7609 // number of clusters and destinations should be enough to decide whether to
7612 // To lower a range with bit tests, the range must fit the bitwidth of a
7614 if (!rangeFitsInWord(Low, High))
7617 // Decide whether it's profitable to lower this range with bit tests. Each
7618 // destination requires a bit test and branch, and there is an overall range
7619 // check branch. For a small number of clusters, separate comparisons might be
7620 // cheaper, and for many destinations, splitting the range might be better.
7621 return (NumDests == 1 && NumCmps >= 3) ||
7622 (NumDests == 2 && NumCmps >= 5) ||
7623 (NumDests == 3 && NumCmps >= 6);
7626 bool SelectionDAGBuilder::buildBitTests(CaseClusterVector &Clusters,
7627 unsigned First, unsigned Last,
7628 const SwitchInst *SI,
7629 CaseCluster &BTCluster) {
7630 assert(First <= Last);
7634 BitVector Dests(FuncInfo.MF->getNumBlockIDs());
7635 unsigned NumCmps = 0;
7636 for (int64_t I = First; I <= Last; ++I) {
7637 assert(Clusters[I].Kind == CC_Range);
7638 Dests.set(Clusters[I].MBB->getNumber());
7639 NumCmps += (Clusters[I].Low == Clusters[I].High) ? 1 : 2;
7641 unsigned NumDests = Dests.count();
7643 APInt Low = Clusters[First].Low->getValue();
7644 APInt High = Clusters[Last].High->getValue();
7645 assert(Low.slt(High));
7647 if (!isSuitableForBitTests(NumDests, NumCmps, Low, High))
7653 const int BitWidth =
7654 DAG.getTargetLoweringInfo().getPointerTy().getSizeInBits();
7655 assert(rangeFitsInWord(Low, High) && "Case range must fit in bit mask!");
7657 if (Low.isNonNegative() && High.slt(BitWidth)) {
7658 // Optimize the case where all the case values fit in a
7659 // word without having to subtract minValue. In this case,
7660 // we can optimize away the subtraction.
7661 LowBound = APInt::getNullValue(Low.getBitWidth());
7665 CmpRange = High - Low;
7669 uint32_t TotalWeight = 0;
7670 for (unsigned i = First; i <= Last; ++i) {
7671 // Find the CaseBits for this destination.
7673 for (j = 0; j < CBV.size(); ++j)
7674 if (CBV[j].BB == Clusters[i].MBB)
7676 if (j == CBV.size())
7677 CBV.push_back(CaseBits(0, Clusters[i].MBB, 0, 0));
7678 CaseBits *CB = &CBV[j];
7680 // Update Mask, Bits and ExtraWeight.
7681 uint64_t Lo = (Clusters[i].Low->getValue() - LowBound).getZExtValue();
7682 uint64_t Hi = (Clusters[i].High->getValue() - LowBound).getZExtValue();
7683 assert(Hi >= Lo && Hi < 64 && "Invalid bit case!");
7684 CB->Mask |= (-1ULL >> (63 - (Hi - Lo))) << Lo;
7685 CB->Bits += Hi - Lo + 1;
7686 CB->ExtraWeight += Clusters[i].Weight;
7687 TotalWeight += Clusters[i].Weight;
7688 assert(TotalWeight >= Clusters[i].Weight && "Weight overflow!");
7692 std::sort(CBV.begin(), CBV.end(), [](const CaseBits &a, const CaseBits &b) {
7693 // Sort by weight first, number of bits second.
7694 if (a.ExtraWeight != b.ExtraWeight)
7695 return a.ExtraWeight > b.ExtraWeight;
7696 return a.Bits > b.Bits;
7699 for (auto &CB : CBV) {
7700 MachineBasicBlock *BitTestBB =
7701 FuncInfo.MF->CreateMachineBasicBlock(SI->getParent());
7702 BTI.push_back(BitTestCase(CB.Mask, BitTestBB, CB.BB, CB.ExtraWeight));
7704 BitTestCases.emplace_back(std::move(LowBound), std::move(CmpRange),
7705 SI->getCondition(), -1U, MVT::Other, false, nullptr,
7706 nullptr, std::move(BTI));
7708 BTCluster = CaseCluster::bitTests(Clusters[First].Low, Clusters[Last].High,
7709 BitTestCases.size() - 1, TotalWeight);
7713 void SelectionDAGBuilder::findBitTestClusters(CaseClusterVector &Clusters,
7714 const SwitchInst *SI) {
7715 // Partition Clusters into as few subsets as possible, where each subset has a
7716 // range that fits in a machine word and has <= 3 unique destinations.
7719 // Clusters must be sorted and contain Range or JumpTable clusters.
7720 assert(!Clusters.empty());
7721 assert(Clusters[0].Kind == CC_Range || Clusters[0].Kind == CC_JumpTable);
7722 for (const CaseCluster &C : Clusters)
7723 assert(C.Kind == CC_Range || C.Kind == CC_JumpTable);
7724 for (unsigned i = 1; i < Clusters.size(); ++i)
7725 assert(Clusters[i-1].High->getValue().slt(Clusters[i].Low->getValue()));
7728 // The algorithm below is not suitable for -O0.
7729 if (TM.getOptLevel() == CodeGenOpt::None)
7732 // If target does not have legal shift left, do not emit bit tests at all.
7733 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7734 EVT PTy = TLI.getPointerTy();
7735 if (!TLI.isOperationLegal(ISD::SHL, PTy))
7738 int BitWidth = PTy.getSizeInBits();
7739 const int64_t N = Clusters.size();
7741 // MinPartitions[i] is the minimum nbr of partitions of Clusters[i..N-1].
7742 SmallVector<unsigned, 8> MinPartitions(N);
7743 // LastElement[i] is the last element of the partition starting at i.
7744 SmallVector<unsigned, 8> LastElement(N);
7746 // FIXME: This might not be the best algorithm for finding bit test clusters.
7748 // Base case: There is only one way to partition Clusters[N-1].
7749 MinPartitions[N - 1] = 1;
7750 LastElement[N - 1] = N - 1;
7752 // Note: loop indexes are signed to avoid underflow.
7753 for (int64_t i = N - 2; i >= 0; --i) {
7754 // Find optimal partitioning of Clusters[i..N-1].
7755 // Baseline: Put Clusters[i] into a partition on its own.
7756 MinPartitions[i] = MinPartitions[i + 1] + 1;
7759 // Search for a solution that results in fewer partitions.
7760 // Note: the search is limited by BitWidth, reducing time complexity.
7761 for (int64_t j = std::min(N - 1, i + BitWidth - 1); j > i; --j) {
7762 // Try building a partition from Clusters[i..j].
7765 if (!rangeFitsInWord(Clusters[i].Low->getValue(),
7766 Clusters[j].High->getValue()))
7769 // Check nbr of destinations and cluster types.
7770 // FIXME: This works, but doesn't seem very efficient.
7771 bool RangesOnly = true;
7772 BitVector Dests(FuncInfo.MF->getNumBlockIDs());
7773 for (int64_t k = i; k <= j; k++) {
7774 if (Clusters[k].Kind != CC_Range) {
7778 Dests.set(Clusters[k].MBB->getNumber());
7780 if (!RangesOnly || Dests.count() > 3)
7783 // Check if it's a better partition.
7784 unsigned NumPartitions = 1 + (j == N - 1 ? 0 : MinPartitions[j + 1]);
7785 if (NumPartitions < MinPartitions[i]) {
7786 // Found a better partition.
7787 MinPartitions[i] = NumPartitions;
7793 // Iterate over the partitions, replacing with bit-test clusters in-place.
7794 unsigned DstIndex = 0;
7795 for (unsigned First = 0, Last; First < N; First = Last + 1) {
7796 Last = LastElement[First];
7797 assert(First <= Last);
7798 assert(DstIndex <= First);
7800 CaseCluster BitTestCluster;
7801 if (buildBitTests(Clusters, First, Last, SI, BitTestCluster)) {
7802 Clusters[DstIndex++] = BitTestCluster;
7804 size_t NumClusters = Last - First + 1;
7805 std::memmove(&Clusters[DstIndex], &Clusters[First],
7806 sizeof(Clusters[0]) * NumClusters);
7807 DstIndex += NumClusters;
7810 Clusters.resize(DstIndex);
7813 void SelectionDAGBuilder::lowerWorkItem(SwitchWorkListItem W, Value *Cond,
7814 MachineBasicBlock *SwitchMBB,
7815 MachineBasicBlock *DefaultMBB) {
7816 MachineFunction *CurMF = FuncInfo.MF;
7817 MachineBasicBlock *NextMBB = nullptr;
7818 MachineFunction::iterator BBI = W.MBB;
7819 if (++BBI != FuncInfo.MF->end())
7822 unsigned Size = W.LastCluster - W.FirstCluster + 1;
7824 BranchProbabilityInfo *BPI = FuncInfo.BPI;
7826 if (Size == 2 && W.MBB == SwitchMBB) {
7827 // If any two of the cases has the same destination, and if one value
7828 // is the same as the other, but has one bit unset that the other has set,
7829 // use bit manipulation to do two compares at once. For example:
7830 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
7831 // TODO: This could be extended to merge any 2 cases in switches with 3
7833 // TODO: Handle cases where W.CaseBB != SwitchBB.
7834 CaseCluster &Small = *W.FirstCluster;
7835 CaseCluster &Big = *W.LastCluster;
7837 if (Small.Low == Small.High && Big.Low == Big.High &&
7838 Small.MBB == Big.MBB) {
7839 const APInt &SmallValue = Small.Low->getValue();
7840 const APInt &BigValue = Big.Low->getValue();
7842 // Check that there is only one bit different.
7843 APInt CommonBit = BigValue ^ SmallValue;
7844 if (CommonBit.isPowerOf2()) {
7845 SDValue CondLHS = getValue(Cond);
7846 EVT VT = CondLHS.getValueType();
7847 SDLoc DL = getCurSDLoc();
7849 SDValue Or = DAG.getNode(ISD::OR, DL, VT, CondLHS,
7850 DAG.getConstant(CommonBit, DL, VT));
7851 SDValue Cond = DAG.getSetCC(
7852 DL, MVT::i1, Or, DAG.getConstant(BigValue | SmallValue, DL, VT),
7855 // Update successor info.
7856 // Both Small and Big will jump to Small.BB, so we sum up the weights.
7857 addSuccessorWithWeight(SwitchMBB, Small.MBB, Small.Weight + Big.Weight);
7858 addSuccessorWithWeight(
7859 SwitchMBB, DefaultMBB,
7860 // The default destination is the first successor in IR.
7861 BPI ? BPI->getEdgeWeight(SwitchMBB->getBasicBlock(), (unsigned)0)
7864 // Insert the true branch.
7866 DAG.getNode(ISD::BRCOND, DL, MVT::Other, getControlRoot(), Cond,
7867 DAG.getBasicBlock(Small.MBB));
7868 // Insert the false branch.
7869 BrCond = DAG.getNode(ISD::BR, DL, MVT::Other, BrCond,
7870 DAG.getBasicBlock(DefaultMBB));
7872 DAG.setRoot(BrCond);
7878 if (TM.getOptLevel() != CodeGenOpt::None) {
7879 // Order cases by weight so the most likely case will be checked first.
7880 std::sort(W.FirstCluster, W.LastCluster + 1,
7881 [](const CaseCluster &a, const CaseCluster &b) {
7882 return a.Weight > b.Weight;
7885 // Rearrange the case blocks so that the last one falls through if possible
7886 // without without changing the order of weights.
7887 for (CaseClusterIt I = W.LastCluster; I > W.FirstCluster; ) {
7889 if (I->Weight > W.LastCluster->Weight)
7891 if (I->Kind == CC_Range && I->MBB == NextMBB) {
7892 std::swap(*I, *W.LastCluster);
7898 // Compute total weight.
7899 uint32_t UnhandledWeights = 0;
7900 for (CaseClusterIt I = W.FirstCluster; I <= W.LastCluster; ++I) {
7901 UnhandledWeights += I->Weight;
7902 assert(UnhandledWeights >= I->Weight && "Weight overflow!");
7905 MachineBasicBlock *CurMBB = W.MBB;
7906 for (CaseClusterIt I = W.FirstCluster, E = W.LastCluster; I <= E; ++I) {
7907 MachineBasicBlock *Fallthrough;
7908 if (I == W.LastCluster) {
7909 // For the last cluster, fall through to the default destination.
7910 Fallthrough = DefaultMBB;
7912 Fallthrough = CurMF->CreateMachineBasicBlock(CurMBB->getBasicBlock());
7913 CurMF->insert(BBI, Fallthrough);
7914 // Put Cond in a virtual register to make it available from the new blocks.
7915 ExportFromCurrentBlock(Cond);
7919 case CC_JumpTable: {
7920 // FIXME: Optimize away range check based on pivot comparisons.
7921 JumpTableHeader *JTH = &JTCases[I->JTCasesIndex].first;
7922 JumpTable *JT = &JTCases[I->JTCasesIndex].second;
7924 // The jump block hasn't been inserted yet; insert it here.
7925 MachineBasicBlock *JumpMBB = JT->MBB;
7926 CurMF->insert(BBI, JumpMBB);
7927 addSuccessorWithWeight(CurMBB, Fallthrough);
7928 addSuccessorWithWeight(CurMBB, JumpMBB);
7930 // The jump table header will be inserted in our current block, do the
7931 // range check, and fall through to our fallthrough block.
7932 JTH->HeaderBB = CurMBB;
7933 JT->Default = Fallthrough; // FIXME: Move Default to JumpTableHeader.
7935 // If we're in the right place, emit the jump table header right now.
7936 if (CurMBB == SwitchMBB) {
7937 visitJumpTableHeader(*JT, *JTH, SwitchMBB);
7938 JTH->Emitted = true;
7943 // FIXME: Optimize away range check based on pivot comparisons.
7944 BitTestBlock *BTB = &BitTestCases[I->BTCasesIndex];
7946 // The bit test blocks haven't been inserted yet; insert them here.
7947 for (BitTestCase &BTC : BTB->Cases)
7948 CurMF->insert(BBI, BTC.ThisBB);
7950 // Fill in fields of the BitTestBlock.
7951 BTB->Parent = CurMBB;
7952 BTB->Default = Fallthrough;
7954 // If we're in the right place, emit the bit test header header right now.
7955 if (CurMBB ==SwitchMBB) {
7956 visitBitTestHeader(*BTB, SwitchMBB);
7957 BTB->Emitted = true;
7962 const Value *RHS, *LHS, *MHS;
7964 if (I->Low == I->High) {
7965 // Check Cond == I->Low.
7971 // Check I->Low <= Cond <= I->High.
7978 // The false weight is the sum of all unhandled cases.
7979 UnhandledWeights -= I->Weight;
7980 CaseBlock CB(CC, LHS, RHS, MHS, I->MBB, Fallthrough, CurMBB, I->Weight,
7983 if (CurMBB == SwitchMBB)
7984 visitSwitchCase(CB, SwitchMBB);
7986 SwitchCases.push_back(CB);
7991 CurMBB = Fallthrough;
7995 unsigned SelectionDAGBuilder::caseClusterRank(const CaseCluster &CC,
7996 CaseClusterIt First,
7997 CaseClusterIt Last) {
7998 return std::count_if(First, Last + 1, [&](const CaseCluster &X) {
7999 if (X.Weight != CC.Weight)
8000 return X.Weight > CC.Weight;
8002 // Ties are broken by comparing the case value.
8003 return X.Low->getValue().slt(CC.Low->getValue());
8007 void SelectionDAGBuilder::splitWorkItem(SwitchWorkList &WorkList,
8008 const SwitchWorkListItem &W,
8010 MachineBasicBlock *SwitchMBB) {
8011 assert(W.FirstCluster->Low->getValue().slt(W.LastCluster->Low->getValue()) &&
8012 "Clusters not sorted?");
8014 assert(W.LastCluster - W.FirstCluster + 1 >= 2 && "Too small to split!");
8016 // Balance the tree based on branch weights to create a near-optimal (in terms
8017 // of search time given key frequency) binary search tree. See e.g. Kurt
8018 // Mehlhorn "Nearly Optimal Binary Search Trees" (1975).
8019 CaseClusterIt LastLeft = W.FirstCluster;
8020 CaseClusterIt FirstRight = W.LastCluster;
8021 uint32_t LeftWeight = LastLeft->Weight;
8022 uint32_t RightWeight = FirstRight->Weight;
8024 // Move LastLeft and FirstRight towards each other from opposite directions to
8025 // find a partitioning of the clusters which balances the weight on both
8026 // sides. If LeftWeight and RightWeight are equal, alternate which side is
8027 // taken to ensure 0-weight nodes are distributed evenly.
8029 while (LastLeft + 1 < FirstRight) {
8030 if (LeftWeight < RightWeight || (LeftWeight == RightWeight && (I & 1)))
8031 LeftWeight += (++LastLeft)->Weight;
8033 RightWeight += (--FirstRight)->Weight;
8038 // Our binary search tree differs from a typical BST in that ours can have up
8039 // to three values in each leaf. The pivot selection above doesn't take that
8040 // into account, which means the tree might require more nodes and be less
8041 // efficient. We compensate for this here.
8043 unsigned NumLeft = LastLeft - W.FirstCluster + 1;
8044 unsigned NumRight = W.LastCluster - FirstRight + 1;
8046 if (std::min(NumLeft, NumRight) < 3 && std::max(NumLeft, NumRight) > 3) {
8047 // If one side has less than 3 clusters, and the other has more than 3,
8048 // consider taking a cluster from the other side.
8050 if (NumLeft < NumRight) {
8051 // Consider moving the first cluster on the right to the left side.
8052 CaseCluster &CC = *FirstRight;
8053 unsigned RightSideRank = caseClusterRank(CC, FirstRight, W.LastCluster);
8054 unsigned LeftSideRank = caseClusterRank(CC, W.FirstCluster, LastLeft);
8055 if (LeftSideRank <= RightSideRank) {
8056 // Moving the cluster to the left does not demote it.
8062 assert(NumRight < NumLeft);
8063 // Consider moving the last element on the left to the right side.
8064 CaseCluster &CC = *LastLeft;
8065 unsigned LeftSideRank = caseClusterRank(CC, W.FirstCluster, LastLeft);
8066 unsigned RightSideRank = caseClusterRank(CC, FirstRight, W.LastCluster);
8067 if (RightSideRank <= LeftSideRank) {
8068 // Moving the cluster to the right does not demot it.
8078 assert(LastLeft + 1 == FirstRight);
8079 assert(LastLeft >= W.FirstCluster);
8080 assert(FirstRight <= W.LastCluster);
8082 // Use the first element on the right as pivot since we will make less-than
8083 // comparisons against it.
8084 CaseClusterIt PivotCluster = FirstRight;
8085 assert(PivotCluster > W.FirstCluster);
8086 assert(PivotCluster <= W.LastCluster);
8088 CaseClusterIt FirstLeft = W.FirstCluster;
8089 CaseClusterIt LastRight = W.LastCluster;
8091 const ConstantInt *Pivot = PivotCluster->Low;
8093 // New blocks will be inserted immediately after the current one.
8094 MachineFunction::iterator BBI = W.MBB;
8097 // We will branch to the LHS if Value < Pivot. If LHS is a single cluster,
8098 // we can branch to its destination directly if it's squeezed exactly in
8099 // between the known lower bound and Pivot - 1.
8100 MachineBasicBlock *LeftMBB;
8101 if (FirstLeft == LastLeft && FirstLeft->Kind == CC_Range &&
8102 FirstLeft->Low == W.GE &&
8103 (FirstLeft->High->getValue() + 1LL) == Pivot->getValue()) {
8104 LeftMBB = FirstLeft->MBB;
8106 LeftMBB = FuncInfo.MF->CreateMachineBasicBlock(W.MBB->getBasicBlock());
8107 FuncInfo.MF->insert(BBI, LeftMBB);
8108 WorkList.push_back({LeftMBB, FirstLeft, LastLeft, W.GE, Pivot});
8109 // Put Cond in a virtual register to make it available from the new blocks.
8110 ExportFromCurrentBlock(Cond);
8113 // Similarly, we will branch to the RHS if Value >= Pivot. If RHS is a
8114 // single cluster, RHS.Low == Pivot, and we can branch to its destination
8115 // directly if RHS.High equals the current upper bound.
8116 MachineBasicBlock *RightMBB;
8117 if (FirstRight == LastRight && FirstRight->Kind == CC_Range &&
8118 W.LT && (FirstRight->High->getValue() + 1ULL) == W.LT->getValue()) {
8119 RightMBB = FirstRight->MBB;
8121 RightMBB = FuncInfo.MF->CreateMachineBasicBlock(W.MBB->getBasicBlock());
8122 FuncInfo.MF->insert(BBI, RightMBB);
8123 WorkList.push_back({RightMBB, FirstRight, LastRight, Pivot, W.LT});
8124 // Put Cond in a virtual register to make it available from the new blocks.
8125 ExportFromCurrentBlock(Cond);
8128 // Create the CaseBlock record that will be used to lower the branch.
8129 CaseBlock CB(ISD::SETLT, Cond, Pivot, nullptr, LeftMBB, RightMBB, W.MBB,
8130 LeftWeight, RightWeight);
8132 if (W.MBB == SwitchMBB)
8133 visitSwitchCase(CB, SwitchMBB);
8135 SwitchCases.push_back(CB);
8138 void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) {
8139 // Extract cases from the switch.
8140 BranchProbabilityInfo *BPI = FuncInfo.BPI;
8141 CaseClusterVector Clusters;
8142 Clusters.reserve(SI.getNumCases());
8143 for (auto I : SI.cases()) {
8144 MachineBasicBlock *Succ = FuncInfo.MBBMap[I.getCaseSuccessor()];
8145 const ConstantInt *CaseVal = I.getCaseValue();
8147 BPI ? BPI->getEdgeWeight(SI.getParent(), I.getSuccessorIndex()) : 0;
8148 Clusters.push_back(CaseCluster::range(CaseVal, CaseVal, Succ, Weight));
8151 MachineBasicBlock *DefaultMBB = FuncInfo.MBBMap[SI.getDefaultDest()];
8153 // Cluster adjacent cases with the same destination. We do this at all
8154 // optimization levels because it's cheap to do and will make codegen faster
8155 // if there are many clusters.
8156 sortAndRangeify(Clusters);
8158 if (TM.getOptLevel() != CodeGenOpt::None) {
8159 // Replace an unreachable default with the most popular destination.
8160 // FIXME: Exploit unreachable default more aggressively.
8161 bool UnreachableDefault =
8162 isa<UnreachableInst>(SI.getDefaultDest()->getFirstNonPHIOrDbg());
8163 if (UnreachableDefault && !Clusters.empty()) {
8164 DenseMap<const BasicBlock *, unsigned> Popularity;
8165 unsigned MaxPop = 0;
8166 const BasicBlock *MaxBB = nullptr;
8167 for (auto I : SI.cases()) {
8168 const BasicBlock *BB = I.getCaseSuccessor();
8169 if (++Popularity[BB] > MaxPop) {
8170 MaxPop = Popularity[BB];
8175 assert(MaxPop > 0 && MaxBB);
8176 DefaultMBB = FuncInfo.MBBMap[MaxBB];
8178 // Remove cases that were pointing to the destination that is now the
8180 CaseClusterVector New;
8181 New.reserve(Clusters.size());
8182 for (CaseCluster &CC : Clusters) {
8183 if (CC.MBB != DefaultMBB)
8186 Clusters = std::move(New);
8190 // If there is only the default destination, jump there directly.
8191 MachineBasicBlock *SwitchMBB = FuncInfo.MBB;
8192 if (Clusters.empty()) {
8193 SwitchMBB->addSuccessor(DefaultMBB);
8194 if (DefaultMBB != NextBlock(SwitchMBB)) {
8195 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other,
8196 getControlRoot(), DAG.getBasicBlock(DefaultMBB)));
8201 findJumpTables(Clusters, &SI, DefaultMBB);
8202 findBitTestClusters(Clusters, &SI);
8205 dbgs() << "Case clusters: ";
8206 for (const CaseCluster &C : Clusters) {
8207 if (C.Kind == CC_JumpTable) dbgs() << "JT:";
8208 if (C.Kind == CC_BitTests) dbgs() << "BT:";
8210 C.Low->getValue().print(dbgs(), true);
8211 if (C.Low != C.High) {
8213 C.High->getValue().print(dbgs(), true);
8220 assert(!Clusters.empty());
8221 SwitchWorkList WorkList;
8222 CaseClusterIt First = Clusters.begin();
8223 CaseClusterIt Last = Clusters.end() - 1;
8224 WorkList.push_back({SwitchMBB, First, Last, nullptr, nullptr});
8226 while (!WorkList.empty()) {
8227 SwitchWorkListItem W = WorkList.back();
8228 WorkList.pop_back();
8229 unsigned NumClusters = W.LastCluster - W.FirstCluster + 1;
8231 if (NumClusters > 3 && TM.getOptLevel() != CodeGenOpt::None) {
8232 // For optimized builds, lower large range as a balanced binary tree.
8233 splitWorkItem(WorkList, W, SI.getCondition(), SwitchMBB);
8237 lowerWorkItem(W, SI.getCondition(), SwitchMBB, DefaultMBB);