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 (DAG.getDataLayout().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 (DAG.getDataLayout().isBigEndian())
165 EVT TotalVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
166 Hi = DAG.getNode(ISD::ANY_EXTEND, DL, TotalVT, Hi);
168 DAG.getNode(ISD::SHL, DL, TotalVT, Hi,
169 DAG.getConstant(Lo.getValueType().getSizeInBits(), DL,
170 TLI.getPointerTy(DAG.getDataLayout())));
171 Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, TotalVT, Lo);
172 Val = DAG.getNode(ISD::OR, DL, TotalVT, Lo, Hi);
174 } else if (PartVT.isFloatingPoint()) {
175 // FP split into multiple FP parts (for ppcf128)
176 assert(ValueVT == EVT(MVT::ppcf128) && PartVT == MVT::f64 &&
179 Lo = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[0]);
180 Hi = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[1]);
181 if (TLI.hasBigEndianPartOrdering(ValueVT, DAG.getDataLayout()))
183 Val = DAG.getNode(ISD::BUILD_PAIR, DL, ValueVT, Lo, Hi);
185 // FP split into integer parts (soft fp)
186 assert(ValueVT.isFloatingPoint() && PartVT.isInteger() &&
187 !PartVT.isVector() && "Unexpected split");
188 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits());
189 Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT, V);
193 // There is now one part, held in Val. Correct it to match ValueVT.
194 EVT PartEVT = Val.getValueType();
196 if (PartEVT == ValueVT)
199 if (PartEVT.isInteger() && ValueVT.isInteger()) {
200 if (ValueVT.bitsLT(PartEVT)) {
201 // For a truncate, see if we have any information to
202 // indicate whether the truncated bits will always be
203 // zero or sign-extension.
204 if (AssertOp != ISD::DELETED_NODE)
205 Val = DAG.getNode(AssertOp, DL, PartEVT, Val,
206 DAG.getValueType(ValueVT));
207 return DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
209 return DAG.getNode(ISD::ANY_EXTEND, DL, ValueVT, Val);
212 if (PartEVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
213 // FP_ROUND's are always exact here.
214 if (ValueVT.bitsLT(Val.getValueType()))
216 ISD::FP_ROUND, DL, ValueVT, Val,
217 DAG.getTargetConstant(1, DL, TLI.getPointerTy(DAG.getDataLayout())));
219 return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val);
222 if (PartEVT.getSizeInBits() == ValueVT.getSizeInBits())
223 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
225 llvm_unreachable("Unknown mismatch!");
228 static void diagnosePossiblyInvalidConstraint(LLVMContext &Ctx, const Value *V,
229 const Twine &ErrMsg) {
230 const Instruction *I = dyn_cast_or_null<Instruction>(V);
232 return Ctx.emitError(ErrMsg);
234 const char *AsmError = ", possible invalid constraint for vector type";
235 if (const CallInst *CI = dyn_cast<CallInst>(I))
236 if (isa<InlineAsm>(CI->getCalledValue()))
237 return Ctx.emitError(I, ErrMsg + AsmError);
239 return Ctx.emitError(I, ErrMsg);
242 /// getCopyFromPartsVector - Create a value that contains the specified legal
243 /// parts combined into the value they represent. If the parts combine to a
244 /// type larger then ValueVT then AssertOp can be used to specify whether the
245 /// extra bits are known to be zero (ISD::AssertZext) or sign extended from
246 /// ValueVT (ISD::AssertSext).
247 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, SDLoc DL,
248 const SDValue *Parts, unsigned NumParts,
249 MVT PartVT, EVT ValueVT, const Value *V) {
250 assert(ValueVT.isVector() && "Not a vector value");
251 assert(NumParts > 0 && "No parts to assemble!");
252 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
253 SDValue Val = Parts[0];
255 // Handle a multi-element vector.
259 unsigned NumIntermediates;
261 TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT,
262 NumIntermediates, RegisterVT);
263 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
264 NumParts = NumRegs; // Silence a compiler warning.
265 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
266 assert(RegisterVT.getSizeInBits() ==
267 Parts[0].getSimpleValueType().getSizeInBits() &&
268 "Part type sizes don't match!");
270 // Assemble the parts into intermediate operands.
271 SmallVector<SDValue, 8> Ops(NumIntermediates);
272 if (NumIntermediates == NumParts) {
273 // If the register was not expanded, truncate or copy the value,
275 for (unsigned i = 0; i != NumParts; ++i)
276 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i], 1,
277 PartVT, IntermediateVT, V);
278 } else if (NumParts > 0) {
279 // If the intermediate type was expanded, build the intermediate
280 // operands from the parts.
281 assert(NumParts % NumIntermediates == 0 &&
282 "Must expand into a divisible number of parts!");
283 unsigned Factor = NumParts / NumIntermediates;
284 for (unsigned i = 0; i != NumIntermediates; ++i)
285 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor,
286 PartVT, IntermediateVT, V);
289 // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the
290 // intermediate operands.
291 Val = DAG.getNode(IntermediateVT.isVector() ? ISD::CONCAT_VECTORS
296 // There is now one part, held in Val. Correct it to match ValueVT.
297 EVT PartEVT = Val.getValueType();
299 if (PartEVT == ValueVT)
302 if (PartEVT.isVector()) {
303 // If the element type of the source/dest vectors are the same, but the
304 // parts vector has more elements than the value vector, then we have a
305 // vector widening case (e.g. <2 x float> -> <4 x float>). Extract the
307 if (PartEVT.getVectorElementType() == ValueVT.getVectorElementType()) {
308 assert(PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements() &&
309 "Cannot narrow, it would be a lossy transformation");
311 ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val,
312 DAG.getConstant(0, DL, TLI.getVectorIdxTy(DAG.getDataLayout())));
315 // Vector/Vector bitcast.
316 if (ValueVT.getSizeInBits() == PartEVT.getSizeInBits())
317 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
319 assert(PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements() &&
320 "Cannot handle this kind of promotion");
321 // Promoted vector extract
322 bool Smaller = ValueVT.bitsLE(PartEVT);
323 return DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
328 // Trivial bitcast if the types are the same size and the destination
329 // vector type is legal.
330 if (PartEVT.getSizeInBits() == ValueVT.getSizeInBits() &&
331 TLI.isTypeLegal(ValueVT))
332 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
334 // Handle cases such as i8 -> <1 x i1>
335 if (ValueVT.getVectorNumElements() != 1) {
336 diagnosePossiblyInvalidConstraint(*DAG.getContext(), V,
337 "non-trivial scalar-to-vector conversion");
338 return DAG.getUNDEF(ValueVT);
341 if (ValueVT.getVectorNumElements() == 1 &&
342 ValueVT.getVectorElementType() != PartEVT) {
343 bool Smaller = ValueVT.bitsLE(PartEVT);
344 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
345 DL, ValueVT.getScalarType(), Val);
348 return DAG.getNode(ISD::BUILD_VECTOR, DL, ValueVT, Val);
351 static void getCopyToPartsVector(SelectionDAG &DAG, SDLoc dl,
352 SDValue Val, SDValue *Parts, unsigned NumParts,
353 MVT PartVT, const Value *V);
355 /// getCopyToParts - Create a series of nodes that contain the specified value
356 /// split into legal parts. If the parts contain more bits than Val, then, for
357 /// integers, ExtendKind can be used to specify how to generate the extra bits.
358 static void getCopyToParts(SelectionDAG &DAG, SDLoc DL,
359 SDValue Val, SDValue *Parts, unsigned NumParts,
360 MVT PartVT, const Value *V,
361 ISD::NodeType ExtendKind = ISD::ANY_EXTEND) {
362 EVT ValueVT = Val.getValueType();
364 // Handle the vector case separately.
365 if (ValueVT.isVector())
366 return getCopyToPartsVector(DAG, DL, Val, Parts, NumParts, PartVT, V);
368 unsigned PartBits = PartVT.getSizeInBits();
369 unsigned OrigNumParts = NumParts;
370 assert(DAG.getTargetLoweringInfo().isTypeLegal(PartVT) &&
371 "Copying to an illegal type!");
376 assert(!ValueVT.isVector() && "Vector case handled elsewhere");
377 EVT PartEVT = PartVT;
378 if (PartEVT == ValueVT) {
379 assert(NumParts == 1 && "No-op copy with multiple parts!");
384 if (NumParts * PartBits > ValueVT.getSizeInBits()) {
385 // If the parts cover more bits than the value has, promote the value.
386 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
387 assert(NumParts == 1 && "Do not know what to promote to!");
388 Val = DAG.getNode(ISD::FP_EXTEND, DL, PartVT, Val);
390 assert((PartVT.isInteger() || PartVT == MVT::x86mmx) &&
391 ValueVT.isInteger() &&
392 "Unknown mismatch!");
393 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
394 Val = DAG.getNode(ExtendKind, DL, ValueVT, Val);
395 if (PartVT == MVT::x86mmx)
396 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
398 } else if (PartBits == ValueVT.getSizeInBits()) {
399 // Different types of the same size.
400 assert(NumParts == 1 && PartEVT != ValueVT);
401 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
402 } else if (NumParts * PartBits < ValueVT.getSizeInBits()) {
403 // If the parts cover less bits than value has, truncate the value.
404 assert((PartVT.isInteger() || PartVT == MVT::x86mmx) &&
405 ValueVT.isInteger() &&
406 "Unknown mismatch!");
407 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
408 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
409 if (PartVT == MVT::x86mmx)
410 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
413 // The value may have changed - recompute ValueVT.
414 ValueVT = Val.getValueType();
415 assert(NumParts * PartBits == ValueVT.getSizeInBits() &&
416 "Failed to tile the value with PartVT!");
419 if (PartEVT != ValueVT)
420 diagnosePossiblyInvalidConstraint(*DAG.getContext(), V,
421 "scalar-to-vector conversion failed");
427 // Expand the value into multiple parts.
428 if (NumParts & (NumParts - 1)) {
429 // The number of parts is not a power of 2. Split off and copy the tail.
430 assert(PartVT.isInteger() && ValueVT.isInteger() &&
431 "Do not know what to expand to!");
432 unsigned RoundParts = 1 << Log2_32(NumParts);
433 unsigned RoundBits = RoundParts * PartBits;
434 unsigned OddParts = NumParts - RoundParts;
435 SDValue OddVal = DAG.getNode(ISD::SRL, DL, ValueVT, Val,
436 DAG.getIntPtrConstant(RoundBits, DL));
437 getCopyToParts(DAG, DL, OddVal, Parts + RoundParts, OddParts, PartVT, V);
439 if (DAG.getDataLayout().isBigEndian())
440 // The odd parts were reversed by getCopyToParts - unreverse them.
441 std::reverse(Parts + RoundParts, Parts + NumParts);
443 NumParts = RoundParts;
444 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
445 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
448 // The number of parts is a power of 2. Repeatedly bisect the value using
450 Parts[0] = DAG.getNode(ISD::BITCAST, DL,
451 EVT::getIntegerVT(*DAG.getContext(),
452 ValueVT.getSizeInBits()),
455 for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) {
456 for (unsigned i = 0; i < NumParts; i += StepSize) {
457 unsigned ThisBits = StepSize * PartBits / 2;
458 EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits);
459 SDValue &Part0 = Parts[i];
460 SDValue &Part1 = Parts[i+StepSize/2];
462 Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
463 ThisVT, Part0, DAG.getIntPtrConstant(1, DL));
464 Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
465 ThisVT, Part0, DAG.getIntPtrConstant(0, DL));
467 if (ThisBits == PartBits && ThisVT != PartVT) {
468 Part0 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part0);
469 Part1 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part1);
474 if (DAG.getDataLayout().isBigEndian())
475 std::reverse(Parts, Parts + OrigNumParts);
479 /// getCopyToPartsVector - Create a series of nodes that contain the specified
480 /// value split into legal parts.
481 static void getCopyToPartsVector(SelectionDAG &DAG, SDLoc DL,
482 SDValue Val, SDValue *Parts, unsigned NumParts,
483 MVT PartVT, const Value *V) {
484 EVT ValueVT = Val.getValueType();
485 assert(ValueVT.isVector() && "Not a vector");
486 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
489 EVT PartEVT = PartVT;
490 if (PartEVT == ValueVT) {
492 } else if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) {
493 // Bitconvert vector->vector case.
494 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
495 } else if (PartVT.isVector() &&
496 PartEVT.getVectorElementType() == ValueVT.getVectorElementType() &&
497 PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements()) {
498 EVT ElementVT = PartVT.getVectorElementType();
499 // Vector widening case, e.g. <2 x float> -> <4 x float>. Shuffle in
501 SmallVector<SDValue, 16> Ops;
502 for (unsigned i = 0, e = ValueVT.getVectorNumElements(); i != e; ++i)
503 Ops.push_back(DAG.getNode(
504 ISD::EXTRACT_VECTOR_ELT, DL, ElementVT, Val,
505 DAG.getConstant(i, DL, TLI.getVectorIdxTy(DAG.getDataLayout()))));
507 for (unsigned i = ValueVT.getVectorNumElements(),
508 e = PartVT.getVectorNumElements(); i != e; ++i)
509 Ops.push_back(DAG.getUNDEF(ElementVT));
511 Val = DAG.getNode(ISD::BUILD_VECTOR, DL, PartVT, Ops);
513 // FIXME: Use CONCAT for 2x -> 4x.
515 //SDValue UndefElts = DAG.getUNDEF(VectorTy);
516 //Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts);
517 } else if (PartVT.isVector() &&
518 PartEVT.getVectorElementType().bitsGE(
519 ValueVT.getVectorElementType()) &&
520 PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements()) {
522 // Promoted vector extract
523 bool Smaller = PartEVT.bitsLE(ValueVT);
524 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
527 // Vector -> scalar conversion.
528 assert(ValueVT.getVectorNumElements() == 1 &&
529 "Only trivial vector-to-scalar conversions should get here!");
531 ISD::EXTRACT_VECTOR_ELT, DL, PartVT, Val,
532 DAG.getConstant(0, DL, TLI.getVectorIdxTy(DAG.getDataLayout())));
534 bool Smaller = ValueVT.bitsLE(PartVT);
535 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
543 // Handle a multi-element vector.
546 unsigned NumIntermediates;
547 unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT,
549 NumIntermediates, RegisterVT);
550 unsigned NumElements = ValueVT.getVectorNumElements();
552 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
553 NumParts = NumRegs; // Silence a compiler warning.
554 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
556 // Split the vector into intermediate operands.
557 SmallVector<SDValue, 8> Ops(NumIntermediates);
558 for (unsigned i = 0; i != NumIntermediates; ++i) {
559 if (IntermediateVT.isVector())
561 DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, IntermediateVT, Val,
562 DAG.getConstant(i * (NumElements / NumIntermediates), DL,
563 TLI.getVectorIdxTy(DAG.getDataLayout())));
565 Ops[i] = DAG.getNode(
566 ISD::EXTRACT_VECTOR_ELT, DL, IntermediateVT, Val,
567 DAG.getConstant(i, DL, TLI.getVectorIdxTy(DAG.getDataLayout())));
570 // Split the intermediate operands into legal parts.
571 if (NumParts == NumIntermediates) {
572 // If the register was not expanded, promote or copy the value,
574 for (unsigned i = 0; i != NumParts; ++i)
575 getCopyToParts(DAG, DL, Ops[i], &Parts[i], 1, PartVT, V);
576 } else if (NumParts > 0) {
577 // If the intermediate type was expanded, split each the value into
579 assert(NumIntermediates != 0 && "division by zero");
580 assert(NumParts % NumIntermediates == 0 &&
581 "Must expand into a divisible number of parts!");
582 unsigned Factor = NumParts / NumIntermediates;
583 for (unsigned i = 0; i != NumIntermediates; ++i)
584 getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT, V);
588 RegsForValue::RegsForValue() {}
590 RegsForValue::RegsForValue(const SmallVector<unsigned, 4> ®s, MVT regvt,
592 : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {}
594 RegsForValue::RegsForValue(LLVMContext &Context, const TargetLowering &TLI,
595 const DataLayout &DL, unsigned Reg, Type *Ty) {
596 ComputeValueVTs(TLI, DL, Ty, ValueVTs);
598 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
599 EVT ValueVT = ValueVTs[Value];
600 unsigned NumRegs = TLI.getNumRegisters(Context, ValueVT);
601 MVT RegisterVT = TLI.getRegisterType(Context, ValueVT);
602 for (unsigned i = 0; i != NumRegs; ++i)
603 Regs.push_back(Reg + i);
604 RegVTs.push_back(RegisterVT);
609 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
610 /// this value and returns the result as a ValueVT value. This uses
611 /// Chain/Flag as the input and updates them for the output Chain/Flag.
612 /// If the Flag pointer is NULL, no flag is used.
613 SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
614 FunctionLoweringInfo &FuncInfo,
616 SDValue &Chain, SDValue *Flag,
617 const Value *V) const {
618 // A Value with type {} or [0 x %t] needs no registers.
619 if (ValueVTs.empty())
622 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
624 // Assemble the legal parts into the final values.
625 SmallVector<SDValue, 4> Values(ValueVTs.size());
626 SmallVector<SDValue, 8> Parts;
627 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
628 // Copy the legal parts from the registers.
629 EVT ValueVT = ValueVTs[Value];
630 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
631 MVT RegisterVT = RegVTs[Value];
633 Parts.resize(NumRegs);
634 for (unsigned i = 0; i != NumRegs; ++i) {
637 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT);
639 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag);
640 *Flag = P.getValue(2);
643 Chain = P.getValue(1);
646 // If the source register was virtual and if we know something about it,
647 // add an assert node.
648 if (!TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) ||
649 !RegisterVT.isInteger() || RegisterVT.isVector())
652 const FunctionLoweringInfo::LiveOutInfo *LOI =
653 FuncInfo.GetLiveOutRegInfo(Regs[Part+i]);
657 unsigned RegSize = RegisterVT.getSizeInBits();
658 unsigned NumSignBits = LOI->NumSignBits;
659 unsigned NumZeroBits = LOI->KnownZero.countLeadingOnes();
661 if (NumZeroBits == RegSize) {
662 // The current value is a zero.
663 // Explicitly express that as it would be easier for
664 // optimizations to kick in.
665 Parts[i] = DAG.getConstant(0, dl, RegisterVT);
669 // FIXME: We capture more information than the dag can represent. For
670 // now, just use the tightest assertzext/assertsext possible.
672 EVT FromVT(MVT::Other);
673 if (NumSignBits == RegSize)
674 isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1
675 else if (NumZeroBits >= RegSize-1)
676 isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1
677 else if (NumSignBits > RegSize-8)
678 isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8
679 else if (NumZeroBits >= RegSize-8)
680 isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8
681 else if (NumSignBits > RegSize-16)
682 isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16
683 else if (NumZeroBits >= RegSize-16)
684 isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16
685 else if (NumSignBits > RegSize-32)
686 isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32
687 else if (NumZeroBits >= RegSize-32)
688 isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32
692 // Add an assertion node.
693 assert(FromVT != MVT::Other);
694 Parts[i] = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl,
695 RegisterVT, P, DAG.getValueType(FromVT));
698 Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(),
699 NumRegs, RegisterVT, ValueVT, V);
704 return DAG.getNode(ISD::MERGE_VALUES, dl, DAG.getVTList(ValueVTs), Values);
707 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
708 /// specified value into the registers specified by this object. This uses
709 /// Chain/Flag as the input and updates them for the output Chain/Flag.
710 /// If the Flag pointer is NULL, no flag is used.
711 void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, SDLoc dl,
712 SDValue &Chain, SDValue *Flag, const Value *V,
713 ISD::NodeType PreferredExtendType) const {
714 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
715 ISD::NodeType ExtendKind = PreferredExtendType;
717 // Get the list of the values's legal parts.
718 unsigned NumRegs = Regs.size();
719 SmallVector<SDValue, 8> Parts(NumRegs);
720 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
721 EVT ValueVT = ValueVTs[Value];
722 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
723 MVT RegisterVT = RegVTs[Value];
725 if (ExtendKind == ISD::ANY_EXTEND && TLI.isZExtFree(Val, RegisterVT))
726 ExtendKind = ISD::ZERO_EXTEND;
728 getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value),
729 &Parts[Part], NumParts, RegisterVT, V, ExtendKind);
733 // Copy the parts into the registers.
734 SmallVector<SDValue, 8> Chains(NumRegs);
735 for (unsigned i = 0; i != NumRegs; ++i) {
738 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]);
740 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag);
741 *Flag = Part.getValue(1);
744 Chains[i] = Part.getValue(0);
747 if (NumRegs == 1 || Flag)
748 // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is
749 // flagged to it. That is the CopyToReg nodes and the user are considered
750 // a single scheduling unit. If we create a TokenFactor and return it as
751 // chain, then the TokenFactor is both a predecessor (operand) of the
752 // user as well as a successor (the TF operands are flagged to the user).
753 // c1, f1 = CopyToReg
754 // c2, f2 = CopyToReg
755 // c3 = TokenFactor c1, c2
758 Chain = Chains[NumRegs-1];
760 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
763 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
764 /// operand list. This adds the code marker and includes the number of
765 /// values added into it.
766 void RegsForValue::AddInlineAsmOperands(unsigned Code, bool HasMatching,
767 unsigned MatchingIdx, SDLoc dl,
769 std::vector<SDValue> &Ops) const {
770 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
772 unsigned Flag = InlineAsm::getFlagWord(Code, Regs.size());
774 Flag = InlineAsm::getFlagWordForMatchingOp(Flag, MatchingIdx);
775 else if (!Regs.empty() &&
776 TargetRegisterInfo::isVirtualRegister(Regs.front())) {
777 // Put the register class of the virtual registers in the flag word. That
778 // way, later passes can recompute register class constraints for inline
779 // assembly as well as normal instructions.
780 // Don't do this for tied operands that can use the regclass information
782 const MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
783 const TargetRegisterClass *RC = MRI.getRegClass(Regs.front());
784 Flag = InlineAsm::getFlagWordForRegClass(Flag, RC->getID());
787 SDValue Res = DAG.getTargetConstant(Flag, dl, MVT::i32);
790 unsigned SP = TLI.getStackPointerRegisterToSaveRestore();
791 for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) {
792 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVTs[Value]);
793 MVT RegisterVT = RegVTs[Value];
794 for (unsigned i = 0; i != NumRegs; ++i) {
795 assert(Reg < Regs.size() && "Mismatch in # registers expected");
796 unsigned TheReg = Regs[Reg++];
797 Ops.push_back(DAG.getRegister(TheReg, RegisterVT));
799 if (TheReg == SP && Code == InlineAsm::Kind_Clobber) {
800 // If we clobbered the stack pointer, MFI should know about it.
801 assert(DAG.getMachineFunction().getFrameInfo()->
802 hasOpaqueSPAdjustment());
808 void SelectionDAGBuilder::init(GCFunctionInfo *gfi, AliasAnalysis &aa,
809 const TargetLibraryInfo *li) {
813 DL = &DAG.getDataLayout();
814 Context = DAG.getContext();
815 LPadToCallSiteMap.clear();
818 /// clear - Clear out the current SelectionDAG and the associated
819 /// state and prepare this SelectionDAGBuilder object to be used
820 /// for a new block. This doesn't clear out information about
821 /// additional blocks that are needed to complete switch lowering
822 /// or PHI node updating; that information is cleared out as it is
824 void SelectionDAGBuilder::clear() {
826 UnusedArgNodeMap.clear();
827 PendingLoads.clear();
828 PendingExports.clear();
831 SDNodeOrder = LowestSDNodeOrder;
832 StatepointLowering.clear();
835 /// clearDanglingDebugInfo - Clear the dangling debug information
836 /// map. This function is separated from the clear so that debug
837 /// information that is dangling in a basic block can be properly
838 /// resolved in a different basic block. This allows the
839 /// SelectionDAG to resolve dangling debug information attached
841 void SelectionDAGBuilder::clearDanglingDebugInfo() {
842 DanglingDebugInfoMap.clear();
845 /// getRoot - Return the current virtual root of the Selection DAG,
846 /// flushing any PendingLoad items. This must be done before emitting
847 /// a store or any other node that may need to be ordered after any
848 /// prior load instructions.
850 SDValue SelectionDAGBuilder::getRoot() {
851 if (PendingLoads.empty())
852 return DAG.getRoot();
854 if (PendingLoads.size() == 1) {
855 SDValue Root = PendingLoads[0];
857 PendingLoads.clear();
861 // Otherwise, we have to make a token factor node.
862 SDValue Root = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
864 PendingLoads.clear();
869 /// getControlRoot - Similar to getRoot, but instead of flushing all the
870 /// PendingLoad items, flush all the PendingExports items. It is necessary
871 /// to do this before emitting a terminator instruction.
873 SDValue SelectionDAGBuilder::getControlRoot() {
874 SDValue Root = DAG.getRoot();
876 if (PendingExports.empty())
879 // Turn all of the CopyToReg chains into one factored node.
880 if (Root.getOpcode() != ISD::EntryToken) {
881 unsigned i = 0, e = PendingExports.size();
882 for (; i != e; ++i) {
883 assert(PendingExports[i].getNode()->getNumOperands() > 1);
884 if (PendingExports[i].getNode()->getOperand(0) == Root)
885 break; // Don't add the root if we already indirectly depend on it.
889 PendingExports.push_back(Root);
892 Root = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
894 PendingExports.clear();
899 void SelectionDAGBuilder::visit(const Instruction &I) {
900 // Set up outgoing PHI node register values before emitting the terminator.
901 if (isa<TerminatorInst>(&I))
902 HandlePHINodesInSuccessorBlocks(I.getParent());
908 visit(I.getOpcode(), I);
910 if (!isa<TerminatorInst>(&I) && !HasTailCall)
911 CopyToExportRegsIfNeeded(&I);
916 void SelectionDAGBuilder::visitPHI(const PHINode &) {
917 llvm_unreachable("SelectionDAGBuilder shouldn't visit PHI nodes!");
920 void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) {
921 // Note: this doesn't use InstVisitor, because it has to work with
922 // ConstantExpr's in addition to instructions.
924 default: llvm_unreachable("Unknown instruction type encountered!");
925 // Build the switch statement using the Instruction.def file.
926 #define HANDLE_INST(NUM, OPCODE, CLASS) \
927 case Instruction::OPCODE: visit##OPCODE((const CLASS&)I); break;
928 #include "llvm/IR/Instruction.def"
932 // resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V,
933 // generate the debug data structures now that we've seen its definition.
934 void SelectionDAGBuilder::resolveDanglingDebugInfo(const Value *V,
936 DanglingDebugInfo &DDI = DanglingDebugInfoMap[V];
938 const DbgValueInst *DI = DDI.getDI();
939 DebugLoc dl = DDI.getdl();
940 unsigned DbgSDNodeOrder = DDI.getSDNodeOrder();
941 DILocalVariable *Variable = DI->getVariable();
942 DIExpression *Expr = DI->getExpression();
943 assert(Variable->isValidLocationForIntrinsic(dl) &&
944 "Expected inlined-at fields to agree");
945 uint64_t Offset = DI->getOffset();
946 // A dbg.value for an alloca is always indirect.
947 bool IsIndirect = isa<AllocaInst>(V) || Offset != 0;
950 if (!EmitFuncArgumentDbgValue(V, Variable, Expr, dl, Offset, IsIndirect,
952 SDV = DAG.getDbgValue(Variable, Expr, Val.getNode(), Val.getResNo(),
953 IsIndirect, Offset, dl, DbgSDNodeOrder);
954 DAG.AddDbgValue(SDV, Val.getNode(), false);
957 DEBUG(dbgs() << "Dropping debug info for " << *DI << "\n");
958 DanglingDebugInfoMap[V] = DanglingDebugInfo();
962 /// getCopyFromRegs - If there was virtual register allocated for the value V
963 /// emit CopyFromReg of the specified type Ty. Return empty SDValue() otherwise.
964 SDValue SelectionDAGBuilder::getCopyFromRegs(const Value *V, Type *Ty) {
965 DenseMap<const Value *, unsigned>::iterator It = FuncInfo.ValueMap.find(V);
968 if (It != FuncInfo.ValueMap.end()) {
969 unsigned InReg = It->second;
970 RegsForValue RFV(*DAG.getContext(), DAG.getTargetLoweringInfo(),
971 DAG.getDataLayout(), InReg, Ty);
972 SDValue Chain = DAG.getEntryNode();
973 Result = RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, nullptr, V);
974 resolveDanglingDebugInfo(V, Result);
980 /// getValue - Return an SDValue for the given Value.
981 SDValue SelectionDAGBuilder::getValue(const Value *V) {
982 // If we already have an SDValue for this value, use it. It's important
983 // to do this first, so that we don't create a CopyFromReg if we already
984 // have a regular SDValue.
985 SDValue &N = NodeMap[V];
986 if (N.getNode()) return N;
988 // If there's a virtual register allocated and initialized for this
990 SDValue copyFromReg = getCopyFromRegs(V, V->getType());
991 if (copyFromReg.getNode()) {
995 // Otherwise create a new SDValue and remember it.
996 SDValue Val = getValueImpl(V);
998 resolveDanglingDebugInfo(V, Val);
1002 // Return true if SDValue exists for the given Value
1003 bool SelectionDAGBuilder::findValue(const Value *V) const {
1004 return (NodeMap.find(V) != NodeMap.end()) ||
1005 (FuncInfo.ValueMap.find(V) != FuncInfo.ValueMap.end());
1008 /// getNonRegisterValue - Return an SDValue for the given Value, but
1009 /// don't look in FuncInfo.ValueMap for a virtual register.
1010 SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) {
1011 // If we already have an SDValue for this value, use it.
1012 SDValue &N = NodeMap[V];
1014 if (isa<ConstantSDNode>(N) || isa<ConstantFPSDNode>(N)) {
1015 // Remove the debug location from the node as the node is about to be used
1016 // in a location which may differ from the original debug location. This
1017 // is relevant to Constant and ConstantFP nodes because they can appear
1018 // as constant expressions inside PHI nodes.
1019 N->setDebugLoc(DebugLoc());
1024 // Otherwise create a new SDValue and remember it.
1025 SDValue Val = getValueImpl(V);
1027 resolveDanglingDebugInfo(V, Val);
1031 /// getValueImpl - Helper function for getValue and getNonRegisterValue.
1032 /// Create an SDValue for the given value.
1033 SDValue SelectionDAGBuilder::getValueImpl(const Value *V) {
1034 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1036 if (const Constant *C = dyn_cast<Constant>(V)) {
1037 EVT VT = TLI.getValueType(DAG.getDataLayout(), V->getType(), true);
1039 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C))
1040 return DAG.getConstant(*CI, getCurSDLoc(), VT);
1042 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
1043 return DAG.getGlobalAddress(GV, getCurSDLoc(), VT);
1045 if (isa<ConstantPointerNull>(C)) {
1046 unsigned AS = V->getType()->getPointerAddressSpace();
1047 return DAG.getConstant(0, getCurSDLoc(),
1048 TLI.getPointerTy(DAG.getDataLayout(), AS));
1051 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
1052 return DAG.getConstantFP(*CFP, getCurSDLoc(), VT);
1054 if (isa<UndefValue>(C) && !V->getType()->isAggregateType())
1055 return DAG.getUNDEF(VT);
1057 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
1058 visit(CE->getOpcode(), *CE);
1059 SDValue N1 = NodeMap[V];
1060 assert(N1.getNode() && "visit didn't populate the NodeMap!");
1064 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) {
1065 SmallVector<SDValue, 4> Constants;
1066 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end();
1068 SDNode *Val = getValue(*OI).getNode();
1069 // If the operand is an empty aggregate, there are no values.
1071 // Add each leaf value from the operand to the Constants list
1072 // to form a flattened list of all the values.
1073 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
1074 Constants.push_back(SDValue(Val, i));
1077 return DAG.getMergeValues(Constants, getCurSDLoc());
1080 if (const ConstantDataSequential *CDS =
1081 dyn_cast<ConstantDataSequential>(C)) {
1082 SmallVector<SDValue, 4> Ops;
1083 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1084 SDNode *Val = getValue(CDS->getElementAsConstant(i)).getNode();
1085 // Add each leaf value from the operand to the Constants list
1086 // to form a flattened list of all the values.
1087 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
1088 Ops.push_back(SDValue(Val, i));
1091 if (isa<ArrayType>(CDS->getType()))
1092 return DAG.getMergeValues(Ops, getCurSDLoc());
1093 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(),
1097 if (C->getType()->isStructTy() || C->getType()->isArrayTy()) {
1098 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
1099 "Unknown struct or array constant!");
1101 SmallVector<EVT, 4> ValueVTs;
1102 ComputeValueVTs(TLI, DAG.getDataLayout(), C->getType(), ValueVTs);
1103 unsigned NumElts = ValueVTs.size();
1105 return SDValue(); // empty struct
1106 SmallVector<SDValue, 4> Constants(NumElts);
1107 for (unsigned i = 0; i != NumElts; ++i) {
1108 EVT EltVT = ValueVTs[i];
1109 if (isa<UndefValue>(C))
1110 Constants[i] = DAG.getUNDEF(EltVT);
1111 else if (EltVT.isFloatingPoint())
1112 Constants[i] = DAG.getConstantFP(0, getCurSDLoc(), EltVT);
1114 Constants[i] = DAG.getConstant(0, getCurSDLoc(), EltVT);
1117 return DAG.getMergeValues(Constants, getCurSDLoc());
1120 if (const BlockAddress *BA = dyn_cast<BlockAddress>(C))
1121 return DAG.getBlockAddress(BA, VT);
1123 VectorType *VecTy = cast<VectorType>(V->getType());
1124 unsigned NumElements = VecTy->getNumElements();
1126 // Now that we know the number and type of the elements, get that number of
1127 // elements into the Ops array based on what kind of constant it is.
1128 SmallVector<SDValue, 16> Ops;
1129 if (const ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
1130 for (unsigned i = 0; i != NumElements; ++i)
1131 Ops.push_back(getValue(CV->getOperand(i)));
1133 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!");
1135 TLI.getValueType(DAG.getDataLayout(), VecTy->getElementType());
1138 if (EltVT.isFloatingPoint())
1139 Op = DAG.getConstantFP(0, getCurSDLoc(), EltVT);
1141 Op = DAG.getConstant(0, getCurSDLoc(), EltVT);
1142 Ops.assign(NumElements, Op);
1145 // Create a BUILD_VECTOR node.
1146 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(), VT, Ops);
1149 // If this is a static alloca, generate it as the frameindex instead of
1151 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
1152 DenseMap<const AllocaInst*, int>::iterator SI =
1153 FuncInfo.StaticAllocaMap.find(AI);
1154 if (SI != FuncInfo.StaticAllocaMap.end())
1155 return DAG.getFrameIndex(SI->second,
1156 TLI.getPointerTy(DAG.getDataLayout()));
1159 // If this is an instruction which fast-isel has deferred, select it now.
1160 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
1161 unsigned InReg = FuncInfo.InitializeRegForValue(Inst);
1162 RegsForValue RFV(*DAG.getContext(), TLI, DAG.getDataLayout(), InReg,
1164 SDValue Chain = DAG.getEntryNode();
1165 return RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, nullptr, V);
1168 llvm_unreachable("Can't get register for value!");
1171 void SelectionDAGBuilder::visitRet(const ReturnInst &I) {
1172 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1173 auto &DL = DAG.getDataLayout();
1174 SDValue Chain = getControlRoot();
1175 SmallVector<ISD::OutputArg, 8> Outs;
1176 SmallVector<SDValue, 8> OutVals;
1178 if (!FuncInfo.CanLowerReturn) {
1179 unsigned DemoteReg = FuncInfo.DemoteRegister;
1180 const Function *F = I.getParent()->getParent();
1182 // Emit a store of the return value through the virtual register.
1183 // Leave Outs empty so that LowerReturn won't try to load return
1184 // registers the usual way.
1185 SmallVector<EVT, 1> PtrValueVTs;
1186 ComputeValueVTs(TLI, DL, PointerType::getUnqual(F->getReturnType()),
1189 SDValue RetPtr = DAG.getRegister(DemoteReg, PtrValueVTs[0]);
1190 SDValue RetOp = getValue(I.getOperand(0));
1192 SmallVector<EVT, 4> ValueVTs;
1193 SmallVector<uint64_t, 4> Offsets;
1194 ComputeValueVTs(TLI, DL, I.getOperand(0)->getType(), ValueVTs, &Offsets);
1195 unsigned NumValues = ValueVTs.size();
1197 SmallVector<SDValue, 4> Chains(NumValues);
1198 for (unsigned i = 0; i != NumValues; ++i) {
1199 SDValue Add = DAG.getNode(ISD::ADD, getCurSDLoc(),
1200 RetPtr.getValueType(), RetPtr,
1201 DAG.getIntPtrConstant(Offsets[i],
1204 DAG.getStore(Chain, getCurSDLoc(),
1205 SDValue(RetOp.getNode(), RetOp.getResNo() + i),
1206 // FIXME: better loc info would be nice.
1207 Add, MachinePointerInfo(), false, false, 0);
1210 Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(),
1211 MVT::Other, Chains);
1212 } else if (I.getNumOperands() != 0) {
1213 SmallVector<EVT, 4> ValueVTs;
1214 ComputeValueVTs(TLI, DL, I.getOperand(0)->getType(), ValueVTs);
1215 unsigned NumValues = ValueVTs.size();
1217 SDValue RetOp = getValue(I.getOperand(0));
1219 const Function *F = I.getParent()->getParent();
1221 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1222 if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1224 ExtendKind = ISD::SIGN_EXTEND;
1225 else if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1227 ExtendKind = ISD::ZERO_EXTEND;
1229 LLVMContext &Context = F->getContext();
1230 bool RetInReg = F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1233 for (unsigned j = 0; j != NumValues; ++j) {
1234 EVT VT = ValueVTs[j];
1236 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger())
1237 VT = TLI.getTypeForExtArgOrReturn(Context, VT, ExtendKind);
1239 unsigned NumParts = TLI.getNumRegisters(Context, VT);
1240 MVT PartVT = TLI.getRegisterType(Context, VT);
1241 SmallVector<SDValue, 4> Parts(NumParts);
1242 getCopyToParts(DAG, getCurSDLoc(),
1243 SDValue(RetOp.getNode(), RetOp.getResNo() + j),
1244 &Parts[0], NumParts, PartVT, &I, ExtendKind);
1246 // 'inreg' on function refers to return value
1247 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1251 // Propagate extension type if any
1252 if (ExtendKind == ISD::SIGN_EXTEND)
1254 else if (ExtendKind == ISD::ZERO_EXTEND)
1257 for (unsigned i = 0; i < NumParts; ++i) {
1258 Outs.push_back(ISD::OutputArg(Flags, Parts[i].getValueType(),
1259 VT, /*isfixed=*/true, 0, 0));
1260 OutVals.push_back(Parts[i]);
1266 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1267 CallingConv::ID CallConv =
1268 DAG.getMachineFunction().getFunction()->getCallingConv();
1269 Chain = DAG.getTargetLoweringInfo().LowerReturn(
1270 Chain, CallConv, isVarArg, Outs, OutVals, getCurSDLoc(), DAG);
1272 // Verify that the target's LowerReturn behaved as expected.
1273 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
1274 "LowerReturn didn't return a valid chain!");
1276 // Update the DAG with the new chain value resulting from return lowering.
1280 /// CopyToExportRegsIfNeeded - If the given value has virtual registers
1281 /// created for it, emit nodes to copy the value into the virtual
1283 void SelectionDAGBuilder::CopyToExportRegsIfNeeded(const Value *V) {
1285 if (V->getType()->isEmptyTy())
1288 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
1289 if (VMI != FuncInfo.ValueMap.end()) {
1290 assert(!V->use_empty() && "Unused value assigned virtual registers!");
1291 CopyValueToVirtualRegister(V, VMI->second);
1295 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
1296 /// the current basic block, add it to ValueMap now so that we'll get a
1298 void SelectionDAGBuilder::ExportFromCurrentBlock(const Value *V) {
1299 // No need to export constants.
1300 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
1302 // Already exported?
1303 if (FuncInfo.isExportedInst(V)) return;
1305 unsigned Reg = FuncInfo.InitializeRegForValue(V);
1306 CopyValueToVirtualRegister(V, Reg);
1309 bool SelectionDAGBuilder::isExportableFromCurrentBlock(const Value *V,
1310 const BasicBlock *FromBB) {
1311 // The operands of the setcc have to be in this block. We don't know
1312 // how to export them from some other block.
1313 if (const Instruction *VI = dyn_cast<Instruction>(V)) {
1314 // Can export from current BB.
1315 if (VI->getParent() == FromBB)
1318 // Is already exported, noop.
1319 return FuncInfo.isExportedInst(V);
1322 // If this is an argument, we can export it if the BB is the entry block or
1323 // if it is already exported.
1324 if (isa<Argument>(V)) {
1325 if (FromBB == &FromBB->getParent()->getEntryBlock())
1328 // Otherwise, can only export this if it is already exported.
1329 return FuncInfo.isExportedInst(V);
1332 // Otherwise, constants can always be exported.
1336 /// Return branch probability calculated by BranchProbabilityInfo for IR blocks.
1337 uint32_t SelectionDAGBuilder::getEdgeWeight(const MachineBasicBlock *Src,
1338 const MachineBasicBlock *Dst) const {
1339 BranchProbabilityInfo *BPI = FuncInfo.BPI;
1342 const BasicBlock *SrcBB = Src->getBasicBlock();
1343 const BasicBlock *DstBB = Dst->getBasicBlock();
1344 return BPI->getEdgeWeight(SrcBB, DstBB);
1347 void SelectionDAGBuilder::
1348 addSuccessorWithWeight(MachineBasicBlock *Src, MachineBasicBlock *Dst,
1349 uint32_t Weight /* = 0 */) {
1351 Weight = getEdgeWeight(Src, Dst);
1352 Src->addSuccessor(Dst, Weight);
1356 static bool InBlock(const Value *V, const BasicBlock *BB) {
1357 if (const Instruction *I = dyn_cast<Instruction>(V))
1358 return I->getParent() == BB;
1362 /// EmitBranchForMergedCondition - Helper method for FindMergedConditions.
1363 /// This function emits a branch and is used at the leaves of an OR or an
1364 /// AND operator tree.
1367 SelectionDAGBuilder::EmitBranchForMergedCondition(const Value *Cond,
1368 MachineBasicBlock *TBB,
1369 MachineBasicBlock *FBB,
1370 MachineBasicBlock *CurBB,
1371 MachineBasicBlock *SwitchBB,
1374 const BasicBlock *BB = CurBB->getBasicBlock();
1376 // If the leaf of the tree is a comparison, merge the condition into
1378 if (const CmpInst *BOp = dyn_cast<CmpInst>(Cond)) {
1379 // The operands of the cmp have to be in this block. We don't know
1380 // how to export them from some other block. If this is the first block
1381 // of the sequence, no exporting is needed.
1382 if (CurBB == SwitchBB ||
1383 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
1384 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) {
1385 ISD::CondCode Condition;
1386 if (const ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
1387 Condition = getICmpCondCode(IC->getPredicate());
1388 } else if (const FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) {
1389 Condition = getFCmpCondCode(FC->getPredicate());
1390 if (TM.Options.NoNaNsFPMath)
1391 Condition = getFCmpCodeWithoutNaN(Condition);
1393 (void)Condition; // silence warning.
1394 llvm_unreachable("Unknown compare instruction");
1397 CaseBlock CB(Condition, BOp->getOperand(0), BOp->getOperand(1), nullptr,
1398 TBB, FBB, CurBB, TWeight, FWeight);
1399 SwitchCases.push_back(CB);
1404 // Create a CaseBlock record representing this branch.
1405 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()),
1406 nullptr, TBB, FBB, CurBB, TWeight, FWeight);
1407 SwitchCases.push_back(CB);
1410 /// Scale down both weights to fit into uint32_t.
1411 static void ScaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
1412 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
1413 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
1414 NewTrue = NewTrue / Scale;
1415 NewFalse = NewFalse / Scale;
1418 /// FindMergedConditions - If Cond is an expression like
1419 void SelectionDAGBuilder::FindMergedConditions(const Value *Cond,
1420 MachineBasicBlock *TBB,
1421 MachineBasicBlock *FBB,
1422 MachineBasicBlock *CurBB,
1423 MachineBasicBlock *SwitchBB,
1424 unsigned Opc, uint32_t TWeight,
1426 // If this node is not part of the or/and tree, emit it as a branch.
1427 const Instruction *BOp = dyn_cast<Instruction>(Cond);
1428 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
1429 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
1430 BOp->getParent() != CurBB->getBasicBlock() ||
1431 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
1432 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
1433 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB,
1438 // Create TmpBB after CurBB.
1439 MachineFunction::iterator BBI = CurBB;
1440 MachineFunction &MF = DAG.getMachineFunction();
1441 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock());
1442 CurBB->getParent()->insert(++BBI, TmpBB);
1444 if (Opc == Instruction::Or) {
1445 // Codegen X | Y as:
1454 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
1455 // The requirement is that
1456 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
1457 // = TrueProb for original BB.
1458 // Assuming the original weights are A and B, one choice is to set BB1's
1459 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
1461 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
1462 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
1463 // TmpBB, but the math is more complicated.
1465 uint64_t NewTrueWeight = TWeight;
1466 uint64_t NewFalseWeight = (uint64_t)TWeight + 2 * (uint64_t)FWeight;
1467 ScaleWeights(NewTrueWeight, NewFalseWeight);
1468 // Emit the LHS condition.
1469 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, SwitchBB, Opc,
1470 NewTrueWeight, NewFalseWeight);
1472 NewTrueWeight = TWeight;
1473 NewFalseWeight = 2 * (uint64_t)FWeight;
1474 ScaleWeights(NewTrueWeight, NewFalseWeight);
1475 // Emit the RHS condition into TmpBB.
1476 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc,
1477 NewTrueWeight, NewFalseWeight);
1479 assert(Opc == Instruction::And && "Unknown merge op!");
1480 // Codegen X & Y as:
1488 // This requires creation of TmpBB after CurBB.
1490 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
1491 // The requirement is that
1492 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
1493 // = FalseProb for original BB.
1494 // Assuming the original weights are A and B, one choice is to set BB1's
1495 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
1497 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
1499 uint64_t NewTrueWeight = 2 * (uint64_t)TWeight + (uint64_t)FWeight;
1500 uint64_t NewFalseWeight = FWeight;
1501 ScaleWeights(NewTrueWeight, NewFalseWeight);
1502 // Emit the LHS condition.
1503 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, SwitchBB, Opc,
1504 NewTrueWeight, NewFalseWeight);
1506 NewTrueWeight = 2 * (uint64_t)TWeight;
1507 NewFalseWeight = FWeight;
1508 ScaleWeights(NewTrueWeight, NewFalseWeight);
1509 // Emit the RHS condition into TmpBB.
1510 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc,
1511 NewTrueWeight, NewFalseWeight);
1515 /// If the set of cases should be emitted as a series of branches, return true.
1516 /// If we should emit this as a bunch of and/or'd together conditions, return
1519 SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases) {
1520 if (Cases.size() != 2) return true;
1522 // If this is two comparisons of the same values or'd or and'd together, they
1523 // will get folded into a single comparison, so don't emit two blocks.
1524 if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
1525 Cases[0].CmpRHS == Cases[1].CmpRHS) ||
1526 (Cases[0].CmpRHS == Cases[1].CmpLHS &&
1527 Cases[0].CmpLHS == Cases[1].CmpRHS)) {
1531 // Handle: (X != null) | (Y != null) --> (X|Y) != 0
1532 // Handle: (X == null) & (Y == null) --> (X|Y) == 0
1533 if (Cases[0].CmpRHS == Cases[1].CmpRHS &&
1534 Cases[0].CC == Cases[1].CC &&
1535 isa<Constant>(Cases[0].CmpRHS) &&
1536 cast<Constant>(Cases[0].CmpRHS)->isNullValue()) {
1537 if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB)
1539 if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB)
1546 void SelectionDAGBuilder::visitBr(const BranchInst &I) {
1547 MachineBasicBlock *BrMBB = FuncInfo.MBB;
1549 // Update machine-CFG edges.
1550 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
1552 if (I.isUnconditional()) {
1553 // Update machine-CFG edges.
1554 BrMBB->addSuccessor(Succ0MBB);
1556 // If this is not a fall-through branch or optimizations are switched off,
1558 if (Succ0MBB != NextBlock(BrMBB) || TM.getOptLevel() == CodeGenOpt::None)
1559 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(),
1560 MVT::Other, getControlRoot(),
1561 DAG.getBasicBlock(Succ0MBB)));
1566 // If this condition is one of the special cases we handle, do special stuff
1568 const Value *CondVal = I.getCondition();
1569 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
1571 // If this is a series of conditions that are or'd or and'd together, emit
1572 // this as a sequence of branches instead of setcc's with and/or operations.
1573 // As long as jumps are not expensive, this should improve performance.
1574 // For example, instead of something like:
1587 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
1588 if (!DAG.getTargetLoweringInfo().isJumpExpensive() &&
1589 BOp->hasOneUse() && (BOp->getOpcode() == Instruction::And ||
1590 BOp->getOpcode() == Instruction::Or)) {
1591 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB,
1592 BOp->getOpcode(), getEdgeWeight(BrMBB, Succ0MBB),
1593 getEdgeWeight(BrMBB, Succ1MBB));
1594 // If the compares in later blocks need to use values not currently
1595 // exported from this block, export them now. This block should always
1596 // be the first entry.
1597 assert(SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!");
1599 // Allow some cases to be rejected.
1600 if (ShouldEmitAsBranches(SwitchCases)) {
1601 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1602 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1603 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1606 // Emit the branch for this block.
1607 visitSwitchCase(SwitchCases[0], BrMBB);
1608 SwitchCases.erase(SwitchCases.begin());
1612 // Okay, we decided not to do this, remove any inserted MBB's and clear
1614 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
1615 FuncInfo.MF->erase(SwitchCases[i].ThisBB);
1617 SwitchCases.clear();
1621 // Create a CaseBlock record representing this branch.
1622 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()),
1623 nullptr, Succ0MBB, Succ1MBB, BrMBB);
1625 // Use visitSwitchCase to actually insert the fast branch sequence for this
1627 visitSwitchCase(CB, BrMBB);
1630 /// visitSwitchCase - Emits the necessary code to represent a single node in
1631 /// the binary search tree resulting from lowering a switch instruction.
1632 void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB,
1633 MachineBasicBlock *SwitchBB) {
1635 SDValue CondLHS = getValue(CB.CmpLHS);
1636 SDLoc dl = getCurSDLoc();
1638 // Build the setcc now.
1640 // Fold "(X == true)" to X and "(X == false)" to !X to
1641 // handle common cases produced by branch lowering.
1642 if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) &&
1643 CB.CC == ISD::SETEQ)
1645 else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) &&
1646 CB.CC == ISD::SETEQ) {
1647 SDValue True = DAG.getConstant(1, dl, CondLHS.getValueType());
1648 Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True);
1650 Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1652 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now");
1654 const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue();
1655 const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue();
1657 SDValue CmpOp = getValue(CB.CmpMHS);
1658 EVT VT = CmpOp.getValueType();
1660 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) {
1661 Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, dl, VT),
1664 SDValue SUB = DAG.getNode(ISD::SUB, dl,
1665 VT, CmpOp, DAG.getConstant(Low, dl, VT));
1666 Cond = DAG.getSetCC(dl, MVT::i1, SUB,
1667 DAG.getConstant(High-Low, dl, VT), ISD::SETULE);
1671 // Update successor info
1672 addSuccessorWithWeight(SwitchBB, CB.TrueBB, CB.TrueWeight);
1673 // TrueBB and FalseBB are always different unless the incoming IR is
1674 // degenerate. This only happens when running llc on weird IR.
1675 if (CB.TrueBB != CB.FalseBB)
1676 addSuccessorWithWeight(SwitchBB, CB.FalseBB, CB.FalseWeight);
1678 // If the lhs block is the next block, invert the condition so that we can
1679 // fall through to the lhs instead of the rhs block.
1680 if (CB.TrueBB == NextBlock(SwitchBB)) {
1681 std::swap(CB.TrueBB, CB.FalseBB);
1682 SDValue True = DAG.getConstant(1, dl, Cond.getValueType());
1683 Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True);
1686 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1687 MVT::Other, getControlRoot(), Cond,
1688 DAG.getBasicBlock(CB.TrueBB));
1690 // Insert the false branch. Do this even if it's a fall through branch,
1691 // this makes it easier to do DAG optimizations which require inverting
1692 // the branch condition.
1693 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1694 DAG.getBasicBlock(CB.FalseBB));
1696 DAG.setRoot(BrCond);
1699 /// visitJumpTable - Emit JumpTable node in the current MBB
1700 void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) {
1701 // Emit the code for the jump table
1702 assert(JT.Reg != -1U && "Should lower JT Header first!");
1703 EVT PTy = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
1704 SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurSDLoc(),
1706 SDValue Table = DAG.getJumpTable(JT.JTI, PTy);
1707 SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurSDLoc(),
1708 MVT::Other, Index.getValue(1),
1710 DAG.setRoot(BrJumpTable);
1713 /// visitJumpTableHeader - This function emits necessary code to produce index
1714 /// in the JumpTable from switch case.
1715 void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT,
1716 JumpTableHeader &JTH,
1717 MachineBasicBlock *SwitchBB) {
1718 SDLoc dl = getCurSDLoc();
1720 // Subtract the lowest switch case value from the value being switched on and
1721 // conditional branch to default mbb if the result is greater than the
1722 // difference between smallest and largest cases.
1723 SDValue SwitchOp = getValue(JTH.SValue);
1724 EVT VT = SwitchOp.getValueType();
1725 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, SwitchOp,
1726 DAG.getConstant(JTH.First, dl, VT));
1728 // The SDNode we just created, which holds the value being switched on minus
1729 // the smallest case value, needs to be copied to a virtual register so it
1730 // can be used as an index into the jump table in a subsequent basic block.
1731 // This value may be smaller or larger than the target's pointer type, and
1732 // therefore require extension or truncating.
1733 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1734 SwitchOp = DAG.getZExtOrTrunc(Sub, dl, TLI.getPointerTy(DAG.getDataLayout()));
1736 unsigned JumpTableReg =
1737 FuncInfo.CreateReg(TLI.getPointerTy(DAG.getDataLayout()));
1738 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), dl,
1739 JumpTableReg, SwitchOp);
1740 JT.Reg = JumpTableReg;
1742 // Emit the range check for the jump table, and branch to the default block
1743 // for the switch statement if the value being switched on exceeds the largest
1744 // case in the switch.
1745 SDValue CMP = DAG.getSetCC(
1746 dl, TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(),
1747 Sub.getValueType()),
1748 Sub, DAG.getConstant(JTH.Last - JTH.First, dl, VT), ISD::SETUGT);
1750 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1751 MVT::Other, CopyTo, CMP,
1752 DAG.getBasicBlock(JT.Default));
1754 // Avoid emitting unnecessary branches to the next block.
1755 if (JT.MBB != NextBlock(SwitchBB))
1756 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1757 DAG.getBasicBlock(JT.MBB));
1759 DAG.setRoot(BrCond);
1762 /// Codegen a new tail for a stack protector check ParentMBB which has had its
1763 /// tail spliced into a stack protector check success bb.
1765 /// For a high level explanation of how this fits into the stack protector
1766 /// generation see the comment on the declaration of class
1767 /// StackProtectorDescriptor.
1768 void SelectionDAGBuilder::visitSPDescriptorParent(StackProtectorDescriptor &SPD,
1769 MachineBasicBlock *ParentBB) {
1771 // First create the loads to the guard/stack slot for the comparison.
1772 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1773 EVT PtrTy = TLI.getPointerTy(DAG.getDataLayout());
1775 MachineFrameInfo *MFI = ParentBB->getParent()->getFrameInfo();
1776 int FI = MFI->getStackProtectorIndex();
1778 const Value *IRGuard = SPD.getGuard();
1779 SDValue GuardPtr = getValue(IRGuard);
1780 SDValue StackSlotPtr = DAG.getFrameIndex(FI, PtrTy);
1782 unsigned Align = DL->getPrefTypeAlignment(IRGuard->getType());
1785 SDLoc dl = getCurSDLoc();
1787 // If GuardReg is set and useLoadStackGuardNode returns true, retrieve the
1788 // guard value from the virtual register holding the value. Otherwise, emit a
1789 // volatile load to retrieve the stack guard value.
1790 unsigned GuardReg = SPD.getGuardReg();
1792 if (GuardReg && TLI.useLoadStackGuardNode())
1793 Guard = DAG.getCopyFromReg(DAG.getEntryNode(), dl, GuardReg,
1796 Guard = DAG.getLoad(PtrTy, dl, DAG.getEntryNode(),
1797 GuardPtr, MachinePointerInfo(IRGuard, 0),
1798 true, false, false, Align);
1800 SDValue StackSlot = DAG.getLoad(PtrTy, dl, DAG.getEntryNode(),
1802 MachinePointerInfo::getFixedStack(FI),
1803 true, false, false, Align);
1805 // Perform the comparison via a subtract/getsetcc.
1806 EVT VT = Guard.getValueType();
1807 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, Guard, StackSlot);
1809 SDValue Cmp = DAG.getSetCC(dl, TLI.getSetCCResultType(DAG.getDataLayout(),
1811 Sub.getValueType()),
1812 Sub, DAG.getConstant(0, dl, VT), ISD::SETNE);
1814 // If the sub is not 0, then we know the guard/stackslot do not equal, so
1815 // branch to failure MBB.
1816 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1817 MVT::Other, StackSlot.getOperand(0),
1818 Cmp, DAG.getBasicBlock(SPD.getFailureMBB()));
1819 // Otherwise branch to success MBB.
1820 SDValue Br = DAG.getNode(ISD::BR, dl,
1822 DAG.getBasicBlock(SPD.getSuccessMBB()));
1827 /// Codegen the failure basic block for a stack protector check.
1829 /// A failure stack protector machine basic block consists simply of a call to
1830 /// __stack_chk_fail().
1832 /// For a high level explanation of how this fits into the stack protector
1833 /// generation see the comment on the declaration of class
1834 /// StackProtectorDescriptor.
1836 SelectionDAGBuilder::visitSPDescriptorFailure(StackProtectorDescriptor &SPD) {
1837 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1839 TLI.makeLibCall(DAG, RTLIB::STACKPROTECTOR_CHECK_FAIL, MVT::isVoid,
1840 nullptr, 0, false, getCurSDLoc(), false, false).second;
1844 /// visitBitTestHeader - This function emits necessary code to produce value
1845 /// suitable for "bit tests"
1846 void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B,
1847 MachineBasicBlock *SwitchBB) {
1848 SDLoc dl = getCurSDLoc();
1850 // Subtract the minimum value
1851 SDValue SwitchOp = getValue(B.SValue);
1852 EVT VT = SwitchOp.getValueType();
1853 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, SwitchOp,
1854 DAG.getConstant(B.First, dl, VT));
1857 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1858 SDValue RangeCmp = DAG.getSetCC(
1859 dl, TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(),
1860 Sub.getValueType()),
1861 Sub, DAG.getConstant(B.Range, dl, VT), ISD::SETUGT);
1863 // Determine the type of the test operands.
1864 bool UsePtrType = false;
1865 if (!TLI.isTypeLegal(VT))
1868 for (unsigned i = 0, e = B.Cases.size(); i != e; ++i)
1869 if (!isUIntN(VT.getSizeInBits(), B.Cases[i].Mask)) {
1870 // Switch table case range are encoded into series of masks.
1871 // Just use pointer type, it's guaranteed to fit.
1877 VT = TLI.getPointerTy(DAG.getDataLayout());
1878 Sub = DAG.getZExtOrTrunc(Sub, dl, VT);
1881 B.RegVT = VT.getSimpleVT();
1882 B.Reg = FuncInfo.CreateReg(B.RegVT);
1883 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), dl, B.Reg, Sub);
1885 MachineBasicBlock* MBB = B.Cases[0].ThisBB;
1887 addSuccessorWithWeight(SwitchBB, B.Default);
1888 addSuccessorWithWeight(SwitchBB, MBB);
1890 SDValue BrRange = DAG.getNode(ISD::BRCOND, dl,
1891 MVT::Other, CopyTo, RangeCmp,
1892 DAG.getBasicBlock(B.Default));
1894 // Avoid emitting unnecessary branches to the next block.
1895 if (MBB != NextBlock(SwitchBB))
1896 BrRange = DAG.getNode(ISD::BR, dl, MVT::Other, BrRange,
1897 DAG.getBasicBlock(MBB));
1899 DAG.setRoot(BrRange);
1902 /// visitBitTestCase - this function produces one "bit test"
1903 void SelectionDAGBuilder::visitBitTestCase(BitTestBlock &BB,
1904 MachineBasicBlock* NextMBB,
1905 uint32_t BranchWeightToNext,
1908 MachineBasicBlock *SwitchBB) {
1909 SDLoc dl = getCurSDLoc();
1911 SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), dl, Reg, VT);
1913 unsigned PopCount = countPopulation(B.Mask);
1914 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1915 if (PopCount == 1) {
1916 // Testing for a single bit; just compare the shift count with what it
1917 // would need to be to shift a 1 bit in that position.
1919 dl, TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT),
1920 ShiftOp, DAG.getConstant(countTrailingZeros(B.Mask), dl, VT),
1922 } else if (PopCount == BB.Range) {
1923 // There is only one zero bit in the range, test for it directly.
1925 dl, TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT),
1926 ShiftOp, DAG.getConstant(countTrailingOnes(B.Mask), dl, VT),
1929 // Make desired shift
1930 SDValue SwitchVal = DAG.getNode(ISD::SHL, dl, VT,
1931 DAG.getConstant(1, dl, VT), ShiftOp);
1933 // Emit bit tests and jumps
1934 SDValue AndOp = DAG.getNode(ISD::AND, dl,
1935 VT, SwitchVal, DAG.getConstant(B.Mask, dl, VT));
1937 dl, TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT),
1938 AndOp, DAG.getConstant(0, dl, VT), ISD::SETNE);
1941 // The branch weight from SwitchBB to B.TargetBB is B.ExtraWeight.
1942 addSuccessorWithWeight(SwitchBB, B.TargetBB, B.ExtraWeight);
1943 // The branch weight from SwitchBB to NextMBB is BranchWeightToNext.
1944 addSuccessorWithWeight(SwitchBB, NextMBB, BranchWeightToNext);
1946 SDValue BrAnd = DAG.getNode(ISD::BRCOND, dl,
1947 MVT::Other, getControlRoot(),
1948 Cmp, DAG.getBasicBlock(B.TargetBB));
1950 // Avoid emitting unnecessary branches to the next block.
1951 if (NextMBB != NextBlock(SwitchBB))
1952 BrAnd = DAG.getNode(ISD::BR, dl, MVT::Other, BrAnd,
1953 DAG.getBasicBlock(NextMBB));
1958 void SelectionDAGBuilder::visitInvoke(const InvokeInst &I) {
1959 MachineBasicBlock *InvokeMBB = FuncInfo.MBB;
1961 // Retrieve successors.
1962 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
1963 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)];
1965 const Value *Callee(I.getCalledValue());
1966 const Function *Fn = dyn_cast<Function>(Callee);
1967 if (isa<InlineAsm>(Callee))
1969 else if (Fn && Fn->isIntrinsic()) {
1970 switch (Fn->getIntrinsicID()) {
1972 llvm_unreachable("Cannot invoke this intrinsic");
1973 case Intrinsic::donothing:
1974 // Ignore invokes to @llvm.donothing: jump directly to the next BB.
1976 case Intrinsic::experimental_patchpoint_void:
1977 case Intrinsic::experimental_patchpoint_i64:
1978 visitPatchpoint(&I, LandingPad);
1980 case Intrinsic::experimental_gc_statepoint:
1981 LowerStatepoint(ImmutableStatepoint(&I), LandingPad);
1985 LowerCallTo(&I, getValue(Callee), false, LandingPad);
1987 // If the value of the invoke is used outside of its defining block, make it
1988 // available as a virtual register.
1989 // We already took care of the exported value for the statepoint instruction
1990 // during call to the LowerStatepoint.
1991 if (!isStatepoint(I)) {
1992 CopyToExportRegsIfNeeded(&I);
1995 // Update successor info
1996 addSuccessorWithWeight(InvokeMBB, Return);
1997 addSuccessorWithWeight(InvokeMBB, LandingPad);
1999 // Drop into normal successor.
2000 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(),
2001 MVT::Other, getControlRoot(),
2002 DAG.getBasicBlock(Return)));
2005 void SelectionDAGBuilder::visitResume(const ResumeInst &RI) {
2006 llvm_unreachable("SelectionDAGBuilder shouldn't visit resume instructions!");
2009 void SelectionDAGBuilder::visitLandingPad(const LandingPadInst &LP) {
2010 assert(FuncInfo.MBB->isLandingPad() &&
2011 "Call to landingpad not in landing pad!");
2013 MachineBasicBlock *MBB = FuncInfo.MBB;
2014 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
2015 AddLandingPadInfo(LP, MMI, MBB);
2017 // If there aren't registers to copy the values into (e.g., during SjLj
2018 // exceptions), then don't bother to create these DAG nodes.
2019 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2020 if (TLI.getExceptionPointerRegister() == 0 &&
2021 TLI.getExceptionSelectorRegister() == 0)
2024 SmallVector<EVT, 2> ValueVTs;
2025 SDLoc dl = getCurSDLoc();
2026 ComputeValueVTs(TLI, DAG.getDataLayout(), LP.getType(), ValueVTs);
2027 assert(ValueVTs.size() == 2 && "Only two-valued landingpads are supported");
2029 // Get the two live-in registers as SDValues. The physregs have already been
2030 // copied into virtual registers.
2032 if (FuncInfo.ExceptionPointerVirtReg) {
2033 Ops[0] = DAG.getZExtOrTrunc(
2034 DAG.getCopyFromReg(DAG.getEntryNode(), dl,
2035 FuncInfo.ExceptionPointerVirtReg,
2036 TLI.getPointerTy(DAG.getDataLayout())),
2039 Ops[0] = DAG.getConstant(0, dl, TLI.getPointerTy(DAG.getDataLayout()));
2041 Ops[1] = DAG.getZExtOrTrunc(
2042 DAG.getCopyFromReg(DAG.getEntryNode(), dl,
2043 FuncInfo.ExceptionSelectorVirtReg,
2044 TLI.getPointerTy(DAG.getDataLayout())),
2048 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, dl,
2049 DAG.getVTList(ValueVTs), Ops);
2054 SelectionDAGBuilder::visitLandingPadClauseBB(GlobalValue *ClauseGV,
2055 MachineBasicBlock *LPadBB) {
2056 SDValue Chain = getControlRoot();
2057 SDLoc dl = getCurSDLoc();
2059 // Get the typeid that we will dispatch on later.
2060 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2061 const TargetRegisterClass *RC =
2062 TLI.getRegClassFor(TLI.getPointerTy(DAG.getDataLayout()));
2063 unsigned VReg = FuncInfo.MF->getRegInfo().createVirtualRegister(RC);
2064 unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(ClauseGV);
2066 DAG.getConstant(TypeID, dl, TLI.getPointerTy(DAG.getDataLayout()));
2067 Chain = DAG.getCopyToReg(Chain, dl, VReg, Sel);
2069 // Branch to the main landing pad block.
2070 MachineBasicBlock *ClauseMBB = FuncInfo.MBB;
2071 ClauseMBB->addSuccessor(LPadBB);
2072 DAG.setRoot(DAG.getNode(ISD::BR, dl, MVT::Other, Chain,
2073 DAG.getBasicBlock(LPadBB)));
2077 void SelectionDAGBuilder::sortAndRangeify(CaseClusterVector &Clusters) {
2079 for (const CaseCluster &CC : Clusters)
2080 assert(CC.Low == CC.High && "Input clusters must be single-case");
2083 std::sort(Clusters.begin(), Clusters.end(),
2084 [](const CaseCluster &a, const CaseCluster &b) {
2085 return a.Low->getValue().slt(b.Low->getValue());
2088 // Merge adjacent clusters with the same destination.
2089 const unsigned N = Clusters.size();
2090 unsigned DstIndex = 0;
2091 for (unsigned SrcIndex = 0; SrcIndex < N; ++SrcIndex) {
2092 CaseCluster &CC = Clusters[SrcIndex];
2093 const ConstantInt *CaseVal = CC.Low;
2094 MachineBasicBlock *Succ = CC.MBB;
2096 if (DstIndex != 0 && Clusters[DstIndex - 1].MBB == Succ &&
2097 (CaseVal->getValue() - Clusters[DstIndex - 1].High->getValue()) == 1) {
2098 // If this case has the same successor and is a neighbour, merge it into
2099 // the previous cluster.
2100 Clusters[DstIndex - 1].High = CaseVal;
2101 Clusters[DstIndex - 1].Weight += CC.Weight;
2102 assert(Clusters[DstIndex - 1].Weight >= CC.Weight && "Weight overflow!");
2104 std::memmove(&Clusters[DstIndex++], &Clusters[SrcIndex],
2105 sizeof(Clusters[SrcIndex]));
2108 Clusters.resize(DstIndex);
2111 void SelectionDAGBuilder::UpdateSplitBlock(MachineBasicBlock *First,
2112 MachineBasicBlock *Last) {
2114 for (unsigned i = 0, e = JTCases.size(); i != e; ++i)
2115 if (JTCases[i].first.HeaderBB == First)
2116 JTCases[i].first.HeaderBB = Last;
2118 // Update BitTestCases.
2119 for (unsigned i = 0, e = BitTestCases.size(); i != e; ++i)
2120 if (BitTestCases[i].Parent == First)
2121 BitTestCases[i].Parent = Last;
2124 void SelectionDAGBuilder::visitIndirectBr(const IndirectBrInst &I) {
2125 MachineBasicBlock *IndirectBrMBB = FuncInfo.MBB;
2127 // Update machine-CFG edges with unique successors.
2128 SmallSet<BasicBlock*, 32> Done;
2129 for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i) {
2130 BasicBlock *BB = I.getSuccessor(i);
2131 bool Inserted = Done.insert(BB).second;
2135 MachineBasicBlock *Succ = FuncInfo.MBBMap[BB];
2136 addSuccessorWithWeight(IndirectBrMBB, Succ);
2139 DAG.setRoot(DAG.getNode(ISD::BRIND, getCurSDLoc(),
2140 MVT::Other, getControlRoot(),
2141 getValue(I.getAddress())));
2144 void SelectionDAGBuilder::visitUnreachable(const UnreachableInst &I) {
2145 if (DAG.getTarget().Options.TrapUnreachable)
2146 DAG.setRoot(DAG.getNode(ISD::TRAP, getCurSDLoc(), MVT::Other, DAG.getRoot()));
2149 void SelectionDAGBuilder::visitFSub(const User &I) {
2150 // -0.0 - X --> fneg
2151 Type *Ty = I.getType();
2152 if (isa<Constant>(I.getOperand(0)) &&
2153 I.getOperand(0) == ConstantFP::getZeroValueForNegation(Ty)) {
2154 SDValue Op2 = getValue(I.getOperand(1));
2155 setValue(&I, DAG.getNode(ISD::FNEG, getCurSDLoc(),
2156 Op2.getValueType(), Op2));
2160 visitBinary(I, ISD::FSUB);
2163 void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) {
2164 SDValue Op1 = getValue(I.getOperand(0));
2165 SDValue Op2 = getValue(I.getOperand(1));
2172 if (const OverflowingBinaryOperator *OFBinOp =
2173 dyn_cast<const OverflowingBinaryOperator>(&I)) {
2174 nuw = OFBinOp->hasNoUnsignedWrap();
2175 nsw = OFBinOp->hasNoSignedWrap();
2177 if (const PossiblyExactOperator *ExactOp =
2178 dyn_cast<const PossiblyExactOperator>(&I))
2179 exact = ExactOp->isExact();
2180 if (const FPMathOperator *FPOp = dyn_cast<const FPMathOperator>(&I))
2181 FMF = FPOp->getFastMathFlags();
2184 Flags.setExact(exact);
2185 Flags.setNoSignedWrap(nsw);
2186 Flags.setNoUnsignedWrap(nuw);
2187 if (EnableFMFInDAG) {
2188 Flags.setAllowReciprocal(FMF.allowReciprocal());
2189 Flags.setNoInfs(FMF.noInfs());
2190 Flags.setNoNaNs(FMF.noNaNs());
2191 Flags.setNoSignedZeros(FMF.noSignedZeros());
2192 Flags.setUnsafeAlgebra(FMF.unsafeAlgebra());
2194 SDValue BinNodeValue = DAG.getNode(OpCode, getCurSDLoc(), Op1.getValueType(),
2196 setValue(&I, BinNodeValue);
2199 void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) {
2200 SDValue Op1 = getValue(I.getOperand(0));
2201 SDValue Op2 = getValue(I.getOperand(1));
2204 DAG.getTargetLoweringInfo().getShiftAmountTy(Op2.getValueType());
2206 // Coerce the shift amount to the right type if we can.
2207 if (!I.getType()->isVectorTy() && Op2.getValueType() != ShiftTy) {
2208 unsigned ShiftSize = ShiftTy.getSizeInBits();
2209 unsigned Op2Size = Op2.getValueType().getSizeInBits();
2210 SDLoc DL = getCurSDLoc();
2212 // If the operand is smaller than the shift count type, promote it.
2213 if (ShiftSize > Op2Size)
2214 Op2 = DAG.getNode(ISD::ZERO_EXTEND, DL, ShiftTy, Op2);
2216 // If the operand is larger than the shift count type but the shift
2217 // count type has enough bits to represent any shift value, truncate
2218 // it now. This is a common case and it exposes the truncate to
2219 // optimization early.
2220 else if (ShiftSize >= Log2_32_Ceil(Op2.getValueType().getSizeInBits()))
2221 Op2 = DAG.getNode(ISD::TRUNCATE, DL, ShiftTy, Op2);
2222 // Otherwise we'll need to temporarily settle for some other convenient
2223 // type. Type legalization will make adjustments once the shiftee is split.
2225 Op2 = DAG.getZExtOrTrunc(Op2, DL, MVT::i32);
2232 if (Opcode == ISD::SRL || Opcode == ISD::SRA || Opcode == ISD::SHL) {
2234 if (const OverflowingBinaryOperator *OFBinOp =
2235 dyn_cast<const OverflowingBinaryOperator>(&I)) {
2236 nuw = OFBinOp->hasNoUnsignedWrap();
2237 nsw = OFBinOp->hasNoSignedWrap();
2239 if (const PossiblyExactOperator *ExactOp =
2240 dyn_cast<const PossiblyExactOperator>(&I))
2241 exact = ExactOp->isExact();
2244 Flags.setExact(exact);
2245 Flags.setNoSignedWrap(nsw);
2246 Flags.setNoUnsignedWrap(nuw);
2247 SDValue Res = DAG.getNode(Opcode, getCurSDLoc(), Op1.getValueType(), Op1, Op2,
2252 void SelectionDAGBuilder::visitSDiv(const User &I) {
2253 SDValue Op1 = getValue(I.getOperand(0));
2254 SDValue Op2 = getValue(I.getOperand(1));
2257 Flags.setExact(isa<PossiblyExactOperator>(&I) &&
2258 cast<PossiblyExactOperator>(&I)->isExact());
2259 setValue(&I, DAG.getNode(ISD::SDIV, getCurSDLoc(), Op1.getValueType(), Op1,
2263 void SelectionDAGBuilder::visitICmp(const User &I) {
2264 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2265 if (const ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2266 predicate = IC->getPredicate();
2267 else if (const ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2268 predicate = ICmpInst::Predicate(IC->getPredicate());
2269 SDValue Op1 = getValue(I.getOperand(0));
2270 SDValue Op2 = getValue(I.getOperand(1));
2271 ISD::CondCode Opcode = getICmpCondCode(predicate);
2273 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2275 setValue(&I, DAG.getSetCC(getCurSDLoc(), DestVT, Op1, Op2, Opcode));
2278 void SelectionDAGBuilder::visitFCmp(const User &I) {
2279 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2280 if (const FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2281 predicate = FC->getPredicate();
2282 else if (const ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2283 predicate = FCmpInst::Predicate(FC->getPredicate());
2284 SDValue Op1 = getValue(I.getOperand(0));
2285 SDValue Op2 = getValue(I.getOperand(1));
2286 ISD::CondCode Condition = getFCmpCondCode(predicate);
2287 if (TM.Options.NoNaNsFPMath)
2288 Condition = getFCmpCodeWithoutNaN(Condition);
2289 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2291 setValue(&I, DAG.getSetCC(getCurSDLoc(), DestVT, Op1, Op2, Condition));
2294 void SelectionDAGBuilder::visitSelect(const User &I) {
2295 SmallVector<EVT, 4> ValueVTs;
2296 ComputeValueVTs(DAG.getTargetLoweringInfo(), DAG.getDataLayout(), I.getType(),
2298 unsigned NumValues = ValueVTs.size();
2299 if (NumValues == 0) return;
2301 SmallVector<SDValue, 4> Values(NumValues);
2302 SDValue Cond = getValue(I.getOperand(0));
2303 SDValue LHSVal = getValue(I.getOperand(1));
2304 SDValue RHSVal = getValue(I.getOperand(2));
2305 auto BaseOps = {Cond};
2306 ISD::NodeType OpCode = Cond.getValueType().isVector() ?
2307 ISD::VSELECT : ISD::SELECT;
2309 // Min/max matching is only viable if all output VTs are the same.
2310 if (std::equal(ValueVTs.begin(), ValueVTs.end(), ValueVTs.begin())) {
2312 SelectPatternFlavor SPF = matchSelectPattern(const_cast<User*>(&I), LHS, RHS);
2313 ISD::NodeType Opc = ISD::DELETED_NODE;
2315 case SPF_UMAX: Opc = ISD::UMAX; break;
2316 case SPF_UMIN: Opc = ISD::UMIN; break;
2317 case SPF_SMAX: Opc = ISD::SMAX; break;
2318 case SPF_SMIN: Opc = ISD::SMIN; break;
2322 EVT VT = ValueVTs[0];
2323 LLVMContext &Ctx = *DAG.getContext();
2324 auto &TLI = DAG.getTargetLoweringInfo();
2325 while (TLI.getTypeAction(Ctx, VT) == TargetLoweringBase::TypeSplitVector)
2326 VT = TLI.getTypeToTransformTo(Ctx, VT);
2328 if (Opc != ISD::DELETED_NODE && TLI.isOperationLegalOrCustom(Opc, VT) &&
2329 // If the underlying comparison instruction is used by any other instruction,
2330 // the consumed instructions won't be destroyed, so it is not profitable
2331 // to convert to a min/max.
2332 cast<SelectInst>(&I)->getCondition()->hasOneUse()) {
2334 LHSVal = getValue(LHS);
2335 RHSVal = getValue(RHS);
2340 for (unsigned i = 0; i != NumValues; ++i) {
2341 SmallVector<SDValue, 3> Ops(BaseOps.begin(), BaseOps.end());
2342 Ops.push_back(SDValue(LHSVal.getNode(), LHSVal.getResNo() + i));
2343 Ops.push_back(SDValue(RHSVal.getNode(), RHSVal.getResNo() + i));
2344 Values[i] = DAG.getNode(OpCode, getCurSDLoc(),
2345 LHSVal.getNode()->getValueType(LHSVal.getResNo()+i),
2349 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
2350 DAG.getVTList(ValueVTs), Values));
2353 void SelectionDAGBuilder::visitTrunc(const User &I) {
2354 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2355 SDValue N = getValue(I.getOperand(0));
2356 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2358 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), DestVT, N));
2361 void SelectionDAGBuilder::visitZExt(const User &I) {
2362 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2363 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2364 SDValue N = getValue(I.getOperand(0));
2365 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2367 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurSDLoc(), DestVT, N));
2370 void SelectionDAGBuilder::visitSExt(const User &I) {
2371 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2372 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2373 SDValue N = getValue(I.getOperand(0));
2374 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2376 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurSDLoc(), DestVT, N));
2379 void SelectionDAGBuilder::visitFPTrunc(const User &I) {
2380 // FPTrunc is never a no-op cast, no need to check
2381 SDValue N = getValue(I.getOperand(0));
2382 SDLoc dl = getCurSDLoc();
2383 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2384 EVT DestVT = TLI.getValueType(DAG.getDataLayout(), I.getType());
2385 setValue(&I, DAG.getNode(ISD::FP_ROUND, dl, DestVT, N,
2386 DAG.getTargetConstant(
2387 0, dl, TLI.getPointerTy(DAG.getDataLayout()))));
2390 void SelectionDAGBuilder::visitFPExt(const User &I) {
2391 // FPExt is never a no-op cast, no need to check
2392 SDValue N = getValue(I.getOperand(0));
2393 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2395 setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurSDLoc(), DestVT, N));
2398 void SelectionDAGBuilder::visitFPToUI(const User &I) {
2399 // FPToUI is never a no-op cast, no need to check
2400 SDValue N = getValue(I.getOperand(0));
2401 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2403 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurSDLoc(), DestVT, N));
2406 void SelectionDAGBuilder::visitFPToSI(const User &I) {
2407 // FPToSI is never a no-op cast, no need to check
2408 SDValue N = getValue(I.getOperand(0));
2409 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2411 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurSDLoc(), DestVT, N));
2414 void SelectionDAGBuilder::visitUIToFP(const User &I) {
2415 // UIToFP is never a no-op cast, no need to check
2416 SDValue N = getValue(I.getOperand(0));
2417 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2419 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurSDLoc(), DestVT, N));
2422 void SelectionDAGBuilder::visitSIToFP(const User &I) {
2423 // SIToFP is never a no-op cast, no need to check
2424 SDValue N = getValue(I.getOperand(0));
2425 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2427 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurSDLoc(), DestVT, N));
2430 void SelectionDAGBuilder::visitPtrToInt(const User &I) {
2431 // What to do depends on the size of the integer and the size of the pointer.
2432 // We can either truncate, zero extend, or no-op, accordingly.
2433 SDValue N = getValue(I.getOperand(0));
2434 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2436 setValue(&I, DAG.getZExtOrTrunc(N, getCurSDLoc(), DestVT));
2439 void SelectionDAGBuilder::visitIntToPtr(const User &I) {
2440 // What to do depends on the size of the integer and the size of the pointer.
2441 // We can either truncate, zero extend, or no-op, accordingly.
2442 SDValue N = getValue(I.getOperand(0));
2443 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2445 setValue(&I, DAG.getZExtOrTrunc(N, getCurSDLoc(), DestVT));
2448 void SelectionDAGBuilder::visitBitCast(const User &I) {
2449 SDValue N = getValue(I.getOperand(0));
2450 SDLoc dl = getCurSDLoc();
2451 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2454 // BitCast assures us that source and destination are the same size so this is
2455 // either a BITCAST or a no-op.
2456 if (DestVT != N.getValueType())
2457 setValue(&I, DAG.getNode(ISD::BITCAST, dl,
2458 DestVT, N)); // convert types.
2459 // Check if the original LLVM IR Operand was a ConstantInt, because getValue()
2460 // might fold any kind of constant expression to an integer constant and that
2461 // is not what we are looking for. Only regcognize a bitcast of a genuine
2462 // constant integer as an opaque constant.
2463 else if(ConstantInt *C = dyn_cast<ConstantInt>(I.getOperand(0)))
2464 setValue(&I, DAG.getConstant(C->getValue(), dl, DestVT, /*isTarget=*/false,
2467 setValue(&I, N); // noop cast.
2470 void SelectionDAGBuilder::visitAddrSpaceCast(const User &I) {
2471 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2472 const Value *SV = I.getOperand(0);
2473 SDValue N = getValue(SV);
2474 EVT DestVT = TLI.getValueType(DAG.getDataLayout(), I.getType());
2476 unsigned SrcAS = SV->getType()->getPointerAddressSpace();
2477 unsigned DestAS = I.getType()->getPointerAddressSpace();
2479 if (!TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
2480 N = DAG.getAddrSpaceCast(getCurSDLoc(), DestVT, N, SrcAS, DestAS);
2485 void SelectionDAGBuilder::visitInsertElement(const User &I) {
2486 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2487 SDValue InVec = getValue(I.getOperand(0));
2488 SDValue InVal = getValue(I.getOperand(1));
2489 SDValue InIdx = DAG.getSExtOrTrunc(getValue(I.getOperand(2)), getCurSDLoc(),
2490 TLI.getVectorIdxTy(DAG.getDataLayout()));
2491 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurSDLoc(),
2492 TLI.getValueType(DAG.getDataLayout(), I.getType()),
2493 InVec, InVal, InIdx));
2496 void SelectionDAGBuilder::visitExtractElement(const User &I) {
2497 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2498 SDValue InVec = getValue(I.getOperand(0));
2499 SDValue InIdx = DAG.getSExtOrTrunc(getValue(I.getOperand(1)), getCurSDLoc(),
2500 TLI.getVectorIdxTy(DAG.getDataLayout()));
2501 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurSDLoc(),
2502 TLI.getValueType(DAG.getDataLayout(), I.getType()),
2506 // Utility for visitShuffleVector - Return true if every element in Mask,
2507 // beginning from position Pos and ending in Pos+Size, falls within the
2508 // specified sequential range [L, L+Pos). or is undef.
2509 static bool isSequentialInRange(const SmallVectorImpl<int> &Mask,
2510 unsigned Pos, unsigned Size, int Low) {
2511 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
2512 if (Mask[i] >= 0 && Mask[i] != Low)
2517 void SelectionDAGBuilder::visitShuffleVector(const User &I) {
2518 SDValue Src1 = getValue(I.getOperand(0));
2519 SDValue Src2 = getValue(I.getOperand(1));
2521 SmallVector<int, 8> Mask;
2522 ShuffleVectorInst::getShuffleMask(cast<Constant>(I.getOperand(2)), Mask);
2523 unsigned MaskNumElts = Mask.size();
2525 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2526 EVT VT = TLI.getValueType(DAG.getDataLayout(), I.getType());
2527 EVT SrcVT = Src1.getValueType();
2528 unsigned SrcNumElts = SrcVT.getVectorNumElements();
2530 if (SrcNumElts == MaskNumElts) {
2531 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
2536 // Normalize the shuffle vector since mask and vector length don't match.
2537 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) {
2538 // Mask is longer than the source vectors and is a multiple of the source
2539 // vectors. We can use concatenate vector to make the mask and vectors
2541 if (SrcNumElts*2 == MaskNumElts) {
2542 // First check for Src1 in low and Src2 in high
2543 if (isSequentialInRange(Mask, 0, SrcNumElts, 0) &&
2544 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, SrcNumElts)) {
2545 // The shuffle is concatenating two vectors together.
2546 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurSDLoc(),
2550 // Then check for Src2 in low and Src1 in high
2551 if (isSequentialInRange(Mask, 0, SrcNumElts, SrcNumElts) &&
2552 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, 0)) {
2553 // The shuffle is concatenating two vectors together.
2554 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurSDLoc(),
2560 // Pad both vectors with undefs to make them the same length as the mask.
2561 unsigned NumConcat = MaskNumElts / SrcNumElts;
2562 bool Src1U = Src1.getOpcode() == ISD::UNDEF;
2563 bool Src2U = Src2.getOpcode() == ISD::UNDEF;
2564 SDValue UndefVal = DAG.getUNDEF(SrcVT);
2566 SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal);
2567 SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal);
2571 Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2572 getCurSDLoc(), VT, MOps1);
2573 Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2574 getCurSDLoc(), VT, MOps2);
2576 // Readjust mask for new input vector length.
2577 SmallVector<int, 8> MappedOps;
2578 for (unsigned i = 0; i != MaskNumElts; ++i) {
2580 if (Idx >= (int)SrcNumElts)
2581 Idx -= SrcNumElts - MaskNumElts;
2582 MappedOps.push_back(Idx);
2585 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
2590 if (SrcNumElts > MaskNumElts) {
2591 // Analyze the access pattern of the vector to see if we can extract
2592 // two subvectors and do the shuffle. The analysis is done by calculating
2593 // the range of elements the mask access on both vectors.
2594 int MinRange[2] = { static_cast<int>(SrcNumElts),
2595 static_cast<int>(SrcNumElts)};
2596 int MaxRange[2] = {-1, -1};
2598 for (unsigned i = 0; i != MaskNumElts; ++i) {
2604 if (Idx >= (int)SrcNumElts) {
2608 if (Idx > MaxRange[Input])
2609 MaxRange[Input] = Idx;
2610 if (Idx < MinRange[Input])
2611 MinRange[Input] = Idx;
2614 // Check if the access is smaller than the vector size and can we find
2615 // a reasonable extract index.
2616 int RangeUse[2] = { -1, -1 }; // 0 = Unused, 1 = Extract, -1 = Can not
2618 int StartIdx[2]; // StartIdx to extract from
2619 for (unsigned Input = 0; Input < 2; ++Input) {
2620 if (MinRange[Input] >= (int)SrcNumElts && MaxRange[Input] < 0) {
2621 RangeUse[Input] = 0; // Unused
2622 StartIdx[Input] = 0;
2626 // Find a good start index that is a multiple of the mask length. Then
2627 // see if the rest of the elements are in range.
2628 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts;
2629 if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts &&
2630 StartIdx[Input] + MaskNumElts <= SrcNumElts)
2631 RangeUse[Input] = 1; // Extract from a multiple of the mask length.
2634 if (RangeUse[0] == 0 && RangeUse[1] == 0) {
2635 setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used.
2638 if (RangeUse[0] >= 0 && RangeUse[1] >= 0) {
2639 // Extract appropriate subvector and generate a vector shuffle
2640 for (unsigned Input = 0; Input < 2; ++Input) {
2641 SDValue &Src = Input == 0 ? Src1 : Src2;
2642 if (RangeUse[Input] == 0)
2643 Src = DAG.getUNDEF(VT);
2645 SDLoc dl = getCurSDLoc();
2647 ISD::EXTRACT_SUBVECTOR, dl, VT, Src,
2648 DAG.getConstant(StartIdx[Input], dl,
2649 TLI.getVectorIdxTy(DAG.getDataLayout())));
2653 // Calculate new mask.
2654 SmallVector<int, 8> MappedOps;
2655 for (unsigned i = 0; i != MaskNumElts; ++i) {
2658 if (Idx < (int)SrcNumElts)
2661 Idx -= SrcNumElts + StartIdx[1] - MaskNumElts;
2663 MappedOps.push_back(Idx);
2666 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
2672 // We can't use either concat vectors or extract subvectors so fall back to
2673 // replacing the shuffle with extract and build vector.
2674 // to insert and build vector.
2675 EVT EltVT = VT.getVectorElementType();
2676 EVT IdxVT = TLI.getVectorIdxTy(DAG.getDataLayout());
2677 SDLoc dl = getCurSDLoc();
2678 SmallVector<SDValue,8> Ops;
2679 for (unsigned i = 0; i != MaskNumElts; ++i) {
2684 Res = DAG.getUNDEF(EltVT);
2686 SDValue &Src = Idx < (int)SrcNumElts ? Src1 : Src2;
2687 if (Idx >= (int)SrcNumElts) Idx -= SrcNumElts;
2689 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
2690 EltVT, Src, DAG.getConstant(Idx, dl, IdxVT));
2696 setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops));
2699 void SelectionDAGBuilder::visitInsertValue(const InsertValueInst &I) {
2700 const Value *Op0 = I.getOperand(0);
2701 const Value *Op1 = I.getOperand(1);
2702 Type *AggTy = I.getType();
2703 Type *ValTy = Op1->getType();
2704 bool IntoUndef = isa<UndefValue>(Op0);
2705 bool FromUndef = isa<UndefValue>(Op1);
2707 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
2709 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2710 SmallVector<EVT, 4> AggValueVTs;
2711 ComputeValueVTs(TLI, DAG.getDataLayout(), AggTy, AggValueVTs);
2712 SmallVector<EVT, 4> ValValueVTs;
2713 ComputeValueVTs(TLI, DAG.getDataLayout(), ValTy, ValValueVTs);
2715 unsigned NumAggValues = AggValueVTs.size();
2716 unsigned NumValValues = ValValueVTs.size();
2717 SmallVector<SDValue, 4> Values(NumAggValues);
2719 // Ignore an insertvalue that produces an empty object
2720 if (!NumAggValues) {
2721 setValue(&I, DAG.getUNDEF(MVT(MVT::Other)));
2725 SDValue Agg = getValue(Op0);
2727 // Copy the beginning value(s) from the original aggregate.
2728 for (; i != LinearIndex; ++i)
2729 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2730 SDValue(Agg.getNode(), Agg.getResNo() + i);
2731 // Copy values from the inserted value(s).
2733 SDValue Val = getValue(Op1);
2734 for (; i != LinearIndex + NumValValues; ++i)
2735 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2736 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex);
2738 // Copy remaining value(s) from the original aggregate.
2739 for (; i != NumAggValues; ++i)
2740 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2741 SDValue(Agg.getNode(), Agg.getResNo() + i);
2743 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
2744 DAG.getVTList(AggValueVTs), Values));
2747 void SelectionDAGBuilder::visitExtractValue(const ExtractValueInst &I) {
2748 const Value *Op0 = I.getOperand(0);
2749 Type *AggTy = Op0->getType();
2750 Type *ValTy = I.getType();
2751 bool OutOfUndef = isa<UndefValue>(Op0);
2753 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
2755 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2756 SmallVector<EVT, 4> ValValueVTs;
2757 ComputeValueVTs(TLI, DAG.getDataLayout(), ValTy, ValValueVTs);
2759 unsigned NumValValues = ValValueVTs.size();
2761 // Ignore a extractvalue that produces an empty object
2762 if (!NumValValues) {
2763 setValue(&I, DAG.getUNDEF(MVT(MVT::Other)));
2767 SmallVector<SDValue, 4> Values(NumValValues);
2769 SDValue Agg = getValue(Op0);
2770 // Copy out the selected value(s).
2771 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i)
2772 Values[i - LinearIndex] =
2774 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) :
2775 SDValue(Agg.getNode(), Agg.getResNo() + i);
2777 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
2778 DAG.getVTList(ValValueVTs), Values));
2781 void SelectionDAGBuilder::visitGetElementPtr(const User &I) {
2782 Value *Op0 = I.getOperand(0);
2783 // Note that the pointer operand may be a vector of pointers. Take the scalar
2784 // element which holds a pointer.
2785 Type *Ty = Op0->getType()->getScalarType();
2786 unsigned AS = Ty->getPointerAddressSpace();
2787 SDValue N = getValue(Op0);
2788 SDLoc dl = getCurSDLoc();
2790 for (GetElementPtrInst::const_op_iterator OI = I.op_begin()+1, E = I.op_end();
2792 const Value *Idx = *OI;
2793 if (StructType *StTy = dyn_cast<StructType>(Ty)) {
2794 unsigned Field = cast<Constant>(Idx)->getUniqueInteger().getZExtValue();
2797 uint64_t Offset = DL->getStructLayout(StTy)->getElementOffset(Field);
2798 N = DAG.getNode(ISD::ADD, dl, N.getValueType(), N,
2799 DAG.getConstant(Offset, dl, N.getValueType()));
2802 Ty = StTy->getElementType(Field);
2804 Ty = cast<SequentialType>(Ty)->getElementType();
2806 DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout(), AS);
2807 unsigned PtrSize = PtrTy.getSizeInBits();
2808 APInt ElementSize(PtrSize, DL->getTypeAllocSize(Ty));
2810 // If this is a constant subscript, handle it quickly.
2811 if (const auto *CI = dyn_cast<ConstantInt>(Idx)) {
2814 APInt Offs = ElementSize * CI->getValue().sextOrTrunc(PtrSize);
2815 SDValue OffsVal = DAG.getConstant(Offs, dl, PtrTy);
2816 N = DAG.getNode(ISD::ADD, dl, N.getValueType(), N, OffsVal);
2820 // N = N + Idx * ElementSize;
2821 SDValue IdxN = getValue(Idx);
2823 // If the index is smaller or larger than intptr_t, truncate or extend
2825 IdxN = DAG.getSExtOrTrunc(IdxN, dl, N.getValueType());
2827 // If this is a multiply by a power of two, turn it into a shl
2828 // immediately. This is a very common case.
2829 if (ElementSize != 1) {
2830 if (ElementSize.isPowerOf2()) {
2831 unsigned Amt = ElementSize.logBase2();
2832 IdxN = DAG.getNode(ISD::SHL, dl,
2833 N.getValueType(), IdxN,
2834 DAG.getConstant(Amt, dl, IdxN.getValueType()));
2836 SDValue Scale = DAG.getConstant(ElementSize, dl, IdxN.getValueType());
2837 IdxN = DAG.getNode(ISD::MUL, dl,
2838 N.getValueType(), IdxN, Scale);
2842 N = DAG.getNode(ISD::ADD, dl,
2843 N.getValueType(), N, IdxN);
2850 void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) {
2851 // If this is a fixed sized alloca in the entry block of the function,
2852 // allocate it statically on the stack.
2853 if (FuncInfo.StaticAllocaMap.count(&I))
2854 return; // getValue will auto-populate this.
2856 SDLoc dl = getCurSDLoc();
2857 Type *Ty = I.getAllocatedType();
2858 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2859 auto &DL = DAG.getDataLayout();
2860 uint64_t TySize = DL.getTypeAllocSize(Ty);
2862 std::max((unsigned)DL.getPrefTypeAlignment(Ty), I.getAlignment());
2864 SDValue AllocSize = getValue(I.getArraySize());
2866 EVT IntPtr = TLI.getPointerTy(DAG.getDataLayout());
2867 if (AllocSize.getValueType() != IntPtr)
2868 AllocSize = DAG.getZExtOrTrunc(AllocSize, dl, IntPtr);
2870 AllocSize = DAG.getNode(ISD::MUL, dl, IntPtr,
2872 DAG.getConstant(TySize, dl, IntPtr));
2874 // Handle alignment. If the requested alignment is less than or equal to
2875 // the stack alignment, ignore it. If the size is greater than or equal to
2876 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
2877 unsigned StackAlign =
2878 DAG.getSubtarget().getFrameLowering()->getStackAlignment();
2879 if (Align <= StackAlign)
2882 // Round the size of the allocation up to the stack alignment size
2883 // by add SA-1 to the size.
2884 AllocSize = DAG.getNode(ISD::ADD, dl,
2885 AllocSize.getValueType(), AllocSize,
2886 DAG.getIntPtrConstant(StackAlign - 1, dl));
2888 // Mask out the low bits for alignment purposes.
2889 AllocSize = DAG.getNode(ISD::AND, dl,
2890 AllocSize.getValueType(), AllocSize,
2891 DAG.getIntPtrConstant(~(uint64_t)(StackAlign - 1),
2894 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align, dl) };
2895 SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other);
2896 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, dl, VTs, Ops);
2898 DAG.setRoot(DSA.getValue(1));
2900 assert(FuncInfo.MF->getFrameInfo()->hasVarSizedObjects());
2903 void SelectionDAGBuilder::visitLoad(const LoadInst &I) {
2905 return visitAtomicLoad(I);
2907 const Value *SV = I.getOperand(0);
2908 SDValue Ptr = getValue(SV);
2910 Type *Ty = I.getType();
2912 bool isVolatile = I.isVolatile();
2913 bool isNonTemporal = I.getMetadata(LLVMContext::MD_nontemporal) != nullptr;
2915 // The IR notion of invariant_load only guarantees that all *non-faulting*
2916 // invariant loads result in the same value. The MI notion of invariant load
2917 // guarantees that the load can be legally moved to any location within its
2918 // containing function. The MI notion of invariant_load is stronger than the
2919 // IR notion of invariant_load -- an MI invariant_load is an IR invariant_load
2920 // with a guarantee that the location being loaded from is dereferenceable
2921 // throughout the function's lifetime.
2923 bool isInvariant = I.getMetadata(LLVMContext::MD_invariant_load) != nullptr &&
2924 isDereferenceablePointer(SV, *DAG.getTarget().getDataLayout());
2925 unsigned Alignment = I.getAlignment();
2928 I.getAAMetadata(AAInfo);
2929 const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range);
2931 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2932 SmallVector<EVT, 4> ValueVTs;
2933 SmallVector<uint64_t, 4> Offsets;
2934 ComputeValueVTs(TLI, DAG.getDataLayout(), Ty, ValueVTs, &Offsets);
2935 unsigned NumValues = ValueVTs.size();
2940 bool ConstantMemory = false;
2941 if (isVolatile || NumValues > MaxParallelChains)
2942 // Serialize volatile loads with other side effects.
2944 else if (AA->pointsToConstantMemory(
2945 MemoryLocation(SV, AA->getTypeStoreSize(Ty), AAInfo))) {
2946 // Do not serialize (non-volatile) loads of constant memory with anything.
2947 Root = DAG.getEntryNode();
2948 ConstantMemory = true;
2950 // Do not serialize non-volatile loads against each other.
2951 Root = DAG.getRoot();
2954 SDLoc dl = getCurSDLoc();
2957 Root = TLI.prepareVolatileOrAtomicLoad(Root, dl, DAG);
2959 SmallVector<SDValue, 4> Values(NumValues);
2960 SmallVector<SDValue, 4> Chains(std::min(MaxParallelChains, NumValues));
2961 EVT PtrVT = Ptr.getValueType();
2962 unsigned ChainI = 0;
2963 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
2964 // Serializing loads here may result in excessive register pressure, and
2965 // TokenFactor places arbitrary choke points on the scheduler. SD scheduling
2966 // could recover a bit by hoisting nodes upward in the chain by recognizing
2967 // they are side-effect free or do not alias. The optimizer should really
2968 // avoid this case by converting large object/array copies to llvm.memcpy
2969 // (MaxParallelChains should always remain as failsafe).
2970 if (ChainI == MaxParallelChains) {
2971 assert(PendingLoads.empty() && "PendingLoads must be serialized first");
2972 SDValue Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2973 makeArrayRef(Chains.data(), ChainI));
2977 SDValue A = DAG.getNode(ISD::ADD, dl,
2979 DAG.getConstant(Offsets[i], dl, PtrVT));
2980 SDValue L = DAG.getLoad(ValueVTs[i], dl, Root,
2981 A, MachinePointerInfo(SV, Offsets[i]), isVolatile,
2982 isNonTemporal, isInvariant, Alignment, AAInfo,
2986 Chains[ChainI] = L.getValue(1);
2989 if (!ConstantMemory) {
2990 SDValue Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2991 makeArrayRef(Chains.data(), ChainI));
2995 PendingLoads.push_back(Chain);
2998 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, dl,
2999 DAG.getVTList(ValueVTs), Values));
3002 void SelectionDAGBuilder::visitStore(const StoreInst &I) {
3004 return visitAtomicStore(I);
3006 const Value *SrcV = I.getOperand(0);
3007 const Value *PtrV = I.getOperand(1);
3009 SmallVector<EVT, 4> ValueVTs;
3010 SmallVector<uint64_t, 4> Offsets;
3011 ComputeValueVTs(DAG.getTargetLoweringInfo(), DAG.getDataLayout(),
3012 SrcV->getType(), ValueVTs, &Offsets);
3013 unsigned NumValues = ValueVTs.size();
3017 // Get the lowered operands. Note that we do this after
3018 // checking if NumResults is zero, because with zero results
3019 // the operands won't have values in the map.
3020 SDValue Src = getValue(SrcV);
3021 SDValue Ptr = getValue(PtrV);
3023 SDValue Root = getRoot();
3024 SmallVector<SDValue, 4> Chains(std::min(MaxParallelChains, NumValues));
3025 EVT PtrVT = Ptr.getValueType();
3026 bool isVolatile = I.isVolatile();
3027 bool isNonTemporal = I.getMetadata(LLVMContext::MD_nontemporal) != nullptr;
3028 unsigned Alignment = I.getAlignment();
3029 SDLoc dl = getCurSDLoc();
3032 I.getAAMetadata(AAInfo);
3034 unsigned ChainI = 0;
3035 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3036 // See visitLoad comments.
3037 if (ChainI == MaxParallelChains) {
3038 SDValue Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3039 makeArrayRef(Chains.data(), ChainI));
3043 SDValue Add = DAG.getNode(ISD::ADD, dl, PtrVT, Ptr,
3044 DAG.getConstant(Offsets[i], dl, PtrVT));
3045 SDValue St = DAG.getStore(Root, dl,
3046 SDValue(Src.getNode(), Src.getResNo() + i),
3047 Add, MachinePointerInfo(PtrV, Offsets[i]),
3048 isVolatile, isNonTemporal, Alignment, AAInfo);
3049 Chains[ChainI] = St;
3052 SDValue StoreNode = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3053 makeArrayRef(Chains.data(), ChainI));
3054 DAG.setRoot(StoreNode);
3057 void SelectionDAGBuilder::visitMaskedStore(const CallInst &I) {
3058 SDLoc sdl = getCurSDLoc();
3060 // llvm.masked.store.*(Src0, Ptr, alignemt, Mask)
3061 Value *PtrOperand = I.getArgOperand(1);
3062 SDValue Ptr = getValue(PtrOperand);
3063 SDValue Src0 = getValue(I.getArgOperand(0));
3064 SDValue Mask = getValue(I.getArgOperand(3));
3065 EVT VT = Src0.getValueType();
3066 unsigned Alignment = (cast<ConstantInt>(I.getArgOperand(2)))->getZExtValue();
3068 Alignment = DAG.getEVTAlignment(VT);
3071 I.getAAMetadata(AAInfo);
3073 MachineMemOperand *MMO =
3074 DAG.getMachineFunction().
3075 getMachineMemOperand(MachinePointerInfo(PtrOperand),
3076 MachineMemOperand::MOStore, VT.getStoreSize(),
3078 SDValue StoreNode = DAG.getMaskedStore(getRoot(), sdl, Src0, Ptr, Mask, VT,
3080 DAG.setRoot(StoreNode);
3081 setValue(&I, StoreNode);
3084 // Gather/scatter receive a vector of pointers.
3085 // This vector of pointers may be represented as a base pointer + vector of
3086 // indices, it depends on GEP and instruction preceeding GEP
3087 // that calculates indices
3088 static bool getUniformBase(Value *& Ptr, SDValue& Base, SDValue& Index,
3089 SelectionDAGBuilder* SDB) {
3091 assert (Ptr->getType()->isVectorTy() && "Uexpected pointer type");
3092 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
3093 if (!Gep || Gep->getNumOperands() > 2)
3095 ShuffleVectorInst *ShuffleInst =
3096 dyn_cast<ShuffleVectorInst>(Gep->getPointerOperand());
3097 if (!ShuffleInst || !ShuffleInst->getMask()->isNullValue() ||
3098 cast<Instruction>(ShuffleInst->getOperand(0))->getOpcode() !=
3099 Instruction::InsertElement)
3102 Ptr = cast<InsertElementInst>(ShuffleInst->getOperand(0))->getOperand(1);
3104 SelectionDAG& DAG = SDB->DAG;
3105 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3106 // Check is the Ptr is inside current basic block
3107 // If not, look for the shuffle instruction
3108 if (SDB->findValue(Ptr))
3109 Base = SDB->getValue(Ptr);
3110 else if (SDB->findValue(ShuffleInst)) {
3111 SDValue ShuffleNode = SDB->getValue(ShuffleInst);
3112 SDLoc sdl = ShuffleNode;
3114 ISD::EXTRACT_VECTOR_ELT, sdl,
3115 ShuffleNode.getValueType().getScalarType(), ShuffleNode,
3116 DAG.getConstant(0, sdl, TLI.getVectorIdxTy(DAG.getDataLayout())));
3117 SDB->setValue(Ptr, Base);
3122 Value *IndexVal = Gep->getOperand(1);
3123 if (SDB->findValue(IndexVal)) {
3124 Index = SDB->getValue(IndexVal);
3126 if (SExtInst* Sext = dyn_cast<SExtInst>(IndexVal)) {
3127 IndexVal = Sext->getOperand(0);
3128 if (SDB->findValue(IndexVal))
3129 Index = SDB->getValue(IndexVal);
3136 void SelectionDAGBuilder::visitMaskedScatter(const CallInst &I) {
3137 SDLoc sdl = getCurSDLoc();
3139 // llvm.masked.scatter.*(Src0, Ptrs, alignemt, Mask)
3140 Value *Ptr = I.getArgOperand(1);
3141 SDValue Src0 = getValue(I.getArgOperand(0));
3142 SDValue Mask = getValue(I.getArgOperand(3));
3143 EVT VT = Src0.getValueType();
3144 unsigned Alignment = (cast<ConstantInt>(I.getArgOperand(2)))->getZExtValue();
3146 Alignment = DAG.getEVTAlignment(VT);
3147 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3150 I.getAAMetadata(AAInfo);
3154 Value *BasePtr = Ptr;
3155 bool UniformBase = getUniformBase(BasePtr, Base, Index, this);
3157 Value *MemOpBasePtr = UniformBase ? BasePtr : nullptr;
3158 MachineMemOperand *MMO = DAG.getMachineFunction().
3159 getMachineMemOperand(MachinePointerInfo(MemOpBasePtr),
3160 MachineMemOperand::MOStore, VT.getStoreSize(),
3163 Base = DAG.getTargetConstant(0, sdl, TLI.getPointerTy(DAG.getDataLayout()));
3164 Index = getValue(Ptr);
3166 SDValue Ops[] = { getRoot(), Src0, Mask, Base, Index };
3167 SDValue Scatter = DAG.getMaskedScatter(DAG.getVTList(MVT::Other), VT, sdl,
3169 DAG.setRoot(Scatter);
3170 setValue(&I, Scatter);
3173 void SelectionDAGBuilder::visitMaskedLoad(const CallInst &I) {
3174 SDLoc sdl = getCurSDLoc();
3176 // @llvm.masked.load.*(Ptr, alignment, Mask, Src0)
3177 Value *PtrOperand = I.getArgOperand(0);
3178 SDValue Ptr = getValue(PtrOperand);
3179 SDValue Src0 = getValue(I.getArgOperand(3));
3180 SDValue Mask = getValue(I.getArgOperand(2));
3182 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3183 EVT VT = TLI.getValueType(DAG.getDataLayout(), I.getType());
3184 unsigned Alignment = (cast<ConstantInt>(I.getArgOperand(1)))->getZExtValue();
3186 Alignment = DAG.getEVTAlignment(VT);
3189 I.getAAMetadata(AAInfo);
3190 const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range);
3192 SDValue InChain = DAG.getRoot();
3193 if (AA->pointsToConstantMemory(MemoryLocation(
3194 PtrOperand, AA->getTypeStoreSize(I.getType()), AAInfo))) {
3195 // Do not serialize (non-volatile) loads of constant memory with anything.
3196 InChain = DAG.getEntryNode();
3199 MachineMemOperand *MMO =
3200 DAG.getMachineFunction().
3201 getMachineMemOperand(MachinePointerInfo(PtrOperand),
3202 MachineMemOperand::MOLoad, VT.getStoreSize(),
3203 Alignment, AAInfo, Ranges);
3205 SDValue Load = DAG.getMaskedLoad(VT, sdl, InChain, Ptr, Mask, Src0, VT, MMO,
3207 SDValue OutChain = Load.getValue(1);
3208 DAG.setRoot(OutChain);
3212 void SelectionDAGBuilder::visitMaskedGather(const CallInst &I) {
3213 SDLoc sdl = getCurSDLoc();
3215 // @llvm.masked.gather.*(Ptrs, alignment, Mask, Src0)
3216 Value *Ptr = I.getArgOperand(0);
3217 SDValue Src0 = getValue(I.getArgOperand(3));
3218 SDValue Mask = getValue(I.getArgOperand(2));
3220 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3221 EVT VT = TLI.getValueType(DAG.getDataLayout(), I.getType());
3222 unsigned Alignment = (cast<ConstantInt>(I.getArgOperand(1)))->getZExtValue();
3224 Alignment = DAG.getEVTAlignment(VT);
3227 I.getAAMetadata(AAInfo);
3228 const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range);
3230 SDValue Root = DAG.getRoot();
3233 Value *BasePtr = Ptr;
3234 bool UniformBase = getUniformBase(BasePtr, Base, Index, this);
3235 bool ConstantMemory = false;
3237 AA->pointsToConstantMemory(
3238 MemoryLocation(BasePtr, AA->getTypeStoreSize(I.getType()), AAInfo))) {
3239 // Do not serialize (non-volatile) loads of constant memory with anything.
3240 Root = DAG.getEntryNode();
3241 ConstantMemory = true;
3244 MachineMemOperand *MMO =
3245 DAG.getMachineFunction().
3246 getMachineMemOperand(MachinePointerInfo(UniformBase ? BasePtr : nullptr),
3247 MachineMemOperand::MOLoad, VT.getStoreSize(),
3248 Alignment, AAInfo, Ranges);
3251 Base = DAG.getTargetConstant(0, sdl, TLI.getPointerTy(DAG.getDataLayout()));
3252 Index = getValue(Ptr);
3254 SDValue Ops[] = { Root, Src0, Mask, Base, Index };
3255 SDValue Gather = DAG.getMaskedGather(DAG.getVTList(VT, MVT::Other), VT, sdl,
3258 SDValue OutChain = Gather.getValue(1);
3259 if (!ConstantMemory)
3260 PendingLoads.push_back(OutChain);
3261 setValue(&I, Gather);
3264 void SelectionDAGBuilder::visitAtomicCmpXchg(const AtomicCmpXchgInst &I) {
3265 SDLoc dl = getCurSDLoc();
3266 AtomicOrdering SuccessOrder = I.getSuccessOrdering();
3267 AtomicOrdering FailureOrder = I.getFailureOrdering();
3268 SynchronizationScope Scope = I.getSynchScope();
3270 SDValue InChain = getRoot();
3272 MVT MemVT = getValue(I.getCompareOperand()).getSimpleValueType();
3273 SDVTList VTs = DAG.getVTList(MemVT, MVT::i1, MVT::Other);
3274 SDValue L = DAG.getAtomicCmpSwap(
3275 ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, dl, MemVT, VTs, InChain,
3276 getValue(I.getPointerOperand()), getValue(I.getCompareOperand()),
3277 getValue(I.getNewValOperand()), MachinePointerInfo(I.getPointerOperand()),
3278 /*Alignment=*/ 0, SuccessOrder, FailureOrder, Scope);
3280 SDValue OutChain = L.getValue(2);
3283 DAG.setRoot(OutChain);
3286 void SelectionDAGBuilder::visitAtomicRMW(const AtomicRMWInst &I) {
3287 SDLoc dl = getCurSDLoc();
3289 switch (I.getOperation()) {
3290 default: llvm_unreachable("Unknown atomicrmw operation");
3291 case AtomicRMWInst::Xchg: NT = ISD::ATOMIC_SWAP; break;
3292 case AtomicRMWInst::Add: NT = ISD::ATOMIC_LOAD_ADD; break;
3293 case AtomicRMWInst::Sub: NT = ISD::ATOMIC_LOAD_SUB; break;
3294 case AtomicRMWInst::And: NT = ISD::ATOMIC_LOAD_AND; break;
3295 case AtomicRMWInst::Nand: NT = ISD::ATOMIC_LOAD_NAND; break;
3296 case AtomicRMWInst::Or: NT = ISD::ATOMIC_LOAD_OR; break;
3297 case AtomicRMWInst::Xor: NT = ISD::ATOMIC_LOAD_XOR; break;
3298 case AtomicRMWInst::Max: NT = ISD::ATOMIC_LOAD_MAX; break;
3299 case AtomicRMWInst::Min: NT = ISD::ATOMIC_LOAD_MIN; break;
3300 case AtomicRMWInst::UMax: NT = ISD::ATOMIC_LOAD_UMAX; break;
3301 case AtomicRMWInst::UMin: NT = ISD::ATOMIC_LOAD_UMIN; break;
3303 AtomicOrdering Order = I.getOrdering();
3304 SynchronizationScope Scope = I.getSynchScope();
3306 SDValue InChain = getRoot();
3309 DAG.getAtomic(NT, dl,
3310 getValue(I.getValOperand()).getSimpleValueType(),
3312 getValue(I.getPointerOperand()),
3313 getValue(I.getValOperand()),
3314 I.getPointerOperand(),
3315 /* Alignment=*/ 0, Order, Scope);
3317 SDValue OutChain = L.getValue(1);
3320 DAG.setRoot(OutChain);
3323 void SelectionDAGBuilder::visitFence(const FenceInst &I) {
3324 SDLoc dl = getCurSDLoc();
3325 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3328 Ops[1] = DAG.getConstant(I.getOrdering(), dl,
3329 TLI.getPointerTy(DAG.getDataLayout()));
3330 Ops[2] = DAG.getConstant(I.getSynchScope(), dl,
3331 TLI.getPointerTy(DAG.getDataLayout()));
3332 DAG.setRoot(DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops));
3335 void SelectionDAGBuilder::visitAtomicLoad(const LoadInst &I) {
3336 SDLoc dl = getCurSDLoc();
3337 AtomicOrdering Order = I.getOrdering();
3338 SynchronizationScope Scope = I.getSynchScope();
3340 SDValue InChain = getRoot();
3342 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3343 EVT VT = TLI.getValueType(DAG.getDataLayout(), I.getType());
3345 if (I.getAlignment() < VT.getSizeInBits() / 8)
3346 report_fatal_error("Cannot generate unaligned atomic load");
3348 MachineMemOperand *MMO =
3349 DAG.getMachineFunction().
3350 getMachineMemOperand(MachinePointerInfo(I.getPointerOperand()),
3351 MachineMemOperand::MOVolatile |
3352 MachineMemOperand::MOLoad,
3354 I.getAlignment() ? I.getAlignment() :
3355 DAG.getEVTAlignment(VT));
3357 InChain = TLI.prepareVolatileOrAtomicLoad(InChain, dl, DAG);
3359 DAG.getAtomic(ISD::ATOMIC_LOAD, dl, VT, VT, InChain,
3360 getValue(I.getPointerOperand()), MMO,
3363 SDValue OutChain = L.getValue(1);
3366 DAG.setRoot(OutChain);
3369 void SelectionDAGBuilder::visitAtomicStore(const StoreInst &I) {
3370 SDLoc dl = getCurSDLoc();
3372 AtomicOrdering Order = I.getOrdering();
3373 SynchronizationScope Scope = I.getSynchScope();
3375 SDValue InChain = getRoot();
3377 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3379 TLI.getValueType(DAG.getDataLayout(), I.getValueOperand()->getType());
3381 if (I.getAlignment() < VT.getSizeInBits() / 8)
3382 report_fatal_error("Cannot generate unaligned atomic store");
3385 DAG.getAtomic(ISD::ATOMIC_STORE, dl, VT,
3387 getValue(I.getPointerOperand()),
3388 getValue(I.getValueOperand()),
3389 I.getPointerOperand(), I.getAlignment(),
3392 DAG.setRoot(OutChain);
3395 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
3397 void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I,
3398 unsigned Intrinsic) {
3399 bool HasChain = !I.doesNotAccessMemory();
3400 bool OnlyLoad = HasChain && I.onlyReadsMemory();
3402 // Build the operand list.
3403 SmallVector<SDValue, 8> Ops;
3404 if (HasChain) { // If this intrinsic has side-effects, chainify it.
3406 // We don't need to serialize loads against other loads.
3407 Ops.push_back(DAG.getRoot());
3409 Ops.push_back(getRoot());
3413 // Info is set by getTgtMemInstrinsic
3414 TargetLowering::IntrinsicInfo Info;
3415 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3416 bool IsTgtIntrinsic = TLI.getTgtMemIntrinsic(Info, I, Intrinsic);
3418 // Add the intrinsic ID as an integer operand if it's not a target intrinsic.
3419 if (!IsTgtIntrinsic || Info.opc == ISD::INTRINSIC_VOID ||
3420 Info.opc == ISD::INTRINSIC_W_CHAIN)
3421 Ops.push_back(DAG.getTargetConstant(Intrinsic, getCurSDLoc(),
3422 TLI.getPointerTy(DAG.getDataLayout())));
3424 // Add all operands of the call to the operand list.
3425 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) {
3426 SDValue Op = getValue(I.getArgOperand(i));
3430 SmallVector<EVT, 4> ValueVTs;
3431 ComputeValueVTs(TLI, DAG.getDataLayout(), I.getType(), ValueVTs);
3434 ValueVTs.push_back(MVT::Other);
3436 SDVTList VTs = DAG.getVTList(ValueVTs);
3440 if (IsTgtIntrinsic) {
3441 // This is target intrinsic that touches memory
3442 Result = DAG.getMemIntrinsicNode(Info.opc, getCurSDLoc(),
3443 VTs, Ops, Info.memVT,
3444 MachinePointerInfo(Info.ptrVal, Info.offset),
3445 Info.align, Info.vol,
3446 Info.readMem, Info.writeMem, Info.size);
3447 } else if (!HasChain) {
3448 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurSDLoc(), VTs, Ops);
3449 } else if (!I.getType()->isVoidTy()) {
3450 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurSDLoc(), VTs, Ops);
3452 Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurSDLoc(), VTs, Ops);
3456 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1);
3458 PendingLoads.push_back(Chain);
3463 if (!I.getType()->isVoidTy()) {
3464 if (VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
3465 EVT VT = TLI.getValueType(DAG.getDataLayout(), PTy);
3466 Result = DAG.getNode(ISD::BITCAST, getCurSDLoc(), VT, Result);
3469 setValue(&I, Result);
3473 /// GetSignificand - Get the significand and build it into a floating-point
3474 /// number with exponent of 1:
3476 /// Op = (Op & 0x007fffff) | 0x3f800000;
3478 /// where Op is the hexadecimal representation of floating point value.
3480 GetSignificand(SelectionDAG &DAG, SDValue Op, SDLoc dl) {
3481 SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3482 DAG.getConstant(0x007fffff, dl, MVT::i32));
3483 SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1,
3484 DAG.getConstant(0x3f800000, dl, MVT::i32));
3485 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, t2);
3488 /// GetExponent - Get the exponent:
3490 /// (float)(int)(((Op & 0x7f800000) >> 23) - 127);
3492 /// where Op is the hexadecimal representation of floating point value.
3494 GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI,
3496 SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3497 DAG.getConstant(0x7f800000, dl, MVT::i32));
3498 SDValue t1 = DAG.getNode(
3499 ISD::SRL, dl, MVT::i32, t0,
3500 DAG.getConstant(23, dl, TLI.getPointerTy(DAG.getDataLayout())));
3501 SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1,
3502 DAG.getConstant(127, dl, MVT::i32));
3503 return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2);
3506 /// getF32Constant - Get 32-bit floating point constant.
3508 getF32Constant(SelectionDAG &DAG, unsigned Flt, SDLoc dl) {
3509 return DAG.getConstantFP(APFloat(APFloat::IEEEsingle, APInt(32, Flt)), dl,
3513 static SDValue getLimitedPrecisionExp2(SDValue t0, SDLoc dl,
3514 SelectionDAG &DAG) {
3515 // IntegerPartOfX = ((int32_t)(t0);
3516 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3518 // FractionalPartOfX = t0 - (float)IntegerPartOfX;
3519 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3520 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3522 // IntegerPartOfX <<= 23;
3523 IntegerPartOfX = DAG.getNode(
3524 ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3525 DAG.getConstant(23, dl, DAG.getTargetLoweringInfo().getPointerTy(
3526 DAG.getDataLayout())));
3528 SDValue TwoToFractionalPartOfX;
3529 if (LimitFloatPrecision <= 6) {
3530 // For floating-point precision of 6:
3532 // TwoToFractionalPartOfX =
3534 // (0.735607626f + 0.252464424f * x) * x;
3536 // error 0.0144103317, which is 6 bits
3537 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3538 getF32Constant(DAG, 0x3e814304, dl));
3539 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3540 getF32Constant(DAG, 0x3f3c50c8, dl));
3541 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3542 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3543 getF32Constant(DAG, 0x3f7f5e7e, dl));
3544 } else if (LimitFloatPrecision <= 12) {
3545 // For floating-point precision of 12:
3547 // TwoToFractionalPartOfX =
3550 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3552 // error 0.000107046256, which is 13 to 14 bits
3553 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3554 getF32Constant(DAG, 0x3da235e3, dl));
3555 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3556 getF32Constant(DAG, 0x3e65b8f3, dl));
3557 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3558 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3559 getF32Constant(DAG, 0x3f324b07, dl));
3560 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3561 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3562 getF32Constant(DAG, 0x3f7ff8fd, dl));
3563 } else { // LimitFloatPrecision <= 18
3564 // For floating-point precision of 18:
3566 // TwoToFractionalPartOfX =
3570 // (0.554906021e-1f +
3571 // (0.961591928e-2f +
3572 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3573 // error 2.47208000*10^(-7), which is better than 18 bits
3574 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3575 getF32Constant(DAG, 0x3924b03e, dl));
3576 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3577 getF32Constant(DAG, 0x3ab24b87, dl));
3578 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3579 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3580 getF32Constant(DAG, 0x3c1d8c17, dl));
3581 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3582 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3583 getF32Constant(DAG, 0x3d634a1d, dl));
3584 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3585 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3586 getF32Constant(DAG, 0x3e75fe14, dl));
3587 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3588 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3589 getF32Constant(DAG, 0x3f317234, dl));
3590 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3591 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3592 getF32Constant(DAG, 0x3f800000, dl));
3595 // Add the exponent into the result in integer domain.
3596 SDValue t13 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, TwoToFractionalPartOfX);
3597 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3598 DAG.getNode(ISD::ADD, dl, MVT::i32, t13, IntegerPartOfX));
3601 /// expandExp - Lower an exp intrinsic. Handles the special sequences for
3602 /// limited-precision mode.
3603 static SDValue expandExp(SDLoc dl, SDValue Op, SelectionDAG &DAG,
3604 const TargetLowering &TLI) {
3605 if (Op.getValueType() == MVT::f32 &&
3606 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3608 // Put the exponent in the right bit position for later addition to the
3611 // #define LOG2OFe 1.4426950f
3612 // t0 = Op * LOG2OFe
3613 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3614 getF32Constant(DAG, 0x3fb8aa3b, dl));
3615 return getLimitedPrecisionExp2(t0, dl, DAG);
3618 // No special expansion.
3619 return DAG.getNode(ISD::FEXP, dl, Op.getValueType(), Op);
3622 /// expandLog - Lower a log intrinsic. Handles the special sequences for
3623 /// limited-precision mode.
3624 static SDValue expandLog(SDLoc dl, SDValue Op, SelectionDAG &DAG,
3625 const TargetLowering &TLI) {
3626 if (Op.getValueType() == MVT::f32 &&
3627 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3628 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3630 // Scale the exponent by log(2) [0.69314718f].
3631 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3632 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3633 getF32Constant(DAG, 0x3f317218, dl));
3635 // Get the significand and build it into a floating-point number with
3637 SDValue X = GetSignificand(DAG, Op1, dl);
3639 SDValue LogOfMantissa;
3640 if (LimitFloatPrecision <= 6) {
3641 // For floating-point precision of 6:
3645 // (1.4034025f - 0.23903021f * x) * x;
3647 // error 0.0034276066, which is better than 8 bits
3648 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3649 getF32Constant(DAG, 0xbe74c456, dl));
3650 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3651 getF32Constant(DAG, 0x3fb3a2b1, dl));
3652 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3653 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3654 getF32Constant(DAG, 0x3f949a29, dl));
3655 } else if (LimitFloatPrecision <= 12) {
3656 // For floating-point precision of 12:
3662 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x;
3664 // error 0.000061011436, which is 14 bits
3665 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3666 getF32Constant(DAG, 0xbd67b6d6, dl));
3667 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3668 getF32Constant(DAG, 0x3ee4f4b8, dl));
3669 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3670 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3671 getF32Constant(DAG, 0x3fbc278b, dl));
3672 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3673 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3674 getF32Constant(DAG, 0x40348e95, dl));
3675 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3676 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3677 getF32Constant(DAG, 0x3fdef31a, dl));
3678 } else { // LimitFloatPrecision <= 18
3679 // For floating-point precision of 18:
3687 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x;
3689 // error 0.0000023660568, which is better than 18 bits
3690 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3691 getF32Constant(DAG, 0xbc91e5ac, dl));
3692 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3693 getF32Constant(DAG, 0x3e4350aa, dl));
3694 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3695 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3696 getF32Constant(DAG, 0x3f60d3e3, dl));
3697 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3698 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3699 getF32Constant(DAG, 0x4011cdf0, dl));
3700 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3701 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3702 getF32Constant(DAG, 0x406cfd1c, dl));
3703 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3704 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3705 getF32Constant(DAG, 0x408797cb, dl));
3706 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3707 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3708 getF32Constant(DAG, 0x4006dcab, dl));
3711 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, LogOfMantissa);
3714 // No special expansion.
3715 return DAG.getNode(ISD::FLOG, dl, Op.getValueType(), Op);
3718 /// expandLog2 - Lower a log2 intrinsic. Handles the special sequences for
3719 /// limited-precision mode.
3720 static SDValue expandLog2(SDLoc dl, SDValue Op, SelectionDAG &DAG,
3721 const TargetLowering &TLI) {
3722 if (Op.getValueType() == MVT::f32 &&
3723 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3724 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3726 // Get the exponent.
3727 SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl);
3729 // Get the significand and build it into a floating-point number with
3731 SDValue X = GetSignificand(DAG, Op1, dl);
3733 // Different possible minimax approximations of significand in
3734 // floating-point for various degrees of accuracy over [1,2].
3735 SDValue Log2ofMantissa;
3736 if (LimitFloatPrecision <= 6) {
3737 // For floating-point precision of 6:
3739 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x;
3741 // error 0.0049451742, which is more than 7 bits
3742 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3743 getF32Constant(DAG, 0xbeb08fe0, dl));
3744 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3745 getF32Constant(DAG, 0x40019463, dl));
3746 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3747 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3748 getF32Constant(DAG, 0x3fd6633d, dl));
3749 } else if (LimitFloatPrecision <= 12) {
3750 // For floating-point precision of 12:
3756 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x;
3758 // error 0.0000876136000, which is better than 13 bits
3759 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3760 getF32Constant(DAG, 0xbda7262e, dl));
3761 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3762 getF32Constant(DAG, 0x3f25280b, dl));
3763 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3764 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3765 getF32Constant(DAG, 0x4007b923, dl));
3766 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3767 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3768 getF32Constant(DAG, 0x40823e2f, dl));
3769 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3770 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3771 getF32Constant(DAG, 0x4020d29c, dl));
3772 } else { // LimitFloatPrecision <= 18
3773 // For floating-point precision of 18:
3782 // 0.25691327e-1f * x) * x) * x) * x) * x) * x;
3784 // error 0.0000018516, which is better than 18 bits
3785 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3786 getF32Constant(DAG, 0xbcd2769e, dl));
3787 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3788 getF32Constant(DAG, 0x3e8ce0b9, dl));
3789 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3790 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3791 getF32Constant(DAG, 0x3fa22ae7, dl));
3792 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3793 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3794 getF32Constant(DAG, 0x40525723, dl));
3795 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3796 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3797 getF32Constant(DAG, 0x40aaf200, dl));
3798 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3799 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3800 getF32Constant(DAG, 0x40c39dad, dl));
3801 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3802 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3803 getF32Constant(DAG, 0x4042902c, dl));
3806 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, Log2ofMantissa);
3809 // No special expansion.
3810 return DAG.getNode(ISD::FLOG2, dl, Op.getValueType(), Op);
3813 /// expandLog10 - Lower a log10 intrinsic. Handles the special sequences for
3814 /// limited-precision mode.
3815 static SDValue expandLog10(SDLoc dl, SDValue Op, SelectionDAG &DAG,
3816 const TargetLowering &TLI) {
3817 if (Op.getValueType() == MVT::f32 &&
3818 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3819 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3821 // Scale the exponent by log10(2) [0.30102999f].
3822 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3823 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3824 getF32Constant(DAG, 0x3e9a209a, dl));
3826 // Get the significand and build it into a floating-point number with
3828 SDValue X = GetSignificand(DAG, Op1, dl);
3830 SDValue Log10ofMantissa;
3831 if (LimitFloatPrecision <= 6) {
3832 // For floating-point precision of 6:
3834 // Log10ofMantissa =
3836 // (0.60948995f - 0.10380950f * x) * x;
3838 // error 0.0014886165, which is 6 bits
3839 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3840 getF32Constant(DAG, 0xbdd49a13, dl));
3841 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3842 getF32Constant(DAG, 0x3f1c0789, dl));
3843 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3844 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3845 getF32Constant(DAG, 0x3f011300, dl));
3846 } else if (LimitFloatPrecision <= 12) {
3847 // For floating-point precision of 12:
3849 // Log10ofMantissa =
3852 // (-0.31664806f + 0.47637168e-1f * x) * x) * x;
3854 // error 0.00019228036, which is better than 12 bits
3855 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3856 getF32Constant(DAG, 0x3d431f31, dl));
3857 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
3858 getF32Constant(DAG, 0x3ea21fb2, dl));
3859 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3860 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3861 getF32Constant(DAG, 0x3f6ae232, dl));
3862 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3863 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
3864 getF32Constant(DAG, 0x3f25f7c3, dl));
3865 } else { // LimitFloatPrecision <= 18
3866 // For floating-point precision of 18:
3868 // Log10ofMantissa =
3873 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x;
3875 // error 0.0000037995730, which is better than 18 bits
3876 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3877 getF32Constant(DAG, 0x3c5d51ce, dl));
3878 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
3879 getF32Constant(DAG, 0x3e00685a, dl));
3880 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3881 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3882 getF32Constant(DAG, 0x3efb6798, dl));
3883 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3884 SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
3885 getF32Constant(DAG, 0x3f88d192, dl));
3886 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3887 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3888 getF32Constant(DAG, 0x3fc4316c, dl));
3889 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3890 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8,
3891 getF32Constant(DAG, 0x3f57ce70, dl));
3894 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, Log10ofMantissa);
3897 // No special expansion.
3898 return DAG.getNode(ISD::FLOG10, dl, Op.getValueType(), Op);
3901 /// expandExp2 - Lower an exp2 intrinsic. Handles the special sequences for
3902 /// limited-precision mode.
3903 static SDValue expandExp2(SDLoc dl, SDValue Op, SelectionDAG &DAG,
3904 const TargetLowering &TLI) {
3905 if (Op.getValueType() == MVT::f32 &&
3906 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18)
3907 return getLimitedPrecisionExp2(Op, dl, DAG);
3909 // No special expansion.
3910 return DAG.getNode(ISD::FEXP2, dl, Op.getValueType(), Op);
3913 /// visitPow - Lower a pow intrinsic. Handles the special sequences for
3914 /// limited-precision mode with x == 10.0f.
3915 static SDValue expandPow(SDLoc dl, SDValue LHS, SDValue RHS,
3916 SelectionDAG &DAG, const TargetLowering &TLI) {
3917 bool IsExp10 = false;
3918 if (LHS.getValueType() == MVT::f32 && RHS.getValueType() == MVT::f32 &&
3919 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3920 if (ConstantFPSDNode *LHSC = dyn_cast<ConstantFPSDNode>(LHS)) {
3922 IsExp10 = LHSC->isExactlyValue(Ten);
3927 // Put the exponent in the right bit position for later addition to the
3930 // #define LOG2OF10 3.3219281f
3931 // t0 = Op * LOG2OF10;
3932 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, RHS,
3933 getF32Constant(DAG, 0x40549a78, dl));
3934 return getLimitedPrecisionExp2(t0, dl, DAG);
3937 // No special expansion.
3938 return DAG.getNode(ISD::FPOW, dl, LHS.getValueType(), LHS, RHS);
3942 /// ExpandPowI - Expand a llvm.powi intrinsic.
3943 static SDValue ExpandPowI(SDLoc DL, SDValue LHS, SDValue RHS,
3944 SelectionDAG &DAG) {
3945 // If RHS is a constant, we can expand this out to a multiplication tree,
3946 // otherwise we end up lowering to a call to __powidf2 (for example). When
3947 // optimizing for size, we only want to do this if the expansion would produce
3948 // a small number of multiplies, otherwise we do the full expansion.
3949 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3950 // Get the exponent as a positive value.
3951 unsigned Val = RHSC->getSExtValue();
3952 if ((int)Val < 0) Val = -Val;
3954 // powi(x, 0) -> 1.0
3956 return DAG.getConstantFP(1.0, DL, LHS.getValueType());
3958 const Function *F = DAG.getMachineFunction().getFunction();
3959 if (!F->hasFnAttribute(Attribute::OptimizeForSize) ||
3960 // If optimizing for size, don't insert too many multiplies. This
3961 // inserts up to 5 multiplies.
3962 countPopulation(Val) + Log2_32(Val) < 7) {
3963 // We use the simple binary decomposition method to generate the multiply
3964 // sequence. There are more optimal ways to do this (for example,
3965 // powi(x,15) generates one more multiply than it should), but this has
3966 // the benefit of being both really simple and much better than a libcall.
3967 SDValue Res; // Logically starts equal to 1.0
3968 SDValue CurSquare = LHS;
3972 Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare);
3974 Res = CurSquare; // 1.0*CurSquare.
3977 CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(),
3978 CurSquare, CurSquare);
3982 // If the original was negative, invert the result, producing 1/(x*x*x).
3983 if (RHSC->getSExtValue() < 0)
3984 Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(),
3985 DAG.getConstantFP(1.0, DL, LHS.getValueType()), Res);
3990 // Otherwise, expand to a libcall.
3991 return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS);
3994 // getTruncatedArgReg - Find underlying register used for an truncated
3996 static unsigned getTruncatedArgReg(const SDValue &N) {
3997 if (N.getOpcode() != ISD::TRUNCATE)
4000 const SDValue &Ext = N.getOperand(0);
4001 if (Ext.getOpcode() == ISD::AssertZext ||
4002 Ext.getOpcode() == ISD::AssertSext) {
4003 const SDValue &CFR = Ext.getOperand(0);
4004 if (CFR.getOpcode() == ISD::CopyFromReg)
4005 return cast<RegisterSDNode>(CFR.getOperand(1))->getReg();
4006 if (CFR.getOpcode() == ISD::TRUNCATE)
4007 return getTruncatedArgReg(CFR);
4012 /// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a function
4013 /// argument, create the corresponding DBG_VALUE machine instruction for it now.
4014 /// At the end of instruction selection, they will be inserted to the entry BB.
4015 bool SelectionDAGBuilder::EmitFuncArgumentDbgValue(
4016 const Value *V, DILocalVariable *Variable, DIExpression *Expr,
4017 DILocation *DL, int64_t Offset, bool IsIndirect, const SDValue &N) {
4018 const Argument *Arg = dyn_cast<Argument>(V);
4022 MachineFunction &MF = DAG.getMachineFunction();
4023 const TargetInstrInfo *TII = DAG.getSubtarget().getInstrInfo();
4025 // Ignore inlined function arguments here.
4027 // FIXME: Should we be checking DL->inlinedAt() to determine this?
4028 if (!Variable->getScope()->getSubprogram()->describes(MF.getFunction()))
4031 Optional<MachineOperand> Op;
4032 // Some arguments' frame index is recorded during argument lowering.
4033 if (int FI = FuncInfo.getArgumentFrameIndex(Arg))
4034 Op = MachineOperand::CreateFI(FI);
4036 if (!Op && N.getNode()) {
4038 if (N.getOpcode() == ISD::CopyFromReg)
4039 Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg();
4041 Reg = getTruncatedArgReg(N);
4042 if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) {
4043 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4044 unsigned PR = RegInfo.getLiveInPhysReg(Reg);
4049 Op = MachineOperand::CreateReg(Reg, false);
4053 // Check if ValueMap has reg number.
4054 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
4055 if (VMI != FuncInfo.ValueMap.end())
4056 Op = MachineOperand::CreateReg(VMI->second, false);
4059 if (!Op && N.getNode())
4060 // Check if frame index is available.
4061 if (LoadSDNode *LNode = dyn_cast<LoadSDNode>(N.getNode()))
4062 if (FrameIndexSDNode *FINode =
4063 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode()))
4064 Op = MachineOperand::CreateFI(FINode->getIndex());
4069 assert(Variable->isValidLocationForIntrinsic(DL) &&
4070 "Expected inlined-at fields to agree");
4072 FuncInfo.ArgDbgValues.push_back(
4073 BuildMI(MF, DL, TII->get(TargetOpcode::DBG_VALUE), IsIndirect,
4074 Op->getReg(), Offset, Variable, Expr));
4076 FuncInfo.ArgDbgValues.push_back(
4077 BuildMI(MF, DL, TII->get(TargetOpcode::DBG_VALUE))
4080 .addMetadata(Variable)
4081 .addMetadata(Expr));
4086 // VisualStudio defines setjmp as _setjmp
4087 #if defined(_MSC_VER) && defined(setjmp) && \
4088 !defined(setjmp_undefined_for_msvc)
4089 # pragma push_macro("setjmp")
4091 # define setjmp_undefined_for_msvc
4094 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
4095 /// we want to emit this as a call to a named external function, return the name
4096 /// otherwise lower it and return null.
4098 SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) {
4099 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4100 SDLoc sdl = getCurSDLoc();
4101 DebugLoc dl = getCurDebugLoc();
4104 switch (Intrinsic) {
4106 // By default, turn this into a target intrinsic node.
4107 visitTargetIntrinsic(I, Intrinsic);
4109 case Intrinsic::vastart: visitVAStart(I); return nullptr;
4110 case Intrinsic::vaend: visitVAEnd(I); return nullptr;
4111 case Intrinsic::vacopy: visitVACopy(I); return nullptr;
4112 case Intrinsic::returnaddress:
4113 setValue(&I, DAG.getNode(ISD::RETURNADDR, sdl,
4114 TLI.getPointerTy(DAG.getDataLayout()),
4115 getValue(I.getArgOperand(0))));
4117 case Intrinsic::frameaddress:
4118 setValue(&I, DAG.getNode(ISD::FRAMEADDR, sdl,
4119 TLI.getPointerTy(DAG.getDataLayout()),
4120 getValue(I.getArgOperand(0))));
4122 case Intrinsic::read_register: {
4123 Value *Reg = I.getArgOperand(0);
4124 SDValue Chain = getRoot();
4126 DAG.getMDNode(cast<MDNode>(cast<MetadataAsValue>(Reg)->getMetadata()));
4127 EVT VT = TLI.getValueType(DAG.getDataLayout(), I.getType());
4128 Res = DAG.getNode(ISD::READ_REGISTER, sdl,
4129 DAG.getVTList(VT, MVT::Other), Chain, RegName);
4131 DAG.setRoot(Res.getValue(1));
4134 case Intrinsic::write_register: {
4135 Value *Reg = I.getArgOperand(0);
4136 Value *RegValue = I.getArgOperand(1);
4137 SDValue Chain = getRoot();
4139 DAG.getMDNode(cast<MDNode>(cast<MetadataAsValue>(Reg)->getMetadata()));
4140 DAG.setRoot(DAG.getNode(ISD::WRITE_REGISTER, sdl, MVT::Other, Chain,
4141 RegName, getValue(RegValue)));
4144 case Intrinsic::setjmp:
4145 return &"_setjmp"[!TLI.usesUnderscoreSetJmp()];
4146 case Intrinsic::longjmp:
4147 return &"_longjmp"[!TLI.usesUnderscoreLongJmp()];
4148 case Intrinsic::memcpy: {
4149 // FIXME: this definition of "user defined address space" is x86-specific
4150 // Assert for address < 256 since we support only user defined address
4152 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4154 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4156 "Unknown address space");
4157 SDValue Op1 = getValue(I.getArgOperand(0));
4158 SDValue Op2 = getValue(I.getArgOperand(1));
4159 SDValue Op3 = getValue(I.getArgOperand(2));
4160 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4162 Align = 1; // @llvm.memcpy defines 0 and 1 to both mean no alignment.
4163 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4164 bool isTC = I.isTailCall() && isInTailCallPosition(&I, DAG.getTarget());
4165 SDValue MC = DAG.getMemcpy(getRoot(), sdl, Op1, Op2, Op3, Align, isVol,
4167 MachinePointerInfo(I.getArgOperand(0)),
4168 MachinePointerInfo(I.getArgOperand(1)));
4169 updateDAGForMaybeTailCall(MC);
4172 case Intrinsic::memset: {
4173 // FIXME: this definition of "user defined address space" is x86-specific
4174 // Assert for address < 256 since we support only user defined address
4176 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4178 "Unknown address space");
4179 SDValue Op1 = getValue(I.getArgOperand(0));
4180 SDValue Op2 = getValue(I.getArgOperand(1));
4181 SDValue Op3 = getValue(I.getArgOperand(2));
4182 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4184 Align = 1; // @llvm.memset defines 0 and 1 to both mean no alignment.
4185 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4186 bool isTC = I.isTailCall() && isInTailCallPosition(&I, DAG.getTarget());
4187 SDValue MS = DAG.getMemset(getRoot(), sdl, Op1, Op2, Op3, Align, isVol,
4188 isTC, MachinePointerInfo(I.getArgOperand(0)));
4189 updateDAGForMaybeTailCall(MS);
4192 case Intrinsic::memmove: {
4193 // FIXME: this definition of "user defined address space" is x86-specific
4194 // Assert for address < 256 since we support only user defined address
4196 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4198 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4200 "Unknown address space");
4201 SDValue Op1 = getValue(I.getArgOperand(0));
4202 SDValue Op2 = getValue(I.getArgOperand(1));
4203 SDValue Op3 = getValue(I.getArgOperand(2));
4204 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4206 Align = 1; // @llvm.memmove defines 0 and 1 to both mean no alignment.
4207 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4208 bool isTC = I.isTailCall() && isInTailCallPosition(&I, DAG.getTarget());
4209 SDValue MM = DAG.getMemmove(getRoot(), sdl, Op1, Op2, Op3, Align, isVol,
4210 isTC, MachinePointerInfo(I.getArgOperand(0)),
4211 MachinePointerInfo(I.getArgOperand(1)));
4212 updateDAGForMaybeTailCall(MM);
4215 case Intrinsic::dbg_declare: {
4216 const DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
4217 DILocalVariable *Variable = DI.getVariable();
4218 DIExpression *Expression = DI.getExpression();
4219 const Value *Address = DI.getAddress();
4220 assert(Variable && "Missing variable");
4222 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4226 // Check if address has undef value.
4227 if (isa<UndefValue>(Address) ||
4228 (Address->use_empty() && !isa<Argument>(Address))) {
4229 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4233 SDValue &N = NodeMap[Address];
4234 if (!N.getNode() && isa<Argument>(Address))
4235 // Check unused arguments map.
4236 N = UnusedArgNodeMap[Address];
4239 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
4240 Address = BCI->getOperand(0);
4241 // Parameters are handled specially.
4242 bool isParameter = Variable->getTag() == dwarf::DW_TAG_arg_variable ||
4243 isa<Argument>(Address);
4245 const AllocaInst *AI = dyn_cast<AllocaInst>(Address);
4247 if (isParameter && !AI) {
4248 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode());
4250 // Byval parameter. We have a frame index at this point.
4251 SDV = DAG.getFrameIndexDbgValue(
4252 Variable, Expression, FINode->getIndex(), 0, dl, SDNodeOrder);
4254 // Address is an argument, so try to emit its dbg value using
4255 // virtual register info from the FuncInfo.ValueMap.
4256 EmitFuncArgumentDbgValue(Address, Variable, Expression, dl, 0, false,
4261 SDV = DAG.getDbgValue(Variable, Expression, N.getNode(), N.getResNo(),
4262 true, 0, dl, SDNodeOrder);
4264 // Can't do anything with other non-AI cases yet.
4265 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4266 DEBUG(dbgs() << "non-AllocaInst issue for Address: \n\t");
4267 DEBUG(Address->dump());
4270 DAG.AddDbgValue(SDV, N.getNode(), isParameter);
4272 // If Address is an argument then try to emit its dbg value using
4273 // virtual register info from the FuncInfo.ValueMap.
4274 if (!EmitFuncArgumentDbgValue(Address, Variable, Expression, dl, 0, false,
4276 // If variable is pinned by a alloca in dominating bb then
4277 // use StaticAllocaMap.
4278 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Address)) {
4279 if (AI->getParent() != DI.getParent()) {
4280 DenseMap<const AllocaInst*, int>::iterator SI =
4281 FuncInfo.StaticAllocaMap.find(AI);
4282 if (SI != FuncInfo.StaticAllocaMap.end()) {
4283 SDV = DAG.getFrameIndexDbgValue(Variable, Expression, SI->second,
4284 0, dl, SDNodeOrder);
4285 DAG.AddDbgValue(SDV, nullptr, false);
4290 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4295 case Intrinsic::dbg_value: {
4296 const DbgValueInst &DI = cast<DbgValueInst>(I);
4297 assert(DI.getVariable() && "Missing variable");
4299 DILocalVariable *Variable = DI.getVariable();
4300 DIExpression *Expression = DI.getExpression();
4301 uint64_t Offset = DI.getOffset();
4302 const Value *V = DI.getValue();
4307 if (isa<ConstantInt>(V) || isa<ConstantFP>(V) || isa<UndefValue>(V)) {
4308 SDV = DAG.getConstantDbgValue(Variable, Expression, V, Offset, dl,
4310 DAG.AddDbgValue(SDV, nullptr, false);
4312 // Do not use getValue() in here; we don't want to generate code at
4313 // this point if it hasn't been done yet.
4314 SDValue N = NodeMap[V];
4315 if (!N.getNode() && isa<Argument>(V))
4316 // Check unused arguments map.
4317 N = UnusedArgNodeMap[V];
4319 // A dbg.value for an alloca is always indirect.
4320 bool IsIndirect = isa<AllocaInst>(V) || Offset != 0;
4321 if (!EmitFuncArgumentDbgValue(V, Variable, Expression, dl, Offset,
4323 SDV = DAG.getDbgValue(Variable, Expression, N.getNode(), N.getResNo(),
4324 IsIndirect, Offset, dl, SDNodeOrder);
4325 DAG.AddDbgValue(SDV, N.getNode(), false);
4327 } else if (!V->use_empty() ) {
4328 // Do not call getValue(V) yet, as we don't want to generate code.
4329 // Remember it for later.
4330 DanglingDebugInfo DDI(&DI, dl, SDNodeOrder);
4331 DanglingDebugInfoMap[V] = DDI;
4333 // We may expand this to cover more cases. One case where we have no
4334 // data available is an unreferenced parameter.
4335 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4339 // Build a debug info table entry.
4340 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
4341 V = BCI->getOperand(0);
4342 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
4343 // Don't handle byval struct arguments or VLAs, for example.
4345 DEBUG(dbgs() << "Dropping debug location info for:\n " << DI << "\n");
4346 DEBUG(dbgs() << " Last seen at:\n " << *V << "\n");
4349 DenseMap<const AllocaInst*, int>::iterator SI =
4350 FuncInfo.StaticAllocaMap.find(AI);
4351 if (SI == FuncInfo.StaticAllocaMap.end())
4352 return nullptr; // VLAs.
4356 case Intrinsic::eh_typeid_for: {
4357 // Find the type id for the given typeinfo.
4358 GlobalValue *GV = ExtractTypeInfo(I.getArgOperand(0));
4359 unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(GV);
4360 Res = DAG.getConstant(TypeID, sdl, MVT::i32);
4365 case Intrinsic::eh_return_i32:
4366 case Intrinsic::eh_return_i64:
4367 DAG.getMachineFunction().getMMI().setCallsEHReturn(true);
4368 DAG.setRoot(DAG.getNode(ISD::EH_RETURN, sdl,
4371 getValue(I.getArgOperand(0)),
4372 getValue(I.getArgOperand(1))));
4374 case Intrinsic::eh_unwind_init:
4375 DAG.getMachineFunction().getMMI().setCallsUnwindInit(true);
4377 case Intrinsic::eh_dwarf_cfa: {
4378 SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getArgOperand(0)), sdl,
4379 TLI.getPointerTy(DAG.getDataLayout()));
4380 SDValue Offset = DAG.getNode(ISD::ADD, sdl,
4381 CfaArg.getValueType(),
4382 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, sdl,
4383 CfaArg.getValueType()),
4385 SDValue FA = DAG.getNode(
4386 ISD::FRAMEADDR, sdl, TLI.getPointerTy(DAG.getDataLayout()),
4387 DAG.getConstant(0, sdl, TLI.getPointerTy(DAG.getDataLayout())));
4388 setValue(&I, DAG.getNode(ISD::ADD, sdl, FA.getValueType(),
4392 case Intrinsic::eh_sjlj_callsite: {
4393 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4394 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(0));
4395 assert(CI && "Non-constant call site value in eh.sjlj.callsite!");
4396 assert(MMI.getCurrentCallSite() == 0 && "Overlapping call sites!");
4398 MMI.setCurrentCallSite(CI->getZExtValue());
4401 case Intrinsic::eh_sjlj_functioncontext: {
4402 // Get and store the index of the function context.
4403 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4405 cast<AllocaInst>(I.getArgOperand(0)->stripPointerCasts());
4406 int FI = FuncInfo.StaticAllocaMap[FnCtx];
4407 MFI->setFunctionContextIndex(FI);
4410 case Intrinsic::eh_sjlj_setjmp: {
4413 Ops[1] = getValue(I.getArgOperand(0));
4414 SDValue Op = DAG.getNode(ISD::EH_SJLJ_SETJMP, sdl,
4415 DAG.getVTList(MVT::i32, MVT::Other), Ops);
4416 setValue(&I, Op.getValue(0));
4417 DAG.setRoot(Op.getValue(1));
4420 case Intrinsic::eh_sjlj_longjmp: {
4421 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, sdl, MVT::Other,
4422 getRoot(), getValue(I.getArgOperand(0))));
4426 case Intrinsic::masked_gather:
4427 visitMaskedGather(I);
4429 case Intrinsic::masked_load:
4432 case Intrinsic::masked_scatter:
4433 visitMaskedScatter(I);
4435 case Intrinsic::masked_store:
4436 visitMaskedStore(I);
4438 case Intrinsic::x86_mmx_pslli_w:
4439 case Intrinsic::x86_mmx_pslli_d:
4440 case Intrinsic::x86_mmx_pslli_q:
4441 case Intrinsic::x86_mmx_psrli_w:
4442 case Intrinsic::x86_mmx_psrli_d:
4443 case Intrinsic::x86_mmx_psrli_q:
4444 case Intrinsic::x86_mmx_psrai_w:
4445 case Intrinsic::x86_mmx_psrai_d: {
4446 SDValue ShAmt = getValue(I.getArgOperand(1));
4447 if (isa<ConstantSDNode>(ShAmt)) {
4448 visitTargetIntrinsic(I, Intrinsic);
4451 unsigned NewIntrinsic = 0;
4452 EVT ShAmtVT = MVT::v2i32;
4453 switch (Intrinsic) {
4454 case Intrinsic::x86_mmx_pslli_w:
4455 NewIntrinsic = Intrinsic::x86_mmx_psll_w;
4457 case Intrinsic::x86_mmx_pslli_d:
4458 NewIntrinsic = Intrinsic::x86_mmx_psll_d;
4460 case Intrinsic::x86_mmx_pslli_q:
4461 NewIntrinsic = Intrinsic::x86_mmx_psll_q;
4463 case Intrinsic::x86_mmx_psrli_w:
4464 NewIntrinsic = Intrinsic::x86_mmx_psrl_w;
4466 case Intrinsic::x86_mmx_psrli_d:
4467 NewIntrinsic = Intrinsic::x86_mmx_psrl_d;
4469 case Intrinsic::x86_mmx_psrli_q:
4470 NewIntrinsic = Intrinsic::x86_mmx_psrl_q;
4472 case Intrinsic::x86_mmx_psrai_w:
4473 NewIntrinsic = Intrinsic::x86_mmx_psra_w;
4475 case Intrinsic::x86_mmx_psrai_d:
4476 NewIntrinsic = Intrinsic::x86_mmx_psra_d;
4478 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4481 // The vector shift intrinsics with scalars uses 32b shift amounts but
4482 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
4484 // We must do this early because v2i32 is not a legal type.
4487 ShOps[1] = DAG.getConstant(0, sdl, MVT::i32);
4488 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, sdl, ShAmtVT, ShOps);
4489 EVT DestVT = TLI.getValueType(DAG.getDataLayout(), I.getType());
4490 ShAmt = DAG.getNode(ISD::BITCAST, sdl, DestVT, ShAmt);
4491 Res = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, sdl, DestVT,
4492 DAG.getConstant(NewIntrinsic, sdl, MVT::i32),
4493 getValue(I.getArgOperand(0)), ShAmt);
4497 case Intrinsic::convertff:
4498 case Intrinsic::convertfsi:
4499 case Intrinsic::convertfui:
4500 case Intrinsic::convertsif:
4501 case Intrinsic::convertuif:
4502 case Intrinsic::convertss:
4503 case Intrinsic::convertsu:
4504 case Intrinsic::convertus:
4505 case Intrinsic::convertuu: {
4506 ISD::CvtCode Code = ISD::CVT_INVALID;
4507 switch (Intrinsic) {
4508 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4509 case Intrinsic::convertff: Code = ISD::CVT_FF; break;
4510 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break;
4511 case Intrinsic::convertfui: Code = ISD::CVT_FU; break;
4512 case Intrinsic::convertsif: Code = ISD::CVT_SF; break;
4513 case Intrinsic::convertuif: Code = ISD::CVT_UF; break;
4514 case Intrinsic::convertss: Code = ISD::CVT_SS; break;
4515 case Intrinsic::convertsu: Code = ISD::CVT_SU; break;
4516 case Intrinsic::convertus: Code = ISD::CVT_US; break;
4517 case Intrinsic::convertuu: Code = ISD::CVT_UU; break;
4519 EVT DestVT = TLI.getValueType(DAG.getDataLayout(), I.getType());
4520 const Value *Op1 = I.getArgOperand(0);
4521 Res = DAG.getConvertRndSat(DestVT, sdl, getValue(Op1),
4522 DAG.getValueType(DestVT),
4523 DAG.getValueType(getValue(Op1).getValueType()),
4524 getValue(I.getArgOperand(1)),
4525 getValue(I.getArgOperand(2)),
4530 case Intrinsic::powi:
4531 setValue(&I, ExpandPowI(sdl, getValue(I.getArgOperand(0)),
4532 getValue(I.getArgOperand(1)), DAG));
4534 case Intrinsic::log:
4535 setValue(&I, expandLog(sdl, getValue(I.getArgOperand(0)), DAG, TLI));
4537 case Intrinsic::log2:
4538 setValue(&I, expandLog2(sdl, getValue(I.getArgOperand(0)), DAG, TLI));
4540 case Intrinsic::log10:
4541 setValue(&I, expandLog10(sdl, getValue(I.getArgOperand(0)), DAG, TLI));
4543 case Intrinsic::exp:
4544 setValue(&I, expandExp(sdl, getValue(I.getArgOperand(0)), DAG, TLI));
4546 case Intrinsic::exp2:
4547 setValue(&I, expandExp2(sdl, getValue(I.getArgOperand(0)), DAG, TLI));
4549 case Intrinsic::pow:
4550 setValue(&I, expandPow(sdl, getValue(I.getArgOperand(0)),
4551 getValue(I.getArgOperand(1)), DAG, TLI));
4553 case Intrinsic::sqrt:
4554 case Intrinsic::fabs:
4555 case Intrinsic::sin:
4556 case Intrinsic::cos:
4557 case Intrinsic::floor:
4558 case Intrinsic::ceil:
4559 case Intrinsic::trunc:
4560 case Intrinsic::rint:
4561 case Intrinsic::nearbyint:
4562 case Intrinsic::round: {
4564 switch (Intrinsic) {
4565 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4566 case Intrinsic::sqrt: Opcode = ISD::FSQRT; break;
4567 case Intrinsic::fabs: Opcode = ISD::FABS; break;
4568 case Intrinsic::sin: Opcode = ISD::FSIN; break;
4569 case Intrinsic::cos: Opcode = ISD::FCOS; break;
4570 case Intrinsic::floor: Opcode = ISD::FFLOOR; break;
4571 case Intrinsic::ceil: Opcode = ISD::FCEIL; break;
4572 case Intrinsic::trunc: Opcode = ISD::FTRUNC; break;
4573 case Intrinsic::rint: Opcode = ISD::FRINT; break;
4574 case Intrinsic::nearbyint: Opcode = ISD::FNEARBYINT; break;
4575 case Intrinsic::round: Opcode = ISD::FROUND; break;
4578 setValue(&I, DAG.getNode(Opcode, sdl,
4579 getValue(I.getArgOperand(0)).getValueType(),
4580 getValue(I.getArgOperand(0))));
4583 case Intrinsic::minnum:
4584 setValue(&I, DAG.getNode(ISD::FMINNUM, sdl,
4585 getValue(I.getArgOperand(0)).getValueType(),
4586 getValue(I.getArgOperand(0)),
4587 getValue(I.getArgOperand(1))));
4589 case Intrinsic::maxnum:
4590 setValue(&I, DAG.getNode(ISD::FMAXNUM, sdl,
4591 getValue(I.getArgOperand(0)).getValueType(),
4592 getValue(I.getArgOperand(0)),
4593 getValue(I.getArgOperand(1))));
4595 case Intrinsic::copysign:
4596 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, sdl,
4597 getValue(I.getArgOperand(0)).getValueType(),
4598 getValue(I.getArgOperand(0)),
4599 getValue(I.getArgOperand(1))));
4601 case Intrinsic::fma:
4602 setValue(&I, DAG.getNode(ISD::FMA, sdl,
4603 getValue(I.getArgOperand(0)).getValueType(),
4604 getValue(I.getArgOperand(0)),
4605 getValue(I.getArgOperand(1)),
4606 getValue(I.getArgOperand(2))));
4608 case Intrinsic::fmuladd: {
4609 EVT VT = TLI.getValueType(DAG.getDataLayout(), I.getType());
4610 if (TM.Options.AllowFPOpFusion != FPOpFusion::Strict &&
4611 TLI.isFMAFasterThanFMulAndFAdd(VT)) {
4612 setValue(&I, DAG.getNode(ISD::FMA, sdl,
4613 getValue(I.getArgOperand(0)).getValueType(),
4614 getValue(I.getArgOperand(0)),
4615 getValue(I.getArgOperand(1)),
4616 getValue(I.getArgOperand(2))));
4618 SDValue Mul = DAG.getNode(ISD::FMUL, sdl,
4619 getValue(I.getArgOperand(0)).getValueType(),
4620 getValue(I.getArgOperand(0)),
4621 getValue(I.getArgOperand(1)));
4622 SDValue Add = DAG.getNode(ISD::FADD, sdl,
4623 getValue(I.getArgOperand(0)).getValueType(),
4625 getValue(I.getArgOperand(2)));
4630 case Intrinsic::convert_to_fp16:
4631 setValue(&I, DAG.getNode(ISD::BITCAST, sdl, MVT::i16,
4632 DAG.getNode(ISD::FP_ROUND, sdl, MVT::f16,
4633 getValue(I.getArgOperand(0)),
4634 DAG.getTargetConstant(0, sdl,
4637 case Intrinsic::convert_from_fp16:
4638 setValue(&I, DAG.getNode(ISD::FP_EXTEND, sdl,
4639 TLI.getValueType(DAG.getDataLayout(), I.getType()),
4640 DAG.getNode(ISD::BITCAST, sdl, MVT::f16,
4641 getValue(I.getArgOperand(0)))));
4643 case Intrinsic::pcmarker: {
4644 SDValue Tmp = getValue(I.getArgOperand(0));
4645 DAG.setRoot(DAG.getNode(ISD::PCMARKER, sdl, MVT::Other, getRoot(), Tmp));
4648 case Intrinsic::readcyclecounter: {
4649 SDValue Op = getRoot();
4650 Res = DAG.getNode(ISD::READCYCLECOUNTER, sdl,
4651 DAG.getVTList(MVT::i64, MVT::Other), Op);
4653 DAG.setRoot(Res.getValue(1));
4656 case Intrinsic::bswap:
4657 setValue(&I, DAG.getNode(ISD::BSWAP, sdl,
4658 getValue(I.getArgOperand(0)).getValueType(),
4659 getValue(I.getArgOperand(0))));
4661 case Intrinsic::cttz: {
4662 SDValue Arg = getValue(I.getArgOperand(0));
4663 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1));
4664 EVT Ty = Arg.getValueType();
4665 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTTZ : ISD::CTTZ_ZERO_UNDEF,
4669 case Intrinsic::ctlz: {
4670 SDValue Arg = getValue(I.getArgOperand(0));
4671 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1));
4672 EVT Ty = Arg.getValueType();
4673 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTLZ : ISD::CTLZ_ZERO_UNDEF,
4677 case Intrinsic::ctpop: {
4678 SDValue Arg = getValue(I.getArgOperand(0));
4679 EVT Ty = Arg.getValueType();
4680 setValue(&I, DAG.getNode(ISD::CTPOP, sdl, Ty, Arg));
4683 case Intrinsic::stacksave: {
4684 SDValue Op = getRoot();
4686 ISD::STACKSAVE, sdl,
4687 DAG.getVTList(TLI.getPointerTy(DAG.getDataLayout()), MVT::Other), Op);
4689 DAG.setRoot(Res.getValue(1));
4692 case Intrinsic::stackrestore: {
4693 Res = getValue(I.getArgOperand(0));
4694 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, sdl, MVT::Other, getRoot(), Res));
4697 case Intrinsic::stackprotector: {
4698 // Emit code into the DAG to store the stack guard onto the stack.
4699 MachineFunction &MF = DAG.getMachineFunction();
4700 MachineFrameInfo *MFI = MF.getFrameInfo();
4701 EVT PtrTy = TLI.getPointerTy(DAG.getDataLayout());
4702 SDValue Src, Chain = getRoot();
4703 const Value *Ptr = cast<LoadInst>(I.getArgOperand(0))->getPointerOperand();
4704 const GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr);
4706 // See if Ptr is a bitcast. If it is, look through it and see if we can get
4707 // global variable __stack_chk_guard.
4709 if (const Operator *BC = dyn_cast<Operator>(Ptr))
4710 if (BC->getOpcode() == Instruction::BitCast)
4711 GV = dyn_cast<GlobalVariable>(BC->getOperand(0));
4713 if (GV && TLI.useLoadStackGuardNode()) {
4714 // Emit a LOAD_STACK_GUARD node.
4715 MachineSDNode *Node = DAG.getMachineNode(TargetOpcode::LOAD_STACK_GUARD,
4717 MachinePointerInfo MPInfo(GV);
4718 MachineInstr::mmo_iterator MemRefs = MF.allocateMemRefsArray(1);
4719 unsigned Flags = MachineMemOperand::MOLoad |
4720 MachineMemOperand::MOInvariant;
4721 *MemRefs = MF.getMachineMemOperand(MPInfo, Flags,
4722 PtrTy.getSizeInBits() / 8,
4723 DAG.getEVTAlignment(PtrTy));
4724 Node->setMemRefs(MemRefs, MemRefs + 1);
4726 // Copy the guard value to a virtual register so that it can be
4727 // retrieved in the epilogue.
4728 Src = SDValue(Node, 0);
4729 const TargetRegisterClass *RC =
4730 TLI.getRegClassFor(Src.getSimpleValueType());
4731 unsigned Reg = MF.getRegInfo().createVirtualRegister(RC);
4733 SPDescriptor.setGuardReg(Reg);
4734 Chain = DAG.getCopyToReg(Chain, sdl, Reg, Src);
4736 Src = getValue(I.getArgOperand(0)); // The guard's value.
4739 AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
4741 int FI = FuncInfo.StaticAllocaMap[Slot];
4742 MFI->setStackProtectorIndex(FI);
4744 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
4746 // Store the stack protector onto the stack.
4747 Res = DAG.getStore(Chain, sdl, Src, FIN,
4748 MachinePointerInfo::getFixedStack(FI),
4754 case Intrinsic::objectsize: {
4755 // If we don't know by now, we're never going to know.
4756 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1));
4758 assert(CI && "Non-constant type in __builtin_object_size?");
4760 SDValue Arg = getValue(I.getCalledValue());
4761 EVT Ty = Arg.getValueType();
4764 Res = DAG.getConstant(-1ULL, sdl, Ty);
4766 Res = DAG.getConstant(0, sdl, Ty);
4771 case Intrinsic::annotation:
4772 case Intrinsic::ptr_annotation:
4773 // Drop the intrinsic, but forward the value
4774 setValue(&I, getValue(I.getOperand(0)));
4776 case Intrinsic::assume:
4777 case Intrinsic::var_annotation:
4778 // Discard annotate attributes and assumptions
4781 case Intrinsic::init_trampoline: {
4782 const Function *F = cast<Function>(I.getArgOperand(1)->stripPointerCasts());
4786 Ops[1] = getValue(I.getArgOperand(0));
4787 Ops[2] = getValue(I.getArgOperand(1));
4788 Ops[3] = getValue(I.getArgOperand(2));
4789 Ops[4] = DAG.getSrcValue(I.getArgOperand(0));
4790 Ops[5] = DAG.getSrcValue(F);
4792 Res = DAG.getNode(ISD::INIT_TRAMPOLINE, sdl, MVT::Other, Ops);
4797 case Intrinsic::adjust_trampoline: {
4798 setValue(&I, DAG.getNode(ISD::ADJUST_TRAMPOLINE, sdl,
4799 TLI.getPointerTy(DAG.getDataLayout()),
4800 getValue(I.getArgOperand(0))));
4803 case Intrinsic::gcroot:
4805 const Value *Alloca = I.getArgOperand(0)->stripPointerCasts();
4806 const Constant *TypeMap = cast<Constant>(I.getArgOperand(1));
4808 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode());
4809 GFI->addStackRoot(FI->getIndex(), TypeMap);
4812 case Intrinsic::gcread:
4813 case Intrinsic::gcwrite:
4814 llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!");
4815 case Intrinsic::flt_rounds:
4816 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, sdl, MVT::i32));
4819 case Intrinsic::expect: {
4820 // Just replace __builtin_expect(exp, c) with EXP.
4821 setValue(&I, getValue(I.getArgOperand(0)));
4825 case Intrinsic::debugtrap:
4826 case Intrinsic::trap: {
4827 StringRef TrapFuncName =
4829 .getAttribute(AttributeSet::FunctionIndex, "trap-func-name")
4830 .getValueAsString();
4831 if (TrapFuncName.empty()) {
4832 ISD::NodeType Op = (Intrinsic == Intrinsic::trap) ?
4833 ISD::TRAP : ISD::DEBUGTRAP;
4834 DAG.setRoot(DAG.getNode(Op, sdl,MVT::Other, getRoot()));
4837 TargetLowering::ArgListTy Args;
4839 TargetLowering::CallLoweringInfo CLI(DAG);
4840 CLI.setDebugLoc(sdl).setChain(getRoot()).setCallee(
4841 CallingConv::C, I.getType(),
4842 DAG.getExternalSymbol(TrapFuncName.data(),
4843 TLI.getPointerTy(DAG.getDataLayout())),
4844 std::move(Args), 0);
4846 std::pair<SDValue, SDValue> Result = TLI.LowerCallTo(CLI);
4847 DAG.setRoot(Result.second);
4851 case Intrinsic::uadd_with_overflow:
4852 case Intrinsic::sadd_with_overflow:
4853 case Intrinsic::usub_with_overflow:
4854 case Intrinsic::ssub_with_overflow:
4855 case Intrinsic::umul_with_overflow:
4856 case Intrinsic::smul_with_overflow: {
4858 switch (Intrinsic) {
4859 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4860 case Intrinsic::uadd_with_overflow: Op = ISD::UADDO; break;
4861 case Intrinsic::sadd_with_overflow: Op = ISD::SADDO; break;
4862 case Intrinsic::usub_with_overflow: Op = ISD::USUBO; break;
4863 case Intrinsic::ssub_with_overflow: Op = ISD::SSUBO; break;
4864 case Intrinsic::umul_with_overflow: Op = ISD::UMULO; break;
4865 case Intrinsic::smul_with_overflow: Op = ISD::SMULO; break;
4867 SDValue Op1 = getValue(I.getArgOperand(0));
4868 SDValue Op2 = getValue(I.getArgOperand(1));
4870 SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1);
4871 setValue(&I, DAG.getNode(Op, sdl, VTs, Op1, Op2));
4874 case Intrinsic::prefetch: {
4876 unsigned rw = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
4878 Ops[1] = getValue(I.getArgOperand(0));
4879 Ops[2] = getValue(I.getArgOperand(1));
4880 Ops[3] = getValue(I.getArgOperand(2));
4881 Ops[4] = getValue(I.getArgOperand(3));
4882 DAG.setRoot(DAG.getMemIntrinsicNode(ISD::PREFETCH, sdl,
4883 DAG.getVTList(MVT::Other), Ops,
4884 EVT::getIntegerVT(*Context, 8),
4885 MachinePointerInfo(I.getArgOperand(0)),
4887 false, /* volatile */
4889 rw==1)); /* write */
4892 case Intrinsic::lifetime_start:
4893 case Intrinsic::lifetime_end: {
4894 bool IsStart = (Intrinsic == Intrinsic::lifetime_start);
4895 // Stack coloring is not enabled in O0, discard region information.
4896 if (TM.getOptLevel() == CodeGenOpt::None)
4899 SmallVector<Value *, 4> Allocas;
4900 GetUnderlyingObjects(I.getArgOperand(1), Allocas, *DL);
4902 for (SmallVectorImpl<Value*>::iterator Object = Allocas.begin(),
4903 E = Allocas.end(); Object != E; ++Object) {
4904 AllocaInst *LifetimeObject = dyn_cast_or_null<AllocaInst>(*Object);
4906 // Could not find an Alloca.
4907 if (!LifetimeObject)
4910 // First check that the Alloca is static, otherwise it won't have a
4911 // valid frame index.
4912 auto SI = FuncInfo.StaticAllocaMap.find(LifetimeObject);
4913 if (SI == FuncInfo.StaticAllocaMap.end())
4916 int FI = SI->second;
4921 DAG.getFrameIndex(FI, TLI.getPointerTy(DAG.getDataLayout()), true);
4922 unsigned Opcode = (IsStart ? ISD::LIFETIME_START : ISD::LIFETIME_END);
4924 Res = DAG.getNode(Opcode, sdl, MVT::Other, Ops);
4929 case Intrinsic::invariant_start:
4930 // Discard region information.
4931 setValue(&I, DAG.getUNDEF(TLI.getPointerTy(DAG.getDataLayout())));
4933 case Intrinsic::invariant_end:
4934 // Discard region information.
4936 case Intrinsic::stackprotectorcheck: {
4937 // Do not actually emit anything for this basic block. Instead we initialize
4938 // the stack protector descriptor and export the guard variable so we can
4939 // access it in FinishBasicBlock.
4940 const BasicBlock *BB = I.getParent();
4941 SPDescriptor.initialize(BB, FuncInfo.MBBMap[BB], I);
4942 ExportFromCurrentBlock(SPDescriptor.getGuard());
4944 // Flush our exports since we are going to process a terminator.
4945 (void)getControlRoot();
4948 case Intrinsic::clear_cache:
4949 return TLI.getClearCacheBuiltinName();
4950 case Intrinsic::eh_actions:
4951 setValue(&I, DAG.getUNDEF(TLI.getPointerTy(DAG.getDataLayout())));
4953 case Intrinsic::donothing:
4956 case Intrinsic::experimental_stackmap: {
4960 case Intrinsic::experimental_patchpoint_void:
4961 case Intrinsic::experimental_patchpoint_i64: {
4962 visitPatchpoint(&I);
4965 case Intrinsic::experimental_gc_statepoint: {
4969 case Intrinsic::experimental_gc_result_int:
4970 case Intrinsic::experimental_gc_result_float:
4971 case Intrinsic::experimental_gc_result_ptr:
4972 case Intrinsic::experimental_gc_result: {
4976 case Intrinsic::experimental_gc_relocate: {
4980 case Intrinsic::instrprof_increment:
4981 llvm_unreachable("instrprof failed to lower an increment");
4983 case Intrinsic::localescape: {
4984 MachineFunction &MF = DAG.getMachineFunction();
4985 const TargetInstrInfo *TII = DAG.getSubtarget().getInstrInfo();
4987 // Directly emit some LOCAL_ESCAPE machine instrs. Label assignment emission
4988 // is the same on all targets.
4989 for (unsigned Idx = 0, E = I.getNumArgOperands(); Idx < E; ++Idx) {
4990 Value *Arg = I.getArgOperand(Idx)->stripPointerCasts();
4991 if (isa<ConstantPointerNull>(Arg))
4992 continue; // Skip null pointers. They represent a hole in index space.
4993 AllocaInst *Slot = cast<AllocaInst>(Arg);
4994 assert(FuncInfo.StaticAllocaMap.count(Slot) &&
4995 "can only escape static allocas");
4996 int FI = FuncInfo.StaticAllocaMap[Slot];
4997 MCSymbol *FrameAllocSym =
4998 MF.getMMI().getContext().getOrCreateFrameAllocSymbol(
4999 GlobalValue::getRealLinkageName(MF.getName()), Idx);
5000 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, dl,
5001 TII->get(TargetOpcode::LOCAL_ESCAPE))
5002 .addSym(FrameAllocSym)
5009 case Intrinsic::localrecover: {
5010 // i8* @llvm.localrecover(i8* %fn, i8* %fp, i32 %idx)
5011 MachineFunction &MF = DAG.getMachineFunction();
5012 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout(), 0);
5014 // Get the symbol that defines the frame offset.
5015 auto *Fn = cast<Function>(I.getArgOperand(0)->stripPointerCasts());
5016 auto *Idx = cast<ConstantInt>(I.getArgOperand(2));
5017 unsigned IdxVal = unsigned(Idx->getLimitedValue(INT_MAX));
5018 MCSymbol *FrameAllocSym =
5019 MF.getMMI().getContext().getOrCreateFrameAllocSymbol(
5020 GlobalValue::getRealLinkageName(Fn->getName()), IdxVal);
5022 // Create a MCSymbol for the label to avoid any target lowering
5023 // that would make this PC relative.
5024 SDValue OffsetSym = DAG.getMCSymbol(FrameAllocSym, PtrVT);
5026 DAG.getNode(ISD::LOCAL_RECOVER, sdl, PtrVT, OffsetSym);
5028 // Add the offset to the FP.
5029 Value *FP = I.getArgOperand(1);
5030 SDValue FPVal = getValue(FP);
5031 SDValue Add = DAG.getNode(ISD::ADD, sdl, PtrVT, FPVal, OffsetVal);
5036 case Intrinsic::eh_begincatch:
5037 case Intrinsic::eh_endcatch:
5038 llvm_unreachable("begin/end catch intrinsics not lowered in codegen");
5039 case Intrinsic::eh_exceptioncode: {
5040 unsigned Reg = TLI.getExceptionPointerRegister();
5041 assert(Reg && "cannot get exception code on this platform");
5042 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
5043 const TargetRegisterClass *PtrRC = TLI.getRegClassFor(PtrVT);
5044 assert(FuncInfo.MBB->isLandingPad() && "eh.exceptioncode in non-lpad");
5045 unsigned VReg = FuncInfo.MBB->addLiveIn(Reg, PtrRC);
5047 DAG.getCopyFromReg(DAG.getEntryNode(), getCurSDLoc(), VReg, PtrVT);
5048 N = DAG.getZExtOrTrunc(N, getCurSDLoc(), MVT::i32);
5055 std::pair<SDValue, SDValue>
5056 SelectionDAGBuilder::lowerInvokable(TargetLowering::CallLoweringInfo &CLI,
5057 MachineBasicBlock *LandingPad) {
5058 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5059 MCSymbol *BeginLabel = nullptr;
5062 // Insert a label before the invoke call to mark the try range. This can be
5063 // used to detect deletion of the invoke via the MachineModuleInfo.
5064 BeginLabel = MMI.getContext().createTempSymbol();
5066 // For SjLj, keep track of which landing pads go with which invokes
5067 // so as to maintain the ordering of pads in the LSDA.
5068 unsigned CallSiteIndex = MMI.getCurrentCallSite();
5069 if (CallSiteIndex) {
5070 MMI.setCallSiteBeginLabel(BeginLabel, CallSiteIndex);
5071 LPadToCallSiteMap[LandingPad].push_back(CallSiteIndex);
5073 // Now that the call site is handled, stop tracking it.
5074 MMI.setCurrentCallSite(0);
5077 // Both PendingLoads and PendingExports must be flushed here;
5078 // this call might not return.
5080 DAG.setRoot(DAG.getEHLabel(getCurSDLoc(), getControlRoot(), BeginLabel));
5082 CLI.setChain(getRoot());
5084 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5085 std::pair<SDValue, SDValue> Result = TLI.LowerCallTo(CLI);
5087 assert((CLI.IsTailCall || Result.second.getNode()) &&
5088 "Non-null chain expected with non-tail call!");
5089 assert((Result.second.getNode() || !Result.first.getNode()) &&
5090 "Null value expected with tail call!");
5092 if (!Result.second.getNode()) {
5093 // As a special case, a null chain means that a tail call has been emitted
5094 // and the DAG root is already updated.
5097 // Since there's no actual continuation from this block, nothing can be
5098 // relying on us setting vregs for them.
5099 PendingExports.clear();
5101 DAG.setRoot(Result.second);
5105 // Insert a label at the end of the invoke call to mark the try range. This
5106 // can be used to detect deletion of the invoke via the MachineModuleInfo.
5107 MCSymbol *EndLabel = MMI.getContext().createTempSymbol();
5108 DAG.setRoot(DAG.getEHLabel(getCurSDLoc(), getRoot(), EndLabel));
5110 // Inform MachineModuleInfo of range.
5111 MMI.addInvoke(LandingPad, BeginLabel, EndLabel);
5117 void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee,
5119 MachineBasicBlock *LandingPad) {
5120 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
5121 FunctionType *FTy = cast<FunctionType>(PT->getElementType());
5122 Type *RetTy = FTy->getReturnType();
5124 TargetLowering::ArgListTy Args;
5125 TargetLowering::ArgListEntry Entry;
5126 Args.reserve(CS.arg_size());
5128 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
5130 const Value *V = *i;
5133 if (V->getType()->isEmptyTy())
5136 SDValue ArgNode = getValue(V);
5137 Entry.Node = ArgNode; Entry.Ty = V->getType();
5139 // Skip the first return-type Attribute to get to params.
5140 Entry.setAttributes(&CS, i - CS.arg_begin() + 1);
5141 Args.push_back(Entry);
5143 // If we have an explicit sret argument that is an Instruction, (i.e., it
5144 // might point to function-local memory), we can't meaningfully tail-call.
5145 if (Entry.isSRet && isa<Instruction>(V))
5149 // Check if target-independent constraints permit a tail call here.
5150 // Target-dependent constraints are checked within TLI->LowerCallTo.
5151 if (isTailCall && !isInTailCallPosition(CS, DAG.getTarget()))
5154 TargetLowering::CallLoweringInfo CLI(DAG);
5155 CLI.setDebugLoc(getCurSDLoc()).setChain(getRoot())
5156 .setCallee(RetTy, FTy, Callee, std::move(Args), CS)
5157 .setTailCall(isTailCall);
5158 std::pair<SDValue,SDValue> Result = lowerInvokable(CLI, LandingPad);
5160 if (Result.first.getNode())
5161 setValue(CS.getInstruction(), Result.first);
5164 /// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the
5165 /// value is equal or not-equal to zero.
5166 static bool IsOnlyUsedInZeroEqualityComparison(const Value *V) {
5167 for (const User *U : V->users()) {
5168 if (const ICmpInst *IC = dyn_cast<ICmpInst>(U))
5169 if (IC->isEquality())
5170 if (const Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
5171 if (C->isNullValue())
5173 // Unknown instruction.
5179 static SDValue getMemCmpLoad(const Value *PtrVal, MVT LoadVT,
5181 SelectionDAGBuilder &Builder) {
5183 // Check to see if this load can be trivially constant folded, e.g. if the
5184 // input is from a string literal.
5185 if (const Constant *LoadInput = dyn_cast<Constant>(PtrVal)) {
5186 // Cast pointer to the type we really want to load.
5187 LoadInput = ConstantExpr::getBitCast(const_cast<Constant *>(LoadInput),
5188 PointerType::getUnqual(LoadTy));
5190 if (const Constant *LoadCst = ConstantFoldLoadFromConstPtr(
5191 const_cast<Constant *>(LoadInput), *Builder.DL))
5192 return Builder.getValue(LoadCst);
5195 // Otherwise, we have to emit the load. If the pointer is to unfoldable but
5196 // still constant memory, the input chain can be the entry node.
5198 bool ConstantMemory = false;
5200 // Do not serialize (non-volatile) loads of constant memory with anything.
5201 if (Builder.AA->pointsToConstantMemory(PtrVal)) {
5202 Root = Builder.DAG.getEntryNode();
5203 ConstantMemory = true;
5205 // Do not serialize non-volatile loads against each other.
5206 Root = Builder.DAG.getRoot();
5209 SDValue Ptr = Builder.getValue(PtrVal);
5210 SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurSDLoc(), Root,
5211 Ptr, MachinePointerInfo(PtrVal),
5213 false /*nontemporal*/,
5214 false /*isinvariant*/, 1 /* align=1 */);
5216 if (!ConstantMemory)
5217 Builder.PendingLoads.push_back(LoadVal.getValue(1));
5221 /// processIntegerCallValue - Record the value for an instruction that
5222 /// produces an integer result, converting the type where necessary.
5223 void SelectionDAGBuilder::processIntegerCallValue(const Instruction &I,
5226 EVT VT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
5229 Value = DAG.getSExtOrTrunc(Value, getCurSDLoc(), VT);
5231 Value = DAG.getZExtOrTrunc(Value, getCurSDLoc(), VT);
5232 setValue(&I, Value);
5235 /// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form.
5236 /// If so, return true and lower it, otherwise return false and it will be
5237 /// lowered like a normal call.
5238 bool SelectionDAGBuilder::visitMemCmpCall(const CallInst &I) {
5239 // Verify that the prototype makes sense. int memcmp(void*,void*,size_t)
5240 if (I.getNumArgOperands() != 3)
5243 const Value *LHS = I.getArgOperand(0), *RHS = I.getArgOperand(1);
5244 if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() ||
5245 !I.getArgOperand(2)->getType()->isIntegerTy() ||
5246 !I.getType()->isIntegerTy())
5249 const Value *Size = I.getArgOperand(2);
5250 const ConstantInt *CSize = dyn_cast<ConstantInt>(Size);
5251 if (CSize && CSize->getZExtValue() == 0) {
5252 EVT CallVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
5254 setValue(&I, DAG.getConstant(0, getCurSDLoc(), CallVT));
5258 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5259 std::pair<SDValue, SDValue> Res =
5260 TSI.EmitTargetCodeForMemcmp(DAG, getCurSDLoc(), DAG.getRoot(),
5261 getValue(LHS), getValue(RHS), getValue(Size),
5262 MachinePointerInfo(LHS),
5263 MachinePointerInfo(RHS));
5264 if (Res.first.getNode()) {
5265 processIntegerCallValue(I, Res.first, true);
5266 PendingLoads.push_back(Res.second);
5270 // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0
5271 // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0
5272 if (CSize && IsOnlyUsedInZeroEqualityComparison(&I)) {
5273 bool ActuallyDoIt = true;
5276 switch (CSize->getZExtValue()) {
5278 LoadVT = MVT::Other;
5280 ActuallyDoIt = false;
5284 LoadTy = Type::getInt16Ty(CSize->getContext());
5288 LoadTy = Type::getInt32Ty(CSize->getContext());
5292 LoadTy = Type::getInt64Ty(CSize->getContext());
5296 LoadVT = MVT::v4i32;
5297 LoadTy = Type::getInt32Ty(CSize->getContext());
5298 LoadTy = VectorType::get(LoadTy, 4);
5303 // This turns into unaligned loads. We only do this if the target natively
5304 // supports the MVT we'll be loading or if it is small enough (<= 4) that
5305 // we'll only produce a small number of byte loads.
5307 // Require that we can find a legal MVT, and only do this if the target
5308 // supports unaligned loads of that type. Expanding into byte loads would
5310 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5311 if (ActuallyDoIt && CSize->getZExtValue() > 4) {
5312 unsigned DstAS = LHS->getType()->getPointerAddressSpace();
5313 unsigned SrcAS = RHS->getType()->getPointerAddressSpace();
5314 // TODO: Handle 5 byte compare as 4-byte + 1 byte.
5315 // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads.
5316 // TODO: Check alignment of src and dest ptrs.
5317 if (!TLI.isTypeLegal(LoadVT) ||
5318 !TLI.allowsMisalignedMemoryAccesses(LoadVT, SrcAS) ||
5319 !TLI.allowsMisalignedMemoryAccesses(LoadVT, DstAS))
5320 ActuallyDoIt = false;
5324 SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this);
5325 SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this);
5327 SDValue Res = DAG.getSetCC(getCurSDLoc(), MVT::i1, LHSVal, RHSVal,
5329 processIntegerCallValue(I, Res, false);
5338 /// visitMemChrCall -- See if we can lower a memchr call into an optimized
5339 /// form. If so, return true and lower it, otherwise return false and it
5340 /// will be lowered like a normal call.
5341 bool SelectionDAGBuilder::visitMemChrCall(const CallInst &I) {
5342 // Verify that the prototype makes sense. void *memchr(void *, int, size_t)
5343 if (I.getNumArgOperands() != 3)
5346 const Value *Src = I.getArgOperand(0);
5347 const Value *Char = I.getArgOperand(1);
5348 const Value *Length = I.getArgOperand(2);
5349 if (!Src->getType()->isPointerTy() ||
5350 !Char->getType()->isIntegerTy() ||
5351 !Length->getType()->isIntegerTy() ||
5352 !I.getType()->isPointerTy())
5355 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5356 std::pair<SDValue, SDValue> Res =
5357 TSI.EmitTargetCodeForMemchr(DAG, getCurSDLoc(), DAG.getRoot(),
5358 getValue(Src), getValue(Char), getValue(Length),
5359 MachinePointerInfo(Src));
5360 if (Res.first.getNode()) {
5361 setValue(&I, Res.first);
5362 PendingLoads.push_back(Res.second);
5369 /// visitStrCpyCall -- See if we can lower a strcpy or stpcpy call into an
5370 /// optimized form. If so, return true and lower it, otherwise return false
5371 /// and it will be lowered like a normal call.
5372 bool SelectionDAGBuilder::visitStrCpyCall(const CallInst &I, bool isStpcpy) {
5373 // Verify that the prototype makes sense. char *strcpy(char *, char *)
5374 if (I.getNumArgOperands() != 2)
5377 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5378 if (!Arg0->getType()->isPointerTy() ||
5379 !Arg1->getType()->isPointerTy() ||
5380 !I.getType()->isPointerTy())
5383 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5384 std::pair<SDValue, SDValue> Res =
5385 TSI.EmitTargetCodeForStrcpy(DAG, getCurSDLoc(), getRoot(),
5386 getValue(Arg0), getValue(Arg1),
5387 MachinePointerInfo(Arg0),
5388 MachinePointerInfo(Arg1), isStpcpy);
5389 if (Res.first.getNode()) {
5390 setValue(&I, Res.first);
5391 DAG.setRoot(Res.second);
5398 /// visitStrCmpCall - See if we can lower a call to strcmp in an optimized form.
5399 /// If so, return true and lower it, otherwise return false and it will be
5400 /// lowered like a normal call.
5401 bool SelectionDAGBuilder::visitStrCmpCall(const CallInst &I) {
5402 // Verify that the prototype makes sense. int strcmp(void*,void*)
5403 if (I.getNumArgOperands() != 2)
5406 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5407 if (!Arg0->getType()->isPointerTy() ||
5408 !Arg1->getType()->isPointerTy() ||
5409 !I.getType()->isIntegerTy())
5412 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5413 std::pair<SDValue, SDValue> Res =
5414 TSI.EmitTargetCodeForStrcmp(DAG, getCurSDLoc(), DAG.getRoot(),
5415 getValue(Arg0), getValue(Arg1),
5416 MachinePointerInfo(Arg0),
5417 MachinePointerInfo(Arg1));
5418 if (Res.first.getNode()) {
5419 processIntegerCallValue(I, Res.first, true);
5420 PendingLoads.push_back(Res.second);
5427 /// visitStrLenCall -- See if we can lower a strlen call into an optimized
5428 /// form. If so, return true and lower it, otherwise return false and it
5429 /// will be lowered like a normal call.
5430 bool SelectionDAGBuilder::visitStrLenCall(const CallInst &I) {
5431 // Verify that the prototype makes sense. size_t strlen(char *)
5432 if (I.getNumArgOperands() != 1)
5435 const Value *Arg0 = I.getArgOperand(0);
5436 if (!Arg0->getType()->isPointerTy() || !I.getType()->isIntegerTy())
5439 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5440 std::pair<SDValue, SDValue> Res =
5441 TSI.EmitTargetCodeForStrlen(DAG, getCurSDLoc(), DAG.getRoot(),
5442 getValue(Arg0), MachinePointerInfo(Arg0));
5443 if (Res.first.getNode()) {
5444 processIntegerCallValue(I, Res.first, false);
5445 PendingLoads.push_back(Res.second);
5452 /// visitStrNLenCall -- See if we can lower a strnlen call into an optimized
5453 /// form. If so, return true and lower it, otherwise return false and it
5454 /// will be lowered like a normal call.
5455 bool SelectionDAGBuilder::visitStrNLenCall(const CallInst &I) {
5456 // Verify that the prototype makes sense. size_t strnlen(char *, size_t)
5457 if (I.getNumArgOperands() != 2)
5460 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5461 if (!Arg0->getType()->isPointerTy() ||
5462 !Arg1->getType()->isIntegerTy() ||
5463 !I.getType()->isIntegerTy())
5466 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5467 std::pair<SDValue, SDValue> Res =
5468 TSI.EmitTargetCodeForStrnlen(DAG, getCurSDLoc(), DAG.getRoot(),
5469 getValue(Arg0), getValue(Arg1),
5470 MachinePointerInfo(Arg0));
5471 if (Res.first.getNode()) {
5472 processIntegerCallValue(I, Res.first, false);
5473 PendingLoads.push_back(Res.second);
5480 /// visitUnaryFloatCall - If a call instruction is a unary floating-point
5481 /// operation (as expected), translate it to an SDNode with the specified opcode
5482 /// and return true.
5483 bool SelectionDAGBuilder::visitUnaryFloatCall(const CallInst &I,
5485 // Sanity check that it really is a unary floating-point call.
5486 if (I.getNumArgOperands() != 1 ||
5487 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
5488 I.getType() != I.getArgOperand(0)->getType() ||
5489 !I.onlyReadsMemory())
5492 SDValue Tmp = getValue(I.getArgOperand(0));
5493 setValue(&I, DAG.getNode(Opcode, getCurSDLoc(), Tmp.getValueType(), Tmp));
5497 /// visitBinaryFloatCall - If a call instruction is a binary floating-point
5498 /// operation (as expected), translate it to an SDNode with the specified opcode
5499 /// and return true.
5500 bool SelectionDAGBuilder::visitBinaryFloatCall(const CallInst &I,
5502 // Sanity check that it really is a binary floating-point call.
5503 if (I.getNumArgOperands() != 2 ||
5504 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
5505 I.getType() != I.getArgOperand(0)->getType() ||
5506 I.getType() != I.getArgOperand(1)->getType() ||
5507 !I.onlyReadsMemory())
5510 SDValue Tmp0 = getValue(I.getArgOperand(0));
5511 SDValue Tmp1 = getValue(I.getArgOperand(1));
5512 EVT VT = Tmp0.getValueType();
5513 setValue(&I, DAG.getNode(Opcode, getCurSDLoc(), VT, Tmp0, Tmp1));
5517 void SelectionDAGBuilder::visitCall(const CallInst &I) {
5518 // Handle inline assembly differently.
5519 if (isa<InlineAsm>(I.getCalledValue())) {
5524 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5525 ComputeUsesVAFloatArgument(I, &MMI);
5527 const char *RenameFn = nullptr;
5528 if (Function *F = I.getCalledFunction()) {
5529 if (F->isDeclaration()) {
5530 if (const TargetIntrinsicInfo *II = TM.getIntrinsicInfo()) {
5531 if (unsigned IID = II->getIntrinsicID(F)) {
5532 RenameFn = visitIntrinsicCall(I, IID);
5537 if (Intrinsic::ID IID = F->getIntrinsicID()) {
5538 RenameFn = visitIntrinsicCall(I, IID);
5544 // Check for well-known libc/libm calls. If the function is internal, it
5545 // can't be a library call.
5547 if (!F->hasLocalLinkage() && F->hasName() &&
5548 LibInfo->getLibFunc(F->getName(), Func) &&
5549 LibInfo->hasOptimizedCodeGen(Func)) {
5552 case LibFunc::copysign:
5553 case LibFunc::copysignf:
5554 case LibFunc::copysignl:
5555 if (I.getNumArgOperands() == 2 && // Basic sanity checks.
5556 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5557 I.getType() == I.getArgOperand(0)->getType() &&
5558 I.getType() == I.getArgOperand(1)->getType() &&
5559 I.onlyReadsMemory()) {
5560 SDValue LHS = getValue(I.getArgOperand(0));
5561 SDValue RHS = getValue(I.getArgOperand(1));
5562 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurSDLoc(),
5563 LHS.getValueType(), LHS, RHS));
5568 case LibFunc::fabsf:
5569 case LibFunc::fabsl:
5570 if (visitUnaryFloatCall(I, ISD::FABS))
5574 case LibFunc::fminf:
5575 case LibFunc::fminl:
5576 if (visitBinaryFloatCall(I, ISD::FMINNUM))
5580 case LibFunc::fmaxf:
5581 case LibFunc::fmaxl:
5582 if (visitBinaryFloatCall(I, ISD::FMAXNUM))
5588 if (visitUnaryFloatCall(I, ISD::FSIN))
5594 if (visitUnaryFloatCall(I, ISD::FCOS))
5598 case LibFunc::sqrtf:
5599 case LibFunc::sqrtl:
5600 case LibFunc::sqrt_finite:
5601 case LibFunc::sqrtf_finite:
5602 case LibFunc::sqrtl_finite:
5603 if (visitUnaryFloatCall(I, ISD::FSQRT))
5606 case LibFunc::floor:
5607 case LibFunc::floorf:
5608 case LibFunc::floorl:
5609 if (visitUnaryFloatCall(I, ISD::FFLOOR))
5612 case LibFunc::nearbyint:
5613 case LibFunc::nearbyintf:
5614 case LibFunc::nearbyintl:
5615 if (visitUnaryFloatCall(I, ISD::FNEARBYINT))
5619 case LibFunc::ceilf:
5620 case LibFunc::ceill:
5621 if (visitUnaryFloatCall(I, ISD::FCEIL))
5625 case LibFunc::rintf:
5626 case LibFunc::rintl:
5627 if (visitUnaryFloatCall(I, ISD::FRINT))
5630 case LibFunc::round:
5631 case LibFunc::roundf:
5632 case LibFunc::roundl:
5633 if (visitUnaryFloatCall(I, ISD::FROUND))
5636 case LibFunc::trunc:
5637 case LibFunc::truncf:
5638 case LibFunc::truncl:
5639 if (visitUnaryFloatCall(I, ISD::FTRUNC))
5643 case LibFunc::log2f:
5644 case LibFunc::log2l:
5645 if (visitUnaryFloatCall(I, ISD::FLOG2))
5649 case LibFunc::exp2f:
5650 case LibFunc::exp2l:
5651 if (visitUnaryFloatCall(I, ISD::FEXP2))
5654 case LibFunc::memcmp:
5655 if (visitMemCmpCall(I))
5658 case LibFunc::memchr:
5659 if (visitMemChrCall(I))
5662 case LibFunc::strcpy:
5663 if (visitStrCpyCall(I, false))
5666 case LibFunc::stpcpy:
5667 if (visitStrCpyCall(I, true))
5670 case LibFunc::strcmp:
5671 if (visitStrCmpCall(I))
5674 case LibFunc::strlen:
5675 if (visitStrLenCall(I))
5678 case LibFunc::strnlen:
5679 if (visitStrNLenCall(I))
5688 Callee = getValue(I.getCalledValue());
5690 Callee = DAG.getExternalSymbol(
5692 DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout()));
5694 // Check if we can potentially perform a tail call. More detailed checking is
5695 // be done within LowerCallTo, after more information about the call is known.
5696 LowerCallTo(&I, Callee, I.isTailCall());
5701 /// AsmOperandInfo - This contains information for each constraint that we are
5703 class SDISelAsmOperandInfo : public TargetLowering::AsmOperandInfo {
5705 /// CallOperand - If this is the result output operand or a clobber
5706 /// this is null, otherwise it is the incoming operand to the CallInst.
5707 /// This gets modified as the asm is processed.
5708 SDValue CallOperand;
5710 /// AssignedRegs - If this is a register or register class operand, this
5711 /// contains the set of register corresponding to the operand.
5712 RegsForValue AssignedRegs;
5714 explicit SDISelAsmOperandInfo(const TargetLowering::AsmOperandInfo &info)
5715 : TargetLowering::AsmOperandInfo(info), CallOperand(nullptr,0) {
5718 /// getCallOperandValEVT - Return the EVT of the Value* that this operand
5719 /// corresponds to. If there is no Value* for this operand, it returns
5721 EVT getCallOperandValEVT(LLVMContext &Context, const TargetLowering &TLI,
5722 const DataLayout &DL) const {
5723 if (!CallOperandVal) return MVT::Other;
5725 if (isa<BasicBlock>(CallOperandVal))
5726 return TLI.getPointerTy(DL);
5728 llvm::Type *OpTy = CallOperandVal->getType();
5730 // FIXME: code duplicated from TargetLowering::ParseConstraints().
5731 // If this is an indirect operand, the operand is a pointer to the
5734 llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
5736 report_fatal_error("Indirect operand for inline asm not a pointer!");
5737 OpTy = PtrTy->getElementType();
5740 // Look for vector wrapped in a struct. e.g. { <16 x i8> }.
5741 if (StructType *STy = dyn_cast<StructType>(OpTy))
5742 if (STy->getNumElements() == 1)
5743 OpTy = STy->getElementType(0);
5745 // If OpTy is not a single value, it may be a struct/union that we
5746 // can tile with integers.
5747 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
5748 unsigned BitSize = DL.getTypeSizeInBits(OpTy);
5757 OpTy = IntegerType::get(Context, BitSize);
5762 return TLI.getValueType(DL, OpTy, true);
5766 typedef SmallVector<SDISelAsmOperandInfo,16> SDISelAsmOperandInfoVector;
5768 } // end anonymous namespace
5770 /// GetRegistersForValue - Assign registers (virtual or physical) for the
5771 /// specified operand. We prefer to assign virtual registers, to allow the
5772 /// register allocator to handle the assignment process. However, if the asm
5773 /// uses features that we can't model on machineinstrs, we have SDISel do the
5774 /// allocation. This produces generally horrible, but correct, code.
5776 /// OpInfo describes the operand.
5778 static void GetRegistersForValue(SelectionDAG &DAG,
5779 const TargetLowering &TLI,
5781 SDISelAsmOperandInfo &OpInfo) {
5782 LLVMContext &Context = *DAG.getContext();
5784 MachineFunction &MF = DAG.getMachineFunction();
5785 SmallVector<unsigned, 4> Regs;
5787 // If this is a constraint for a single physreg, or a constraint for a
5788 // register class, find it.
5789 std::pair<unsigned, const TargetRegisterClass *> PhysReg =
5790 TLI.getRegForInlineAsmConstraint(MF.getSubtarget().getRegisterInfo(),
5791 OpInfo.ConstraintCode,
5792 OpInfo.ConstraintVT);
5794 unsigned NumRegs = 1;
5795 if (OpInfo.ConstraintVT != MVT::Other) {
5796 // If this is a FP input in an integer register (or visa versa) insert a bit
5797 // cast of the input value. More generally, handle any case where the input
5798 // value disagrees with the register class we plan to stick this in.
5799 if (OpInfo.Type == InlineAsm::isInput &&
5800 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) {
5801 // Try to convert to the first EVT that the reg class contains. If the
5802 // types are identical size, use a bitcast to convert (e.g. two differing
5804 MVT RegVT = *PhysReg.second->vt_begin();
5805 if (RegVT.getSizeInBits() == OpInfo.CallOperand.getValueSizeInBits()) {
5806 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
5807 RegVT, OpInfo.CallOperand);
5808 OpInfo.ConstraintVT = RegVT;
5809 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) {
5810 // If the input is a FP value and we want it in FP registers, do a
5811 // bitcast to the corresponding integer type. This turns an f64 value
5812 // into i64, which can be passed with two i32 values on a 32-bit
5814 RegVT = MVT::getIntegerVT(OpInfo.ConstraintVT.getSizeInBits());
5815 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
5816 RegVT, OpInfo.CallOperand);
5817 OpInfo.ConstraintVT = RegVT;
5821 NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT);
5825 EVT ValueVT = OpInfo.ConstraintVT;
5827 // If this is a constraint for a specific physical register, like {r17},
5829 if (unsigned AssignedReg = PhysReg.first) {
5830 const TargetRegisterClass *RC = PhysReg.second;
5831 if (OpInfo.ConstraintVT == MVT::Other)
5832 ValueVT = *RC->vt_begin();
5834 // Get the actual register value type. This is important, because the user
5835 // may have asked for (e.g.) the AX register in i32 type. We need to
5836 // remember that AX is actually i16 to get the right extension.
5837 RegVT = *RC->vt_begin();
5839 // This is a explicit reference to a physical register.
5840 Regs.push_back(AssignedReg);
5842 // If this is an expanded reference, add the rest of the regs to Regs.
5844 TargetRegisterClass::iterator I = RC->begin();
5845 for (; *I != AssignedReg; ++I)
5846 assert(I != RC->end() && "Didn't find reg!");
5848 // Already added the first reg.
5850 for (; NumRegs; --NumRegs, ++I) {
5851 assert(I != RC->end() && "Ran out of registers to allocate!");
5856 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5860 // Otherwise, if this was a reference to an LLVM register class, create vregs
5861 // for this reference.
5862 if (const TargetRegisterClass *RC = PhysReg.second) {
5863 RegVT = *RC->vt_begin();
5864 if (OpInfo.ConstraintVT == MVT::Other)
5867 // Create the appropriate number of virtual registers.
5868 MachineRegisterInfo &RegInfo = MF.getRegInfo();
5869 for (; NumRegs; --NumRegs)
5870 Regs.push_back(RegInfo.createVirtualRegister(RC));
5872 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5876 // Otherwise, we couldn't allocate enough registers for this.
5879 /// visitInlineAsm - Handle a call to an InlineAsm object.
5881 void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) {
5882 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
5884 /// ConstraintOperands - Information about all of the constraints.
5885 SDISelAsmOperandInfoVector ConstraintOperands;
5887 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5888 TargetLowering::AsmOperandInfoVector TargetConstraints = TLI.ParseConstraints(
5889 DAG.getDataLayout(), DAG.getSubtarget().getRegisterInfo(), CS);
5891 bool hasMemory = false;
5893 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
5894 unsigned ResNo = 0; // ResNo - The result number of the next output.
5895 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
5896 ConstraintOperands.push_back(SDISelAsmOperandInfo(TargetConstraints[i]));
5897 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
5899 MVT OpVT = MVT::Other;
5901 // Compute the value type for each operand.
5902 switch (OpInfo.Type) {
5903 case InlineAsm::isOutput:
5904 // Indirect outputs just consume an argument.
5905 if (OpInfo.isIndirect) {
5906 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5910 // The return value of the call is this value. As such, there is no
5911 // corresponding argument.
5912 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
5913 if (StructType *STy = dyn_cast<StructType>(CS.getType())) {
5914 OpVT = TLI.getSimpleValueType(DAG.getDataLayout(),
5915 STy->getElementType(ResNo));
5917 assert(ResNo == 0 && "Asm only has one result!");
5918 OpVT = TLI.getSimpleValueType(DAG.getDataLayout(), CS.getType());
5922 case InlineAsm::isInput:
5923 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5925 case InlineAsm::isClobber:
5930 // If this is an input or an indirect output, process the call argument.
5931 // BasicBlocks are labels, currently appearing only in asm's.
5932 if (OpInfo.CallOperandVal) {
5933 if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) {
5934 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]);
5936 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
5939 OpVT = OpInfo.getCallOperandValEVT(*DAG.getContext(), TLI,
5940 DAG.getDataLayout()).getSimpleVT();
5943 OpInfo.ConstraintVT = OpVT;
5945 // Indirect operand accesses access memory.
5946 if (OpInfo.isIndirect)
5949 for (unsigned j = 0, ee = OpInfo.Codes.size(); j != ee; ++j) {
5950 TargetLowering::ConstraintType
5951 CType = TLI.getConstraintType(OpInfo.Codes[j]);
5952 if (CType == TargetLowering::C_Memory) {
5960 SDValue Chain, Flag;
5962 // We won't need to flush pending loads if this asm doesn't touch
5963 // memory and is nonvolatile.
5964 if (hasMemory || IA->hasSideEffects())
5967 Chain = DAG.getRoot();
5969 // Second pass over the constraints: compute which constraint option to use
5970 // and assign registers to constraints that want a specific physreg.
5971 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5972 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5974 // If this is an output operand with a matching input operand, look up the
5975 // matching input. If their types mismatch, e.g. one is an integer, the
5976 // other is floating point, or their sizes are different, flag it as an
5978 if (OpInfo.hasMatchingInput()) {
5979 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
5981 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
5982 const TargetRegisterInfo *TRI = DAG.getSubtarget().getRegisterInfo();
5983 std::pair<unsigned, const TargetRegisterClass *> MatchRC =
5984 TLI.getRegForInlineAsmConstraint(TRI, OpInfo.ConstraintCode,
5985 OpInfo.ConstraintVT);
5986 std::pair<unsigned, const TargetRegisterClass *> InputRC =
5987 TLI.getRegForInlineAsmConstraint(TRI, Input.ConstraintCode,
5988 Input.ConstraintVT);
5989 if ((OpInfo.ConstraintVT.isInteger() !=
5990 Input.ConstraintVT.isInteger()) ||
5991 (MatchRC.second != InputRC.second)) {
5992 report_fatal_error("Unsupported asm: input constraint"
5993 " with a matching output constraint of"
5994 " incompatible type!");
5996 Input.ConstraintVT = OpInfo.ConstraintVT;
6000 // Compute the constraint code and ConstraintType to use.
6001 TLI.ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, &DAG);
6003 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
6004 OpInfo.Type == InlineAsm::isClobber)
6007 // If this is a memory input, and if the operand is not indirect, do what we
6008 // need to to provide an address for the memory input.
6009 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
6010 !OpInfo.isIndirect) {
6011 assert((OpInfo.isMultipleAlternative ||
6012 (OpInfo.Type == InlineAsm::isInput)) &&
6013 "Can only indirectify direct input operands!");
6015 // Memory operands really want the address of the value. If we don't have
6016 // an indirect input, put it in the constpool if we can, otherwise spill
6017 // it to a stack slot.
6018 // TODO: This isn't quite right. We need to handle these according to
6019 // the addressing mode that the constraint wants. Also, this may take
6020 // an additional register for the computation and we don't want that
6023 // If the operand is a float, integer, or vector constant, spill to a
6024 // constant pool entry to get its address.
6025 const Value *OpVal = OpInfo.CallOperandVal;
6026 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
6027 isa<ConstantVector>(OpVal) || isa<ConstantDataVector>(OpVal)) {
6028 OpInfo.CallOperand = DAG.getConstantPool(
6029 cast<Constant>(OpVal), TLI.getPointerTy(DAG.getDataLayout()));
6031 // Otherwise, create a stack slot and emit a store to it before the
6033 Type *Ty = OpVal->getType();
6034 auto &DL = DAG.getDataLayout();
6035 uint64_t TySize = DL.getTypeAllocSize(Ty);
6036 unsigned Align = DL.getPrefTypeAlignment(Ty);
6037 MachineFunction &MF = DAG.getMachineFunction();
6038 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
6040 DAG.getFrameIndex(SSFI, TLI.getPointerTy(DAG.getDataLayout()));
6041 Chain = DAG.getStore(Chain, getCurSDLoc(),
6042 OpInfo.CallOperand, StackSlot,
6043 MachinePointerInfo::getFixedStack(SSFI),
6045 OpInfo.CallOperand = StackSlot;
6048 // There is no longer a Value* corresponding to this operand.
6049 OpInfo.CallOperandVal = nullptr;
6051 // It is now an indirect operand.
6052 OpInfo.isIndirect = true;
6055 // If this constraint is for a specific register, allocate it before
6057 if (OpInfo.ConstraintType == TargetLowering::C_Register)
6058 GetRegistersForValue(DAG, TLI, getCurSDLoc(), OpInfo);
6061 // Second pass - Loop over all of the operands, assigning virtual or physregs
6062 // to register class operands.
6063 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6064 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6066 // C_Register operands have already been allocated, Other/Memory don't need
6068 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
6069 GetRegistersForValue(DAG, TLI, getCurSDLoc(), OpInfo);
6072 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
6073 std::vector<SDValue> AsmNodeOperands;
6074 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain
6075 AsmNodeOperands.push_back(DAG.getTargetExternalSymbol(
6076 IA->getAsmString().c_str(), TLI.getPointerTy(DAG.getDataLayout())));
6078 // If we have a !srcloc metadata node associated with it, we want to attach
6079 // this to the ultimately generated inline asm machineinstr. To do this, we
6080 // pass in the third operand as this (potentially null) inline asm MDNode.
6081 const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc");
6082 AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc));
6084 // Remember the HasSideEffect, AlignStack, AsmDialect, MayLoad and MayStore
6085 // bits as operand 3.
6086 unsigned ExtraInfo = 0;
6087 if (IA->hasSideEffects())
6088 ExtraInfo |= InlineAsm::Extra_HasSideEffects;
6089 if (IA->isAlignStack())
6090 ExtraInfo |= InlineAsm::Extra_IsAlignStack;
6091 // Set the asm dialect.
6092 ExtraInfo |= IA->getDialect() * InlineAsm::Extra_AsmDialect;
6094 // Determine if this InlineAsm MayLoad or MayStore based on the constraints.
6095 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
6096 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
6098 // Compute the constraint code and ConstraintType to use.
6099 TLI.ComputeConstraintToUse(OpInfo, SDValue());
6101 // Ideally, we would only check against memory constraints. However, the
6102 // meaning of an other constraint can be target-specific and we can't easily
6103 // reason about it. Therefore, be conservative and set MayLoad/MayStore
6104 // for other constriants as well.
6105 if (OpInfo.ConstraintType == TargetLowering::C_Memory ||
6106 OpInfo.ConstraintType == TargetLowering::C_Other) {
6107 if (OpInfo.Type == InlineAsm::isInput)
6108 ExtraInfo |= InlineAsm::Extra_MayLoad;
6109 else if (OpInfo.Type == InlineAsm::isOutput)
6110 ExtraInfo |= InlineAsm::Extra_MayStore;
6111 else if (OpInfo.Type == InlineAsm::isClobber)
6112 ExtraInfo |= (InlineAsm::Extra_MayLoad | InlineAsm::Extra_MayStore);
6116 AsmNodeOperands.push_back(DAG.getTargetConstant(
6117 ExtraInfo, getCurSDLoc(), TLI.getPointerTy(DAG.getDataLayout())));
6119 // Loop over all of the inputs, copying the operand values into the
6120 // appropriate registers and processing the output regs.
6121 RegsForValue RetValRegs;
6123 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
6124 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
6126 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6127 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6129 switch (OpInfo.Type) {
6130 case InlineAsm::isOutput: {
6131 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
6132 OpInfo.ConstraintType != TargetLowering::C_Register) {
6133 // Memory output, or 'other' output (e.g. 'X' constraint).
6134 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
6136 unsigned ConstraintID =
6137 TLI.getInlineAsmMemConstraint(OpInfo.ConstraintCode);
6138 assert(ConstraintID != InlineAsm::Constraint_Unknown &&
6139 "Failed to convert memory constraint code to constraint id.");
6141 // Add information to the INLINEASM node to know about this output.
6142 unsigned OpFlags = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
6143 OpFlags = InlineAsm::getFlagWordForMem(OpFlags, ConstraintID);
6144 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlags, getCurSDLoc(),
6146 AsmNodeOperands.push_back(OpInfo.CallOperand);
6150 // Otherwise, this is a register or register class output.
6152 // Copy the output from the appropriate register. Find a register that
6154 if (OpInfo.AssignedRegs.Regs.empty()) {
6155 LLVMContext &Ctx = *DAG.getContext();
6156 Ctx.emitError(CS.getInstruction(),
6157 "couldn't allocate output register for constraint '" +
6158 Twine(OpInfo.ConstraintCode) + "'");
6162 // If this is an indirect operand, store through the pointer after the
6164 if (OpInfo.isIndirect) {
6165 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
6166 OpInfo.CallOperandVal));
6168 // This is the result value of the call.
6169 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
6170 // Concatenate this output onto the outputs list.
6171 RetValRegs.append(OpInfo.AssignedRegs);
6174 // Add information to the INLINEASM node to know that this register is
6177 .AddInlineAsmOperands(OpInfo.isEarlyClobber
6178 ? InlineAsm::Kind_RegDefEarlyClobber
6179 : InlineAsm::Kind_RegDef,
6180 false, 0, getCurSDLoc(), DAG, AsmNodeOperands);
6183 case InlineAsm::isInput: {
6184 SDValue InOperandVal = OpInfo.CallOperand;
6186 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint?
6187 // If this is required to match an output register we have already set,
6188 // just use its register.
6189 unsigned OperandNo = OpInfo.getMatchedOperand();
6191 // Scan until we find the definition we already emitted of this operand.
6192 // When we find it, create a RegsForValue operand.
6193 unsigned CurOp = InlineAsm::Op_FirstOperand;
6194 for (; OperandNo; --OperandNo) {
6195 // Advance to the next operand.
6197 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6198 assert((InlineAsm::isRegDefKind(OpFlag) ||
6199 InlineAsm::isRegDefEarlyClobberKind(OpFlag) ||
6200 InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?");
6201 CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1;
6205 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6206 if (InlineAsm::isRegDefKind(OpFlag) ||
6207 InlineAsm::isRegDefEarlyClobberKind(OpFlag)) {
6208 // Add (OpFlag&0xffff)>>3 registers to MatchedRegs.
6209 if (OpInfo.isIndirect) {
6210 // This happens on gcc/testsuite/gcc.dg/pr8788-1.c
6211 LLVMContext &Ctx = *DAG.getContext();
6212 Ctx.emitError(CS.getInstruction(), "inline asm not supported yet:"
6213 " don't know how to handle tied "
6214 "indirect register inputs");
6218 RegsForValue MatchedRegs;
6219 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
6220 MVT RegVT = AsmNodeOperands[CurOp+1].getSimpleValueType();
6221 MatchedRegs.RegVTs.push_back(RegVT);
6222 MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo();
6223 for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag);
6225 if (const TargetRegisterClass *RC = TLI.getRegClassFor(RegVT))
6226 MatchedRegs.Regs.push_back(RegInfo.createVirtualRegister(RC));
6228 LLVMContext &Ctx = *DAG.getContext();
6229 Ctx.emitError(CS.getInstruction(),
6230 "inline asm error: This value"
6231 " type register class is not natively supported!");
6235 SDLoc dl = getCurSDLoc();
6236 // Use the produced MatchedRegs object to
6237 MatchedRegs.getCopyToRegs(InOperandVal, DAG, dl,
6238 Chain, &Flag, CS.getInstruction());
6239 MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse,
6240 true, OpInfo.getMatchedOperand(), dl,
6241 DAG, AsmNodeOperands);
6245 assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!");
6246 assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 &&
6247 "Unexpected number of operands");
6248 // Add information to the INLINEASM node to know about this input.
6249 // See InlineAsm.h isUseOperandTiedToDef.
6250 OpFlag = InlineAsm::convertMemFlagWordToMatchingFlagWord(OpFlag);
6251 OpFlag = InlineAsm::getFlagWordForMatchingOp(OpFlag,
6252 OpInfo.getMatchedOperand());
6253 AsmNodeOperands.push_back(DAG.getTargetConstant(
6254 OpFlag, getCurSDLoc(), TLI.getPointerTy(DAG.getDataLayout())));
6255 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]);
6259 // Treat indirect 'X' constraint as memory.
6260 if (OpInfo.ConstraintType == TargetLowering::C_Other &&
6262 OpInfo.ConstraintType = TargetLowering::C_Memory;
6264 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
6265 std::vector<SDValue> Ops;
6266 TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode,
6269 LLVMContext &Ctx = *DAG.getContext();
6270 Ctx.emitError(CS.getInstruction(),
6271 "invalid operand for inline asm constraint '" +
6272 Twine(OpInfo.ConstraintCode) + "'");
6276 // Add information to the INLINEASM node to know about this input.
6277 unsigned ResOpType =
6278 InlineAsm::getFlagWord(InlineAsm::Kind_Imm, Ops.size());
6279 AsmNodeOperands.push_back(DAG.getTargetConstant(
6280 ResOpType, getCurSDLoc(), TLI.getPointerTy(DAG.getDataLayout())));
6281 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
6285 if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
6286 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
6287 assert(InOperandVal.getValueType() ==
6288 TLI.getPointerTy(DAG.getDataLayout()) &&
6289 "Memory operands expect pointer values");
6291 unsigned ConstraintID =
6292 TLI.getInlineAsmMemConstraint(OpInfo.ConstraintCode);
6293 assert(ConstraintID != InlineAsm::Constraint_Unknown &&
6294 "Failed to convert memory constraint code to constraint id.");
6296 // Add information to the INLINEASM node to know about this input.
6297 unsigned ResOpType = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
6298 ResOpType = InlineAsm::getFlagWordForMem(ResOpType, ConstraintID);
6299 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
6302 AsmNodeOperands.push_back(InOperandVal);
6306 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
6307 OpInfo.ConstraintType == TargetLowering::C_Register) &&
6308 "Unknown constraint type!");
6310 // TODO: Support this.
6311 if (OpInfo.isIndirect) {
6312 LLVMContext &Ctx = *DAG.getContext();
6313 Ctx.emitError(CS.getInstruction(),
6314 "Don't know how to handle indirect register inputs yet "
6315 "for constraint '" +
6316 Twine(OpInfo.ConstraintCode) + "'");
6320 // Copy the input into the appropriate registers.
6321 if (OpInfo.AssignedRegs.Regs.empty()) {
6322 LLVMContext &Ctx = *DAG.getContext();
6323 Ctx.emitError(CS.getInstruction(),
6324 "couldn't allocate input reg for constraint '" +
6325 Twine(OpInfo.ConstraintCode) + "'");
6329 SDLoc dl = getCurSDLoc();
6331 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, dl,
6332 Chain, &Flag, CS.getInstruction());
6334 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, false, 0,
6335 dl, DAG, AsmNodeOperands);
6338 case InlineAsm::isClobber: {
6339 // Add the clobbered value to the operand list, so that the register
6340 // allocator is aware that the physreg got clobbered.
6341 if (!OpInfo.AssignedRegs.Regs.empty())
6342 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_Clobber,
6343 false, 0, getCurSDLoc(), DAG,
6350 // Finish up input operands. Set the input chain and add the flag last.
6351 AsmNodeOperands[InlineAsm::Op_InputChain] = Chain;
6352 if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
6354 Chain = DAG.getNode(ISD::INLINEASM, getCurSDLoc(),
6355 DAG.getVTList(MVT::Other, MVT::Glue), AsmNodeOperands);
6356 Flag = Chain.getValue(1);
6358 // If this asm returns a register value, copy the result from that register
6359 // and set it as the value of the call.
6360 if (!RetValRegs.Regs.empty()) {
6361 SDValue Val = RetValRegs.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(),
6362 Chain, &Flag, CS.getInstruction());
6364 // FIXME: Why don't we do this for inline asms with MRVs?
6365 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) {
6366 EVT ResultType = TLI.getValueType(DAG.getDataLayout(), CS.getType());
6368 // If any of the results of the inline asm is a vector, it may have the
6369 // wrong width/num elts. This can happen for register classes that can
6370 // contain multiple different value types. The preg or vreg allocated may
6371 // not have the same VT as was expected. Convert it to the right type
6372 // with bit_convert.
6373 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) {
6374 Val = DAG.getNode(ISD::BITCAST, getCurSDLoc(),
6377 } else if (ResultType != Val.getValueType() &&
6378 ResultType.isInteger() && Val.getValueType().isInteger()) {
6379 // If a result value was tied to an input value, the computed result may
6380 // have a wider width than the expected result. Extract the relevant
6382 Val = DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), ResultType, Val);
6385 assert(ResultType == Val.getValueType() && "Asm result value mismatch!");
6388 setValue(CS.getInstruction(), Val);
6389 // Don't need to use this as a chain in this case.
6390 if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty())
6394 std::vector<std::pair<SDValue, const Value *> > StoresToEmit;
6396 // Process indirect outputs, first output all of the flagged copies out of
6398 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
6399 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
6400 const Value *Ptr = IndirectStoresToEmit[i].second;
6401 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(),
6403 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
6406 // Emit the non-flagged stores from the physregs.
6407 SmallVector<SDValue, 8> OutChains;
6408 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) {
6409 SDValue Val = DAG.getStore(Chain, getCurSDLoc(),
6410 StoresToEmit[i].first,
6411 getValue(StoresToEmit[i].second),
6412 MachinePointerInfo(StoresToEmit[i].second),
6414 OutChains.push_back(Val);
6417 if (!OutChains.empty())
6418 Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other, OutChains);
6423 void SelectionDAGBuilder::visitVAStart(const CallInst &I) {
6424 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurSDLoc(),
6425 MVT::Other, getRoot(),
6426 getValue(I.getArgOperand(0)),
6427 DAG.getSrcValue(I.getArgOperand(0))));
6430 void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) {
6431 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6432 const DataLayout &DL = DAG.getDataLayout();
6433 SDValue V = DAG.getVAArg(TLI.getValueType(DAG.getDataLayout(), I.getType()),
6434 getCurSDLoc(), getRoot(), getValue(I.getOperand(0)),
6435 DAG.getSrcValue(I.getOperand(0)),
6436 DL.getABITypeAlignment(I.getType()));
6438 DAG.setRoot(V.getValue(1));
6441 void SelectionDAGBuilder::visitVAEnd(const CallInst &I) {
6442 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurSDLoc(),
6443 MVT::Other, getRoot(),
6444 getValue(I.getArgOperand(0)),
6445 DAG.getSrcValue(I.getArgOperand(0))));
6448 void SelectionDAGBuilder::visitVACopy(const CallInst &I) {
6449 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurSDLoc(),
6450 MVT::Other, getRoot(),
6451 getValue(I.getArgOperand(0)),
6452 getValue(I.getArgOperand(1)),
6453 DAG.getSrcValue(I.getArgOperand(0)),
6454 DAG.getSrcValue(I.getArgOperand(1))));
6457 /// \brief Lower an argument list according to the target calling convention.
6459 /// \return A tuple of <return-value, token-chain>
6461 /// This is a helper for lowering intrinsics that follow a target calling
6462 /// convention or require stack pointer adjustment. Only a subset of the
6463 /// intrinsic's operands need to participate in the calling convention.
6464 std::pair<SDValue, SDValue>
6465 SelectionDAGBuilder::lowerCallOperands(ImmutableCallSite CS, unsigned ArgIdx,
6466 unsigned NumArgs, SDValue Callee,
6468 MachineBasicBlock *LandingPad,
6469 bool IsPatchPoint) {
6470 TargetLowering::ArgListTy Args;
6471 Args.reserve(NumArgs);
6473 // Populate the argument list.
6474 // Attributes for args start at offset 1, after the return attribute.
6475 for (unsigned ArgI = ArgIdx, ArgE = ArgIdx + NumArgs, AttrI = ArgIdx + 1;
6476 ArgI != ArgE; ++ArgI) {
6477 const Value *V = CS->getOperand(ArgI);
6479 assert(!V->getType()->isEmptyTy() && "Empty type passed to intrinsic.");
6481 TargetLowering::ArgListEntry Entry;
6482 Entry.Node = getValue(V);
6483 Entry.Ty = V->getType();
6484 Entry.setAttributes(&CS, AttrI);
6485 Args.push_back(Entry);
6488 TargetLowering::CallLoweringInfo CLI(DAG);
6489 CLI.setDebugLoc(getCurSDLoc()).setChain(getRoot())
6490 .setCallee(CS.getCallingConv(), ReturnTy, Callee, std::move(Args), NumArgs)
6491 .setDiscardResult(CS->use_empty()).setIsPatchPoint(IsPatchPoint);
6493 return lowerInvokable(CLI, LandingPad);
6496 /// \brief Add a stack map intrinsic call's live variable operands to a stackmap
6497 /// or patchpoint target node's operand list.
6499 /// Constants are converted to TargetConstants purely as an optimization to
6500 /// avoid constant materialization and register allocation.
6502 /// FrameIndex operands are converted to TargetFrameIndex so that ISEL does not
6503 /// generate addess computation nodes, and so ExpandISelPseudo can convert the
6504 /// TargetFrameIndex into a DirectMemRefOp StackMap location. This avoids
6505 /// address materialization and register allocation, but may also be required
6506 /// for correctness. If a StackMap (or PatchPoint) intrinsic directly uses an
6507 /// alloca in the entry block, then the runtime may assume that the alloca's
6508 /// StackMap location can be read immediately after compilation and that the
6509 /// location is valid at any point during execution (this is similar to the
6510 /// assumption made by the llvm.gcroot intrinsic). If the alloca's location were
6511 /// only available in a register, then the runtime would need to trap when
6512 /// execution reaches the StackMap in order to read the alloca's location.
6513 static void addStackMapLiveVars(ImmutableCallSite CS, unsigned StartIdx,
6514 SDLoc DL, SmallVectorImpl<SDValue> &Ops,
6515 SelectionDAGBuilder &Builder) {
6516 for (unsigned i = StartIdx, e = CS.arg_size(); i != e; ++i) {
6517 SDValue OpVal = Builder.getValue(CS.getArgument(i));
6518 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(OpVal)) {
6520 Builder.DAG.getTargetConstant(StackMaps::ConstantOp, DL, MVT::i64));
6522 Builder.DAG.getTargetConstant(C->getSExtValue(), DL, MVT::i64));
6523 } else if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(OpVal)) {
6524 const TargetLowering &TLI = Builder.DAG.getTargetLoweringInfo();
6525 Ops.push_back(Builder.DAG.getTargetFrameIndex(
6526 FI->getIndex(), TLI.getPointerTy(Builder.DAG.getDataLayout())));
6528 Ops.push_back(OpVal);
6532 /// \brief Lower llvm.experimental.stackmap directly to its target opcode.
6533 void SelectionDAGBuilder::visitStackmap(const CallInst &CI) {
6534 // void @llvm.experimental.stackmap(i32 <id>, i32 <numShadowBytes>,
6535 // [live variables...])
6537 assert(CI.getType()->isVoidTy() && "Stackmap cannot return a value.");
6539 SDValue Chain, InFlag, Callee, NullPtr;
6540 SmallVector<SDValue, 32> Ops;
6542 SDLoc DL = getCurSDLoc();
6543 Callee = getValue(CI.getCalledValue());
6544 NullPtr = DAG.getIntPtrConstant(0, DL, true);
6546 // The stackmap intrinsic only records the live variables (the arguemnts
6547 // passed to it) and emits NOPS (if requested). Unlike the patchpoint
6548 // intrinsic, this won't be lowered to a function call. This means we don't
6549 // have to worry about calling conventions and target specific lowering code.
6550 // Instead we perform the call lowering right here.
6552 // chain, flag = CALLSEQ_START(chain, 0)
6553 // chain, flag = STACKMAP(id, nbytes, ..., chain, flag)
6554 // chain, flag = CALLSEQ_END(chain, 0, 0, flag)
6556 Chain = DAG.getCALLSEQ_START(getRoot(), NullPtr, DL);
6557 InFlag = Chain.getValue(1);
6559 // Add the <id> and <numBytes> constants.
6560 SDValue IDVal = getValue(CI.getOperand(PatchPointOpers::IDPos));
6561 Ops.push_back(DAG.getTargetConstant(
6562 cast<ConstantSDNode>(IDVal)->getZExtValue(), DL, MVT::i64));
6563 SDValue NBytesVal = getValue(CI.getOperand(PatchPointOpers::NBytesPos));
6564 Ops.push_back(DAG.getTargetConstant(
6565 cast<ConstantSDNode>(NBytesVal)->getZExtValue(), DL,
6568 // Push live variables for the stack map.
6569 addStackMapLiveVars(&CI, 2, DL, Ops, *this);
6571 // We are not pushing any register mask info here on the operands list,
6572 // because the stackmap doesn't clobber anything.
6574 // Push the chain and the glue flag.
6575 Ops.push_back(Chain);
6576 Ops.push_back(InFlag);
6578 // Create the STACKMAP node.
6579 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6580 SDNode *SM = DAG.getMachineNode(TargetOpcode::STACKMAP, DL, NodeTys, Ops);
6581 Chain = SDValue(SM, 0);
6582 InFlag = Chain.getValue(1);
6584 Chain = DAG.getCALLSEQ_END(Chain, NullPtr, NullPtr, InFlag, DL);
6586 // Stackmaps don't generate values, so nothing goes into the NodeMap.
6588 // Set the root to the target-lowered call chain.
6591 // Inform the Frame Information that we have a stackmap in this function.
6592 FuncInfo.MF->getFrameInfo()->setHasStackMap();
6595 /// \brief Lower llvm.experimental.patchpoint directly to its target opcode.
6596 void SelectionDAGBuilder::visitPatchpoint(ImmutableCallSite CS,
6597 MachineBasicBlock *LandingPad) {
6598 // void|i64 @llvm.experimental.patchpoint.void|i64(i64 <id>,
6603 // [live variables...])
6605 CallingConv::ID CC = CS.getCallingConv();
6606 bool IsAnyRegCC = CC == CallingConv::AnyReg;
6607 bool HasDef = !CS->getType()->isVoidTy();
6608 SDLoc dl = getCurSDLoc();
6609 SDValue Callee = getValue(CS->getOperand(PatchPointOpers::TargetPos));
6611 // Handle immediate and symbolic callees.
6612 if (auto* ConstCallee = dyn_cast<ConstantSDNode>(Callee))
6613 Callee = DAG.getIntPtrConstant(ConstCallee->getZExtValue(), dl,
6615 else if (auto* SymbolicCallee = dyn_cast<GlobalAddressSDNode>(Callee))
6616 Callee = DAG.getTargetGlobalAddress(SymbolicCallee->getGlobal(),
6617 SDLoc(SymbolicCallee),
6618 SymbolicCallee->getValueType(0));
6620 // Get the real number of arguments participating in the call <numArgs>
6621 SDValue NArgVal = getValue(CS.getArgument(PatchPointOpers::NArgPos));
6622 unsigned NumArgs = cast<ConstantSDNode>(NArgVal)->getZExtValue();
6624 // Skip the four meta args: <id>, <numNopBytes>, <target>, <numArgs>
6625 // Intrinsics include all meta-operands up to but not including CC.
6626 unsigned NumMetaOpers = PatchPointOpers::CCPos;
6627 assert(CS.arg_size() >= NumMetaOpers + NumArgs &&
6628 "Not enough arguments provided to the patchpoint intrinsic");
6630 // For AnyRegCC the arguments are lowered later on manually.
6631 unsigned NumCallArgs = IsAnyRegCC ? 0 : NumArgs;
6633 IsAnyRegCC ? Type::getVoidTy(*DAG.getContext()) : CS->getType();
6634 std::pair<SDValue, SDValue> Result =
6635 lowerCallOperands(CS, NumMetaOpers, NumCallArgs, Callee, ReturnTy,
6638 SDNode *CallEnd = Result.second.getNode();
6639 if (HasDef && (CallEnd->getOpcode() == ISD::CopyFromReg))
6640 CallEnd = CallEnd->getOperand(0).getNode();
6642 /// Get a call instruction from the call sequence chain.
6643 /// Tail calls are not allowed.
6644 assert(CallEnd->getOpcode() == ISD::CALLSEQ_END &&
6645 "Expected a callseq node.");
6646 SDNode *Call = CallEnd->getOperand(0).getNode();
6647 bool HasGlue = Call->getGluedNode();
6649 // Replace the target specific call node with the patchable intrinsic.
6650 SmallVector<SDValue, 8> Ops;
6652 // Add the <id> and <numBytes> constants.
6653 SDValue IDVal = getValue(CS->getOperand(PatchPointOpers::IDPos));
6654 Ops.push_back(DAG.getTargetConstant(
6655 cast<ConstantSDNode>(IDVal)->getZExtValue(), dl, MVT::i64));
6656 SDValue NBytesVal = getValue(CS->getOperand(PatchPointOpers::NBytesPos));
6657 Ops.push_back(DAG.getTargetConstant(
6658 cast<ConstantSDNode>(NBytesVal)->getZExtValue(), dl,
6662 Ops.push_back(Callee);
6664 // Adjust <numArgs> to account for any arguments that have been passed on the
6666 // Call Node: Chain, Target, {Args}, RegMask, [Glue]
6667 unsigned NumCallRegArgs = Call->getNumOperands() - (HasGlue ? 4 : 3);
6668 NumCallRegArgs = IsAnyRegCC ? NumArgs : NumCallRegArgs;
6669 Ops.push_back(DAG.getTargetConstant(NumCallRegArgs, dl, MVT::i32));
6671 // Add the calling convention
6672 Ops.push_back(DAG.getTargetConstant((unsigned)CC, dl, MVT::i32));
6674 // Add the arguments we omitted previously. The register allocator should
6675 // place these in any free register.
6677 for (unsigned i = NumMetaOpers, e = NumMetaOpers + NumArgs; i != e; ++i)
6678 Ops.push_back(getValue(CS.getArgument(i)));
6680 // Push the arguments from the call instruction up to the register mask.
6681 SDNode::op_iterator e = HasGlue ? Call->op_end()-2 : Call->op_end()-1;
6682 Ops.append(Call->op_begin() + 2, e);
6684 // Push live variables for the stack map.
6685 addStackMapLiveVars(CS, NumMetaOpers + NumArgs, dl, Ops, *this);
6687 // Push the register mask info.
6689 Ops.push_back(*(Call->op_end()-2));
6691 Ops.push_back(*(Call->op_end()-1));
6693 // Push the chain (this is originally the first operand of the call, but
6694 // becomes now the last or second to last operand).
6695 Ops.push_back(*(Call->op_begin()));
6697 // Push the glue flag (last operand).
6699 Ops.push_back(*(Call->op_end()-1));
6702 if (IsAnyRegCC && HasDef) {
6703 // Create the return types based on the intrinsic definition
6704 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6705 SmallVector<EVT, 3> ValueVTs;
6706 ComputeValueVTs(TLI, DAG.getDataLayout(), CS->getType(), ValueVTs);
6707 assert(ValueVTs.size() == 1 && "Expected only one return value type.");
6709 // There is always a chain and a glue type at the end
6710 ValueVTs.push_back(MVT::Other);
6711 ValueVTs.push_back(MVT::Glue);
6712 NodeTys = DAG.getVTList(ValueVTs);
6714 NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6716 // Replace the target specific call node with a PATCHPOINT node.
6717 MachineSDNode *MN = DAG.getMachineNode(TargetOpcode::PATCHPOINT,
6720 // Update the NodeMap.
6723 setValue(CS.getInstruction(), SDValue(MN, 0));
6725 setValue(CS.getInstruction(), Result.first);
6728 // Fixup the consumers of the intrinsic. The chain and glue may be used in the
6729 // call sequence. Furthermore the location of the chain and glue can change
6730 // when the AnyReg calling convention is used and the intrinsic returns a
6732 if (IsAnyRegCC && HasDef) {
6733 SDValue From[] = {SDValue(Call, 0), SDValue(Call, 1)};
6734 SDValue To[] = {SDValue(MN, 1), SDValue(MN, 2)};
6735 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
6737 DAG.ReplaceAllUsesWith(Call, MN);
6738 DAG.DeleteNode(Call);
6740 // Inform the Frame Information that we have a patchpoint in this function.
6741 FuncInfo.MF->getFrameInfo()->setHasPatchPoint();
6744 /// Returns an AttributeSet representing the attributes applied to the return
6745 /// value of the given call.
6746 static AttributeSet getReturnAttrs(TargetLowering::CallLoweringInfo &CLI) {
6747 SmallVector<Attribute::AttrKind, 2> Attrs;
6749 Attrs.push_back(Attribute::SExt);
6751 Attrs.push_back(Attribute::ZExt);
6753 Attrs.push_back(Attribute::InReg);
6755 return AttributeSet::get(CLI.RetTy->getContext(), AttributeSet::ReturnIndex,
6759 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
6760 /// implementation, which just calls LowerCall.
6761 /// FIXME: When all targets are
6762 /// migrated to using LowerCall, this hook should be integrated into SDISel.
6763 std::pair<SDValue, SDValue>
6764 TargetLowering::LowerCallTo(TargetLowering::CallLoweringInfo &CLI) const {
6765 // Handle the incoming return values from the call.
6767 Type *OrigRetTy = CLI.RetTy;
6768 SmallVector<EVT, 4> RetTys;
6769 SmallVector<uint64_t, 4> Offsets;
6770 auto &DL = CLI.DAG.getDataLayout();
6771 ComputeValueVTs(*this, DL, CLI.RetTy, RetTys, &Offsets);
6773 SmallVector<ISD::OutputArg, 4> Outs;
6774 GetReturnInfo(CLI.RetTy, getReturnAttrs(CLI), Outs, *this, DL);
6776 bool CanLowerReturn =
6777 this->CanLowerReturn(CLI.CallConv, CLI.DAG.getMachineFunction(),
6778 CLI.IsVarArg, Outs, CLI.RetTy->getContext());
6780 SDValue DemoteStackSlot;
6781 int DemoteStackIdx = -100;
6782 if (!CanLowerReturn) {
6783 // FIXME: equivalent assert?
6784 // assert(!CS.hasInAllocaArgument() &&
6785 // "sret demotion is incompatible with inalloca");
6786 uint64_t TySize = DL.getTypeAllocSize(CLI.RetTy);
6787 unsigned Align = DL.getPrefTypeAlignment(CLI.RetTy);
6788 MachineFunction &MF = CLI.DAG.getMachineFunction();
6789 DemoteStackIdx = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
6790 Type *StackSlotPtrType = PointerType::getUnqual(CLI.RetTy);
6792 DemoteStackSlot = CLI.DAG.getFrameIndex(DemoteStackIdx, getPointerTy(DL));
6794 Entry.Node = DemoteStackSlot;
6795 Entry.Ty = StackSlotPtrType;
6796 Entry.isSExt = false;
6797 Entry.isZExt = false;
6798 Entry.isInReg = false;
6799 Entry.isSRet = true;
6800 Entry.isNest = false;
6801 Entry.isByVal = false;
6802 Entry.isReturned = false;
6803 Entry.Alignment = Align;
6804 CLI.getArgs().insert(CLI.getArgs().begin(), Entry);
6805 CLI.RetTy = Type::getVoidTy(CLI.RetTy->getContext());
6807 // sret demotion isn't compatible with tail-calls, since the sret argument
6808 // points into the callers stack frame.
6809 CLI.IsTailCall = false;
6811 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
6813 MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
6814 unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
6815 for (unsigned i = 0; i != NumRegs; ++i) {
6816 ISD::InputArg MyFlags;
6817 MyFlags.VT = RegisterVT;
6819 MyFlags.Used = CLI.IsReturnValueUsed;
6821 MyFlags.Flags.setSExt();
6823 MyFlags.Flags.setZExt();
6825 MyFlags.Flags.setInReg();
6826 CLI.Ins.push_back(MyFlags);
6831 // Handle all of the outgoing arguments.
6833 CLI.OutVals.clear();
6834 ArgListTy &Args = CLI.getArgs();
6835 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
6836 SmallVector<EVT, 4> ValueVTs;
6837 ComputeValueVTs(*this, DL, Args[i].Ty, ValueVTs);
6838 Type *FinalType = Args[i].Ty;
6839 if (Args[i].isByVal)
6840 FinalType = cast<PointerType>(Args[i].Ty)->getElementType();
6841 bool NeedsRegBlock = functionArgumentNeedsConsecutiveRegisters(
6842 FinalType, CLI.CallConv, CLI.IsVarArg);
6843 for (unsigned Value = 0, NumValues = ValueVTs.size(); Value != NumValues;
6845 EVT VT = ValueVTs[Value];
6846 Type *ArgTy = VT.getTypeForEVT(CLI.RetTy->getContext());
6847 SDValue Op = SDValue(Args[i].Node.getNode(),
6848 Args[i].Node.getResNo() + Value);
6849 ISD::ArgFlagsTy Flags;
6850 unsigned OriginalAlignment = DL.getABITypeAlignment(ArgTy);
6856 if (Args[i].isInReg)
6860 if (Args[i].isByVal)
6862 if (Args[i].isInAlloca) {
6863 Flags.setInAlloca();
6864 // Set the byval flag for CCAssignFn callbacks that don't know about
6865 // inalloca. This way we can know how many bytes we should've allocated
6866 // and how many bytes a callee cleanup function will pop. If we port
6867 // inalloca to more targets, we'll have to add custom inalloca handling
6868 // in the various CC lowering callbacks.
6871 if (Args[i].isByVal || Args[i].isInAlloca) {
6872 PointerType *Ty = cast<PointerType>(Args[i].Ty);
6873 Type *ElementTy = Ty->getElementType();
6874 Flags.setByValSize(DL.getTypeAllocSize(ElementTy));
6875 // For ByVal, alignment should come from FE. BE will guess if this
6876 // info is not there but there are cases it cannot get right.
6877 unsigned FrameAlign;
6878 if (Args[i].Alignment)
6879 FrameAlign = Args[i].Alignment;
6881 FrameAlign = getByValTypeAlignment(ElementTy);
6882 Flags.setByValAlign(FrameAlign);
6887 Flags.setInConsecutiveRegs();
6888 Flags.setOrigAlign(OriginalAlignment);
6890 MVT PartVT = getRegisterType(CLI.RetTy->getContext(), VT);
6891 unsigned NumParts = getNumRegisters(CLI.RetTy->getContext(), VT);
6892 SmallVector<SDValue, 4> Parts(NumParts);
6893 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
6896 ExtendKind = ISD::SIGN_EXTEND;
6897 else if (Args[i].isZExt)
6898 ExtendKind = ISD::ZERO_EXTEND;
6900 // Conservatively only handle 'returned' on non-vectors for now
6901 if (Args[i].isReturned && !Op.getValueType().isVector()) {
6902 assert(CLI.RetTy == Args[i].Ty && RetTys.size() == NumValues &&
6903 "unexpected use of 'returned'");
6904 // Before passing 'returned' to the target lowering code, ensure that
6905 // either the register MVT and the actual EVT are the same size or that
6906 // the return value and argument are extended in the same way; in these
6907 // cases it's safe to pass the argument register value unchanged as the
6908 // return register value (although it's at the target's option whether
6910 // TODO: allow code generation to take advantage of partially preserved
6911 // registers rather than clobbering the entire register when the
6912 // parameter extension method is not compatible with the return
6914 if ((NumParts * PartVT.getSizeInBits() == VT.getSizeInBits()) ||
6915 (ExtendKind != ISD::ANY_EXTEND &&
6916 CLI.RetSExt == Args[i].isSExt && CLI.RetZExt == Args[i].isZExt))
6917 Flags.setReturned();
6920 getCopyToParts(CLI.DAG, CLI.DL, Op, &Parts[0], NumParts, PartVT,
6921 CLI.CS ? CLI.CS->getInstruction() : nullptr, ExtendKind);
6923 for (unsigned j = 0; j != NumParts; ++j) {
6924 // if it isn't first piece, alignment must be 1
6925 ISD::OutputArg MyFlags(Flags, Parts[j].getValueType(), VT,
6926 i < CLI.NumFixedArgs,
6927 i, j*Parts[j].getValueType().getStoreSize());
6928 if (NumParts > 1 && j == 0)
6929 MyFlags.Flags.setSplit();
6931 MyFlags.Flags.setOrigAlign(1);
6933 CLI.Outs.push_back(MyFlags);
6934 CLI.OutVals.push_back(Parts[j]);
6937 if (NeedsRegBlock && Value == NumValues - 1)
6938 CLI.Outs[CLI.Outs.size() - 1].Flags.setInConsecutiveRegsLast();
6942 SmallVector<SDValue, 4> InVals;
6943 CLI.Chain = LowerCall(CLI, InVals);
6945 // Verify that the target's LowerCall behaved as expected.
6946 assert(CLI.Chain.getNode() && CLI.Chain.getValueType() == MVT::Other &&
6947 "LowerCall didn't return a valid chain!");
6948 assert((!CLI.IsTailCall || InVals.empty()) &&
6949 "LowerCall emitted a return value for a tail call!");
6950 assert((CLI.IsTailCall || InVals.size() == CLI.Ins.size()) &&
6951 "LowerCall didn't emit the correct number of values!");
6953 // For a tail call, the return value is merely live-out and there aren't
6954 // any nodes in the DAG representing it. Return a special value to
6955 // indicate that a tail call has been emitted and no more Instructions
6956 // should be processed in the current block.
6957 if (CLI.IsTailCall) {
6958 CLI.DAG.setRoot(CLI.Chain);
6959 return std::make_pair(SDValue(), SDValue());
6962 DEBUG(for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++i) {
6963 assert(InVals[i].getNode() &&
6964 "LowerCall emitted a null value!");
6965 assert(EVT(CLI.Ins[i].VT) == InVals[i].getValueType() &&
6966 "LowerCall emitted a value with the wrong type!");
6969 SmallVector<SDValue, 4> ReturnValues;
6970 if (!CanLowerReturn) {
6971 // The instruction result is the result of loading from the
6972 // hidden sret parameter.
6973 SmallVector<EVT, 1> PVTs;
6974 Type *PtrRetTy = PointerType::getUnqual(OrigRetTy);
6976 ComputeValueVTs(*this, DL, PtrRetTy, PVTs);
6977 assert(PVTs.size() == 1 && "Pointers should fit in one register");
6978 EVT PtrVT = PVTs[0];
6980 unsigned NumValues = RetTys.size();
6981 ReturnValues.resize(NumValues);
6982 SmallVector<SDValue, 4> Chains(NumValues);
6984 for (unsigned i = 0; i < NumValues; ++i) {
6985 SDValue Add = CLI.DAG.getNode(ISD::ADD, CLI.DL, PtrVT, DemoteStackSlot,
6986 CLI.DAG.getConstant(Offsets[i], CLI.DL,
6988 SDValue L = CLI.DAG.getLoad(
6989 RetTys[i], CLI.DL, CLI.Chain, Add,
6990 MachinePointerInfo::getFixedStack(DemoteStackIdx, Offsets[i]), false,
6992 ReturnValues[i] = L;
6993 Chains[i] = L.getValue(1);
6996 CLI.Chain = CLI.DAG.getNode(ISD::TokenFactor, CLI.DL, MVT::Other, Chains);
6998 // Collect the legal value parts into potentially illegal values
6999 // that correspond to the original function's return values.
7000 ISD::NodeType AssertOp = ISD::DELETED_NODE;
7002 AssertOp = ISD::AssertSext;
7003 else if (CLI.RetZExt)
7004 AssertOp = ISD::AssertZext;
7005 unsigned CurReg = 0;
7006 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
7008 MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
7009 unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
7011 ReturnValues.push_back(getCopyFromParts(CLI.DAG, CLI.DL, &InVals[CurReg],
7012 NumRegs, RegisterVT, VT, nullptr,
7017 // For a function returning void, there is no return value. We can't create
7018 // such a node, so we just return a null return value in that case. In
7019 // that case, nothing will actually look at the value.
7020 if (ReturnValues.empty())
7021 return std::make_pair(SDValue(), CLI.Chain);
7024 SDValue Res = CLI.DAG.getNode(ISD::MERGE_VALUES, CLI.DL,
7025 CLI.DAG.getVTList(RetTys), ReturnValues);
7026 return std::make_pair(Res, CLI.Chain);
7029 void TargetLowering::LowerOperationWrapper(SDNode *N,
7030 SmallVectorImpl<SDValue> &Results,
7031 SelectionDAG &DAG) const {
7032 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
7034 Results.push_back(Res);
7037 SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
7038 llvm_unreachable("LowerOperation not implemented for this target!");
7042 SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) {
7043 SDValue Op = getNonRegisterValue(V);
7044 assert((Op.getOpcode() != ISD::CopyFromReg ||
7045 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
7046 "Copy from a reg to the same reg!");
7047 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg");
7049 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7050 RegsForValue RFV(V->getContext(), TLI, DAG.getDataLayout(), Reg,
7052 SDValue Chain = DAG.getEntryNode();
7054 ISD::NodeType ExtendType = (FuncInfo.PreferredExtendType.find(V) ==
7055 FuncInfo.PreferredExtendType.end())
7057 : FuncInfo.PreferredExtendType[V];
7058 RFV.getCopyToRegs(Op, DAG, getCurSDLoc(), Chain, nullptr, V, ExtendType);
7059 PendingExports.push_back(Chain);
7062 #include "llvm/CodeGen/SelectionDAGISel.h"
7064 /// isOnlyUsedInEntryBlock - If the specified argument is only used in the
7065 /// entry block, return true. This includes arguments used by switches, since
7066 /// the switch may expand into multiple basic blocks.
7067 static bool isOnlyUsedInEntryBlock(const Argument *A, bool FastISel) {
7068 // With FastISel active, we may be splitting blocks, so force creation
7069 // of virtual registers for all non-dead arguments.
7071 return A->use_empty();
7073 const BasicBlock *Entry = A->getParent()->begin();
7074 for (const User *U : A->users())
7075 if (cast<Instruction>(U)->getParent() != Entry || isa<SwitchInst>(U))
7076 return false; // Use not in entry block.
7081 void SelectionDAGISel::LowerArguments(const Function &F) {
7082 SelectionDAG &DAG = SDB->DAG;
7083 SDLoc dl = SDB->getCurSDLoc();
7084 const DataLayout &DL = DAG.getDataLayout();
7085 SmallVector<ISD::InputArg, 16> Ins;
7087 if (!FuncInfo->CanLowerReturn) {
7088 // Put in an sret pointer parameter before all the other parameters.
7089 SmallVector<EVT, 1> ValueVTs;
7090 ComputeValueVTs(*TLI, DAG.getDataLayout(),
7091 PointerType::getUnqual(F.getReturnType()), ValueVTs);
7093 // NOTE: Assuming that a pointer will never break down to more than one VT
7095 ISD::ArgFlagsTy Flags;
7097 MVT RegisterVT = TLI->getRegisterType(*DAG.getContext(), ValueVTs[0]);
7098 ISD::InputArg RetArg(Flags, RegisterVT, ValueVTs[0], true,
7099 ISD::InputArg::NoArgIndex, 0);
7100 Ins.push_back(RetArg);
7103 // Set up the incoming argument description vector.
7105 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end();
7106 I != E; ++I, ++Idx) {
7107 SmallVector<EVT, 4> ValueVTs;
7108 ComputeValueVTs(*TLI, DAG.getDataLayout(), I->getType(), ValueVTs);
7109 bool isArgValueUsed = !I->use_empty();
7110 unsigned PartBase = 0;
7111 Type *FinalType = I->getType();
7112 if (F.getAttributes().hasAttribute(Idx, Attribute::ByVal))
7113 FinalType = cast<PointerType>(FinalType)->getElementType();
7114 bool NeedsRegBlock = TLI->functionArgumentNeedsConsecutiveRegisters(
7115 FinalType, F.getCallingConv(), F.isVarArg());
7116 for (unsigned Value = 0, NumValues = ValueVTs.size();
7117 Value != NumValues; ++Value) {
7118 EVT VT = ValueVTs[Value];
7119 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
7120 ISD::ArgFlagsTy Flags;
7121 unsigned OriginalAlignment = DL.getABITypeAlignment(ArgTy);
7123 if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt))
7125 if (F.getAttributes().hasAttribute(Idx, Attribute::SExt))
7127 if (F.getAttributes().hasAttribute(Idx, Attribute::InReg))
7129 if (F.getAttributes().hasAttribute(Idx, Attribute::StructRet))
7131 if (F.getAttributes().hasAttribute(Idx, Attribute::ByVal))
7133 if (F.getAttributes().hasAttribute(Idx, Attribute::InAlloca)) {
7134 Flags.setInAlloca();
7135 // Set the byval flag for CCAssignFn callbacks that don't know about
7136 // inalloca. This way we can know how many bytes we should've allocated
7137 // and how many bytes a callee cleanup function will pop. If we port
7138 // inalloca to more targets, we'll have to add custom inalloca handling
7139 // in the various CC lowering callbacks.
7142 if (Flags.isByVal() || Flags.isInAlloca()) {
7143 PointerType *Ty = cast<PointerType>(I->getType());
7144 Type *ElementTy = Ty->getElementType();
7145 Flags.setByValSize(DL.getTypeAllocSize(ElementTy));
7146 // For ByVal, alignment should be passed from FE. BE will guess if
7147 // this info is not there but there are cases it cannot get right.
7148 unsigned FrameAlign;
7149 if (F.getParamAlignment(Idx))
7150 FrameAlign = F.getParamAlignment(Idx);
7152 FrameAlign = TLI->getByValTypeAlignment(ElementTy);
7153 Flags.setByValAlign(FrameAlign);
7155 if (F.getAttributes().hasAttribute(Idx, Attribute::Nest))
7158 Flags.setInConsecutiveRegs();
7159 Flags.setOrigAlign(OriginalAlignment);
7161 MVT RegisterVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7162 unsigned NumRegs = TLI->getNumRegisters(*CurDAG->getContext(), VT);
7163 for (unsigned i = 0; i != NumRegs; ++i) {
7164 ISD::InputArg MyFlags(Flags, RegisterVT, VT, isArgValueUsed,
7165 Idx-1, PartBase+i*RegisterVT.getStoreSize());
7166 if (NumRegs > 1 && i == 0)
7167 MyFlags.Flags.setSplit();
7168 // if it isn't first piece, alignment must be 1
7170 MyFlags.Flags.setOrigAlign(1);
7171 Ins.push_back(MyFlags);
7173 if (NeedsRegBlock && Value == NumValues - 1)
7174 Ins[Ins.size() - 1].Flags.setInConsecutiveRegsLast();
7175 PartBase += VT.getStoreSize();
7179 // Call the target to set up the argument values.
7180 SmallVector<SDValue, 8> InVals;
7181 SDValue NewRoot = TLI->LowerFormalArguments(
7182 DAG.getRoot(), F.getCallingConv(), F.isVarArg(), Ins, dl, DAG, InVals);
7184 // Verify that the target's LowerFormalArguments behaved as expected.
7185 assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other &&
7186 "LowerFormalArguments didn't return a valid chain!");
7187 assert(InVals.size() == Ins.size() &&
7188 "LowerFormalArguments didn't emit the correct number of values!");
7190 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
7191 assert(InVals[i].getNode() &&
7192 "LowerFormalArguments emitted a null value!");
7193 assert(EVT(Ins[i].VT) == InVals[i].getValueType() &&
7194 "LowerFormalArguments emitted a value with the wrong type!");
7198 // Update the DAG with the new chain value resulting from argument lowering.
7199 DAG.setRoot(NewRoot);
7201 // Set up the argument values.
7204 if (!FuncInfo->CanLowerReturn) {
7205 // Create a virtual register for the sret pointer, and put in a copy
7206 // from the sret argument into it.
7207 SmallVector<EVT, 1> ValueVTs;
7208 ComputeValueVTs(*TLI, DAG.getDataLayout(),
7209 PointerType::getUnqual(F.getReturnType()), ValueVTs);
7210 MVT VT = ValueVTs[0].getSimpleVT();
7211 MVT RegVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7212 ISD::NodeType AssertOp = ISD::DELETED_NODE;
7213 SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1,
7214 RegVT, VT, nullptr, AssertOp);
7216 MachineFunction& MF = SDB->DAG.getMachineFunction();
7217 MachineRegisterInfo& RegInfo = MF.getRegInfo();
7218 unsigned SRetReg = RegInfo.createVirtualRegister(TLI->getRegClassFor(RegVT));
7219 FuncInfo->DemoteRegister = SRetReg;
7221 SDB->DAG.getCopyToReg(NewRoot, SDB->getCurSDLoc(), SRetReg, ArgValue);
7222 DAG.setRoot(NewRoot);
7224 // i indexes lowered arguments. Bump it past the hidden sret argument.
7225 // Idx indexes LLVM arguments. Don't touch it.
7229 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
7231 SmallVector<SDValue, 4> ArgValues;
7232 SmallVector<EVT, 4> ValueVTs;
7233 ComputeValueVTs(*TLI, DAG.getDataLayout(), I->getType(), ValueVTs);
7234 unsigned NumValues = ValueVTs.size();
7236 // If this argument is unused then remember its value. It is used to generate
7237 // debugging information.
7238 if (I->use_empty() && NumValues) {
7239 SDB->setUnusedArgValue(I, InVals[i]);
7241 // Also remember any frame index for use in FastISel.
7242 if (FrameIndexSDNode *FI =
7243 dyn_cast<FrameIndexSDNode>(InVals[i].getNode()))
7244 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
7247 for (unsigned Val = 0; Val != NumValues; ++Val) {
7248 EVT VT = ValueVTs[Val];
7249 MVT PartVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7250 unsigned NumParts = TLI->getNumRegisters(*CurDAG->getContext(), VT);
7252 if (!I->use_empty()) {
7253 ISD::NodeType AssertOp = ISD::DELETED_NODE;
7254 if (F.getAttributes().hasAttribute(Idx, Attribute::SExt))
7255 AssertOp = ISD::AssertSext;
7256 else if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt))
7257 AssertOp = ISD::AssertZext;
7259 ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i],
7260 NumParts, PartVT, VT,
7261 nullptr, AssertOp));
7267 // We don't need to do anything else for unused arguments.
7268 if (ArgValues.empty())
7271 // Note down frame index.
7272 if (FrameIndexSDNode *FI =
7273 dyn_cast<FrameIndexSDNode>(ArgValues[0].getNode()))
7274 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
7276 SDValue Res = DAG.getMergeValues(makeArrayRef(ArgValues.data(), NumValues),
7277 SDB->getCurSDLoc());
7279 SDB->setValue(I, Res);
7280 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::BUILD_PAIR) {
7281 if (LoadSDNode *LNode =
7282 dyn_cast<LoadSDNode>(Res.getOperand(0).getNode()))
7283 if (FrameIndexSDNode *FI =
7284 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode()))
7285 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
7288 // If this argument is live outside of the entry block, insert a copy from
7289 // wherever we got it to the vreg that other BB's will reference it as.
7290 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::CopyFromReg) {
7291 // If we can, though, try to skip creating an unnecessary vreg.
7292 // FIXME: This isn't very clean... it would be nice to make this more
7293 // general. It's also subtly incompatible with the hacks FastISel
7295 unsigned Reg = cast<RegisterSDNode>(Res.getOperand(1))->getReg();
7296 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
7297 FuncInfo->ValueMap[I] = Reg;
7301 if (!isOnlyUsedInEntryBlock(I, TM.Options.EnableFastISel)) {
7302 FuncInfo->InitializeRegForValue(I);
7303 SDB->CopyToExportRegsIfNeeded(I);
7307 assert(i == InVals.size() && "Argument register count mismatch!");
7309 // Finally, if the target has anything special to do, allow it to do so.
7310 EmitFunctionEntryCode();
7313 /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
7314 /// ensure constants are generated when needed. Remember the virtual registers
7315 /// that need to be added to the Machine PHI nodes as input. We cannot just
7316 /// directly add them, because expansion might result in multiple MBB's for one
7317 /// BB. As such, the start of the BB might correspond to a different MBB than
7321 SelectionDAGBuilder::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
7322 const TerminatorInst *TI = LLVMBB->getTerminator();
7324 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
7326 // Check PHI nodes in successors that expect a value to be available from this
7328 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
7329 const BasicBlock *SuccBB = TI->getSuccessor(succ);
7330 if (!isa<PHINode>(SuccBB->begin())) continue;
7331 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
7333 // If this terminator has multiple identical successors (common for
7334 // switches), only handle each succ once.
7335 if (!SuccsHandled.insert(SuccMBB).second)
7338 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
7340 // At this point we know that there is a 1-1 correspondence between LLVM PHI
7341 // nodes and Machine PHI nodes, but the incoming operands have not been
7343 for (BasicBlock::const_iterator I = SuccBB->begin();
7344 const PHINode *PN = dyn_cast<PHINode>(I); ++I) {
7345 // Ignore dead phi's.
7346 if (PN->use_empty()) continue;
7349 if (PN->getType()->isEmptyTy())
7353 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
7355 if (const Constant *C = dyn_cast<Constant>(PHIOp)) {
7356 unsigned &RegOut = ConstantsOut[C];
7358 RegOut = FuncInfo.CreateRegs(C->getType());
7359 CopyValueToVirtualRegister(C, RegOut);
7363 DenseMap<const Value *, unsigned>::iterator I =
7364 FuncInfo.ValueMap.find(PHIOp);
7365 if (I != FuncInfo.ValueMap.end())
7368 assert(isa<AllocaInst>(PHIOp) &&
7369 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
7370 "Didn't codegen value into a register!??");
7371 Reg = FuncInfo.CreateRegs(PHIOp->getType());
7372 CopyValueToVirtualRegister(PHIOp, Reg);
7376 // Remember that this register needs to added to the machine PHI node as
7377 // the input for this MBB.
7378 SmallVector<EVT, 4> ValueVTs;
7379 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7380 ComputeValueVTs(TLI, DAG.getDataLayout(), PN->getType(), ValueVTs);
7381 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
7382 EVT VT = ValueVTs[vti];
7383 unsigned NumRegisters = TLI.getNumRegisters(*DAG.getContext(), VT);
7384 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
7385 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
7386 Reg += NumRegisters;
7391 ConstantsOut.clear();
7394 /// Add a successor MBB to ParentMBB< creating a new MachineBB for BB if SuccMBB
7397 SelectionDAGBuilder::StackProtectorDescriptor::
7398 AddSuccessorMBB(const BasicBlock *BB,
7399 MachineBasicBlock *ParentMBB,
7401 MachineBasicBlock *SuccMBB) {
7402 // If SuccBB has not been created yet, create it.
7404 MachineFunction *MF = ParentMBB->getParent();
7405 MachineFunction::iterator BBI = ParentMBB;
7406 SuccMBB = MF->CreateMachineBasicBlock(BB);
7407 MF->insert(++BBI, SuccMBB);
7409 // Add it as a successor of ParentMBB.
7410 ParentMBB->addSuccessor(
7411 SuccMBB, BranchProbabilityInfo::getBranchWeightStackProtector(IsLikely));
7415 MachineBasicBlock *SelectionDAGBuilder::NextBlock(MachineBasicBlock *MBB) {
7416 MachineFunction::iterator I = MBB;
7417 if (++I == FuncInfo.MF->end())
7422 /// During lowering new call nodes can be created (such as memset, etc.).
7423 /// Those will become new roots of the current DAG, but complications arise
7424 /// when they are tail calls. In such cases, the call lowering will update
7425 /// the root, but the builder still needs to know that a tail call has been
7426 /// lowered in order to avoid generating an additional return.
7427 void SelectionDAGBuilder::updateDAGForMaybeTailCall(SDValue MaybeTC) {
7428 // If the node is null, we do have a tail call.
7429 if (MaybeTC.getNode() != nullptr)
7430 DAG.setRoot(MaybeTC);
7435 bool SelectionDAGBuilder::isDense(const CaseClusterVector &Clusters,
7436 unsigned *TotalCases, unsigned First,
7438 assert(Last >= First);
7439 assert(TotalCases[Last] >= TotalCases[First]);
7441 APInt LowCase = Clusters[First].Low->getValue();
7442 APInt HighCase = Clusters[Last].High->getValue();
7443 assert(LowCase.getBitWidth() == HighCase.getBitWidth());
7445 // FIXME: A range of consecutive cases has 100% density, but only requires one
7446 // comparison to lower. We should discriminate against such consecutive ranges
7449 uint64_t Diff = (HighCase - LowCase).getLimitedValue((UINT64_MAX - 1) / 100);
7450 uint64_t Range = Diff + 1;
7453 TotalCases[Last] - (First == 0 ? 0 : TotalCases[First - 1]);
7455 assert(NumCases < UINT64_MAX / 100);
7456 assert(Range >= NumCases);
7458 return NumCases * 100 >= Range * MinJumpTableDensity;
7461 static inline bool areJTsAllowed(const TargetLowering &TLI) {
7462 return TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
7463 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other);
7466 bool SelectionDAGBuilder::buildJumpTable(CaseClusterVector &Clusters,
7467 unsigned First, unsigned Last,
7468 const SwitchInst *SI,
7469 MachineBasicBlock *DefaultMBB,
7470 CaseCluster &JTCluster) {
7471 assert(First <= Last);
7473 uint32_t Weight = 0;
7474 unsigned NumCmps = 0;
7475 std::vector<MachineBasicBlock*> Table;
7476 DenseMap<MachineBasicBlock*, uint32_t> JTWeights;
7477 for (unsigned I = First; I <= Last; ++I) {
7478 assert(Clusters[I].Kind == CC_Range);
7479 Weight += Clusters[I].Weight;
7480 assert(Weight >= Clusters[I].Weight && "Weight overflow!");
7481 APInt Low = Clusters[I].Low->getValue();
7482 APInt High = Clusters[I].High->getValue();
7483 NumCmps += (Low == High) ? 1 : 2;
7485 // Fill the gap between this and the previous cluster.
7486 APInt PreviousHigh = Clusters[I - 1].High->getValue();
7487 assert(PreviousHigh.slt(Low));
7488 uint64_t Gap = (Low - PreviousHigh).getLimitedValue() - 1;
7489 for (uint64_t J = 0; J < Gap; J++)
7490 Table.push_back(DefaultMBB);
7492 uint64_t ClusterSize = (High - Low).getLimitedValue() + 1;
7493 for (uint64_t J = 0; J < ClusterSize; ++J)
7494 Table.push_back(Clusters[I].MBB);
7495 JTWeights[Clusters[I].MBB] += Clusters[I].Weight;
7498 unsigned NumDests = JTWeights.size();
7499 if (isSuitableForBitTests(NumDests, NumCmps,
7500 Clusters[First].Low->getValue(),
7501 Clusters[Last].High->getValue())) {
7502 // Clusters[First..Last] should be lowered as bit tests instead.
7506 // Create the MBB that will load from and jump through the table.
7507 // Note: We create it here, but it's not inserted into the function yet.
7508 MachineFunction *CurMF = FuncInfo.MF;
7509 MachineBasicBlock *JumpTableMBB =
7510 CurMF->CreateMachineBasicBlock(SI->getParent());
7512 // Add successors. Note: use table order for determinism.
7513 SmallPtrSet<MachineBasicBlock *, 8> Done;
7514 for (MachineBasicBlock *Succ : Table) {
7515 if (Done.count(Succ))
7517 addSuccessorWithWeight(JumpTableMBB, Succ, JTWeights[Succ]);
7521 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7522 unsigned JTI = CurMF->getOrCreateJumpTableInfo(TLI.getJumpTableEncoding())
7523 ->createJumpTableIndex(Table);
7525 // Set up the jump table info.
7526 JumpTable JT(-1U, JTI, JumpTableMBB, nullptr);
7527 JumpTableHeader JTH(Clusters[First].Low->getValue(),
7528 Clusters[Last].High->getValue(), SI->getCondition(),
7530 JTCases.emplace_back(std::move(JTH), std::move(JT));
7532 JTCluster = CaseCluster::jumpTable(Clusters[First].Low, Clusters[Last].High,
7533 JTCases.size() - 1, Weight);
7537 void SelectionDAGBuilder::findJumpTables(CaseClusterVector &Clusters,
7538 const SwitchInst *SI,
7539 MachineBasicBlock *DefaultMBB) {
7541 // Clusters must be non-empty, sorted, and only contain Range clusters.
7542 assert(!Clusters.empty());
7543 for (CaseCluster &C : Clusters)
7544 assert(C.Kind == CC_Range);
7545 for (unsigned i = 1, e = Clusters.size(); i < e; ++i)
7546 assert(Clusters[i - 1].High->getValue().slt(Clusters[i].Low->getValue()));
7549 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7550 if (!areJTsAllowed(TLI))
7553 const int64_t N = Clusters.size();
7554 const unsigned MinJumpTableSize = TLI.getMinimumJumpTableEntries();
7556 // TotalCases[i]: Total nbr of cases in Clusters[0..i].
7557 SmallVector<unsigned, 8> TotalCases(N);
7559 for (unsigned i = 0; i < N; ++i) {
7560 APInt Hi = Clusters[i].High->getValue();
7561 APInt Lo = Clusters[i].Low->getValue();
7562 TotalCases[i] = (Hi - Lo).getLimitedValue() + 1;
7564 TotalCases[i] += TotalCases[i - 1];
7567 if (N >= MinJumpTableSize && isDense(Clusters, &TotalCases[0], 0, N - 1)) {
7568 // Cheap case: the whole range might be suitable for jump table.
7569 CaseCluster JTCluster;
7570 if (buildJumpTable(Clusters, 0, N - 1, SI, DefaultMBB, JTCluster)) {
7571 Clusters[0] = JTCluster;
7577 // The algorithm below is not suitable for -O0.
7578 if (TM.getOptLevel() == CodeGenOpt::None)
7581 // Split Clusters into minimum number of dense partitions. The algorithm uses
7582 // the same idea as Kannan & Proebsting "Correction to 'Producing Good Code
7583 // for the Case Statement'" (1994), but builds the MinPartitions array in
7584 // reverse order to make it easier to reconstruct the partitions in ascending
7585 // order. In the choice between two optimal partitionings, it picks the one
7586 // which yields more jump tables.
7588 // MinPartitions[i] is the minimum nbr of partitions of Clusters[i..N-1].
7589 SmallVector<unsigned, 8> MinPartitions(N);
7590 // LastElement[i] is the last element of the partition starting at i.
7591 SmallVector<unsigned, 8> LastElement(N);
7592 // NumTables[i]: nbr of >= MinJumpTableSize partitions from Clusters[i..N-1].
7593 SmallVector<unsigned, 8> NumTables(N);
7595 // Base case: There is only one way to partition Clusters[N-1].
7596 MinPartitions[N - 1] = 1;
7597 LastElement[N - 1] = N - 1;
7598 assert(MinJumpTableSize > 1);
7599 NumTables[N - 1] = 0;
7601 // Note: loop indexes are signed to avoid underflow.
7602 for (int64_t i = N - 2; i >= 0; i--) {
7603 // Find optimal partitioning of Clusters[i..N-1].
7604 // Baseline: Put Clusters[i] into a partition on its own.
7605 MinPartitions[i] = MinPartitions[i + 1] + 1;
7607 NumTables[i] = NumTables[i + 1];
7609 // Search for a solution that results in fewer partitions.
7610 for (int64_t j = N - 1; j > i; j--) {
7611 // Try building a partition from Clusters[i..j].
7612 if (isDense(Clusters, &TotalCases[0], i, j)) {
7613 unsigned NumPartitions = 1 + (j == N - 1 ? 0 : MinPartitions[j + 1]);
7614 bool IsTable = j - i + 1 >= MinJumpTableSize;
7615 unsigned Tables = IsTable + (j == N - 1 ? 0 : NumTables[j + 1]);
7617 // If this j leads to fewer partitions, or same number of partitions
7618 // with more lookup tables, it is a better partitioning.
7619 if (NumPartitions < MinPartitions[i] ||
7620 (NumPartitions == MinPartitions[i] && Tables > NumTables[i])) {
7621 MinPartitions[i] = NumPartitions;
7623 NumTables[i] = Tables;
7629 // Iterate over the partitions, replacing some with jump tables in-place.
7630 unsigned DstIndex = 0;
7631 for (unsigned First = 0, Last; First < N; First = Last + 1) {
7632 Last = LastElement[First];
7633 assert(Last >= First);
7634 assert(DstIndex <= First);
7635 unsigned NumClusters = Last - First + 1;
7637 CaseCluster JTCluster;
7638 if (NumClusters >= MinJumpTableSize &&
7639 buildJumpTable(Clusters, First, Last, SI, DefaultMBB, JTCluster)) {
7640 Clusters[DstIndex++] = JTCluster;
7642 for (unsigned I = First; I <= Last; ++I)
7643 std::memmove(&Clusters[DstIndex++], &Clusters[I], sizeof(Clusters[I]));
7646 Clusters.resize(DstIndex);
7649 bool SelectionDAGBuilder::rangeFitsInWord(const APInt &Low, const APInt &High) {
7650 // FIXME: Using the pointer type doesn't seem ideal.
7651 uint64_t BW = DAG.getDataLayout().getPointerSizeInBits();
7652 uint64_t Range = (High - Low).getLimitedValue(UINT64_MAX - 1) + 1;
7656 bool SelectionDAGBuilder::isSuitableForBitTests(unsigned NumDests,
7659 const APInt &High) {
7660 // FIXME: I don't think NumCmps is the correct metric: a single case and a
7661 // range of cases both require only one branch to lower. Just looking at the
7662 // number of clusters and destinations should be enough to decide whether to
7665 // To lower a range with bit tests, the range must fit the bitwidth of a
7667 if (!rangeFitsInWord(Low, High))
7670 // Decide whether it's profitable to lower this range with bit tests. Each
7671 // destination requires a bit test and branch, and there is an overall range
7672 // check branch. For a small number of clusters, separate comparisons might be
7673 // cheaper, and for many destinations, splitting the range might be better.
7674 return (NumDests == 1 && NumCmps >= 3) ||
7675 (NumDests == 2 && NumCmps >= 5) ||
7676 (NumDests == 3 && NumCmps >= 6);
7679 bool SelectionDAGBuilder::buildBitTests(CaseClusterVector &Clusters,
7680 unsigned First, unsigned Last,
7681 const SwitchInst *SI,
7682 CaseCluster &BTCluster) {
7683 assert(First <= Last);
7687 BitVector Dests(FuncInfo.MF->getNumBlockIDs());
7688 unsigned NumCmps = 0;
7689 for (int64_t I = First; I <= Last; ++I) {
7690 assert(Clusters[I].Kind == CC_Range);
7691 Dests.set(Clusters[I].MBB->getNumber());
7692 NumCmps += (Clusters[I].Low == Clusters[I].High) ? 1 : 2;
7694 unsigned NumDests = Dests.count();
7696 APInt Low = Clusters[First].Low->getValue();
7697 APInt High = Clusters[Last].High->getValue();
7698 assert(Low.slt(High));
7700 if (!isSuitableForBitTests(NumDests, NumCmps, Low, High))
7706 const int BitWidth = DAG.getTargetLoweringInfo()
7707 .getPointerTy(DAG.getDataLayout())
7709 assert(rangeFitsInWord(Low, High) && "Case range must fit in bit mask!");
7711 if (Low.isNonNegative() && High.slt(BitWidth)) {
7712 // Optimize the case where all the case values fit in a
7713 // word without having to subtract minValue. In this case,
7714 // we can optimize away the subtraction.
7715 LowBound = APInt::getNullValue(Low.getBitWidth());
7719 CmpRange = High - Low;
7723 uint32_t TotalWeight = 0;
7724 for (unsigned i = First; i <= Last; ++i) {
7725 // Find the CaseBits for this destination.
7727 for (j = 0; j < CBV.size(); ++j)
7728 if (CBV[j].BB == Clusters[i].MBB)
7730 if (j == CBV.size())
7731 CBV.push_back(CaseBits(0, Clusters[i].MBB, 0, 0));
7732 CaseBits *CB = &CBV[j];
7734 // Update Mask, Bits and ExtraWeight.
7735 uint64_t Lo = (Clusters[i].Low->getValue() - LowBound).getZExtValue();
7736 uint64_t Hi = (Clusters[i].High->getValue() - LowBound).getZExtValue();
7737 assert(Hi >= Lo && Hi < 64 && "Invalid bit case!");
7738 CB->Mask |= (-1ULL >> (63 - (Hi - Lo))) << Lo;
7739 CB->Bits += Hi - Lo + 1;
7740 CB->ExtraWeight += Clusters[i].Weight;
7741 TotalWeight += Clusters[i].Weight;
7742 assert(TotalWeight >= Clusters[i].Weight && "Weight overflow!");
7746 std::sort(CBV.begin(), CBV.end(), [](const CaseBits &a, const CaseBits &b) {
7747 // Sort by weight first, number of bits second.
7748 if (a.ExtraWeight != b.ExtraWeight)
7749 return a.ExtraWeight > b.ExtraWeight;
7750 return a.Bits > b.Bits;
7753 for (auto &CB : CBV) {
7754 MachineBasicBlock *BitTestBB =
7755 FuncInfo.MF->CreateMachineBasicBlock(SI->getParent());
7756 BTI.push_back(BitTestCase(CB.Mask, BitTestBB, CB.BB, CB.ExtraWeight));
7758 BitTestCases.emplace_back(std::move(LowBound), std::move(CmpRange),
7759 SI->getCondition(), -1U, MVT::Other, false, nullptr,
7760 nullptr, std::move(BTI));
7762 BTCluster = CaseCluster::bitTests(Clusters[First].Low, Clusters[Last].High,
7763 BitTestCases.size() - 1, TotalWeight);
7767 void SelectionDAGBuilder::findBitTestClusters(CaseClusterVector &Clusters,
7768 const SwitchInst *SI) {
7769 // Partition Clusters into as few subsets as possible, where each subset has a
7770 // range that fits in a machine word and has <= 3 unique destinations.
7773 // Clusters must be sorted and contain Range or JumpTable clusters.
7774 assert(!Clusters.empty());
7775 assert(Clusters[0].Kind == CC_Range || Clusters[0].Kind == CC_JumpTable);
7776 for (const CaseCluster &C : Clusters)
7777 assert(C.Kind == CC_Range || C.Kind == CC_JumpTable);
7778 for (unsigned i = 1; i < Clusters.size(); ++i)
7779 assert(Clusters[i-1].High->getValue().slt(Clusters[i].Low->getValue()));
7782 // The algorithm below is not suitable for -O0.
7783 if (TM.getOptLevel() == CodeGenOpt::None)
7786 // If target does not have legal shift left, do not emit bit tests at all.
7787 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7788 EVT PTy = TLI.getPointerTy(DAG.getDataLayout());
7789 if (!TLI.isOperationLegal(ISD::SHL, PTy))
7792 int BitWidth = PTy.getSizeInBits();
7793 const int64_t N = Clusters.size();
7795 // MinPartitions[i] is the minimum nbr of partitions of Clusters[i..N-1].
7796 SmallVector<unsigned, 8> MinPartitions(N);
7797 // LastElement[i] is the last element of the partition starting at i.
7798 SmallVector<unsigned, 8> LastElement(N);
7800 // FIXME: This might not be the best algorithm for finding bit test clusters.
7802 // Base case: There is only one way to partition Clusters[N-1].
7803 MinPartitions[N - 1] = 1;
7804 LastElement[N - 1] = N - 1;
7806 // Note: loop indexes are signed to avoid underflow.
7807 for (int64_t i = N - 2; i >= 0; --i) {
7808 // Find optimal partitioning of Clusters[i..N-1].
7809 // Baseline: Put Clusters[i] into a partition on its own.
7810 MinPartitions[i] = MinPartitions[i + 1] + 1;
7813 // Search for a solution that results in fewer partitions.
7814 // Note: the search is limited by BitWidth, reducing time complexity.
7815 for (int64_t j = std::min(N - 1, i + BitWidth - 1); j > i; --j) {
7816 // Try building a partition from Clusters[i..j].
7819 if (!rangeFitsInWord(Clusters[i].Low->getValue(),
7820 Clusters[j].High->getValue()))
7823 // Check nbr of destinations and cluster types.
7824 // FIXME: This works, but doesn't seem very efficient.
7825 bool RangesOnly = true;
7826 BitVector Dests(FuncInfo.MF->getNumBlockIDs());
7827 for (int64_t k = i; k <= j; k++) {
7828 if (Clusters[k].Kind != CC_Range) {
7832 Dests.set(Clusters[k].MBB->getNumber());
7834 if (!RangesOnly || Dests.count() > 3)
7837 // Check if it's a better partition.
7838 unsigned NumPartitions = 1 + (j == N - 1 ? 0 : MinPartitions[j + 1]);
7839 if (NumPartitions < MinPartitions[i]) {
7840 // Found a better partition.
7841 MinPartitions[i] = NumPartitions;
7847 // Iterate over the partitions, replacing with bit-test clusters in-place.
7848 unsigned DstIndex = 0;
7849 for (unsigned First = 0, Last; First < N; First = Last + 1) {
7850 Last = LastElement[First];
7851 assert(First <= Last);
7852 assert(DstIndex <= First);
7854 CaseCluster BitTestCluster;
7855 if (buildBitTests(Clusters, First, Last, SI, BitTestCluster)) {
7856 Clusters[DstIndex++] = BitTestCluster;
7858 size_t NumClusters = Last - First + 1;
7859 std::memmove(&Clusters[DstIndex], &Clusters[First],
7860 sizeof(Clusters[0]) * NumClusters);
7861 DstIndex += NumClusters;
7864 Clusters.resize(DstIndex);
7867 void SelectionDAGBuilder::lowerWorkItem(SwitchWorkListItem W, Value *Cond,
7868 MachineBasicBlock *SwitchMBB,
7869 MachineBasicBlock *DefaultMBB) {
7870 MachineFunction *CurMF = FuncInfo.MF;
7871 MachineBasicBlock *NextMBB = nullptr;
7872 MachineFunction::iterator BBI = W.MBB;
7873 if (++BBI != FuncInfo.MF->end())
7876 unsigned Size = W.LastCluster - W.FirstCluster + 1;
7878 BranchProbabilityInfo *BPI = FuncInfo.BPI;
7880 if (Size == 2 && W.MBB == SwitchMBB) {
7881 // If any two of the cases has the same destination, and if one value
7882 // is the same as the other, but has one bit unset that the other has set,
7883 // use bit manipulation to do two compares at once. For example:
7884 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
7885 // TODO: This could be extended to merge any 2 cases in switches with 3
7887 // TODO: Handle cases where W.CaseBB != SwitchBB.
7888 CaseCluster &Small = *W.FirstCluster;
7889 CaseCluster &Big = *W.LastCluster;
7891 if (Small.Low == Small.High && Big.Low == Big.High &&
7892 Small.MBB == Big.MBB) {
7893 const APInt &SmallValue = Small.Low->getValue();
7894 const APInt &BigValue = Big.Low->getValue();
7896 // Check that there is only one bit different.
7897 APInt CommonBit = BigValue ^ SmallValue;
7898 if (CommonBit.isPowerOf2()) {
7899 SDValue CondLHS = getValue(Cond);
7900 EVT VT = CondLHS.getValueType();
7901 SDLoc DL = getCurSDLoc();
7903 SDValue Or = DAG.getNode(ISD::OR, DL, VT, CondLHS,
7904 DAG.getConstant(CommonBit, DL, VT));
7905 SDValue Cond = DAG.getSetCC(
7906 DL, MVT::i1, Or, DAG.getConstant(BigValue | SmallValue, DL, VT),
7909 // Update successor info.
7910 // Both Small and Big will jump to Small.BB, so we sum up the weights.
7911 addSuccessorWithWeight(SwitchMBB, Small.MBB, Small.Weight + Big.Weight);
7912 addSuccessorWithWeight(
7913 SwitchMBB, DefaultMBB,
7914 // The default destination is the first successor in IR.
7915 BPI ? BPI->getEdgeWeight(SwitchMBB->getBasicBlock(), (unsigned)0)
7918 // Insert the true branch.
7920 DAG.getNode(ISD::BRCOND, DL, MVT::Other, getControlRoot(), Cond,
7921 DAG.getBasicBlock(Small.MBB));
7922 // Insert the false branch.
7923 BrCond = DAG.getNode(ISD::BR, DL, MVT::Other, BrCond,
7924 DAG.getBasicBlock(DefaultMBB));
7926 DAG.setRoot(BrCond);
7932 if (TM.getOptLevel() != CodeGenOpt::None) {
7933 // Order cases by weight so the most likely case will be checked first.
7934 std::sort(W.FirstCluster, W.LastCluster + 1,
7935 [](const CaseCluster &a, const CaseCluster &b) {
7936 return a.Weight > b.Weight;
7939 // Rearrange the case blocks so that the last one falls through if possible
7940 // without without changing the order of weights.
7941 for (CaseClusterIt I = W.LastCluster; I > W.FirstCluster; ) {
7943 if (I->Weight > W.LastCluster->Weight)
7945 if (I->Kind == CC_Range && I->MBB == NextMBB) {
7946 std::swap(*I, *W.LastCluster);
7952 // Compute total weight.
7953 uint32_t UnhandledWeights = 0;
7954 for (CaseClusterIt I = W.FirstCluster; I <= W.LastCluster; ++I) {
7955 UnhandledWeights += I->Weight;
7956 assert(UnhandledWeights >= I->Weight && "Weight overflow!");
7959 MachineBasicBlock *CurMBB = W.MBB;
7960 for (CaseClusterIt I = W.FirstCluster, E = W.LastCluster; I <= E; ++I) {
7961 MachineBasicBlock *Fallthrough;
7962 if (I == W.LastCluster) {
7963 // For the last cluster, fall through to the default destination.
7964 Fallthrough = DefaultMBB;
7966 Fallthrough = CurMF->CreateMachineBasicBlock(CurMBB->getBasicBlock());
7967 CurMF->insert(BBI, Fallthrough);
7968 // Put Cond in a virtual register to make it available from the new blocks.
7969 ExportFromCurrentBlock(Cond);
7973 case CC_JumpTable: {
7974 // FIXME: Optimize away range check based on pivot comparisons.
7975 JumpTableHeader *JTH = &JTCases[I->JTCasesIndex].first;
7976 JumpTable *JT = &JTCases[I->JTCasesIndex].second;
7978 // The jump block hasn't been inserted yet; insert it here.
7979 MachineBasicBlock *JumpMBB = JT->MBB;
7980 CurMF->insert(BBI, JumpMBB);
7981 addSuccessorWithWeight(CurMBB, Fallthrough);
7982 addSuccessorWithWeight(CurMBB, JumpMBB);
7984 // The jump table header will be inserted in our current block, do the
7985 // range check, and fall through to our fallthrough block.
7986 JTH->HeaderBB = CurMBB;
7987 JT->Default = Fallthrough; // FIXME: Move Default to JumpTableHeader.
7989 // If we're in the right place, emit the jump table header right now.
7990 if (CurMBB == SwitchMBB) {
7991 visitJumpTableHeader(*JT, *JTH, SwitchMBB);
7992 JTH->Emitted = true;
7997 // FIXME: Optimize away range check based on pivot comparisons.
7998 BitTestBlock *BTB = &BitTestCases[I->BTCasesIndex];
8000 // The bit test blocks haven't been inserted yet; insert them here.
8001 for (BitTestCase &BTC : BTB->Cases)
8002 CurMF->insert(BBI, BTC.ThisBB);
8004 // Fill in fields of the BitTestBlock.
8005 BTB->Parent = CurMBB;
8006 BTB->Default = Fallthrough;
8008 // If we're in the right place, emit the bit test header header right now.
8009 if (CurMBB ==SwitchMBB) {
8010 visitBitTestHeader(*BTB, SwitchMBB);
8011 BTB->Emitted = true;
8016 const Value *RHS, *LHS, *MHS;
8018 if (I->Low == I->High) {
8019 // Check Cond == I->Low.
8025 // Check I->Low <= Cond <= I->High.
8032 // The false weight is the sum of all unhandled cases.
8033 UnhandledWeights -= I->Weight;
8034 CaseBlock CB(CC, LHS, RHS, MHS, I->MBB, Fallthrough, CurMBB, I->Weight,
8037 if (CurMBB == SwitchMBB)
8038 visitSwitchCase(CB, SwitchMBB);
8040 SwitchCases.push_back(CB);
8045 CurMBB = Fallthrough;
8049 unsigned SelectionDAGBuilder::caseClusterRank(const CaseCluster &CC,
8050 CaseClusterIt First,
8051 CaseClusterIt Last) {
8052 return std::count_if(First, Last + 1, [&](const CaseCluster &X) {
8053 if (X.Weight != CC.Weight)
8054 return X.Weight > CC.Weight;
8056 // Ties are broken by comparing the case value.
8057 return X.Low->getValue().slt(CC.Low->getValue());
8061 void SelectionDAGBuilder::splitWorkItem(SwitchWorkList &WorkList,
8062 const SwitchWorkListItem &W,
8064 MachineBasicBlock *SwitchMBB) {
8065 assert(W.FirstCluster->Low->getValue().slt(W.LastCluster->Low->getValue()) &&
8066 "Clusters not sorted?");
8068 assert(W.LastCluster - W.FirstCluster + 1 >= 2 && "Too small to split!");
8070 // Balance the tree based on branch weights to create a near-optimal (in terms
8071 // of search time given key frequency) binary search tree. See e.g. Kurt
8072 // Mehlhorn "Nearly Optimal Binary Search Trees" (1975).
8073 CaseClusterIt LastLeft = W.FirstCluster;
8074 CaseClusterIt FirstRight = W.LastCluster;
8075 uint32_t LeftWeight = LastLeft->Weight;
8076 uint32_t RightWeight = FirstRight->Weight;
8078 // Move LastLeft and FirstRight towards each other from opposite directions to
8079 // find a partitioning of the clusters which balances the weight on both
8080 // sides. If LeftWeight and RightWeight are equal, alternate which side is
8081 // taken to ensure 0-weight nodes are distributed evenly.
8083 while (LastLeft + 1 < FirstRight) {
8084 if (LeftWeight < RightWeight || (LeftWeight == RightWeight && (I & 1)))
8085 LeftWeight += (++LastLeft)->Weight;
8087 RightWeight += (--FirstRight)->Weight;
8092 // Our binary search tree differs from a typical BST in that ours can have up
8093 // to three values in each leaf. The pivot selection above doesn't take that
8094 // into account, which means the tree might require more nodes and be less
8095 // efficient. We compensate for this here.
8097 unsigned NumLeft = LastLeft - W.FirstCluster + 1;
8098 unsigned NumRight = W.LastCluster - FirstRight + 1;
8100 if (std::min(NumLeft, NumRight) < 3 && std::max(NumLeft, NumRight) > 3) {
8101 // If one side has less than 3 clusters, and the other has more than 3,
8102 // consider taking a cluster from the other side.
8104 if (NumLeft < NumRight) {
8105 // Consider moving the first cluster on the right to the left side.
8106 CaseCluster &CC = *FirstRight;
8107 unsigned RightSideRank = caseClusterRank(CC, FirstRight, W.LastCluster);
8108 unsigned LeftSideRank = caseClusterRank(CC, W.FirstCluster, LastLeft);
8109 if (LeftSideRank <= RightSideRank) {
8110 // Moving the cluster to the left does not demote it.
8116 assert(NumRight < NumLeft);
8117 // Consider moving the last element on the left to the right side.
8118 CaseCluster &CC = *LastLeft;
8119 unsigned LeftSideRank = caseClusterRank(CC, W.FirstCluster, LastLeft);
8120 unsigned RightSideRank = caseClusterRank(CC, FirstRight, W.LastCluster);
8121 if (RightSideRank <= LeftSideRank) {
8122 // Moving the cluster to the right does not demot it.
8132 assert(LastLeft + 1 == FirstRight);
8133 assert(LastLeft >= W.FirstCluster);
8134 assert(FirstRight <= W.LastCluster);
8136 // Use the first element on the right as pivot since we will make less-than
8137 // comparisons against it.
8138 CaseClusterIt PivotCluster = FirstRight;
8139 assert(PivotCluster > W.FirstCluster);
8140 assert(PivotCluster <= W.LastCluster);
8142 CaseClusterIt FirstLeft = W.FirstCluster;
8143 CaseClusterIt LastRight = W.LastCluster;
8145 const ConstantInt *Pivot = PivotCluster->Low;
8147 // New blocks will be inserted immediately after the current one.
8148 MachineFunction::iterator BBI = W.MBB;
8151 // We will branch to the LHS if Value < Pivot. If LHS is a single cluster,
8152 // we can branch to its destination directly if it's squeezed exactly in
8153 // between the known lower bound and Pivot - 1.
8154 MachineBasicBlock *LeftMBB;
8155 if (FirstLeft == LastLeft && FirstLeft->Kind == CC_Range &&
8156 FirstLeft->Low == W.GE &&
8157 (FirstLeft->High->getValue() + 1LL) == Pivot->getValue()) {
8158 LeftMBB = FirstLeft->MBB;
8160 LeftMBB = FuncInfo.MF->CreateMachineBasicBlock(W.MBB->getBasicBlock());
8161 FuncInfo.MF->insert(BBI, LeftMBB);
8162 WorkList.push_back({LeftMBB, FirstLeft, LastLeft, W.GE, Pivot});
8163 // Put Cond in a virtual register to make it available from the new blocks.
8164 ExportFromCurrentBlock(Cond);
8167 // Similarly, we will branch to the RHS if Value >= Pivot. If RHS is a
8168 // single cluster, RHS.Low == Pivot, and we can branch to its destination
8169 // directly if RHS.High equals the current upper bound.
8170 MachineBasicBlock *RightMBB;
8171 if (FirstRight == LastRight && FirstRight->Kind == CC_Range &&
8172 W.LT && (FirstRight->High->getValue() + 1ULL) == W.LT->getValue()) {
8173 RightMBB = FirstRight->MBB;
8175 RightMBB = FuncInfo.MF->CreateMachineBasicBlock(W.MBB->getBasicBlock());
8176 FuncInfo.MF->insert(BBI, RightMBB);
8177 WorkList.push_back({RightMBB, FirstRight, LastRight, Pivot, W.LT});
8178 // Put Cond in a virtual register to make it available from the new blocks.
8179 ExportFromCurrentBlock(Cond);
8182 // Create the CaseBlock record that will be used to lower the branch.
8183 CaseBlock CB(ISD::SETLT, Cond, Pivot, nullptr, LeftMBB, RightMBB, W.MBB,
8184 LeftWeight, RightWeight);
8186 if (W.MBB == SwitchMBB)
8187 visitSwitchCase(CB, SwitchMBB);
8189 SwitchCases.push_back(CB);
8192 void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) {
8193 // Extract cases from the switch.
8194 BranchProbabilityInfo *BPI = FuncInfo.BPI;
8195 CaseClusterVector Clusters;
8196 Clusters.reserve(SI.getNumCases());
8197 for (auto I : SI.cases()) {
8198 MachineBasicBlock *Succ = FuncInfo.MBBMap[I.getCaseSuccessor()];
8199 const ConstantInt *CaseVal = I.getCaseValue();
8201 BPI ? BPI->getEdgeWeight(SI.getParent(), I.getSuccessorIndex()) : 0;
8202 Clusters.push_back(CaseCluster::range(CaseVal, CaseVal, Succ, Weight));
8205 MachineBasicBlock *DefaultMBB = FuncInfo.MBBMap[SI.getDefaultDest()];
8207 // Cluster adjacent cases with the same destination. We do this at all
8208 // optimization levels because it's cheap to do and will make codegen faster
8209 // if there are many clusters.
8210 sortAndRangeify(Clusters);
8212 if (TM.getOptLevel() != CodeGenOpt::None) {
8213 // Replace an unreachable default with the most popular destination.
8214 // FIXME: Exploit unreachable default more aggressively.
8215 bool UnreachableDefault =
8216 isa<UnreachableInst>(SI.getDefaultDest()->getFirstNonPHIOrDbg());
8217 if (UnreachableDefault && !Clusters.empty()) {
8218 DenseMap<const BasicBlock *, unsigned> Popularity;
8219 unsigned MaxPop = 0;
8220 const BasicBlock *MaxBB = nullptr;
8221 for (auto I : SI.cases()) {
8222 const BasicBlock *BB = I.getCaseSuccessor();
8223 if (++Popularity[BB] > MaxPop) {
8224 MaxPop = Popularity[BB];
8229 assert(MaxPop > 0 && MaxBB);
8230 DefaultMBB = FuncInfo.MBBMap[MaxBB];
8232 // Remove cases that were pointing to the destination that is now the
8234 CaseClusterVector New;
8235 New.reserve(Clusters.size());
8236 for (CaseCluster &CC : Clusters) {
8237 if (CC.MBB != DefaultMBB)
8240 Clusters = std::move(New);
8244 // If there is only the default destination, jump there directly.
8245 MachineBasicBlock *SwitchMBB = FuncInfo.MBB;
8246 if (Clusters.empty()) {
8247 SwitchMBB->addSuccessor(DefaultMBB);
8248 if (DefaultMBB != NextBlock(SwitchMBB)) {
8249 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other,
8250 getControlRoot(), DAG.getBasicBlock(DefaultMBB)));
8255 findJumpTables(Clusters, &SI, DefaultMBB);
8256 findBitTestClusters(Clusters, &SI);
8259 dbgs() << "Case clusters: ";
8260 for (const CaseCluster &C : Clusters) {
8261 if (C.Kind == CC_JumpTable) dbgs() << "JT:";
8262 if (C.Kind == CC_BitTests) dbgs() << "BT:";
8264 C.Low->getValue().print(dbgs(), true);
8265 if (C.Low != C.High) {
8267 C.High->getValue().print(dbgs(), true);
8274 assert(!Clusters.empty());
8275 SwitchWorkList WorkList;
8276 CaseClusterIt First = Clusters.begin();
8277 CaseClusterIt Last = Clusters.end() - 1;
8278 WorkList.push_back({SwitchMBB, First, Last, nullptr, nullptr});
8280 while (!WorkList.empty()) {
8281 SwitchWorkListItem W = WorkList.back();
8282 WorkList.pop_back();
8283 unsigned NumClusters = W.LastCluster - W.FirstCluster + 1;
8285 if (NumClusters > 3 && TM.getOptLevel() != CodeGenOpt::None) {
8286 // For optimized builds, lower large range as a balanced binary tree.
8287 splitWorkItem(WorkList, W, SI.getCondition(), SwitchMBB);
8291 lowerWorkItem(W, SI.getCondition(), SwitchMBB, DefaultMBB);