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/Analysis/VectorUtils.h"
26 #include "llvm/CodeGen/FastISel.h"
27 #include "llvm/CodeGen/FunctionLoweringInfo.h"
28 #include "llvm/CodeGen/GCMetadata.h"
29 #include "llvm/CodeGen/GCStrategy.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/SelectionDAG.h"
37 #include "llvm/CodeGen/StackMaps.h"
38 #include "llvm/CodeGen/WinEHFuncInfo.h"
39 #include "llvm/IR/CallingConv.h"
40 #include "llvm/IR/Constants.h"
41 #include "llvm/IR/DataLayout.h"
42 #include "llvm/IR/DebugInfo.h"
43 #include "llvm/IR/DerivedTypes.h"
44 #include "llvm/IR/Function.h"
45 #include "llvm/IR/GlobalVariable.h"
46 #include "llvm/IR/InlineAsm.h"
47 #include "llvm/IR/Instructions.h"
48 #include "llvm/IR/IntrinsicInst.h"
49 #include "llvm/IR/Intrinsics.h"
50 #include "llvm/IR/LLVMContext.h"
51 #include "llvm/IR/Module.h"
52 #include "llvm/IR/Statepoint.h"
53 #include "llvm/MC/MCSymbol.h"
54 #include "llvm/Support/CommandLine.h"
55 #include "llvm/Support/Debug.h"
56 #include "llvm/Support/ErrorHandling.h"
57 #include "llvm/Support/MathExtras.h"
58 #include "llvm/Support/raw_ostream.h"
59 #include "llvm/Target/TargetFrameLowering.h"
60 #include "llvm/Target/TargetInstrInfo.h"
61 #include "llvm/Target/TargetIntrinsicInfo.h"
62 #include "llvm/Target/TargetLowering.h"
63 #include "llvm/Target/TargetOptions.h"
64 #include "llvm/Target/TargetSelectionDAGInfo.h"
65 #include "llvm/Target/TargetSubtargetInfo.h"
70 #define DEBUG_TYPE "isel"
72 /// LimitFloatPrecision - Generate low-precision inline sequences for
73 /// some float libcalls (6, 8 or 12 bits).
74 static unsigned LimitFloatPrecision;
76 static cl::opt<unsigned, true>
77 LimitFPPrecision("limit-float-precision",
78 cl::desc("Generate low-precision inline sequences "
79 "for some float libcalls"),
80 cl::location(LimitFloatPrecision),
84 EnableFMFInDAG("enable-fmf-dag", cl::init(true), cl::Hidden,
85 cl::desc("Enable fast-math-flags for DAG nodes"));
87 // Limit the width of DAG chains. This is important in general to prevent
88 // DAG-based analysis from blowing up. For example, alias analysis and
89 // load clustering may not complete in reasonable time. It is difficult to
90 // recognize and avoid this situation within each individual analysis, and
91 // future analyses are likely to have the same behavior. Limiting DAG width is
92 // the safe approach and will be especially important with global DAGs.
94 // MaxParallelChains default is arbitrarily high to avoid affecting
95 // optimization, but could be lowered to improve compile time. Any ld-ld-st-st
96 // sequence over this should have been converted to llvm.memcpy by the
97 // frontend. It easy to induce this behavior with .ll code such as:
98 // %buffer = alloca [4096 x i8]
99 // %data = load [4096 x i8]* %argPtr
100 // store [4096 x i8] %data, [4096 x i8]* %buffer
101 static const unsigned MaxParallelChains = 64;
103 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, SDLoc DL,
104 const SDValue *Parts, unsigned NumParts,
105 MVT PartVT, EVT ValueVT, const Value *V);
107 /// getCopyFromParts - Create a value that contains the specified legal parts
108 /// combined into the value they represent. If the parts combine to a type
109 /// larger then ValueVT then AssertOp can be used to specify whether the extra
110 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
111 /// (ISD::AssertSext).
112 static SDValue getCopyFromParts(SelectionDAG &DAG, SDLoc DL,
113 const SDValue *Parts,
114 unsigned NumParts, MVT PartVT, EVT ValueVT,
116 ISD::NodeType AssertOp = ISD::DELETED_NODE) {
117 if (ValueVT.isVector())
118 return getCopyFromPartsVector(DAG, DL, Parts, NumParts,
121 assert(NumParts > 0 && "No parts to assemble!");
122 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
123 SDValue Val = Parts[0];
126 // Assemble the value from multiple parts.
127 if (ValueVT.isInteger()) {
128 unsigned PartBits = PartVT.getSizeInBits();
129 unsigned ValueBits = ValueVT.getSizeInBits();
131 // Assemble the power of 2 part.
132 unsigned RoundParts = NumParts & (NumParts - 1) ?
133 1 << Log2_32(NumParts) : NumParts;
134 unsigned RoundBits = PartBits * RoundParts;
135 EVT RoundVT = RoundBits == ValueBits ?
136 ValueVT : EVT::getIntegerVT(*DAG.getContext(), RoundBits);
139 EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2);
141 if (RoundParts > 2) {
142 Lo = getCopyFromParts(DAG, DL, Parts, RoundParts / 2,
144 Hi = getCopyFromParts(DAG, DL, Parts + RoundParts / 2,
145 RoundParts / 2, PartVT, HalfVT, V);
147 Lo = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[0]);
148 Hi = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[1]);
151 if (DAG.getDataLayout().isBigEndian())
154 Val = DAG.getNode(ISD::BUILD_PAIR, DL, RoundVT, Lo, Hi);
156 if (RoundParts < NumParts) {
157 // Assemble the trailing non-power-of-2 part.
158 unsigned OddParts = NumParts - RoundParts;
159 EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits);
160 Hi = getCopyFromParts(DAG, DL,
161 Parts + RoundParts, OddParts, PartVT, OddVT, V);
163 // Combine the round and odd parts.
165 if (DAG.getDataLayout().isBigEndian())
167 EVT TotalVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
168 Hi = DAG.getNode(ISD::ANY_EXTEND, DL, TotalVT, Hi);
170 DAG.getNode(ISD::SHL, DL, TotalVT, Hi,
171 DAG.getConstant(Lo.getValueType().getSizeInBits(), DL,
172 TLI.getPointerTy(DAG.getDataLayout())));
173 Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, TotalVT, Lo);
174 Val = DAG.getNode(ISD::OR, DL, TotalVT, Lo, Hi);
176 } else if (PartVT.isFloatingPoint()) {
177 // FP split into multiple FP parts (for ppcf128)
178 assert(ValueVT == EVT(MVT::ppcf128) && PartVT == MVT::f64 &&
181 Lo = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[0]);
182 Hi = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[1]);
183 if (TLI.hasBigEndianPartOrdering(ValueVT, DAG.getDataLayout()))
185 Val = DAG.getNode(ISD::BUILD_PAIR, DL, ValueVT, Lo, Hi);
187 // FP split into integer parts (soft fp)
188 assert(ValueVT.isFloatingPoint() && PartVT.isInteger() &&
189 !PartVT.isVector() && "Unexpected split");
190 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits());
191 Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT, V);
195 // There is now one part, held in Val. Correct it to match ValueVT.
196 EVT PartEVT = Val.getValueType();
198 if (PartEVT == ValueVT)
201 if (PartEVT.isInteger() && ValueVT.isInteger()) {
202 if (ValueVT.bitsLT(PartEVT)) {
203 // For a truncate, see if we have any information to
204 // indicate whether the truncated bits will always be
205 // zero or sign-extension.
206 if (AssertOp != ISD::DELETED_NODE)
207 Val = DAG.getNode(AssertOp, DL, PartEVT, Val,
208 DAG.getValueType(ValueVT));
209 return DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
211 return DAG.getNode(ISD::ANY_EXTEND, DL, ValueVT, Val);
214 if (PartEVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
215 // FP_ROUND's are always exact here.
216 if (ValueVT.bitsLT(Val.getValueType()))
218 ISD::FP_ROUND, DL, ValueVT, Val,
219 DAG.getTargetConstant(1, DL, TLI.getPointerTy(DAG.getDataLayout())));
221 return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val);
224 if (PartEVT.getSizeInBits() == ValueVT.getSizeInBits())
225 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
227 llvm_unreachable("Unknown mismatch!");
230 static void diagnosePossiblyInvalidConstraint(LLVMContext &Ctx, const Value *V,
231 const Twine &ErrMsg) {
232 const Instruction *I = dyn_cast_or_null<Instruction>(V);
234 return Ctx.emitError(ErrMsg);
236 const char *AsmError = ", possible invalid constraint for vector type";
237 if (const CallInst *CI = dyn_cast<CallInst>(I))
238 if (isa<InlineAsm>(CI->getCalledValue()))
239 return Ctx.emitError(I, ErrMsg + AsmError);
241 return Ctx.emitError(I, ErrMsg);
244 /// getCopyFromPartsVector - Create a value that contains the specified legal
245 /// parts combined into the value they represent. If the parts combine to a
246 /// type larger then ValueVT then AssertOp can be used to specify whether the
247 /// extra bits are known to be zero (ISD::AssertZext) or sign extended from
248 /// ValueVT (ISD::AssertSext).
249 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, SDLoc DL,
250 const SDValue *Parts, unsigned NumParts,
251 MVT PartVT, EVT ValueVT, const Value *V) {
252 assert(ValueVT.isVector() && "Not a vector value");
253 assert(NumParts > 0 && "No parts to assemble!");
254 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
255 SDValue Val = Parts[0];
257 // Handle a multi-element vector.
261 unsigned NumIntermediates;
263 TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT,
264 NumIntermediates, RegisterVT);
265 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
266 NumParts = NumRegs; // Silence a compiler warning.
267 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
268 assert(RegisterVT.getSizeInBits() ==
269 Parts[0].getSimpleValueType().getSizeInBits() &&
270 "Part type sizes don't match!");
272 // Assemble the parts into intermediate operands.
273 SmallVector<SDValue, 8> Ops(NumIntermediates);
274 if (NumIntermediates == NumParts) {
275 // If the register was not expanded, truncate or copy the value,
277 for (unsigned i = 0; i != NumParts; ++i)
278 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i], 1,
279 PartVT, IntermediateVT, V);
280 } else if (NumParts > 0) {
281 // If the intermediate type was expanded, build the intermediate
282 // operands from the parts.
283 assert(NumParts % NumIntermediates == 0 &&
284 "Must expand into a divisible number of parts!");
285 unsigned Factor = NumParts / NumIntermediates;
286 for (unsigned i = 0; i != NumIntermediates; ++i)
287 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor,
288 PartVT, IntermediateVT, V);
291 // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the
292 // intermediate operands.
293 Val = DAG.getNode(IntermediateVT.isVector() ? ISD::CONCAT_VECTORS
298 // There is now one part, held in Val. Correct it to match ValueVT.
299 EVT PartEVT = Val.getValueType();
301 if (PartEVT == ValueVT)
304 if (PartEVT.isVector()) {
305 // If the element type of the source/dest vectors are the same, but the
306 // parts vector has more elements than the value vector, then we have a
307 // vector widening case (e.g. <2 x float> -> <4 x float>). Extract the
309 if (PartEVT.getVectorElementType() == ValueVT.getVectorElementType()) {
310 assert(PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements() &&
311 "Cannot narrow, it would be a lossy transformation");
313 ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val,
314 DAG.getConstant(0, DL, TLI.getVectorIdxTy(DAG.getDataLayout())));
317 // Vector/Vector bitcast.
318 if (ValueVT.getSizeInBits() == PartEVT.getSizeInBits())
319 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
321 assert(PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements() &&
322 "Cannot handle this kind of promotion");
323 // Promoted vector extract
324 return DAG.getAnyExtOrTrunc(Val, DL, ValueVT);
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 Val = DAG.getAnyExtOrTrunc(Val, DL, ValueVT.getScalarType());
345 return DAG.getNode(ISD::BUILD_VECTOR, DL, ValueVT, Val);
348 static void getCopyToPartsVector(SelectionDAG &DAG, SDLoc dl,
349 SDValue Val, SDValue *Parts, unsigned NumParts,
350 MVT PartVT, const Value *V);
352 /// getCopyToParts - Create a series of nodes that contain the specified value
353 /// split into legal parts. If the parts contain more bits than Val, then, for
354 /// integers, ExtendKind can be used to specify how to generate the extra bits.
355 static void getCopyToParts(SelectionDAG &DAG, SDLoc DL,
356 SDValue Val, SDValue *Parts, unsigned NumParts,
357 MVT PartVT, const Value *V,
358 ISD::NodeType ExtendKind = ISD::ANY_EXTEND) {
359 EVT ValueVT = Val.getValueType();
361 // Handle the vector case separately.
362 if (ValueVT.isVector())
363 return getCopyToPartsVector(DAG, DL, Val, Parts, NumParts, PartVT, V);
365 unsigned PartBits = PartVT.getSizeInBits();
366 unsigned OrigNumParts = NumParts;
367 assert(DAG.getTargetLoweringInfo().isTypeLegal(PartVT) &&
368 "Copying to an illegal type!");
373 assert(!ValueVT.isVector() && "Vector case handled elsewhere");
374 EVT PartEVT = PartVT;
375 if (PartEVT == ValueVT) {
376 assert(NumParts == 1 && "No-op copy with multiple parts!");
381 if (NumParts * PartBits > ValueVT.getSizeInBits()) {
382 // If the parts cover more bits than the value has, promote the value.
383 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
384 assert(NumParts == 1 && "Do not know what to promote to!");
385 Val = DAG.getNode(ISD::FP_EXTEND, DL, PartVT, Val);
387 assert((PartVT.isInteger() || PartVT == MVT::x86mmx) &&
388 ValueVT.isInteger() &&
389 "Unknown mismatch!");
390 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
391 Val = DAG.getNode(ExtendKind, DL, ValueVT, Val);
392 if (PartVT == MVT::x86mmx)
393 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
395 } else if (PartBits == ValueVT.getSizeInBits()) {
396 // Different types of the same size.
397 assert(NumParts == 1 && PartEVT != ValueVT);
398 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
399 } else if (NumParts * PartBits < ValueVT.getSizeInBits()) {
400 // If the parts cover less bits than value has, truncate the value.
401 assert((PartVT.isInteger() || PartVT == MVT::x86mmx) &&
402 ValueVT.isInteger() &&
403 "Unknown mismatch!");
404 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
405 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
406 if (PartVT == MVT::x86mmx)
407 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
410 // The value may have changed - recompute ValueVT.
411 ValueVT = Val.getValueType();
412 assert(NumParts * PartBits == ValueVT.getSizeInBits() &&
413 "Failed to tile the value with PartVT!");
416 if (PartEVT != ValueVT)
417 diagnosePossiblyInvalidConstraint(*DAG.getContext(), V,
418 "scalar-to-vector conversion failed");
424 // Expand the value into multiple parts.
425 if (NumParts & (NumParts - 1)) {
426 // The number of parts is not a power of 2. Split off and copy the tail.
427 assert(PartVT.isInteger() && ValueVT.isInteger() &&
428 "Do not know what to expand to!");
429 unsigned RoundParts = 1 << Log2_32(NumParts);
430 unsigned RoundBits = RoundParts * PartBits;
431 unsigned OddParts = NumParts - RoundParts;
432 SDValue OddVal = DAG.getNode(ISD::SRL, DL, ValueVT, Val,
433 DAG.getIntPtrConstant(RoundBits, DL));
434 getCopyToParts(DAG, DL, OddVal, Parts + RoundParts, OddParts, PartVT, V);
436 if (DAG.getDataLayout().isBigEndian())
437 // The odd parts were reversed by getCopyToParts - unreverse them.
438 std::reverse(Parts + RoundParts, Parts + NumParts);
440 NumParts = RoundParts;
441 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
442 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
445 // The number of parts is a power of 2. Repeatedly bisect the value using
447 Parts[0] = DAG.getNode(ISD::BITCAST, DL,
448 EVT::getIntegerVT(*DAG.getContext(),
449 ValueVT.getSizeInBits()),
452 for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) {
453 for (unsigned i = 0; i < NumParts; i += StepSize) {
454 unsigned ThisBits = StepSize * PartBits / 2;
455 EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits);
456 SDValue &Part0 = Parts[i];
457 SDValue &Part1 = Parts[i+StepSize/2];
459 Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
460 ThisVT, Part0, DAG.getIntPtrConstant(1, DL));
461 Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
462 ThisVT, Part0, DAG.getIntPtrConstant(0, DL));
464 if (ThisBits == PartBits && ThisVT != PartVT) {
465 Part0 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part0);
466 Part1 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part1);
471 if (DAG.getDataLayout().isBigEndian())
472 std::reverse(Parts, Parts + OrigNumParts);
476 /// getCopyToPartsVector - Create a series of nodes that contain the specified
477 /// value split into legal parts.
478 static void getCopyToPartsVector(SelectionDAG &DAG, SDLoc DL,
479 SDValue Val, SDValue *Parts, unsigned NumParts,
480 MVT PartVT, const Value *V) {
481 EVT ValueVT = Val.getValueType();
482 assert(ValueVT.isVector() && "Not a vector");
483 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
486 EVT PartEVT = PartVT;
487 if (PartEVT == ValueVT) {
489 } else if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) {
490 // Bitconvert vector->vector case.
491 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
492 } else if (PartVT.isVector() &&
493 PartEVT.getVectorElementType() == ValueVT.getVectorElementType() &&
494 PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements()) {
495 EVT ElementVT = PartVT.getVectorElementType();
496 // Vector widening case, e.g. <2 x float> -> <4 x float>. Shuffle in
498 SmallVector<SDValue, 16> Ops;
499 for (unsigned i = 0, e = ValueVT.getVectorNumElements(); i != e; ++i)
500 Ops.push_back(DAG.getNode(
501 ISD::EXTRACT_VECTOR_ELT, DL, ElementVT, Val,
502 DAG.getConstant(i, DL, TLI.getVectorIdxTy(DAG.getDataLayout()))));
504 for (unsigned i = ValueVT.getVectorNumElements(),
505 e = PartVT.getVectorNumElements(); i != e; ++i)
506 Ops.push_back(DAG.getUNDEF(ElementVT));
508 Val = DAG.getNode(ISD::BUILD_VECTOR, DL, PartVT, Ops);
510 // FIXME: Use CONCAT for 2x -> 4x.
512 //SDValue UndefElts = DAG.getUNDEF(VectorTy);
513 //Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts);
514 } else if (PartVT.isVector() &&
515 PartEVT.getVectorElementType().bitsGE(
516 ValueVT.getVectorElementType()) &&
517 PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements()) {
519 // Promoted vector extract
520 Val = DAG.getAnyExtOrTrunc(Val, DL, PartVT);
522 // Vector -> scalar conversion.
523 assert(ValueVT.getVectorNumElements() == 1 &&
524 "Only trivial vector-to-scalar conversions should get here!");
526 ISD::EXTRACT_VECTOR_ELT, DL, PartVT, Val,
527 DAG.getConstant(0, DL, TLI.getVectorIdxTy(DAG.getDataLayout())));
529 Val = DAG.getAnyExtOrTrunc(Val, DL, PartVT);
536 // Handle a multi-element vector.
539 unsigned NumIntermediates;
540 unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT,
542 NumIntermediates, RegisterVT);
543 unsigned NumElements = ValueVT.getVectorNumElements();
545 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
546 NumParts = NumRegs; // Silence a compiler warning.
547 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
549 // Split the vector into intermediate operands.
550 SmallVector<SDValue, 8> Ops(NumIntermediates);
551 for (unsigned i = 0; i != NumIntermediates; ++i) {
552 if (IntermediateVT.isVector())
554 DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, IntermediateVT, Val,
555 DAG.getConstant(i * (NumElements / NumIntermediates), DL,
556 TLI.getVectorIdxTy(DAG.getDataLayout())));
558 Ops[i] = DAG.getNode(
559 ISD::EXTRACT_VECTOR_ELT, DL, IntermediateVT, Val,
560 DAG.getConstant(i, DL, TLI.getVectorIdxTy(DAG.getDataLayout())));
563 // Split the intermediate operands into legal parts.
564 if (NumParts == NumIntermediates) {
565 // If the register was not expanded, promote or copy the value,
567 for (unsigned i = 0; i != NumParts; ++i)
568 getCopyToParts(DAG, DL, Ops[i], &Parts[i], 1, PartVT, V);
569 } else if (NumParts > 0) {
570 // If the intermediate type was expanded, split each the value into
572 assert(NumIntermediates != 0 && "division by zero");
573 assert(NumParts % NumIntermediates == 0 &&
574 "Must expand into a divisible number of parts!");
575 unsigned Factor = NumParts / NumIntermediates;
576 for (unsigned i = 0; i != NumIntermediates; ++i)
577 getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT, V);
581 RegsForValue::RegsForValue() {}
583 RegsForValue::RegsForValue(const SmallVector<unsigned, 4> ®s, MVT regvt,
585 : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {}
587 RegsForValue::RegsForValue(LLVMContext &Context, const TargetLowering &TLI,
588 const DataLayout &DL, unsigned Reg, Type *Ty) {
589 ComputeValueVTs(TLI, DL, Ty, ValueVTs);
591 for (EVT ValueVT : ValueVTs) {
592 unsigned NumRegs = TLI.getNumRegisters(Context, ValueVT);
593 MVT RegisterVT = TLI.getRegisterType(Context, ValueVT);
594 for (unsigned i = 0; i != NumRegs; ++i)
595 Regs.push_back(Reg + i);
596 RegVTs.push_back(RegisterVT);
601 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
602 /// this value and returns the result as a ValueVT value. This uses
603 /// Chain/Flag as the input and updates them for the output Chain/Flag.
604 /// If the Flag pointer is NULL, no flag is used.
605 SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
606 FunctionLoweringInfo &FuncInfo,
608 SDValue &Chain, SDValue *Flag,
609 const Value *V) const {
610 // A Value with type {} or [0 x %t] needs no registers.
611 if (ValueVTs.empty())
614 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
616 // Assemble the legal parts into the final values.
617 SmallVector<SDValue, 4> Values(ValueVTs.size());
618 SmallVector<SDValue, 8> Parts;
619 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
620 // Copy the legal parts from the registers.
621 EVT ValueVT = ValueVTs[Value];
622 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
623 MVT RegisterVT = RegVTs[Value];
625 Parts.resize(NumRegs);
626 for (unsigned i = 0; i != NumRegs; ++i) {
629 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT);
631 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag);
632 *Flag = P.getValue(2);
635 Chain = P.getValue(1);
638 // If the source register was virtual and if we know something about it,
639 // add an assert node.
640 if (!TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) ||
641 !RegisterVT.isInteger() || RegisterVT.isVector())
644 const FunctionLoweringInfo::LiveOutInfo *LOI =
645 FuncInfo.GetLiveOutRegInfo(Regs[Part+i]);
649 unsigned RegSize = RegisterVT.getSizeInBits();
650 unsigned NumSignBits = LOI->NumSignBits;
651 unsigned NumZeroBits = LOI->KnownZero.countLeadingOnes();
653 if (NumZeroBits == RegSize) {
654 // The current value is a zero.
655 // Explicitly express that as it would be easier for
656 // optimizations to kick in.
657 Parts[i] = DAG.getConstant(0, dl, RegisterVT);
661 // FIXME: We capture more information than the dag can represent. For
662 // now, just use the tightest assertzext/assertsext possible.
664 EVT FromVT(MVT::Other);
665 if (NumSignBits == RegSize)
666 isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1
667 else if (NumZeroBits >= RegSize-1)
668 isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1
669 else if (NumSignBits > RegSize-8)
670 isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8
671 else if (NumZeroBits >= RegSize-8)
672 isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8
673 else if (NumSignBits > RegSize-16)
674 isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16
675 else if (NumZeroBits >= RegSize-16)
676 isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16
677 else if (NumSignBits > RegSize-32)
678 isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32
679 else if (NumZeroBits >= RegSize-32)
680 isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32
684 // Add an assertion node.
685 assert(FromVT != MVT::Other);
686 Parts[i] = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl,
687 RegisterVT, P, DAG.getValueType(FromVT));
690 Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(),
691 NumRegs, RegisterVT, ValueVT, V);
696 return DAG.getNode(ISD::MERGE_VALUES, dl, DAG.getVTList(ValueVTs), Values);
699 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
700 /// specified value into the registers specified by this object. This uses
701 /// Chain/Flag as the input and updates them for the output Chain/Flag.
702 /// If the Flag pointer is NULL, no flag is used.
703 void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, SDLoc dl,
704 SDValue &Chain, SDValue *Flag, const Value *V,
705 ISD::NodeType PreferredExtendType) const {
706 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
707 ISD::NodeType ExtendKind = PreferredExtendType;
709 // Get the list of the values's legal parts.
710 unsigned NumRegs = Regs.size();
711 SmallVector<SDValue, 8> Parts(NumRegs);
712 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
713 EVT ValueVT = ValueVTs[Value];
714 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
715 MVT RegisterVT = RegVTs[Value];
717 if (ExtendKind == ISD::ANY_EXTEND && TLI.isZExtFree(Val, RegisterVT))
718 ExtendKind = ISD::ZERO_EXTEND;
720 getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value),
721 &Parts[Part], NumParts, RegisterVT, V, ExtendKind);
725 // Copy the parts into the registers.
726 SmallVector<SDValue, 8> Chains(NumRegs);
727 for (unsigned i = 0; i != NumRegs; ++i) {
730 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]);
732 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag);
733 *Flag = Part.getValue(1);
736 Chains[i] = Part.getValue(0);
739 if (NumRegs == 1 || Flag)
740 // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is
741 // flagged to it. That is the CopyToReg nodes and the user are considered
742 // a single scheduling unit. If we create a TokenFactor and return it as
743 // chain, then the TokenFactor is both a predecessor (operand) of the
744 // user as well as a successor (the TF operands are flagged to the user).
745 // c1, f1 = CopyToReg
746 // c2, f2 = CopyToReg
747 // c3 = TokenFactor c1, c2
750 Chain = Chains[NumRegs-1];
752 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
755 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
756 /// operand list. This adds the code marker and includes the number of
757 /// values added into it.
758 void RegsForValue::AddInlineAsmOperands(unsigned Code, bool HasMatching,
759 unsigned MatchingIdx, SDLoc dl,
761 std::vector<SDValue> &Ops) const {
762 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
764 unsigned Flag = InlineAsm::getFlagWord(Code, Regs.size());
766 Flag = InlineAsm::getFlagWordForMatchingOp(Flag, MatchingIdx);
767 else if (!Regs.empty() &&
768 TargetRegisterInfo::isVirtualRegister(Regs.front())) {
769 // Put the register class of the virtual registers in the flag word. That
770 // way, later passes can recompute register class constraints for inline
771 // assembly as well as normal instructions.
772 // Don't do this for tied operands that can use the regclass information
774 const MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
775 const TargetRegisterClass *RC = MRI.getRegClass(Regs.front());
776 Flag = InlineAsm::getFlagWordForRegClass(Flag, RC->getID());
779 SDValue Res = DAG.getTargetConstant(Flag, dl, MVT::i32);
782 unsigned SP = TLI.getStackPointerRegisterToSaveRestore();
783 for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) {
784 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVTs[Value]);
785 MVT RegisterVT = RegVTs[Value];
786 for (unsigned i = 0; i != NumRegs; ++i) {
787 assert(Reg < Regs.size() && "Mismatch in # registers expected");
788 unsigned TheReg = Regs[Reg++];
789 Ops.push_back(DAG.getRegister(TheReg, RegisterVT));
791 if (TheReg == SP && Code == InlineAsm::Kind_Clobber) {
792 // If we clobbered the stack pointer, MFI should know about it.
793 assert(DAG.getMachineFunction().getFrameInfo()->
794 hasOpaqueSPAdjustment());
800 void SelectionDAGBuilder::init(GCFunctionInfo *gfi, AliasAnalysis &aa,
801 const TargetLibraryInfo *li) {
805 DL = &DAG.getDataLayout();
806 Context = DAG.getContext();
807 LPadToCallSiteMap.clear();
810 /// clear - Clear out the current SelectionDAG and the associated
811 /// state and prepare this SelectionDAGBuilder object to be used
812 /// for a new block. This doesn't clear out information about
813 /// additional blocks that are needed to complete switch lowering
814 /// or PHI node updating; that information is cleared out as it is
816 void SelectionDAGBuilder::clear() {
818 UnusedArgNodeMap.clear();
819 PendingLoads.clear();
820 PendingExports.clear();
823 SDNodeOrder = LowestSDNodeOrder;
824 StatepointLowering.clear();
827 /// clearDanglingDebugInfo - Clear the dangling debug information
828 /// map. This function is separated from the clear so that debug
829 /// information that is dangling in a basic block can be properly
830 /// resolved in a different basic block. This allows the
831 /// SelectionDAG to resolve dangling debug information attached
833 void SelectionDAGBuilder::clearDanglingDebugInfo() {
834 DanglingDebugInfoMap.clear();
837 /// getRoot - Return the current virtual root of the Selection DAG,
838 /// flushing any PendingLoad items. This must be done before emitting
839 /// a store or any other node that may need to be ordered after any
840 /// prior load instructions.
842 SDValue SelectionDAGBuilder::getRoot() {
843 if (PendingLoads.empty())
844 return DAG.getRoot();
846 if (PendingLoads.size() == 1) {
847 SDValue Root = PendingLoads[0];
849 PendingLoads.clear();
853 // Otherwise, we have to make a token factor node.
854 SDValue Root = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
856 PendingLoads.clear();
861 /// getControlRoot - Similar to getRoot, but instead of flushing all the
862 /// PendingLoad items, flush all the PendingExports items. It is necessary
863 /// to do this before emitting a terminator instruction.
865 SDValue SelectionDAGBuilder::getControlRoot() {
866 SDValue Root = DAG.getRoot();
868 if (PendingExports.empty())
871 // Turn all of the CopyToReg chains into one factored node.
872 if (Root.getOpcode() != ISD::EntryToken) {
873 unsigned i = 0, e = PendingExports.size();
874 for (; i != e; ++i) {
875 assert(PendingExports[i].getNode()->getNumOperands() > 1);
876 if (PendingExports[i].getNode()->getOperand(0) == Root)
877 break; // Don't add the root if we already indirectly depend on it.
881 PendingExports.push_back(Root);
884 Root = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
886 PendingExports.clear();
891 void SelectionDAGBuilder::visit(const Instruction &I) {
892 // Set up outgoing PHI node register values before emitting the terminator.
893 if (isa<TerminatorInst>(&I))
894 HandlePHINodesInSuccessorBlocks(I.getParent());
900 visit(I.getOpcode(), I);
902 if (!isa<TerminatorInst>(&I) && !HasTailCall &&
903 !isStatepoint(&I)) // statepoints handle their exports internally
904 CopyToExportRegsIfNeeded(&I);
909 void SelectionDAGBuilder::visitPHI(const PHINode &) {
910 llvm_unreachable("SelectionDAGBuilder shouldn't visit PHI nodes!");
913 void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) {
914 // Note: this doesn't use InstVisitor, because it has to work with
915 // ConstantExpr's in addition to instructions.
917 default: llvm_unreachable("Unknown instruction type encountered!");
918 // Build the switch statement using the Instruction.def file.
919 #define HANDLE_INST(NUM, OPCODE, CLASS) \
920 case Instruction::OPCODE: visit##OPCODE((const CLASS&)I); break;
921 #include "llvm/IR/Instruction.def"
925 // resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V,
926 // generate the debug data structures now that we've seen its definition.
927 void SelectionDAGBuilder::resolveDanglingDebugInfo(const Value *V,
929 DanglingDebugInfo &DDI = DanglingDebugInfoMap[V];
931 const DbgValueInst *DI = DDI.getDI();
932 DebugLoc dl = DDI.getdl();
933 unsigned DbgSDNodeOrder = DDI.getSDNodeOrder();
934 DILocalVariable *Variable = DI->getVariable();
935 DIExpression *Expr = DI->getExpression();
936 assert(Variable->isValidLocationForIntrinsic(dl) &&
937 "Expected inlined-at fields to agree");
938 uint64_t Offset = DI->getOffset();
939 // A dbg.value for an alloca is always indirect.
940 bool IsIndirect = isa<AllocaInst>(V) || Offset != 0;
943 if (!EmitFuncArgumentDbgValue(V, Variable, Expr, dl, Offset, IsIndirect,
945 SDV = DAG.getDbgValue(Variable, Expr, Val.getNode(), Val.getResNo(),
946 IsIndirect, Offset, dl, DbgSDNodeOrder);
947 DAG.AddDbgValue(SDV, Val.getNode(), false);
950 DEBUG(dbgs() << "Dropping debug info for " << *DI << "\n");
951 DanglingDebugInfoMap[V] = DanglingDebugInfo();
955 /// getCopyFromRegs - If there was virtual register allocated for the value V
956 /// emit CopyFromReg of the specified type Ty. Return empty SDValue() otherwise.
957 SDValue SelectionDAGBuilder::getCopyFromRegs(const Value *V, Type *Ty) {
958 DenseMap<const Value *, unsigned>::iterator It = FuncInfo.ValueMap.find(V);
961 if (It != FuncInfo.ValueMap.end()) {
962 unsigned InReg = It->second;
963 RegsForValue RFV(*DAG.getContext(), DAG.getTargetLoweringInfo(),
964 DAG.getDataLayout(), InReg, Ty);
965 SDValue Chain = DAG.getEntryNode();
966 Result = RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, nullptr, V);
967 resolveDanglingDebugInfo(V, Result);
973 /// getValue - Return an SDValue for the given Value.
974 SDValue SelectionDAGBuilder::getValue(const Value *V) {
975 // If we already have an SDValue for this value, use it. It's important
976 // to do this first, so that we don't create a CopyFromReg if we already
977 // have a regular SDValue.
978 SDValue &N = NodeMap[V];
979 if (N.getNode()) return N;
981 // If there's a virtual register allocated and initialized for this
983 SDValue copyFromReg = getCopyFromRegs(V, V->getType());
984 if (copyFromReg.getNode()) {
988 // Otherwise create a new SDValue and remember it.
989 SDValue Val = getValueImpl(V);
991 resolveDanglingDebugInfo(V, Val);
995 // Return true if SDValue exists for the given Value
996 bool SelectionDAGBuilder::findValue(const Value *V) const {
997 return (NodeMap.find(V) != NodeMap.end()) ||
998 (FuncInfo.ValueMap.find(V) != FuncInfo.ValueMap.end());
1001 /// getNonRegisterValue - Return an SDValue for the given Value, but
1002 /// don't look in FuncInfo.ValueMap for a virtual register.
1003 SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) {
1004 // If we already have an SDValue for this value, use it.
1005 SDValue &N = NodeMap[V];
1007 if (isa<ConstantSDNode>(N) || isa<ConstantFPSDNode>(N)) {
1008 // Remove the debug location from the node as the node is about to be used
1009 // in a location which may differ from the original debug location. This
1010 // is relevant to Constant and ConstantFP nodes because they can appear
1011 // as constant expressions inside PHI nodes.
1012 N->setDebugLoc(DebugLoc());
1017 // Otherwise create a new SDValue and remember it.
1018 SDValue Val = getValueImpl(V);
1020 resolveDanglingDebugInfo(V, Val);
1024 /// getValueImpl - Helper function for getValue and getNonRegisterValue.
1025 /// Create an SDValue for the given value.
1026 SDValue SelectionDAGBuilder::getValueImpl(const Value *V) {
1027 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1029 if (const Constant *C = dyn_cast<Constant>(V)) {
1030 EVT VT = TLI.getValueType(DAG.getDataLayout(), V->getType(), true);
1032 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C))
1033 return DAG.getConstant(*CI, getCurSDLoc(), VT);
1035 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
1036 return DAG.getGlobalAddress(GV, getCurSDLoc(), VT);
1038 if (isa<ConstantPointerNull>(C)) {
1039 unsigned AS = V->getType()->getPointerAddressSpace();
1040 return DAG.getConstant(0, getCurSDLoc(),
1041 TLI.getPointerTy(DAG.getDataLayout(), AS));
1044 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
1045 return DAG.getConstantFP(*CFP, getCurSDLoc(), VT);
1047 if (isa<UndefValue>(C) && !V->getType()->isAggregateType())
1048 return DAG.getUNDEF(VT);
1050 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
1051 visit(CE->getOpcode(), *CE);
1052 SDValue N1 = NodeMap[V];
1053 assert(N1.getNode() && "visit didn't populate the NodeMap!");
1057 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) {
1058 SmallVector<SDValue, 4> Constants;
1059 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end();
1061 SDNode *Val = getValue(*OI).getNode();
1062 // If the operand is an empty aggregate, there are no values.
1064 // Add each leaf value from the operand to the Constants list
1065 // to form a flattened list of all the values.
1066 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
1067 Constants.push_back(SDValue(Val, i));
1070 return DAG.getMergeValues(Constants, getCurSDLoc());
1073 if (const ConstantDataSequential *CDS =
1074 dyn_cast<ConstantDataSequential>(C)) {
1075 SmallVector<SDValue, 4> Ops;
1076 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1077 SDNode *Val = getValue(CDS->getElementAsConstant(i)).getNode();
1078 // Add each leaf value from the operand to the Constants list
1079 // to form a flattened list of all the values.
1080 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
1081 Ops.push_back(SDValue(Val, i));
1084 if (isa<ArrayType>(CDS->getType()))
1085 return DAG.getMergeValues(Ops, getCurSDLoc());
1086 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(),
1090 if (C->getType()->isStructTy() || C->getType()->isArrayTy()) {
1091 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
1092 "Unknown struct or array constant!");
1094 SmallVector<EVT, 4> ValueVTs;
1095 ComputeValueVTs(TLI, DAG.getDataLayout(), C->getType(), ValueVTs);
1096 unsigned NumElts = ValueVTs.size();
1098 return SDValue(); // empty struct
1099 SmallVector<SDValue, 4> Constants(NumElts);
1100 for (unsigned i = 0; i != NumElts; ++i) {
1101 EVT EltVT = ValueVTs[i];
1102 if (isa<UndefValue>(C))
1103 Constants[i] = DAG.getUNDEF(EltVT);
1104 else if (EltVT.isFloatingPoint())
1105 Constants[i] = DAG.getConstantFP(0, getCurSDLoc(), EltVT);
1107 Constants[i] = DAG.getConstant(0, getCurSDLoc(), EltVT);
1110 return DAG.getMergeValues(Constants, getCurSDLoc());
1113 if (const BlockAddress *BA = dyn_cast<BlockAddress>(C))
1114 return DAG.getBlockAddress(BA, VT);
1116 VectorType *VecTy = cast<VectorType>(V->getType());
1117 unsigned NumElements = VecTy->getNumElements();
1119 // Now that we know the number and type of the elements, get that number of
1120 // elements into the Ops array based on what kind of constant it is.
1121 SmallVector<SDValue, 16> Ops;
1122 if (const ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
1123 for (unsigned i = 0; i != NumElements; ++i)
1124 Ops.push_back(getValue(CV->getOperand(i)));
1126 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!");
1128 TLI.getValueType(DAG.getDataLayout(), VecTy->getElementType());
1131 if (EltVT.isFloatingPoint())
1132 Op = DAG.getConstantFP(0, getCurSDLoc(), EltVT);
1134 Op = DAG.getConstant(0, getCurSDLoc(), EltVT);
1135 Ops.assign(NumElements, Op);
1138 // Create a BUILD_VECTOR node.
1139 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(), VT, Ops);
1142 // If this is a static alloca, generate it as the frameindex instead of
1144 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
1145 DenseMap<const AllocaInst*, int>::iterator SI =
1146 FuncInfo.StaticAllocaMap.find(AI);
1147 if (SI != FuncInfo.StaticAllocaMap.end())
1148 return DAG.getFrameIndex(SI->second,
1149 TLI.getPointerTy(DAG.getDataLayout()));
1152 // If this is an instruction which fast-isel has deferred, select it now.
1153 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
1154 unsigned InReg = FuncInfo.InitializeRegForValue(Inst);
1155 RegsForValue RFV(*DAG.getContext(), TLI, DAG.getDataLayout(), InReg,
1157 SDValue Chain = DAG.getEntryNode();
1158 return RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, nullptr, V);
1161 llvm_unreachable("Can't get register for value!");
1164 void SelectionDAGBuilder::visitCatchPad(const CatchPadInst &I) {
1165 auto Pers = classifyEHPersonality(FuncInfo.Fn->getPersonalityFn());
1166 bool IsMSVCCXX = Pers == EHPersonality::MSVC_CXX;
1167 bool IsSEH = isAsynchronousEHPersonality(Pers);
1168 bool IsCoreCLR = Pers == EHPersonality::CoreCLR;
1169 MachineBasicBlock *CatchPadMBB = FuncInfo.MBB;
1170 // In MSVC C++ and CoreCLR, catchblocks are funclets and need prologues.
1171 if (IsMSVCCXX || IsCoreCLR)
1172 CatchPadMBB->setIsEHFuncletEntry();
1174 MachineBasicBlock *NormalDestMBB = FuncInfo.MBBMap[I.getNormalDest()];
1176 // Update machine-CFG edge.
1177 FuncInfo.MBB->addSuccessor(NormalDestMBB);
1179 // CatchPads in SEH are not funclets, they are merely markers which indicate
1180 // where to insert register restoration code.
1182 DAG.setRoot(DAG.getNode(ISD::CATCHRET, getCurSDLoc(), MVT::Other,
1183 getControlRoot(), DAG.getBasicBlock(NormalDestMBB),
1184 DAG.getBasicBlock(&FuncInfo.MF->front())));
1188 // If this is not a fall-through branch or optimizations are switched off,
1190 if (NormalDestMBB != NextBlock(CatchPadMBB) ||
1191 TM.getOptLevel() == CodeGenOpt::None)
1192 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other,
1194 DAG.getBasicBlock(NormalDestMBB)));
1197 void SelectionDAGBuilder::visitCatchRet(const CatchReturnInst &I) {
1198 // Update machine-CFG edge.
1199 MachineBasicBlock *TargetMBB = FuncInfo.MBBMap[I.getSuccessor()];
1200 FuncInfo.MBB->addSuccessor(TargetMBB);
1202 auto Pers = classifyEHPersonality(FuncInfo.Fn->getPersonalityFn());
1203 bool IsSEH = isAsynchronousEHPersonality(Pers);
1205 // If this is not a fall-through branch or optimizations are switched off,
1207 if (TargetMBB != NextBlock(FuncInfo.MBB) ||
1208 TM.getOptLevel() == CodeGenOpt::None)
1209 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other,
1210 getControlRoot(), DAG.getBasicBlock(TargetMBB)));
1214 // Figure out the funclet membership for the catchret's successor.
1215 // This will be used by the FuncletLayout pass to determine how to order the
1217 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
1218 WinEHFuncInfo &EHInfo =
1219 MMI.getWinEHFuncInfo(DAG.getMachineFunction().getFunction());
1220 const BasicBlock *SuccessorColor = EHInfo.CatchRetSuccessorColorMap[&I];
1221 assert(SuccessorColor && "No parent funclet for catchret!");
1222 MachineBasicBlock *SuccessorColorMBB = FuncInfo.MBBMap[SuccessorColor];
1223 assert(SuccessorColorMBB && "No MBB for SuccessorColor!");
1225 // Create the terminator node.
1226 SDValue Ret = DAG.getNode(ISD::CATCHRET, getCurSDLoc(), MVT::Other,
1227 getControlRoot(), DAG.getBasicBlock(TargetMBB),
1228 DAG.getBasicBlock(SuccessorColorMBB));
1232 void SelectionDAGBuilder::visitCatchEndPad(const CatchEndPadInst &I) {
1233 llvm_unreachable("should never codegen catchendpads");
1236 void SelectionDAGBuilder::visitCleanupPad(const CleanupPadInst &CPI) {
1237 // Don't emit any special code for the cleanuppad instruction. It just marks
1238 // the start of a funclet.
1239 FuncInfo.MBB->setIsEHFuncletEntry();
1240 FuncInfo.MBB->setIsCleanupFuncletEntry();
1243 /// When an invoke or a cleanupret unwinds to the next EH pad, there are
1244 /// many places it could ultimately go. In the IR, we have a single unwind
1245 /// destination, but in the machine CFG, we enumerate all the possible blocks.
1246 /// This function skips over imaginary basic blocks that hold catchpad,
1247 /// terminatepad, or catchendpad instructions, and finds all the "real" machine
1248 /// basic block destinations. As those destinations may not be successors of
1249 /// EHPadBB, here we also calculate the edge weight to those destinations. The
1250 /// passed-in Weight is the edge weight to EHPadBB.
1251 static void findUnwindDestinations(
1252 FunctionLoweringInfo &FuncInfo, const BasicBlock *EHPadBB, uint32_t Weight,
1253 SmallVectorImpl<std::pair<MachineBasicBlock *, uint32_t>> &UnwindDests) {
1254 EHPersonality Personality =
1255 classifyEHPersonality(FuncInfo.Fn->getPersonalityFn());
1256 bool IsMSVCCXX = Personality == EHPersonality::MSVC_CXX;
1257 bool IsCoreCLR = Personality == EHPersonality::CoreCLR;
1260 const Instruction *Pad = EHPadBB->getFirstNonPHI();
1261 BasicBlock *NewEHPadBB = nullptr;
1262 if (isa<LandingPadInst>(Pad)) {
1263 // Stop on landingpads. They are not funclets.
1264 UnwindDests.emplace_back(FuncInfo.MBBMap[EHPadBB], Weight);
1266 } else if (isa<CleanupPadInst>(Pad)) {
1267 // Stop on cleanup pads. Cleanups are always funclet entries for all known
1269 UnwindDests.emplace_back(FuncInfo.MBBMap[EHPadBB], Weight);
1270 UnwindDests.back().first->setIsEHFuncletEntry();
1272 } else if (const auto *CPI = dyn_cast<CatchPadInst>(Pad)) {
1273 // Add the catchpad handler to the possible destinations.
1274 UnwindDests.emplace_back(FuncInfo.MBBMap[EHPadBB], Weight);
1275 // In MSVC C++, catchblocks are funclets and need prologues.
1276 if (IsMSVCCXX || IsCoreCLR)
1277 UnwindDests.back().first->setIsEHFuncletEntry();
1278 NewEHPadBB = CPI->getUnwindDest();
1279 } else if (const auto *CEPI = dyn_cast<CatchEndPadInst>(Pad))
1280 NewEHPadBB = CEPI->getUnwindDest();
1281 else if (const auto *CEPI = dyn_cast<CleanupEndPadInst>(Pad))
1282 NewEHPadBB = CEPI->getUnwindDest();
1286 BranchProbabilityInfo *BPI = FuncInfo.BPI;
1287 if (BPI && NewEHPadBB) {
1288 // When BPI is available, the calculated weight cannot be zero as zero
1289 // will be turned to a default weight in MachineBlockFrequencyInfo.
1290 Weight = std::max<uint32_t>(
1291 BPI->getEdgeProbability(EHPadBB, NewEHPadBB).scale(Weight), 1);
1293 EHPadBB = NewEHPadBB;
1297 void SelectionDAGBuilder::visitCleanupRet(const CleanupReturnInst &I) {
1298 // Update successor info.
1299 SmallVector<std::pair<MachineBasicBlock *, uint32_t>, 1> UnwindDests;
1300 auto UnwindDest = I.getUnwindDest();
1301 BranchProbabilityInfo *BPI = FuncInfo.BPI;
1302 uint32_t UnwindDestWeight =
1303 BPI ? BPI->getEdgeWeight(FuncInfo.MBB->getBasicBlock(), UnwindDest) : 0;
1304 findUnwindDestinations(FuncInfo, UnwindDest, UnwindDestWeight, UnwindDests);
1305 for (auto &UnwindDest : UnwindDests) {
1306 UnwindDest.first->setIsEHPad();
1307 addSuccessorWithWeight(FuncInfo.MBB, UnwindDest.first, UnwindDest.second);
1310 // Create the terminator node.
1312 DAG.getNode(ISD::CLEANUPRET, getCurSDLoc(), MVT::Other, getControlRoot());
1316 void SelectionDAGBuilder::visitCleanupEndPad(const CleanupEndPadInst &I) {
1317 report_fatal_error("visitCleanupEndPad not yet implemented!");
1320 void SelectionDAGBuilder::visitTerminatePad(const TerminatePadInst &TPI) {
1321 report_fatal_error("visitTerminatePad not yet implemented!");
1324 void SelectionDAGBuilder::visitRet(const ReturnInst &I) {
1325 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1326 auto &DL = DAG.getDataLayout();
1327 SDValue Chain = getControlRoot();
1328 SmallVector<ISD::OutputArg, 8> Outs;
1329 SmallVector<SDValue, 8> OutVals;
1331 if (!FuncInfo.CanLowerReturn) {
1332 unsigned DemoteReg = FuncInfo.DemoteRegister;
1333 const Function *F = I.getParent()->getParent();
1335 // Emit a store of the return value through the virtual register.
1336 // Leave Outs empty so that LowerReturn won't try to load return
1337 // registers the usual way.
1338 SmallVector<EVT, 1> PtrValueVTs;
1339 ComputeValueVTs(TLI, DL, PointerType::getUnqual(F->getReturnType()),
1342 SDValue RetPtr = DAG.getRegister(DemoteReg, PtrValueVTs[0]);
1343 SDValue RetOp = getValue(I.getOperand(0));
1345 SmallVector<EVT, 4> ValueVTs;
1346 SmallVector<uint64_t, 4> Offsets;
1347 ComputeValueVTs(TLI, DL, I.getOperand(0)->getType(), ValueVTs, &Offsets);
1348 unsigned NumValues = ValueVTs.size();
1350 SmallVector<SDValue, 4> Chains(NumValues);
1351 for (unsigned i = 0; i != NumValues; ++i) {
1352 SDValue Add = DAG.getNode(ISD::ADD, getCurSDLoc(),
1353 RetPtr.getValueType(), RetPtr,
1354 DAG.getIntPtrConstant(Offsets[i],
1357 DAG.getStore(Chain, getCurSDLoc(),
1358 SDValue(RetOp.getNode(), RetOp.getResNo() + i),
1359 // FIXME: better loc info would be nice.
1360 Add, MachinePointerInfo(), false, false, 0);
1363 Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(),
1364 MVT::Other, Chains);
1365 } else if (I.getNumOperands() != 0) {
1366 SmallVector<EVT, 4> ValueVTs;
1367 ComputeValueVTs(TLI, DL, I.getOperand(0)->getType(), ValueVTs);
1368 unsigned NumValues = ValueVTs.size();
1370 SDValue RetOp = getValue(I.getOperand(0));
1372 const Function *F = I.getParent()->getParent();
1374 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1375 if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1377 ExtendKind = ISD::SIGN_EXTEND;
1378 else if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1380 ExtendKind = ISD::ZERO_EXTEND;
1382 LLVMContext &Context = F->getContext();
1383 bool RetInReg = F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1386 for (unsigned j = 0; j != NumValues; ++j) {
1387 EVT VT = ValueVTs[j];
1389 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger())
1390 VT = TLI.getTypeForExtArgOrReturn(Context, VT, ExtendKind);
1392 unsigned NumParts = TLI.getNumRegisters(Context, VT);
1393 MVT PartVT = TLI.getRegisterType(Context, VT);
1394 SmallVector<SDValue, 4> Parts(NumParts);
1395 getCopyToParts(DAG, getCurSDLoc(),
1396 SDValue(RetOp.getNode(), RetOp.getResNo() + j),
1397 &Parts[0], NumParts, PartVT, &I, ExtendKind);
1399 // 'inreg' on function refers to return value
1400 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1404 // Propagate extension type if any
1405 if (ExtendKind == ISD::SIGN_EXTEND)
1407 else if (ExtendKind == ISD::ZERO_EXTEND)
1410 for (unsigned i = 0; i < NumParts; ++i) {
1411 Outs.push_back(ISD::OutputArg(Flags, Parts[i].getValueType(),
1412 VT, /*isfixed=*/true, 0, 0));
1413 OutVals.push_back(Parts[i]);
1419 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1420 CallingConv::ID CallConv =
1421 DAG.getMachineFunction().getFunction()->getCallingConv();
1422 Chain = DAG.getTargetLoweringInfo().LowerReturn(
1423 Chain, CallConv, isVarArg, Outs, OutVals, getCurSDLoc(), DAG);
1425 // Verify that the target's LowerReturn behaved as expected.
1426 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
1427 "LowerReturn didn't return a valid chain!");
1429 // Update the DAG with the new chain value resulting from return lowering.
1433 /// CopyToExportRegsIfNeeded - If the given value has virtual registers
1434 /// created for it, emit nodes to copy the value into the virtual
1436 void SelectionDAGBuilder::CopyToExportRegsIfNeeded(const Value *V) {
1438 if (V->getType()->isEmptyTy())
1441 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
1442 if (VMI != FuncInfo.ValueMap.end()) {
1443 assert(!V->use_empty() && "Unused value assigned virtual registers!");
1444 CopyValueToVirtualRegister(V, VMI->second);
1448 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
1449 /// the current basic block, add it to ValueMap now so that we'll get a
1451 void SelectionDAGBuilder::ExportFromCurrentBlock(const Value *V) {
1452 // No need to export constants.
1453 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
1455 // Already exported?
1456 if (FuncInfo.isExportedInst(V)) return;
1458 unsigned Reg = FuncInfo.InitializeRegForValue(V);
1459 CopyValueToVirtualRegister(V, Reg);
1462 bool SelectionDAGBuilder::isExportableFromCurrentBlock(const Value *V,
1463 const BasicBlock *FromBB) {
1464 // The operands of the setcc have to be in this block. We don't know
1465 // how to export them from some other block.
1466 if (const Instruction *VI = dyn_cast<Instruction>(V)) {
1467 // Can export from current BB.
1468 if (VI->getParent() == FromBB)
1471 // Is already exported, noop.
1472 return FuncInfo.isExportedInst(V);
1475 // If this is an argument, we can export it if the BB is the entry block or
1476 // if it is already exported.
1477 if (isa<Argument>(V)) {
1478 if (FromBB == &FromBB->getParent()->getEntryBlock())
1481 // Otherwise, can only export this if it is already exported.
1482 return FuncInfo.isExportedInst(V);
1485 // Otherwise, constants can always be exported.
1489 /// Return branch probability calculated by BranchProbabilityInfo for IR blocks.
1490 uint32_t SelectionDAGBuilder::getEdgeWeight(const MachineBasicBlock *Src,
1491 const MachineBasicBlock *Dst) const {
1492 BranchProbabilityInfo *BPI = FuncInfo.BPI;
1495 const BasicBlock *SrcBB = Src->getBasicBlock();
1496 const BasicBlock *DstBB = Dst->getBasicBlock();
1497 return BPI->getEdgeWeight(SrcBB, DstBB);
1500 void SelectionDAGBuilder::
1501 addSuccessorWithWeight(MachineBasicBlock *Src, MachineBasicBlock *Dst,
1502 uint32_t Weight /* = 0 */) {
1504 Src->addSuccessorWithoutWeight(Dst);
1507 Weight = getEdgeWeight(Src, Dst);
1508 Src->addSuccessor(Dst, Weight);
1513 static bool InBlock(const Value *V, const BasicBlock *BB) {
1514 if (const Instruction *I = dyn_cast<Instruction>(V))
1515 return I->getParent() == BB;
1519 /// EmitBranchForMergedCondition - Helper method for FindMergedConditions.
1520 /// This function emits a branch and is used at the leaves of an OR or an
1521 /// AND operator tree.
1524 SelectionDAGBuilder::EmitBranchForMergedCondition(const Value *Cond,
1525 MachineBasicBlock *TBB,
1526 MachineBasicBlock *FBB,
1527 MachineBasicBlock *CurBB,
1528 MachineBasicBlock *SwitchBB,
1531 const BasicBlock *BB = CurBB->getBasicBlock();
1533 // If the leaf of the tree is a comparison, merge the condition into
1535 if (const CmpInst *BOp = dyn_cast<CmpInst>(Cond)) {
1536 // The operands of the cmp have to be in this block. We don't know
1537 // how to export them from some other block. If this is the first block
1538 // of the sequence, no exporting is needed.
1539 if (CurBB == SwitchBB ||
1540 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
1541 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) {
1542 ISD::CondCode Condition;
1543 if (const ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
1544 Condition = getICmpCondCode(IC->getPredicate());
1546 const FCmpInst *FC = cast<FCmpInst>(Cond);
1547 Condition = getFCmpCondCode(FC->getPredicate());
1548 if (TM.Options.NoNaNsFPMath)
1549 Condition = getFCmpCodeWithoutNaN(Condition);
1552 CaseBlock CB(Condition, BOp->getOperand(0), BOp->getOperand(1), nullptr,
1553 TBB, FBB, CurBB, TWeight, FWeight);
1554 SwitchCases.push_back(CB);
1559 // Create a CaseBlock record representing this branch.
1560 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()),
1561 nullptr, TBB, FBB, CurBB, TWeight, FWeight);
1562 SwitchCases.push_back(CB);
1565 /// Scale down both weights to fit into uint32_t.
1566 static void ScaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
1567 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
1568 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
1569 NewTrue = NewTrue / Scale;
1570 NewFalse = NewFalse / Scale;
1573 /// FindMergedConditions - If Cond is an expression like
1574 void SelectionDAGBuilder::FindMergedConditions(const Value *Cond,
1575 MachineBasicBlock *TBB,
1576 MachineBasicBlock *FBB,
1577 MachineBasicBlock *CurBB,
1578 MachineBasicBlock *SwitchBB,
1579 Instruction::BinaryOps Opc,
1582 // If this node is not part of the or/and tree, emit it as a branch.
1583 const Instruction *BOp = dyn_cast<Instruction>(Cond);
1584 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
1585 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
1586 BOp->getParent() != CurBB->getBasicBlock() ||
1587 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
1588 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
1589 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB,
1594 // Create TmpBB after CurBB.
1595 MachineFunction::iterator BBI(CurBB);
1596 MachineFunction &MF = DAG.getMachineFunction();
1597 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock());
1598 CurBB->getParent()->insert(++BBI, TmpBB);
1600 if (Opc == Instruction::Or) {
1601 // Codegen X | Y as:
1610 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
1611 // The requirement is that
1612 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
1613 // = TrueProb for original BB.
1614 // Assuming the original weights are A and B, one choice is to set BB1's
1615 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
1617 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
1618 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
1619 // TmpBB, but the math is more complicated.
1621 uint64_t NewTrueWeight = TWeight;
1622 uint64_t NewFalseWeight = (uint64_t)TWeight + 2 * (uint64_t)FWeight;
1623 ScaleWeights(NewTrueWeight, NewFalseWeight);
1624 // Emit the LHS condition.
1625 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, SwitchBB, Opc,
1626 NewTrueWeight, NewFalseWeight);
1628 NewTrueWeight = TWeight;
1629 NewFalseWeight = 2 * (uint64_t)FWeight;
1630 ScaleWeights(NewTrueWeight, NewFalseWeight);
1631 // Emit the RHS condition into TmpBB.
1632 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc,
1633 NewTrueWeight, NewFalseWeight);
1635 assert(Opc == Instruction::And && "Unknown merge op!");
1636 // Codegen X & Y as:
1644 // This requires creation of TmpBB after CurBB.
1646 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
1647 // The requirement is that
1648 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
1649 // = FalseProb for original BB.
1650 // Assuming the original weights are A and B, one choice is to set BB1's
1651 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
1653 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
1655 uint64_t NewTrueWeight = 2 * (uint64_t)TWeight + (uint64_t)FWeight;
1656 uint64_t NewFalseWeight = FWeight;
1657 ScaleWeights(NewTrueWeight, NewFalseWeight);
1658 // Emit the LHS condition.
1659 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, SwitchBB, Opc,
1660 NewTrueWeight, NewFalseWeight);
1662 NewTrueWeight = 2 * (uint64_t)TWeight;
1663 NewFalseWeight = FWeight;
1664 ScaleWeights(NewTrueWeight, NewFalseWeight);
1665 // Emit the RHS condition into TmpBB.
1666 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc,
1667 NewTrueWeight, NewFalseWeight);
1671 /// If the set of cases should be emitted as a series of branches, return true.
1672 /// If we should emit this as a bunch of and/or'd together conditions, return
1675 SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases) {
1676 if (Cases.size() != 2) return true;
1678 // If this is two comparisons of the same values or'd or and'd together, they
1679 // will get folded into a single comparison, so don't emit two blocks.
1680 if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
1681 Cases[0].CmpRHS == Cases[1].CmpRHS) ||
1682 (Cases[0].CmpRHS == Cases[1].CmpLHS &&
1683 Cases[0].CmpLHS == Cases[1].CmpRHS)) {
1687 // Handle: (X != null) | (Y != null) --> (X|Y) != 0
1688 // Handle: (X == null) & (Y == null) --> (X|Y) == 0
1689 if (Cases[0].CmpRHS == Cases[1].CmpRHS &&
1690 Cases[0].CC == Cases[1].CC &&
1691 isa<Constant>(Cases[0].CmpRHS) &&
1692 cast<Constant>(Cases[0].CmpRHS)->isNullValue()) {
1693 if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB)
1695 if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB)
1702 void SelectionDAGBuilder::visitBr(const BranchInst &I) {
1703 MachineBasicBlock *BrMBB = FuncInfo.MBB;
1705 // Update machine-CFG edges.
1706 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
1708 if (I.isUnconditional()) {
1709 // Update machine-CFG edges.
1710 BrMBB->addSuccessor(Succ0MBB);
1712 // If this is not a fall-through branch or optimizations are switched off,
1714 if (Succ0MBB != NextBlock(BrMBB) || TM.getOptLevel() == CodeGenOpt::None)
1715 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(),
1716 MVT::Other, getControlRoot(),
1717 DAG.getBasicBlock(Succ0MBB)));
1722 // If this condition is one of the special cases we handle, do special stuff
1724 const Value *CondVal = I.getCondition();
1725 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
1727 // If this is a series of conditions that are or'd or and'd together, emit
1728 // this as a sequence of branches instead of setcc's with and/or operations.
1729 // As long as jumps are not expensive, this should improve performance.
1730 // For example, instead of something like:
1743 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
1744 Instruction::BinaryOps Opcode = BOp->getOpcode();
1745 if (!DAG.getTargetLoweringInfo().isJumpExpensive() && BOp->hasOneUse() &&
1746 !I.getMetadata(LLVMContext::MD_unpredictable) &&
1747 (Opcode == Instruction::And || Opcode == Instruction::Or)) {
1748 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB,
1749 Opcode, getEdgeWeight(BrMBB, Succ0MBB),
1750 getEdgeWeight(BrMBB, Succ1MBB));
1751 // If the compares in later blocks need to use values not currently
1752 // exported from this block, export them now. This block should always
1753 // be the first entry.
1754 assert(SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!");
1756 // Allow some cases to be rejected.
1757 if (ShouldEmitAsBranches(SwitchCases)) {
1758 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1759 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1760 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1763 // Emit the branch for this block.
1764 visitSwitchCase(SwitchCases[0], BrMBB);
1765 SwitchCases.erase(SwitchCases.begin());
1769 // Okay, we decided not to do this, remove any inserted MBB's and clear
1771 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
1772 FuncInfo.MF->erase(SwitchCases[i].ThisBB);
1774 SwitchCases.clear();
1778 // Create a CaseBlock record representing this branch.
1779 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()),
1780 nullptr, Succ0MBB, Succ1MBB, BrMBB);
1782 // Use visitSwitchCase to actually insert the fast branch sequence for this
1784 visitSwitchCase(CB, BrMBB);
1787 /// visitSwitchCase - Emits the necessary code to represent a single node in
1788 /// the binary search tree resulting from lowering a switch instruction.
1789 void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB,
1790 MachineBasicBlock *SwitchBB) {
1792 SDValue CondLHS = getValue(CB.CmpLHS);
1793 SDLoc dl = getCurSDLoc();
1795 // Build the setcc now.
1797 // Fold "(X == true)" to X and "(X == false)" to !X to
1798 // handle common cases produced by branch lowering.
1799 if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) &&
1800 CB.CC == ISD::SETEQ)
1802 else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) &&
1803 CB.CC == ISD::SETEQ) {
1804 SDValue True = DAG.getConstant(1, dl, CondLHS.getValueType());
1805 Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True);
1807 Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1809 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now");
1811 const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue();
1812 const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue();
1814 SDValue CmpOp = getValue(CB.CmpMHS);
1815 EVT VT = CmpOp.getValueType();
1817 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) {
1818 Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, dl, VT),
1821 SDValue SUB = DAG.getNode(ISD::SUB, dl,
1822 VT, CmpOp, DAG.getConstant(Low, dl, VT));
1823 Cond = DAG.getSetCC(dl, MVT::i1, SUB,
1824 DAG.getConstant(High-Low, dl, VT), ISD::SETULE);
1828 // Update successor info
1829 addSuccessorWithWeight(SwitchBB, CB.TrueBB, CB.TrueWeight);
1830 // TrueBB and FalseBB are always different unless the incoming IR is
1831 // degenerate. This only happens when running llc on weird IR.
1832 if (CB.TrueBB != CB.FalseBB)
1833 addSuccessorWithWeight(SwitchBB, CB.FalseBB, CB.FalseWeight);
1835 // If the lhs block is the next block, invert the condition so that we can
1836 // fall through to the lhs instead of the rhs block.
1837 if (CB.TrueBB == NextBlock(SwitchBB)) {
1838 std::swap(CB.TrueBB, CB.FalseBB);
1839 SDValue True = DAG.getConstant(1, dl, Cond.getValueType());
1840 Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True);
1843 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1844 MVT::Other, getControlRoot(), Cond,
1845 DAG.getBasicBlock(CB.TrueBB));
1847 // Insert the false branch. Do this even if it's a fall through branch,
1848 // this makes it easier to do DAG optimizations which require inverting
1849 // the branch condition.
1850 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1851 DAG.getBasicBlock(CB.FalseBB));
1853 DAG.setRoot(BrCond);
1856 /// visitJumpTable - Emit JumpTable node in the current MBB
1857 void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) {
1858 // Emit the code for the jump table
1859 assert(JT.Reg != -1U && "Should lower JT Header first!");
1860 EVT PTy = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
1861 SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurSDLoc(),
1863 SDValue Table = DAG.getJumpTable(JT.JTI, PTy);
1864 SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurSDLoc(),
1865 MVT::Other, Index.getValue(1),
1867 DAG.setRoot(BrJumpTable);
1870 /// visitJumpTableHeader - This function emits necessary code to produce index
1871 /// in the JumpTable from switch case.
1872 void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT,
1873 JumpTableHeader &JTH,
1874 MachineBasicBlock *SwitchBB) {
1875 SDLoc dl = getCurSDLoc();
1877 // Subtract the lowest switch case value from the value being switched on and
1878 // conditional branch to default mbb if the result is greater than the
1879 // difference between smallest and largest cases.
1880 SDValue SwitchOp = getValue(JTH.SValue);
1881 EVT VT = SwitchOp.getValueType();
1882 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, SwitchOp,
1883 DAG.getConstant(JTH.First, dl, VT));
1885 // The SDNode we just created, which holds the value being switched on minus
1886 // the smallest case value, needs to be copied to a virtual register so it
1887 // can be used as an index into the jump table in a subsequent basic block.
1888 // This value may be smaller or larger than the target's pointer type, and
1889 // therefore require extension or truncating.
1890 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1891 SwitchOp = DAG.getZExtOrTrunc(Sub, dl, TLI.getPointerTy(DAG.getDataLayout()));
1893 unsigned JumpTableReg =
1894 FuncInfo.CreateReg(TLI.getPointerTy(DAG.getDataLayout()));
1895 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), dl,
1896 JumpTableReg, SwitchOp);
1897 JT.Reg = JumpTableReg;
1899 // Emit the range check for the jump table, and branch to the default block
1900 // for the switch statement if the value being switched on exceeds the largest
1901 // case in the switch.
1902 SDValue CMP = DAG.getSetCC(
1903 dl, TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(),
1904 Sub.getValueType()),
1905 Sub, DAG.getConstant(JTH.Last - JTH.First, dl, VT), ISD::SETUGT);
1907 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1908 MVT::Other, CopyTo, CMP,
1909 DAG.getBasicBlock(JT.Default));
1911 // Avoid emitting unnecessary branches to the next block.
1912 if (JT.MBB != NextBlock(SwitchBB))
1913 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1914 DAG.getBasicBlock(JT.MBB));
1916 DAG.setRoot(BrCond);
1919 /// Codegen a new tail for a stack protector check ParentMBB which has had its
1920 /// tail spliced into a stack protector check success bb.
1922 /// For a high level explanation of how this fits into the stack protector
1923 /// generation see the comment on the declaration of class
1924 /// StackProtectorDescriptor.
1925 void SelectionDAGBuilder::visitSPDescriptorParent(StackProtectorDescriptor &SPD,
1926 MachineBasicBlock *ParentBB) {
1928 // First create the loads to the guard/stack slot for the comparison.
1929 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1930 EVT PtrTy = TLI.getPointerTy(DAG.getDataLayout());
1932 MachineFrameInfo *MFI = ParentBB->getParent()->getFrameInfo();
1933 int FI = MFI->getStackProtectorIndex();
1935 const Value *IRGuard = SPD.getGuard();
1936 SDValue GuardPtr = getValue(IRGuard);
1937 SDValue StackSlotPtr = DAG.getFrameIndex(FI, PtrTy);
1939 unsigned Align = DL->getPrefTypeAlignment(IRGuard->getType());
1942 SDLoc dl = getCurSDLoc();
1944 // If GuardReg is set and useLoadStackGuardNode returns true, retrieve the
1945 // guard value from the virtual register holding the value. Otherwise, emit a
1946 // volatile load to retrieve the stack guard value.
1947 unsigned GuardReg = SPD.getGuardReg();
1949 if (GuardReg && TLI.useLoadStackGuardNode())
1950 Guard = DAG.getCopyFromReg(DAG.getEntryNode(), dl, GuardReg,
1953 Guard = DAG.getLoad(PtrTy, dl, DAG.getEntryNode(),
1954 GuardPtr, MachinePointerInfo(IRGuard, 0),
1955 true, false, false, Align);
1957 SDValue StackSlot = DAG.getLoad(
1958 PtrTy, dl, DAG.getEntryNode(), StackSlotPtr,
1959 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), true,
1960 false, false, Align);
1962 // Perform the comparison via a subtract/getsetcc.
1963 EVT VT = Guard.getValueType();
1964 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, Guard, StackSlot);
1966 SDValue Cmp = DAG.getSetCC(dl, TLI.getSetCCResultType(DAG.getDataLayout(),
1968 Sub.getValueType()),
1969 Sub, DAG.getConstant(0, dl, VT), ISD::SETNE);
1971 // If the sub is not 0, then we know the guard/stackslot do not equal, so
1972 // branch to failure MBB.
1973 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1974 MVT::Other, StackSlot.getOperand(0),
1975 Cmp, DAG.getBasicBlock(SPD.getFailureMBB()));
1976 // Otherwise branch to success MBB.
1977 SDValue Br = DAG.getNode(ISD::BR, dl,
1979 DAG.getBasicBlock(SPD.getSuccessMBB()));
1984 /// Codegen the failure basic block for a stack protector check.
1986 /// A failure stack protector machine basic block consists simply of a call to
1987 /// __stack_chk_fail().
1989 /// For a high level explanation of how this fits into the stack protector
1990 /// generation see the comment on the declaration of class
1991 /// StackProtectorDescriptor.
1993 SelectionDAGBuilder::visitSPDescriptorFailure(StackProtectorDescriptor &SPD) {
1994 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1996 TLI.makeLibCall(DAG, RTLIB::STACKPROTECTOR_CHECK_FAIL, MVT::isVoid,
1997 None, false, getCurSDLoc(), false, false).second;
2001 /// visitBitTestHeader - This function emits necessary code to produce value
2002 /// suitable for "bit tests"
2003 void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B,
2004 MachineBasicBlock *SwitchBB) {
2005 SDLoc dl = getCurSDLoc();
2007 // Subtract the minimum value
2008 SDValue SwitchOp = getValue(B.SValue);
2009 EVT VT = SwitchOp.getValueType();
2010 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, SwitchOp,
2011 DAG.getConstant(B.First, dl, VT));
2014 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2015 SDValue RangeCmp = DAG.getSetCC(
2016 dl, TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(),
2017 Sub.getValueType()),
2018 Sub, DAG.getConstant(B.Range, dl, VT), ISD::SETUGT);
2020 // Determine the type of the test operands.
2021 bool UsePtrType = false;
2022 if (!TLI.isTypeLegal(VT))
2025 for (unsigned i = 0, e = B.Cases.size(); i != e; ++i)
2026 if (!isUIntN(VT.getSizeInBits(), B.Cases[i].Mask)) {
2027 // Switch table case range are encoded into series of masks.
2028 // Just use pointer type, it's guaranteed to fit.
2034 VT = TLI.getPointerTy(DAG.getDataLayout());
2035 Sub = DAG.getZExtOrTrunc(Sub, dl, VT);
2038 B.RegVT = VT.getSimpleVT();
2039 B.Reg = FuncInfo.CreateReg(B.RegVT);
2040 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), dl, B.Reg, Sub);
2042 MachineBasicBlock* MBB = B.Cases[0].ThisBB;
2044 addSuccessorWithWeight(SwitchBB, B.Default, B.DefaultWeight);
2045 addSuccessorWithWeight(SwitchBB, MBB, B.Weight);
2047 SDValue BrRange = DAG.getNode(ISD::BRCOND, dl,
2048 MVT::Other, CopyTo, RangeCmp,
2049 DAG.getBasicBlock(B.Default));
2051 // Avoid emitting unnecessary branches to the next block.
2052 if (MBB != NextBlock(SwitchBB))
2053 BrRange = DAG.getNode(ISD::BR, dl, MVT::Other, BrRange,
2054 DAG.getBasicBlock(MBB));
2056 DAG.setRoot(BrRange);
2059 /// visitBitTestCase - this function produces one "bit test"
2060 void SelectionDAGBuilder::visitBitTestCase(BitTestBlock &BB,
2061 MachineBasicBlock* NextMBB,
2062 uint32_t BranchWeightToNext,
2065 MachineBasicBlock *SwitchBB) {
2066 SDLoc dl = getCurSDLoc();
2068 SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), dl, Reg, VT);
2070 unsigned PopCount = countPopulation(B.Mask);
2071 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2072 if (PopCount == 1) {
2073 // Testing for a single bit; just compare the shift count with what it
2074 // would need to be to shift a 1 bit in that position.
2076 dl, TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT),
2077 ShiftOp, DAG.getConstant(countTrailingZeros(B.Mask), dl, VT),
2079 } else if (PopCount == BB.Range) {
2080 // There is only one zero bit in the range, test for it directly.
2082 dl, TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT),
2083 ShiftOp, DAG.getConstant(countTrailingOnes(B.Mask), dl, VT),
2086 // Make desired shift
2087 SDValue SwitchVal = DAG.getNode(ISD::SHL, dl, VT,
2088 DAG.getConstant(1, dl, VT), ShiftOp);
2090 // Emit bit tests and jumps
2091 SDValue AndOp = DAG.getNode(ISD::AND, dl,
2092 VT, SwitchVal, DAG.getConstant(B.Mask, dl, VT));
2094 dl, TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT),
2095 AndOp, DAG.getConstant(0, dl, VT), ISD::SETNE);
2098 // The branch weight from SwitchBB to B.TargetBB is B.ExtraWeight.
2099 addSuccessorWithWeight(SwitchBB, B.TargetBB, B.ExtraWeight);
2100 // The branch weight from SwitchBB to NextMBB is BranchWeightToNext.
2101 addSuccessorWithWeight(SwitchBB, NextMBB, BranchWeightToNext);
2103 SDValue BrAnd = DAG.getNode(ISD::BRCOND, dl,
2104 MVT::Other, getControlRoot(),
2105 Cmp, DAG.getBasicBlock(B.TargetBB));
2107 // Avoid emitting unnecessary branches to the next block.
2108 if (NextMBB != NextBlock(SwitchBB))
2109 BrAnd = DAG.getNode(ISD::BR, dl, MVT::Other, BrAnd,
2110 DAG.getBasicBlock(NextMBB));
2115 void SelectionDAGBuilder::visitInvoke(const InvokeInst &I) {
2116 MachineBasicBlock *InvokeMBB = FuncInfo.MBB;
2118 // Retrieve successors. Look through artificial IR level blocks like catchpads
2119 // and catchendpads for successors.
2120 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
2121 const BasicBlock *EHPadBB = I.getSuccessor(1);
2123 const Value *Callee(I.getCalledValue());
2124 const Function *Fn = dyn_cast<Function>(Callee);
2125 if (isa<InlineAsm>(Callee))
2127 else if (Fn && Fn->isIntrinsic()) {
2128 switch (Fn->getIntrinsicID()) {
2130 llvm_unreachable("Cannot invoke this intrinsic");
2131 case Intrinsic::donothing:
2132 // Ignore invokes to @llvm.donothing: jump directly to the next BB.
2134 case Intrinsic::experimental_patchpoint_void:
2135 case Intrinsic::experimental_patchpoint_i64:
2136 visitPatchpoint(&I, EHPadBB);
2138 case Intrinsic::experimental_gc_statepoint:
2139 LowerStatepoint(ImmutableStatepoint(&I), EHPadBB);
2143 LowerCallTo(&I, getValue(Callee), false, EHPadBB);
2145 // If the value of the invoke is used outside of its defining block, make it
2146 // available as a virtual register.
2147 // We already took care of the exported value for the statepoint instruction
2148 // during call to the LowerStatepoint.
2149 if (!isStatepoint(I)) {
2150 CopyToExportRegsIfNeeded(&I);
2153 SmallVector<std::pair<MachineBasicBlock *, uint32_t>, 1> UnwindDests;
2154 BranchProbabilityInfo *BPI = FuncInfo.BPI;
2155 uint32_t EHPadBBWeight =
2156 BPI ? BPI->getEdgeWeight(InvokeMBB->getBasicBlock(), EHPadBB) : 0;
2157 findUnwindDestinations(FuncInfo, EHPadBB, EHPadBBWeight, UnwindDests);
2159 // Update successor info.
2160 addSuccessorWithWeight(InvokeMBB, Return);
2161 for (auto &UnwindDest : UnwindDests) {
2162 UnwindDest.first->setIsEHPad();
2163 addSuccessorWithWeight(InvokeMBB, UnwindDest.first, UnwindDest.second);
2166 // Drop into normal successor.
2167 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(),
2168 MVT::Other, getControlRoot(),
2169 DAG.getBasicBlock(Return)));
2172 void SelectionDAGBuilder::visitResume(const ResumeInst &RI) {
2173 llvm_unreachable("SelectionDAGBuilder shouldn't visit resume instructions!");
2176 void SelectionDAGBuilder::visitLandingPad(const LandingPadInst &LP) {
2177 assert(FuncInfo.MBB->isEHPad() &&
2178 "Call to landingpad not in landing pad!");
2180 MachineBasicBlock *MBB = FuncInfo.MBB;
2181 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
2182 AddLandingPadInfo(LP, MMI, MBB);
2184 // If there aren't registers to copy the values into (e.g., during SjLj
2185 // exceptions), then don't bother to create these DAG nodes.
2186 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2187 const Constant *PersonalityFn = FuncInfo.Fn->getPersonalityFn();
2188 if (TLI.getExceptionPointerRegister(PersonalityFn) == 0 &&
2189 TLI.getExceptionSelectorRegister(PersonalityFn) == 0)
2192 SmallVector<EVT, 2> ValueVTs;
2193 SDLoc dl = getCurSDLoc();
2194 ComputeValueVTs(TLI, DAG.getDataLayout(), LP.getType(), ValueVTs);
2195 assert(ValueVTs.size() == 2 && "Only two-valued landingpads are supported");
2197 // Get the two live-in registers as SDValues. The physregs have already been
2198 // copied into virtual registers.
2200 if (FuncInfo.ExceptionPointerVirtReg) {
2201 Ops[0] = DAG.getZExtOrTrunc(
2202 DAG.getCopyFromReg(DAG.getEntryNode(), dl,
2203 FuncInfo.ExceptionPointerVirtReg,
2204 TLI.getPointerTy(DAG.getDataLayout())),
2207 Ops[0] = DAG.getConstant(0, dl, TLI.getPointerTy(DAG.getDataLayout()));
2209 Ops[1] = DAG.getZExtOrTrunc(
2210 DAG.getCopyFromReg(DAG.getEntryNode(), dl,
2211 FuncInfo.ExceptionSelectorVirtReg,
2212 TLI.getPointerTy(DAG.getDataLayout())),
2216 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, dl,
2217 DAG.getVTList(ValueVTs), Ops);
2221 void SelectionDAGBuilder::sortAndRangeify(CaseClusterVector &Clusters) {
2223 for (const CaseCluster &CC : Clusters)
2224 assert(CC.Low == CC.High && "Input clusters must be single-case");
2227 std::sort(Clusters.begin(), Clusters.end(),
2228 [](const CaseCluster &a, const CaseCluster &b) {
2229 return a.Low->getValue().slt(b.Low->getValue());
2232 // Merge adjacent clusters with the same destination.
2233 const unsigned N = Clusters.size();
2234 unsigned DstIndex = 0;
2235 for (unsigned SrcIndex = 0; SrcIndex < N; ++SrcIndex) {
2236 CaseCluster &CC = Clusters[SrcIndex];
2237 const ConstantInt *CaseVal = CC.Low;
2238 MachineBasicBlock *Succ = CC.MBB;
2240 if (DstIndex != 0 && Clusters[DstIndex - 1].MBB == Succ &&
2241 (CaseVal->getValue() - Clusters[DstIndex - 1].High->getValue()) == 1) {
2242 // If this case has the same successor and is a neighbour, merge it into
2243 // the previous cluster.
2244 Clusters[DstIndex - 1].High = CaseVal;
2245 Clusters[DstIndex - 1].Weight += CC.Weight;
2246 assert(Clusters[DstIndex - 1].Weight >= CC.Weight && "Weight overflow!");
2248 std::memmove(&Clusters[DstIndex++], &Clusters[SrcIndex],
2249 sizeof(Clusters[SrcIndex]));
2252 Clusters.resize(DstIndex);
2255 void SelectionDAGBuilder::UpdateSplitBlock(MachineBasicBlock *First,
2256 MachineBasicBlock *Last) {
2258 for (unsigned i = 0, e = JTCases.size(); i != e; ++i)
2259 if (JTCases[i].first.HeaderBB == First)
2260 JTCases[i].first.HeaderBB = Last;
2262 // Update BitTestCases.
2263 for (unsigned i = 0, e = BitTestCases.size(); i != e; ++i)
2264 if (BitTestCases[i].Parent == First)
2265 BitTestCases[i].Parent = Last;
2268 void SelectionDAGBuilder::visitIndirectBr(const IndirectBrInst &I) {
2269 MachineBasicBlock *IndirectBrMBB = FuncInfo.MBB;
2271 // Update machine-CFG edges with unique successors.
2272 SmallSet<BasicBlock*, 32> Done;
2273 for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i) {
2274 BasicBlock *BB = I.getSuccessor(i);
2275 bool Inserted = Done.insert(BB).second;
2279 MachineBasicBlock *Succ = FuncInfo.MBBMap[BB];
2280 addSuccessorWithWeight(IndirectBrMBB, Succ);
2283 DAG.setRoot(DAG.getNode(ISD::BRIND, getCurSDLoc(),
2284 MVT::Other, getControlRoot(),
2285 getValue(I.getAddress())));
2288 void SelectionDAGBuilder::visitUnreachable(const UnreachableInst &I) {
2289 if (DAG.getTarget().Options.TrapUnreachable)
2291 DAG.getNode(ISD::TRAP, getCurSDLoc(), MVT::Other, DAG.getRoot()));
2294 void SelectionDAGBuilder::visitFSub(const User &I) {
2295 // -0.0 - X --> fneg
2296 Type *Ty = I.getType();
2297 if (isa<Constant>(I.getOperand(0)) &&
2298 I.getOperand(0) == ConstantFP::getZeroValueForNegation(Ty)) {
2299 SDValue Op2 = getValue(I.getOperand(1));
2300 setValue(&I, DAG.getNode(ISD::FNEG, getCurSDLoc(),
2301 Op2.getValueType(), Op2));
2305 visitBinary(I, ISD::FSUB);
2308 void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) {
2309 SDValue Op1 = getValue(I.getOperand(0));
2310 SDValue Op2 = getValue(I.getOperand(1));
2317 if (const OverflowingBinaryOperator *OFBinOp =
2318 dyn_cast<const OverflowingBinaryOperator>(&I)) {
2319 nuw = OFBinOp->hasNoUnsignedWrap();
2320 nsw = OFBinOp->hasNoSignedWrap();
2322 if (const PossiblyExactOperator *ExactOp =
2323 dyn_cast<const PossiblyExactOperator>(&I))
2324 exact = ExactOp->isExact();
2325 if (const FPMathOperator *FPOp = dyn_cast<const FPMathOperator>(&I))
2326 FMF = FPOp->getFastMathFlags();
2329 Flags.setExact(exact);
2330 Flags.setNoSignedWrap(nsw);
2331 Flags.setNoUnsignedWrap(nuw);
2332 if (EnableFMFInDAG) {
2333 Flags.setAllowReciprocal(FMF.allowReciprocal());
2334 Flags.setNoInfs(FMF.noInfs());
2335 Flags.setNoNaNs(FMF.noNaNs());
2336 Flags.setNoSignedZeros(FMF.noSignedZeros());
2337 Flags.setUnsafeAlgebra(FMF.unsafeAlgebra());
2339 SDValue BinNodeValue = DAG.getNode(OpCode, getCurSDLoc(), Op1.getValueType(),
2341 setValue(&I, BinNodeValue);
2344 void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) {
2345 SDValue Op1 = getValue(I.getOperand(0));
2346 SDValue Op2 = getValue(I.getOperand(1));
2348 EVT ShiftTy = DAG.getTargetLoweringInfo().getShiftAmountTy(
2349 Op2.getValueType(), DAG.getDataLayout());
2351 // Coerce the shift amount to the right type if we can.
2352 if (!I.getType()->isVectorTy() && Op2.getValueType() != ShiftTy) {
2353 unsigned ShiftSize = ShiftTy.getSizeInBits();
2354 unsigned Op2Size = Op2.getValueType().getSizeInBits();
2355 SDLoc DL = getCurSDLoc();
2357 // If the operand is smaller than the shift count type, promote it.
2358 if (ShiftSize > Op2Size)
2359 Op2 = DAG.getNode(ISD::ZERO_EXTEND, DL, ShiftTy, Op2);
2361 // If the operand is larger than the shift count type but the shift
2362 // count type has enough bits to represent any shift value, truncate
2363 // it now. This is a common case and it exposes the truncate to
2364 // optimization early.
2365 else if (ShiftSize >= Log2_32_Ceil(Op2.getValueType().getSizeInBits()))
2366 Op2 = DAG.getNode(ISD::TRUNCATE, DL, ShiftTy, Op2);
2367 // Otherwise we'll need to temporarily settle for some other convenient
2368 // type. Type legalization will make adjustments once the shiftee is split.
2370 Op2 = DAG.getZExtOrTrunc(Op2, DL, MVT::i32);
2377 if (Opcode == ISD::SRL || Opcode == ISD::SRA || Opcode == ISD::SHL) {
2379 if (const OverflowingBinaryOperator *OFBinOp =
2380 dyn_cast<const OverflowingBinaryOperator>(&I)) {
2381 nuw = OFBinOp->hasNoUnsignedWrap();
2382 nsw = OFBinOp->hasNoSignedWrap();
2384 if (const PossiblyExactOperator *ExactOp =
2385 dyn_cast<const PossiblyExactOperator>(&I))
2386 exact = ExactOp->isExact();
2389 Flags.setExact(exact);
2390 Flags.setNoSignedWrap(nsw);
2391 Flags.setNoUnsignedWrap(nuw);
2392 SDValue Res = DAG.getNode(Opcode, getCurSDLoc(), Op1.getValueType(), Op1, Op2,
2397 void SelectionDAGBuilder::visitSDiv(const User &I) {
2398 SDValue Op1 = getValue(I.getOperand(0));
2399 SDValue Op2 = getValue(I.getOperand(1));
2402 Flags.setExact(isa<PossiblyExactOperator>(&I) &&
2403 cast<PossiblyExactOperator>(&I)->isExact());
2404 setValue(&I, DAG.getNode(ISD::SDIV, getCurSDLoc(), Op1.getValueType(), Op1,
2408 void SelectionDAGBuilder::visitICmp(const User &I) {
2409 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2410 if (const ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2411 predicate = IC->getPredicate();
2412 else if (const ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2413 predicate = ICmpInst::Predicate(IC->getPredicate());
2414 SDValue Op1 = getValue(I.getOperand(0));
2415 SDValue Op2 = getValue(I.getOperand(1));
2416 ISD::CondCode Opcode = getICmpCondCode(predicate);
2418 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2420 setValue(&I, DAG.getSetCC(getCurSDLoc(), DestVT, Op1, Op2, Opcode));
2423 void SelectionDAGBuilder::visitFCmp(const User &I) {
2424 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2425 if (const FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2426 predicate = FC->getPredicate();
2427 else if (const ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2428 predicate = FCmpInst::Predicate(FC->getPredicate());
2429 SDValue Op1 = getValue(I.getOperand(0));
2430 SDValue Op2 = getValue(I.getOperand(1));
2431 ISD::CondCode Condition = getFCmpCondCode(predicate);
2433 // FIXME: Fcmp instructions have fast-math-flags in IR, so we should use them.
2434 // FIXME: We should propagate the fast-math-flags to the DAG node itself for
2435 // further optimization, but currently FMF is only applicable to binary nodes.
2436 if (TM.Options.NoNaNsFPMath)
2437 Condition = getFCmpCodeWithoutNaN(Condition);
2438 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2440 setValue(&I, DAG.getSetCC(getCurSDLoc(), DestVT, Op1, Op2, Condition));
2443 void SelectionDAGBuilder::visitSelect(const User &I) {
2444 SmallVector<EVT, 4> ValueVTs;
2445 ComputeValueVTs(DAG.getTargetLoweringInfo(), DAG.getDataLayout(), I.getType(),
2447 unsigned NumValues = ValueVTs.size();
2448 if (NumValues == 0) return;
2450 SmallVector<SDValue, 4> Values(NumValues);
2451 SDValue Cond = getValue(I.getOperand(0));
2452 SDValue LHSVal = getValue(I.getOperand(1));
2453 SDValue RHSVal = getValue(I.getOperand(2));
2454 auto BaseOps = {Cond};
2455 ISD::NodeType OpCode = Cond.getValueType().isVector() ?
2456 ISD::VSELECT : ISD::SELECT;
2458 // Min/max matching is only viable if all output VTs are the same.
2459 if (std::equal(ValueVTs.begin(), ValueVTs.end(), ValueVTs.begin())) {
2460 EVT VT = ValueVTs[0];
2461 LLVMContext &Ctx = *DAG.getContext();
2462 auto &TLI = DAG.getTargetLoweringInfo();
2463 while (TLI.getTypeAction(Ctx, VT) == TargetLoweringBase::TypeSplitVector)
2464 VT = TLI.getTypeToTransformTo(Ctx, VT);
2467 auto SPR = matchSelectPattern(const_cast<User*>(&I), LHS, RHS);
2468 ISD::NodeType Opc = ISD::DELETED_NODE;
2469 switch (SPR.Flavor) {
2470 case SPF_UMAX: Opc = ISD::UMAX; break;
2471 case SPF_UMIN: Opc = ISD::UMIN; break;
2472 case SPF_SMAX: Opc = ISD::SMAX; break;
2473 case SPF_SMIN: Opc = ISD::SMIN; break;
2475 switch (SPR.NaNBehavior) {
2476 case SPNB_NA: llvm_unreachable("No NaN behavior for FP op?");
2477 case SPNB_RETURNS_NAN: Opc = ISD::FMINNAN; break;
2478 case SPNB_RETURNS_OTHER: Opc = ISD::FMINNUM; break;
2479 case SPNB_RETURNS_ANY:
2480 Opc = TLI.isOperationLegalOrCustom(ISD::FMINNUM, VT) ? ISD::FMINNUM
2486 switch (SPR.NaNBehavior) {
2487 case SPNB_NA: llvm_unreachable("No NaN behavior for FP op?");
2488 case SPNB_RETURNS_NAN: Opc = ISD::FMAXNAN; break;
2489 case SPNB_RETURNS_OTHER: Opc = ISD::FMAXNUM; break;
2490 case SPNB_RETURNS_ANY:
2491 Opc = TLI.isOperationLegalOrCustom(ISD::FMAXNUM, VT) ? ISD::FMAXNUM
2499 if (Opc != ISD::DELETED_NODE && TLI.isOperationLegalOrCustom(Opc, VT) &&
2500 // If the underlying comparison instruction is used by any other instruction,
2501 // the consumed instructions won't be destroyed, so it is not profitable
2502 // to convert to a min/max.
2503 cast<SelectInst>(&I)->getCondition()->hasOneUse()) {
2505 LHSVal = getValue(LHS);
2506 RHSVal = getValue(RHS);
2511 for (unsigned i = 0; i != NumValues; ++i) {
2512 SmallVector<SDValue, 3> Ops(BaseOps.begin(), BaseOps.end());
2513 Ops.push_back(SDValue(LHSVal.getNode(), LHSVal.getResNo() + i));
2514 Ops.push_back(SDValue(RHSVal.getNode(), RHSVal.getResNo() + i));
2515 Values[i] = DAG.getNode(OpCode, getCurSDLoc(),
2516 LHSVal.getNode()->getValueType(LHSVal.getResNo()+i),
2520 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
2521 DAG.getVTList(ValueVTs), Values));
2524 void SelectionDAGBuilder::visitTrunc(const User &I) {
2525 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2526 SDValue N = getValue(I.getOperand(0));
2527 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2529 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), DestVT, N));
2532 void SelectionDAGBuilder::visitZExt(const User &I) {
2533 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2534 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2535 SDValue N = getValue(I.getOperand(0));
2536 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2538 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurSDLoc(), DestVT, N));
2541 void SelectionDAGBuilder::visitSExt(const User &I) {
2542 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2543 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2544 SDValue N = getValue(I.getOperand(0));
2545 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2547 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurSDLoc(), DestVT, N));
2550 void SelectionDAGBuilder::visitFPTrunc(const User &I) {
2551 // FPTrunc is never a no-op cast, no need to check
2552 SDValue N = getValue(I.getOperand(0));
2553 SDLoc dl = getCurSDLoc();
2554 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2555 EVT DestVT = TLI.getValueType(DAG.getDataLayout(), I.getType());
2556 setValue(&I, DAG.getNode(ISD::FP_ROUND, dl, DestVT, N,
2557 DAG.getTargetConstant(
2558 0, dl, TLI.getPointerTy(DAG.getDataLayout()))));
2561 void SelectionDAGBuilder::visitFPExt(const User &I) {
2562 // FPExt is never a no-op cast, no need to check
2563 SDValue N = getValue(I.getOperand(0));
2564 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2566 setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurSDLoc(), DestVT, N));
2569 void SelectionDAGBuilder::visitFPToUI(const User &I) {
2570 // FPToUI is never a no-op cast, no need to check
2571 SDValue N = getValue(I.getOperand(0));
2572 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2574 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurSDLoc(), DestVT, N));
2577 void SelectionDAGBuilder::visitFPToSI(const User &I) {
2578 // FPToSI is never a no-op cast, no need to check
2579 SDValue N = getValue(I.getOperand(0));
2580 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2582 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurSDLoc(), DestVT, N));
2585 void SelectionDAGBuilder::visitUIToFP(const User &I) {
2586 // UIToFP is never a no-op cast, no need to check
2587 SDValue N = getValue(I.getOperand(0));
2588 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2590 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurSDLoc(), DestVT, N));
2593 void SelectionDAGBuilder::visitSIToFP(const User &I) {
2594 // SIToFP is never a no-op cast, no need to check
2595 SDValue N = getValue(I.getOperand(0));
2596 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2598 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurSDLoc(), DestVT, N));
2601 void SelectionDAGBuilder::visitPtrToInt(const User &I) {
2602 // What to do depends on the size of the integer and the size of the pointer.
2603 // We can either truncate, zero extend, or no-op, accordingly.
2604 SDValue N = getValue(I.getOperand(0));
2605 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2607 setValue(&I, DAG.getZExtOrTrunc(N, getCurSDLoc(), DestVT));
2610 void SelectionDAGBuilder::visitIntToPtr(const User &I) {
2611 // What to do depends on the size of the integer and the size of the pointer.
2612 // We can either truncate, zero extend, or no-op, accordingly.
2613 SDValue N = getValue(I.getOperand(0));
2614 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2616 setValue(&I, DAG.getZExtOrTrunc(N, getCurSDLoc(), DestVT));
2619 void SelectionDAGBuilder::visitBitCast(const User &I) {
2620 SDValue N = getValue(I.getOperand(0));
2621 SDLoc dl = getCurSDLoc();
2622 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2625 // BitCast assures us that source and destination are the same size so this is
2626 // either a BITCAST or a no-op.
2627 if (DestVT != N.getValueType())
2628 setValue(&I, DAG.getNode(ISD::BITCAST, dl,
2629 DestVT, N)); // convert types.
2630 // Check if the original LLVM IR Operand was a ConstantInt, because getValue()
2631 // might fold any kind of constant expression to an integer constant and that
2632 // is not what we are looking for. Only regcognize a bitcast of a genuine
2633 // constant integer as an opaque constant.
2634 else if(ConstantInt *C = dyn_cast<ConstantInt>(I.getOperand(0)))
2635 setValue(&I, DAG.getConstant(C->getValue(), dl, DestVT, /*isTarget=*/false,
2638 setValue(&I, N); // noop cast.
2641 void SelectionDAGBuilder::visitAddrSpaceCast(const User &I) {
2642 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2643 const Value *SV = I.getOperand(0);
2644 SDValue N = getValue(SV);
2645 EVT DestVT = TLI.getValueType(DAG.getDataLayout(), I.getType());
2647 unsigned SrcAS = SV->getType()->getPointerAddressSpace();
2648 unsigned DestAS = I.getType()->getPointerAddressSpace();
2650 if (!TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
2651 N = DAG.getAddrSpaceCast(getCurSDLoc(), DestVT, N, SrcAS, DestAS);
2656 void SelectionDAGBuilder::visitInsertElement(const User &I) {
2657 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2658 SDValue InVec = getValue(I.getOperand(0));
2659 SDValue InVal = getValue(I.getOperand(1));
2660 SDValue InIdx = DAG.getSExtOrTrunc(getValue(I.getOperand(2)), getCurSDLoc(),
2661 TLI.getVectorIdxTy(DAG.getDataLayout()));
2662 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurSDLoc(),
2663 TLI.getValueType(DAG.getDataLayout(), I.getType()),
2664 InVec, InVal, InIdx));
2667 void SelectionDAGBuilder::visitExtractElement(const User &I) {
2668 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2669 SDValue InVec = getValue(I.getOperand(0));
2670 SDValue InIdx = DAG.getSExtOrTrunc(getValue(I.getOperand(1)), getCurSDLoc(),
2671 TLI.getVectorIdxTy(DAG.getDataLayout()));
2672 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurSDLoc(),
2673 TLI.getValueType(DAG.getDataLayout(), I.getType()),
2677 // Utility for visitShuffleVector - Return true if every element in Mask,
2678 // beginning from position Pos and ending in Pos+Size, falls within the
2679 // specified sequential range [L, L+Pos). or is undef.
2680 static bool isSequentialInRange(const SmallVectorImpl<int> &Mask,
2681 unsigned Pos, unsigned Size, int Low) {
2682 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
2683 if (Mask[i] >= 0 && Mask[i] != Low)
2688 void SelectionDAGBuilder::visitShuffleVector(const User &I) {
2689 SDValue Src1 = getValue(I.getOperand(0));
2690 SDValue Src2 = getValue(I.getOperand(1));
2692 SmallVector<int, 8> Mask;
2693 ShuffleVectorInst::getShuffleMask(cast<Constant>(I.getOperand(2)), Mask);
2694 unsigned MaskNumElts = Mask.size();
2696 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2697 EVT VT = TLI.getValueType(DAG.getDataLayout(), I.getType());
2698 EVT SrcVT = Src1.getValueType();
2699 unsigned SrcNumElts = SrcVT.getVectorNumElements();
2701 if (SrcNumElts == MaskNumElts) {
2702 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
2707 // Normalize the shuffle vector since mask and vector length don't match.
2708 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) {
2709 // Mask is longer than the source vectors and is a multiple of the source
2710 // vectors. We can use concatenate vector to make the mask and vectors
2712 if (SrcNumElts*2 == MaskNumElts) {
2713 // First check for Src1 in low and Src2 in high
2714 if (isSequentialInRange(Mask, 0, SrcNumElts, 0) &&
2715 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, SrcNumElts)) {
2716 // The shuffle is concatenating two vectors together.
2717 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurSDLoc(),
2721 // Then check for Src2 in low and Src1 in high
2722 if (isSequentialInRange(Mask, 0, SrcNumElts, SrcNumElts) &&
2723 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, 0)) {
2724 // The shuffle is concatenating two vectors together.
2725 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurSDLoc(),
2731 // Pad both vectors with undefs to make them the same length as the mask.
2732 unsigned NumConcat = MaskNumElts / SrcNumElts;
2733 bool Src1U = Src1.getOpcode() == ISD::UNDEF;
2734 bool Src2U = Src2.getOpcode() == ISD::UNDEF;
2735 SDValue UndefVal = DAG.getUNDEF(SrcVT);
2737 SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal);
2738 SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal);
2742 Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2743 getCurSDLoc(), VT, MOps1);
2744 Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2745 getCurSDLoc(), VT, MOps2);
2747 // Readjust mask for new input vector length.
2748 SmallVector<int, 8> MappedOps;
2749 for (unsigned i = 0; i != MaskNumElts; ++i) {
2751 if (Idx >= (int)SrcNumElts)
2752 Idx -= SrcNumElts - MaskNumElts;
2753 MappedOps.push_back(Idx);
2756 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
2761 if (SrcNumElts > MaskNumElts) {
2762 // Analyze the access pattern of the vector to see if we can extract
2763 // two subvectors and do the shuffle. The analysis is done by calculating
2764 // the range of elements the mask access on both vectors.
2765 int MinRange[2] = { static_cast<int>(SrcNumElts),
2766 static_cast<int>(SrcNumElts)};
2767 int MaxRange[2] = {-1, -1};
2769 for (unsigned i = 0; i != MaskNumElts; ++i) {
2775 if (Idx >= (int)SrcNumElts) {
2779 if (Idx > MaxRange[Input])
2780 MaxRange[Input] = Idx;
2781 if (Idx < MinRange[Input])
2782 MinRange[Input] = Idx;
2785 // Check if the access is smaller than the vector size and can we find
2786 // a reasonable extract index.
2787 int RangeUse[2] = { -1, -1 }; // 0 = Unused, 1 = Extract, -1 = Can not
2789 int StartIdx[2]; // StartIdx to extract from
2790 for (unsigned Input = 0; Input < 2; ++Input) {
2791 if (MinRange[Input] >= (int)SrcNumElts && MaxRange[Input] < 0) {
2792 RangeUse[Input] = 0; // Unused
2793 StartIdx[Input] = 0;
2797 // Find a good start index that is a multiple of the mask length. Then
2798 // see if the rest of the elements are in range.
2799 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts;
2800 if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts &&
2801 StartIdx[Input] + MaskNumElts <= SrcNumElts)
2802 RangeUse[Input] = 1; // Extract from a multiple of the mask length.
2805 if (RangeUse[0] == 0 && RangeUse[1] == 0) {
2806 setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used.
2809 if (RangeUse[0] >= 0 && RangeUse[1] >= 0) {
2810 // Extract appropriate subvector and generate a vector shuffle
2811 for (unsigned Input = 0; Input < 2; ++Input) {
2812 SDValue &Src = Input == 0 ? Src1 : Src2;
2813 if (RangeUse[Input] == 0)
2814 Src = DAG.getUNDEF(VT);
2816 SDLoc dl = getCurSDLoc();
2818 ISD::EXTRACT_SUBVECTOR, dl, VT, Src,
2819 DAG.getConstant(StartIdx[Input], dl,
2820 TLI.getVectorIdxTy(DAG.getDataLayout())));
2824 // Calculate new mask.
2825 SmallVector<int, 8> MappedOps;
2826 for (unsigned i = 0; i != MaskNumElts; ++i) {
2829 if (Idx < (int)SrcNumElts)
2832 Idx -= SrcNumElts + StartIdx[1] - MaskNumElts;
2834 MappedOps.push_back(Idx);
2837 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
2843 // We can't use either concat vectors or extract subvectors so fall back to
2844 // replacing the shuffle with extract and build vector.
2845 // to insert and build vector.
2846 EVT EltVT = VT.getVectorElementType();
2847 EVT IdxVT = TLI.getVectorIdxTy(DAG.getDataLayout());
2848 SDLoc dl = getCurSDLoc();
2849 SmallVector<SDValue,8> Ops;
2850 for (unsigned i = 0; i != MaskNumElts; ++i) {
2855 Res = DAG.getUNDEF(EltVT);
2857 SDValue &Src = Idx < (int)SrcNumElts ? Src1 : Src2;
2858 if (Idx >= (int)SrcNumElts) Idx -= SrcNumElts;
2860 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
2861 EltVT, Src, DAG.getConstant(Idx, dl, IdxVT));
2867 setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops));
2870 void SelectionDAGBuilder::visitInsertValue(const InsertValueInst &I) {
2871 const Value *Op0 = I.getOperand(0);
2872 const Value *Op1 = I.getOperand(1);
2873 Type *AggTy = I.getType();
2874 Type *ValTy = Op1->getType();
2875 bool IntoUndef = isa<UndefValue>(Op0);
2876 bool FromUndef = isa<UndefValue>(Op1);
2878 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
2880 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2881 SmallVector<EVT, 4> AggValueVTs;
2882 ComputeValueVTs(TLI, DAG.getDataLayout(), AggTy, AggValueVTs);
2883 SmallVector<EVT, 4> ValValueVTs;
2884 ComputeValueVTs(TLI, DAG.getDataLayout(), ValTy, ValValueVTs);
2886 unsigned NumAggValues = AggValueVTs.size();
2887 unsigned NumValValues = ValValueVTs.size();
2888 SmallVector<SDValue, 4> Values(NumAggValues);
2890 // Ignore an insertvalue that produces an empty object
2891 if (!NumAggValues) {
2892 setValue(&I, DAG.getUNDEF(MVT(MVT::Other)));
2896 SDValue Agg = getValue(Op0);
2898 // Copy the beginning value(s) from the original aggregate.
2899 for (; i != LinearIndex; ++i)
2900 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2901 SDValue(Agg.getNode(), Agg.getResNo() + i);
2902 // Copy values from the inserted value(s).
2904 SDValue Val = getValue(Op1);
2905 for (; i != LinearIndex + NumValValues; ++i)
2906 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2907 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex);
2909 // Copy remaining value(s) from the original aggregate.
2910 for (; i != NumAggValues; ++i)
2911 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2912 SDValue(Agg.getNode(), Agg.getResNo() + i);
2914 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
2915 DAG.getVTList(AggValueVTs), Values));
2918 void SelectionDAGBuilder::visitExtractValue(const ExtractValueInst &I) {
2919 const Value *Op0 = I.getOperand(0);
2920 Type *AggTy = Op0->getType();
2921 Type *ValTy = I.getType();
2922 bool OutOfUndef = isa<UndefValue>(Op0);
2924 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
2926 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2927 SmallVector<EVT, 4> ValValueVTs;
2928 ComputeValueVTs(TLI, DAG.getDataLayout(), ValTy, ValValueVTs);
2930 unsigned NumValValues = ValValueVTs.size();
2932 // Ignore a extractvalue that produces an empty object
2933 if (!NumValValues) {
2934 setValue(&I, DAG.getUNDEF(MVT(MVT::Other)));
2938 SmallVector<SDValue, 4> Values(NumValValues);
2940 SDValue Agg = getValue(Op0);
2941 // Copy out the selected value(s).
2942 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i)
2943 Values[i - LinearIndex] =
2945 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) :
2946 SDValue(Agg.getNode(), Agg.getResNo() + i);
2948 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
2949 DAG.getVTList(ValValueVTs), Values));
2952 void SelectionDAGBuilder::visitGetElementPtr(const User &I) {
2953 Value *Op0 = I.getOperand(0);
2954 // Note that the pointer operand may be a vector of pointers. Take the scalar
2955 // element which holds a pointer.
2956 Type *Ty = Op0->getType()->getScalarType();
2957 unsigned AS = Ty->getPointerAddressSpace();
2958 SDValue N = getValue(Op0);
2959 SDLoc dl = getCurSDLoc();
2961 // Normalize Vector GEP - all scalar operands should be converted to the
2963 unsigned VectorWidth = I.getType()->isVectorTy() ?
2964 cast<VectorType>(I.getType())->getVectorNumElements() : 0;
2966 if (VectorWidth && !N.getValueType().isVector()) {
2967 MVT VT = MVT::getVectorVT(N.getValueType().getSimpleVT(), VectorWidth);
2968 SmallVector<SDValue, 16> Ops(VectorWidth, N);
2969 N = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
2971 for (GetElementPtrInst::const_op_iterator OI = I.op_begin()+1, E = I.op_end();
2973 const Value *Idx = *OI;
2974 if (StructType *StTy = dyn_cast<StructType>(Ty)) {
2975 unsigned Field = cast<Constant>(Idx)->getUniqueInteger().getZExtValue();
2978 uint64_t Offset = DL->getStructLayout(StTy)->getElementOffset(Field);
2979 N = DAG.getNode(ISD::ADD, dl, N.getValueType(), N,
2980 DAG.getConstant(Offset, dl, N.getValueType()));
2983 Ty = StTy->getElementType(Field);
2985 Ty = cast<SequentialType>(Ty)->getElementType();
2987 DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout(), AS);
2988 unsigned PtrSize = PtrTy.getSizeInBits();
2989 APInt ElementSize(PtrSize, DL->getTypeAllocSize(Ty));
2991 // If this is a scalar constant or a splat vector of constants,
2992 // handle it quickly.
2993 const auto *CI = dyn_cast<ConstantInt>(Idx);
2994 if (!CI && isa<ConstantDataVector>(Idx) &&
2995 cast<ConstantDataVector>(Idx)->getSplatValue())
2996 CI = cast<ConstantInt>(cast<ConstantDataVector>(Idx)->getSplatValue());
3001 APInt Offs = ElementSize * CI->getValue().sextOrTrunc(PtrSize);
3002 SDValue OffsVal = VectorWidth ?
3003 DAG.getConstant(Offs, dl, MVT::getVectorVT(PtrTy, VectorWidth)) :
3004 DAG.getConstant(Offs, dl, PtrTy);
3005 N = DAG.getNode(ISD::ADD, dl, N.getValueType(), N, OffsVal);
3009 // N = N + Idx * ElementSize;
3010 SDValue IdxN = getValue(Idx);
3012 if (!IdxN.getValueType().isVector() && VectorWidth) {
3013 MVT VT = MVT::getVectorVT(IdxN.getValueType().getSimpleVT(), VectorWidth);
3014 SmallVector<SDValue, 16> Ops(VectorWidth, IdxN);
3015 IdxN = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
3017 // If the index is smaller or larger than intptr_t, truncate or extend
3019 IdxN = DAG.getSExtOrTrunc(IdxN, dl, N.getValueType());
3021 // If this is a multiply by a power of two, turn it into a shl
3022 // immediately. This is a very common case.
3023 if (ElementSize != 1) {
3024 if (ElementSize.isPowerOf2()) {
3025 unsigned Amt = ElementSize.logBase2();
3026 IdxN = DAG.getNode(ISD::SHL, dl,
3027 N.getValueType(), IdxN,
3028 DAG.getConstant(Amt, dl, IdxN.getValueType()));
3030 SDValue Scale = DAG.getConstant(ElementSize, dl, IdxN.getValueType());
3031 IdxN = DAG.getNode(ISD::MUL, dl,
3032 N.getValueType(), IdxN, Scale);
3036 N = DAG.getNode(ISD::ADD, dl,
3037 N.getValueType(), N, IdxN);
3044 void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) {
3045 // If this is a fixed sized alloca in the entry block of the function,
3046 // allocate it statically on the stack.
3047 if (FuncInfo.StaticAllocaMap.count(&I))
3048 return; // getValue will auto-populate this.
3050 SDLoc dl = getCurSDLoc();
3051 Type *Ty = I.getAllocatedType();
3052 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3053 auto &DL = DAG.getDataLayout();
3054 uint64_t TySize = DL.getTypeAllocSize(Ty);
3056 std::max((unsigned)DL.getPrefTypeAlignment(Ty), I.getAlignment());
3058 SDValue AllocSize = getValue(I.getArraySize());
3060 EVT IntPtr = TLI.getPointerTy(DAG.getDataLayout());
3061 if (AllocSize.getValueType() != IntPtr)
3062 AllocSize = DAG.getZExtOrTrunc(AllocSize, dl, IntPtr);
3064 AllocSize = DAG.getNode(ISD::MUL, dl, IntPtr,
3066 DAG.getConstant(TySize, dl, IntPtr));
3068 // Handle alignment. If the requested alignment is less than or equal to
3069 // the stack alignment, ignore it. If the size is greater than or equal to
3070 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
3071 unsigned StackAlign =
3072 DAG.getSubtarget().getFrameLowering()->getStackAlignment();
3073 if (Align <= StackAlign)
3076 // Round the size of the allocation up to the stack alignment size
3077 // by add SA-1 to the size.
3078 AllocSize = DAG.getNode(ISD::ADD, dl,
3079 AllocSize.getValueType(), AllocSize,
3080 DAG.getIntPtrConstant(StackAlign - 1, dl));
3082 // Mask out the low bits for alignment purposes.
3083 AllocSize = DAG.getNode(ISD::AND, dl,
3084 AllocSize.getValueType(), AllocSize,
3085 DAG.getIntPtrConstant(~(uint64_t)(StackAlign - 1),
3088 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align, dl) };
3089 SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other);
3090 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, dl, VTs, Ops);
3092 DAG.setRoot(DSA.getValue(1));
3094 assert(FuncInfo.MF->getFrameInfo()->hasVarSizedObjects());
3097 void SelectionDAGBuilder::visitLoad(const LoadInst &I) {
3099 return visitAtomicLoad(I);
3101 const Value *SV = I.getOperand(0);
3102 SDValue Ptr = getValue(SV);
3104 Type *Ty = I.getType();
3106 bool isVolatile = I.isVolatile();
3107 bool isNonTemporal = I.getMetadata(LLVMContext::MD_nontemporal) != nullptr;
3109 // The IR notion of invariant_load only guarantees that all *non-faulting*
3110 // invariant loads result in the same value. The MI notion of invariant load
3111 // guarantees that the load can be legally moved to any location within its
3112 // containing function. The MI notion of invariant_load is stronger than the
3113 // IR notion of invariant_load -- an MI invariant_load is an IR invariant_load
3114 // with a guarantee that the location being loaded from is dereferenceable
3115 // throughout the function's lifetime.
3117 bool isInvariant = I.getMetadata(LLVMContext::MD_invariant_load) != nullptr &&
3118 isDereferenceablePointer(SV, DAG.getDataLayout());
3119 unsigned Alignment = I.getAlignment();
3122 I.getAAMetadata(AAInfo);
3123 const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range);
3125 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3126 SmallVector<EVT, 4> ValueVTs;
3127 SmallVector<uint64_t, 4> Offsets;
3128 ComputeValueVTs(TLI, DAG.getDataLayout(), Ty, ValueVTs, &Offsets);
3129 unsigned NumValues = ValueVTs.size();
3134 bool ConstantMemory = false;
3135 if (isVolatile || NumValues > MaxParallelChains)
3136 // Serialize volatile loads with other side effects.
3138 else if (AA->pointsToConstantMemory(MemoryLocation(
3139 SV, DAG.getDataLayout().getTypeStoreSize(Ty), AAInfo))) {
3140 // Do not serialize (non-volatile) loads of constant memory with anything.
3141 Root = DAG.getEntryNode();
3142 ConstantMemory = true;
3144 // Do not serialize non-volatile loads against each other.
3145 Root = DAG.getRoot();
3148 SDLoc dl = getCurSDLoc();
3151 Root = TLI.prepareVolatileOrAtomicLoad(Root, dl, DAG);
3153 SmallVector<SDValue, 4> Values(NumValues);
3154 SmallVector<SDValue, 4> Chains(std::min(MaxParallelChains, NumValues));
3155 EVT PtrVT = Ptr.getValueType();
3156 unsigned ChainI = 0;
3157 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3158 // Serializing loads here may result in excessive register pressure, and
3159 // TokenFactor places arbitrary choke points on the scheduler. SD scheduling
3160 // could recover a bit by hoisting nodes upward in the chain by recognizing
3161 // they are side-effect free or do not alias. The optimizer should really
3162 // avoid this case by converting large object/array copies to llvm.memcpy
3163 // (MaxParallelChains should always remain as failsafe).
3164 if (ChainI == MaxParallelChains) {
3165 assert(PendingLoads.empty() && "PendingLoads must be serialized first");
3166 SDValue Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3167 makeArrayRef(Chains.data(), ChainI));
3171 SDValue A = DAG.getNode(ISD::ADD, dl,
3173 DAG.getConstant(Offsets[i], dl, PtrVT));
3174 SDValue L = DAG.getLoad(ValueVTs[i], dl, Root,
3175 A, MachinePointerInfo(SV, Offsets[i]), isVolatile,
3176 isNonTemporal, isInvariant, Alignment, AAInfo,
3180 Chains[ChainI] = L.getValue(1);
3183 if (!ConstantMemory) {
3184 SDValue Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3185 makeArrayRef(Chains.data(), ChainI));
3189 PendingLoads.push_back(Chain);
3192 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, dl,
3193 DAG.getVTList(ValueVTs), Values));
3196 void SelectionDAGBuilder::visitStore(const StoreInst &I) {
3198 return visitAtomicStore(I);
3200 const Value *SrcV = I.getOperand(0);
3201 const Value *PtrV = I.getOperand(1);
3203 SmallVector<EVT, 4> ValueVTs;
3204 SmallVector<uint64_t, 4> Offsets;
3205 ComputeValueVTs(DAG.getTargetLoweringInfo(), DAG.getDataLayout(),
3206 SrcV->getType(), ValueVTs, &Offsets);
3207 unsigned NumValues = ValueVTs.size();
3211 // Get the lowered operands. Note that we do this after
3212 // checking if NumResults is zero, because with zero results
3213 // the operands won't have values in the map.
3214 SDValue Src = getValue(SrcV);
3215 SDValue Ptr = getValue(PtrV);
3217 SDValue Root = getRoot();
3218 SmallVector<SDValue, 4> Chains(std::min(MaxParallelChains, NumValues));
3219 EVT PtrVT = Ptr.getValueType();
3220 bool isVolatile = I.isVolatile();
3221 bool isNonTemporal = I.getMetadata(LLVMContext::MD_nontemporal) != nullptr;
3222 unsigned Alignment = I.getAlignment();
3223 SDLoc dl = getCurSDLoc();
3226 I.getAAMetadata(AAInfo);
3228 unsigned ChainI = 0;
3229 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3230 // See visitLoad comments.
3231 if (ChainI == MaxParallelChains) {
3232 SDValue Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3233 makeArrayRef(Chains.data(), ChainI));
3237 SDValue Add = DAG.getNode(ISD::ADD, dl, PtrVT, Ptr,
3238 DAG.getConstant(Offsets[i], dl, PtrVT));
3239 SDValue St = DAG.getStore(Root, dl,
3240 SDValue(Src.getNode(), Src.getResNo() + i),
3241 Add, MachinePointerInfo(PtrV, Offsets[i]),
3242 isVolatile, isNonTemporal, Alignment, AAInfo);
3243 Chains[ChainI] = St;
3246 SDValue StoreNode = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3247 makeArrayRef(Chains.data(), ChainI));
3248 DAG.setRoot(StoreNode);
3251 void SelectionDAGBuilder::visitMaskedStore(const CallInst &I) {
3252 SDLoc sdl = getCurSDLoc();
3254 // llvm.masked.store.*(Src0, Ptr, alignment, Mask)
3255 Value *PtrOperand = I.getArgOperand(1);
3256 SDValue Ptr = getValue(PtrOperand);
3257 SDValue Src0 = getValue(I.getArgOperand(0));
3258 SDValue Mask = getValue(I.getArgOperand(3));
3259 EVT VT = Src0.getValueType();
3260 unsigned Alignment = (cast<ConstantInt>(I.getArgOperand(2)))->getZExtValue();
3262 Alignment = DAG.getEVTAlignment(VT);
3265 I.getAAMetadata(AAInfo);
3267 MachineMemOperand *MMO =
3268 DAG.getMachineFunction().
3269 getMachineMemOperand(MachinePointerInfo(PtrOperand),
3270 MachineMemOperand::MOStore, VT.getStoreSize(),
3272 SDValue StoreNode = DAG.getMaskedStore(getRoot(), sdl, Src0, Ptr, Mask, VT,
3274 DAG.setRoot(StoreNode);
3275 setValue(&I, StoreNode);
3278 // Get a uniform base for the Gather/Scatter intrinsic.
3279 // The first argument of the Gather/Scatter intrinsic is a vector of pointers.
3280 // We try to represent it as a base pointer + vector of indices.
3281 // Usually, the vector of pointers comes from a 'getelementptr' instruction.
3282 // The first operand of the GEP may be a single pointer or a vector of pointers
3284 // %gep.ptr = getelementptr i32, <8 x i32*> %vptr, <8 x i32> %ind
3286 // %gep.ptr = getelementptr i32, i32* %ptr, <8 x i32> %ind
3287 // %res = call <8 x i32> @llvm.masked.gather.v8i32(<8 x i32*> %gep.ptr, ..
3289 // When the first GEP operand is a single pointer - it is the uniform base we
3290 // are looking for. If first operand of the GEP is a splat vector - we
3291 // extract the spalt value and use it as a uniform base.
3292 // In all other cases the function returns 'false'.
3294 static bool getUniformBase(Value *& Ptr, SDValue& Base, SDValue& Index,
3295 SelectionDAGBuilder* SDB) {
3297 SelectionDAG& DAG = SDB->DAG;
3298 LLVMContext &Context = *DAG.getContext();
3300 assert(Ptr->getType()->isVectorTy() && "Uexpected pointer type");
3301 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3302 if (!GEP || GEP->getNumOperands() > 2)
3305 Value *GEPPtr = GEP->getPointerOperand();
3306 if (!GEPPtr->getType()->isVectorTy())
3308 else if (!(Ptr = getSplatValue(GEPPtr)))
3311 Value *IndexVal = GEP->getOperand(1);
3313 // The operands of the GEP may be defined in another basic block.
3314 // In this case we'll not find nodes for the operands.
3315 if (!SDB->findValue(Ptr) || !SDB->findValue(IndexVal))
3318 Base = SDB->getValue(Ptr);
3319 Index = SDB->getValue(IndexVal);
3321 // Suppress sign extension.
3322 if (SExtInst* Sext = dyn_cast<SExtInst>(IndexVal)) {
3323 if (SDB->findValue(Sext->getOperand(0))) {
3324 IndexVal = Sext->getOperand(0);
3325 Index = SDB->getValue(IndexVal);
3328 if (!Index.getValueType().isVector()) {
3329 unsigned GEPWidth = GEP->getType()->getVectorNumElements();
3330 EVT VT = EVT::getVectorVT(Context, Index.getValueType(), GEPWidth);
3331 SmallVector<SDValue, 16> Ops(GEPWidth, Index);
3332 Index = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Index), VT, Ops);
3337 void SelectionDAGBuilder::visitMaskedScatter(const CallInst &I) {
3338 SDLoc sdl = getCurSDLoc();
3340 // llvm.masked.scatter.*(Src0, Ptrs, alignemt, Mask)
3341 Value *Ptr = I.getArgOperand(1);
3342 SDValue Src0 = getValue(I.getArgOperand(0));
3343 SDValue Mask = getValue(I.getArgOperand(3));
3344 EVT VT = Src0.getValueType();
3345 unsigned Alignment = (cast<ConstantInt>(I.getArgOperand(2)))->getZExtValue();
3347 Alignment = DAG.getEVTAlignment(VT);
3348 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3351 I.getAAMetadata(AAInfo);
3355 Value *BasePtr = Ptr;
3356 bool UniformBase = getUniformBase(BasePtr, Base, Index, this);
3358 Value *MemOpBasePtr = UniformBase ? BasePtr : nullptr;
3359 MachineMemOperand *MMO = DAG.getMachineFunction().
3360 getMachineMemOperand(MachinePointerInfo(MemOpBasePtr),
3361 MachineMemOperand::MOStore, VT.getStoreSize(),
3364 Base = DAG.getTargetConstant(0, sdl, TLI.getPointerTy(DAG.getDataLayout()));
3365 Index = getValue(Ptr);
3367 SDValue Ops[] = { getRoot(), Src0, Mask, Base, Index };
3368 SDValue Scatter = DAG.getMaskedScatter(DAG.getVTList(MVT::Other), VT, sdl,
3370 DAG.setRoot(Scatter);
3371 setValue(&I, Scatter);
3374 void SelectionDAGBuilder::visitMaskedLoad(const CallInst &I) {
3375 SDLoc sdl = getCurSDLoc();
3377 // @llvm.masked.load.*(Ptr, alignment, Mask, Src0)
3378 Value *PtrOperand = I.getArgOperand(0);
3379 SDValue Ptr = getValue(PtrOperand);
3380 SDValue Src0 = getValue(I.getArgOperand(3));
3381 SDValue Mask = getValue(I.getArgOperand(2));
3383 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3384 EVT VT = TLI.getValueType(DAG.getDataLayout(), I.getType());
3385 unsigned Alignment = (cast<ConstantInt>(I.getArgOperand(1)))->getZExtValue();
3387 Alignment = DAG.getEVTAlignment(VT);
3390 I.getAAMetadata(AAInfo);
3391 const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range);
3393 SDValue InChain = DAG.getRoot();
3394 if (AA->pointsToConstantMemory(MemoryLocation(
3395 PtrOperand, DAG.getDataLayout().getTypeStoreSize(I.getType()),
3397 // Do not serialize (non-volatile) loads of constant memory with anything.
3398 InChain = DAG.getEntryNode();
3401 MachineMemOperand *MMO =
3402 DAG.getMachineFunction().
3403 getMachineMemOperand(MachinePointerInfo(PtrOperand),
3404 MachineMemOperand::MOLoad, VT.getStoreSize(),
3405 Alignment, AAInfo, Ranges);
3407 SDValue Load = DAG.getMaskedLoad(VT, sdl, InChain, Ptr, Mask, Src0, VT, MMO,
3409 SDValue OutChain = Load.getValue(1);
3410 DAG.setRoot(OutChain);
3414 void SelectionDAGBuilder::visitMaskedGather(const CallInst &I) {
3415 SDLoc sdl = getCurSDLoc();
3417 // @llvm.masked.gather.*(Ptrs, alignment, Mask, Src0)
3418 Value *Ptr = I.getArgOperand(0);
3419 SDValue Src0 = getValue(I.getArgOperand(3));
3420 SDValue Mask = getValue(I.getArgOperand(2));
3422 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3423 EVT VT = TLI.getValueType(DAG.getDataLayout(), I.getType());
3424 unsigned Alignment = (cast<ConstantInt>(I.getArgOperand(1)))->getZExtValue();
3426 Alignment = DAG.getEVTAlignment(VT);
3429 I.getAAMetadata(AAInfo);
3430 const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range);
3432 SDValue Root = DAG.getRoot();
3435 Value *BasePtr = Ptr;
3436 bool UniformBase = getUniformBase(BasePtr, Base, Index, this);
3437 bool ConstantMemory = false;
3439 AA->pointsToConstantMemory(MemoryLocation(
3440 BasePtr, DAG.getDataLayout().getTypeStoreSize(I.getType()),
3442 // Do not serialize (non-volatile) loads of constant memory with anything.
3443 Root = DAG.getEntryNode();
3444 ConstantMemory = true;
3447 MachineMemOperand *MMO =
3448 DAG.getMachineFunction().
3449 getMachineMemOperand(MachinePointerInfo(UniformBase ? BasePtr : nullptr),
3450 MachineMemOperand::MOLoad, VT.getStoreSize(),
3451 Alignment, AAInfo, Ranges);
3454 Base = DAG.getTargetConstant(0, sdl, TLI.getPointerTy(DAG.getDataLayout()));
3455 Index = getValue(Ptr);
3457 SDValue Ops[] = { Root, Src0, Mask, Base, Index };
3458 SDValue Gather = DAG.getMaskedGather(DAG.getVTList(VT, MVT::Other), VT, sdl,
3461 SDValue OutChain = Gather.getValue(1);
3462 if (!ConstantMemory)
3463 PendingLoads.push_back(OutChain);
3464 setValue(&I, Gather);
3467 void SelectionDAGBuilder::visitAtomicCmpXchg(const AtomicCmpXchgInst &I) {
3468 SDLoc dl = getCurSDLoc();
3469 AtomicOrdering SuccessOrder = I.getSuccessOrdering();
3470 AtomicOrdering FailureOrder = I.getFailureOrdering();
3471 SynchronizationScope Scope = I.getSynchScope();
3473 SDValue InChain = getRoot();
3475 MVT MemVT = getValue(I.getCompareOperand()).getSimpleValueType();
3476 SDVTList VTs = DAG.getVTList(MemVT, MVT::i1, MVT::Other);
3477 SDValue L = DAG.getAtomicCmpSwap(
3478 ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, dl, MemVT, VTs, InChain,
3479 getValue(I.getPointerOperand()), getValue(I.getCompareOperand()),
3480 getValue(I.getNewValOperand()), MachinePointerInfo(I.getPointerOperand()),
3481 /*Alignment=*/ 0, SuccessOrder, FailureOrder, Scope);
3483 SDValue OutChain = L.getValue(2);
3486 DAG.setRoot(OutChain);
3489 void SelectionDAGBuilder::visitAtomicRMW(const AtomicRMWInst &I) {
3490 SDLoc dl = getCurSDLoc();
3492 switch (I.getOperation()) {
3493 default: llvm_unreachable("Unknown atomicrmw operation");
3494 case AtomicRMWInst::Xchg: NT = ISD::ATOMIC_SWAP; break;
3495 case AtomicRMWInst::Add: NT = ISD::ATOMIC_LOAD_ADD; break;
3496 case AtomicRMWInst::Sub: NT = ISD::ATOMIC_LOAD_SUB; break;
3497 case AtomicRMWInst::And: NT = ISD::ATOMIC_LOAD_AND; break;
3498 case AtomicRMWInst::Nand: NT = ISD::ATOMIC_LOAD_NAND; break;
3499 case AtomicRMWInst::Or: NT = ISD::ATOMIC_LOAD_OR; break;
3500 case AtomicRMWInst::Xor: NT = ISD::ATOMIC_LOAD_XOR; break;
3501 case AtomicRMWInst::Max: NT = ISD::ATOMIC_LOAD_MAX; break;
3502 case AtomicRMWInst::Min: NT = ISD::ATOMIC_LOAD_MIN; break;
3503 case AtomicRMWInst::UMax: NT = ISD::ATOMIC_LOAD_UMAX; break;
3504 case AtomicRMWInst::UMin: NT = ISD::ATOMIC_LOAD_UMIN; break;
3506 AtomicOrdering Order = I.getOrdering();
3507 SynchronizationScope Scope = I.getSynchScope();
3509 SDValue InChain = getRoot();
3512 DAG.getAtomic(NT, dl,
3513 getValue(I.getValOperand()).getSimpleValueType(),
3515 getValue(I.getPointerOperand()),
3516 getValue(I.getValOperand()),
3517 I.getPointerOperand(),
3518 /* Alignment=*/ 0, Order, Scope);
3520 SDValue OutChain = L.getValue(1);
3523 DAG.setRoot(OutChain);
3526 void SelectionDAGBuilder::visitFence(const FenceInst &I) {
3527 SDLoc dl = getCurSDLoc();
3528 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3531 Ops[1] = DAG.getConstant(I.getOrdering(), dl,
3532 TLI.getPointerTy(DAG.getDataLayout()));
3533 Ops[2] = DAG.getConstant(I.getSynchScope(), dl,
3534 TLI.getPointerTy(DAG.getDataLayout()));
3535 DAG.setRoot(DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops));
3538 void SelectionDAGBuilder::visitAtomicLoad(const LoadInst &I) {
3539 SDLoc dl = getCurSDLoc();
3540 AtomicOrdering Order = I.getOrdering();
3541 SynchronizationScope Scope = I.getSynchScope();
3543 SDValue InChain = getRoot();
3545 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3546 EVT VT = TLI.getValueType(DAG.getDataLayout(), I.getType());
3548 if (I.getAlignment() < VT.getSizeInBits() / 8)
3549 report_fatal_error("Cannot generate unaligned atomic load");
3551 MachineMemOperand *MMO =
3552 DAG.getMachineFunction().
3553 getMachineMemOperand(MachinePointerInfo(I.getPointerOperand()),
3554 MachineMemOperand::MOVolatile |
3555 MachineMemOperand::MOLoad,
3557 I.getAlignment() ? I.getAlignment() :
3558 DAG.getEVTAlignment(VT));
3560 InChain = TLI.prepareVolatileOrAtomicLoad(InChain, dl, DAG);
3562 DAG.getAtomic(ISD::ATOMIC_LOAD, dl, VT, VT, InChain,
3563 getValue(I.getPointerOperand()), MMO,
3566 SDValue OutChain = L.getValue(1);
3569 DAG.setRoot(OutChain);
3572 void SelectionDAGBuilder::visitAtomicStore(const StoreInst &I) {
3573 SDLoc dl = getCurSDLoc();
3575 AtomicOrdering Order = I.getOrdering();
3576 SynchronizationScope Scope = I.getSynchScope();
3578 SDValue InChain = getRoot();
3580 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3582 TLI.getValueType(DAG.getDataLayout(), I.getValueOperand()->getType());
3584 if (I.getAlignment() < VT.getSizeInBits() / 8)
3585 report_fatal_error("Cannot generate unaligned atomic store");
3588 DAG.getAtomic(ISD::ATOMIC_STORE, dl, VT,
3590 getValue(I.getPointerOperand()),
3591 getValue(I.getValueOperand()),
3592 I.getPointerOperand(), I.getAlignment(),
3595 DAG.setRoot(OutChain);
3598 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
3600 void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I,
3601 unsigned Intrinsic) {
3602 bool HasChain = !I.doesNotAccessMemory();
3603 bool OnlyLoad = HasChain && I.onlyReadsMemory();
3605 // Build the operand list.
3606 SmallVector<SDValue, 8> Ops;
3607 if (HasChain) { // If this intrinsic has side-effects, chainify it.
3609 // We don't need to serialize loads against other loads.
3610 Ops.push_back(DAG.getRoot());
3612 Ops.push_back(getRoot());
3616 // Info is set by getTgtMemInstrinsic
3617 TargetLowering::IntrinsicInfo Info;
3618 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3619 bool IsTgtIntrinsic = TLI.getTgtMemIntrinsic(Info, I, Intrinsic);
3621 // Add the intrinsic ID as an integer operand if it's not a target intrinsic.
3622 if (!IsTgtIntrinsic || Info.opc == ISD::INTRINSIC_VOID ||
3623 Info.opc == ISD::INTRINSIC_W_CHAIN)
3624 Ops.push_back(DAG.getTargetConstant(Intrinsic, getCurSDLoc(),
3625 TLI.getPointerTy(DAG.getDataLayout())));
3627 // Add all operands of the call to the operand list.
3628 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) {
3629 SDValue Op = getValue(I.getArgOperand(i));
3633 SmallVector<EVT, 4> ValueVTs;
3634 ComputeValueVTs(TLI, DAG.getDataLayout(), I.getType(), ValueVTs);
3637 ValueVTs.push_back(MVT::Other);
3639 SDVTList VTs = DAG.getVTList(ValueVTs);
3643 if (IsTgtIntrinsic) {
3644 // This is target intrinsic that touches memory
3645 Result = DAG.getMemIntrinsicNode(Info.opc, getCurSDLoc(),
3646 VTs, Ops, Info.memVT,
3647 MachinePointerInfo(Info.ptrVal, Info.offset),
3648 Info.align, Info.vol,
3649 Info.readMem, Info.writeMem, Info.size);
3650 } else if (!HasChain) {
3651 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurSDLoc(), VTs, Ops);
3652 } else if (!I.getType()->isVoidTy()) {
3653 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurSDLoc(), VTs, Ops);
3655 Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurSDLoc(), VTs, Ops);
3659 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1);
3661 PendingLoads.push_back(Chain);
3666 if (!I.getType()->isVoidTy()) {
3667 if (VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
3668 EVT VT = TLI.getValueType(DAG.getDataLayout(), PTy);
3669 Result = DAG.getNode(ISD::BITCAST, getCurSDLoc(), VT, Result);
3672 setValue(&I, Result);
3676 /// GetSignificand - Get the significand and build it into a floating-point
3677 /// number with exponent of 1:
3679 /// Op = (Op & 0x007fffff) | 0x3f800000;
3681 /// where Op is the hexadecimal representation of floating point value.
3683 GetSignificand(SelectionDAG &DAG, SDValue Op, SDLoc dl) {
3684 SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3685 DAG.getConstant(0x007fffff, dl, MVT::i32));
3686 SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1,
3687 DAG.getConstant(0x3f800000, dl, MVT::i32));
3688 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, t2);
3691 /// GetExponent - Get the exponent:
3693 /// (float)(int)(((Op & 0x7f800000) >> 23) - 127);
3695 /// where Op is the hexadecimal representation of floating point value.
3697 GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI,
3699 SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3700 DAG.getConstant(0x7f800000, dl, MVT::i32));
3701 SDValue t1 = DAG.getNode(
3702 ISD::SRL, dl, MVT::i32, t0,
3703 DAG.getConstant(23, dl, TLI.getPointerTy(DAG.getDataLayout())));
3704 SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1,
3705 DAG.getConstant(127, dl, MVT::i32));
3706 return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2);
3709 /// getF32Constant - Get 32-bit floating point constant.
3711 getF32Constant(SelectionDAG &DAG, unsigned Flt, SDLoc dl) {
3712 return DAG.getConstantFP(APFloat(APFloat::IEEEsingle, APInt(32, Flt)), dl,
3716 static SDValue getLimitedPrecisionExp2(SDValue t0, SDLoc dl,
3717 SelectionDAG &DAG) {
3718 // TODO: What fast-math-flags should be set on the floating-point nodes?
3720 // IntegerPartOfX = ((int32_t)(t0);
3721 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3723 // FractionalPartOfX = t0 - (float)IntegerPartOfX;
3724 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3725 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3727 // IntegerPartOfX <<= 23;
3728 IntegerPartOfX = DAG.getNode(
3729 ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3730 DAG.getConstant(23, dl, DAG.getTargetLoweringInfo().getPointerTy(
3731 DAG.getDataLayout())));
3733 SDValue TwoToFractionalPartOfX;
3734 if (LimitFloatPrecision <= 6) {
3735 // For floating-point precision of 6:
3737 // TwoToFractionalPartOfX =
3739 // (0.735607626f + 0.252464424f * x) * x;
3741 // error 0.0144103317, which is 6 bits
3742 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3743 getF32Constant(DAG, 0x3e814304, dl));
3744 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3745 getF32Constant(DAG, 0x3f3c50c8, dl));
3746 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3747 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3748 getF32Constant(DAG, 0x3f7f5e7e, dl));
3749 } else if (LimitFloatPrecision <= 12) {
3750 // For floating-point precision of 12:
3752 // TwoToFractionalPartOfX =
3755 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3757 // error 0.000107046256, which is 13 to 14 bits
3758 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3759 getF32Constant(DAG, 0x3da235e3, dl));
3760 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3761 getF32Constant(DAG, 0x3e65b8f3, dl));
3762 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3763 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3764 getF32Constant(DAG, 0x3f324b07, dl));
3765 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3766 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3767 getF32Constant(DAG, 0x3f7ff8fd, dl));
3768 } else { // LimitFloatPrecision <= 18
3769 // For floating-point precision of 18:
3771 // TwoToFractionalPartOfX =
3775 // (0.554906021e-1f +
3776 // (0.961591928e-2f +
3777 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3778 // error 2.47208000*10^(-7), which is better than 18 bits
3779 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3780 getF32Constant(DAG, 0x3924b03e, dl));
3781 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3782 getF32Constant(DAG, 0x3ab24b87, dl));
3783 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3784 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3785 getF32Constant(DAG, 0x3c1d8c17, dl));
3786 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3787 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3788 getF32Constant(DAG, 0x3d634a1d, dl));
3789 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3790 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3791 getF32Constant(DAG, 0x3e75fe14, dl));
3792 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3793 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3794 getF32Constant(DAG, 0x3f317234, dl));
3795 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3796 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3797 getF32Constant(DAG, 0x3f800000, dl));
3800 // Add the exponent into the result in integer domain.
3801 SDValue t13 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, TwoToFractionalPartOfX);
3802 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3803 DAG.getNode(ISD::ADD, dl, MVT::i32, t13, IntegerPartOfX));
3806 /// expandExp - Lower an exp intrinsic. Handles the special sequences for
3807 /// limited-precision mode.
3808 static SDValue expandExp(SDLoc dl, SDValue Op, SelectionDAG &DAG,
3809 const TargetLowering &TLI) {
3810 if (Op.getValueType() == MVT::f32 &&
3811 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3813 // Put the exponent in the right bit position for later addition to the
3816 // #define LOG2OFe 1.4426950f
3817 // t0 = Op * LOG2OFe
3819 // TODO: What fast-math-flags should be set here?
3820 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3821 getF32Constant(DAG, 0x3fb8aa3b, dl));
3822 return getLimitedPrecisionExp2(t0, dl, DAG);
3825 // No special expansion.
3826 return DAG.getNode(ISD::FEXP, dl, Op.getValueType(), Op);
3829 /// expandLog - Lower a log intrinsic. Handles the special sequences for
3830 /// limited-precision mode.
3831 static SDValue expandLog(SDLoc dl, SDValue Op, SelectionDAG &DAG,
3832 const TargetLowering &TLI) {
3834 // TODO: What fast-math-flags should be set on the floating-point nodes?
3836 if (Op.getValueType() == MVT::f32 &&
3837 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3838 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3840 // Scale the exponent by log(2) [0.69314718f].
3841 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3842 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3843 getF32Constant(DAG, 0x3f317218, dl));
3845 // Get the significand and build it into a floating-point number with
3847 SDValue X = GetSignificand(DAG, Op1, dl);
3849 SDValue LogOfMantissa;
3850 if (LimitFloatPrecision <= 6) {
3851 // For floating-point precision of 6:
3855 // (1.4034025f - 0.23903021f * x) * x;
3857 // error 0.0034276066, which is better than 8 bits
3858 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3859 getF32Constant(DAG, 0xbe74c456, dl));
3860 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3861 getF32Constant(DAG, 0x3fb3a2b1, dl));
3862 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3863 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3864 getF32Constant(DAG, 0x3f949a29, dl));
3865 } else if (LimitFloatPrecision <= 12) {
3866 // For floating-point precision of 12:
3872 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x;
3874 // error 0.000061011436, which is 14 bits
3875 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3876 getF32Constant(DAG, 0xbd67b6d6, dl));
3877 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3878 getF32Constant(DAG, 0x3ee4f4b8, dl));
3879 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3880 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3881 getF32Constant(DAG, 0x3fbc278b, dl));
3882 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3883 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3884 getF32Constant(DAG, 0x40348e95, dl));
3885 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3886 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3887 getF32Constant(DAG, 0x3fdef31a, dl));
3888 } else { // LimitFloatPrecision <= 18
3889 // For floating-point precision of 18:
3897 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x;
3899 // error 0.0000023660568, which is better than 18 bits
3900 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3901 getF32Constant(DAG, 0xbc91e5ac, dl));
3902 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3903 getF32Constant(DAG, 0x3e4350aa, dl));
3904 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3905 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3906 getF32Constant(DAG, 0x3f60d3e3, dl));
3907 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3908 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3909 getF32Constant(DAG, 0x4011cdf0, dl));
3910 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3911 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3912 getF32Constant(DAG, 0x406cfd1c, dl));
3913 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3914 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3915 getF32Constant(DAG, 0x408797cb, dl));
3916 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3917 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3918 getF32Constant(DAG, 0x4006dcab, dl));
3921 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, LogOfMantissa);
3924 // No special expansion.
3925 return DAG.getNode(ISD::FLOG, dl, Op.getValueType(), Op);
3928 /// expandLog2 - Lower a log2 intrinsic. Handles the special sequences for
3929 /// limited-precision mode.
3930 static SDValue expandLog2(SDLoc dl, SDValue Op, SelectionDAG &DAG,
3931 const TargetLowering &TLI) {
3933 // TODO: What fast-math-flags should be set on the floating-point nodes?
3935 if (Op.getValueType() == MVT::f32 &&
3936 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3937 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3939 // Get the exponent.
3940 SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl);
3942 // Get the significand and build it into a floating-point number with
3944 SDValue X = GetSignificand(DAG, Op1, dl);
3946 // Different possible minimax approximations of significand in
3947 // floating-point for various degrees of accuracy over [1,2].
3948 SDValue Log2ofMantissa;
3949 if (LimitFloatPrecision <= 6) {
3950 // For floating-point precision of 6:
3952 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x;
3954 // error 0.0049451742, which is more than 7 bits
3955 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3956 getF32Constant(DAG, 0xbeb08fe0, dl));
3957 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3958 getF32Constant(DAG, 0x40019463, dl));
3959 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3960 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3961 getF32Constant(DAG, 0x3fd6633d, dl));
3962 } else if (LimitFloatPrecision <= 12) {
3963 // For floating-point precision of 12:
3969 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x;
3971 // error 0.0000876136000, which is better than 13 bits
3972 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3973 getF32Constant(DAG, 0xbda7262e, dl));
3974 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3975 getF32Constant(DAG, 0x3f25280b, dl));
3976 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3977 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3978 getF32Constant(DAG, 0x4007b923, dl));
3979 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3980 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3981 getF32Constant(DAG, 0x40823e2f, dl));
3982 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3983 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3984 getF32Constant(DAG, 0x4020d29c, dl));
3985 } else { // LimitFloatPrecision <= 18
3986 // For floating-point precision of 18:
3995 // 0.25691327e-1f * x) * x) * x) * x) * x) * x;
3997 // error 0.0000018516, which is better than 18 bits
3998 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3999 getF32Constant(DAG, 0xbcd2769e, dl));
4000 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4001 getF32Constant(DAG, 0x3e8ce0b9, dl));
4002 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4003 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4004 getF32Constant(DAG, 0x3fa22ae7, dl));
4005 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4006 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4007 getF32Constant(DAG, 0x40525723, dl));
4008 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4009 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
4010 getF32Constant(DAG, 0x40aaf200, dl));
4011 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4012 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4013 getF32Constant(DAG, 0x40c39dad, dl));
4014 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4015 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
4016 getF32Constant(DAG, 0x4042902c, dl));
4019 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, Log2ofMantissa);
4022 // No special expansion.
4023 return DAG.getNode(ISD::FLOG2, dl, Op.getValueType(), Op);
4026 /// expandLog10 - Lower a log10 intrinsic. Handles the special sequences for
4027 /// limited-precision mode.
4028 static SDValue expandLog10(SDLoc dl, SDValue Op, SelectionDAG &DAG,
4029 const TargetLowering &TLI) {
4031 // TODO: What fast-math-flags should be set on the floating-point nodes?
4033 if (Op.getValueType() == MVT::f32 &&
4034 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4035 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
4037 // Scale the exponent by log10(2) [0.30102999f].
4038 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
4039 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
4040 getF32Constant(DAG, 0x3e9a209a, dl));
4042 // Get the significand and build it into a floating-point number with
4044 SDValue X = GetSignificand(DAG, Op1, dl);
4046 SDValue Log10ofMantissa;
4047 if (LimitFloatPrecision <= 6) {
4048 // For floating-point precision of 6:
4050 // Log10ofMantissa =
4052 // (0.60948995f - 0.10380950f * x) * x;
4054 // error 0.0014886165, which is 6 bits
4055 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4056 getF32Constant(DAG, 0xbdd49a13, dl));
4057 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4058 getF32Constant(DAG, 0x3f1c0789, dl));
4059 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4060 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4061 getF32Constant(DAG, 0x3f011300, dl));
4062 } else if (LimitFloatPrecision <= 12) {
4063 // For floating-point precision of 12:
4065 // Log10ofMantissa =
4068 // (-0.31664806f + 0.47637168e-1f * x) * x) * x;
4070 // error 0.00019228036, which is better than 12 bits
4071 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4072 getF32Constant(DAG, 0x3d431f31, dl));
4073 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
4074 getF32Constant(DAG, 0x3ea21fb2, dl));
4075 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4076 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4077 getF32Constant(DAG, 0x3f6ae232, dl));
4078 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4079 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
4080 getF32Constant(DAG, 0x3f25f7c3, dl));
4081 } else { // LimitFloatPrecision <= 18
4082 // For floating-point precision of 18:
4084 // Log10ofMantissa =
4089 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x;
4091 // error 0.0000037995730, which is better than 18 bits
4092 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4093 getF32Constant(DAG, 0x3c5d51ce, dl));
4094 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
4095 getF32Constant(DAG, 0x3e00685a, dl));
4096 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4097 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4098 getF32Constant(DAG, 0x3efb6798, dl));
4099 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4100 SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
4101 getF32Constant(DAG, 0x3f88d192, dl));
4102 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4103 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4104 getF32Constant(DAG, 0x3fc4316c, dl));
4105 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4106 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8,
4107 getF32Constant(DAG, 0x3f57ce70, dl));
4110 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, Log10ofMantissa);
4113 // No special expansion.
4114 return DAG.getNode(ISD::FLOG10, dl, Op.getValueType(), Op);
4117 /// expandExp2 - Lower an exp2 intrinsic. Handles the special sequences for
4118 /// limited-precision mode.
4119 static SDValue expandExp2(SDLoc dl, SDValue Op, SelectionDAG &DAG,
4120 const TargetLowering &TLI) {
4121 if (Op.getValueType() == MVT::f32 &&
4122 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18)
4123 return getLimitedPrecisionExp2(Op, dl, DAG);
4125 // No special expansion.
4126 return DAG.getNode(ISD::FEXP2, dl, Op.getValueType(), Op);
4129 /// visitPow - Lower a pow intrinsic. Handles the special sequences for
4130 /// limited-precision mode with x == 10.0f.
4131 static SDValue expandPow(SDLoc dl, SDValue LHS, SDValue RHS,
4132 SelectionDAG &DAG, const TargetLowering &TLI) {
4133 bool IsExp10 = false;
4134 if (LHS.getValueType() == MVT::f32 && RHS.getValueType() == MVT::f32 &&
4135 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4136 if (ConstantFPSDNode *LHSC = dyn_cast<ConstantFPSDNode>(LHS)) {
4138 IsExp10 = LHSC->isExactlyValue(Ten);
4142 // TODO: What fast-math-flags should be set on the FMUL node?
4144 // Put the exponent in the right bit position for later addition to the
4147 // #define LOG2OF10 3.3219281f
4148 // t0 = Op * LOG2OF10;
4149 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, RHS,
4150 getF32Constant(DAG, 0x40549a78, dl));
4151 return getLimitedPrecisionExp2(t0, dl, DAG);
4154 // No special expansion.
4155 return DAG.getNode(ISD::FPOW, dl, LHS.getValueType(), LHS, RHS);
4159 /// ExpandPowI - Expand a llvm.powi intrinsic.
4160 static SDValue ExpandPowI(SDLoc DL, SDValue LHS, SDValue RHS,
4161 SelectionDAG &DAG) {
4162 // If RHS is a constant, we can expand this out to a multiplication tree,
4163 // otherwise we end up lowering to a call to __powidf2 (for example). When
4164 // optimizing for size, we only want to do this if the expansion would produce
4165 // a small number of multiplies, otherwise we do the full expansion.
4166 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
4167 // Get the exponent as a positive value.
4168 unsigned Val = RHSC->getSExtValue();
4169 if ((int)Val < 0) Val = -Val;
4171 // powi(x, 0) -> 1.0
4173 return DAG.getConstantFP(1.0, DL, LHS.getValueType());
4175 const Function *F = DAG.getMachineFunction().getFunction();
4176 if (!F->optForSize() ||
4177 // If optimizing for size, don't insert too many multiplies.
4178 // This inserts up to 5 multiplies.
4179 countPopulation(Val) + Log2_32(Val) < 7) {
4180 // We use the simple binary decomposition method to generate the multiply
4181 // sequence. There are more optimal ways to do this (for example,
4182 // powi(x,15) generates one more multiply than it should), but this has
4183 // the benefit of being both really simple and much better than a libcall.
4184 SDValue Res; // Logically starts equal to 1.0
4185 SDValue CurSquare = LHS;
4186 // TODO: Intrinsics should have fast-math-flags that propagate to these
4191 Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare);
4193 Res = CurSquare; // 1.0*CurSquare.
4196 CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(),
4197 CurSquare, CurSquare);
4201 // If the original was negative, invert the result, producing 1/(x*x*x).
4202 if (RHSC->getSExtValue() < 0)
4203 Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(),
4204 DAG.getConstantFP(1.0, DL, LHS.getValueType()), Res);
4209 // Otherwise, expand to a libcall.
4210 return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS);
4213 // getUnderlyingArgReg - Find underlying register used for a truncated or
4214 // bitcasted argument.
4215 static unsigned getUnderlyingArgReg(const SDValue &N) {
4216 switch (N.getOpcode()) {
4217 case ISD::CopyFromReg:
4218 return cast<RegisterSDNode>(N.getOperand(1))->getReg();
4220 case ISD::AssertZext:
4221 case ISD::AssertSext:
4223 return getUnderlyingArgReg(N.getOperand(0));
4229 /// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a function
4230 /// argument, create the corresponding DBG_VALUE machine instruction for it now.
4231 /// At the end of instruction selection, they will be inserted to the entry BB.
4232 bool SelectionDAGBuilder::EmitFuncArgumentDbgValue(
4233 const Value *V, DILocalVariable *Variable, DIExpression *Expr,
4234 DILocation *DL, int64_t Offset, bool IsIndirect, const SDValue &N) {
4235 const Argument *Arg = dyn_cast<Argument>(V);
4239 MachineFunction &MF = DAG.getMachineFunction();
4240 const TargetInstrInfo *TII = DAG.getSubtarget().getInstrInfo();
4242 // Ignore inlined function arguments here.
4244 // FIXME: Should we be checking DL->inlinedAt() to determine this?
4245 if (!Variable->getScope()->getSubprogram()->describes(MF.getFunction()))
4248 Optional<MachineOperand> Op;
4249 // Some arguments' frame index is recorded during argument lowering.
4250 if (int FI = FuncInfo.getArgumentFrameIndex(Arg))
4251 Op = MachineOperand::CreateFI(FI);
4253 if (!Op && N.getNode()) {
4254 unsigned Reg = getUnderlyingArgReg(N);
4255 if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) {
4256 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4257 unsigned PR = RegInfo.getLiveInPhysReg(Reg);
4262 Op = MachineOperand::CreateReg(Reg, false);
4266 // Check if ValueMap has reg number.
4267 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
4268 if (VMI != FuncInfo.ValueMap.end())
4269 Op = MachineOperand::CreateReg(VMI->second, false);
4272 if (!Op && N.getNode())
4273 // Check if frame index is available.
4274 if (LoadSDNode *LNode = dyn_cast<LoadSDNode>(N.getNode()))
4275 if (FrameIndexSDNode *FINode =
4276 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode()))
4277 Op = MachineOperand::CreateFI(FINode->getIndex());
4282 assert(Variable->isValidLocationForIntrinsic(DL) &&
4283 "Expected inlined-at fields to agree");
4285 FuncInfo.ArgDbgValues.push_back(
4286 BuildMI(MF, DL, TII->get(TargetOpcode::DBG_VALUE), IsIndirect,
4287 Op->getReg(), Offset, Variable, Expr));
4289 FuncInfo.ArgDbgValues.push_back(
4290 BuildMI(MF, DL, TII->get(TargetOpcode::DBG_VALUE))
4293 .addMetadata(Variable)
4294 .addMetadata(Expr));
4299 // VisualStudio defines setjmp as _setjmp
4300 #if defined(_MSC_VER) && defined(setjmp) && \
4301 !defined(setjmp_undefined_for_msvc)
4302 # pragma push_macro("setjmp")
4304 # define setjmp_undefined_for_msvc
4307 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
4308 /// we want to emit this as a call to a named external function, return the name
4309 /// otherwise lower it and return null.
4311 SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) {
4312 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4313 SDLoc sdl = getCurSDLoc();
4314 DebugLoc dl = getCurDebugLoc();
4317 switch (Intrinsic) {
4319 // By default, turn this into a target intrinsic node.
4320 visitTargetIntrinsic(I, Intrinsic);
4322 case Intrinsic::vastart: visitVAStart(I); return nullptr;
4323 case Intrinsic::vaend: visitVAEnd(I); return nullptr;
4324 case Intrinsic::vacopy: visitVACopy(I); return nullptr;
4325 case Intrinsic::returnaddress:
4326 setValue(&I, DAG.getNode(ISD::RETURNADDR, sdl,
4327 TLI.getPointerTy(DAG.getDataLayout()),
4328 getValue(I.getArgOperand(0))));
4330 case Intrinsic::frameaddress:
4331 setValue(&I, DAG.getNode(ISD::FRAMEADDR, sdl,
4332 TLI.getPointerTy(DAG.getDataLayout()),
4333 getValue(I.getArgOperand(0))));
4335 case Intrinsic::read_register: {
4336 Value *Reg = I.getArgOperand(0);
4337 SDValue Chain = getRoot();
4339 DAG.getMDNode(cast<MDNode>(cast<MetadataAsValue>(Reg)->getMetadata()));
4340 EVT VT = TLI.getValueType(DAG.getDataLayout(), I.getType());
4341 Res = DAG.getNode(ISD::READ_REGISTER, sdl,
4342 DAG.getVTList(VT, MVT::Other), Chain, RegName);
4344 DAG.setRoot(Res.getValue(1));
4347 case Intrinsic::write_register: {
4348 Value *Reg = I.getArgOperand(0);
4349 Value *RegValue = I.getArgOperand(1);
4350 SDValue Chain = getRoot();
4352 DAG.getMDNode(cast<MDNode>(cast<MetadataAsValue>(Reg)->getMetadata()));
4353 DAG.setRoot(DAG.getNode(ISD::WRITE_REGISTER, sdl, MVT::Other, Chain,
4354 RegName, getValue(RegValue)));
4357 case Intrinsic::setjmp:
4358 return &"_setjmp"[!TLI.usesUnderscoreSetJmp()];
4359 case Intrinsic::longjmp:
4360 return &"_longjmp"[!TLI.usesUnderscoreLongJmp()];
4361 case Intrinsic::memcpy: {
4362 // FIXME: this definition of "user defined address space" is x86-specific
4363 // Assert for address < 256 since we support only user defined address
4365 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4367 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4369 "Unknown address space");
4370 SDValue Op1 = getValue(I.getArgOperand(0));
4371 SDValue Op2 = getValue(I.getArgOperand(1));
4372 SDValue Op3 = getValue(I.getArgOperand(2));
4373 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4375 Align = 1; // @llvm.memcpy defines 0 and 1 to both mean no alignment.
4376 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4377 bool isTC = I.isTailCall() && isInTailCallPosition(&I, DAG.getTarget());
4378 SDValue MC = DAG.getMemcpy(getRoot(), sdl, Op1, Op2, Op3, Align, isVol,
4380 MachinePointerInfo(I.getArgOperand(0)),
4381 MachinePointerInfo(I.getArgOperand(1)));
4382 updateDAGForMaybeTailCall(MC);
4385 case Intrinsic::memset: {
4386 // FIXME: this definition of "user defined address space" is x86-specific
4387 // Assert for address < 256 since we support only user defined address
4389 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4391 "Unknown address space");
4392 SDValue Op1 = getValue(I.getArgOperand(0));
4393 SDValue Op2 = getValue(I.getArgOperand(1));
4394 SDValue Op3 = getValue(I.getArgOperand(2));
4395 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4397 Align = 1; // @llvm.memset defines 0 and 1 to both mean no alignment.
4398 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4399 bool isTC = I.isTailCall() && isInTailCallPosition(&I, DAG.getTarget());
4400 SDValue MS = DAG.getMemset(getRoot(), sdl, Op1, Op2, Op3, Align, isVol,
4401 isTC, MachinePointerInfo(I.getArgOperand(0)));
4402 updateDAGForMaybeTailCall(MS);
4405 case Intrinsic::memmove: {
4406 // FIXME: this definition of "user defined address space" is x86-specific
4407 // Assert for address < 256 since we support only user defined address
4409 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4411 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4413 "Unknown address space");
4414 SDValue Op1 = getValue(I.getArgOperand(0));
4415 SDValue Op2 = getValue(I.getArgOperand(1));
4416 SDValue Op3 = getValue(I.getArgOperand(2));
4417 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4419 Align = 1; // @llvm.memmove defines 0 and 1 to both mean no alignment.
4420 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4421 bool isTC = I.isTailCall() && isInTailCallPosition(&I, DAG.getTarget());
4422 SDValue MM = DAG.getMemmove(getRoot(), sdl, Op1, Op2, Op3, Align, isVol,
4423 isTC, MachinePointerInfo(I.getArgOperand(0)),
4424 MachinePointerInfo(I.getArgOperand(1)));
4425 updateDAGForMaybeTailCall(MM);
4428 case Intrinsic::dbg_declare: {
4429 const DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
4430 DILocalVariable *Variable = DI.getVariable();
4431 DIExpression *Expression = DI.getExpression();
4432 const Value *Address = DI.getAddress();
4433 assert(Variable && "Missing variable");
4435 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4439 // Check if address has undef value.
4440 if (isa<UndefValue>(Address) ||
4441 (Address->use_empty() && !isa<Argument>(Address))) {
4442 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4446 SDValue &N = NodeMap[Address];
4447 if (!N.getNode() && isa<Argument>(Address))
4448 // Check unused arguments map.
4449 N = UnusedArgNodeMap[Address];
4452 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
4453 Address = BCI->getOperand(0);
4454 // Parameters are handled specially.
4455 bool isParameter = Variable->isParameter() || isa<Argument>(Address);
4457 const AllocaInst *AI = dyn_cast<AllocaInst>(Address);
4459 if (isParameter && !AI) {
4460 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode());
4462 // Byval parameter. We have a frame index at this point.
4463 SDV = DAG.getFrameIndexDbgValue(
4464 Variable, Expression, FINode->getIndex(), 0, dl, SDNodeOrder);
4466 // Address is an argument, so try to emit its dbg value using
4467 // virtual register info from the FuncInfo.ValueMap.
4468 EmitFuncArgumentDbgValue(Address, Variable, Expression, dl, 0, false,
4473 SDV = DAG.getDbgValue(Variable, Expression, N.getNode(), N.getResNo(),
4474 true, 0, dl, SDNodeOrder);
4476 DAG.AddDbgValue(SDV, N.getNode(), isParameter);
4478 // If Address is an argument then try to emit its dbg value using
4479 // virtual register info from the FuncInfo.ValueMap.
4480 if (!EmitFuncArgumentDbgValue(Address, Variable, Expression, dl, 0, false,
4482 // If variable is pinned by a alloca in dominating bb then
4483 // use StaticAllocaMap.
4484 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Address)) {
4485 if (AI->getParent() != DI.getParent()) {
4486 DenseMap<const AllocaInst*, int>::iterator SI =
4487 FuncInfo.StaticAllocaMap.find(AI);
4488 if (SI != FuncInfo.StaticAllocaMap.end()) {
4489 SDV = DAG.getFrameIndexDbgValue(Variable, Expression, SI->second,
4490 0, dl, SDNodeOrder);
4491 DAG.AddDbgValue(SDV, nullptr, false);
4496 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4501 case Intrinsic::dbg_value: {
4502 const DbgValueInst &DI = cast<DbgValueInst>(I);
4503 assert(DI.getVariable() && "Missing variable");
4505 DILocalVariable *Variable = DI.getVariable();
4506 DIExpression *Expression = DI.getExpression();
4507 uint64_t Offset = DI.getOffset();
4508 const Value *V = DI.getValue();
4513 if (isa<ConstantInt>(V) || isa<ConstantFP>(V) || isa<UndefValue>(V)) {
4514 SDV = DAG.getConstantDbgValue(Variable, Expression, V, Offset, dl,
4516 DAG.AddDbgValue(SDV, nullptr, false);
4518 // Do not use getValue() in here; we don't want to generate code at
4519 // this point if it hasn't been done yet.
4520 SDValue N = NodeMap[V];
4521 if (!N.getNode() && isa<Argument>(V))
4522 // Check unused arguments map.
4523 N = UnusedArgNodeMap[V];
4525 // A dbg.value for an alloca is always indirect.
4526 bool IsIndirect = isa<AllocaInst>(V) || Offset != 0;
4527 if (!EmitFuncArgumentDbgValue(V, Variable, Expression, dl, Offset,
4529 SDV = DAG.getDbgValue(Variable, Expression, N.getNode(), N.getResNo(),
4530 IsIndirect, Offset, dl, SDNodeOrder);
4531 DAG.AddDbgValue(SDV, N.getNode(), false);
4533 } else if (!V->use_empty() ) {
4534 // Do not call getValue(V) yet, as we don't want to generate code.
4535 // Remember it for later.
4536 DanglingDebugInfo DDI(&DI, dl, SDNodeOrder);
4537 DanglingDebugInfoMap[V] = DDI;
4539 // We may expand this to cover more cases. One case where we have no
4540 // data available is an unreferenced parameter.
4541 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4545 // Build a debug info table entry.
4546 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
4547 V = BCI->getOperand(0);
4548 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
4549 // Don't handle byval struct arguments or VLAs, for example.
4551 DEBUG(dbgs() << "Dropping debug location info for:\n " << DI << "\n");
4552 DEBUG(dbgs() << " Last seen at:\n " << *V << "\n");
4555 DenseMap<const AllocaInst*, int>::iterator SI =
4556 FuncInfo.StaticAllocaMap.find(AI);
4557 if (SI == FuncInfo.StaticAllocaMap.end())
4558 return nullptr; // VLAs.
4562 case Intrinsic::eh_typeid_for: {
4563 // Find the type id for the given typeinfo.
4564 GlobalValue *GV = ExtractTypeInfo(I.getArgOperand(0));
4565 unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(GV);
4566 Res = DAG.getConstant(TypeID, sdl, MVT::i32);
4571 case Intrinsic::eh_return_i32:
4572 case Intrinsic::eh_return_i64:
4573 DAG.getMachineFunction().getMMI().setCallsEHReturn(true);
4574 DAG.setRoot(DAG.getNode(ISD::EH_RETURN, sdl,
4577 getValue(I.getArgOperand(0)),
4578 getValue(I.getArgOperand(1))));
4580 case Intrinsic::eh_unwind_init:
4581 DAG.getMachineFunction().getMMI().setCallsUnwindInit(true);
4583 case Intrinsic::eh_dwarf_cfa: {
4584 SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getArgOperand(0)), sdl,
4585 TLI.getPointerTy(DAG.getDataLayout()));
4586 SDValue Offset = DAG.getNode(ISD::ADD, sdl,
4587 CfaArg.getValueType(),
4588 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, sdl,
4589 CfaArg.getValueType()),
4591 SDValue FA = DAG.getNode(
4592 ISD::FRAMEADDR, sdl, TLI.getPointerTy(DAG.getDataLayout()),
4593 DAG.getConstant(0, sdl, TLI.getPointerTy(DAG.getDataLayout())));
4594 setValue(&I, DAG.getNode(ISD::ADD, sdl, FA.getValueType(),
4598 case Intrinsic::eh_sjlj_callsite: {
4599 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4600 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(0));
4601 assert(CI && "Non-constant call site value in eh.sjlj.callsite!");
4602 assert(MMI.getCurrentCallSite() == 0 && "Overlapping call sites!");
4604 MMI.setCurrentCallSite(CI->getZExtValue());
4607 case Intrinsic::eh_sjlj_functioncontext: {
4608 // Get and store the index of the function context.
4609 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4611 cast<AllocaInst>(I.getArgOperand(0)->stripPointerCasts());
4612 int FI = FuncInfo.StaticAllocaMap[FnCtx];
4613 MFI->setFunctionContextIndex(FI);
4616 case Intrinsic::eh_sjlj_setjmp: {
4619 Ops[1] = getValue(I.getArgOperand(0));
4620 SDValue Op = DAG.getNode(ISD::EH_SJLJ_SETJMP, sdl,
4621 DAG.getVTList(MVT::i32, MVT::Other), Ops);
4622 setValue(&I, Op.getValue(0));
4623 DAG.setRoot(Op.getValue(1));
4626 case Intrinsic::eh_sjlj_longjmp: {
4627 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, sdl, MVT::Other,
4628 getRoot(), getValue(I.getArgOperand(0))));
4631 case Intrinsic::eh_sjlj_setup_dispatch: {
4632 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_SETUP_DISPATCH, sdl, MVT::Other,
4637 case Intrinsic::masked_gather:
4638 visitMaskedGather(I);
4640 case Intrinsic::masked_load:
4643 case Intrinsic::masked_scatter:
4644 visitMaskedScatter(I);
4646 case Intrinsic::masked_store:
4647 visitMaskedStore(I);
4649 case Intrinsic::x86_mmx_pslli_w:
4650 case Intrinsic::x86_mmx_pslli_d:
4651 case Intrinsic::x86_mmx_pslli_q:
4652 case Intrinsic::x86_mmx_psrli_w:
4653 case Intrinsic::x86_mmx_psrli_d:
4654 case Intrinsic::x86_mmx_psrli_q:
4655 case Intrinsic::x86_mmx_psrai_w:
4656 case Intrinsic::x86_mmx_psrai_d: {
4657 SDValue ShAmt = getValue(I.getArgOperand(1));
4658 if (isa<ConstantSDNode>(ShAmt)) {
4659 visitTargetIntrinsic(I, Intrinsic);
4662 unsigned NewIntrinsic = 0;
4663 EVT ShAmtVT = MVT::v2i32;
4664 switch (Intrinsic) {
4665 case Intrinsic::x86_mmx_pslli_w:
4666 NewIntrinsic = Intrinsic::x86_mmx_psll_w;
4668 case Intrinsic::x86_mmx_pslli_d:
4669 NewIntrinsic = Intrinsic::x86_mmx_psll_d;
4671 case Intrinsic::x86_mmx_pslli_q:
4672 NewIntrinsic = Intrinsic::x86_mmx_psll_q;
4674 case Intrinsic::x86_mmx_psrli_w:
4675 NewIntrinsic = Intrinsic::x86_mmx_psrl_w;
4677 case Intrinsic::x86_mmx_psrli_d:
4678 NewIntrinsic = Intrinsic::x86_mmx_psrl_d;
4680 case Intrinsic::x86_mmx_psrli_q:
4681 NewIntrinsic = Intrinsic::x86_mmx_psrl_q;
4683 case Intrinsic::x86_mmx_psrai_w:
4684 NewIntrinsic = Intrinsic::x86_mmx_psra_w;
4686 case Intrinsic::x86_mmx_psrai_d:
4687 NewIntrinsic = Intrinsic::x86_mmx_psra_d;
4689 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4692 // The vector shift intrinsics with scalars uses 32b shift amounts but
4693 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
4695 // We must do this early because v2i32 is not a legal type.
4698 ShOps[1] = DAG.getConstant(0, sdl, MVT::i32);
4699 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, sdl, ShAmtVT, ShOps);
4700 EVT DestVT = TLI.getValueType(DAG.getDataLayout(), I.getType());
4701 ShAmt = DAG.getNode(ISD::BITCAST, sdl, DestVT, ShAmt);
4702 Res = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, sdl, DestVT,
4703 DAG.getConstant(NewIntrinsic, sdl, MVT::i32),
4704 getValue(I.getArgOperand(0)), ShAmt);
4708 case Intrinsic::convertff:
4709 case Intrinsic::convertfsi:
4710 case Intrinsic::convertfui:
4711 case Intrinsic::convertsif:
4712 case Intrinsic::convertuif:
4713 case Intrinsic::convertss:
4714 case Intrinsic::convertsu:
4715 case Intrinsic::convertus:
4716 case Intrinsic::convertuu: {
4717 ISD::CvtCode Code = ISD::CVT_INVALID;
4718 switch (Intrinsic) {
4719 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4720 case Intrinsic::convertff: Code = ISD::CVT_FF; break;
4721 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break;
4722 case Intrinsic::convertfui: Code = ISD::CVT_FU; break;
4723 case Intrinsic::convertsif: Code = ISD::CVT_SF; break;
4724 case Intrinsic::convertuif: Code = ISD::CVT_UF; break;
4725 case Intrinsic::convertss: Code = ISD::CVT_SS; break;
4726 case Intrinsic::convertsu: Code = ISD::CVT_SU; break;
4727 case Intrinsic::convertus: Code = ISD::CVT_US; break;
4728 case Intrinsic::convertuu: Code = ISD::CVT_UU; break;
4730 EVT DestVT = TLI.getValueType(DAG.getDataLayout(), I.getType());
4731 const Value *Op1 = I.getArgOperand(0);
4732 Res = DAG.getConvertRndSat(DestVT, sdl, getValue(Op1),
4733 DAG.getValueType(DestVT),
4734 DAG.getValueType(getValue(Op1).getValueType()),
4735 getValue(I.getArgOperand(1)),
4736 getValue(I.getArgOperand(2)),
4741 case Intrinsic::powi:
4742 setValue(&I, ExpandPowI(sdl, getValue(I.getArgOperand(0)),
4743 getValue(I.getArgOperand(1)), DAG));
4745 case Intrinsic::log:
4746 setValue(&I, expandLog(sdl, getValue(I.getArgOperand(0)), DAG, TLI));
4748 case Intrinsic::log2:
4749 setValue(&I, expandLog2(sdl, getValue(I.getArgOperand(0)), DAG, TLI));
4751 case Intrinsic::log10:
4752 setValue(&I, expandLog10(sdl, getValue(I.getArgOperand(0)), DAG, TLI));
4754 case Intrinsic::exp:
4755 setValue(&I, expandExp(sdl, getValue(I.getArgOperand(0)), DAG, TLI));
4757 case Intrinsic::exp2:
4758 setValue(&I, expandExp2(sdl, getValue(I.getArgOperand(0)), DAG, TLI));
4760 case Intrinsic::pow:
4761 setValue(&I, expandPow(sdl, getValue(I.getArgOperand(0)),
4762 getValue(I.getArgOperand(1)), DAG, TLI));
4764 case Intrinsic::sqrt:
4765 case Intrinsic::fabs:
4766 case Intrinsic::sin:
4767 case Intrinsic::cos:
4768 case Intrinsic::floor:
4769 case Intrinsic::ceil:
4770 case Intrinsic::trunc:
4771 case Intrinsic::rint:
4772 case Intrinsic::nearbyint:
4773 case Intrinsic::round: {
4775 switch (Intrinsic) {
4776 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4777 case Intrinsic::sqrt: Opcode = ISD::FSQRT; break;
4778 case Intrinsic::fabs: Opcode = ISD::FABS; break;
4779 case Intrinsic::sin: Opcode = ISD::FSIN; break;
4780 case Intrinsic::cos: Opcode = ISD::FCOS; break;
4781 case Intrinsic::floor: Opcode = ISD::FFLOOR; break;
4782 case Intrinsic::ceil: Opcode = ISD::FCEIL; break;
4783 case Intrinsic::trunc: Opcode = ISD::FTRUNC; break;
4784 case Intrinsic::rint: Opcode = ISD::FRINT; break;
4785 case Intrinsic::nearbyint: Opcode = ISD::FNEARBYINT; break;
4786 case Intrinsic::round: Opcode = ISD::FROUND; break;
4789 setValue(&I, DAG.getNode(Opcode, sdl,
4790 getValue(I.getArgOperand(0)).getValueType(),
4791 getValue(I.getArgOperand(0))));
4794 case Intrinsic::minnum:
4795 setValue(&I, DAG.getNode(ISD::FMINNUM, sdl,
4796 getValue(I.getArgOperand(0)).getValueType(),
4797 getValue(I.getArgOperand(0)),
4798 getValue(I.getArgOperand(1))));
4800 case Intrinsic::maxnum:
4801 setValue(&I, DAG.getNode(ISD::FMAXNUM, sdl,
4802 getValue(I.getArgOperand(0)).getValueType(),
4803 getValue(I.getArgOperand(0)),
4804 getValue(I.getArgOperand(1))));
4806 case Intrinsic::copysign:
4807 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, sdl,
4808 getValue(I.getArgOperand(0)).getValueType(),
4809 getValue(I.getArgOperand(0)),
4810 getValue(I.getArgOperand(1))));
4812 case Intrinsic::fma:
4813 setValue(&I, DAG.getNode(ISD::FMA, sdl,
4814 getValue(I.getArgOperand(0)).getValueType(),
4815 getValue(I.getArgOperand(0)),
4816 getValue(I.getArgOperand(1)),
4817 getValue(I.getArgOperand(2))));
4819 case Intrinsic::fmuladd: {
4820 EVT VT = TLI.getValueType(DAG.getDataLayout(), I.getType());
4821 if (TM.Options.AllowFPOpFusion != FPOpFusion::Strict &&
4822 TLI.isFMAFasterThanFMulAndFAdd(VT)) {
4823 setValue(&I, DAG.getNode(ISD::FMA, sdl,
4824 getValue(I.getArgOperand(0)).getValueType(),
4825 getValue(I.getArgOperand(0)),
4826 getValue(I.getArgOperand(1)),
4827 getValue(I.getArgOperand(2))));
4829 // TODO: Intrinsic calls should have fast-math-flags.
4830 SDValue Mul = DAG.getNode(ISD::FMUL, sdl,
4831 getValue(I.getArgOperand(0)).getValueType(),
4832 getValue(I.getArgOperand(0)),
4833 getValue(I.getArgOperand(1)));
4834 SDValue Add = DAG.getNode(ISD::FADD, sdl,
4835 getValue(I.getArgOperand(0)).getValueType(),
4837 getValue(I.getArgOperand(2)));
4842 case Intrinsic::convert_to_fp16:
4843 setValue(&I, DAG.getNode(ISD::BITCAST, sdl, MVT::i16,
4844 DAG.getNode(ISD::FP_ROUND, sdl, MVT::f16,
4845 getValue(I.getArgOperand(0)),
4846 DAG.getTargetConstant(0, sdl,
4849 case Intrinsic::convert_from_fp16:
4850 setValue(&I, DAG.getNode(ISD::FP_EXTEND, sdl,
4851 TLI.getValueType(DAG.getDataLayout(), I.getType()),
4852 DAG.getNode(ISD::BITCAST, sdl, MVT::f16,
4853 getValue(I.getArgOperand(0)))));
4855 case Intrinsic::pcmarker: {
4856 SDValue Tmp = getValue(I.getArgOperand(0));
4857 DAG.setRoot(DAG.getNode(ISD::PCMARKER, sdl, MVT::Other, getRoot(), Tmp));
4860 case Intrinsic::readcyclecounter: {
4861 SDValue Op = getRoot();
4862 Res = DAG.getNode(ISD::READCYCLECOUNTER, sdl,
4863 DAG.getVTList(MVT::i64, MVT::Other), Op);
4865 DAG.setRoot(Res.getValue(1));
4868 case Intrinsic::bswap:
4869 setValue(&I, DAG.getNode(ISD::BSWAP, sdl,
4870 getValue(I.getArgOperand(0)).getValueType(),
4871 getValue(I.getArgOperand(0))));
4873 case Intrinsic::uabsdiff:
4874 setValue(&I, DAG.getNode(ISD::UABSDIFF, sdl,
4875 getValue(I.getArgOperand(0)).getValueType(),
4876 getValue(I.getArgOperand(0)),
4877 getValue(I.getArgOperand(1))));
4879 case Intrinsic::sabsdiff:
4880 setValue(&I, DAG.getNode(ISD::SABSDIFF, sdl,
4881 getValue(I.getArgOperand(0)).getValueType(),
4882 getValue(I.getArgOperand(0)),
4883 getValue(I.getArgOperand(1))));
4885 case Intrinsic::cttz: {
4886 SDValue Arg = getValue(I.getArgOperand(0));
4887 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1));
4888 EVT Ty = Arg.getValueType();
4889 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTTZ : ISD::CTTZ_ZERO_UNDEF,
4893 case Intrinsic::ctlz: {
4894 SDValue Arg = getValue(I.getArgOperand(0));
4895 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1));
4896 EVT Ty = Arg.getValueType();
4897 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTLZ : ISD::CTLZ_ZERO_UNDEF,
4901 case Intrinsic::ctpop: {
4902 SDValue Arg = getValue(I.getArgOperand(0));
4903 EVT Ty = Arg.getValueType();
4904 setValue(&I, DAG.getNode(ISD::CTPOP, sdl, Ty, Arg));
4907 case Intrinsic::stacksave: {
4908 SDValue Op = getRoot();
4910 ISD::STACKSAVE, sdl,
4911 DAG.getVTList(TLI.getPointerTy(DAG.getDataLayout()), MVT::Other), Op);
4913 DAG.setRoot(Res.getValue(1));
4916 case Intrinsic::stackrestore: {
4917 Res = getValue(I.getArgOperand(0));
4918 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, sdl, MVT::Other, getRoot(), Res));
4921 case Intrinsic::stackprotector: {
4922 // Emit code into the DAG to store the stack guard onto the stack.
4923 MachineFunction &MF = DAG.getMachineFunction();
4924 MachineFrameInfo *MFI = MF.getFrameInfo();
4925 EVT PtrTy = TLI.getPointerTy(DAG.getDataLayout());
4926 SDValue Src, Chain = getRoot();
4927 const Value *Ptr = cast<LoadInst>(I.getArgOperand(0))->getPointerOperand();
4928 const GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr);
4930 // See if Ptr is a bitcast. If it is, look through it and see if we can get
4931 // global variable __stack_chk_guard.
4933 if (const Operator *BC = dyn_cast<Operator>(Ptr))
4934 if (BC->getOpcode() == Instruction::BitCast)
4935 GV = dyn_cast<GlobalVariable>(BC->getOperand(0));
4937 if (GV && TLI.useLoadStackGuardNode()) {
4938 // Emit a LOAD_STACK_GUARD node.
4939 MachineSDNode *Node = DAG.getMachineNode(TargetOpcode::LOAD_STACK_GUARD,
4941 MachinePointerInfo MPInfo(GV);
4942 MachineInstr::mmo_iterator MemRefs = MF.allocateMemRefsArray(1);
4943 unsigned Flags = MachineMemOperand::MOLoad |
4944 MachineMemOperand::MOInvariant;
4945 *MemRefs = MF.getMachineMemOperand(MPInfo, Flags,
4946 PtrTy.getSizeInBits() / 8,
4947 DAG.getEVTAlignment(PtrTy));
4948 Node->setMemRefs(MemRefs, MemRefs + 1);
4950 // Copy the guard value to a virtual register so that it can be
4951 // retrieved in the epilogue.
4952 Src = SDValue(Node, 0);
4953 const TargetRegisterClass *RC =
4954 TLI.getRegClassFor(Src.getSimpleValueType());
4955 unsigned Reg = MF.getRegInfo().createVirtualRegister(RC);
4957 SPDescriptor.setGuardReg(Reg);
4958 Chain = DAG.getCopyToReg(Chain, sdl, Reg, Src);
4960 Src = getValue(I.getArgOperand(0)); // The guard's value.
4963 AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
4965 int FI = FuncInfo.StaticAllocaMap[Slot];
4966 MFI->setStackProtectorIndex(FI);
4968 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
4970 // Store the stack protector onto the stack.
4971 Res = DAG.getStore(Chain, sdl, Src, FIN, MachinePointerInfo::getFixedStack(
4972 DAG.getMachineFunction(), FI),
4978 case Intrinsic::objectsize: {
4979 // If we don't know by now, we're never going to know.
4980 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1));
4982 assert(CI && "Non-constant type in __builtin_object_size?");
4984 SDValue Arg = getValue(I.getCalledValue());
4985 EVT Ty = Arg.getValueType();
4988 Res = DAG.getConstant(-1ULL, sdl, Ty);
4990 Res = DAG.getConstant(0, sdl, Ty);
4995 case Intrinsic::annotation:
4996 case Intrinsic::ptr_annotation:
4997 // Drop the intrinsic, but forward the value
4998 setValue(&I, getValue(I.getOperand(0)));
5000 case Intrinsic::assume:
5001 case Intrinsic::var_annotation:
5002 // Discard annotate attributes and assumptions
5005 case Intrinsic::init_trampoline: {
5006 const Function *F = cast<Function>(I.getArgOperand(1)->stripPointerCasts());
5010 Ops[1] = getValue(I.getArgOperand(0));
5011 Ops[2] = getValue(I.getArgOperand(1));
5012 Ops[3] = getValue(I.getArgOperand(2));
5013 Ops[4] = DAG.getSrcValue(I.getArgOperand(0));
5014 Ops[5] = DAG.getSrcValue(F);
5016 Res = DAG.getNode(ISD::INIT_TRAMPOLINE, sdl, MVT::Other, Ops);
5021 case Intrinsic::adjust_trampoline: {
5022 setValue(&I, DAG.getNode(ISD::ADJUST_TRAMPOLINE, sdl,
5023 TLI.getPointerTy(DAG.getDataLayout()),
5024 getValue(I.getArgOperand(0))));
5027 case Intrinsic::gcroot:
5029 const Value *Alloca = I.getArgOperand(0)->stripPointerCasts();
5030 const Constant *TypeMap = cast<Constant>(I.getArgOperand(1));
5032 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode());
5033 GFI->addStackRoot(FI->getIndex(), TypeMap);
5036 case Intrinsic::gcread:
5037 case Intrinsic::gcwrite:
5038 llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!");
5039 case Intrinsic::flt_rounds:
5040 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, sdl, MVT::i32));
5043 case Intrinsic::expect: {
5044 // Just replace __builtin_expect(exp, c) with EXP.
5045 setValue(&I, getValue(I.getArgOperand(0)));
5049 case Intrinsic::debugtrap:
5050 case Intrinsic::trap: {
5051 StringRef TrapFuncName =
5053 .getAttribute(AttributeSet::FunctionIndex, "trap-func-name")
5054 .getValueAsString();
5055 if (TrapFuncName.empty()) {
5056 ISD::NodeType Op = (Intrinsic == Intrinsic::trap) ?
5057 ISD::TRAP : ISD::DEBUGTRAP;
5058 DAG.setRoot(DAG.getNode(Op, sdl,MVT::Other, getRoot()));
5061 TargetLowering::ArgListTy Args;
5063 TargetLowering::CallLoweringInfo CLI(DAG);
5064 CLI.setDebugLoc(sdl).setChain(getRoot()).setCallee(
5065 CallingConv::C, I.getType(),
5066 DAG.getExternalSymbol(TrapFuncName.data(),
5067 TLI.getPointerTy(DAG.getDataLayout())),
5068 std::move(Args), 0);
5070 std::pair<SDValue, SDValue> Result = TLI.LowerCallTo(CLI);
5071 DAG.setRoot(Result.second);
5075 case Intrinsic::uadd_with_overflow:
5076 case Intrinsic::sadd_with_overflow:
5077 case Intrinsic::usub_with_overflow:
5078 case Intrinsic::ssub_with_overflow:
5079 case Intrinsic::umul_with_overflow:
5080 case Intrinsic::smul_with_overflow: {
5082 switch (Intrinsic) {
5083 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
5084 case Intrinsic::uadd_with_overflow: Op = ISD::UADDO; break;
5085 case Intrinsic::sadd_with_overflow: Op = ISD::SADDO; break;
5086 case Intrinsic::usub_with_overflow: Op = ISD::USUBO; break;
5087 case Intrinsic::ssub_with_overflow: Op = ISD::SSUBO; break;
5088 case Intrinsic::umul_with_overflow: Op = ISD::UMULO; break;
5089 case Intrinsic::smul_with_overflow: Op = ISD::SMULO; break;
5091 SDValue Op1 = getValue(I.getArgOperand(0));
5092 SDValue Op2 = getValue(I.getArgOperand(1));
5094 SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1);
5095 setValue(&I, DAG.getNode(Op, sdl, VTs, Op1, Op2));
5098 case Intrinsic::prefetch: {
5100 unsigned rw = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
5102 Ops[1] = getValue(I.getArgOperand(0));
5103 Ops[2] = getValue(I.getArgOperand(1));
5104 Ops[3] = getValue(I.getArgOperand(2));
5105 Ops[4] = getValue(I.getArgOperand(3));
5106 DAG.setRoot(DAG.getMemIntrinsicNode(ISD::PREFETCH, sdl,
5107 DAG.getVTList(MVT::Other), Ops,
5108 EVT::getIntegerVT(*Context, 8),
5109 MachinePointerInfo(I.getArgOperand(0)),
5111 false, /* volatile */
5113 rw==1)); /* write */
5116 case Intrinsic::lifetime_start:
5117 case Intrinsic::lifetime_end: {
5118 bool IsStart = (Intrinsic == Intrinsic::lifetime_start);
5119 // Stack coloring is not enabled in O0, discard region information.
5120 if (TM.getOptLevel() == CodeGenOpt::None)
5123 SmallVector<Value *, 4> Allocas;
5124 GetUnderlyingObjects(I.getArgOperand(1), Allocas, *DL);
5126 for (SmallVectorImpl<Value*>::iterator Object = Allocas.begin(),
5127 E = Allocas.end(); Object != E; ++Object) {
5128 AllocaInst *LifetimeObject = dyn_cast_or_null<AllocaInst>(*Object);
5130 // Could not find an Alloca.
5131 if (!LifetimeObject)
5134 // First check that the Alloca is static, otherwise it won't have a
5135 // valid frame index.
5136 auto SI = FuncInfo.StaticAllocaMap.find(LifetimeObject);
5137 if (SI == FuncInfo.StaticAllocaMap.end())
5140 int FI = SI->second;
5145 DAG.getFrameIndex(FI, TLI.getPointerTy(DAG.getDataLayout()), true);
5146 unsigned Opcode = (IsStart ? ISD::LIFETIME_START : ISD::LIFETIME_END);
5148 Res = DAG.getNode(Opcode, sdl, MVT::Other, Ops);
5153 case Intrinsic::invariant_start:
5154 // Discard region information.
5155 setValue(&I, DAG.getUNDEF(TLI.getPointerTy(DAG.getDataLayout())));
5157 case Intrinsic::invariant_end:
5158 // Discard region information.
5160 case Intrinsic::stackprotectorcheck: {
5161 // Do not actually emit anything for this basic block. Instead we initialize
5162 // the stack protector descriptor and export the guard variable so we can
5163 // access it in FinishBasicBlock.
5164 const BasicBlock *BB = I.getParent();
5165 SPDescriptor.initialize(BB, FuncInfo.MBBMap[BB], I);
5166 ExportFromCurrentBlock(SPDescriptor.getGuard());
5168 // Flush our exports since we are going to process a terminator.
5169 (void)getControlRoot();
5172 case Intrinsic::clear_cache:
5173 return TLI.getClearCacheBuiltinName();
5174 case Intrinsic::donothing:
5177 case Intrinsic::experimental_stackmap: {
5181 case Intrinsic::experimental_patchpoint_void:
5182 case Intrinsic::experimental_patchpoint_i64: {
5183 visitPatchpoint(&I);
5186 case Intrinsic::experimental_gc_statepoint: {
5190 case Intrinsic::experimental_gc_result_int:
5191 case Intrinsic::experimental_gc_result_float:
5192 case Intrinsic::experimental_gc_result_ptr:
5193 case Intrinsic::experimental_gc_result: {
5197 case Intrinsic::experimental_gc_relocate: {
5201 case Intrinsic::instrprof_increment:
5202 llvm_unreachable("instrprof failed to lower an increment");
5204 case Intrinsic::localescape: {
5205 MachineFunction &MF = DAG.getMachineFunction();
5206 const TargetInstrInfo *TII = DAG.getSubtarget().getInstrInfo();
5208 // Directly emit some LOCAL_ESCAPE machine instrs. Label assignment emission
5209 // is the same on all targets.
5210 for (unsigned Idx = 0, E = I.getNumArgOperands(); Idx < E; ++Idx) {
5211 Value *Arg = I.getArgOperand(Idx)->stripPointerCasts();
5212 if (isa<ConstantPointerNull>(Arg))
5213 continue; // Skip null pointers. They represent a hole in index space.
5214 AllocaInst *Slot = cast<AllocaInst>(Arg);
5215 assert(FuncInfo.StaticAllocaMap.count(Slot) &&
5216 "can only escape static allocas");
5217 int FI = FuncInfo.StaticAllocaMap[Slot];
5218 MCSymbol *FrameAllocSym =
5219 MF.getMMI().getContext().getOrCreateFrameAllocSymbol(
5220 GlobalValue::getRealLinkageName(MF.getName()), Idx);
5221 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, dl,
5222 TII->get(TargetOpcode::LOCAL_ESCAPE))
5223 .addSym(FrameAllocSym)
5230 case Intrinsic::localrecover: {
5231 // i8* @llvm.localrecover(i8* %fn, i8* %fp, i32 %idx)
5232 MachineFunction &MF = DAG.getMachineFunction();
5233 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout(), 0);
5235 // Get the symbol that defines the frame offset.
5236 auto *Fn = cast<Function>(I.getArgOperand(0)->stripPointerCasts());
5237 auto *Idx = cast<ConstantInt>(I.getArgOperand(2));
5238 unsigned IdxVal = unsigned(Idx->getLimitedValue(INT_MAX));
5239 MCSymbol *FrameAllocSym =
5240 MF.getMMI().getContext().getOrCreateFrameAllocSymbol(
5241 GlobalValue::getRealLinkageName(Fn->getName()), IdxVal);
5243 // Create a MCSymbol for the label to avoid any target lowering
5244 // that would make this PC relative.
5245 SDValue OffsetSym = DAG.getMCSymbol(FrameAllocSym, PtrVT);
5247 DAG.getNode(ISD::LOCAL_RECOVER, sdl, PtrVT, OffsetSym);
5249 // Add the offset to the FP.
5250 Value *FP = I.getArgOperand(1);
5251 SDValue FPVal = getValue(FP);
5252 SDValue Add = DAG.getNode(ISD::ADD, sdl, PtrVT, FPVal, OffsetVal);
5258 case Intrinsic::eh_exceptionpointer:
5259 case Intrinsic::eh_exceptioncode: {
5260 // Get the exception pointer vreg, copy from it, and resize it to fit.
5261 const auto *CPI = cast<CatchPadInst>(I.getArgOperand(0));
5262 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
5263 const TargetRegisterClass *PtrRC = TLI.getRegClassFor(PtrVT);
5264 unsigned VReg = FuncInfo.getCatchPadExceptionPointerVReg(CPI, PtrRC);
5266 DAG.getCopyFromReg(DAG.getEntryNode(), getCurSDLoc(), VReg, PtrVT);
5267 if (Intrinsic == Intrinsic::eh_exceptioncode)
5268 N = DAG.getZExtOrTrunc(N, getCurSDLoc(), MVT::i32);
5275 std::pair<SDValue, SDValue>
5276 SelectionDAGBuilder::lowerInvokable(TargetLowering::CallLoweringInfo &CLI,
5277 const BasicBlock *EHPadBB) {
5278 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5279 MCSymbol *BeginLabel = nullptr;
5282 // Insert a label before the invoke call to mark the try range. This can be
5283 // used to detect deletion of the invoke via the MachineModuleInfo.
5284 BeginLabel = MMI.getContext().createTempSymbol();
5286 // For SjLj, keep track of which landing pads go with which invokes
5287 // so as to maintain the ordering of pads in the LSDA.
5288 unsigned CallSiteIndex = MMI.getCurrentCallSite();
5289 if (CallSiteIndex) {
5290 MMI.setCallSiteBeginLabel(BeginLabel, CallSiteIndex);
5291 LPadToCallSiteMap[FuncInfo.MBBMap[EHPadBB]].push_back(CallSiteIndex);
5293 // Now that the call site is handled, stop tracking it.
5294 MMI.setCurrentCallSite(0);
5297 // Both PendingLoads and PendingExports must be flushed here;
5298 // this call might not return.
5300 DAG.setRoot(DAG.getEHLabel(getCurSDLoc(), getControlRoot(), BeginLabel));
5302 CLI.setChain(getRoot());
5304 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5305 std::pair<SDValue, SDValue> Result = TLI.LowerCallTo(CLI);
5307 assert((CLI.IsTailCall || Result.second.getNode()) &&
5308 "Non-null chain expected with non-tail call!");
5309 assert((Result.second.getNode() || !Result.first.getNode()) &&
5310 "Null value expected with tail call!");
5312 if (!Result.second.getNode()) {
5313 // As a special case, a null chain means that a tail call has been emitted
5314 // and the DAG root is already updated.
5317 // Since there's no actual continuation from this block, nothing can be
5318 // relying on us setting vregs for them.
5319 PendingExports.clear();
5321 DAG.setRoot(Result.second);
5325 // Insert a label at the end of the invoke call to mark the try range. This
5326 // can be used to detect deletion of the invoke via the MachineModuleInfo.
5327 MCSymbol *EndLabel = MMI.getContext().createTempSymbol();
5328 DAG.setRoot(DAG.getEHLabel(getCurSDLoc(), getRoot(), EndLabel));
5330 // Inform MachineModuleInfo of range.
5331 if (MMI.hasEHFunclets()) {
5332 WinEHFuncInfo &EHInfo =
5333 MMI.getWinEHFuncInfo(DAG.getMachineFunction().getFunction());
5334 EHInfo.addIPToStateRange(EHPadBB, BeginLabel, EndLabel);
5336 MMI.addInvoke(FuncInfo.MBBMap[EHPadBB], BeginLabel, EndLabel);
5343 void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee,
5345 const BasicBlock *EHPadBB) {
5346 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
5347 FunctionType *FTy = cast<FunctionType>(PT->getElementType());
5348 Type *RetTy = FTy->getReturnType();
5350 TargetLowering::ArgListTy Args;
5351 TargetLowering::ArgListEntry Entry;
5352 Args.reserve(CS.arg_size());
5354 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
5356 const Value *V = *i;
5359 if (V->getType()->isEmptyTy())
5362 SDValue ArgNode = getValue(V);
5363 Entry.Node = ArgNode; Entry.Ty = V->getType();
5365 // Skip the first return-type Attribute to get to params.
5366 Entry.setAttributes(&CS, i - CS.arg_begin() + 1);
5367 Args.push_back(Entry);
5369 // If we have an explicit sret argument that is an Instruction, (i.e., it
5370 // might point to function-local memory), we can't meaningfully tail-call.
5371 if (Entry.isSRet && isa<Instruction>(V))
5375 // Check if target-independent constraints permit a tail call here.
5376 // Target-dependent constraints are checked within TLI->LowerCallTo.
5377 if (isTailCall && !isInTailCallPosition(CS, DAG.getTarget()))
5380 TargetLowering::CallLoweringInfo CLI(DAG);
5381 CLI.setDebugLoc(getCurSDLoc()).setChain(getRoot())
5382 .setCallee(RetTy, FTy, Callee, std::move(Args), CS)
5383 .setTailCall(isTailCall);
5384 std::pair<SDValue, SDValue> Result = lowerInvokable(CLI, EHPadBB);
5386 if (Result.first.getNode())
5387 setValue(CS.getInstruction(), Result.first);
5390 /// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the
5391 /// value is equal or not-equal to zero.
5392 static bool IsOnlyUsedInZeroEqualityComparison(const Value *V) {
5393 for (const User *U : V->users()) {
5394 if (const ICmpInst *IC = dyn_cast<ICmpInst>(U))
5395 if (IC->isEquality())
5396 if (const Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
5397 if (C->isNullValue())
5399 // Unknown instruction.
5405 static SDValue getMemCmpLoad(const Value *PtrVal, MVT LoadVT,
5407 SelectionDAGBuilder &Builder) {
5409 // Check to see if this load can be trivially constant folded, e.g. if the
5410 // input is from a string literal.
5411 if (const Constant *LoadInput = dyn_cast<Constant>(PtrVal)) {
5412 // Cast pointer to the type we really want to load.
5413 LoadInput = ConstantExpr::getBitCast(const_cast<Constant *>(LoadInput),
5414 PointerType::getUnqual(LoadTy));
5416 if (const Constant *LoadCst = ConstantFoldLoadFromConstPtr(
5417 const_cast<Constant *>(LoadInput), *Builder.DL))
5418 return Builder.getValue(LoadCst);
5421 // Otherwise, we have to emit the load. If the pointer is to unfoldable but
5422 // still constant memory, the input chain can be the entry node.
5424 bool ConstantMemory = false;
5426 // Do not serialize (non-volatile) loads of constant memory with anything.
5427 if (Builder.AA->pointsToConstantMemory(PtrVal)) {
5428 Root = Builder.DAG.getEntryNode();
5429 ConstantMemory = true;
5431 // Do not serialize non-volatile loads against each other.
5432 Root = Builder.DAG.getRoot();
5435 SDValue Ptr = Builder.getValue(PtrVal);
5436 SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurSDLoc(), Root,
5437 Ptr, MachinePointerInfo(PtrVal),
5439 false /*nontemporal*/,
5440 false /*isinvariant*/, 1 /* align=1 */);
5442 if (!ConstantMemory)
5443 Builder.PendingLoads.push_back(LoadVal.getValue(1));
5447 /// processIntegerCallValue - Record the value for an instruction that
5448 /// produces an integer result, converting the type where necessary.
5449 void SelectionDAGBuilder::processIntegerCallValue(const Instruction &I,
5452 EVT VT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
5455 Value = DAG.getSExtOrTrunc(Value, getCurSDLoc(), VT);
5457 Value = DAG.getZExtOrTrunc(Value, getCurSDLoc(), VT);
5458 setValue(&I, Value);
5461 /// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form.
5462 /// If so, return true and lower it, otherwise return false and it will be
5463 /// lowered like a normal call.
5464 bool SelectionDAGBuilder::visitMemCmpCall(const CallInst &I) {
5465 // Verify that the prototype makes sense. int memcmp(void*,void*,size_t)
5466 if (I.getNumArgOperands() != 3)
5469 const Value *LHS = I.getArgOperand(0), *RHS = I.getArgOperand(1);
5470 if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() ||
5471 !I.getArgOperand(2)->getType()->isIntegerTy() ||
5472 !I.getType()->isIntegerTy())
5475 const Value *Size = I.getArgOperand(2);
5476 const ConstantInt *CSize = dyn_cast<ConstantInt>(Size);
5477 if (CSize && CSize->getZExtValue() == 0) {
5478 EVT CallVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
5480 setValue(&I, DAG.getConstant(0, getCurSDLoc(), CallVT));
5484 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5485 std::pair<SDValue, SDValue> Res =
5486 TSI.EmitTargetCodeForMemcmp(DAG, getCurSDLoc(), DAG.getRoot(),
5487 getValue(LHS), getValue(RHS), getValue(Size),
5488 MachinePointerInfo(LHS),
5489 MachinePointerInfo(RHS));
5490 if (Res.first.getNode()) {
5491 processIntegerCallValue(I, Res.first, true);
5492 PendingLoads.push_back(Res.second);
5496 // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0
5497 // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0
5498 if (CSize && IsOnlyUsedInZeroEqualityComparison(&I)) {
5499 bool ActuallyDoIt = true;
5502 switch (CSize->getZExtValue()) {
5504 LoadVT = MVT::Other;
5506 ActuallyDoIt = false;
5510 LoadTy = Type::getInt16Ty(CSize->getContext());
5514 LoadTy = Type::getInt32Ty(CSize->getContext());
5518 LoadTy = Type::getInt64Ty(CSize->getContext());
5522 LoadVT = MVT::v4i32;
5523 LoadTy = Type::getInt32Ty(CSize->getContext());
5524 LoadTy = VectorType::get(LoadTy, 4);
5529 // This turns into unaligned loads. We only do this if the target natively
5530 // supports the MVT we'll be loading or if it is small enough (<= 4) that
5531 // we'll only produce a small number of byte loads.
5533 // Require that we can find a legal MVT, and only do this if the target
5534 // supports unaligned loads of that type. Expanding into byte loads would
5536 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5537 if (ActuallyDoIt && CSize->getZExtValue() > 4) {
5538 unsigned DstAS = LHS->getType()->getPointerAddressSpace();
5539 unsigned SrcAS = RHS->getType()->getPointerAddressSpace();
5540 // TODO: Handle 5 byte compare as 4-byte + 1 byte.
5541 // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads.
5542 // TODO: Check alignment of src and dest ptrs.
5543 if (!TLI.isTypeLegal(LoadVT) ||
5544 !TLI.allowsMisalignedMemoryAccesses(LoadVT, SrcAS) ||
5545 !TLI.allowsMisalignedMemoryAccesses(LoadVT, DstAS))
5546 ActuallyDoIt = false;
5550 SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this);
5551 SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this);
5553 SDValue Res = DAG.getSetCC(getCurSDLoc(), MVT::i1, LHSVal, RHSVal,
5555 processIntegerCallValue(I, Res, false);
5564 /// visitMemChrCall -- See if we can lower a memchr call into an optimized
5565 /// form. If so, return true and lower it, otherwise return false and it
5566 /// will be lowered like a normal call.
5567 bool SelectionDAGBuilder::visitMemChrCall(const CallInst &I) {
5568 // Verify that the prototype makes sense. void *memchr(void *, int, size_t)
5569 if (I.getNumArgOperands() != 3)
5572 const Value *Src = I.getArgOperand(0);
5573 const Value *Char = I.getArgOperand(1);
5574 const Value *Length = I.getArgOperand(2);
5575 if (!Src->getType()->isPointerTy() ||
5576 !Char->getType()->isIntegerTy() ||
5577 !Length->getType()->isIntegerTy() ||
5578 !I.getType()->isPointerTy())
5581 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5582 std::pair<SDValue, SDValue> Res =
5583 TSI.EmitTargetCodeForMemchr(DAG, getCurSDLoc(), DAG.getRoot(),
5584 getValue(Src), getValue(Char), getValue(Length),
5585 MachinePointerInfo(Src));
5586 if (Res.first.getNode()) {
5587 setValue(&I, Res.first);
5588 PendingLoads.push_back(Res.second);
5595 /// visitStrCpyCall -- See if we can lower a strcpy or stpcpy call into an
5596 /// optimized form. If so, return true and lower it, otherwise return false
5597 /// and it will be lowered like a normal call.
5598 bool SelectionDAGBuilder::visitStrCpyCall(const CallInst &I, bool isStpcpy) {
5599 // Verify that the prototype makes sense. char *strcpy(char *, char *)
5600 if (I.getNumArgOperands() != 2)
5603 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5604 if (!Arg0->getType()->isPointerTy() ||
5605 !Arg1->getType()->isPointerTy() ||
5606 !I.getType()->isPointerTy())
5609 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5610 std::pair<SDValue, SDValue> Res =
5611 TSI.EmitTargetCodeForStrcpy(DAG, getCurSDLoc(), getRoot(),
5612 getValue(Arg0), getValue(Arg1),
5613 MachinePointerInfo(Arg0),
5614 MachinePointerInfo(Arg1), isStpcpy);
5615 if (Res.first.getNode()) {
5616 setValue(&I, Res.first);
5617 DAG.setRoot(Res.second);
5624 /// visitStrCmpCall - See if we can lower a call to strcmp in an optimized form.
5625 /// If so, return true and lower it, otherwise return false and it will be
5626 /// lowered like a normal call.
5627 bool SelectionDAGBuilder::visitStrCmpCall(const CallInst &I) {
5628 // Verify that the prototype makes sense. int strcmp(void*,void*)
5629 if (I.getNumArgOperands() != 2)
5632 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5633 if (!Arg0->getType()->isPointerTy() ||
5634 !Arg1->getType()->isPointerTy() ||
5635 !I.getType()->isIntegerTy())
5638 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5639 std::pair<SDValue, SDValue> Res =
5640 TSI.EmitTargetCodeForStrcmp(DAG, getCurSDLoc(), DAG.getRoot(),
5641 getValue(Arg0), getValue(Arg1),
5642 MachinePointerInfo(Arg0),
5643 MachinePointerInfo(Arg1));
5644 if (Res.first.getNode()) {
5645 processIntegerCallValue(I, Res.first, true);
5646 PendingLoads.push_back(Res.second);
5653 /// visitStrLenCall -- See if we can lower a strlen call into an optimized
5654 /// form. If so, return true and lower it, otherwise return false and it
5655 /// will be lowered like a normal call.
5656 bool SelectionDAGBuilder::visitStrLenCall(const CallInst &I) {
5657 // Verify that the prototype makes sense. size_t strlen(char *)
5658 if (I.getNumArgOperands() != 1)
5661 const Value *Arg0 = I.getArgOperand(0);
5662 if (!Arg0->getType()->isPointerTy() || !I.getType()->isIntegerTy())
5665 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5666 std::pair<SDValue, SDValue> Res =
5667 TSI.EmitTargetCodeForStrlen(DAG, getCurSDLoc(), DAG.getRoot(),
5668 getValue(Arg0), MachinePointerInfo(Arg0));
5669 if (Res.first.getNode()) {
5670 processIntegerCallValue(I, Res.first, false);
5671 PendingLoads.push_back(Res.second);
5678 /// visitStrNLenCall -- See if we can lower a strnlen call into an optimized
5679 /// form. If so, return true and lower it, otherwise return false and it
5680 /// will be lowered like a normal call.
5681 bool SelectionDAGBuilder::visitStrNLenCall(const CallInst &I) {
5682 // Verify that the prototype makes sense. size_t strnlen(char *, size_t)
5683 if (I.getNumArgOperands() != 2)
5686 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5687 if (!Arg0->getType()->isPointerTy() ||
5688 !Arg1->getType()->isIntegerTy() ||
5689 !I.getType()->isIntegerTy())
5692 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5693 std::pair<SDValue, SDValue> Res =
5694 TSI.EmitTargetCodeForStrnlen(DAG, getCurSDLoc(), DAG.getRoot(),
5695 getValue(Arg0), getValue(Arg1),
5696 MachinePointerInfo(Arg0));
5697 if (Res.first.getNode()) {
5698 processIntegerCallValue(I, Res.first, false);
5699 PendingLoads.push_back(Res.second);
5706 /// visitUnaryFloatCall - If a call instruction is a unary floating-point
5707 /// operation (as expected), translate it to an SDNode with the specified opcode
5708 /// and return true.
5709 bool SelectionDAGBuilder::visitUnaryFloatCall(const CallInst &I,
5711 // Sanity check that it really is a unary floating-point call.
5712 if (I.getNumArgOperands() != 1 ||
5713 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
5714 I.getType() != I.getArgOperand(0)->getType() ||
5715 !I.onlyReadsMemory())
5718 SDValue Tmp = getValue(I.getArgOperand(0));
5719 setValue(&I, DAG.getNode(Opcode, getCurSDLoc(), Tmp.getValueType(), Tmp));
5723 /// visitBinaryFloatCall - If a call instruction is a binary floating-point
5724 /// operation (as expected), translate it to an SDNode with the specified opcode
5725 /// and return true.
5726 bool SelectionDAGBuilder::visitBinaryFloatCall(const CallInst &I,
5728 // Sanity check that it really is a binary floating-point call.
5729 if (I.getNumArgOperands() != 2 ||
5730 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
5731 I.getType() != I.getArgOperand(0)->getType() ||
5732 I.getType() != I.getArgOperand(1)->getType() ||
5733 !I.onlyReadsMemory())
5736 SDValue Tmp0 = getValue(I.getArgOperand(0));
5737 SDValue Tmp1 = getValue(I.getArgOperand(1));
5738 EVT VT = Tmp0.getValueType();
5739 setValue(&I, DAG.getNode(Opcode, getCurSDLoc(), VT, Tmp0, Tmp1));
5743 void SelectionDAGBuilder::visitCall(const CallInst &I) {
5744 // Handle inline assembly differently.
5745 if (isa<InlineAsm>(I.getCalledValue())) {
5750 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5751 ComputeUsesVAFloatArgument(I, &MMI);
5753 const char *RenameFn = nullptr;
5754 if (Function *F = I.getCalledFunction()) {
5755 if (F->isDeclaration()) {
5756 if (const TargetIntrinsicInfo *II = TM.getIntrinsicInfo()) {
5757 if (unsigned IID = II->getIntrinsicID(F)) {
5758 RenameFn = visitIntrinsicCall(I, IID);
5763 if (Intrinsic::ID IID = F->getIntrinsicID()) {
5764 RenameFn = visitIntrinsicCall(I, IID);
5770 // Check for well-known libc/libm calls. If the function is internal, it
5771 // can't be a library call.
5773 if (!F->hasLocalLinkage() && F->hasName() &&
5774 LibInfo->getLibFunc(F->getName(), Func) &&
5775 LibInfo->hasOptimizedCodeGen(Func)) {
5778 case LibFunc::copysign:
5779 case LibFunc::copysignf:
5780 case LibFunc::copysignl:
5781 if (I.getNumArgOperands() == 2 && // Basic sanity checks.
5782 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5783 I.getType() == I.getArgOperand(0)->getType() &&
5784 I.getType() == I.getArgOperand(1)->getType() &&
5785 I.onlyReadsMemory()) {
5786 SDValue LHS = getValue(I.getArgOperand(0));
5787 SDValue RHS = getValue(I.getArgOperand(1));
5788 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurSDLoc(),
5789 LHS.getValueType(), LHS, RHS));
5794 case LibFunc::fabsf:
5795 case LibFunc::fabsl:
5796 if (visitUnaryFloatCall(I, ISD::FABS))
5800 case LibFunc::fminf:
5801 case LibFunc::fminl:
5802 if (visitBinaryFloatCall(I, ISD::FMINNUM))
5806 case LibFunc::fmaxf:
5807 case LibFunc::fmaxl:
5808 if (visitBinaryFloatCall(I, ISD::FMAXNUM))
5814 if (visitUnaryFloatCall(I, ISD::FSIN))
5820 if (visitUnaryFloatCall(I, ISD::FCOS))
5824 case LibFunc::sqrtf:
5825 case LibFunc::sqrtl:
5826 case LibFunc::sqrt_finite:
5827 case LibFunc::sqrtf_finite:
5828 case LibFunc::sqrtl_finite:
5829 if (visitUnaryFloatCall(I, ISD::FSQRT))
5832 case LibFunc::floor:
5833 case LibFunc::floorf:
5834 case LibFunc::floorl:
5835 if (visitUnaryFloatCall(I, ISD::FFLOOR))
5838 case LibFunc::nearbyint:
5839 case LibFunc::nearbyintf:
5840 case LibFunc::nearbyintl:
5841 if (visitUnaryFloatCall(I, ISD::FNEARBYINT))
5845 case LibFunc::ceilf:
5846 case LibFunc::ceill:
5847 if (visitUnaryFloatCall(I, ISD::FCEIL))
5851 case LibFunc::rintf:
5852 case LibFunc::rintl:
5853 if (visitUnaryFloatCall(I, ISD::FRINT))
5856 case LibFunc::round:
5857 case LibFunc::roundf:
5858 case LibFunc::roundl:
5859 if (visitUnaryFloatCall(I, ISD::FROUND))
5862 case LibFunc::trunc:
5863 case LibFunc::truncf:
5864 case LibFunc::truncl:
5865 if (visitUnaryFloatCall(I, ISD::FTRUNC))
5869 case LibFunc::log2f:
5870 case LibFunc::log2l:
5871 if (visitUnaryFloatCall(I, ISD::FLOG2))
5875 case LibFunc::exp2f:
5876 case LibFunc::exp2l:
5877 if (visitUnaryFloatCall(I, ISD::FEXP2))
5880 case LibFunc::memcmp:
5881 if (visitMemCmpCall(I))
5884 case LibFunc::memchr:
5885 if (visitMemChrCall(I))
5888 case LibFunc::strcpy:
5889 if (visitStrCpyCall(I, false))
5892 case LibFunc::stpcpy:
5893 if (visitStrCpyCall(I, true))
5896 case LibFunc::strcmp:
5897 if (visitStrCmpCall(I))
5900 case LibFunc::strlen:
5901 if (visitStrLenCall(I))
5904 case LibFunc::strnlen:
5905 if (visitStrNLenCall(I))
5914 Callee = getValue(I.getCalledValue());
5916 Callee = DAG.getExternalSymbol(
5918 DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout()));
5920 // Check if we can potentially perform a tail call. More detailed checking is
5921 // be done within LowerCallTo, after more information about the call is known.
5922 LowerCallTo(&I, Callee, I.isTailCall());
5927 /// AsmOperandInfo - This contains information for each constraint that we are
5929 class SDISelAsmOperandInfo : public TargetLowering::AsmOperandInfo {
5931 /// CallOperand - If this is the result output operand or a clobber
5932 /// this is null, otherwise it is the incoming operand to the CallInst.
5933 /// This gets modified as the asm is processed.
5934 SDValue CallOperand;
5936 /// AssignedRegs - If this is a register or register class operand, this
5937 /// contains the set of register corresponding to the operand.
5938 RegsForValue AssignedRegs;
5940 explicit SDISelAsmOperandInfo(const TargetLowering::AsmOperandInfo &info)
5941 : TargetLowering::AsmOperandInfo(info), CallOperand(nullptr,0) {
5944 /// getCallOperandValEVT - Return the EVT of the Value* that this operand
5945 /// corresponds to. If there is no Value* for this operand, it returns
5947 EVT getCallOperandValEVT(LLVMContext &Context, const TargetLowering &TLI,
5948 const DataLayout &DL) const {
5949 if (!CallOperandVal) return MVT::Other;
5951 if (isa<BasicBlock>(CallOperandVal))
5952 return TLI.getPointerTy(DL);
5954 llvm::Type *OpTy = CallOperandVal->getType();
5956 // FIXME: code duplicated from TargetLowering::ParseConstraints().
5957 // If this is an indirect operand, the operand is a pointer to the
5960 llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
5962 report_fatal_error("Indirect operand for inline asm not a pointer!");
5963 OpTy = PtrTy->getElementType();
5966 // Look for vector wrapped in a struct. e.g. { <16 x i8> }.
5967 if (StructType *STy = dyn_cast<StructType>(OpTy))
5968 if (STy->getNumElements() == 1)
5969 OpTy = STy->getElementType(0);
5971 // If OpTy is not a single value, it may be a struct/union that we
5972 // can tile with integers.
5973 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
5974 unsigned BitSize = DL.getTypeSizeInBits(OpTy);
5983 OpTy = IntegerType::get(Context, BitSize);
5988 return TLI.getValueType(DL, OpTy, true);
5992 typedef SmallVector<SDISelAsmOperandInfo,16> SDISelAsmOperandInfoVector;
5994 } // end anonymous namespace
5996 /// GetRegistersForValue - Assign registers (virtual or physical) for the
5997 /// specified operand. We prefer to assign virtual registers, to allow the
5998 /// register allocator to handle the assignment process. However, if the asm
5999 /// uses features that we can't model on machineinstrs, we have SDISel do the
6000 /// allocation. This produces generally horrible, but correct, code.
6002 /// OpInfo describes the operand.
6004 static void GetRegistersForValue(SelectionDAG &DAG,
6005 const TargetLowering &TLI,
6007 SDISelAsmOperandInfo &OpInfo) {
6008 LLVMContext &Context = *DAG.getContext();
6010 MachineFunction &MF = DAG.getMachineFunction();
6011 SmallVector<unsigned, 4> Regs;
6013 // If this is a constraint for a single physreg, or a constraint for a
6014 // register class, find it.
6015 std::pair<unsigned, const TargetRegisterClass *> PhysReg =
6016 TLI.getRegForInlineAsmConstraint(MF.getSubtarget().getRegisterInfo(),
6017 OpInfo.ConstraintCode,
6018 OpInfo.ConstraintVT);
6020 unsigned NumRegs = 1;
6021 if (OpInfo.ConstraintVT != MVT::Other) {
6022 // If this is a FP input in an integer register (or visa versa) insert a bit
6023 // cast of the input value. More generally, handle any case where the input
6024 // value disagrees with the register class we plan to stick this in.
6025 if (OpInfo.Type == InlineAsm::isInput &&
6026 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) {
6027 // Try to convert to the first EVT that the reg class contains. If the
6028 // types are identical size, use a bitcast to convert (e.g. two differing
6030 MVT RegVT = *PhysReg.second->vt_begin();
6031 if (RegVT.getSizeInBits() == OpInfo.CallOperand.getValueSizeInBits()) {
6032 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
6033 RegVT, OpInfo.CallOperand);
6034 OpInfo.ConstraintVT = RegVT;
6035 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) {
6036 // If the input is a FP value and we want it in FP registers, do a
6037 // bitcast to the corresponding integer type. This turns an f64 value
6038 // into i64, which can be passed with two i32 values on a 32-bit
6040 RegVT = MVT::getIntegerVT(OpInfo.ConstraintVT.getSizeInBits());
6041 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
6042 RegVT, OpInfo.CallOperand);
6043 OpInfo.ConstraintVT = RegVT;
6047 NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT);
6051 EVT ValueVT = OpInfo.ConstraintVT;
6053 // If this is a constraint for a specific physical register, like {r17},
6055 if (unsigned AssignedReg = PhysReg.first) {
6056 const TargetRegisterClass *RC = PhysReg.second;
6057 if (OpInfo.ConstraintVT == MVT::Other)
6058 ValueVT = *RC->vt_begin();
6060 // Get the actual register value type. This is important, because the user
6061 // may have asked for (e.g.) the AX register in i32 type. We need to
6062 // remember that AX is actually i16 to get the right extension.
6063 RegVT = *RC->vt_begin();
6065 // This is a explicit reference to a physical register.
6066 Regs.push_back(AssignedReg);
6068 // If this is an expanded reference, add the rest of the regs to Regs.
6070 TargetRegisterClass::iterator I = RC->begin();
6071 for (; *I != AssignedReg; ++I)
6072 assert(I != RC->end() && "Didn't find reg!");
6074 // Already added the first reg.
6076 for (; NumRegs; --NumRegs, ++I) {
6077 assert(I != RC->end() && "Ran out of registers to allocate!");
6082 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
6086 // Otherwise, if this was a reference to an LLVM register class, create vregs
6087 // for this reference.
6088 if (const TargetRegisterClass *RC = PhysReg.second) {
6089 RegVT = *RC->vt_begin();
6090 if (OpInfo.ConstraintVT == MVT::Other)
6093 // Create the appropriate number of virtual registers.
6094 MachineRegisterInfo &RegInfo = MF.getRegInfo();
6095 for (; NumRegs; --NumRegs)
6096 Regs.push_back(RegInfo.createVirtualRegister(RC));
6098 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
6102 // Otherwise, we couldn't allocate enough registers for this.
6105 /// visitInlineAsm - Handle a call to an InlineAsm object.
6107 void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) {
6108 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
6110 /// ConstraintOperands - Information about all of the constraints.
6111 SDISelAsmOperandInfoVector ConstraintOperands;
6113 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6114 TargetLowering::AsmOperandInfoVector TargetConstraints = TLI.ParseConstraints(
6115 DAG.getDataLayout(), DAG.getSubtarget().getRegisterInfo(), CS);
6117 bool hasMemory = false;
6119 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
6120 unsigned ResNo = 0; // ResNo - The result number of the next output.
6121 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
6122 ConstraintOperands.push_back(SDISelAsmOperandInfo(TargetConstraints[i]));
6123 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
6125 MVT OpVT = MVT::Other;
6127 // Compute the value type for each operand.
6128 switch (OpInfo.Type) {
6129 case InlineAsm::isOutput:
6130 // Indirect outputs just consume an argument.
6131 if (OpInfo.isIndirect) {
6132 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
6136 // The return value of the call is this value. As such, there is no
6137 // corresponding argument.
6138 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
6139 if (StructType *STy = dyn_cast<StructType>(CS.getType())) {
6140 OpVT = TLI.getSimpleValueType(DAG.getDataLayout(),
6141 STy->getElementType(ResNo));
6143 assert(ResNo == 0 && "Asm only has one result!");
6144 OpVT = TLI.getSimpleValueType(DAG.getDataLayout(), CS.getType());
6148 case InlineAsm::isInput:
6149 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
6151 case InlineAsm::isClobber:
6156 // If this is an input or an indirect output, process the call argument.
6157 // BasicBlocks are labels, currently appearing only in asm's.
6158 if (OpInfo.CallOperandVal) {
6159 if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) {
6160 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]);
6162 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
6165 OpVT = OpInfo.getCallOperandValEVT(*DAG.getContext(), TLI,
6166 DAG.getDataLayout()).getSimpleVT();
6169 OpInfo.ConstraintVT = OpVT;
6171 // Indirect operand accesses access memory.
6172 if (OpInfo.isIndirect)
6175 for (unsigned j = 0, ee = OpInfo.Codes.size(); j != ee; ++j) {
6176 TargetLowering::ConstraintType
6177 CType = TLI.getConstraintType(OpInfo.Codes[j]);
6178 if (CType == TargetLowering::C_Memory) {
6186 SDValue Chain, Flag;
6188 // We won't need to flush pending loads if this asm doesn't touch
6189 // memory and is nonvolatile.
6190 if (hasMemory || IA->hasSideEffects())
6193 Chain = DAG.getRoot();
6195 // Second pass over the constraints: compute which constraint option to use
6196 // and assign registers to constraints that want a specific physreg.
6197 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6198 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6200 // If this is an output operand with a matching input operand, look up the
6201 // matching input. If their types mismatch, e.g. one is an integer, the
6202 // other is floating point, or their sizes are different, flag it as an
6204 if (OpInfo.hasMatchingInput()) {
6205 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
6207 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
6208 const TargetRegisterInfo *TRI = DAG.getSubtarget().getRegisterInfo();
6209 std::pair<unsigned, const TargetRegisterClass *> MatchRC =
6210 TLI.getRegForInlineAsmConstraint(TRI, OpInfo.ConstraintCode,
6211 OpInfo.ConstraintVT);
6212 std::pair<unsigned, const TargetRegisterClass *> InputRC =
6213 TLI.getRegForInlineAsmConstraint(TRI, Input.ConstraintCode,
6214 Input.ConstraintVT);
6215 if ((OpInfo.ConstraintVT.isInteger() !=
6216 Input.ConstraintVT.isInteger()) ||
6217 (MatchRC.second != InputRC.second)) {
6218 report_fatal_error("Unsupported asm: input constraint"
6219 " with a matching output constraint of"
6220 " incompatible type!");
6222 Input.ConstraintVT = OpInfo.ConstraintVT;
6226 // Compute the constraint code and ConstraintType to use.
6227 TLI.ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, &DAG);
6229 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
6230 OpInfo.Type == InlineAsm::isClobber)
6233 // If this is a memory input, and if the operand is not indirect, do what we
6234 // need to to provide an address for the memory input.
6235 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
6236 !OpInfo.isIndirect) {
6237 assert((OpInfo.isMultipleAlternative ||
6238 (OpInfo.Type == InlineAsm::isInput)) &&
6239 "Can only indirectify direct input operands!");
6241 // Memory operands really want the address of the value. If we don't have
6242 // an indirect input, put it in the constpool if we can, otherwise spill
6243 // it to a stack slot.
6244 // TODO: This isn't quite right. We need to handle these according to
6245 // the addressing mode that the constraint wants. Also, this may take
6246 // an additional register for the computation and we don't want that
6249 // If the operand is a float, integer, or vector constant, spill to a
6250 // constant pool entry to get its address.
6251 const Value *OpVal = OpInfo.CallOperandVal;
6252 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
6253 isa<ConstantVector>(OpVal) || isa<ConstantDataVector>(OpVal)) {
6254 OpInfo.CallOperand = DAG.getConstantPool(
6255 cast<Constant>(OpVal), TLI.getPointerTy(DAG.getDataLayout()));
6257 // Otherwise, create a stack slot and emit a store to it before the
6259 Type *Ty = OpVal->getType();
6260 auto &DL = DAG.getDataLayout();
6261 uint64_t TySize = DL.getTypeAllocSize(Ty);
6262 unsigned Align = DL.getPrefTypeAlignment(Ty);
6263 MachineFunction &MF = DAG.getMachineFunction();
6264 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
6266 DAG.getFrameIndex(SSFI, TLI.getPointerTy(DAG.getDataLayout()));
6267 Chain = DAG.getStore(
6268 Chain, getCurSDLoc(), OpInfo.CallOperand, StackSlot,
6269 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
6271 OpInfo.CallOperand = StackSlot;
6274 // There is no longer a Value* corresponding to this operand.
6275 OpInfo.CallOperandVal = nullptr;
6277 // It is now an indirect operand.
6278 OpInfo.isIndirect = true;
6281 // If this constraint is for a specific register, allocate it before
6283 if (OpInfo.ConstraintType == TargetLowering::C_Register)
6284 GetRegistersForValue(DAG, TLI, getCurSDLoc(), OpInfo);
6287 // Second pass - Loop over all of the operands, assigning virtual or physregs
6288 // to register class operands.
6289 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6290 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6292 // C_Register operands have already been allocated, Other/Memory don't need
6294 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
6295 GetRegistersForValue(DAG, TLI, getCurSDLoc(), OpInfo);
6298 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
6299 std::vector<SDValue> AsmNodeOperands;
6300 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain
6301 AsmNodeOperands.push_back(DAG.getTargetExternalSymbol(
6302 IA->getAsmString().c_str(), TLI.getPointerTy(DAG.getDataLayout())));
6304 // If we have a !srcloc metadata node associated with it, we want to attach
6305 // this to the ultimately generated inline asm machineinstr. To do this, we
6306 // pass in the third operand as this (potentially null) inline asm MDNode.
6307 const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc");
6308 AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc));
6310 // Remember the HasSideEffect, AlignStack, AsmDialect, MayLoad and MayStore
6311 // bits as operand 3.
6312 unsigned ExtraInfo = 0;
6313 if (IA->hasSideEffects())
6314 ExtraInfo |= InlineAsm::Extra_HasSideEffects;
6315 if (IA->isAlignStack())
6316 ExtraInfo |= InlineAsm::Extra_IsAlignStack;
6317 // Set the asm dialect.
6318 ExtraInfo |= IA->getDialect() * InlineAsm::Extra_AsmDialect;
6320 // Determine if this InlineAsm MayLoad or MayStore based on the constraints.
6321 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
6322 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
6324 // Compute the constraint code and ConstraintType to use.
6325 TLI.ComputeConstraintToUse(OpInfo, SDValue());
6327 // Ideally, we would only check against memory constraints. However, the
6328 // meaning of an other constraint can be target-specific and we can't easily
6329 // reason about it. Therefore, be conservative and set MayLoad/MayStore
6330 // for other constriants as well.
6331 if (OpInfo.ConstraintType == TargetLowering::C_Memory ||
6332 OpInfo.ConstraintType == TargetLowering::C_Other) {
6333 if (OpInfo.Type == InlineAsm::isInput)
6334 ExtraInfo |= InlineAsm::Extra_MayLoad;
6335 else if (OpInfo.Type == InlineAsm::isOutput)
6336 ExtraInfo |= InlineAsm::Extra_MayStore;
6337 else if (OpInfo.Type == InlineAsm::isClobber)
6338 ExtraInfo |= (InlineAsm::Extra_MayLoad | InlineAsm::Extra_MayStore);
6342 AsmNodeOperands.push_back(DAG.getTargetConstant(
6343 ExtraInfo, getCurSDLoc(), TLI.getPointerTy(DAG.getDataLayout())));
6345 // Loop over all of the inputs, copying the operand values into the
6346 // appropriate registers and processing the output regs.
6347 RegsForValue RetValRegs;
6349 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
6350 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
6352 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6353 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6355 switch (OpInfo.Type) {
6356 case InlineAsm::isOutput: {
6357 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
6358 OpInfo.ConstraintType != TargetLowering::C_Register) {
6359 // Memory output, or 'other' output (e.g. 'X' constraint).
6360 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
6362 unsigned ConstraintID =
6363 TLI.getInlineAsmMemConstraint(OpInfo.ConstraintCode);
6364 assert(ConstraintID != InlineAsm::Constraint_Unknown &&
6365 "Failed to convert memory constraint code to constraint id.");
6367 // Add information to the INLINEASM node to know about this output.
6368 unsigned OpFlags = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
6369 OpFlags = InlineAsm::getFlagWordForMem(OpFlags, ConstraintID);
6370 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlags, getCurSDLoc(),
6372 AsmNodeOperands.push_back(OpInfo.CallOperand);
6376 // Otherwise, this is a register or register class output.
6378 // Copy the output from the appropriate register. Find a register that
6380 if (OpInfo.AssignedRegs.Regs.empty()) {
6381 LLVMContext &Ctx = *DAG.getContext();
6382 Ctx.emitError(CS.getInstruction(),
6383 "couldn't allocate output register for constraint '" +
6384 Twine(OpInfo.ConstraintCode) + "'");
6388 // If this is an indirect operand, store through the pointer after the
6390 if (OpInfo.isIndirect) {
6391 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
6392 OpInfo.CallOperandVal));
6394 // This is the result value of the call.
6395 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
6396 // Concatenate this output onto the outputs list.
6397 RetValRegs.append(OpInfo.AssignedRegs);
6400 // Add information to the INLINEASM node to know that this register is
6403 .AddInlineAsmOperands(OpInfo.isEarlyClobber
6404 ? InlineAsm::Kind_RegDefEarlyClobber
6405 : InlineAsm::Kind_RegDef,
6406 false, 0, getCurSDLoc(), DAG, AsmNodeOperands);
6409 case InlineAsm::isInput: {
6410 SDValue InOperandVal = OpInfo.CallOperand;
6412 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint?
6413 // If this is required to match an output register we have already set,
6414 // just use its register.
6415 unsigned OperandNo = OpInfo.getMatchedOperand();
6417 // Scan until we find the definition we already emitted of this operand.
6418 // When we find it, create a RegsForValue operand.
6419 unsigned CurOp = InlineAsm::Op_FirstOperand;
6420 for (; OperandNo; --OperandNo) {
6421 // Advance to the next operand.
6423 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6424 assert((InlineAsm::isRegDefKind(OpFlag) ||
6425 InlineAsm::isRegDefEarlyClobberKind(OpFlag) ||
6426 InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?");
6427 CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1;
6431 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6432 if (InlineAsm::isRegDefKind(OpFlag) ||
6433 InlineAsm::isRegDefEarlyClobberKind(OpFlag)) {
6434 // Add (OpFlag&0xffff)>>3 registers to MatchedRegs.
6435 if (OpInfo.isIndirect) {
6436 // This happens on gcc/testsuite/gcc.dg/pr8788-1.c
6437 LLVMContext &Ctx = *DAG.getContext();
6438 Ctx.emitError(CS.getInstruction(), "inline asm not supported yet:"
6439 " don't know how to handle tied "
6440 "indirect register inputs");
6444 RegsForValue MatchedRegs;
6445 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
6446 MVT RegVT = AsmNodeOperands[CurOp+1].getSimpleValueType();
6447 MatchedRegs.RegVTs.push_back(RegVT);
6448 MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo();
6449 for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag);
6451 if (const TargetRegisterClass *RC = TLI.getRegClassFor(RegVT))
6452 MatchedRegs.Regs.push_back(RegInfo.createVirtualRegister(RC));
6454 LLVMContext &Ctx = *DAG.getContext();
6455 Ctx.emitError(CS.getInstruction(),
6456 "inline asm error: This value"
6457 " type register class is not natively supported!");
6461 SDLoc dl = getCurSDLoc();
6462 // Use the produced MatchedRegs object to
6463 MatchedRegs.getCopyToRegs(InOperandVal, DAG, dl,
6464 Chain, &Flag, CS.getInstruction());
6465 MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse,
6466 true, OpInfo.getMatchedOperand(), dl,
6467 DAG, AsmNodeOperands);
6471 assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!");
6472 assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 &&
6473 "Unexpected number of operands");
6474 // Add information to the INLINEASM node to know about this input.
6475 // See InlineAsm.h isUseOperandTiedToDef.
6476 OpFlag = InlineAsm::convertMemFlagWordToMatchingFlagWord(OpFlag);
6477 OpFlag = InlineAsm::getFlagWordForMatchingOp(OpFlag,
6478 OpInfo.getMatchedOperand());
6479 AsmNodeOperands.push_back(DAG.getTargetConstant(
6480 OpFlag, getCurSDLoc(), TLI.getPointerTy(DAG.getDataLayout())));
6481 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]);
6485 // Treat indirect 'X' constraint as memory.
6486 if (OpInfo.ConstraintType == TargetLowering::C_Other &&
6488 OpInfo.ConstraintType = TargetLowering::C_Memory;
6490 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
6491 std::vector<SDValue> Ops;
6492 TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode,
6495 LLVMContext &Ctx = *DAG.getContext();
6496 Ctx.emitError(CS.getInstruction(),
6497 "invalid operand for inline asm constraint '" +
6498 Twine(OpInfo.ConstraintCode) + "'");
6502 // Add information to the INLINEASM node to know about this input.
6503 unsigned ResOpType =
6504 InlineAsm::getFlagWord(InlineAsm::Kind_Imm, Ops.size());
6505 AsmNodeOperands.push_back(DAG.getTargetConstant(
6506 ResOpType, getCurSDLoc(), TLI.getPointerTy(DAG.getDataLayout())));
6507 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
6511 if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
6512 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
6513 assert(InOperandVal.getValueType() ==
6514 TLI.getPointerTy(DAG.getDataLayout()) &&
6515 "Memory operands expect pointer values");
6517 unsigned ConstraintID =
6518 TLI.getInlineAsmMemConstraint(OpInfo.ConstraintCode);
6519 assert(ConstraintID != InlineAsm::Constraint_Unknown &&
6520 "Failed to convert memory constraint code to constraint id.");
6522 // Add information to the INLINEASM node to know about this input.
6523 unsigned ResOpType = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
6524 ResOpType = InlineAsm::getFlagWordForMem(ResOpType, ConstraintID);
6525 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
6528 AsmNodeOperands.push_back(InOperandVal);
6532 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
6533 OpInfo.ConstraintType == TargetLowering::C_Register) &&
6534 "Unknown constraint type!");
6536 // TODO: Support this.
6537 if (OpInfo.isIndirect) {
6538 LLVMContext &Ctx = *DAG.getContext();
6539 Ctx.emitError(CS.getInstruction(),
6540 "Don't know how to handle indirect register inputs yet "
6541 "for constraint '" +
6542 Twine(OpInfo.ConstraintCode) + "'");
6546 // Copy the input into the appropriate registers.
6547 if (OpInfo.AssignedRegs.Regs.empty()) {
6548 LLVMContext &Ctx = *DAG.getContext();
6549 Ctx.emitError(CS.getInstruction(),
6550 "couldn't allocate input reg for constraint '" +
6551 Twine(OpInfo.ConstraintCode) + "'");
6555 SDLoc dl = getCurSDLoc();
6557 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, dl,
6558 Chain, &Flag, CS.getInstruction());
6560 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, false, 0,
6561 dl, DAG, AsmNodeOperands);
6564 case InlineAsm::isClobber: {
6565 // Add the clobbered value to the operand list, so that the register
6566 // allocator is aware that the physreg got clobbered.
6567 if (!OpInfo.AssignedRegs.Regs.empty())
6568 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_Clobber,
6569 false, 0, getCurSDLoc(), DAG,
6576 // Finish up input operands. Set the input chain and add the flag last.
6577 AsmNodeOperands[InlineAsm::Op_InputChain] = Chain;
6578 if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
6580 Chain = DAG.getNode(ISD::INLINEASM, getCurSDLoc(),
6581 DAG.getVTList(MVT::Other, MVT::Glue), AsmNodeOperands);
6582 Flag = Chain.getValue(1);
6584 // If this asm returns a register value, copy the result from that register
6585 // and set it as the value of the call.
6586 if (!RetValRegs.Regs.empty()) {
6587 SDValue Val = RetValRegs.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(),
6588 Chain, &Flag, CS.getInstruction());
6590 // FIXME: Why don't we do this for inline asms with MRVs?
6591 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) {
6592 EVT ResultType = TLI.getValueType(DAG.getDataLayout(), CS.getType());
6594 // If any of the results of the inline asm is a vector, it may have the
6595 // wrong width/num elts. This can happen for register classes that can
6596 // contain multiple different value types. The preg or vreg allocated may
6597 // not have the same VT as was expected. Convert it to the right type
6598 // with bit_convert.
6599 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) {
6600 Val = DAG.getNode(ISD::BITCAST, getCurSDLoc(),
6603 } else if (ResultType != Val.getValueType() &&
6604 ResultType.isInteger() && Val.getValueType().isInteger()) {
6605 // If a result value was tied to an input value, the computed result may
6606 // have a wider width than the expected result. Extract the relevant
6608 Val = DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), ResultType, Val);
6611 assert(ResultType == Val.getValueType() && "Asm result value mismatch!");
6614 setValue(CS.getInstruction(), Val);
6615 // Don't need to use this as a chain in this case.
6616 if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty())
6620 std::vector<std::pair<SDValue, const Value *> > StoresToEmit;
6622 // Process indirect outputs, first output all of the flagged copies out of
6624 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
6625 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
6626 const Value *Ptr = IndirectStoresToEmit[i].second;
6627 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(),
6629 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
6632 // Emit the non-flagged stores from the physregs.
6633 SmallVector<SDValue, 8> OutChains;
6634 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) {
6635 SDValue Val = DAG.getStore(Chain, getCurSDLoc(),
6636 StoresToEmit[i].first,
6637 getValue(StoresToEmit[i].second),
6638 MachinePointerInfo(StoresToEmit[i].second),
6640 OutChains.push_back(Val);
6643 if (!OutChains.empty())
6644 Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other, OutChains);
6649 void SelectionDAGBuilder::visitVAStart(const CallInst &I) {
6650 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurSDLoc(),
6651 MVT::Other, getRoot(),
6652 getValue(I.getArgOperand(0)),
6653 DAG.getSrcValue(I.getArgOperand(0))));
6656 void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) {
6657 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6658 const DataLayout &DL = DAG.getDataLayout();
6659 SDValue V = DAG.getVAArg(TLI.getValueType(DAG.getDataLayout(), I.getType()),
6660 getCurSDLoc(), getRoot(), getValue(I.getOperand(0)),
6661 DAG.getSrcValue(I.getOperand(0)),
6662 DL.getABITypeAlignment(I.getType()));
6664 DAG.setRoot(V.getValue(1));
6667 void SelectionDAGBuilder::visitVAEnd(const CallInst &I) {
6668 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurSDLoc(),
6669 MVT::Other, getRoot(),
6670 getValue(I.getArgOperand(0)),
6671 DAG.getSrcValue(I.getArgOperand(0))));
6674 void SelectionDAGBuilder::visitVACopy(const CallInst &I) {
6675 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurSDLoc(),
6676 MVT::Other, getRoot(),
6677 getValue(I.getArgOperand(0)),
6678 getValue(I.getArgOperand(1)),
6679 DAG.getSrcValue(I.getArgOperand(0)),
6680 DAG.getSrcValue(I.getArgOperand(1))));
6683 /// \brief Lower an argument list according to the target calling convention.
6685 /// \return A tuple of <return-value, token-chain>
6687 /// This is a helper for lowering intrinsics that follow a target calling
6688 /// convention or require stack pointer adjustment. Only a subset of the
6689 /// intrinsic's operands need to participate in the calling convention.
6690 std::pair<SDValue, SDValue> SelectionDAGBuilder::lowerCallOperands(
6691 ImmutableCallSite CS, unsigned ArgIdx, unsigned NumArgs, SDValue Callee,
6692 Type *ReturnTy, const BasicBlock *EHPadBB, bool IsPatchPoint) {
6693 TargetLowering::ArgListTy Args;
6694 Args.reserve(NumArgs);
6696 // Populate the argument list.
6697 // Attributes for args start at offset 1, after the return attribute.
6698 for (unsigned ArgI = ArgIdx, ArgE = ArgIdx + NumArgs, AttrI = ArgIdx + 1;
6699 ArgI != ArgE; ++ArgI) {
6700 const Value *V = CS->getOperand(ArgI);
6702 assert(!V->getType()->isEmptyTy() && "Empty type passed to intrinsic.");
6704 TargetLowering::ArgListEntry Entry;
6705 Entry.Node = getValue(V);
6706 Entry.Ty = V->getType();
6707 Entry.setAttributes(&CS, AttrI);
6708 Args.push_back(Entry);
6711 TargetLowering::CallLoweringInfo CLI(DAG);
6712 CLI.setDebugLoc(getCurSDLoc()).setChain(getRoot())
6713 .setCallee(CS.getCallingConv(), ReturnTy, Callee, std::move(Args), NumArgs)
6714 .setDiscardResult(CS->use_empty()).setIsPatchPoint(IsPatchPoint);
6716 return lowerInvokable(CLI, EHPadBB);
6719 /// \brief Add a stack map intrinsic call's live variable operands to a stackmap
6720 /// or patchpoint target node's operand list.
6722 /// Constants are converted to TargetConstants purely as an optimization to
6723 /// avoid constant materialization and register allocation.
6725 /// FrameIndex operands are converted to TargetFrameIndex so that ISEL does not
6726 /// generate addess computation nodes, and so ExpandISelPseudo can convert the
6727 /// TargetFrameIndex into a DirectMemRefOp StackMap location. This avoids
6728 /// address materialization and register allocation, but may also be required
6729 /// for correctness. If a StackMap (or PatchPoint) intrinsic directly uses an
6730 /// alloca in the entry block, then the runtime may assume that the alloca's
6731 /// StackMap location can be read immediately after compilation and that the
6732 /// location is valid at any point during execution (this is similar to the
6733 /// assumption made by the llvm.gcroot intrinsic). If the alloca's location were
6734 /// only available in a register, then the runtime would need to trap when
6735 /// execution reaches the StackMap in order to read the alloca's location.
6736 static void addStackMapLiveVars(ImmutableCallSite CS, unsigned StartIdx,
6737 SDLoc DL, SmallVectorImpl<SDValue> &Ops,
6738 SelectionDAGBuilder &Builder) {
6739 for (unsigned i = StartIdx, e = CS.arg_size(); i != e; ++i) {
6740 SDValue OpVal = Builder.getValue(CS.getArgument(i));
6741 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(OpVal)) {
6743 Builder.DAG.getTargetConstant(StackMaps::ConstantOp, DL, MVT::i64));
6745 Builder.DAG.getTargetConstant(C->getSExtValue(), DL, MVT::i64));
6746 } else if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(OpVal)) {
6747 const TargetLowering &TLI = Builder.DAG.getTargetLoweringInfo();
6748 Ops.push_back(Builder.DAG.getTargetFrameIndex(
6749 FI->getIndex(), TLI.getPointerTy(Builder.DAG.getDataLayout())));
6751 Ops.push_back(OpVal);
6755 /// \brief Lower llvm.experimental.stackmap directly to its target opcode.
6756 void SelectionDAGBuilder::visitStackmap(const CallInst &CI) {
6757 // void @llvm.experimental.stackmap(i32 <id>, i32 <numShadowBytes>,
6758 // [live variables...])
6760 assert(CI.getType()->isVoidTy() && "Stackmap cannot return a value.");
6762 SDValue Chain, InFlag, Callee, NullPtr;
6763 SmallVector<SDValue, 32> Ops;
6765 SDLoc DL = getCurSDLoc();
6766 Callee = getValue(CI.getCalledValue());
6767 NullPtr = DAG.getIntPtrConstant(0, DL, true);
6769 // The stackmap intrinsic only records the live variables (the arguemnts
6770 // passed to it) and emits NOPS (if requested). Unlike the patchpoint
6771 // intrinsic, this won't be lowered to a function call. This means we don't
6772 // have to worry about calling conventions and target specific lowering code.
6773 // Instead we perform the call lowering right here.
6775 // chain, flag = CALLSEQ_START(chain, 0)
6776 // chain, flag = STACKMAP(id, nbytes, ..., chain, flag)
6777 // chain, flag = CALLSEQ_END(chain, 0, 0, flag)
6779 Chain = DAG.getCALLSEQ_START(getRoot(), NullPtr, DL);
6780 InFlag = Chain.getValue(1);
6782 // Add the <id> and <numBytes> constants.
6783 SDValue IDVal = getValue(CI.getOperand(PatchPointOpers::IDPos));
6784 Ops.push_back(DAG.getTargetConstant(
6785 cast<ConstantSDNode>(IDVal)->getZExtValue(), DL, MVT::i64));
6786 SDValue NBytesVal = getValue(CI.getOperand(PatchPointOpers::NBytesPos));
6787 Ops.push_back(DAG.getTargetConstant(
6788 cast<ConstantSDNode>(NBytesVal)->getZExtValue(), DL,
6791 // Push live variables for the stack map.
6792 addStackMapLiveVars(&CI, 2, DL, Ops, *this);
6794 // We are not pushing any register mask info here on the operands list,
6795 // because the stackmap doesn't clobber anything.
6797 // Push the chain and the glue flag.
6798 Ops.push_back(Chain);
6799 Ops.push_back(InFlag);
6801 // Create the STACKMAP node.
6802 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6803 SDNode *SM = DAG.getMachineNode(TargetOpcode::STACKMAP, DL, NodeTys, Ops);
6804 Chain = SDValue(SM, 0);
6805 InFlag = Chain.getValue(1);
6807 Chain = DAG.getCALLSEQ_END(Chain, NullPtr, NullPtr, InFlag, DL);
6809 // Stackmaps don't generate values, so nothing goes into the NodeMap.
6811 // Set the root to the target-lowered call chain.
6814 // Inform the Frame Information that we have a stackmap in this function.
6815 FuncInfo.MF->getFrameInfo()->setHasStackMap();
6818 /// \brief Lower llvm.experimental.patchpoint directly to its target opcode.
6819 void SelectionDAGBuilder::visitPatchpoint(ImmutableCallSite CS,
6820 const BasicBlock *EHPadBB) {
6821 // void|i64 @llvm.experimental.patchpoint.void|i64(i64 <id>,
6826 // [live variables...])
6828 CallingConv::ID CC = CS.getCallingConv();
6829 bool IsAnyRegCC = CC == CallingConv::AnyReg;
6830 bool HasDef = !CS->getType()->isVoidTy();
6831 SDLoc dl = getCurSDLoc();
6832 SDValue Callee = getValue(CS->getOperand(PatchPointOpers::TargetPos));
6834 // Handle immediate and symbolic callees.
6835 if (auto* ConstCallee = dyn_cast<ConstantSDNode>(Callee))
6836 Callee = DAG.getIntPtrConstant(ConstCallee->getZExtValue(), dl,
6838 else if (auto* SymbolicCallee = dyn_cast<GlobalAddressSDNode>(Callee))
6839 Callee = DAG.getTargetGlobalAddress(SymbolicCallee->getGlobal(),
6840 SDLoc(SymbolicCallee),
6841 SymbolicCallee->getValueType(0));
6843 // Get the real number of arguments participating in the call <numArgs>
6844 SDValue NArgVal = getValue(CS.getArgument(PatchPointOpers::NArgPos));
6845 unsigned NumArgs = cast<ConstantSDNode>(NArgVal)->getZExtValue();
6847 // Skip the four meta args: <id>, <numNopBytes>, <target>, <numArgs>
6848 // Intrinsics include all meta-operands up to but not including CC.
6849 unsigned NumMetaOpers = PatchPointOpers::CCPos;
6850 assert(CS.arg_size() >= NumMetaOpers + NumArgs &&
6851 "Not enough arguments provided to the patchpoint intrinsic");
6853 // For AnyRegCC the arguments are lowered later on manually.
6854 unsigned NumCallArgs = IsAnyRegCC ? 0 : NumArgs;
6856 IsAnyRegCC ? Type::getVoidTy(*DAG.getContext()) : CS->getType();
6857 std::pair<SDValue, SDValue> Result = lowerCallOperands(
6858 CS, NumMetaOpers, NumCallArgs, Callee, ReturnTy, EHPadBB, true);
6860 SDNode *CallEnd = Result.second.getNode();
6861 if (HasDef && (CallEnd->getOpcode() == ISD::CopyFromReg))
6862 CallEnd = CallEnd->getOperand(0).getNode();
6864 /// Get a call instruction from the call sequence chain.
6865 /// Tail calls are not allowed.
6866 assert(CallEnd->getOpcode() == ISD::CALLSEQ_END &&
6867 "Expected a callseq node.");
6868 SDNode *Call = CallEnd->getOperand(0).getNode();
6869 bool HasGlue = Call->getGluedNode();
6871 // Replace the target specific call node with the patchable intrinsic.
6872 SmallVector<SDValue, 8> Ops;
6874 // Add the <id> and <numBytes> constants.
6875 SDValue IDVal = getValue(CS->getOperand(PatchPointOpers::IDPos));
6876 Ops.push_back(DAG.getTargetConstant(
6877 cast<ConstantSDNode>(IDVal)->getZExtValue(), dl, MVT::i64));
6878 SDValue NBytesVal = getValue(CS->getOperand(PatchPointOpers::NBytesPos));
6879 Ops.push_back(DAG.getTargetConstant(
6880 cast<ConstantSDNode>(NBytesVal)->getZExtValue(), dl,
6884 Ops.push_back(Callee);
6886 // Adjust <numArgs> to account for any arguments that have been passed on the
6888 // Call Node: Chain, Target, {Args}, RegMask, [Glue]
6889 unsigned NumCallRegArgs = Call->getNumOperands() - (HasGlue ? 4 : 3);
6890 NumCallRegArgs = IsAnyRegCC ? NumArgs : NumCallRegArgs;
6891 Ops.push_back(DAG.getTargetConstant(NumCallRegArgs, dl, MVT::i32));
6893 // Add the calling convention
6894 Ops.push_back(DAG.getTargetConstant((unsigned)CC, dl, MVT::i32));
6896 // Add the arguments we omitted previously. The register allocator should
6897 // place these in any free register.
6899 for (unsigned i = NumMetaOpers, e = NumMetaOpers + NumArgs; i != e; ++i)
6900 Ops.push_back(getValue(CS.getArgument(i)));
6902 // Push the arguments from the call instruction up to the register mask.
6903 SDNode::op_iterator e = HasGlue ? Call->op_end()-2 : Call->op_end()-1;
6904 Ops.append(Call->op_begin() + 2, e);
6906 // Push live variables for the stack map.
6907 addStackMapLiveVars(CS, NumMetaOpers + NumArgs, dl, Ops, *this);
6909 // Push the register mask info.
6911 Ops.push_back(*(Call->op_end()-2));
6913 Ops.push_back(*(Call->op_end()-1));
6915 // Push the chain (this is originally the first operand of the call, but
6916 // becomes now the last or second to last operand).
6917 Ops.push_back(*(Call->op_begin()));
6919 // Push the glue flag (last operand).
6921 Ops.push_back(*(Call->op_end()-1));
6924 if (IsAnyRegCC && HasDef) {
6925 // Create the return types based on the intrinsic definition
6926 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6927 SmallVector<EVT, 3> ValueVTs;
6928 ComputeValueVTs(TLI, DAG.getDataLayout(), CS->getType(), ValueVTs);
6929 assert(ValueVTs.size() == 1 && "Expected only one return value type.");
6931 // There is always a chain and a glue type at the end
6932 ValueVTs.push_back(MVT::Other);
6933 ValueVTs.push_back(MVT::Glue);
6934 NodeTys = DAG.getVTList(ValueVTs);
6936 NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6938 // Replace the target specific call node with a PATCHPOINT node.
6939 MachineSDNode *MN = DAG.getMachineNode(TargetOpcode::PATCHPOINT,
6942 // Update the NodeMap.
6945 setValue(CS.getInstruction(), SDValue(MN, 0));
6947 setValue(CS.getInstruction(), Result.first);
6950 // Fixup the consumers of the intrinsic. The chain and glue may be used in the
6951 // call sequence. Furthermore the location of the chain and glue can change
6952 // when the AnyReg calling convention is used and the intrinsic returns a
6954 if (IsAnyRegCC && HasDef) {
6955 SDValue From[] = {SDValue(Call, 0), SDValue(Call, 1)};
6956 SDValue To[] = {SDValue(MN, 1), SDValue(MN, 2)};
6957 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
6959 DAG.ReplaceAllUsesWith(Call, MN);
6960 DAG.DeleteNode(Call);
6962 // Inform the Frame Information that we have a patchpoint in this function.
6963 FuncInfo.MF->getFrameInfo()->setHasPatchPoint();
6966 /// Returns an AttributeSet representing the attributes applied to the return
6967 /// value of the given call.
6968 static AttributeSet getReturnAttrs(TargetLowering::CallLoweringInfo &CLI) {
6969 SmallVector<Attribute::AttrKind, 2> Attrs;
6971 Attrs.push_back(Attribute::SExt);
6973 Attrs.push_back(Attribute::ZExt);
6975 Attrs.push_back(Attribute::InReg);
6977 return AttributeSet::get(CLI.RetTy->getContext(), AttributeSet::ReturnIndex,
6981 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
6982 /// implementation, which just calls LowerCall.
6983 /// FIXME: When all targets are
6984 /// migrated to using LowerCall, this hook should be integrated into SDISel.
6985 std::pair<SDValue, SDValue>
6986 TargetLowering::LowerCallTo(TargetLowering::CallLoweringInfo &CLI) const {
6987 // Handle the incoming return values from the call.
6989 Type *OrigRetTy = CLI.RetTy;
6990 SmallVector<EVT, 4> RetTys;
6991 SmallVector<uint64_t, 4> Offsets;
6992 auto &DL = CLI.DAG.getDataLayout();
6993 ComputeValueVTs(*this, DL, CLI.RetTy, RetTys, &Offsets);
6995 SmallVector<ISD::OutputArg, 4> Outs;
6996 GetReturnInfo(CLI.RetTy, getReturnAttrs(CLI), Outs, *this, DL);
6998 bool CanLowerReturn =
6999 this->CanLowerReturn(CLI.CallConv, CLI.DAG.getMachineFunction(),
7000 CLI.IsVarArg, Outs, CLI.RetTy->getContext());
7002 SDValue DemoteStackSlot;
7003 int DemoteStackIdx = -100;
7004 if (!CanLowerReturn) {
7005 // FIXME: equivalent assert?
7006 // assert(!CS.hasInAllocaArgument() &&
7007 // "sret demotion is incompatible with inalloca");
7008 uint64_t TySize = DL.getTypeAllocSize(CLI.RetTy);
7009 unsigned Align = DL.getPrefTypeAlignment(CLI.RetTy);
7010 MachineFunction &MF = CLI.DAG.getMachineFunction();
7011 DemoteStackIdx = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
7012 Type *StackSlotPtrType = PointerType::getUnqual(CLI.RetTy);
7014 DemoteStackSlot = CLI.DAG.getFrameIndex(DemoteStackIdx, getPointerTy(DL));
7016 Entry.Node = DemoteStackSlot;
7017 Entry.Ty = StackSlotPtrType;
7018 Entry.isSExt = false;
7019 Entry.isZExt = false;
7020 Entry.isInReg = false;
7021 Entry.isSRet = true;
7022 Entry.isNest = false;
7023 Entry.isByVal = false;
7024 Entry.isReturned = false;
7025 Entry.Alignment = Align;
7026 CLI.getArgs().insert(CLI.getArgs().begin(), Entry);
7027 CLI.RetTy = Type::getVoidTy(CLI.RetTy->getContext());
7029 // sret demotion isn't compatible with tail-calls, since the sret argument
7030 // points into the callers stack frame.
7031 CLI.IsTailCall = false;
7033 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
7035 MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
7036 unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
7037 for (unsigned i = 0; i != NumRegs; ++i) {
7038 ISD::InputArg MyFlags;
7039 MyFlags.VT = RegisterVT;
7041 MyFlags.Used = CLI.IsReturnValueUsed;
7043 MyFlags.Flags.setSExt();
7045 MyFlags.Flags.setZExt();
7047 MyFlags.Flags.setInReg();
7048 CLI.Ins.push_back(MyFlags);
7053 // Handle all of the outgoing arguments.
7055 CLI.OutVals.clear();
7056 ArgListTy &Args = CLI.getArgs();
7057 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
7058 SmallVector<EVT, 4> ValueVTs;
7059 ComputeValueVTs(*this, DL, Args[i].Ty, ValueVTs);
7060 Type *FinalType = Args[i].Ty;
7061 if (Args[i].isByVal)
7062 FinalType = cast<PointerType>(Args[i].Ty)->getElementType();
7063 bool NeedsRegBlock = functionArgumentNeedsConsecutiveRegisters(
7064 FinalType, CLI.CallConv, CLI.IsVarArg);
7065 for (unsigned Value = 0, NumValues = ValueVTs.size(); Value != NumValues;
7067 EVT VT = ValueVTs[Value];
7068 Type *ArgTy = VT.getTypeForEVT(CLI.RetTy->getContext());
7069 SDValue Op = SDValue(Args[i].Node.getNode(),
7070 Args[i].Node.getResNo() + Value);
7071 ISD::ArgFlagsTy Flags;
7072 unsigned OriginalAlignment = DL.getABITypeAlignment(ArgTy);
7078 if (Args[i].isInReg)
7082 if (Args[i].isByVal)
7084 if (Args[i].isInAlloca) {
7085 Flags.setInAlloca();
7086 // Set the byval flag for CCAssignFn callbacks that don't know about
7087 // inalloca. This way we can know how many bytes we should've allocated
7088 // and how many bytes a callee cleanup function will pop. If we port
7089 // inalloca to more targets, we'll have to add custom inalloca handling
7090 // in the various CC lowering callbacks.
7093 if (Args[i].isByVal || Args[i].isInAlloca) {
7094 PointerType *Ty = cast<PointerType>(Args[i].Ty);
7095 Type *ElementTy = Ty->getElementType();
7096 Flags.setByValSize(DL.getTypeAllocSize(ElementTy));
7097 // For ByVal, alignment should come from FE. BE will guess if this
7098 // info is not there but there are cases it cannot get right.
7099 unsigned FrameAlign;
7100 if (Args[i].Alignment)
7101 FrameAlign = Args[i].Alignment;
7103 FrameAlign = getByValTypeAlignment(ElementTy, DL);
7104 Flags.setByValAlign(FrameAlign);
7109 Flags.setInConsecutiveRegs();
7110 Flags.setOrigAlign(OriginalAlignment);
7112 MVT PartVT = getRegisterType(CLI.RetTy->getContext(), VT);
7113 unsigned NumParts = getNumRegisters(CLI.RetTy->getContext(), VT);
7114 SmallVector<SDValue, 4> Parts(NumParts);
7115 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
7118 ExtendKind = ISD::SIGN_EXTEND;
7119 else if (Args[i].isZExt)
7120 ExtendKind = ISD::ZERO_EXTEND;
7122 // Conservatively only handle 'returned' on non-vectors for now
7123 if (Args[i].isReturned && !Op.getValueType().isVector()) {
7124 assert(CLI.RetTy == Args[i].Ty && RetTys.size() == NumValues &&
7125 "unexpected use of 'returned'");
7126 // Before passing 'returned' to the target lowering code, ensure that
7127 // either the register MVT and the actual EVT are the same size or that
7128 // the return value and argument are extended in the same way; in these
7129 // cases it's safe to pass the argument register value unchanged as the
7130 // return register value (although it's at the target's option whether
7132 // TODO: allow code generation to take advantage of partially preserved
7133 // registers rather than clobbering the entire register when the
7134 // parameter extension method is not compatible with the return
7136 if ((NumParts * PartVT.getSizeInBits() == VT.getSizeInBits()) ||
7137 (ExtendKind != ISD::ANY_EXTEND &&
7138 CLI.RetSExt == Args[i].isSExt && CLI.RetZExt == Args[i].isZExt))
7139 Flags.setReturned();
7142 getCopyToParts(CLI.DAG, CLI.DL, Op, &Parts[0], NumParts, PartVT,
7143 CLI.CS ? CLI.CS->getInstruction() : nullptr, ExtendKind);
7145 for (unsigned j = 0; j != NumParts; ++j) {
7146 // if it isn't first piece, alignment must be 1
7147 ISD::OutputArg MyFlags(Flags, Parts[j].getValueType(), VT,
7148 i < CLI.NumFixedArgs,
7149 i, j*Parts[j].getValueType().getStoreSize());
7150 if (NumParts > 1 && j == 0)
7151 MyFlags.Flags.setSplit();
7153 MyFlags.Flags.setOrigAlign(1);
7155 CLI.Outs.push_back(MyFlags);
7156 CLI.OutVals.push_back(Parts[j]);
7159 if (NeedsRegBlock && Value == NumValues - 1)
7160 CLI.Outs[CLI.Outs.size() - 1].Flags.setInConsecutiveRegsLast();
7164 SmallVector<SDValue, 4> InVals;
7165 CLI.Chain = LowerCall(CLI, InVals);
7167 // Verify that the target's LowerCall behaved as expected.
7168 assert(CLI.Chain.getNode() && CLI.Chain.getValueType() == MVT::Other &&
7169 "LowerCall didn't return a valid chain!");
7170 assert((!CLI.IsTailCall || InVals.empty()) &&
7171 "LowerCall emitted a return value for a tail call!");
7172 assert((CLI.IsTailCall || InVals.size() == CLI.Ins.size()) &&
7173 "LowerCall didn't emit the correct number of values!");
7175 // For a tail call, the return value is merely live-out and there aren't
7176 // any nodes in the DAG representing it. Return a special value to
7177 // indicate that a tail call has been emitted and no more Instructions
7178 // should be processed in the current block.
7179 if (CLI.IsTailCall) {
7180 CLI.DAG.setRoot(CLI.Chain);
7181 return std::make_pair(SDValue(), SDValue());
7184 DEBUG(for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++i) {
7185 assert(InVals[i].getNode() &&
7186 "LowerCall emitted a null value!");
7187 assert(EVT(CLI.Ins[i].VT) == InVals[i].getValueType() &&
7188 "LowerCall emitted a value with the wrong type!");
7191 SmallVector<SDValue, 4> ReturnValues;
7192 if (!CanLowerReturn) {
7193 // The instruction result is the result of loading from the
7194 // hidden sret parameter.
7195 SmallVector<EVT, 1> PVTs;
7196 Type *PtrRetTy = PointerType::getUnqual(OrigRetTy);
7198 ComputeValueVTs(*this, DL, PtrRetTy, PVTs);
7199 assert(PVTs.size() == 1 && "Pointers should fit in one register");
7200 EVT PtrVT = PVTs[0];
7202 unsigned NumValues = RetTys.size();
7203 ReturnValues.resize(NumValues);
7204 SmallVector<SDValue, 4> Chains(NumValues);
7206 for (unsigned i = 0; i < NumValues; ++i) {
7207 SDValue Add = CLI.DAG.getNode(ISD::ADD, CLI.DL, PtrVT, DemoteStackSlot,
7208 CLI.DAG.getConstant(Offsets[i], CLI.DL,
7210 SDValue L = CLI.DAG.getLoad(
7211 RetTys[i], CLI.DL, CLI.Chain, Add,
7212 MachinePointerInfo::getFixedStack(CLI.DAG.getMachineFunction(),
7213 DemoteStackIdx, Offsets[i]),
7214 false, false, false, 1);
7215 ReturnValues[i] = L;
7216 Chains[i] = L.getValue(1);
7219 CLI.Chain = CLI.DAG.getNode(ISD::TokenFactor, CLI.DL, MVT::Other, Chains);
7221 // Collect the legal value parts into potentially illegal values
7222 // that correspond to the original function's return values.
7223 ISD::NodeType AssertOp = ISD::DELETED_NODE;
7225 AssertOp = ISD::AssertSext;
7226 else if (CLI.RetZExt)
7227 AssertOp = ISD::AssertZext;
7228 unsigned CurReg = 0;
7229 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
7231 MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
7232 unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
7234 ReturnValues.push_back(getCopyFromParts(CLI.DAG, CLI.DL, &InVals[CurReg],
7235 NumRegs, RegisterVT, VT, nullptr,
7240 // For a function returning void, there is no return value. We can't create
7241 // such a node, so we just return a null return value in that case. In
7242 // that case, nothing will actually look at the value.
7243 if (ReturnValues.empty())
7244 return std::make_pair(SDValue(), CLI.Chain);
7247 SDValue Res = CLI.DAG.getNode(ISD::MERGE_VALUES, CLI.DL,
7248 CLI.DAG.getVTList(RetTys), ReturnValues);
7249 return std::make_pair(Res, CLI.Chain);
7252 void TargetLowering::LowerOperationWrapper(SDNode *N,
7253 SmallVectorImpl<SDValue> &Results,
7254 SelectionDAG &DAG) const {
7255 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
7257 Results.push_back(Res);
7260 SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
7261 llvm_unreachable("LowerOperation not implemented for this target!");
7265 SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) {
7266 SDValue Op = getNonRegisterValue(V);
7267 assert((Op.getOpcode() != ISD::CopyFromReg ||
7268 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
7269 "Copy from a reg to the same reg!");
7270 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg");
7272 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7273 RegsForValue RFV(V->getContext(), TLI, DAG.getDataLayout(), Reg,
7275 SDValue Chain = DAG.getEntryNode();
7277 ISD::NodeType ExtendType = (FuncInfo.PreferredExtendType.find(V) ==
7278 FuncInfo.PreferredExtendType.end())
7280 : FuncInfo.PreferredExtendType[V];
7281 RFV.getCopyToRegs(Op, DAG, getCurSDLoc(), Chain, nullptr, V, ExtendType);
7282 PendingExports.push_back(Chain);
7285 #include "llvm/CodeGen/SelectionDAGISel.h"
7287 /// isOnlyUsedInEntryBlock - If the specified argument is only used in the
7288 /// entry block, return true. This includes arguments used by switches, since
7289 /// the switch may expand into multiple basic blocks.
7290 static bool isOnlyUsedInEntryBlock(const Argument *A, bool FastISel) {
7291 // With FastISel active, we may be splitting blocks, so force creation
7292 // of virtual registers for all non-dead arguments.
7294 return A->use_empty();
7296 const BasicBlock &Entry = A->getParent()->front();
7297 for (const User *U : A->users())
7298 if (cast<Instruction>(U)->getParent() != &Entry || isa<SwitchInst>(U))
7299 return false; // Use not in entry block.
7304 void SelectionDAGISel::LowerArguments(const Function &F) {
7305 SelectionDAG &DAG = SDB->DAG;
7306 SDLoc dl = SDB->getCurSDLoc();
7307 const DataLayout &DL = DAG.getDataLayout();
7308 SmallVector<ISD::InputArg, 16> Ins;
7310 if (!FuncInfo->CanLowerReturn) {
7311 // Put in an sret pointer parameter before all the other parameters.
7312 SmallVector<EVT, 1> ValueVTs;
7313 ComputeValueVTs(*TLI, DAG.getDataLayout(),
7314 PointerType::getUnqual(F.getReturnType()), ValueVTs);
7316 // NOTE: Assuming that a pointer will never break down to more than one VT
7318 ISD::ArgFlagsTy Flags;
7320 MVT RegisterVT = TLI->getRegisterType(*DAG.getContext(), ValueVTs[0]);
7321 ISD::InputArg RetArg(Flags, RegisterVT, ValueVTs[0], true,
7322 ISD::InputArg::NoArgIndex, 0);
7323 Ins.push_back(RetArg);
7326 // Set up the incoming argument description vector.
7328 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end();
7329 I != E; ++I, ++Idx) {
7330 SmallVector<EVT, 4> ValueVTs;
7331 ComputeValueVTs(*TLI, DAG.getDataLayout(), I->getType(), ValueVTs);
7332 bool isArgValueUsed = !I->use_empty();
7333 unsigned PartBase = 0;
7334 Type *FinalType = I->getType();
7335 if (F.getAttributes().hasAttribute(Idx, Attribute::ByVal))
7336 FinalType = cast<PointerType>(FinalType)->getElementType();
7337 bool NeedsRegBlock = TLI->functionArgumentNeedsConsecutiveRegisters(
7338 FinalType, F.getCallingConv(), F.isVarArg());
7339 for (unsigned Value = 0, NumValues = ValueVTs.size();
7340 Value != NumValues; ++Value) {
7341 EVT VT = ValueVTs[Value];
7342 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
7343 ISD::ArgFlagsTy Flags;
7344 unsigned OriginalAlignment = DL.getABITypeAlignment(ArgTy);
7346 if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt))
7348 if (F.getAttributes().hasAttribute(Idx, Attribute::SExt))
7350 if (F.getAttributes().hasAttribute(Idx, Attribute::InReg))
7352 if (F.getAttributes().hasAttribute(Idx, Attribute::StructRet))
7354 if (F.getAttributes().hasAttribute(Idx, Attribute::ByVal))
7356 if (F.getAttributes().hasAttribute(Idx, Attribute::InAlloca)) {
7357 Flags.setInAlloca();
7358 // Set the byval flag for CCAssignFn callbacks that don't know about
7359 // inalloca. This way we can know how many bytes we should've allocated
7360 // and how many bytes a callee cleanup function will pop. If we port
7361 // inalloca to more targets, we'll have to add custom inalloca handling
7362 // in the various CC lowering callbacks.
7365 if (Flags.isByVal() || Flags.isInAlloca()) {
7366 PointerType *Ty = cast<PointerType>(I->getType());
7367 Type *ElementTy = Ty->getElementType();
7368 Flags.setByValSize(DL.getTypeAllocSize(ElementTy));
7369 // For ByVal, alignment should be passed from FE. BE will guess if
7370 // this info is not there but there are cases it cannot get right.
7371 unsigned FrameAlign;
7372 if (F.getParamAlignment(Idx))
7373 FrameAlign = F.getParamAlignment(Idx);
7375 FrameAlign = TLI->getByValTypeAlignment(ElementTy, DL);
7376 Flags.setByValAlign(FrameAlign);
7378 if (F.getAttributes().hasAttribute(Idx, Attribute::Nest))
7381 Flags.setInConsecutiveRegs();
7382 Flags.setOrigAlign(OriginalAlignment);
7384 MVT RegisterVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7385 unsigned NumRegs = TLI->getNumRegisters(*CurDAG->getContext(), VT);
7386 for (unsigned i = 0; i != NumRegs; ++i) {
7387 ISD::InputArg MyFlags(Flags, RegisterVT, VT, isArgValueUsed,
7388 Idx-1, PartBase+i*RegisterVT.getStoreSize());
7389 if (NumRegs > 1 && i == 0)
7390 MyFlags.Flags.setSplit();
7391 // if it isn't first piece, alignment must be 1
7393 MyFlags.Flags.setOrigAlign(1);
7394 Ins.push_back(MyFlags);
7396 if (NeedsRegBlock && Value == NumValues - 1)
7397 Ins[Ins.size() - 1].Flags.setInConsecutiveRegsLast();
7398 PartBase += VT.getStoreSize();
7402 // Call the target to set up the argument values.
7403 SmallVector<SDValue, 8> InVals;
7404 SDValue NewRoot = TLI->LowerFormalArguments(
7405 DAG.getRoot(), F.getCallingConv(), F.isVarArg(), Ins, dl, DAG, InVals);
7407 // Verify that the target's LowerFormalArguments behaved as expected.
7408 assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other &&
7409 "LowerFormalArguments didn't return a valid chain!");
7410 assert(InVals.size() == Ins.size() &&
7411 "LowerFormalArguments didn't emit the correct number of values!");
7413 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
7414 assert(InVals[i].getNode() &&
7415 "LowerFormalArguments emitted a null value!");
7416 assert(EVT(Ins[i].VT) == InVals[i].getValueType() &&
7417 "LowerFormalArguments emitted a value with the wrong type!");
7421 // Update the DAG with the new chain value resulting from argument lowering.
7422 DAG.setRoot(NewRoot);
7424 // Set up the argument values.
7427 if (!FuncInfo->CanLowerReturn) {
7428 // Create a virtual register for the sret pointer, and put in a copy
7429 // from the sret argument into it.
7430 SmallVector<EVT, 1> ValueVTs;
7431 ComputeValueVTs(*TLI, DAG.getDataLayout(),
7432 PointerType::getUnqual(F.getReturnType()), ValueVTs);
7433 MVT VT = ValueVTs[0].getSimpleVT();
7434 MVT RegVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7435 ISD::NodeType AssertOp = ISD::DELETED_NODE;
7436 SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1,
7437 RegVT, VT, nullptr, AssertOp);
7439 MachineFunction& MF = SDB->DAG.getMachineFunction();
7440 MachineRegisterInfo& RegInfo = MF.getRegInfo();
7441 unsigned SRetReg = RegInfo.createVirtualRegister(TLI->getRegClassFor(RegVT));
7442 FuncInfo->DemoteRegister = SRetReg;
7444 SDB->DAG.getCopyToReg(NewRoot, SDB->getCurSDLoc(), SRetReg, ArgValue);
7445 DAG.setRoot(NewRoot);
7447 // i indexes lowered arguments. Bump it past the hidden sret argument.
7448 // Idx indexes LLVM arguments. Don't touch it.
7452 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
7454 SmallVector<SDValue, 4> ArgValues;
7455 SmallVector<EVT, 4> ValueVTs;
7456 ComputeValueVTs(*TLI, DAG.getDataLayout(), I->getType(), ValueVTs);
7457 unsigned NumValues = ValueVTs.size();
7459 // If this argument is unused then remember its value. It is used to generate
7460 // debugging information.
7461 if (I->use_empty() && NumValues) {
7462 SDB->setUnusedArgValue(&*I, InVals[i]);
7464 // Also remember any frame index for use in FastISel.
7465 if (FrameIndexSDNode *FI =
7466 dyn_cast<FrameIndexSDNode>(InVals[i].getNode()))
7467 FuncInfo->setArgumentFrameIndex(&*I, FI->getIndex());
7470 for (unsigned Val = 0; Val != NumValues; ++Val) {
7471 EVT VT = ValueVTs[Val];
7472 MVT PartVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7473 unsigned NumParts = TLI->getNumRegisters(*CurDAG->getContext(), VT);
7475 if (!I->use_empty()) {
7476 ISD::NodeType AssertOp = ISD::DELETED_NODE;
7477 if (F.getAttributes().hasAttribute(Idx, Attribute::SExt))
7478 AssertOp = ISD::AssertSext;
7479 else if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt))
7480 AssertOp = ISD::AssertZext;
7482 ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i],
7483 NumParts, PartVT, VT,
7484 nullptr, AssertOp));
7490 // We don't need to do anything else for unused arguments.
7491 if (ArgValues.empty())
7494 // Note down frame index.
7495 if (FrameIndexSDNode *FI =
7496 dyn_cast<FrameIndexSDNode>(ArgValues[0].getNode()))
7497 FuncInfo->setArgumentFrameIndex(&*I, FI->getIndex());
7499 SDValue Res = DAG.getMergeValues(makeArrayRef(ArgValues.data(), NumValues),
7500 SDB->getCurSDLoc());
7502 SDB->setValue(&*I, Res);
7503 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::BUILD_PAIR) {
7504 if (LoadSDNode *LNode =
7505 dyn_cast<LoadSDNode>(Res.getOperand(0).getNode()))
7506 if (FrameIndexSDNode *FI =
7507 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode()))
7508 FuncInfo->setArgumentFrameIndex(&*I, FI->getIndex());
7511 // If this argument is live outside of the entry block, insert a copy from
7512 // wherever we got it to the vreg that other BB's will reference it as.
7513 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::CopyFromReg) {
7514 // If we can, though, try to skip creating an unnecessary vreg.
7515 // FIXME: This isn't very clean... it would be nice to make this more
7516 // general. It's also subtly incompatible with the hacks FastISel
7518 unsigned Reg = cast<RegisterSDNode>(Res.getOperand(1))->getReg();
7519 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
7520 FuncInfo->ValueMap[&*I] = Reg;
7524 if (!isOnlyUsedInEntryBlock(&*I, TM.Options.EnableFastISel)) {
7525 FuncInfo->InitializeRegForValue(&*I);
7526 SDB->CopyToExportRegsIfNeeded(&*I);
7530 assert(i == InVals.size() && "Argument register count mismatch!");
7532 // Finally, if the target has anything special to do, allow it to do so.
7533 EmitFunctionEntryCode();
7536 /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
7537 /// ensure constants are generated when needed. Remember the virtual registers
7538 /// that need to be added to the Machine PHI nodes as input. We cannot just
7539 /// directly add them, because expansion might result in multiple MBB's for one
7540 /// BB. As such, the start of the BB might correspond to a different MBB than
7544 SelectionDAGBuilder::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
7545 const TerminatorInst *TI = LLVMBB->getTerminator();
7547 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
7549 // Check PHI nodes in successors that expect a value to be available from this
7551 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
7552 const BasicBlock *SuccBB = TI->getSuccessor(succ);
7553 if (!isa<PHINode>(SuccBB->begin())) continue;
7554 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
7556 // If this terminator has multiple identical successors (common for
7557 // switches), only handle each succ once.
7558 if (!SuccsHandled.insert(SuccMBB).second)
7561 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
7563 // At this point we know that there is a 1-1 correspondence between LLVM PHI
7564 // nodes and Machine PHI nodes, but the incoming operands have not been
7566 for (BasicBlock::const_iterator I = SuccBB->begin();
7567 const PHINode *PN = dyn_cast<PHINode>(I); ++I) {
7568 // Ignore dead phi's.
7569 if (PN->use_empty()) continue;
7572 if (PN->getType()->isEmptyTy())
7576 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
7578 if (const Constant *C = dyn_cast<Constant>(PHIOp)) {
7579 unsigned &RegOut = ConstantsOut[C];
7581 RegOut = FuncInfo.CreateRegs(C->getType());
7582 CopyValueToVirtualRegister(C, RegOut);
7586 DenseMap<const Value *, unsigned>::iterator I =
7587 FuncInfo.ValueMap.find(PHIOp);
7588 if (I != FuncInfo.ValueMap.end())
7591 assert(isa<AllocaInst>(PHIOp) &&
7592 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
7593 "Didn't codegen value into a register!??");
7594 Reg = FuncInfo.CreateRegs(PHIOp->getType());
7595 CopyValueToVirtualRegister(PHIOp, Reg);
7599 // Remember that this register needs to added to the machine PHI node as
7600 // the input for this MBB.
7601 SmallVector<EVT, 4> ValueVTs;
7602 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7603 ComputeValueVTs(TLI, DAG.getDataLayout(), PN->getType(), ValueVTs);
7604 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
7605 EVT VT = ValueVTs[vti];
7606 unsigned NumRegisters = TLI.getNumRegisters(*DAG.getContext(), VT);
7607 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
7608 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
7609 Reg += NumRegisters;
7614 ConstantsOut.clear();
7617 /// Add a successor MBB to ParentMBB< creating a new MachineBB for BB if SuccMBB
7620 SelectionDAGBuilder::StackProtectorDescriptor::
7621 AddSuccessorMBB(const BasicBlock *BB,
7622 MachineBasicBlock *ParentMBB,
7624 MachineBasicBlock *SuccMBB) {
7625 // If SuccBB has not been created yet, create it.
7627 MachineFunction *MF = ParentMBB->getParent();
7628 MachineFunction::iterator BBI(ParentMBB);
7629 SuccMBB = MF->CreateMachineBasicBlock(BB);
7630 MF->insert(++BBI, SuccMBB);
7632 // Add it as a successor of ParentMBB.
7633 ParentMBB->addSuccessor(
7634 SuccMBB, BranchProbabilityInfo::getBranchWeightStackProtector(IsLikely));
7638 MachineBasicBlock *SelectionDAGBuilder::NextBlock(MachineBasicBlock *MBB) {
7639 MachineFunction::iterator I(MBB);
7640 if (++I == FuncInfo.MF->end())
7645 /// During lowering new call nodes can be created (such as memset, etc.).
7646 /// Those will become new roots of the current DAG, but complications arise
7647 /// when they are tail calls. In such cases, the call lowering will update
7648 /// the root, but the builder still needs to know that a tail call has been
7649 /// lowered in order to avoid generating an additional return.
7650 void SelectionDAGBuilder::updateDAGForMaybeTailCall(SDValue MaybeTC) {
7651 // If the node is null, we do have a tail call.
7652 if (MaybeTC.getNode() != nullptr)
7653 DAG.setRoot(MaybeTC);
7658 bool SelectionDAGBuilder::isDense(const CaseClusterVector &Clusters,
7659 unsigned *TotalCases, unsigned First,
7661 assert(Last >= First);
7662 assert(TotalCases[Last] >= TotalCases[First]);
7664 APInt LowCase = Clusters[First].Low->getValue();
7665 APInt HighCase = Clusters[Last].High->getValue();
7666 assert(LowCase.getBitWidth() == HighCase.getBitWidth());
7668 // FIXME: A range of consecutive cases has 100% density, but only requires one
7669 // comparison to lower. We should discriminate against such consecutive ranges
7672 uint64_t Diff = (HighCase - LowCase).getLimitedValue((UINT64_MAX - 1) / 100);
7673 uint64_t Range = Diff + 1;
7676 TotalCases[Last] - (First == 0 ? 0 : TotalCases[First - 1]);
7678 assert(NumCases < UINT64_MAX / 100);
7679 assert(Range >= NumCases);
7681 return NumCases * 100 >= Range * MinJumpTableDensity;
7684 static inline bool areJTsAllowed(const TargetLowering &TLI) {
7685 return TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
7686 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other);
7689 bool SelectionDAGBuilder::buildJumpTable(CaseClusterVector &Clusters,
7690 unsigned First, unsigned Last,
7691 const SwitchInst *SI,
7692 MachineBasicBlock *DefaultMBB,
7693 CaseCluster &JTCluster) {
7694 assert(First <= Last);
7696 uint32_t Weight = 0;
7697 unsigned NumCmps = 0;
7698 std::vector<MachineBasicBlock*> Table;
7699 DenseMap<MachineBasicBlock*, uint32_t> JTWeights;
7700 for (unsigned I = First; I <= Last; ++I) {
7701 assert(Clusters[I].Kind == CC_Range);
7702 Weight += Clusters[I].Weight;
7703 assert(Weight >= Clusters[I].Weight && "Weight overflow!");
7704 APInt Low = Clusters[I].Low->getValue();
7705 APInt High = Clusters[I].High->getValue();
7706 NumCmps += (Low == High) ? 1 : 2;
7708 // Fill the gap between this and the previous cluster.
7709 APInt PreviousHigh = Clusters[I - 1].High->getValue();
7710 assert(PreviousHigh.slt(Low));
7711 uint64_t Gap = (Low - PreviousHigh).getLimitedValue() - 1;
7712 for (uint64_t J = 0; J < Gap; J++)
7713 Table.push_back(DefaultMBB);
7715 uint64_t ClusterSize = (High - Low).getLimitedValue() + 1;
7716 for (uint64_t J = 0; J < ClusterSize; ++J)
7717 Table.push_back(Clusters[I].MBB);
7718 JTWeights[Clusters[I].MBB] += Clusters[I].Weight;
7721 unsigned NumDests = JTWeights.size();
7722 if (isSuitableForBitTests(NumDests, NumCmps,
7723 Clusters[First].Low->getValue(),
7724 Clusters[Last].High->getValue())) {
7725 // Clusters[First..Last] should be lowered as bit tests instead.
7729 // Create the MBB that will load from and jump through the table.
7730 // Note: We create it here, but it's not inserted into the function yet.
7731 MachineFunction *CurMF = FuncInfo.MF;
7732 MachineBasicBlock *JumpTableMBB =
7733 CurMF->CreateMachineBasicBlock(SI->getParent());
7735 // Add successors. Note: use table order for determinism.
7736 SmallPtrSet<MachineBasicBlock *, 8> Done;
7737 for (MachineBasicBlock *Succ : Table) {
7738 if (Done.count(Succ))
7740 addSuccessorWithWeight(JumpTableMBB, Succ, JTWeights[Succ]);
7744 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7745 unsigned JTI = CurMF->getOrCreateJumpTableInfo(TLI.getJumpTableEncoding())
7746 ->createJumpTableIndex(Table);
7748 // Set up the jump table info.
7749 JumpTable JT(-1U, JTI, JumpTableMBB, nullptr);
7750 JumpTableHeader JTH(Clusters[First].Low->getValue(),
7751 Clusters[Last].High->getValue(), SI->getCondition(),
7753 JTCases.emplace_back(std::move(JTH), std::move(JT));
7755 JTCluster = CaseCluster::jumpTable(Clusters[First].Low, Clusters[Last].High,
7756 JTCases.size() - 1, Weight);
7760 void SelectionDAGBuilder::findJumpTables(CaseClusterVector &Clusters,
7761 const SwitchInst *SI,
7762 MachineBasicBlock *DefaultMBB) {
7764 // Clusters must be non-empty, sorted, and only contain Range clusters.
7765 assert(!Clusters.empty());
7766 for (CaseCluster &C : Clusters)
7767 assert(C.Kind == CC_Range);
7768 for (unsigned i = 1, e = Clusters.size(); i < e; ++i)
7769 assert(Clusters[i - 1].High->getValue().slt(Clusters[i].Low->getValue()));
7772 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7773 if (!areJTsAllowed(TLI))
7776 const int64_t N = Clusters.size();
7777 const unsigned MinJumpTableSize = TLI.getMinimumJumpTableEntries();
7779 // TotalCases[i]: Total nbr of cases in Clusters[0..i].
7780 SmallVector<unsigned, 8> TotalCases(N);
7782 for (unsigned i = 0; i < N; ++i) {
7783 APInt Hi = Clusters[i].High->getValue();
7784 APInt Lo = Clusters[i].Low->getValue();
7785 TotalCases[i] = (Hi - Lo).getLimitedValue() + 1;
7787 TotalCases[i] += TotalCases[i - 1];
7790 if (N >= MinJumpTableSize && isDense(Clusters, &TotalCases[0], 0, N - 1)) {
7791 // Cheap case: the whole range might be suitable for jump table.
7792 CaseCluster JTCluster;
7793 if (buildJumpTable(Clusters, 0, N - 1, SI, DefaultMBB, JTCluster)) {
7794 Clusters[0] = JTCluster;
7800 // The algorithm below is not suitable for -O0.
7801 if (TM.getOptLevel() == CodeGenOpt::None)
7804 // Split Clusters into minimum number of dense partitions. The algorithm uses
7805 // the same idea as Kannan & Proebsting "Correction to 'Producing Good Code
7806 // for the Case Statement'" (1994), but builds the MinPartitions array in
7807 // reverse order to make it easier to reconstruct the partitions in ascending
7808 // order. In the choice between two optimal partitionings, it picks the one
7809 // which yields more jump tables.
7811 // MinPartitions[i] is the minimum nbr of partitions of Clusters[i..N-1].
7812 SmallVector<unsigned, 8> MinPartitions(N);
7813 // LastElement[i] is the last element of the partition starting at i.
7814 SmallVector<unsigned, 8> LastElement(N);
7815 // NumTables[i]: nbr of >= MinJumpTableSize partitions from Clusters[i..N-1].
7816 SmallVector<unsigned, 8> NumTables(N);
7818 // Base case: There is only one way to partition Clusters[N-1].
7819 MinPartitions[N - 1] = 1;
7820 LastElement[N - 1] = N - 1;
7821 assert(MinJumpTableSize > 1);
7822 NumTables[N - 1] = 0;
7824 // Note: loop indexes are signed to avoid underflow.
7825 for (int64_t i = N - 2; i >= 0; i--) {
7826 // Find optimal partitioning of Clusters[i..N-1].
7827 // Baseline: Put Clusters[i] into a partition on its own.
7828 MinPartitions[i] = MinPartitions[i + 1] + 1;
7830 NumTables[i] = NumTables[i + 1];
7832 // Search for a solution that results in fewer partitions.
7833 for (int64_t j = N - 1; j > i; j--) {
7834 // Try building a partition from Clusters[i..j].
7835 if (isDense(Clusters, &TotalCases[0], i, j)) {
7836 unsigned NumPartitions = 1 + (j == N - 1 ? 0 : MinPartitions[j + 1]);
7837 bool IsTable = j - i + 1 >= MinJumpTableSize;
7838 unsigned Tables = IsTable + (j == N - 1 ? 0 : NumTables[j + 1]);
7840 // If this j leads to fewer partitions, or same number of partitions
7841 // with more lookup tables, it is a better partitioning.
7842 if (NumPartitions < MinPartitions[i] ||
7843 (NumPartitions == MinPartitions[i] && Tables > NumTables[i])) {
7844 MinPartitions[i] = NumPartitions;
7846 NumTables[i] = Tables;
7852 // Iterate over the partitions, replacing some with jump tables in-place.
7853 unsigned DstIndex = 0;
7854 for (unsigned First = 0, Last; First < N; First = Last + 1) {
7855 Last = LastElement[First];
7856 assert(Last >= First);
7857 assert(DstIndex <= First);
7858 unsigned NumClusters = Last - First + 1;
7860 CaseCluster JTCluster;
7861 if (NumClusters >= MinJumpTableSize &&
7862 buildJumpTable(Clusters, First, Last, SI, DefaultMBB, JTCluster)) {
7863 Clusters[DstIndex++] = JTCluster;
7865 for (unsigned I = First; I <= Last; ++I)
7866 std::memmove(&Clusters[DstIndex++], &Clusters[I], sizeof(Clusters[I]));
7869 Clusters.resize(DstIndex);
7872 bool SelectionDAGBuilder::rangeFitsInWord(const APInt &Low, const APInt &High) {
7873 // FIXME: Using the pointer type doesn't seem ideal.
7874 uint64_t BW = DAG.getDataLayout().getPointerSizeInBits();
7875 uint64_t Range = (High - Low).getLimitedValue(UINT64_MAX - 1) + 1;
7879 bool SelectionDAGBuilder::isSuitableForBitTests(unsigned NumDests,
7882 const APInt &High) {
7883 // FIXME: I don't think NumCmps is the correct metric: a single case and a
7884 // range of cases both require only one branch to lower. Just looking at the
7885 // number of clusters and destinations should be enough to decide whether to
7888 // To lower a range with bit tests, the range must fit the bitwidth of a
7890 if (!rangeFitsInWord(Low, High))
7893 // Decide whether it's profitable to lower this range with bit tests. Each
7894 // destination requires a bit test and branch, and there is an overall range
7895 // check branch. For a small number of clusters, separate comparisons might be
7896 // cheaper, and for many destinations, splitting the range might be better.
7897 return (NumDests == 1 && NumCmps >= 3) ||
7898 (NumDests == 2 && NumCmps >= 5) ||
7899 (NumDests == 3 && NumCmps >= 6);
7902 bool SelectionDAGBuilder::buildBitTests(CaseClusterVector &Clusters,
7903 unsigned First, unsigned Last,
7904 const SwitchInst *SI,
7905 CaseCluster &BTCluster) {
7906 assert(First <= Last);
7910 BitVector Dests(FuncInfo.MF->getNumBlockIDs());
7911 unsigned NumCmps = 0;
7912 for (int64_t I = First; I <= Last; ++I) {
7913 assert(Clusters[I].Kind == CC_Range);
7914 Dests.set(Clusters[I].MBB->getNumber());
7915 NumCmps += (Clusters[I].Low == Clusters[I].High) ? 1 : 2;
7917 unsigned NumDests = Dests.count();
7919 APInt Low = Clusters[First].Low->getValue();
7920 APInt High = Clusters[Last].High->getValue();
7921 assert(Low.slt(High));
7923 if (!isSuitableForBitTests(NumDests, NumCmps, Low, High))
7929 const int BitWidth = DAG.getTargetLoweringInfo()
7930 .getPointerTy(DAG.getDataLayout())
7932 assert(rangeFitsInWord(Low, High) && "Case range must fit in bit mask!");
7934 // Check if the clusters cover a contiguous range such that no value in the
7935 // range will jump to the default statement.
7936 bool ContiguousRange = true;
7937 for (int64_t I = First + 1; I <= Last; ++I) {
7938 if (Clusters[I].Low->getValue() != Clusters[I - 1].High->getValue() + 1) {
7939 ContiguousRange = false;
7944 if (Low.isStrictlyPositive() && High.slt(BitWidth)) {
7945 // Optimize the case where all the case values fit in a word without having
7946 // to subtract minValue. In this case, we can optimize away the subtraction.
7947 LowBound = APInt::getNullValue(Low.getBitWidth());
7949 ContiguousRange = false;
7952 CmpRange = High - Low;
7956 uint32_t TotalWeight = 0;
7957 for (unsigned i = First; i <= Last; ++i) {
7958 // Find the CaseBits for this destination.
7960 for (j = 0; j < CBV.size(); ++j)
7961 if (CBV[j].BB == Clusters[i].MBB)
7963 if (j == CBV.size())
7964 CBV.push_back(CaseBits(0, Clusters[i].MBB, 0, 0));
7965 CaseBits *CB = &CBV[j];
7967 // Update Mask, Bits and ExtraWeight.
7968 uint64_t Lo = (Clusters[i].Low->getValue() - LowBound).getZExtValue();
7969 uint64_t Hi = (Clusters[i].High->getValue() - LowBound).getZExtValue();
7970 assert(Hi >= Lo && Hi < 64 && "Invalid bit case!");
7971 CB->Mask |= (-1ULL >> (63 - (Hi - Lo))) << Lo;
7972 CB->Bits += Hi - Lo + 1;
7973 CB->ExtraWeight += Clusters[i].Weight;
7974 TotalWeight += Clusters[i].Weight;
7975 assert(TotalWeight >= Clusters[i].Weight && "Weight overflow!");
7979 std::sort(CBV.begin(), CBV.end(), [](const CaseBits &a, const CaseBits &b) {
7980 // Sort by weight first, number of bits second.
7981 if (a.ExtraWeight != b.ExtraWeight)
7982 return a.ExtraWeight > b.ExtraWeight;
7983 return a.Bits > b.Bits;
7986 for (auto &CB : CBV) {
7987 MachineBasicBlock *BitTestBB =
7988 FuncInfo.MF->CreateMachineBasicBlock(SI->getParent());
7989 BTI.push_back(BitTestCase(CB.Mask, BitTestBB, CB.BB, CB.ExtraWeight));
7991 BitTestCases.emplace_back(std::move(LowBound), std::move(CmpRange),
7992 SI->getCondition(), -1U, MVT::Other, false,
7993 ContiguousRange, nullptr, nullptr, std::move(BTI),
7996 BTCluster = CaseCluster::bitTests(Clusters[First].Low, Clusters[Last].High,
7997 BitTestCases.size() - 1, TotalWeight);
8001 void SelectionDAGBuilder::findBitTestClusters(CaseClusterVector &Clusters,
8002 const SwitchInst *SI) {
8003 // Partition Clusters into as few subsets as possible, where each subset has a
8004 // range that fits in a machine word and has <= 3 unique destinations.
8007 // Clusters must be sorted and contain Range or JumpTable clusters.
8008 assert(!Clusters.empty());
8009 assert(Clusters[0].Kind == CC_Range || Clusters[0].Kind == CC_JumpTable);
8010 for (const CaseCluster &C : Clusters)
8011 assert(C.Kind == CC_Range || C.Kind == CC_JumpTable);
8012 for (unsigned i = 1; i < Clusters.size(); ++i)
8013 assert(Clusters[i-1].High->getValue().slt(Clusters[i].Low->getValue()));
8016 // The algorithm below is not suitable for -O0.
8017 if (TM.getOptLevel() == CodeGenOpt::None)
8020 // If target does not have legal shift left, do not emit bit tests at all.
8021 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8022 EVT PTy = TLI.getPointerTy(DAG.getDataLayout());
8023 if (!TLI.isOperationLegal(ISD::SHL, PTy))
8026 int BitWidth = PTy.getSizeInBits();
8027 const int64_t N = Clusters.size();
8029 // MinPartitions[i] is the minimum nbr of partitions of Clusters[i..N-1].
8030 SmallVector<unsigned, 8> MinPartitions(N);
8031 // LastElement[i] is the last element of the partition starting at i.
8032 SmallVector<unsigned, 8> LastElement(N);
8034 // FIXME: This might not be the best algorithm for finding bit test clusters.
8036 // Base case: There is only one way to partition Clusters[N-1].
8037 MinPartitions[N - 1] = 1;
8038 LastElement[N - 1] = N - 1;
8040 // Note: loop indexes are signed to avoid underflow.
8041 for (int64_t i = N - 2; i >= 0; --i) {
8042 // Find optimal partitioning of Clusters[i..N-1].
8043 // Baseline: Put Clusters[i] into a partition on its own.
8044 MinPartitions[i] = MinPartitions[i + 1] + 1;
8047 // Search for a solution that results in fewer partitions.
8048 // Note: the search is limited by BitWidth, reducing time complexity.
8049 for (int64_t j = std::min(N - 1, i + BitWidth - 1); j > i; --j) {
8050 // Try building a partition from Clusters[i..j].
8053 if (!rangeFitsInWord(Clusters[i].Low->getValue(),
8054 Clusters[j].High->getValue()))
8057 // Check nbr of destinations and cluster types.
8058 // FIXME: This works, but doesn't seem very efficient.
8059 bool RangesOnly = true;
8060 BitVector Dests(FuncInfo.MF->getNumBlockIDs());
8061 for (int64_t k = i; k <= j; k++) {
8062 if (Clusters[k].Kind != CC_Range) {
8066 Dests.set(Clusters[k].MBB->getNumber());
8068 if (!RangesOnly || Dests.count() > 3)
8071 // Check if it's a better partition.
8072 unsigned NumPartitions = 1 + (j == N - 1 ? 0 : MinPartitions[j + 1]);
8073 if (NumPartitions < MinPartitions[i]) {
8074 // Found a better partition.
8075 MinPartitions[i] = NumPartitions;
8081 // Iterate over the partitions, replacing with bit-test clusters in-place.
8082 unsigned DstIndex = 0;
8083 for (unsigned First = 0, Last; First < N; First = Last + 1) {
8084 Last = LastElement[First];
8085 assert(First <= Last);
8086 assert(DstIndex <= First);
8088 CaseCluster BitTestCluster;
8089 if (buildBitTests(Clusters, First, Last, SI, BitTestCluster)) {
8090 Clusters[DstIndex++] = BitTestCluster;
8092 size_t NumClusters = Last - First + 1;
8093 std::memmove(&Clusters[DstIndex], &Clusters[First],
8094 sizeof(Clusters[0]) * NumClusters);
8095 DstIndex += NumClusters;
8098 Clusters.resize(DstIndex);
8101 void SelectionDAGBuilder::lowerWorkItem(SwitchWorkListItem W, Value *Cond,
8102 MachineBasicBlock *SwitchMBB,
8103 MachineBasicBlock *DefaultMBB) {
8104 MachineFunction *CurMF = FuncInfo.MF;
8105 MachineBasicBlock *NextMBB = nullptr;
8106 MachineFunction::iterator BBI(W.MBB);
8107 if (++BBI != FuncInfo.MF->end())
8110 unsigned Size = W.LastCluster - W.FirstCluster + 1;
8112 BranchProbabilityInfo *BPI = FuncInfo.BPI;
8114 if (Size == 2 && W.MBB == SwitchMBB) {
8115 // If any two of the cases has the same destination, and if one value
8116 // is the same as the other, but has one bit unset that the other has set,
8117 // use bit manipulation to do two compares at once. For example:
8118 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
8119 // TODO: This could be extended to merge any 2 cases in switches with 3
8121 // TODO: Handle cases where W.CaseBB != SwitchBB.
8122 CaseCluster &Small = *W.FirstCluster;
8123 CaseCluster &Big = *W.LastCluster;
8125 if (Small.Low == Small.High && Big.Low == Big.High &&
8126 Small.MBB == Big.MBB) {
8127 const APInt &SmallValue = Small.Low->getValue();
8128 const APInt &BigValue = Big.Low->getValue();
8130 // Check that there is only one bit different.
8131 APInt CommonBit = BigValue ^ SmallValue;
8132 if (CommonBit.isPowerOf2()) {
8133 SDValue CondLHS = getValue(Cond);
8134 EVT VT = CondLHS.getValueType();
8135 SDLoc DL = getCurSDLoc();
8137 SDValue Or = DAG.getNode(ISD::OR, DL, VT, CondLHS,
8138 DAG.getConstant(CommonBit, DL, VT));
8139 SDValue Cond = DAG.getSetCC(
8140 DL, MVT::i1, Or, DAG.getConstant(BigValue | SmallValue, DL, VT),
8143 // Update successor info.
8144 // Both Small and Big will jump to Small.BB, so we sum up the weights.
8145 addSuccessorWithWeight(SwitchMBB, Small.MBB, Small.Weight + Big.Weight);
8146 addSuccessorWithWeight(
8147 SwitchMBB, DefaultMBB,
8148 // The default destination is the first successor in IR.
8149 BPI ? BPI->getEdgeWeight(SwitchMBB->getBasicBlock(), (unsigned)0)
8152 // Insert the true branch.
8154 DAG.getNode(ISD::BRCOND, DL, MVT::Other, getControlRoot(), Cond,
8155 DAG.getBasicBlock(Small.MBB));
8156 // Insert the false branch.
8157 BrCond = DAG.getNode(ISD::BR, DL, MVT::Other, BrCond,
8158 DAG.getBasicBlock(DefaultMBB));
8160 DAG.setRoot(BrCond);
8166 if (TM.getOptLevel() != CodeGenOpt::None) {
8167 // Order cases by weight so the most likely case will be checked first.
8168 std::sort(W.FirstCluster, W.LastCluster + 1,
8169 [](const CaseCluster &a, const CaseCluster &b) {
8170 return a.Weight > b.Weight;
8173 // Rearrange the case blocks so that the last one falls through if possible
8174 // without without changing the order of weights.
8175 for (CaseClusterIt I = W.LastCluster; I > W.FirstCluster; ) {
8177 if (I->Weight > W.LastCluster->Weight)
8179 if (I->Kind == CC_Range && I->MBB == NextMBB) {
8180 std::swap(*I, *W.LastCluster);
8186 // Compute total weight.
8187 uint32_t DefaultWeight = W.DefaultWeight;
8188 uint32_t UnhandledWeights = DefaultWeight;
8189 for (CaseClusterIt I = W.FirstCluster; I <= W.LastCluster; ++I) {
8190 UnhandledWeights += I->Weight;
8191 assert(UnhandledWeights >= I->Weight && "Weight overflow!");
8194 MachineBasicBlock *CurMBB = W.MBB;
8195 for (CaseClusterIt I = W.FirstCluster, E = W.LastCluster; I <= E; ++I) {
8196 MachineBasicBlock *Fallthrough;
8197 if (I == W.LastCluster) {
8198 // For the last cluster, fall through to the default destination.
8199 Fallthrough = DefaultMBB;
8201 Fallthrough = CurMF->CreateMachineBasicBlock(CurMBB->getBasicBlock());
8202 CurMF->insert(BBI, Fallthrough);
8203 // Put Cond in a virtual register to make it available from the new blocks.
8204 ExportFromCurrentBlock(Cond);
8206 UnhandledWeights -= I->Weight;
8209 case CC_JumpTable: {
8210 // FIXME: Optimize away range check based on pivot comparisons.
8211 JumpTableHeader *JTH = &JTCases[I->JTCasesIndex].first;
8212 JumpTable *JT = &JTCases[I->JTCasesIndex].second;
8214 // The jump block hasn't been inserted yet; insert it here.
8215 MachineBasicBlock *JumpMBB = JT->MBB;
8216 CurMF->insert(BBI, JumpMBB);
8218 uint32_t JumpWeight = I->Weight;
8219 uint32_t FallthroughWeight = UnhandledWeights;
8221 // If the default statement is a target of the jump table, we evenly
8222 // distribute the default weight to successors of CurMBB. Also update
8223 // the weight on the edge from JumpMBB to Fallthrough.
8224 for (MachineBasicBlock::succ_iterator SI = JumpMBB->succ_begin(),
8225 SE = JumpMBB->succ_end();
8227 if (*SI == DefaultMBB) {
8228 JumpWeight += DefaultWeight / 2;
8229 FallthroughWeight -= DefaultWeight / 2;
8230 JumpMBB->setSuccWeight(SI, DefaultWeight / 2);
8235 addSuccessorWithWeight(CurMBB, Fallthrough, FallthroughWeight);
8236 addSuccessorWithWeight(CurMBB, JumpMBB, JumpWeight);
8238 // The jump table header will be inserted in our current block, do the
8239 // range check, and fall through to our fallthrough block.
8240 JTH->HeaderBB = CurMBB;
8241 JT->Default = Fallthrough; // FIXME: Move Default to JumpTableHeader.
8243 // If we're in the right place, emit the jump table header right now.
8244 if (CurMBB == SwitchMBB) {
8245 visitJumpTableHeader(*JT, *JTH, SwitchMBB);
8246 JTH->Emitted = true;
8251 // FIXME: Optimize away range check based on pivot comparisons.
8252 BitTestBlock *BTB = &BitTestCases[I->BTCasesIndex];
8254 // The bit test blocks haven't been inserted yet; insert them here.
8255 for (BitTestCase &BTC : BTB->Cases)
8256 CurMF->insert(BBI, BTC.ThisBB);
8258 // Fill in fields of the BitTestBlock.
8259 BTB->Parent = CurMBB;
8260 BTB->Default = Fallthrough;
8262 BTB->DefaultWeight = UnhandledWeights;
8263 // If the cases in bit test don't form a contiguous range, we evenly
8264 // distribute the weight on the edge to Fallthrough to two successors
8266 if (!BTB->ContiguousRange) {
8267 BTB->Weight += DefaultWeight / 2;
8268 BTB->DefaultWeight -= DefaultWeight / 2;
8271 // If we're in the right place, emit the bit test header right now.
8272 if (CurMBB == SwitchMBB) {
8273 visitBitTestHeader(*BTB, SwitchMBB);
8274 BTB->Emitted = true;
8279 const Value *RHS, *LHS, *MHS;
8281 if (I->Low == I->High) {
8282 // Check Cond == I->Low.
8288 // Check I->Low <= Cond <= I->High.
8295 // The false weight is the sum of all unhandled cases.
8296 CaseBlock CB(CC, LHS, RHS, MHS, I->MBB, Fallthrough, CurMBB, I->Weight,
8299 if (CurMBB == SwitchMBB)
8300 visitSwitchCase(CB, SwitchMBB);
8302 SwitchCases.push_back(CB);
8307 CurMBB = Fallthrough;
8311 unsigned SelectionDAGBuilder::caseClusterRank(const CaseCluster &CC,
8312 CaseClusterIt First,
8313 CaseClusterIt Last) {
8314 return std::count_if(First, Last + 1, [&](const CaseCluster &X) {
8315 if (X.Weight != CC.Weight)
8316 return X.Weight > CC.Weight;
8318 // Ties are broken by comparing the case value.
8319 return X.Low->getValue().slt(CC.Low->getValue());
8323 void SelectionDAGBuilder::splitWorkItem(SwitchWorkList &WorkList,
8324 const SwitchWorkListItem &W,
8326 MachineBasicBlock *SwitchMBB) {
8327 assert(W.FirstCluster->Low->getValue().slt(W.LastCluster->Low->getValue()) &&
8328 "Clusters not sorted?");
8330 assert(W.LastCluster - W.FirstCluster + 1 >= 2 && "Too small to split!");
8332 // Balance the tree based on branch weights to create a near-optimal (in terms
8333 // of search time given key frequency) binary search tree. See e.g. Kurt
8334 // Mehlhorn "Nearly Optimal Binary Search Trees" (1975).
8335 CaseClusterIt LastLeft = W.FirstCluster;
8336 CaseClusterIt FirstRight = W.LastCluster;
8337 uint32_t LeftWeight = LastLeft->Weight + W.DefaultWeight / 2;
8338 uint32_t RightWeight = FirstRight->Weight + W.DefaultWeight / 2;
8340 // Move LastLeft and FirstRight towards each other from opposite directions to
8341 // find a partitioning of the clusters which balances the weight on both
8342 // sides. If LeftWeight and RightWeight are equal, alternate which side is
8343 // taken to ensure 0-weight nodes are distributed evenly.
8345 while (LastLeft + 1 < FirstRight) {
8346 if (LeftWeight < RightWeight || (LeftWeight == RightWeight && (I & 1)))
8347 LeftWeight += (++LastLeft)->Weight;
8349 RightWeight += (--FirstRight)->Weight;
8354 // Our binary search tree differs from a typical BST in that ours can have up
8355 // to three values in each leaf. The pivot selection above doesn't take that
8356 // into account, which means the tree might require more nodes and be less
8357 // efficient. We compensate for this here.
8359 unsigned NumLeft = LastLeft - W.FirstCluster + 1;
8360 unsigned NumRight = W.LastCluster - FirstRight + 1;
8362 if (std::min(NumLeft, NumRight) < 3 && std::max(NumLeft, NumRight) > 3) {
8363 // If one side has less than 3 clusters, and the other has more than 3,
8364 // consider taking a cluster from the other side.
8366 if (NumLeft < NumRight) {
8367 // Consider moving the first cluster on the right to the left side.
8368 CaseCluster &CC = *FirstRight;
8369 unsigned RightSideRank = caseClusterRank(CC, FirstRight, W.LastCluster);
8370 unsigned LeftSideRank = caseClusterRank(CC, W.FirstCluster, LastLeft);
8371 if (LeftSideRank <= RightSideRank) {
8372 // Moving the cluster to the left does not demote it.
8378 assert(NumRight < NumLeft);
8379 // Consider moving the last element on the left to the right side.
8380 CaseCluster &CC = *LastLeft;
8381 unsigned LeftSideRank = caseClusterRank(CC, W.FirstCluster, LastLeft);
8382 unsigned RightSideRank = caseClusterRank(CC, FirstRight, W.LastCluster);
8383 if (RightSideRank <= LeftSideRank) {
8384 // Moving the cluster to the right does not demot it.
8394 assert(LastLeft + 1 == FirstRight);
8395 assert(LastLeft >= W.FirstCluster);
8396 assert(FirstRight <= W.LastCluster);
8398 // Use the first element on the right as pivot since we will make less-than
8399 // comparisons against it.
8400 CaseClusterIt PivotCluster = FirstRight;
8401 assert(PivotCluster > W.FirstCluster);
8402 assert(PivotCluster <= W.LastCluster);
8404 CaseClusterIt FirstLeft = W.FirstCluster;
8405 CaseClusterIt LastRight = W.LastCluster;
8407 const ConstantInt *Pivot = PivotCluster->Low;
8409 // New blocks will be inserted immediately after the current one.
8410 MachineFunction::iterator BBI(W.MBB);
8413 // We will branch to the LHS if Value < Pivot. If LHS is a single cluster,
8414 // we can branch to its destination directly if it's squeezed exactly in
8415 // between the known lower bound and Pivot - 1.
8416 MachineBasicBlock *LeftMBB;
8417 if (FirstLeft == LastLeft && FirstLeft->Kind == CC_Range &&
8418 FirstLeft->Low == W.GE &&
8419 (FirstLeft->High->getValue() + 1LL) == Pivot->getValue()) {
8420 LeftMBB = FirstLeft->MBB;
8422 LeftMBB = FuncInfo.MF->CreateMachineBasicBlock(W.MBB->getBasicBlock());
8423 FuncInfo.MF->insert(BBI, LeftMBB);
8425 {LeftMBB, FirstLeft, LastLeft, W.GE, Pivot, W.DefaultWeight / 2});
8426 // Put Cond in a virtual register to make it available from the new blocks.
8427 ExportFromCurrentBlock(Cond);
8430 // Similarly, we will branch to the RHS if Value >= Pivot. If RHS is a
8431 // single cluster, RHS.Low == Pivot, and we can branch to its destination
8432 // directly if RHS.High equals the current upper bound.
8433 MachineBasicBlock *RightMBB;
8434 if (FirstRight == LastRight && FirstRight->Kind == CC_Range &&
8435 W.LT && (FirstRight->High->getValue() + 1ULL) == W.LT->getValue()) {
8436 RightMBB = FirstRight->MBB;
8438 RightMBB = FuncInfo.MF->CreateMachineBasicBlock(W.MBB->getBasicBlock());
8439 FuncInfo.MF->insert(BBI, RightMBB);
8441 {RightMBB, FirstRight, LastRight, Pivot, W.LT, W.DefaultWeight / 2});
8442 // Put Cond in a virtual register to make it available from the new blocks.
8443 ExportFromCurrentBlock(Cond);
8446 // Create the CaseBlock record that will be used to lower the branch.
8447 CaseBlock CB(ISD::SETLT, Cond, Pivot, nullptr, LeftMBB, RightMBB, W.MBB,
8448 LeftWeight, RightWeight);
8450 if (W.MBB == SwitchMBB)
8451 visitSwitchCase(CB, SwitchMBB);
8453 SwitchCases.push_back(CB);
8456 void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) {
8457 // Extract cases from the switch.
8458 BranchProbabilityInfo *BPI = FuncInfo.BPI;
8459 CaseClusterVector Clusters;
8460 Clusters.reserve(SI.getNumCases());
8461 for (auto I : SI.cases()) {
8462 MachineBasicBlock *Succ = FuncInfo.MBBMap[I.getCaseSuccessor()];
8463 const ConstantInt *CaseVal = I.getCaseValue();
8465 BPI ? BPI->getEdgeWeight(SI.getParent(), I.getSuccessorIndex()) : 0;
8466 Clusters.push_back(CaseCluster::range(CaseVal, CaseVal, Succ, Weight));
8469 MachineBasicBlock *DefaultMBB = FuncInfo.MBBMap[SI.getDefaultDest()];
8471 // Cluster adjacent cases with the same destination. We do this at all
8472 // optimization levels because it's cheap to do and will make codegen faster
8473 // if there are many clusters.
8474 sortAndRangeify(Clusters);
8476 if (TM.getOptLevel() != CodeGenOpt::None) {
8477 // Replace an unreachable default with the most popular destination.
8478 // FIXME: Exploit unreachable default more aggressively.
8479 bool UnreachableDefault =
8480 isa<UnreachableInst>(SI.getDefaultDest()->getFirstNonPHIOrDbg());
8481 if (UnreachableDefault && !Clusters.empty()) {
8482 DenseMap<const BasicBlock *, unsigned> Popularity;
8483 unsigned MaxPop = 0;
8484 const BasicBlock *MaxBB = nullptr;
8485 for (auto I : SI.cases()) {
8486 const BasicBlock *BB = I.getCaseSuccessor();
8487 if (++Popularity[BB] > MaxPop) {
8488 MaxPop = Popularity[BB];
8493 assert(MaxPop > 0 && MaxBB);
8494 DefaultMBB = FuncInfo.MBBMap[MaxBB];
8496 // Remove cases that were pointing to the destination that is now the
8498 CaseClusterVector New;
8499 New.reserve(Clusters.size());
8500 for (CaseCluster &CC : Clusters) {
8501 if (CC.MBB != DefaultMBB)
8504 Clusters = std::move(New);
8508 // If there is only the default destination, jump there directly.
8509 MachineBasicBlock *SwitchMBB = FuncInfo.MBB;
8510 if (Clusters.empty()) {
8511 SwitchMBB->addSuccessor(DefaultMBB);
8512 if (DefaultMBB != NextBlock(SwitchMBB)) {
8513 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other,
8514 getControlRoot(), DAG.getBasicBlock(DefaultMBB)));
8519 findJumpTables(Clusters, &SI, DefaultMBB);
8520 findBitTestClusters(Clusters, &SI);
8523 dbgs() << "Case clusters: ";
8524 for (const CaseCluster &C : Clusters) {
8525 if (C.Kind == CC_JumpTable) dbgs() << "JT:";
8526 if (C.Kind == CC_BitTests) dbgs() << "BT:";
8528 C.Low->getValue().print(dbgs(), true);
8529 if (C.Low != C.High) {
8531 C.High->getValue().print(dbgs(), true);
8538 assert(!Clusters.empty());
8539 SwitchWorkList WorkList;
8540 CaseClusterIt First = Clusters.begin();
8541 CaseClusterIt Last = Clusters.end() - 1;
8542 uint32_t DefaultWeight = getEdgeWeight(SwitchMBB, DefaultMBB);
8543 WorkList.push_back({SwitchMBB, First, Last, nullptr, nullptr, DefaultWeight});
8545 while (!WorkList.empty()) {
8546 SwitchWorkListItem W = WorkList.back();
8547 WorkList.pop_back();
8548 unsigned NumClusters = W.LastCluster - W.FirstCluster + 1;
8550 if (NumClusters > 3 && TM.getOptLevel() != CodeGenOpt::None) {
8551 // For optimized builds, lower large range as a balanced binary tree.
8552 splitWorkItem(WorkList, W, SI.getCondition(), SwitchMBB);
8556 lowerWorkItem(W, SI.getCondition(), SwitchMBB, DefaultMBB);