1 //===-- SelectionDAGBuilder.cpp - Selection-DAG building ------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements routines for translating from LLVM IR into SelectionDAG IR.
12 //===----------------------------------------------------------------------===//
14 #include "SelectionDAGBuilder.h"
15 #include "SDNodeDbgValue.h"
16 #include "llvm/ADT/BitVector.h"
17 #include "llvm/ADT/Optional.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/Analysis/AliasAnalysis.h"
20 #include "llvm/Analysis/BranchProbabilityInfo.h"
21 #include "llvm/Analysis/ConstantFolding.h"
22 #include "llvm/Analysis/ValueTracking.h"
23 #include "llvm/CodeGen/Analysis.h"
24 #include "llvm/CodeGen/FastISel.h"
25 #include "llvm/CodeGen/FunctionLoweringInfo.h"
26 #include "llvm/CodeGen/GCMetadata.h"
27 #include "llvm/CodeGen/GCStrategy.h"
28 #include "llvm/CodeGen/MachineFrameInfo.h"
29 #include "llvm/CodeGen/MachineFunction.h"
30 #include "llvm/CodeGen/MachineInstrBuilder.h"
31 #include "llvm/CodeGen/MachineJumpTableInfo.h"
32 #include "llvm/CodeGen/MachineModuleInfo.h"
33 #include "llvm/CodeGen/MachineRegisterInfo.h"
34 #include "llvm/CodeGen/SelectionDAG.h"
35 #include "llvm/CodeGen/StackMaps.h"
36 #include "llvm/IR/CallingConv.h"
37 #include "llvm/IR/Constants.h"
38 #include "llvm/IR/DataLayout.h"
39 #include "llvm/IR/DebugInfo.h"
40 #include "llvm/IR/DerivedTypes.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/GlobalVariable.h"
43 #include "llvm/IR/InlineAsm.h"
44 #include "llvm/IR/Instructions.h"
45 #include "llvm/IR/IntrinsicInst.h"
46 #include "llvm/IR/Intrinsics.h"
47 #include "llvm/IR/LLVMContext.h"
48 #include "llvm/IR/Module.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/raw_ostream.h"
54 #include "llvm/Target/TargetFrameLowering.h"
55 #include "llvm/Target/TargetInstrInfo.h"
56 #include "llvm/Target/TargetIntrinsicInfo.h"
57 #include "llvm/Target/TargetLibraryInfo.h"
58 #include "llvm/Target/TargetLowering.h"
59 #include "llvm/Target/TargetOptions.h"
60 #include "llvm/Target/TargetSelectionDAGInfo.h"
61 #include "llvm/Target/TargetSubtargetInfo.h"
65 #define DEBUG_TYPE "isel"
67 /// LimitFloatPrecision - Generate low-precision inline sequences for
68 /// some float libcalls (6, 8 or 12 bits).
69 static unsigned LimitFloatPrecision;
71 static cl::opt<unsigned, true>
72 LimitFPPrecision("limit-float-precision",
73 cl::desc("Generate low-precision inline sequences "
74 "for some float libcalls"),
75 cl::location(LimitFloatPrecision),
78 // Limit the width of DAG chains. This is important in general to prevent
79 // prevent DAG-based analysis from blowing up. For example, alias analysis and
80 // load clustering may not complete in reasonable time. It is difficult to
81 // recognize and avoid this situation within each individual analysis, and
82 // future analyses are likely to have the same behavior. Limiting DAG width is
83 // the safe approach, and will be especially important with global DAGs.
85 // MaxParallelChains default is arbitrarily high to avoid affecting
86 // optimization, but could be lowered to improve compile time. Any ld-ld-st-st
87 // sequence over this should have been converted to llvm.memcpy by the
88 // frontend. It easy to induce this behavior with .ll code such as:
89 // %buffer = alloca [4096 x i8]
90 // %data = load [4096 x i8]* %argPtr
91 // store [4096 x i8] %data, [4096 x i8]* %buffer
92 static const unsigned MaxParallelChains = 64;
94 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, SDLoc DL,
95 const SDValue *Parts, unsigned NumParts,
96 MVT PartVT, EVT ValueVT, const Value *V);
98 /// getCopyFromParts - Create a value that contains the specified legal parts
99 /// combined into the value they represent. If the parts combine to a type
100 /// larger then ValueVT then AssertOp can be used to specify whether the extra
101 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
102 /// (ISD::AssertSext).
103 static SDValue getCopyFromParts(SelectionDAG &DAG, SDLoc DL,
104 const SDValue *Parts,
105 unsigned NumParts, MVT PartVT, EVT ValueVT,
107 ISD::NodeType AssertOp = ISD::DELETED_NODE) {
108 if (ValueVT.isVector())
109 return getCopyFromPartsVector(DAG, DL, Parts, NumParts,
112 assert(NumParts > 0 && "No parts to assemble!");
113 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
114 SDValue Val = Parts[0];
117 // Assemble the value from multiple parts.
118 if (ValueVT.isInteger()) {
119 unsigned PartBits = PartVT.getSizeInBits();
120 unsigned ValueBits = ValueVT.getSizeInBits();
122 // Assemble the power of 2 part.
123 unsigned RoundParts = NumParts & (NumParts - 1) ?
124 1 << Log2_32(NumParts) : NumParts;
125 unsigned RoundBits = PartBits * RoundParts;
126 EVT RoundVT = RoundBits == ValueBits ?
127 ValueVT : EVT::getIntegerVT(*DAG.getContext(), RoundBits);
130 EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2);
132 if (RoundParts > 2) {
133 Lo = getCopyFromParts(DAG, DL, Parts, RoundParts / 2,
135 Hi = getCopyFromParts(DAG, DL, Parts + RoundParts / 2,
136 RoundParts / 2, PartVT, HalfVT, V);
138 Lo = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[0]);
139 Hi = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[1]);
142 if (TLI.isBigEndian())
145 Val = DAG.getNode(ISD::BUILD_PAIR, DL, RoundVT, Lo, Hi);
147 if (RoundParts < NumParts) {
148 // Assemble the trailing non-power-of-2 part.
149 unsigned OddParts = NumParts - RoundParts;
150 EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits);
151 Hi = getCopyFromParts(DAG, DL,
152 Parts + RoundParts, OddParts, PartVT, OddVT, V);
154 // Combine the round and odd parts.
156 if (TLI.isBigEndian())
158 EVT TotalVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
159 Hi = DAG.getNode(ISD::ANY_EXTEND, DL, TotalVT, Hi);
160 Hi = DAG.getNode(ISD::SHL, DL, TotalVT, Hi,
161 DAG.getConstant(Lo.getValueType().getSizeInBits(),
162 TLI.getPointerTy()));
163 Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, TotalVT, Lo);
164 Val = DAG.getNode(ISD::OR, DL, TotalVT, Lo, Hi);
166 } else if (PartVT.isFloatingPoint()) {
167 // FP split into multiple FP parts (for ppcf128)
168 assert(ValueVT == EVT(MVT::ppcf128) && PartVT == MVT::f64 &&
171 Lo = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[0]);
172 Hi = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[1]);
173 if (TLI.hasBigEndianPartOrdering(ValueVT))
175 Val = DAG.getNode(ISD::BUILD_PAIR, DL, ValueVT, Lo, Hi);
177 // FP split into integer parts (soft fp)
178 assert(ValueVT.isFloatingPoint() && PartVT.isInteger() &&
179 !PartVT.isVector() && "Unexpected split");
180 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits());
181 Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT, V);
185 // There is now one part, held in Val. Correct it to match ValueVT.
186 EVT PartEVT = Val.getValueType();
188 if (PartEVT == ValueVT)
191 if (PartEVT.isInteger() && ValueVT.isInteger()) {
192 if (ValueVT.bitsLT(PartEVT)) {
193 // For a truncate, see if we have any information to
194 // indicate whether the truncated bits will always be
195 // zero or sign-extension.
196 if (AssertOp != ISD::DELETED_NODE)
197 Val = DAG.getNode(AssertOp, DL, PartEVT, Val,
198 DAG.getValueType(ValueVT));
199 return DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
201 return DAG.getNode(ISD::ANY_EXTEND, DL, ValueVT, Val);
204 if (PartEVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
205 // FP_ROUND's are always exact here.
206 if (ValueVT.bitsLT(Val.getValueType()))
207 return DAG.getNode(ISD::FP_ROUND, DL, ValueVT, Val,
208 DAG.getTargetConstant(1, TLI.getPointerTy()));
210 return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val);
213 if (PartEVT.getSizeInBits() == ValueVT.getSizeInBits())
214 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
216 llvm_unreachable("Unknown mismatch!");
219 static void diagnosePossiblyInvalidConstraint(LLVMContext &Ctx, const Value *V,
220 const Twine &ErrMsg) {
221 const Instruction *I = dyn_cast_or_null<Instruction>(V);
223 return Ctx.emitError(ErrMsg);
225 const char *AsmError = ", possible invalid constraint for vector type";
226 if (const CallInst *CI = dyn_cast<CallInst>(I))
227 if (isa<InlineAsm>(CI->getCalledValue()))
228 return Ctx.emitError(I, ErrMsg + AsmError);
230 return Ctx.emitError(I, ErrMsg);
233 /// getCopyFromPartsVector - Create a value that contains the specified legal
234 /// parts combined into the value they represent. If the parts combine to a
235 /// type larger then ValueVT then AssertOp can be used to specify whether the
236 /// extra bits are known to be zero (ISD::AssertZext) or sign extended from
237 /// ValueVT (ISD::AssertSext).
238 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, SDLoc DL,
239 const SDValue *Parts, unsigned NumParts,
240 MVT PartVT, EVT ValueVT, const Value *V) {
241 assert(ValueVT.isVector() && "Not a vector value");
242 assert(NumParts > 0 && "No parts to assemble!");
243 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
244 SDValue Val = Parts[0];
246 // Handle a multi-element vector.
250 unsigned NumIntermediates;
252 TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT,
253 NumIntermediates, RegisterVT);
254 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
255 NumParts = NumRegs; // Silence a compiler warning.
256 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
257 assert(RegisterVT == Parts[0].getSimpleValueType() &&
258 "Part type doesn't match part!");
260 // Assemble the parts into intermediate operands.
261 SmallVector<SDValue, 8> Ops(NumIntermediates);
262 if (NumIntermediates == NumParts) {
263 // If the register was not expanded, truncate or copy the value,
265 for (unsigned i = 0; i != NumParts; ++i)
266 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i], 1,
267 PartVT, IntermediateVT, V);
268 } else if (NumParts > 0) {
269 // If the intermediate type was expanded, build the intermediate
270 // operands from the parts.
271 assert(NumParts % NumIntermediates == 0 &&
272 "Must expand into a divisible number of parts!");
273 unsigned Factor = NumParts / NumIntermediates;
274 for (unsigned i = 0; i != NumIntermediates; ++i)
275 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor,
276 PartVT, IntermediateVT, V);
279 // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the
280 // intermediate operands.
281 Val = DAG.getNode(IntermediateVT.isVector() ? ISD::CONCAT_VECTORS
286 // There is now one part, held in Val. Correct it to match ValueVT.
287 EVT PartEVT = Val.getValueType();
289 if (PartEVT == ValueVT)
292 if (PartEVT.isVector()) {
293 // If the element type of the source/dest vectors are the same, but the
294 // parts vector has more elements than the value vector, then we have a
295 // vector widening case (e.g. <2 x float> -> <4 x float>). Extract the
297 if (PartEVT.getVectorElementType() == ValueVT.getVectorElementType()) {
298 assert(PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements() &&
299 "Cannot narrow, it would be a lossy transformation");
300 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val,
301 DAG.getConstant(0, TLI.getVectorIdxTy()));
304 // Vector/Vector bitcast.
305 if (ValueVT.getSizeInBits() == PartEVT.getSizeInBits())
306 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
308 assert(PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements() &&
309 "Cannot handle this kind of promotion");
310 // Promoted vector extract
311 bool Smaller = ValueVT.bitsLE(PartEVT);
312 return DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
317 // Trivial bitcast if the types are the same size and the destination
318 // vector type is legal.
319 if (PartEVT.getSizeInBits() == ValueVT.getSizeInBits() &&
320 TLI.isTypeLegal(ValueVT))
321 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
323 // Handle cases such as i8 -> <1 x i1>
324 if (ValueVT.getVectorNumElements() != 1) {
325 diagnosePossiblyInvalidConstraint(*DAG.getContext(), V,
326 "non-trivial scalar-to-vector conversion");
327 return DAG.getUNDEF(ValueVT);
330 if (ValueVT.getVectorNumElements() == 1 &&
331 ValueVT.getVectorElementType() != PartEVT) {
332 bool Smaller = ValueVT.bitsLE(PartEVT);
333 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
334 DL, ValueVT.getScalarType(), Val);
337 return DAG.getNode(ISD::BUILD_VECTOR, DL, ValueVT, Val);
340 static void getCopyToPartsVector(SelectionDAG &DAG, SDLoc dl,
341 SDValue Val, SDValue *Parts, unsigned NumParts,
342 MVT PartVT, const Value *V);
344 /// getCopyToParts - Create a series of nodes that contain the specified value
345 /// split into legal parts. If the parts contain more bits than Val, then, for
346 /// integers, ExtendKind can be used to specify how to generate the extra bits.
347 static void getCopyToParts(SelectionDAG &DAG, SDLoc DL,
348 SDValue Val, SDValue *Parts, unsigned NumParts,
349 MVT PartVT, const Value *V,
350 ISD::NodeType ExtendKind = ISD::ANY_EXTEND) {
351 EVT ValueVT = Val.getValueType();
353 // Handle the vector case separately.
354 if (ValueVT.isVector())
355 return getCopyToPartsVector(DAG, DL, Val, Parts, NumParts, PartVT, V);
357 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
358 unsigned PartBits = PartVT.getSizeInBits();
359 unsigned OrigNumParts = NumParts;
360 assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!");
365 assert(!ValueVT.isVector() && "Vector case handled elsewhere");
366 EVT PartEVT = PartVT;
367 if (PartEVT == ValueVT) {
368 assert(NumParts == 1 && "No-op copy with multiple parts!");
373 if (NumParts * PartBits > ValueVT.getSizeInBits()) {
374 // If the parts cover more bits than the value has, promote the value.
375 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
376 assert(NumParts == 1 && "Do not know what to promote to!");
377 Val = DAG.getNode(ISD::FP_EXTEND, DL, PartVT, Val);
379 assert((PartVT.isInteger() || PartVT == MVT::x86mmx) &&
380 ValueVT.isInteger() &&
381 "Unknown mismatch!");
382 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
383 Val = DAG.getNode(ExtendKind, DL, ValueVT, Val);
384 if (PartVT == MVT::x86mmx)
385 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
387 } else if (PartBits == ValueVT.getSizeInBits()) {
388 // Different types of the same size.
389 assert(NumParts == 1 && PartEVT != ValueVT);
390 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
391 } else if (NumParts * PartBits < ValueVT.getSizeInBits()) {
392 // If the parts cover less bits than value has, truncate the value.
393 assert((PartVT.isInteger() || PartVT == MVT::x86mmx) &&
394 ValueVT.isInteger() &&
395 "Unknown mismatch!");
396 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
397 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
398 if (PartVT == MVT::x86mmx)
399 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
402 // The value may have changed - recompute ValueVT.
403 ValueVT = Val.getValueType();
404 assert(NumParts * PartBits == ValueVT.getSizeInBits() &&
405 "Failed to tile the value with PartVT!");
408 if (PartEVT != ValueVT)
409 diagnosePossiblyInvalidConstraint(*DAG.getContext(), V,
410 "scalar-to-vector conversion failed");
416 // Expand the value into multiple parts.
417 if (NumParts & (NumParts - 1)) {
418 // The number of parts is not a power of 2. Split off and copy the tail.
419 assert(PartVT.isInteger() && ValueVT.isInteger() &&
420 "Do not know what to expand to!");
421 unsigned RoundParts = 1 << Log2_32(NumParts);
422 unsigned RoundBits = RoundParts * PartBits;
423 unsigned OddParts = NumParts - RoundParts;
424 SDValue OddVal = DAG.getNode(ISD::SRL, DL, ValueVT, Val,
425 DAG.getIntPtrConstant(RoundBits));
426 getCopyToParts(DAG, DL, OddVal, Parts + RoundParts, OddParts, PartVT, V);
428 if (TLI.isBigEndian())
429 // The odd parts were reversed by getCopyToParts - unreverse them.
430 std::reverse(Parts + RoundParts, Parts + NumParts);
432 NumParts = RoundParts;
433 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
434 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
437 // The number of parts is a power of 2. Repeatedly bisect the value using
439 Parts[0] = DAG.getNode(ISD::BITCAST, DL,
440 EVT::getIntegerVT(*DAG.getContext(),
441 ValueVT.getSizeInBits()),
444 for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) {
445 for (unsigned i = 0; i < NumParts; i += StepSize) {
446 unsigned ThisBits = StepSize * PartBits / 2;
447 EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits);
448 SDValue &Part0 = Parts[i];
449 SDValue &Part1 = Parts[i+StepSize/2];
451 Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
452 ThisVT, Part0, DAG.getIntPtrConstant(1));
453 Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
454 ThisVT, Part0, DAG.getIntPtrConstant(0));
456 if (ThisBits == PartBits && ThisVT != PartVT) {
457 Part0 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part0);
458 Part1 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part1);
463 if (TLI.isBigEndian())
464 std::reverse(Parts, Parts + OrigNumParts);
468 /// getCopyToPartsVector - Create a series of nodes that contain the specified
469 /// value split into legal parts.
470 static void getCopyToPartsVector(SelectionDAG &DAG, SDLoc DL,
471 SDValue Val, SDValue *Parts, unsigned NumParts,
472 MVT PartVT, const Value *V) {
473 EVT ValueVT = Val.getValueType();
474 assert(ValueVT.isVector() && "Not a vector");
475 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
478 EVT PartEVT = PartVT;
479 if (PartEVT == ValueVT) {
481 } else if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) {
482 // Bitconvert vector->vector case.
483 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
484 } else if (PartVT.isVector() &&
485 PartEVT.getVectorElementType() == ValueVT.getVectorElementType() &&
486 PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements()) {
487 EVT ElementVT = PartVT.getVectorElementType();
488 // Vector widening case, e.g. <2 x float> -> <4 x float>. Shuffle in
490 SmallVector<SDValue, 16> Ops;
491 for (unsigned i = 0, e = ValueVT.getVectorNumElements(); i != e; ++i)
492 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
493 ElementVT, Val, DAG.getConstant(i,
494 TLI.getVectorIdxTy())));
496 for (unsigned i = ValueVT.getVectorNumElements(),
497 e = PartVT.getVectorNumElements(); i != e; ++i)
498 Ops.push_back(DAG.getUNDEF(ElementVT));
500 Val = DAG.getNode(ISD::BUILD_VECTOR, DL, PartVT, Ops);
502 // FIXME: Use CONCAT for 2x -> 4x.
504 //SDValue UndefElts = DAG.getUNDEF(VectorTy);
505 //Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts);
506 } else if (PartVT.isVector() &&
507 PartEVT.getVectorElementType().bitsGE(
508 ValueVT.getVectorElementType()) &&
509 PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements()) {
511 // Promoted vector extract
512 bool Smaller = PartEVT.bitsLE(ValueVT);
513 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
516 // Vector -> scalar conversion.
517 assert(ValueVT.getVectorNumElements() == 1 &&
518 "Only trivial vector-to-scalar conversions should get here!");
519 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
520 PartVT, Val, DAG.getConstant(0, TLI.getVectorIdxTy()));
522 bool Smaller = ValueVT.bitsLE(PartVT);
523 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
531 // Handle a multi-element vector.
534 unsigned NumIntermediates;
535 unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT,
537 NumIntermediates, RegisterVT);
538 unsigned NumElements = ValueVT.getVectorNumElements();
540 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
541 NumParts = NumRegs; // Silence a compiler warning.
542 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
544 // Split the vector into intermediate operands.
545 SmallVector<SDValue, 8> Ops(NumIntermediates);
546 for (unsigned i = 0; i != NumIntermediates; ++i) {
547 if (IntermediateVT.isVector())
548 Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL,
550 DAG.getConstant(i * (NumElements / NumIntermediates),
551 TLI.getVectorIdxTy()));
553 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
555 DAG.getConstant(i, TLI.getVectorIdxTy()));
558 // Split the intermediate operands into legal parts.
559 if (NumParts == NumIntermediates) {
560 // If the register was not expanded, promote or copy the value,
562 for (unsigned i = 0; i != NumParts; ++i)
563 getCopyToParts(DAG, DL, Ops[i], &Parts[i], 1, PartVT, V);
564 } else if (NumParts > 0) {
565 // If the intermediate type was expanded, split each the value into
567 assert(NumParts % NumIntermediates == 0 &&
568 "Must expand into a divisible number of parts!");
569 unsigned Factor = NumParts / NumIntermediates;
570 for (unsigned i = 0; i != NumIntermediates; ++i)
571 getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT, V);
576 /// RegsForValue - This struct represents the registers (physical or virtual)
577 /// that a particular set of values is assigned, and the type information
578 /// about the value. The most common situation is to represent one value at a
579 /// time, but struct or array values are handled element-wise as multiple
580 /// values. The splitting of aggregates is performed recursively, so that we
581 /// never have aggregate-typed registers. The values at this point do not
582 /// necessarily have legal types, so each value may require one or more
583 /// registers of some legal type.
585 struct RegsForValue {
586 /// ValueVTs - The value types of the values, which may not be legal, and
587 /// may need be promoted or synthesized from one or more registers.
589 SmallVector<EVT, 4> ValueVTs;
591 /// RegVTs - The value types of the registers. This is the same size as
592 /// ValueVTs and it records, for each value, what the type of the assigned
593 /// register or registers are. (Individual values are never synthesized
594 /// from more than one type of register.)
596 /// With virtual registers, the contents of RegVTs is redundant with TLI's
597 /// getRegisterType member function, however when with physical registers
598 /// it is necessary to have a separate record of the types.
600 SmallVector<MVT, 4> RegVTs;
602 /// Regs - This list holds the registers assigned to the values.
603 /// Each legal or promoted value requires one register, and each
604 /// expanded value requires multiple registers.
606 SmallVector<unsigned, 4> Regs;
610 RegsForValue(const SmallVector<unsigned, 4> ®s,
611 MVT regvt, EVT valuevt)
612 : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {}
614 RegsForValue(LLVMContext &Context, const TargetLowering &tli,
615 unsigned Reg, Type *Ty) {
616 ComputeValueVTs(tli, Ty, ValueVTs);
618 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
619 EVT ValueVT = ValueVTs[Value];
620 unsigned NumRegs = tli.getNumRegisters(Context, ValueVT);
621 MVT RegisterVT = tli.getRegisterType(Context, ValueVT);
622 for (unsigned i = 0; i != NumRegs; ++i)
623 Regs.push_back(Reg + i);
624 RegVTs.push_back(RegisterVT);
629 /// append - Add the specified values to this one.
630 void append(const RegsForValue &RHS) {
631 ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end());
632 RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end());
633 Regs.append(RHS.Regs.begin(), RHS.Regs.end());
636 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
637 /// this value and returns the result as a ValueVTs value. This uses
638 /// Chain/Flag as the input and updates them for the output Chain/Flag.
639 /// If the Flag pointer is NULL, no flag is used.
640 SDValue getCopyFromRegs(SelectionDAG &DAG, FunctionLoweringInfo &FuncInfo,
642 SDValue &Chain, SDValue *Flag,
643 const Value *V = nullptr) const;
645 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
646 /// specified value into the registers specified by this object. This uses
647 /// Chain/Flag as the input and updates them for the output Chain/Flag.
648 /// If the Flag pointer is NULL, no flag is used.
649 void getCopyToRegs(SDValue Val, SelectionDAG &DAG, SDLoc dl,
650 SDValue &Chain, SDValue *Flag, const Value *V) const;
652 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
653 /// operand list. This adds the code marker, matching input operand index
654 /// (if applicable), and includes the number of values added into it.
655 void AddInlineAsmOperands(unsigned Kind,
656 bool HasMatching, unsigned MatchingIdx,
658 std::vector<SDValue> &Ops) const;
662 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
663 /// this value and returns the result as a ValueVT value. This uses
664 /// Chain/Flag as the input and updates them for the output Chain/Flag.
665 /// If the Flag pointer is NULL, no flag is used.
666 SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
667 FunctionLoweringInfo &FuncInfo,
669 SDValue &Chain, SDValue *Flag,
670 const Value *V) const {
671 // A Value with type {} or [0 x %t] needs no registers.
672 if (ValueVTs.empty())
675 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
677 // Assemble the legal parts into the final values.
678 SmallVector<SDValue, 4> Values(ValueVTs.size());
679 SmallVector<SDValue, 8> Parts;
680 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
681 // Copy the legal parts from the registers.
682 EVT ValueVT = ValueVTs[Value];
683 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
684 MVT RegisterVT = RegVTs[Value];
686 Parts.resize(NumRegs);
687 for (unsigned i = 0; i != NumRegs; ++i) {
690 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT);
692 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag);
693 *Flag = P.getValue(2);
696 Chain = P.getValue(1);
699 // If the source register was virtual and if we know something about it,
700 // add an assert node.
701 if (!TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) ||
702 !RegisterVT.isInteger() || RegisterVT.isVector())
705 const FunctionLoweringInfo::LiveOutInfo *LOI =
706 FuncInfo.GetLiveOutRegInfo(Regs[Part+i]);
710 unsigned RegSize = RegisterVT.getSizeInBits();
711 unsigned NumSignBits = LOI->NumSignBits;
712 unsigned NumZeroBits = LOI->KnownZero.countLeadingOnes();
714 if (NumZeroBits == RegSize) {
715 // The current value is a zero.
716 // Explicitly express that as it would be easier for
717 // optimizations to kick in.
718 Parts[i] = DAG.getConstant(0, RegisterVT);
722 // FIXME: We capture more information than the dag can represent. For
723 // now, just use the tightest assertzext/assertsext possible.
725 EVT FromVT(MVT::Other);
726 if (NumSignBits == RegSize)
727 isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1
728 else if (NumZeroBits >= RegSize-1)
729 isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1
730 else if (NumSignBits > RegSize-8)
731 isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8
732 else if (NumZeroBits >= RegSize-8)
733 isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8
734 else if (NumSignBits > RegSize-16)
735 isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16
736 else if (NumZeroBits >= RegSize-16)
737 isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16
738 else if (NumSignBits > RegSize-32)
739 isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32
740 else if (NumZeroBits >= RegSize-32)
741 isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32
745 // Add an assertion node.
746 assert(FromVT != MVT::Other);
747 Parts[i] = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl,
748 RegisterVT, P, DAG.getValueType(FromVT));
751 Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(),
752 NumRegs, RegisterVT, ValueVT, V);
757 return DAG.getNode(ISD::MERGE_VALUES, dl, DAG.getVTList(ValueVTs), Values);
760 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
761 /// specified value into the registers specified by this object. This uses
762 /// Chain/Flag as the input and updates them for the output Chain/Flag.
763 /// If the Flag pointer is NULL, no flag is used.
764 void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, SDLoc dl,
765 SDValue &Chain, SDValue *Flag,
766 const Value *V) const {
767 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
769 // Get the list of the values's legal parts.
770 unsigned NumRegs = Regs.size();
771 SmallVector<SDValue, 8> Parts(NumRegs);
772 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
773 EVT ValueVT = ValueVTs[Value];
774 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
775 MVT RegisterVT = RegVTs[Value];
776 ISD::NodeType ExtendKind =
777 TLI.isZExtFree(Val, RegisterVT)? ISD::ZERO_EXTEND: ISD::ANY_EXTEND;
779 getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value),
780 &Parts[Part], NumParts, RegisterVT, V, ExtendKind);
784 // Copy the parts into the registers.
785 SmallVector<SDValue, 8> Chains(NumRegs);
786 for (unsigned i = 0; i != NumRegs; ++i) {
789 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]);
791 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag);
792 *Flag = Part.getValue(1);
795 Chains[i] = Part.getValue(0);
798 if (NumRegs == 1 || Flag)
799 // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is
800 // flagged to it. That is the CopyToReg nodes and the user are considered
801 // a single scheduling unit. If we create a TokenFactor and return it as
802 // chain, then the TokenFactor is both a predecessor (operand) of the
803 // user as well as a successor (the TF operands are flagged to the user).
804 // c1, f1 = CopyToReg
805 // c2, f2 = CopyToReg
806 // c3 = TokenFactor c1, c2
809 Chain = Chains[NumRegs-1];
811 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
814 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
815 /// operand list. This adds the code marker and includes the number of
816 /// values added into it.
817 void RegsForValue::AddInlineAsmOperands(unsigned Code, bool HasMatching,
818 unsigned MatchingIdx,
820 std::vector<SDValue> &Ops) const {
821 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
823 unsigned Flag = InlineAsm::getFlagWord(Code, Regs.size());
825 Flag = InlineAsm::getFlagWordForMatchingOp(Flag, MatchingIdx);
826 else if (!Regs.empty() &&
827 TargetRegisterInfo::isVirtualRegister(Regs.front())) {
828 // Put the register class of the virtual registers in the flag word. That
829 // way, later passes can recompute register class constraints for inline
830 // assembly as well as normal instructions.
831 // Don't do this for tied operands that can use the regclass information
833 const MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
834 const TargetRegisterClass *RC = MRI.getRegClass(Regs.front());
835 Flag = InlineAsm::getFlagWordForRegClass(Flag, RC->getID());
838 SDValue Res = DAG.getTargetConstant(Flag, MVT::i32);
841 unsigned SP = TLI.getStackPointerRegisterToSaveRestore();
842 for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) {
843 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVTs[Value]);
844 MVT RegisterVT = RegVTs[Value];
845 for (unsigned i = 0; i != NumRegs; ++i) {
846 assert(Reg < Regs.size() && "Mismatch in # registers expected");
847 unsigned TheReg = Regs[Reg++];
848 Ops.push_back(DAG.getRegister(TheReg, RegisterVT));
850 if (TheReg == SP && Code == InlineAsm::Kind_Clobber) {
851 // If we clobbered the stack pointer, MFI should know about it.
852 assert(DAG.getMachineFunction().getFrameInfo()->
853 hasInlineAsmWithSPAdjust());
859 void SelectionDAGBuilder::init(GCFunctionInfo *gfi, AliasAnalysis &aa,
860 const TargetLibraryInfo *li) {
864 DL = DAG.getSubtarget().getDataLayout();
865 Context = DAG.getContext();
866 LPadToCallSiteMap.clear();
869 /// clear - Clear out the current SelectionDAG and the associated
870 /// state and prepare this SelectionDAGBuilder object to be used
871 /// for a new block. This doesn't clear out information about
872 /// additional blocks that are needed to complete switch lowering
873 /// or PHI node updating; that information is cleared out as it is
875 void SelectionDAGBuilder::clear() {
877 UnusedArgNodeMap.clear();
878 PendingLoads.clear();
879 PendingExports.clear();
882 SDNodeOrder = LowestSDNodeOrder;
885 /// clearDanglingDebugInfo - Clear the dangling debug information
886 /// map. This function is separated from the clear so that debug
887 /// information that is dangling in a basic block can be properly
888 /// resolved in a different basic block. This allows the
889 /// SelectionDAG to resolve dangling debug information attached
891 void SelectionDAGBuilder::clearDanglingDebugInfo() {
892 DanglingDebugInfoMap.clear();
895 /// getRoot - Return the current virtual root of the Selection DAG,
896 /// flushing any PendingLoad items. This must be done before emitting
897 /// a store or any other node that may need to be ordered after any
898 /// prior load instructions.
900 SDValue SelectionDAGBuilder::getRoot() {
901 if (PendingLoads.empty())
902 return DAG.getRoot();
904 if (PendingLoads.size() == 1) {
905 SDValue Root = PendingLoads[0];
907 PendingLoads.clear();
911 // Otherwise, we have to make a token factor node.
912 SDValue Root = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
914 PendingLoads.clear();
919 /// getControlRoot - Similar to getRoot, but instead of flushing all the
920 /// PendingLoad items, flush all the PendingExports items. It is necessary
921 /// to do this before emitting a terminator instruction.
923 SDValue SelectionDAGBuilder::getControlRoot() {
924 SDValue Root = DAG.getRoot();
926 if (PendingExports.empty())
929 // Turn all of the CopyToReg chains into one factored node.
930 if (Root.getOpcode() != ISD::EntryToken) {
931 unsigned i = 0, e = PendingExports.size();
932 for (; i != e; ++i) {
933 assert(PendingExports[i].getNode()->getNumOperands() > 1);
934 if (PendingExports[i].getNode()->getOperand(0) == Root)
935 break; // Don't add the root if we already indirectly depend on it.
939 PendingExports.push_back(Root);
942 Root = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
944 PendingExports.clear();
949 void SelectionDAGBuilder::visit(const Instruction &I) {
950 // Set up outgoing PHI node register values before emitting the terminator.
951 if (isa<TerminatorInst>(&I))
952 HandlePHINodesInSuccessorBlocks(I.getParent());
958 visit(I.getOpcode(), I);
960 if (!isa<TerminatorInst>(&I) && !HasTailCall)
961 CopyToExportRegsIfNeeded(&I);
966 void SelectionDAGBuilder::visitPHI(const PHINode &) {
967 llvm_unreachable("SelectionDAGBuilder shouldn't visit PHI nodes!");
970 void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) {
971 // Note: this doesn't use InstVisitor, because it has to work with
972 // ConstantExpr's in addition to instructions.
974 default: llvm_unreachable("Unknown instruction type encountered!");
975 // Build the switch statement using the Instruction.def file.
976 #define HANDLE_INST(NUM, OPCODE, CLASS) \
977 case Instruction::OPCODE: visit##OPCODE((const CLASS&)I); break;
978 #include "llvm/IR/Instruction.def"
982 // resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V,
983 // generate the debug data structures now that we've seen its definition.
984 void SelectionDAGBuilder::resolveDanglingDebugInfo(const Value *V,
986 DanglingDebugInfo &DDI = DanglingDebugInfoMap[V];
988 const DbgValueInst *DI = DDI.getDI();
989 DebugLoc dl = DDI.getdl();
990 unsigned DbgSDNodeOrder = DDI.getSDNodeOrder();
991 MDNode *Variable = DI->getVariable();
992 uint64_t Offset = DI->getOffset();
993 // A dbg.value for an alloca is always indirect.
994 bool IsIndirect = isa<AllocaInst>(V) || Offset != 0;
997 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, IsIndirect, Val)) {
998 SDV = DAG.getDbgValue(Variable, Val.getNode(),
999 Val.getResNo(), IsIndirect,
1000 Offset, dl, DbgSDNodeOrder);
1001 DAG.AddDbgValue(SDV, Val.getNode(), false);
1004 DEBUG(dbgs() << "Dropping debug info for " << *DI << "\n");
1005 DanglingDebugInfoMap[V] = DanglingDebugInfo();
1009 /// getValue - Return an SDValue for the given Value.
1010 SDValue SelectionDAGBuilder::getValue(const Value *V) {
1011 // If we already have an SDValue for this value, use it. It's important
1012 // to do this first, so that we don't create a CopyFromReg if we already
1013 // have a regular SDValue.
1014 SDValue &N = NodeMap[V];
1015 if (N.getNode()) return N;
1017 // If there's a virtual register allocated and initialized for this
1019 DenseMap<const Value *, unsigned>::iterator It = FuncInfo.ValueMap.find(V);
1020 if (It != FuncInfo.ValueMap.end()) {
1021 unsigned InReg = It->second;
1022 RegsForValue RFV(*DAG.getContext(),
1023 *TM.getSubtargetImpl()->getTargetLowering(), InReg,
1025 SDValue Chain = DAG.getEntryNode();
1026 N = RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, nullptr, V);
1027 resolveDanglingDebugInfo(V, N);
1031 // Otherwise create a new SDValue and remember it.
1032 SDValue Val = getValueImpl(V);
1034 resolveDanglingDebugInfo(V, Val);
1038 /// getNonRegisterValue - Return an SDValue for the given Value, but
1039 /// don't look in FuncInfo.ValueMap for a virtual register.
1040 SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) {
1041 // If we already have an SDValue for this value, use it.
1042 SDValue &N = NodeMap[V];
1043 if (N.getNode()) return N;
1045 // Otherwise create a new SDValue and remember it.
1046 SDValue Val = getValueImpl(V);
1048 resolveDanglingDebugInfo(V, Val);
1052 /// getValueImpl - Helper function for getValue and getNonRegisterValue.
1053 /// Create an SDValue for the given value.
1054 SDValue SelectionDAGBuilder::getValueImpl(const Value *V) {
1055 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
1057 if (const Constant *C = dyn_cast<Constant>(V)) {
1058 EVT VT = TLI->getValueType(V->getType(), true);
1060 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C))
1061 return DAG.getConstant(*CI, VT);
1063 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
1064 return DAG.getGlobalAddress(GV, getCurSDLoc(), VT);
1066 if (isa<ConstantPointerNull>(C)) {
1067 unsigned AS = V->getType()->getPointerAddressSpace();
1068 return DAG.getConstant(0, TLI->getPointerTy(AS));
1071 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
1072 return DAG.getConstantFP(*CFP, VT);
1074 if (isa<UndefValue>(C) && !V->getType()->isAggregateType())
1075 return DAG.getUNDEF(VT);
1077 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
1078 visit(CE->getOpcode(), *CE);
1079 SDValue N1 = NodeMap[V];
1080 assert(N1.getNode() && "visit didn't populate the NodeMap!");
1084 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) {
1085 SmallVector<SDValue, 4> Constants;
1086 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end();
1088 SDNode *Val = getValue(*OI).getNode();
1089 // If the operand is an empty aggregate, there are no values.
1091 // Add each leaf value from the operand to the Constants list
1092 // to form a flattened list of all the values.
1093 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
1094 Constants.push_back(SDValue(Val, i));
1097 return DAG.getMergeValues(Constants, getCurSDLoc());
1100 if (const ConstantDataSequential *CDS =
1101 dyn_cast<ConstantDataSequential>(C)) {
1102 SmallVector<SDValue, 4> Ops;
1103 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1104 SDNode *Val = getValue(CDS->getElementAsConstant(i)).getNode();
1105 // Add each leaf value from the operand to the Constants list
1106 // to form a flattened list of all the values.
1107 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
1108 Ops.push_back(SDValue(Val, i));
1111 if (isa<ArrayType>(CDS->getType()))
1112 return DAG.getMergeValues(Ops, getCurSDLoc());
1113 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(),
1117 if (C->getType()->isStructTy() || C->getType()->isArrayTy()) {
1118 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
1119 "Unknown struct or array constant!");
1121 SmallVector<EVT, 4> ValueVTs;
1122 ComputeValueVTs(*TLI, C->getType(), ValueVTs);
1123 unsigned NumElts = ValueVTs.size();
1125 return SDValue(); // empty struct
1126 SmallVector<SDValue, 4> Constants(NumElts);
1127 for (unsigned i = 0; i != NumElts; ++i) {
1128 EVT EltVT = ValueVTs[i];
1129 if (isa<UndefValue>(C))
1130 Constants[i] = DAG.getUNDEF(EltVT);
1131 else if (EltVT.isFloatingPoint())
1132 Constants[i] = DAG.getConstantFP(0, EltVT);
1134 Constants[i] = DAG.getConstant(0, EltVT);
1137 return DAG.getMergeValues(Constants, getCurSDLoc());
1140 if (const BlockAddress *BA = dyn_cast<BlockAddress>(C))
1141 return DAG.getBlockAddress(BA, VT);
1143 VectorType *VecTy = cast<VectorType>(V->getType());
1144 unsigned NumElements = VecTy->getNumElements();
1146 // Now that we know the number and type of the elements, get that number of
1147 // elements into the Ops array based on what kind of constant it is.
1148 SmallVector<SDValue, 16> Ops;
1149 if (const ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
1150 for (unsigned i = 0; i != NumElements; ++i)
1151 Ops.push_back(getValue(CV->getOperand(i)));
1153 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!");
1154 EVT EltVT = TLI->getValueType(VecTy->getElementType());
1157 if (EltVT.isFloatingPoint())
1158 Op = DAG.getConstantFP(0, EltVT);
1160 Op = DAG.getConstant(0, EltVT);
1161 Ops.assign(NumElements, Op);
1164 // Create a BUILD_VECTOR node.
1165 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(), VT, Ops);
1168 // If this is a static alloca, generate it as the frameindex instead of
1170 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
1171 DenseMap<const AllocaInst*, int>::iterator SI =
1172 FuncInfo.StaticAllocaMap.find(AI);
1173 if (SI != FuncInfo.StaticAllocaMap.end())
1174 return DAG.getFrameIndex(SI->second, TLI->getPointerTy());
1177 // If this is an instruction which fast-isel has deferred, select it now.
1178 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
1179 unsigned InReg = FuncInfo.InitializeRegForValue(Inst);
1180 RegsForValue RFV(*DAG.getContext(), *TLI, InReg, Inst->getType());
1181 SDValue Chain = DAG.getEntryNode();
1182 return RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, nullptr, V);
1185 llvm_unreachable("Can't get register for value!");
1188 void SelectionDAGBuilder::visitRet(const ReturnInst &I) {
1189 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
1190 SDValue Chain = getControlRoot();
1191 SmallVector<ISD::OutputArg, 8> Outs;
1192 SmallVector<SDValue, 8> OutVals;
1194 if (!FuncInfo.CanLowerReturn) {
1195 unsigned DemoteReg = FuncInfo.DemoteRegister;
1196 const Function *F = I.getParent()->getParent();
1198 // Emit a store of the return value through the virtual register.
1199 // Leave Outs empty so that LowerReturn won't try to load return
1200 // registers the usual way.
1201 SmallVector<EVT, 1> PtrValueVTs;
1202 ComputeValueVTs(*TLI, PointerType::getUnqual(F->getReturnType()),
1205 SDValue RetPtr = DAG.getRegister(DemoteReg, PtrValueVTs[0]);
1206 SDValue RetOp = getValue(I.getOperand(0));
1208 SmallVector<EVT, 4> ValueVTs;
1209 SmallVector<uint64_t, 4> Offsets;
1210 ComputeValueVTs(*TLI, I.getOperand(0)->getType(), ValueVTs, &Offsets);
1211 unsigned NumValues = ValueVTs.size();
1213 SmallVector<SDValue, 4> Chains(NumValues);
1214 for (unsigned i = 0; i != NumValues; ++i) {
1215 SDValue Add = DAG.getNode(ISD::ADD, getCurSDLoc(),
1216 RetPtr.getValueType(), RetPtr,
1217 DAG.getIntPtrConstant(Offsets[i]));
1219 DAG.getStore(Chain, getCurSDLoc(),
1220 SDValue(RetOp.getNode(), RetOp.getResNo() + i),
1221 // FIXME: better loc info would be nice.
1222 Add, MachinePointerInfo(), false, false, 0);
1225 Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(),
1226 MVT::Other, Chains);
1227 } else if (I.getNumOperands() != 0) {
1228 SmallVector<EVT, 4> ValueVTs;
1229 ComputeValueVTs(*TLI, I.getOperand(0)->getType(), ValueVTs);
1230 unsigned NumValues = ValueVTs.size();
1232 SDValue RetOp = getValue(I.getOperand(0));
1233 for (unsigned j = 0, f = NumValues; j != f; ++j) {
1234 EVT VT = ValueVTs[j];
1236 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1238 const Function *F = I.getParent()->getParent();
1239 if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1241 ExtendKind = ISD::SIGN_EXTEND;
1242 else if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1244 ExtendKind = ISD::ZERO_EXTEND;
1246 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger())
1247 VT = TLI->getTypeForExtArgOrReturn(*DAG.getContext(), VT, ExtendKind);
1249 unsigned NumParts = TLI->getNumRegisters(*DAG.getContext(), VT);
1250 MVT PartVT = TLI->getRegisterType(*DAG.getContext(), VT);
1251 SmallVector<SDValue, 4> Parts(NumParts);
1252 getCopyToParts(DAG, getCurSDLoc(),
1253 SDValue(RetOp.getNode(), RetOp.getResNo() + j),
1254 &Parts[0], NumParts, PartVT, &I, ExtendKind);
1256 // 'inreg' on function refers to return value
1257 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1258 if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1262 // Propagate extension type if any
1263 if (ExtendKind == ISD::SIGN_EXTEND)
1265 else if (ExtendKind == ISD::ZERO_EXTEND)
1268 for (unsigned i = 0; i < NumParts; ++i) {
1269 Outs.push_back(ISD::OutputArg(Flags, Parts[i].getValueType(),
1270 VT, /*isfixed=*/true, 0, 0));
1271 OutVals.push_back(Parts[i]);
1277 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1278 CallingConv::ID CallConv =
1279 DAG.getMachineFunction().getFunction()->getCallingConv();
1280 Chain = TM.getSubtargetImpl()->getTargetLowering()->LowerReturn(
1281 Chain, CallConv, isVarArg, Outs, OutVals, getCurSDLoc(), DAG);
1283 // Verify that the target's LowerReturn behaved as expected.
1284 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
1285 "LowerReturn didn't return a valid chain!");
1287 // Update the DAG with the new chain value resulting from return lowering.
1291 /// CopyToExportRegsIfNeeded - If the given value has virtual registers
1292 /// created for it, emit nodes to copy the value into the virtual
1294 void SelectionDAGBuilder::CopyToExportRegsIfNeeded(const Value *V) {
1296 if (V->getType()->isEmptyTy())
1299 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
1300 if (VMI != FuncInfo.ValueMap.end()) {
1301 assert(!V->use_empty() && "Unused value assigned virtual registers!");
1302 CopyValueToVirtualRegister(V, VMI->second);
1306 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
1307 /// the current basic block, add it to ValueMap now so that we'll get a
1309 void SelectionDAGBuilder::ExportFromCurrentBlock(const Value *V) {
1310 // No need to export constants.
1311 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
1313 // Already exported?
1314 if (FuncInfo.isExportedInst(V)) return;
1316 unsigned Reg = FuncInfo.InitializeRegForValue(V);
1317 CopyValueToVirtualRegister(V, Reg);
1320 bool SelectionDAGBuilder::isExportableFromCurrentBlock(const Value *V,
1321 const BasicBlock *FromBB) {
1322 // The operands of the setcc have to be in this block. We don't know
1323 // how to export them from some other block.
1324 if (const Instruction *VI = dyn_cast<Instruction>(V)) {
1325 // Can export from current BB.
1326 if (VI->getParent() == FromBB)
1329 // Is already exported, noop.
1330 return FuncInfo.isExportedInst(V);
1333 // If this is an argument, we can export it if the BB is the entry block or
1334 // if it is already exported.
1335 if (isa<Argument>(V)) {
1336 if (FromBB == &FromBB->getParent()->getEntryBlock())
1339 // Otherwise, can only export this if it is already exported.
1340 return FuncInfo.isExportedInst(V);
1343 // Otherwise, constants can always be exported.
1347 /// Return branch probability calculated by BranchProbabilityInfo for IR blocks.
1348 uint32_t SelectionDAGBuilder::getEdgeWeight(const MachineBasicBlock *Src,
1349 const MachineBasicBlock *Dst) const {
1350 BranchProbabilityInfo *BPI = FuncInfo.BPI;
1353 const BasicBlock *SrcBB = Src->getBasicBlock();
1354 const BasicBlock *DstBB = Dst->getBasicBlock();
1355 return BPI->getEdgeWeight(SrcBB, DstBB);
1358 void SelectionDAGBuilder::
1359 addSuccessorWithWeight(MachineBasicBlock *Src, MachineBasicBlock *Dst,
1360 uint32_t Weight /* = 0 */) {
1362 Weight = getEdgeWeight(Src, Dst);
1363 Src->addSuccessor(Dst, Weight);
1367 static bool InBlock(const Value *V, const BasicBlock *BB) {
1368 if (const Instruction *I = dyn_cast<Instruction>(V))
1369 return I->getParent() == BB;
1373 /// EmitBranchForMergedCondition - Helper method for FindMergedConditions.
1374 /// This function emits a branch and is used at the leaves of an OR or an
1375 /// AND operator tree.
1378 SelectionDAGBuilder::EmitBranchForMergedCondition(const Value *Cond,
1379 MachineBasicBlock *TBB,
1380 MachineBasicBlock *FBB,
1381 MachineBasicBlock *CurBB,
1382 MachineBasicBlock *SwitchBB,
1385 const BasicBlock *BB = CurBB->getBasicBlock();
1387 // If the leaf of the tree is a comparison, merge the condition into
1389 if (const CmpInst *BOp = dyn_cast<CmpInst>(Cond)) {
1390 // The operands of the cmp have to be in this block. We don't know
1391 // how to export them from some other block. If this is the first block
1392 // of the sequence, no exporting is needed.
1393 if (CurBB == SwitchBB ||
1394 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
1395 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) {
1396 ISD::CondCode Condition;
1397 if (const ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
1398 Condition = getICmpCondCode(IC->getPredicate());
1399 } else if (const FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) {
1400 Condition = getFCmpCondCode(FC->getPredicate());
1401 if (TM.Options.NoNaNsFPMath)
1402 Condition = getFCmpCodeWithoutNaN(Condition);
1404 Condition = ISD::SETEQ; // silence warning.
1405 llvm_unreachable("Unknown compare instruction");
1408 CaseBlock CB(Condition, BOp->getOperand(0), BOp->getOperand(1), nullptr,
1409 TBB, FBB, CurBB, TWeight, FWeight);
1410 SwitchCases.push_back(CB);
1415 // Create a CaseBlock record representing this branch.
1416 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()),
1417 nullptr, TBB, FBB, CurBB, TWeight, FWeight);
1418 SwitchCases.push_back(CB);
1421 /// Scale down both weights to fit into uint32_t.
1422 static void ScaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
1423 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
1424 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
1425 NewTrue = NewTrue / Scale;
1426 NewFalse = NewFalse / Scale;
1429 /// FindMergedConditions - If Cond is an expression like
1430 void SelectionDAGBuilder::FindMergedConditions(const Value *Cond,
1431 MachineBasicBlock *TBB,
1432 MachineBasicBlock *FBB,
1433 MachineBasicBlock *CurBB,
1434 MachineBasicBlock *SwitchBB,
1435 unsigned Opc, uint32_t TWeight,
1437 // If this node is not part of the or/and tree, emit it as a branch.
1438 const Instruction *BOp = dyn_cast<Instruction>(Cond);
1439 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
1440 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
1441 BOp->getParent() != CurBB->getBasicBlock() ||
1442 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
1443 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
1444 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB,
1449 // Create TmpBB after CurBB.
1450 MachineFunction::iterator BBI = CurBB;
1451 MachineFunction &MF = DAG.getMachineFunction();
1452 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock());
1453 CurBB->getParent()->insert(++BBI, TmpBB);
1455 if (Opc == Instruction::Or) {
1456 // Codegen X | Y as:
1465 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
1466 // The requirement is that
1467 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
1468 // = TrueProb for orignal BB.
1469 // Assuming the orignal weights are A and B, one choice is to set BB1's
1470 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
1472 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
1473 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
1474 // TmpBB, but the math is more complicated.
1476 uint64_t NewTrueWeight = TWeight;
1477 uint64_t NewFalseWeight = (uint64_t)TWeight + 2 * (uint64_t)FWeight;
1478 ScaleWeights(NewTrueWeight, NewFalseWeight);
1479 // Emit the LHS condition.
1480 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, SwitchBB, Opc,
1481 NewTrueWeight, NewFalseWeight);
1483 NewTrueWeight = TWeight;
1484 NewFalseWeight = 2 * (uint64_t)FWeight;
1485 ScaleWeights(NewTrueWeight, NewFalseWeight);
1486 // Emit the RHS condition into TmpBB.
1487 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc,
1488 NewTrueWeight, NewFalseWeight);
1490 assert(Opc == Instruction::And && "Unknown merge op!");
1491 // Codegen X & Y as:
1499 // This requires creation of TmpBB after CurBB.
1501 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
1502 // The requirement is that
1503 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
1504 // = FalseProb for orignal BB.
1505 // Assuming the orignal weights are A and B, one choice is to set BB1's
1506 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
1508 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
1510 uint64_t NewTrueWeight = 2 * (uint64_t)TWeight + (uint64_t)FWeight;
1511 uint64_t NewFalseWeight = FWeight;
1512 ScaleWeights(NewTrueWeight, NewFalseWeight);
1513 // Emit the LHS condition.
1514 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, SwitchBB, Opc,
1515 NewTrueWeight, NewFalseWeight);
1517 NewTrueWeight = 2 * (uint64_t)TWeight;
1518 NewFalseWeight = FWeight;
1519 ScaleWeights(NewTrueWeight, NewFalseWeight);
1520 // Emit the RHS condition into TmpBB.
1521 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc,
1522 NewTrueWeight, NewFalseWeight);
1526 /// If the set of cases should be emitted as a series of branches, return true.
1527 /// If we should emit this as a bunch of and/or'd together conditions, return
1530 SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases) {
1531 if (Cases.size() != 2) return true;
1533 // If this is two comparisons of the same values or'd or and'd together, they
1534 // will get folded into a single comparison, so don't emit two blocks.
1535 if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
1536 Cases[0].CmpRHS == Cases[1].CmpRHS) ||
1537 (Cases[0].CmpRHS == Cases[1].CmpLHS &&
1538 Cases[0].CmpLHS == Cases[1].CmpRHS)) {
1542 // Handle: (X != null) | (Y != null) --> (X|Y) != 0
1543 // Handle: (X == null) & (Y == null) --> (X|Y) == 0
1544 if (Cases[0].CmpRHS == Cases[1].CmpRHS &&
1545 Cases[0].CC == Cases[1].CC &&
1546 isa<Constant>(Cases[0].CmpRHS) &&
1547 cast<Constant>(Cases[0].CmpRHS)->isNullValue()) {
1548 if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB)
1550 if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB)
1557 void SelectionDAGBuilder::visitBr(const BranchInst &I) {
1558 MachineBasicBlock *BrMBB = FuncInfo.MBB;
1560 // Update machine-CFG edges.
1561 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
1563 // Figure out which block is immediately after the current one.
1564 MachineBasicBlock *NextBlock = nullptr;
1565 MachineFunction::iterator BBI = BrMBB;
1566 if (++BBI != FuncInfo.MF->end())
1569 if (I.isUnconditional()) {
1570 // Update machine-CFG edges.
1571 BrMBB->addSuccessor(Succ0MBB);
1573 // If this is not a fall-through branch or optimizations are switched off,
1575 if (Succ0MBB != NextBlock || TM.getOptLevel() == CodeGenOpt::None)
1576 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(),
1577 MVT::Other, getControlRoot(),
1578 DAG.getBasicBlock(Succ0MBB)));
1583 // If this condition is one of the special cases we handle, do special stuff
1585 const Value *CondVal = I.getCondition();
1586 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
1588 // If this is a series of conditions that are or'd or and'd together, emit
1589 // this as a sequence of branches instead of setcc's with and/or operations.
1590 // As long as jumps are not expensive, this should improve performance.
1591 // For example, instead of something like:
1604 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
1605 if (!TM.getSubtargetImpl()->getTargetLowering()->isJumpExpensive() &&
1606 BOp->hasOneUse() && (BOp->getOpcode() == Instruction::And ||
1607 BOp->getOpcode() == Instruction::Or)) {
1608 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB,
1609 BOp->getOpcode(), getEdgeWeight(BrMBB, Succ0MBB),
1610 getEdgeWeight(BrMBB, Succ1MBB));
1611 // If the compares in later blocks need to use values not currently
1612 // exported from this block, export them now. This block should always
1613 // be the first entry.
1614 assert(SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!");
1616 // Allow some cases to be rejected.
1617 if (ShouldEmitAsBranches(SwitchCases)) {
1618 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1619 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1620 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1623 // Emit the branch for this block.
1624 visitSwitchCase(SwitchCases[0], BrMBB);
1625 SwitchCases.erase(SwitchCases.begin());
1629 // Okay, we decided not to do this, remove any inserted MBB's and clear
1631 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
1632 FuncInfo.MF->erase(SwitchCases[i].ThisBB);
1634 SwitchCases.clear();
1638 // Create a CaseBlock record representing this branch.
1639 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()),
1640 nullptr, Succ0MBB, Succ1MBB, BrMBB);
1642 // Use visitSwitchCase to actually insert the fast branch sequence for this
1644 visitSwitchCase(CB, BrMBB);
1647 /// visitSwitchCase - Emits the necessary code to represent a single node in
1648 /// the binary search tree resulting from lowering a switch instruction.
1649 void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB,
1650 MachineBasicBlock *SwitchBB) {
1652 SDValue CondLHS = getValue(CB.CmpLHS);
1653 SDLoc dl = getCurSDLoc();
1655 // Build the setcc now.
1657 // Fold "(X == true)" to X and "(X == false)" to !X to
1658 // handle common cases produced by branch lowering.
1659 if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) &&
1660 CB.CC == ISD::SETEQ)
1662 else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) &&
1663 CB.CC == ISD::SETEQ) {
1664 SDValue True = DAG.getConstant(1, CondLHS.getValueType());
1665 Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True);
1667 Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1669 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now");
1671 const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue();
1672 const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue();
1674 SDValue CmpOp = getValue(CB.CmpMHS);
1675 EVT VT = CmpOp.getValueType();
1677 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) {
1678 Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, VT),
1681 SDValue SUB = DAG.getNode(ISD::SUB, dl,
1682 VT, CmpOp, DAG.getConstant(Low, VT));
1683 Cond = DAG.getSetCC(dl, MVT::i1, SUB,
1684 DAG.getConstant(High-Low, VT), ISD::SETULE);
1688 // Update successor info
1689 addSuccessorWithWeight(SwitchBB, CB.TrueBB, CB.TrueWeight);
1690 // TrueBB and FalseBB are always different unless the incoming IR is
1691 // degenerate. This only happens when running llc on weird IR.
1692 if (CB.TrueBB != CB.FalseBB)
1693 addSuccessorWithWeight(SwitchBB, CB.FalseBB, CB.FalseWeight);
1695 // Set NextBlock to be the MBB immediately after the current one, if any.
1696 // This is used to avoid emitting unnecessary branches to the next block.
1697 MachineBasicBlock *NextBlock = nullptr;
1698 MachineFunction::iterator BBI = SwitchBB;
1699 if (++BBI != FuncInfo.MF->end())
1702 // If the lhs block is the next block, invert the condition so that we can
1703 // fall through to the lhs instead of the rhs block.
1704 if (CB.TrueBB == NextBlock) {
1705 std::swap(CB.TrueBB, CB.FalseBB);
1706 SDValue True = DAG.getConstant(1, Cond.getValueType());
1707 Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True);
1710 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1711 MVT::Other, getControlRoot(), Cond,
1712 DAG.getBasicBlock(CB.TrueBB));
1714 // Insert the false branch. Do this even if it's a fall through branch,
1715 // this makes it easier to do DAG optimizations which require inverting
1716 // the branch condition.
1717 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1718 DAG.getBasicBlock(CB.FalseBB));
1720 DAG.setRoot(BrCond);
1723 /// visitJumpTable - Emit JumpTable node in the current MBB
1724 void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) {
1725 // Emit the code for the jump table
1726 assert(JT.Reg != -1U && "Should lower JT Header first!");
1727 EVT PTy = TM.getSubtargetImpl()->getTargetLowering()->getPointerTy();
1728 SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurSDLoc(),
1730 SDValue Table = DAG.getJumpTable(JT.JTI, PTy);
1731 SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurSDLoc(),
1732 MVT::Other, Index.getValue(1),
1734 DAG.setRoot(BrJumpTable);
1737 /// visitJumpTableHeader - This function emits necessary code to produce index
1738 /// in the JumpTable from switch case.
1739 void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT,
1740 JumpTableHeader &JTH,
1741 MachineBasicBlock *SwitchBB) {
1742 // Subtract the lowest switch case value from the value being switched on and
1743 // conditional branch to default mbb if the result is greater than the
1744 // difference between smallest and largest cases.
1745 SDValue SwitchOp = getValue(JTH.SValue);
1746 EVT VT = SwitchOp.getValueType();
1747 SDValue Sub = DAG.getNode(ISD::SUB, getCurSDLoc(), VT, SwitchOp,
1748 DAG.getConstant(JTH.First, VT));
1750 // The SDNode we just created, which holds the value being switched on minus
1751 // the smallest case value, needs to be copied to a virtual register so it
1752 // can be used as an index into the jump table in a subsequent basic block.
1753 // This value may be smaller or larger than the target's pointer type, and
1754 // therefore require extension or truncating.
1755 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
1756 SwitchOp = DAG.getZExtOrTrunc(Sub, getCurSDLoc(), TLI->getPointerTy());
1758 unsigned JumpTableReg = FuncInfo.CreateReg(TLI->getPointerTy());
1759 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurSDLoc(),
1760 JumpTableReg, SwitchOp);
1761 JT.Reg = JumpTableReg;
1763 // Emit the range check for the jump table, and branch to the default block
1764 // for the switch statement if the value being switched on exceeds the largest
1765 // case in the switch.
1766 SDValue CMP = DAG.getSetCC(getCurSDLoc(),
1767 TLI->getSetCCResultType(*DAG.getContext(),
1768 Sub.getValueType()),
1770 DAG.getConstant(JTH.Last - JTH.First,VT),
1773 // Set NextBlock to be the MBB immediately after the current one, if any.
1774 // This is used to avoid emitting unnecessary branches to the next block.
1775 MachineBasicBlock *NextBlock = nullptr;
1776 MachineFunction::iterator BBI = SwitchBB;
1778 if (++BBI != FuncInfo.MF->end())
1781 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurSDLoc(),
1782 MVT::Other, CopyTo, CMP,
1783 DAG.getBasicBlock(JT.Default));
1785 if (JT.MBB != NextBlock)
1786 BrCond = DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, BrCond,
1787 DAG.getBasicBlock(JT.MBB));
1789 DAG.setRoot(BrCond);
1792 /// Codegen a new tail for a stack protector check ParentMBB which has had its
1793 /// tail spliced into a stack protector check success bb.
1795 /// For a high level explanation of how this fits into the stack protector
1796 /// generation see the comment on the declaration of class
1797 /// StackProtectorDescriptor.
1798 void SelectionDAGBuilder::visitSPDescriptorParent(StackProtectorDescriptor &SPD,
1799 MachineBasicBlock *ParentBB) {
1801 // First create the loads to the guard/stack slot for the comparison.
1802 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
1803 EVT PtrTy = TLI->getPointerTy();
1805 MachineFrameInfo *MFI = ParentBB->getParent()->getFrameInfo();
1806 int FI = MFI->getStackProtectorIndex();
1808 const Value *IRGuard = SPD.getGuard();
1809 SDValue GuardPtr = getValue(IRGuard);
1810 SDValue StackSlotPtr = DAG.getFrameIndex(FI, PtrTy);
1813 TLI->getDataLayout()->getPrefTypeAlignment(IRGuard->getType());
1817 // If GuardReg is set and useLoadStackGuardNode returns true, retrieve the
1818 // guard value from the virtual register holding the value. Otherwise, emit a
1819 // volatile load to retrieve the stack guard value.
1820 unsigned GuardReg = SPD.getGuardReg();
1822 if (GuardReg && TLI->useLoadStackGuardNode())
1823 Guard = DAG.getCopyFromReg(DAG.getEntryNode(), getCurSDLoc(), GuardReg,
1826 Guard = DAG.getLoad(PtrTy, getCurSDLoc(), DAG.getEntryNode(),
1827 GuardPtr, MachinePointerInfo(IRGuard, 0),
1828 true, false, false, Align);
1830 SDValue StackSlot = DAG.getLoad(PtrTy, getCurSDLoc(), DAG.getEntryNode(),
1832 MachinePointerInfo::getFixedStack(FI),
1833 true, false, false, Align);
1835 // Perform the comparison via a subtract/getsetcc.
1836 EVT VT = Guard.getValueType();
1837 SDValue Sub = DAG.getNode(ISD::SUB, getCurSDLoc(), VT, Guard, StackSlot);
1839 SDValue Cmp = DAG.getSetCC(getCurSDLoc(),
1840 TLI->getSetCCResultType(*DAG.getContext(),
1841 Sub.getValueType()),
1842 Sub, DAG.getConstant(0, VT),
1845 // If the sub is not 0, then we know the guard/stackslot do not equal, so
1846 // branch to failure MBB.
1847 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurSDLoc(),
1848 MVT::Other, StackSlot.getOperand(0),
1849 Cmp, DAG.getBasicBlock(SPD.getFailureMBB()));
1850 // Otherwise branch to success MBB.
1851 SDValue Br = DAG.getNode(ISD::BR, getCurSDLoc(),
1853 DAG.getBasicBlock(SPD.getSuccessMBB()));
1858 /// Codegen the failure basic block for a stack protector check.
1860 /// A failure stack protector machine basic block consists simply of a call to
1861 /// __stack_chk_fail().
1863 /// For a high level explanation of how this fits into the stack protector
1864 /// generation see the comment on the declaration of class
1865 /// StackProtectorDescriptor.
1867 SelectionDAGBuilder::visitSPDescriptorFailure(StackProtectorDescriptor &SPD) {
1868 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
1869 SDValue Chain = TLI->makeLibCall(DAG, RTLIB::STACKPROTECTOR_CHECK_FAIL,
1870 MVT::isVoid, nullptr, 0, false,
1871 getCurSDLoc(), false, false).second;
1875 /// visitBitTestHeader - This function emits necessary code to produce value
1876 /// suitable for "bit tests"
1877 void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B,
1878 MachineBasicBlock *SwitchBB) {
1879 // Subtract the minimum value
1880 SDValue SwitchOp = getValue(B.SValue);
1881 EVT VT = SwitchOp.getValueType();
1882 SDValue Sub = DAG.getNode(ISD::SUB, getCurSDLoc(), VT, SwitchOp,
1883 DAG.getConstant(B.First, VT));
1886 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
1887 SDValue RangeCmp = DAG.getSetCC(getCurSDLoc(),
1888 TLI->getSetCCResultType(*DAG.getContext(),
1889 Sub.getValueType()),
1890 Sub, DAG.getConstant(B.Range, VT),
1893 // Determine the type of the test operands.
1894 bool UsePtrType = false;
1895 if (!TLI->isTypeLegal(VT))
1898 for (unsigned i = 0, e = B.Cases.size(); i != e; ++i)
1899 if (!isUIntN(VT.getSizeInBits(), B.Cases[i].Mask)) {
1900 // Switch table case range are encoded into series of masks.
1901 // Just use pointer type, it's guaranteed to fit.
1907 VT = TLI->getPointerTy();
1908 Sub = DAG.getZExtOrTrunc(Sub, getCurSDLoc(), VT);
1911 B.RegVT = VT.getSimpleVT();
1912 B.Reg = FuncInfo.CreateReg(B.RegVT);
1913 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurSDLoc(),
1916 // Set NextBlock to be the MBB immediately after the current one, if any.
1917 // This is used to avoid emitting unnecessary branches to the next block.
1918 MachineBasicBlock *NextBlock = nullptr;
1919 MachineFunction::iterator BBI = SwitchBB;
1920 if (++BBI != FuncInfo.MF->end())
1923 MachineBasicBlock* MBB = B.Cases[0].ThisBB;
1925 addSuccessorWithWeight(SwitchBB, B.Default);
1926 addSuccessorWithWeight(SwitchBB, MBB);
1928 SDValue BrRange = DAG.getNode(ISD::BRCOND, getCurSDLoc(),
1929 MVT::Other, CopyTo, RangeCmp,
1930 DAG.getBasicBlock(B.Default));
1932 if (MBB != NextBlock)
1933 BrRange = DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, CopyTo,
1934 DAG.getBasicBlock(MBB));
1936 DAG.setRoot(BrRange);
1939 /// visitBitTestCase - this function produces one "bit test"
1940 void SelectionDAGBuilder::visitBitTestCase(BitTestBlock &BB,
1941 MachineBasicBlock* NextMBB,
1942 uint32_t BranchWeightToNext,
1945 MachineBasicBlock *SwitchBB) {
1947 SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurSDLoc(),
1950 unsigned PopCount = CountPopulation_64(B.Mask);
1951 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
1952 if (PopCount == 1) {
1953 // Testing for a single bit; just compare the shift count with what it
1954 // would need to be to shift a 1 bit in that position.
1955 Cmp = DAG.getSetCC(getCurSDLoc(),
1956 TLI->getSetCCResultType(*DAG.getContext(), VT),
1958 DAG.getConstant(countTrailingZeros(B.Mask), VT),
1960 } else if (PopCount == BB.Range) {
1961 // There is only one zero bit in the range, test for it directly.
1962 Cmp = DAG.getSetCC(getCurSDLoc(),
1963 TLI->getSetCCResultType(*DAG.getContext(), VT),
1965 DAG.getConstant(CountTrailingOnes_64(B.Mask), VT),
1968 // Make desired shift
1969 SDValue SwitchVal = DAG.getNode(ISD::SHL, getCurSDLoc(), VT,
1970 DAG.getConstant(1, VT), ShiftOp);
1972 // Emit bit tests and jumps
1973 SDValue AndOp = DAG.getNode(ISD::AND, getCurSDLoc(),
1974 VT, SwitchVal, DAG.getConstant(B.Mask, VT));
1975 Cmp = DAG.getSetCC(getCurSDLoc(),
1976 TLI->getSetCCResultType(*DAG.getContext(), VT),
1977 AndOp, DAG.getConstant(0, VT),
1981 // The branch weight from SwitchBB to B.TargetBB is B.ExtraWeight.
1982 addSuccessorWithWeight(SwitchBB, B.TargetBB, B.ExtraWeight);
1983 // The branch weight from SwitchBB to NextMBB is BranchWeightToNext.
1984 addSuccessorWithWeight(SwitchBB, NextMBB, BranchWeightToNext);
1986 SDValue BrAnd = DAG.getNode(ISD::BRCOND, getCurSDLoc(),
1987 MVT::Other, getControlRoot(),
1988 Cmp, DAG.getBasicBlock(B.TargetBB));
1990 // Set NextBlock to be the MBB immediately after the current one, if any.
1991 // This is used to avoid emitting unnecessary branches to the next block.
1992 MachineBasicBlock *NextBlock = nullptr;
1993 MachineFunction::iterator BBI = SwitchBB;
1994 if (++BBI != FuncInfo.MF->end())
1997 if (NextMBB != NextBlock)
1998 BrAnd = DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, BrAnd,
1999 DAG.getBasicBlock(NextMBB));
2004 void SelectionDAGBuilder::visitInvoke(const InvokeInst &I) {
2005 MachineBasicBlock *InvokeMBB = FuncInfo.MBB;
2007 // Retrieve successors.
2008 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
2009 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)];
2011 const Value *Callee(I.getCalledValue());
2012 const Function *Fn = dyn_cast<Function>(Callee);
2013 if (isa<InlineAsm>(Callee))
2015 else if (Fn && Fn->isIntrinsic()) {
2016 assert(Fn->getIntrinsicID() == Intrinsic::donothing);
2017 // Ignore invokes to @llvm.donothing: jump directly to the next BB.
2019 LowerCallTo(&I, getValue(Callee), false, LandingPad);
2021 // If the value of the invoke is used outside of its defining block, make it
2022 // available as a virtual register.
2023 CopyToExportRegsIfNeeded(&I);
2025 // Update successor info
2026 addSuccessorWithWeight(InvokeMBB, Return);
2027 addSuccessorWithWeight(InvokeMBB, LandingPad);
2029 // Drop into normal successor.
2030 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(),
2031 MVT::Other, getControlRoot(),
2032 DAG.getBasicBlock(Return)));
2035 void SelectionDAGBuilder::visitResume(const ResumeInst &RI) {
2036 llvm_unreachable("SelectionDAGBuilder shouldn't visit resume instructions!");
2039 void SelectionDAGBuilder::visitLandingPad(const LandingPadInst &LP) {
2040 assert(FuncInfo.MBB->isLandingPad() &&
2041 "Call to landingpad not in landing pad!");
2043 MachineBasicBlock *MBB = FuncInfo.MBB;
2044 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
2045 AddLandingPadInfo(LP, MMI, MBB);
2047 // If there aren't registers to copy the values into (e.g., during SjLj
2048 // exceptions), then don't bother to create these DAG nodes.
2049 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
2050 if (TLI->getExceptionPointerRegister() == 0 &&
2051 TLI->getExceptionSelectorRegister() == 0)
2054 SmallVector<EVT, 2> ValueVTs;
2055 ComputeValueVTs(*TLI, LP.getType(), ValueVTs);
2056 assert(ValueVTs.size() == 2 && "Only two-valued landingpads are supported");
2058 // Get the two live-in registers as SDValues. The physregs have already been
2059 // copied into virtual registers.
2061 Ops[0] = DAG.getZExtOrTrunc(
2062 DAG.getCopyFromReg(DAG.getEntryNode(), getCurSDLoc(),
2063 FuncInfo.ExceptionPointerVirtReg, TLI->getPointerTy()),
2064 getCurSDLoc(), ValueVTs[0]);
2065 Ops[1] = DAG.getZExtOrTrunc(
2066 DAG.getCopyFromReg(DAG.getEntryNode(), getCurSDLoc(),
2067 FuncInfo.ExceptionSelectorVirtReg, TLI->getPointerTy()),
2068 getCurSDLoc(), ValueVTs[1]);
2071 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
2072 DAG.getVTList(ValueVTs), Ops);
2076 /// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for
2077 /// small case ranges).
2078 bool SelectionDAGBuilder::handleSmallSwitchRange(CaseRec& CR,
2079 CaseRecVector& WorkList,
2081 MachineBasicBlock *Default,
2082 MachineBasicBlock *SwitchBB) {
2083 // Size is the number of Cases represented by this range.
2084 size_t Size = CR.Range.second - CR.Range.first;
2088 // Get the MachineFunction which holds the current MBB. This is used when
2089 // inserting any additional MBBs necessary to represent the switch.
2090 MachineFunction *CurMF = FuncInfo.MF;
2092 // Figure out which block is immediately after the current one.
2093 MachineBasicBlock *NextBlock = nullptr;
2094 MachineFunction::iterator BBI = CR.CaseBB;
2096 if (++BBI != FuncInfo.MF->end())
2099 BranchProbabilityInfo *BPI = FuncInfo.BPI;
2100 // If any two of the cases has the same destination, and if one value
2101 // is the same as the other, but has one bit unset that the other has set,
2102 // use bit manipulation to do two compares at once. For example:
2103 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
2104 // TODO: This could be extended to merge any 2 cases in switches with 3 cases.
2105 // TODO: Handle cases where CR.CaseBB != SwitchBB.
2106 if (Size == 2 && CR.CaseBB == SwitchBB) {
2107 Case &Small = *CR.Range.first;
2108 Case &Big = *(CR.Range.second-1);
2110 if (Small.Low == Small.High && Big.Low == Big.High && Small.BB == Big.BB) {
2111 const APInt& SmallValue = cast<ConstantInt>(Small.Low)->getValue();
2112 const APInt& BigValue = cast<ConstantInt>(Big.Low)->getValue();
2114 // Check that there is only one bit different.
2115 if (BigValue.countPopulation() == SmallValue.countPopulation() + 1 &&
2116 (SmallValue | BigValue) == BigValue) {
2117 // Isolate the common bit.
2118 APInt CommonBit = BigValue & ~SmallValue;
2119 assert((SmallValue | CommonBit) == BigValue &&
2120 CommonBit.countPopulation() == 1 && "Not a common bit?");
2122 SDValue CondLHS = getValue(SV);
2123 EVT VT = CondLHS.getValueType();
2124 SDLoc DL = getCurSDLoc();
2126 SDValue Or = DAG.getNode(ISD::OR, DL, VT, CondLHS,
2127 DAG.getConstant(CommonBit, VT));
2128 SDValue Cond = DAG.getSetCC(DL, MVT::i1,
2129 Or, DAG.getConstant(BigValue, VT),
2132 // Update successor info.
2133 // Both Small and Big will jump to Small.BB, so we sum up the weights.
2134 addSuccessorWithWeight(SwitchBB, Small.BB,
2135 Small.ExtraWeight + Big.ExtraWeight);
2136 addSuccessorWithWeight(SwitchBB, Default,
2137 // The default destination is the first successor in IR.
2138 BPI ? BPI->getEdgeWeight(SwitchBB->getBasicBlock(), (unsigned)0) : 0);
2140 // Insert the true branch.
2141 SDValue BrCond = DAG.getNode(ISD::BRCOND, DL, MVT::Other,
2142 getControlRoot(), Cond,
2143 DAG.getBasicBlock(Small.BB));
2145 // Insert the false branch.
2146 BrCond = DAG.getNode(ISD::BR, DL, MVT::Other, BrCond,
2147 DAG.getBasicBlock(Default));
2149 DAG.setRoot(BrCond);
2155 // Order cases by weight so the most likely case will be checked first.
2156 uint32_t UnhandledWeights = 0;
2158 for (CaseItr I = CR.Range.first, IE = CR.Range.second; I != IE; ++I) {
2159 uint32_t IWeight = I->ExtraWeight;
2160 UnhandledWeights += IWeight;
2161 for (CaseItr J = CR.Range.first; J < I; ++J) {
2162 uint32_t JWeight = J->ExtraWeight;
2163 if (IWeight > JWeight)
2168 // Rearrange the case blocks so that the last one falls through if possible.
2169 Case &BackCase = *(CR.Range.second-1);
2171 NextBlock && Default != NextBlock && BackCase.BB != NextBlock) {
2172 // The last case block won't fall through into 'NextBlock' if we emit the
2173 // branches in this order. See if rearranging a case value would help.
2174 // We start at the bottom as it's the case with the least weight.
2175 for (Case *I = &*(CR.Range.second-2), *E = &*CR.Range.first-1; I != E; --I)
2176 if (I->BB == NextBlock) {
2177 std::swap(*I, BackCase);
2182 // Create a CaseBlock record representing a conditional branch to
2183 // the Case's target mbb if the value being switched on SV is equal
2185 MachineBasicBlock *CurBlock = CR.CaseBB;
2186 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
2187 MachineBasicBlock *FallThrough;
2189 FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock());
2190 CurMF->insert(BBI, FallThrough);
2192 // Put SV in a virtual register to make it available from the new blocks.
2193 ExportFromCurrentBlock(SV);
2195 // If the last case doesn't match, go to the default block.
2196 FallThrough = Default;
2199 const Value *RHS, *LHS, *MHS;
2201 if (I->High == I->Low) {
2202 // This is just small small case range :) containing exactly 1 case
2204 LHS = SV; RHS = I->High; MHS = nullptr;
2207 LHS = I->Low; MHS = SV; RHS = I->High;
2210 // The false weight should be sum of all un-handled cases.
2211 UnhandledWeights -= I->ExtraWeight;
2212 CaseBlock CB(CC, LHS, RHS, MHS, /* truebb */ I->BB, /* falsebb */ FallThrough,
2214 /* trueweight */ I->ExtraWeight,
2215 /* falseweight */ UnhandledWeights);
2217 // If emitting the first comparison, just call visitSwitchCase to emit the
2218 // code into the current block. Otherwise, push the CaseBlock onto the
2219 // vector to be later processed by SDISel, and insert the node's MBB
2220 // before the next MBB.
2221 if (CurBlock == SwitchBB)
2222 visitSwitchCase(CB, SwitchBB);
2224 SwitchCases.push_back(CB);
2226 CurBlock = FallThrough;
2232 static inline bool areJTsAllowed(const TargetLowering &TLI) {
2233 return TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
2234 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other);
2237 static APInt ComputeRange(const APInt &First, const APInt &Last) {
2238 uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1;
2239 APInt LastExt = Last.sext(BitWidth), FirstExt = First.sext(BitWidth);
2240 return (LastExt - FirstExt + 1ULL);
2243 /// handleJTSwitchCase - Emit jumptable for current switch case range
2244 bool SelectionDAGBuilder::handleJTSwitchCase(CaseRec &CR,
2245 CaseRecVector &WorkList,
2247 MachineBasicBlock *Default,
2248 MachineBasicBlock *SwitchBB) {
2249 Case& FrontCase = *CR.Range.first;
2250 Case& BackCase = *(CR.Range.second-1);
2252 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
2253 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
2255 APInt TSize(First.getBitWidth(), 0);
2256 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I)
2259 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
2260 if (!areJTsAllowed(*TLI) || TSize.ult(TLI->getMinimumJumpTableEntries()))
2263 APInt Range = ComputeRange(First, Last);
2264 // The density is TSize / Range. Require at least 40%.
2265 // It should not be possible for IntTSize to saturate for sane code, but make
2266 // sure we handle Range saturation correctly.
2267 uint64_t IntRange = Range.getLimitedValue(UINT64_MAX/10);
2268 uint64_t IntTSize = TSize.getLimitedValue(UINT64_MAX/10);
2269 if (IntTSize * 10 < IntRange * 4)
2272 DEBUG(dbgs() << "Lowering jump table\n"
2273 << "First entry: " << First << ". Last entry: " << Last << '\n'
2274 << "Range: " << Range << ". Size: " << TSize << ".\n\n");
2276 // Get the MachineFunction which holds the current MBB. This is used when
2277 // inserting any additional MBBs necessary to represent the switch.
2278 MachineFunction *CurMF = FuncInfo.MF;
2280 // Figure out which block is immediately after the current one.
2281 MachineFunction::iterator BBI = CR.CaseBB;
2284 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2286 // Create a new basic block to hold the code for loading the address
2287 // of the jump table, and jumping to it. Update successor information;
2288 // we will either branch to the default case for the switch, or the jump
2290 MachineBasicBlock *JumpTableBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2291 CurMF->insert(BBI, JumpTableBB);
2293 addSuccessorWithWeight(CR.CaseBB, Default);
2294 addSuccessorWithWeight(CR.CaseBB, JumpTableBB);
2296 // Build a vector of destination BBs, corresponding to each target
2297 // of the jump table. If the value of the jump table slot corresponds to
2298 // a case statement, push the case's BB onto the vector, otherwise, push
2300 std::vector<MachineBasicBlock*> DestBBs;
2302 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) {
2303 const APInt &Low = cast<ConstantInt>(I->Low)->getValue();
2304 const APInt &High = cast<ConstantInt>(I->High)->getValue();
2306 if (Low.sle(TEI) && TEI.sle(High)) {
2307 DestBBs.push_back(I->BB);
2311 DestBBs.push_back(Default);
2315 // Calculate weight for each unique destination in CR.
2316 DenseMap<MachineBasicBlock*, uint32_t> DestWeights;
2318 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
2319 DenseMap<MachineBasicBlock*, uint32_t>::iterator Itr =
2320 DestWeights.find(I->BB);
2321 if (Itr != DestWeights.end())
2322 Itr->second += I->ExtraWeight;
2324 DestWeights[I->BB] = I->ExtraWeight;
2327 // Update successor info. Add one edge to each unique successor.
2328 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs());
2329 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(),
2330 E = DestBBs.end(); I != E; ++I) {
2331 if (!SuccsHandled[(*I)->getNumber()]) {
2332 SuccsHandled[(*I)->getNumber()] = true;
2333 DenseMap<MachineBasicBlock*, uint32_t>::iterator Itr =
2334 DestWeights.find(*I);
2335 addSuccessorWithWeight(JumpTableBB, *I,
2336 Itr != DestWeights.end() ? Itr->second : 0);
2340 // Create a jump table index for this jump table.
2341 unsigned JTEncoding = TLI->getJumpTableEncoding();
2342 unsigned JTI = CurMF->getOrCreateJumpTableInfo(JTEncoding)
2343 ->createJumpTableIndex(DestBBs);
2345 // Set the jump table information so that we can codegen it as a second
2346 // MachineBasicBlock
2347 JumpTable JT(-1U, JTI, JumpTableBB, Default);
2348 JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == SwitchBB));
2349 if (CR.CaseBB == SwitchBB)
2350 visitJumpTableHeader(JT, JTH, SwitchBB);
2352 JTCases.push_back(JumpTableBlock(JTH, JT));
2356 /// handleBTSplitSwitchCase - emit comparison and split binary search tree into
2358 bool SelectionDAGBuilder::handleBTSplitSwitchCase(CaseRec& CR,
2359 CaseRecVector& WorkList,
2361 MachineBasicBlock* Default,
2362 MachineBasicBlock* SwitchBB) {
2363 // Get the MachineFunction which holds the current MBB. This is used when
2364 // inserting any additional MBBs necessary to represent the switch.
2365 MachineFunction *CurMF = FuncInfo.MF;
2367 // Figure out which block is immediately after the current one.
2368 MachineFunction::iterator BBI = CR.CaseBB;
2371 Case& FrontCase = *CR.Range.first;
2372 Case& BackCase = *(CR.Range.second-1);
2373 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2375 // Size is the number of Cases represented by this range.
2376 unsigned Size = CR.Range.second - CR.Range.first;
2378 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
2379 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
2381 CaseItr Pivot = CR.Range.first + Size/2;
2383 // Select optimal pivot, maximizing sum density of LHS and RHS. This will
2384 // (heuristically) allow us to emit JumpTable's later.
2385 APInt TSize(First.getBitWidth(), 0);
2386 for (CaseItr I = CR.Range.first, E = CR.Range.second;
2390 APInt LSize = FrontCase.size();
2391 APInt RSize = TSize-LSize;
2392 DEBUG(dbgs() << "Selecting best pivot: \n"
2393 << "First: " << First << ", Last: " << Last <<'\n'
2394 << "LSize: " << LSize << ", RSize: " << RSize << '\n');
2395 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second;
2397 const APInt &LEnd = cast<ConstantInt>(I->High)->getValue();
2398 const APInt &RBegin = cast<ConstantInt>(J->Low)->getValue();
2399 APInt Range = ComputeRange(LEnd, RBegin);
2400 assert((Range - 2ULL).isNonNegative() &&
2401 "Invalid case distance");
2402 // Use volatile double here to avoid excess precision issues on some hosts,
2403 // e.g. that use 80-bit X87 registers.
2404 volatile double LDensity =
2405 (double)LSize.roundToDouble() /
2406 (LEnd - First + 1ULL).roundToDouble();
2407 volatile double RDensity =
2408 (double)RSize.roundToDouble() /
2409 (Last - RBegin + 1ULL).roundToDouble();
2410 volatile double Metric = Range.logBase2()*(LDensity+RDensity);
2411 // Should always split in some non-trivial place
2412 DEBUG(dbgs() <<"=>Step\n"
2413 << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n'
2414 << "LDensity: " << LDensity
2415 << ", RDensity: " << RDensity << '\n'
2416 << "Metric: " << Metric << '\n');
2417 if (FMetric < Metric) {
2420 DEBUG(dbgs() << "Current metric set to: " << FMetric << '\n');
2427 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
2428 if (areJTsAllowed(*TLI)) {
2429 // If our case is dense we *really* should handle it earlier!
2430 assert((FMetric > 0) && "Should handle dense range earlier!");
2432 Pivot = CR.Range.first + Size/2;
2435 CaseRange LHSR(CR.Range.first, Pivot);
2436 CaseRange RHSR(Pivot, CR.Range.second);
2437 const Constant *C = Pivot->Low;
2438 MachineBasicBlock *FalseBB = nullptr, *TrueBB = nullptr;
2440 // We know that we branch to the LHS if the Value being switched on is
2441 // less than the Pivot value, C. We use this to optimize our binary
2442 // tree a bit, by recognizing that if SV is greater than or equal to the
2443 // LHS's Case Value, and that Case Value is exactly one less than the
2444 // Pivot's Value, then we can branch directly to the LHS's Target,
2445 // rather than creating a leaf node for it.
2446 if ((LHSR.second - LHSR.first) == 1 &&
2447 LHSR.first->High == CR.GE &&
2448 cast<ConstantInt>(C)->getValue() ==
2449 (cast<ConstantInt>(CR.GE)->getValue() + 1LL)) {
2450 TrueBB = LHSR.first->BB;
2452 TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2453 CurMF->insert(BBI, TrueBB);
2454 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR));
2456 // Put SV in a virtual register to make it available from the new blocks.
2457 ExportFromCurrentBlock(SV);
2460 // Similar to the optimization above, if the Value being switched on is
2461 // known to be less than the Constant CR.LT, and the current Case Value
2462 // is CR.LT - 1, then we can branch directly to the target block for
2463 // the current Case Value, rather than emitting a RHS leaf node for it.
2464 if ((RHSR.second - RHSR.first) == 1 && CR.LT &&
2465 cast<ConstantInt>(RHSR.first->Low)->getValue() ==
2466 (cast<ConstantInt>(CR.LT)->getValue() - 1LL)) {
2467 FalseBB = RHSR.first->BB;
2469 FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2470 CurMF->insert(BBI, FalseBB);
2471 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR));
2473 // Put SV in a virtual register to make it available from the new blocks.
2474 ExportFromCurrentBlock(SV);
2477 // Create a CaseBlock record representing a conditional branch to
2478 // the LHS node if the value being switched on SV is less than C.
2479 // Otherwise, branch to LHS.
2480 CaseBlock CB(ISD::SETLT, SV, C, nullptr, TrueBB, FalseBB, CR.CaseBB);
2482 if (CR.CaseBB == SwitchBB)
2483 visitSwitchCase(CB, SwitchBB);
2485 SwitchCases.push_back(CB);
2490 /// handleBitTestsSwitchCase - if current case range has few destination and
2491 /// range span less, than machine word bitwidth, encode case range into series
2492 /// of masks and emit bit tests with these masks.
2493 bool SelectionDAGBuilder::handleBitTestsSwitchCase(CaseRec& CR,
2494 CaseRecVector& WorkList,
2496 MachineBasicBlock* Default,
2497 MachineBasicBlock* SwitchBB) {
2498 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
2499 EVT PTy = TLI->getPointerTy();
2500 unsigned IntPtrBits = PTy.getSizeInBits();
2502 Case& FrontCase = *CR.Range.first;
2503 Case& BackCase = *(CR.Range.second-1);
2505 // Get the MachineFunction which holds the current MBB. This is used when
2506 // inserting any additional MBBs necessary to represent the switch.
2507 MachineFunction *CurMF = FuncInfo.MF;
2509 // If target does not have legal shift left, do not emit bit tests at all.
2510 if (!TLI->isOperationLegal(ISD::SHL, PTy))
2514 for (CaseItr I = CR.Range.first, E = CR.Range.second;
2516 // Single case counts one, case range - two.
2517 numCmps += (I->Low == I->High ? 1 : 2);
2520 // Count unique destinations
2521 SmallSet<MachineBasicBlock*, 4> Dests;
2522 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2523 Dests.insert(I->BB);
2524 if (Dests.size() > 3)
2525 // Don't bother the code below, if there are too much unique destinations
2528 DEBUG(dbgs() << "Total number of unique destinations: "
2529 << Dests.size() << '\n'
2530 << "Total number of comparisons: " << numCmps << '\n');
2532 // Compute span of values.
2533 const APInt& minValue = cast<ConstantInt>(FrontCase.Low)->getValue();
2534 const APInt& maxValue = cast<ConstantInt>(BackCase.High)->getValue();
2535 APInt cmpRange = maxValue - minValue;
2537 DEBUG(dbgs() << "Compare range: " << cmpRange << '\n'
2538 << "Low bound: " << minValue << '\n'
2539 << "High bound: " << maxValue << '\n');
2541 if (cmpRange.uge(IntPtrBits) ||
2542 (!(Dests.size() == 1 && numCmps >= 3) &&
2543 !(Dests.size() == 2 && numCmps >= 5) &&
2544 !(Dests.size() >= 3 && numCmps >= 6)))
2547 DEBUG(dbgs() << "Emitting bit tests\n");
2548 APInt lowBound = APInt::getNullValue(cmpRange.getBitWidth());
2550 // Optimize the case where all the case values fit in a
2551 // word without having to subtract minValue. In this case,
2552 // we can optimize away the subtraction.
2553 if (minValue.isNonNegative() && maxValue.slt(IntPtrBits)) {
2554 cmpRange = maxValue;
2556 lowBound = minValue;
2559 CaseBitsVector CasesBits;
2560 unsigned i, count = 0;
2562 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2563 MachineBasicBlock* Dest = I->BB;
2564 for (i = 0; i < count; ++i)
2565 if (Dest == CasesBits[i].BB)
2569 assert((count < 3) && "Too much destinations to test!");
2570 CasesBits.push_back(CaseBits(0, Dest, 0, 0/*Weight*/));
2574 const APInt& lowValue = cast<ConstantInt>(I->Low)->getValue();
2575 const APInt& highValue = cast<ConstantInt>(I->High)->getValue();
2577 uint64_t lo = (lowValue - lowBound).getZExtValue();
2578 uint64_t hi = (highValue - lowBound).getZExtValue();
2579 CasesBits[i].ExtraWeight += I->ExtraWeight;
2581 for (uint64_t j = lo; j <= hi; j++) {
2582 CasesBits[i].Mask |= 1ULL << j;
2583 CasesBits[i].Bits++;
2587 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp());
2591 // Figure out which block is immediately after the current one.
2592 MachineFunction::iterator BBI = CR.CaseBB;
2595 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2597 DEBUG(dbgs() << "Cases:\n");
2598 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) {
2599 DEBUG(dbgs() << "Mask: " << CasesBits[i].Mask
2600 << ", Bits: " << CasesBits[i].Bits
2601 << ", BB: " << CasesBits[i].BB << '\n');
2603 MachineBasicBlock *CaseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2604 CurMF->insert(BBI, CaseBB);
2605 BTC.push_back(BitTestCase(CasesBits[i].Mask,
2607 CasesBits[i].BB, CasesBits[i].ExtraWeight));
2609 // Put SV in a virtual register to make it available from the new blocks.
2610 ExportFromCurrentBlock(SV);
2613 BitTestBlock BTB(lowBound, cmpRange, SV,
2614 -1U, MVT::Other, (CR.CaseBB == SwitchBB),
2615 CR.CaseBB, Default, BTC);
2617 if (CR.CaseBB == SwitchBB)
2618 visitBitTestHeader(BTB, SwitchBB);
2620 BitTestCases.push_back(BTB);
2625 /// Clusterify - Transform simple list of Cases into list of CaseRange's
2626 size_t SelectionDAGBuilder::Clusterify(CaseVector& Cases,
2627 const SwitchInst& SI) {
2630 BranchProbabilityInfo *BPI = FuncInfo.BPI;
2631 // Start with "simple" cases
2632 for (SwitchInst::ConstCaseIt i = SI.case_begin(), e = SI.case_end();
2634 const BasicBlock *SuccBB = i.getCaseSuccessor();
2635 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SuccBB];
2637 uint32_t ExtraWeight =
2638 BPI ? BPI->getEdgeWeight(SI.getParent(), i.getSuccessorIndex()) : 0;
2640 Cases.push_back(Case(i.getCaseValue(), i.getCaseValue(),
2641 SMBB, ExtraWeight));
2643 std::sort(Cases.begin(), Cases.end(), CaseCmp());
2645 // Merge case into clusters
2646 if (Cases.size() >= 2)
2647 // Must recompute end() each iteration because it may be
2648 // invalidated by erase if we hold on to it
2649 for (CaseItr I = Cases.begin(), J = std::next(Cases.begin());
2650 J != Cases.end(); ) {
2651 const APInt& nextValue = cast<ConstantInt>(J->Low)->getValue();
2652 const APInt& currentValue = cast<ConstantInt>(I->High)->getValue();
2653 MachineBasicBlock* nextBB = J->BB;
2654 MachineBasicBlock* currentBB = I->BB;
2656 // If the two neighboring cases go to the same destination, merge them
2657 // into a single case.
2658 if ((nextValue - currentValue == 1) && (currentBB == nextBB)) {
2660 I->ExtraWeight += J->ExtraWeight;
2667 for (CaseItr I=Cases.begin(), E=Cases.end(); I!=E; ++I, ++numCmps) {
2668 if (I->Low != I->High)
2669 // A range counts double, since it requires two compares.
2676 void SelectionDAGBuilder::UpdateSplitBlock(MachineBasicBlock *First,
2677 MachineBasicBlock *Last) {
2679 for (unsigned i = 0, e = JTCases.size(); i != e; ++i)
2680 if (JTCases[i].first.HeaderBB == First)
2681 JTCases[i].first.HeaderBB = Last;
2683 // Update BitTestCases.
2684 for (unsigned i = 0, e = BitTestCases.size(); i != e; ++i)
2685 if (BitTestCases[i].Parent == First)
2686 BitTestCases[i].Parent = Last;
2689 void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) {
2690 MachineBasicBlock *SwitchMBB = FuncInfo.MBB;
2692 // Figure out which block is immediately after the current one.
2693 MachineBasicBlock *NextBlock = nullptr;
2694 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()];
2696 // If there is only the default destination, branch to it if it is not the
2697 // next basic block. Otherwise, just fall through.
2698 if (!SI.getNumCases()) {
2699 // Update machine-CFG edges.
2701 // If this is not a fall-through branch, emit the branch.
2702 SwitchMBB->addSuccessor(Default);
2703 if (Default != NextBlock)
2704 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(),
2705 MVT::Other, getControlRoot(),
2706 DAG.getBasicBlock(Default)));
2711 // If there are any non-default case statements, create a vector of Cases
2712 // representing each one, and sort the vector so that we can efficiently
2713 // create a binary search tree from them.
2715 size_t numCmps = Clusterify(Cases, SI);
2716 DEBUG(dbgs() << "Clusterify finished. Total clusters: " << Cases.size()
2717 << ". Total compares: " << numCmps << '\n');
2720 // Get the Value to be switched on and default basic blocks, which will be
2721 // inserted into CaseBlock records, representing basic blocks in the binary
2723 const Value *SV = SI.getCondition();
2725 // Push the initial CaseRec onto the worklist
2726 CaseRecVector WorkList;
2727 WorkList.push_back(CaseRec(SwitchMBB,nullptr,nullptr,
2728 CaseRange(Cases.begin(),Cases.end())));
2730 while (!WorkList.empty()) {
2731 // Grab a record representing a case range to process off the worklist
2732 CaseRec CR = WorkList.back();
2733 WorkList.pop_back();
2735 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2738 // If the range has few cases (two or less) emit a series of specific
2740 if (handleSmallSwitchRange(CR, WorkList, SV, Default, SwitchMBB))
2743 // If the switch has more than N blocks, and is at least 40% dense, and the
2744 // target supports indirect branches, then emit a jump table rather than
2745 // lowering the switch to a binary tree of conditional branches.
2746 // N defaults to 4 and is controlled via TLS.getMinimumJumpTableEntries().
2747 if (handleJTSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2750 // Emit binary tree. We need to pick a pivot, and push left and right ranges
2751 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call.
2752 handleBTSplitSwitchCase(CR, WorkList, SV, Default, SwitchMBB);
2756 void SelectionDAGBuilder::visitIndirectBr(const IndirectBrInst &I) {
2757 MachineBasicBlock *IndirectBrMBB = FuncInfo.MBB;
2759 // Update machine-CFG edges with unique successors.
2760 SmallSet<BasicBlock*, 32> Done;
2761 for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i) {
2762 BasicBlock *BB = I.getSuccessor(i);
2763 bool Inserted = Done.insert(BB);
2767 MachineBasicBlock *Succ = FuncInfo.MBBMap[BB];
2768 addSuccessorWithWeight(IndirectBrMBB, Succ);
2771 DAG.setRoot(DAG.getNode(ISD::BRIND, getCurSDLoc(),
2772 MVT::Other, getControlRoot(),
2773 getValue(I.getAddress())));
2776 void SelectionDAGBuilder::visitUnreachable(const UnreachableInst &I) {
2777 if (DAG.getTarget().Options.TrapUnreachable)
2778 DAG.setRoot(DAG.getNode(ISD::TRAP, getCurSDLoc(), MVT::Other, DAG.getRoot()));
2781 void SelectionDAGBuilder::visitFSub(const User &I) {
2782 // -0.0 - X --> fneg
2783 Type *Ty = I.getType();
2784 if (isa<Constant>(I.getOperand(0)) &&
2785 I.getOperand(0) == ConstantFP::getZeroValueForNegation(Ty)) {
2786 SDValue Op2 = getValue(I.getOperand(1));
2787 setValue(&I, DAG.getNode(ISD::FNEG, getCurSDLoc(),
2788 Op2.getValueType(), Op2));
2792 visitBinary(I, ISD::FSUB);
2795 void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) {
2796 SDValue Op1 = getValue(I.getOperand(0));
2797 SDValue Op2 = getValue(I.getOperand(1));
2802 if (const OverflowingBinaryOperator *OFBinOp =
2803 dyn_cast<const OverflowingBinaryOperator>(&I)) {
2804 nuw = OFBinOp->hasNoUnsignedWrap();
2805 nsw = OFBinOp->hasNoSignedWrap();
2807 if (const PossiblyExactOperator *ExactOp =
2808 dyn_cast<const PossiblyExactOperator>(&I))
2809 exact = ExactOp->isExact();
2811 SDValue BinNodeValue = DAG.getNode(OpCode, getCurSDLoc(), Op1.getValueType(),
2812 Op1, Op2, nuw, nsw, exact);
2813 setValue(&I, BinNodeValue);
2816 void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) {
2817 SDValue Op1 = getValue(I.getOperand(0));
2818 SDValue Op2 = getValue(I.getOperand(1));
2820 EVT ShiftTy = TM.getSubtargetImpl()->getTargetLowering()->getShiftAmountTy(
2821 Op2.getValueType());
2823 // Coerce the shift amount to the right type if we can.
2824 if (!I.getType()->isVectorTy() && Op2.getValueType() != ShiftTy) {
2825 unsigned ShiftSize = ShiftTy.getSizeInBits();
2826 unsigned Op2Size = Op2.getValueType().getSizeInBits();
2827 SDLoc DL = getCurSDLoc();
2829 // If the operand is smaller than the shift count type, promote it.
2830 if (ShiftSize > Op2Size)
2831 Op2 = DAG.getNode(ISD::ZERO_EXTEND, DL, ShiftTy, Op2);
2833 // If the operand is larger than the shift count type but the shift
2834 // count type has enough bits to represent any shift value, truncate
2835 // it now. This is a common case and it exposes the truncate to
2836 // optimization early.
2837 else if (ShiftSize >= Log2_32_Ceil(Op2.getValueType().getSizeInBits()))
2838 Op2 = DAG.getNode(ISD::TRUNCATE, DL, ShiftTy, Op2);
2839 // Otherwise we'll need to temporarily settle for some other convenient
2840 // type. Type legalization will make adjustments once the shiftee is split.
2842 Op2 = DAG.getZExtOrTrunc(Op2, DL, MVT::i32);
2849 if (Opcode == ISD::SRL || Opcode == ISD::SRA || Opcode == ISD::SHL) {
2851 if (const OverflowingBinaryOperator *OFBinOp =
2852 dyn_cast<const OverflowingBinaryOperator>(&I)) {
2853 nuw = OFBinOp->hasNoUnsignedWrap();
2854 nsw = OFBinOp->hasNoSignedWrap();
2856 if (const PossiblyExactOperator *ExactOp =
2857 dyn_cast<const PossiblyExactOperator>(&I))
2858 exact = ExactOp->isExact();
2861 SDValue Res = DAG.getNode(Opcode, getCurSDLoc(), Op1.getValueType(), Op1, Op2,
2866 void SelectionDAGBuilder::visitSDiv(const User &I) {
2867 SDValue Op1 = getValue(I.getOperand(0));
2868 SDValue Op2 = getValue(I.getOperand(1));
2870 // Turn exact SDivs into multiplications.
2871 // FIXME: This should be in DAGCombiner, but it doesn't have access to the
2873 if (isa<BinaryOperator>(&I) && cast<BinaryOperator>(&I)->isExact() &&
2874 !isa<ConstantSDNode>(Op1) &&
2875 isa<ConstantSDNode>(Op2) && !cast<ConstantSDNode>(Op2)->isNullValue())
2876 setValue(&I, TM.getSubtargetImpl()->getTargetLowering()->BuildExactSDIV(
2877 Op1, Op2, getCurSDLoc(), DAG));
2879 setValue(&I, DAG.getNode(ISD::SDIV, getCurSDLoc(), Op1.getValueType(),
2883 void SelectionDAGBuilder::visitICmp(const User &I) {
2884 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2885 if (const ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2886 predicate = IC->getPredicate();
2887 else if (const ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2888 predicate = ICmpInst::Predicate(IC->getPredicate());
2889 SDValue Op1 = getValue(I.getOperand(0));
2890 SDValue Op2 = getValue(I.getOperand(1));
2891 ISD::CondCode Opcode = getICmpCondCode(predicate);
2894 TM.getSubtargetImpl()->getTargetLowering()->getValueType(I.getType());
2895 setValue(&I, DAG.getSetCC(getCurSDLoc(), DestVT, Op1, Op2, Opcode));
2898 void SelectionDAGBuilder::visitFCmp(const User &I) {
2899 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2900 if (const FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2901 predicate = FC->getPredicate();
2902 else if (const ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2903 predicate = FCmpInst::Predicate(FC->getPredicate());
2904 SDValue Op1 = getValue(I.getOperand(0));
2905 SDValue Op2 = getValue(I.getOperand(1));
2906 ISD::CondCode Condition = getFCmpCondCode(predicate);
2907 if (TM.Options.NoNaNsFPMath)
2908 Condition = getFCmpCodeWithoutNaN(Condition);
2910 TM.getSubtargetImpl()->getTargetLowering()->getValueType(I.getType());
2911 setValue(&I, DAG.getSetCC(getCurSDLoc(), DestVT, Op1, Op2, Condition));
2914 void SelectionDAGBuilder::visitSelect(const User &I) {
2915 SmallVector<EVT, 4> ValueVTs;
2916 ComputeValueVTs(*TM.getSubtargetImpl()->getTargetLowering(), I.getType(),
2918 unsigned NumValues = ValueVTs.size();
2919 if (NumValues == 0) return;
2921 SmallVector<SDValue, 4> Values(NumValues);
2922 SDValue Cond = getValue(I.getOperand(0));
2923 SDValue TrueVal = getValue(I.getOperand(1));
2924 SDValue FalseVal = getValue(I.getOperand(2));
2925 ISD::NodeType OpCode = Cond.getValueType().isVector() ?
2926 ISD::VSELECT : ISD::SELECT;
2928 for (unsigned i = 0; i != NumValues; ++i)
2929 Values[i] = DAG.getNode(OpCode, getCurSDLoc(),
2930 TrueVal.getNode()->getValueType(TrueVal.getResNo()+i),
2932 SDValue(TrueVal.getNode(),
2933 TrueVal.getResNo() + i),
2934 SDValue(FalseVal.getNode(),
2935 FalseVal.getResNo() + i));
2937 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
2938 DAG.getVTList(ValueVTs), Values));
2941 void SelectionDAGBuilder::visitTrunc(const User &I) {
2942 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2943 SDValue N = getValue(I.getOperand(0));
2945 TM.getSubtargetImpl()->getTargetLowering()->getValueType(I.getType());
2946 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), DestVT, N));
2949 void SelectionDAGBuilder::visitZExt(const User &I) {
2950 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2951 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2952 SDValue N = getValue(I.getOperand(0));
2954 TM.getSubtargetImpl()->getTargetLowering()->getValueType(I.getType());
2955 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurSDLoc(), DestVT, N));
2958 void SelectionDAGBuilder::visitSExt(const User &I) {
2959 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2960 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2961 SDValue N = getValue(I.getOperand(0));
2963 TM.getSubtargetImpl()->getTargetLowering()->getValueType(I.getType());
2964 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurSDLoc(), DestVT, N));
2967 void SelectionDAGBuilder::visitFPTrunc(const User &I) {
2968 // FPTrunc is never a no-op cast, no need to check
2969 SDValue N = getValue(I.getOperand(0));
2970 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
2971 EVT DestVT = TLI->getValueType(I.getType());
2972 setValue(&I, DAG.getNode(ISD::FP_ROUND, getCurSDLoc(),
2974 DAG.getTargetConstant(0, TLI->getPointerTy())));
2977 void SelectionDAGBuilder::visitFPExt(const User &I) {
2978 // FPExt is never a no-op cast, no need to check
2979 SDValue N = getValue(I.getOperand(0));
2981 TM.getSubtargetImpl()->getTargetLowering()->getValueType(I.getType());
2982 setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurSDLoc(), DestVT, N));
2985 void SelectionDAGBuilder::visitFPToUI(const User &I) {
2986 // FPToUI is never a no-op cast, no need to check
2987 SDValue N = getValue(I.getOperand(0));
2989 TM.getSubtargetImpl()->getTargetLowering()->getValueType(I.getType());
2990 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurSDLoc(), DestVT, N));
2993 void SelectionDAGBuilder::visitFPToSI(const User &I) {
2994 // FPToSI is never a no-op cast, no need to check
2995 SDValue N = getValue(I.getOperand(0));
2997 TM.getSubtargetImpl()->getTargetLowering()->getValueType(I.getType());
2998 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurSDLoc(), DestVT, N));
3001 void SelectionDAGBuilder::visitUIToFP(const User &I) {
3002 // UIToFP is never a no-op cast, no need to check
3003 SDValue N = getValue(I.getOperand(0));
3005 TM.getSubtargetImpl()->getTargetLowering()->getValueType(I.getType());
3006 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurSDLoc(), DestVT, N));
3009 void SelectionDAGBuilder::visitSIToFP(const User &I) {
3010 // SIToFP is never a no-op cast, no need to check
3011 SDValue N = getValue(I.getOperand(0));
3013 TM.getSubtargetImpl()->getTargetLowering()->getValueType(I.getType());
3014 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurSDLoc(), DestVT, N));
3017 void SelectionDAGBuilder::visitPtrToInt(const User &I) {
3018 // What to do depends on the size of the integer and the size of the pointer.
3019 // We can either truncate, zero extend, or no-op, accordingly.
3020 SDValue N = getValue(I.getOperand(0));
3022 TM.getSubtargetImpl()->getTargetLowering()->getValueType(I.getType());
3023 setValue(&I, DAG.getZExtOrTrunc(N, getCurSDLoc(), DestVT));
3026 void SelectionDAGBuilder::visitIntToPtr(const User &I) {
3027 // What to do depends on the size of the integer and the size of the pointer.
3028 // We can either truncate, zero extend, or no-op, accordingly.
3029 SDValue N = getValue(I.getOperand(0));
3031 TM.getSubtargetImpl()->getTargetLowering()->getValueType(I.getType());
3032 setValue(&I, DAG.getZExtOrTrunc(N, getCurSDLoc(), DestVT));
3035 void SelectionDAGBuilder::visitBitCast(const User &I) {
3036 SDValue N = getValue(I.getOperand(0));
3038 TM.getSubtargetImpl()->getTargetLowering()->getValueType(I.getType());
3040 // BitCast assures us that source and destination are the same size so this is
3041 // either a BITCAST or a no-op.
3042 if (DestVT != N.getValueType())
3043 setValue(&I, DAG.getNode(ISD::BITCAST, getCurSDLoc(),
3044 DestVT, N)); // convert types.
3045 // Check if the original LLVM IR Operand was a ConstantInt, because getValue()
3046 // might fold any kind of constant expression to an integer constant and that
3047 // is not what we are looking for. Only regcognize a bitcast of a genuine
3048 // constant integer as an opaque constant.
3049 else if(ConstantInt *C = dyn_cast<ConstantInt>(I.getOperand(0)))
3050 setValue(&I, DAG.getConstant(C->getValue(), DestVT, /*isTarget=*/false,
3053 setValue(&I, N); // noop cast.
3056 void SelectionDAGBuilder::visitAddrSpaceCast(const User &I) {
3057 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3058 const Value *SV = I.getOperand(0);
3059 SDValue N = getValue(SV);
3061 TM.getSubtargetImpl()->getTargetLowering()->getValueType(I.getType());
3063 unsigned SrcAS = SV->getType()->getPointerAddressSpace();
3064 unsigned DestAS = I.getType()->getPointerAddressSpace();
3066 if (!TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
3067 N = DAG.getAddrSpaceCast(getCurSDLoc(), DestVT, N, SrcAS, DestAS);
3072 void SelectionDAGBuilder::visitInsertElement(const User &I) {
3073 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3074 SDValue InVec = getValue(I.getOperand(0));
3075 SDValue InVal = getValue(I.getOperand(1));
3076 SDValue InIdx = DAG.getSExtOrTrunc(getValue(I.getOperand(2)),
3077 getCurSDLoc(), TLI.getVectorIdxTy());
3079 DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurSDLoc(),
3080 TM.getSubtargetImpl()->getTargetLowering()->getValueType(
3082 InVec, InVal, InIdx));
3085 void SelectionDAGBuilder::visitExtractElement(const User &I) {
3086 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3087 SDValue InVec = getValue(I.getOperand(0));
3088 SDValue InIdx = DAG.getSExtOrTrunc(getValue(I.getOperand(1)),
3089 getCurSDLoc(), TLI.getVectorIdxTy());
3091 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurSDLoc(),
3092 TM.getSubtargetImpl()->getTargetLowering()->getValueType(
3097 // Utility for visitShuffleVector - Return true if every element in Mask,
3098 // beginning from position Pos and ending in Pos+Size, falls within the
3099 // specified sequential range [L, L+Pos). or is undef.
3100 static bool isSequentialInRange(const SmallVectorImpl<int> &Mask,
3101 unsigned Pos, unsigned Size, int Low) {
3102 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
3103 if (Mask[i] >= 0 && Mask[i] != Low)
3108 void SelectionDAGBuilder::visitShuffleVector(const User &I) {
3109 SDValue Src1 = getValue(I.getOperand(0));
3110 SDValue Src2 = getValue(I.getOperand(1));
3112 SmallVector<int, 8> Mask;
3113 ShuffleVectorInst::getShuffleMask(cast<Constant>(I.getOperand(2)), Mask);
3114 unsigned MaskNumElts = Mask.size();
3116 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
3117 EVT VT = TLI->getValueType(I.getType());
3118 EVT SrcVT = Src1.getValueType();
3119 unsigned SrcNumElts = SrcVT.getVectorNumElements();
3121 if (SrcNumElts == MaskNumElts) {
3122 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
3127 // Normalize the shuffle vector since mask and vector length don't match.
3128 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) {
3129 // Mask is longer than the source vectors and is a multiple of the source
3130 // vectors. We can use concatenate vector to make the mask and vectors
3132 if (SrcNumElts*2 == MaskNumElts) {
3133 // First check for Src1 in low and Src2 in high
3134 if (isSequentialInRange(Mask, 0, SrcNumElts, 0) &&
3135 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, SrcNumElts)) {
3136 // The shuffle is concatenating two vectors together.
3137 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurSDLoc(),
3141 // Then check for Src2 in low and Src1 in high
3142 if (isSequentialInRange(Mask, 0, SrcNumElts, SrcNumElts) &&
3143 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, 0)) {
3144 // The shuffle is concatenating two vectors together.
3145 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurSDLoc(),
3151 // Pad both vectors with undefs to make them the same length as the mask.
3152 unsigned NumConcat = MaskNumElts / SrcNumElts;
3153 bool Src1U = Src1.getOpcode() == ISD::UNDEF;
3154 bool Src2U = Src2.getOpcode() == ISD::UNDEF;
3155 SDValue UndefVal = DAG.getUNDEF(SrcVT);
3157 SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal);
3158 SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal);
3162 Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
3163 getCurSDLoc(), VT, MOps1);
3164 Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
3165 getCurSDLoc(), VT, MOps2);
3167 // Readjust mask for new input vector length.
3168 SmallVector<int, 8> MappedOps;
3169 for (unsigned i = 0; i != MaskNumElts; ++i) {
3171 if (Idx >= (int)SrcNumElts)
3172 Idx -= SrcNumElts - MaskNumElts;
3173 MappedOps.push_back(Idx);
3176 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
3181 if (SrcNumElts > MaskNumElts) {
3182 // Analyze the access pattern of the vector to see if we can extract
3183 // two subvectors and do the shuffle. The analysis is done by calculating
3184 // the range of elements the mask access on both vectors.
3185 int MinRange[2] = { static_cast<int>(SrcNumElts),
3186 static_cast<int>(SrcNumElts)};
3187 int MaxRange[2] = {-1, -1};
3189 for (unsigned i = 0; i != MaskNumElts; ++i) {
3195 if (Idx >= (int)SrcNumElts) {
3199 if (Idx > MaxRange[Input])
3200 MaxRange[Input] = Idx;
3201 if (Idx < MinRange[Input])
3202 MinRange[Input] = Idx;
3205 // Check if the access is smaller than the vector size and can we find
3206 // a reasonable extract index.
3207 int RangeUse[2] = { -1, -1 }; // 0 = Unused, 1 = Extract, -1 = Can not
3209 int StartIdx[2]; // StartIdx to extract from
3210 for (unsigned Input = 0; Input < 2; ++Input) {
3211 if (MinRange[Input] >= (int)SrcNumElts && MaxRange[Input] < 0) {
3212 RangeUse[Input] = 0; // Unused
3213 StartIdx[Input] = 0;
3217 // Find a good start index that is a multiple of the mask length. Then
3218 // see if the rest of the elements are in range.
3219 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts;
3220 if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts &&
3221 StartIdx[Input] + MaskNumElts <= SrcNumElts)
3222 RangeUse[Input] = 1; // Extract from a multiple of the mask length.
3225 if (RangeUse[0] == 0 && RangeUse[1] == 0) {
3226 setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used.
3229 if (RangeUse[0] >= 0 && RangeUse[1] >= 0) {
3230 // Extract appropriate subvector and generate a vector shuffle
3231 for (unsigned Input = 0; Input < 2; ++Input) {
3232 SDValue &Src = Input == 0 ? Src1 : Src2;
3233 if (RangeUse[Input] == 0)
3234 Src = DAG.getUNDEF(VT);
3236 Src = DAG.getNode(ISD::EXTRACT_SUBVECTOR, getCurSDLoc(), VT,
3237 Src, DAG.getConstant(StartIdx[Input],
3238 TLI->getVectorIdxTy()));
3241 // Calculate new mask.
3242 SmallVector<int, 8> MappedOps;
3243 for (unsigned i = 0; i != MaskNumElts; ++i) {
3246 if (Idx < (int)SrcNumElts)
3249 Idx -= SrcNumElts + StartIdx[1] - MaskNumElts;
3251 MappedOps.push_back(Idx);
3254 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
3260 // We can't use either concat vectors or extract subvectors so fall back to
3261 // replacing the shuffle with extract and build vector.
3262 // to insert and build vector.
3263 EVT EltVT = VT.getVectorElementType();
3264 EVT IdxVT = TLI->getVectorIdxTy();
3265 SmallVector<SDValue,8> Ops;
3266 for (unsigned i = 0; i != MaskNumElts; ++i) {
3271 Res = DAG.getUNDEF(EltVT);
3273 SDValue &Src = Idx < (int)SrcNumElts ? Src1 : Src2;
3274 if (Idx >= (int)SrcNumElts) Idx -= SrcNumElts;
3276 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurSDLoc(),
3277 EltVT, Src, DAG.getConstant(Idx, IdxVT));
3283 setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(), VT, Ops));
3286 void SelectionDAGBuilder::visitInsertValue(const InsertValueInst &I) {
3287 const Value *Op0 = I.getOperand(0);
3288 const Value *Op1 = I.getOperand(1);
3289 Type *AggTy = I.getType();
3290 Type *ValTy = Op1->getType();
3291 bool IntoUndef = isa<UndefValue>(Op0);
3292 bool FromUndef = isa<UndefValue>(Op1);
3294 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
3296 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
3297 SmallVector<EVT, 4> AggValueVTs;
3298 ComputeValueVTs(*TLI, AggTy, AggValueVTs);
3299 SmallVector<EVT, 4> ValValueVTs;
3300 ComputeValueVTs(*TLI, ValTy, ValValueVTs);
3302 unsigned NumAggValues = AggValueVTs.size();
3303 unsigned NumValValues = ValValueVTs.size();
3304 SmallVector<SDValue, 4> Values(NumAggValues);
3306 SDValue Agg = getValue(Op0);
3308 // Copy the beginning value(s) from the original aggregate.
3309 for (; i != LinearIndex; ++i)
3310 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
3311 SDValue(Agg.getNode(), Agg.getResNo() + i);
3312 // Copy values from the inserted value(s).
3314 SDValue Val = getValue(Op1);
3315 for (; i != LinearIndex + NumValValues; ++i)
3316 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) :
3317 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex);
3319 // Copy remaining value(s) from the original aggregate.
3320 for (; i != NumAggValues; ++i)
3321 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
3322 SDValue(Agg.getNode(), Agg.getResNo() + i);
3324 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
3325 DAG.getVTList(AggValueVTs), Values));
3328 void SelectionDAGBuilder::visitExtractValue(const ExtractValueInst &I) {
3329 const Value *Op0 = I.getOperand(0);
3330 Type *AggTy = Op0->getType();
3331 Type *ValTy = I.getType();
3332 bool OutOfUndef = isa<UndefValue>(Op0);
3334 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
3336 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
3337 SmallVector<EVT, 4> ValValueVTs;
3338 ComputeValueVTs(*TLI, ValTy, ValValueVTs);
3340 unsigned NumValValues = ValValueVTs.size();
3342 // Ignore a extractvalue that produces an empty object
3343 if (!NumValValues) {
3344 setValue(&I, DAG.getUNDEF(MVT(MVT::Other)));
3348 SmallVector<SDValue, 4> Values(NumValValues);
3350 SDValue Agg = getValue(Op0);
3351 // Copy out the selected value(s).
3352 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i)
3353 Values[i - LinearIndex] =
3355 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) :
3356 SDValue(Agg.getNode(), Agg.getResNo() + i);
3358 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
3359 DAG.getVTList(ValValueVTs), Values));
3362 void SelectionDAGBuilder::visitGetElementPtr(const User &I) {
3363 Value *Op0 = I.getOperand(0);
3364 // Note that the pointer operand may be a vector of pointers. Take the scalar
3365 // element which holds a pointer.
3366 Type *Ty = Op0->getType()->getScalarType();
3367 unsigned AS = Ty->getPointerAddressSpace();
3368 SDValue N = getValue(Op0);
3370 for (GetElementPtrInst::const_op_iterator OI = I.op_begin()+1, E = I.op_end();
3372 const Value *Idx = *OI;
3373 if (StructType *StTy = dyn_cast<StructType>(Ty)) {
3374 unsigned Field = cast<Constant>(Idx)->getUniqueInteger().getZExtValue();
3377 uint64_t Offset = DL->getStructLayout(StTy)->getElementOffset(Field);
3378 N = DAG.getNode(ISD::ADD, getCurSDLoc(), N.getValueType(), N,
3379 DAG.getConstant(Offset, N.getValueType()));
3382 Ty = StTy->getElementType(Field);
3384 Ty = cast<SequentialType>(Ty)->getElementType();
3386 // If this is a constant subscript, handle it quickly.
3387 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
3388 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
3389 if (CI->isZero()) continue;
3391 DL->getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
3393 EVT PTy = TLI->getPointerTy(AS);
3394 unsigned PtrBits = PTy.getSizeInBits();
3396 OffsVal = DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), PTy,
3397 DAG.getConstant(Offs, MVT::i64));
3399 OffsVal = DAG.getConstant(Offs, PTy);
3401 N = DAG.getNode(ISD::ADD, getCurSDLoc(), N.getValueType(), N,
3406 // N = N + Idx * ElementSize;
3407 APInt ElementSize = APInt(TLI->getPointerSizeInBits(AS),
3408 DL->getTypeAllocSize(Ty));
3409 SDValue IdxN = getValue(Idx);
3411 // If the index is smaller or larger than intptr_t, truncate or extend
3413 IdxN = DAG.getSExtOrTrunc(IdxN, getCurSDLoc(), N.getValueType());
3415 // If this is a multiply by a power of two, turn it into a shl
3416 // immediately. This is a very common case.
3417 if (ElementSize != 1) {
3418 if (ElementSize.isPowerOf2()) {
3419 unsigned Amt = ElementSize.logBase2();
3420 IdxN = DAG.getNode(ISD::SHL, getCurSDLoc(),
3421 N.getValueType(), IdxN,
3422 DAG.getConstant(Amt, IdxN.getValueType()));
3424 SDValue Scale = DAG.getConstant(ElementSize, IdxN.getValueType());
3425 IdxN = DAG.getNode(ISD::MUL, getCurSDLoc(),
3426 N.getValueType(), IdxN, Scale);
3430 N = DAG.getNode(ISD::ADD, getCurSDLoc(),
3431 N.getValueType(), N, IdxN);
3438 void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) {
3439 // If this is a fixed sized alloca in the entry block of the function,
3440 // allocate it statically on the stack.
3441 if (FuncInfo.StaticAllocaMap.count(&I))
3442 return; // getValue will auto-populate this.
3444 Type *Ty = I.getAllocatedType();
3445 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
3446 uint64_t TySize = TLI->getDataLayout()->getTypeAllocSize(Ty);
3448 std::max((unsigned)TLI->getDataLayout()->getPrefTypeAlignment(Ty),
3451 SDValue AllocSize = getValue(I.getArraySize());
3453 EVT IntPtr = TLI->getPointerTy();
3454 if (AllocSize.getValueType() != IntPtr)
3455 AllocSize = DAG.getZExtOrTrunc(AllocSize, getCurSDLoc(), IntPtr);
3457 AllocSize = DAG.getNode(ISD::MUL, getCurSDLoc(), IntPtr,
3459 DAG.getConstant(TySize, IntPtr));
3461 // Handle alignment. If the requested alignment is less than or equal to
3462 // the stack alignment, ignore it. If the size is greater than or equal to
3463 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
3464 unsigned StackAlign =
3465 TM.getSubtargetImpl()->getFrameLowering()->getStackAlignment();
3466 if (Align <= StackAlign)
3469 // Round the size of the allocation up to the stack alignment size
3470 // by add SA-1 to the size.
3471 AllocSize = DAG.getNode(ISD::ADD, getCurSDLoc(),
3472 AllocSize.getValueType(), AllocSize,
3473 DAG.getIntPtrConstant(StackAlign-1));
3475 // Mask out the low bits for alignment purposes.
3476 AllocSize = DAG.getNode(ISD::AND, getCurSDLoc(),
3477 AllocSize.getValueType(), AllocSize,
3478 DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1)));
3480 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) };
3481 SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other);
3482 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, getCurSDLoc(), VTs, Ops);
3484 DAG.setRoot(DSA.getValue(1));
3486 assert(FuncInfo.MF->getFrameInfo()->hasVarSizedObjects());
3489 void SelectionDAGBuilder::visitLoad(const LoadInst &I) {
3491 return visitAtomicLoad(I);
3493 const Value *SV = I.getOperand(0);
3494 SDValue Ptr = getValue(SV);
3496 Type *Ty = I.getType();
3498 bool isVolatile = I.isVolatile();
3499 bool isNonTemporal = I.getMetadata("nontemporal") != nullptr;
3500 bool isInvariant = I.getMetadata("invariant.load") != nullptr;
3501 unsigned Alignment = I.getAlignment();
3504 I.getAAMetadata(AAInfo);
3505 const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range);
3507 SmallVector<EVT, 4> ValueVTs;
3508 SmallVector<uint64_t, 4> Offsets;
3509 ComputeValueVTs(*TM.getSubtargetImpl()->getTargetLowering(), Ty, ValueVTs,
3511 unsigned NumValues = ValueVTs.size();
3516 bool ConstantMemory = false;
3517 if (isVolatile || NumValues > MaxParallelChains)
3518 // Serialize volatile loads with other side effects.
3520 else if (AA->pointsToConstantMemory(
3521 AliasAnalysis::Location(SV, AA->getTypeStoreSize(Ty), AAInfo))) {
3522 // Do not serialize (non-volatile) loads of constant memory with anything.
3523 Root = DAG.getEntryNode();
3524 ConstantMemory = true;
3526 // Do not serialize non-volatile loads against each other.
3527 Root = DAG.getRoot();
3530 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
3532 Root = TLI->prepareVolatileOrAtomicLoad(Root, getCurSDLoc(), DAG);
3534 SmallVector<SDValue, 4> Values(NumValues);
3535 SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains),
3537 EVT PtrVT = Ptr.getValueType();
3538 unsigned ChainI = 0;
3539 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3540 // Serializing loads here may result in excessive register pressure, and
3541 // TokenFactor places arbitrary choke points on the scheduler. SD scheduling
3542 // could recover a bit by hoisting nodes upward in the chain by recognizing
3543 // they are side-effect free or do not alias. The optimizer should really
3544 // avoid this case by converting large object/array copies to llvm.memcpy
3545 // (MaxParallelChains should always remain as failsafe).
3546 if (ChainI == MaxParallelChains) {
3547 assert(PendingLoads.empty() && "PendingLoads must be serialized first");
3548 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
3549 makeArrayRef(Chains.data(), ChainI));
3553 SDValue A = DAG.getNode(ISD::ADD, getCurSDLoc(),
3555 DAG.getConstant(Offsets[i], PtrVT));
3556 SDValue L = DAG.getLoad(ValueVTs[i], getCurSDLoc(), Root,
3557 A, MachinePointerInfo(SV, Offsets[i]), isVolatile,
3558 isNonTemporal, isInvariant, Alignment, AAInfo,
3562 Chains[ChainI] = L.getValue(1);
3565 if (!ConstantMemory) {
3566 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
3567 makeArrayRef(Chains.data(), ChainI));
3571 PendingLoads.push_back(Chain);
3574 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
3575 DAG.getVTList(ValueVTs), Values));
3578 void SelectionDAGBuilder::visitStore(const StoreInst &I) {
3580 return visitAtomicStore(I);
3582 const Value *SrcV = I.getOperand(0);
3583 const Value *PtrV = I.getOperand(1);
3585 SmallVector<EVT, 4> ValueVTs;
3586 SmallVector<uint64_t, 4> Offsets;
3587 ComputeValueVTs(*TM.getSubtargetImpl()->getTargetLowering(), SrcV->getType(),
3588 ValueVTs, &Offsets);
3589 unsigned NumValues = ValueVTs.size();
3593 // Get the lowered operands. Note that we do this after
3594 // checking if NumResults is zero, because with zero results
3595 // the operands won't have values in the map.
3596 SDValue Src = getValue(SrcV);
3597 SDValue Ptr = getValue(PtrV);
3599 SDValue Root = getRoot();
3600 SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains),
3602 EVT PtrVT = Ptr.getValueType();
3603 bool isVolatile = I.isVolatile();
3604 bool isNonTemporal = I.getMetadata("nontemporal") != nullptr;
3605 unsigned Alignment = I.getAlignment();
3608 I.getAAMetadata(AAInfo);
3610 unsigned ChainI = 0;
3611 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3612 // See visitLoad comments.
3613 if (ChainI == MaxParallelChains) {
3614 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
3615 makeArrayRef(Chains.data(), ChainI));
3619 SDValue Add = DAG.getNode(ISD::ADD, getCurSDLoc(), PtrVT, Ptr,
3620 DAG.getConstant(Offsets[i], PtrVT));
3621 SDValue St = DAG.getStore(Root, getCurSDLoc(),
3622 SDValue(Src.getNode(), Src.getResNo() + i),
3623 Add, MachinePointerInfo(PtrV, Offsets[i]),
3624 isVolatile, isNonTemporal, Alignment, AAInfo);
3625 Chains[ChainI] = St;
3628 SDValue StoreNode = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
3629 makeArrayRef(Chains.data(), ChainI));
3630 DAG.setRoot(StoreNode);
3633 static SDValue InsertFenceForAtomic(SDValue Chain, AtomicOrdering Order,
3634 SynchronizationScope Scope,
3635 bool Before, SDLoc dl,
3637 const TargetLowering &TLI) {
3638 // Fence, if necessary
3640 if (Order == AcquireRelease || Order == SequentiallyConsistent)
3642 else if (Order == Acquire || Order == Monotonic || Order == Unordered)
3645 if (Order == AcquireRelease)
3647 else if (Order == Release || Order == Monotonic || Order == Unordered)
3652 Ops[1] = DAG.getConstant(Order, TLI.getPointerTy());
3653 Ops[2] = DAG.getConstant(Scope, TLI.getPointerTy());
3654 return DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops);
3657 void SelectionDAGBuilder::visitAtomicCmpXchg(const AtomicCmpXchgInst &I) {
3658 SDLoc dl = getCurSDLoc();
3659 AtomicOrdering SuccessOrder = I.getSuccessOrdering();
3660 AtomicOrdering FailureOrder = I.getFailureOrdering();
3661 SynchronizationScope Scope = I.getSynchScope();
3663 SDValue InChain = getRoot();
3665 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
3666 if (TLI->getInsertFencesForAtomic())
3667 InChain = InsertFenceForAtomic(InChain, SuccessOrder, Scope, true, dl,
3670 MVT MemVT = getValue(I.getCompareOperand()).getSimpleValueType();
3671 SDVTList VTs = DAG.getVTList(MemVT, MVT::i1, MVT::Other);
3672 SDValue L = DAG.getAtomicCmpSwap(
3673 ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, dl, MemVT, VTs, InChain,
3674 getValue(I.getPointerOperand()), getValue(I.getCompareOperand()),
3675 getValue(I.getNewValOperand()), MachinePointerInfo(I.getPointerOperand()),
3677 TLI->getInsertFencesForAtomic() ? Monotonic : SuccessOrder,
3678 TLI->getInsertFencesForAtomic() ? Monotonic : FailureOrder, Scope);
3680 SDValue OutChain = L.getValue(2);
3682 if (TLI->getInsertFencesForAtomic())
3683 OutChain = InsertFenceForAtomic(OutChain, SuccessOrder, Scope, false, dl,
3687 DAG.setRoot(OutChain);
3690 void SelectionDAGBuilder::visitAtomicRMW(const AtomicRMWInst &I) {
3691 SDLoc dl = getCurSDLoc();
3693 switch (I.getOperation()) {
3694 default: llvm_unreachable("Unknown atomicrmw operation");
3695 case AtomicRMWInst::Xchg: NT = ISD::ATOMIC_SWAP; break;
3696 case AtomicRMWInst::Add: NT = ISD::ATOMIC_LOAD_ADD; break;
3697 case AtomicRMWInst::Sub: NT = ISD::ATOMIC_LOAD_SUB; break;
3698 case AtomicRMWInst::And: NT = ISD::ATOMIC_LOAD_AND; break;
3699 case AtomicRMWInst::Nand: NT = ISD::ATOMIC_LOAD_NAND; break;
3700 case AtomicRMWInst::Or: NT = ISD::ATOMIC_LOAD_OR; break;
3701 case AtomicRMWInst::Xor: NT = ISD::ATOMIC_LOAD_XOR; break;
3702 case AtomicRMWInst::Max: NT = ISD::ATOMIC_LOAD_MAX; break;
3703 case AtomicRMWInst::Min: NT = ISD::ATOMIC_LOAD_MIN; break;
3704 case AtomicRMWInst::UMax: NT = ISD::ATOMIC_LOAD_UMAX; break;
3705 case AtomicRMWInst::UMin: NT = ISD::ATOMIC_LOAD_UMIN; break;
3707 AtomicOrdering Order = I.getOrdering();
3708 SynchronizationScope Scope = I.getSynchScope();
3710 SDValue InChain = getRoot();
3712 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
3713 if (TLI->getInsertFencesForAtomic())
3714 InChain = InsertFenceForAtomic(InChain, Order, Scope, true, dl,
3718 DAG.getAtomic(NT, dl,
3719 getValue(I.getValOperand()).getSimpleValueType(),
3721 getValue(I.getPointerOperand()),
3722 getValue(I.getValOperand()),
3723 I.getPointerOperand(), 0 /* Alignment */,
3724 TLI->getInsertFencesForAtomic() ? Monotonic : Order,
3727 SDValue OutChain = L.getValue(1);
3729 if (TLI->getInsertFencesForAtomic())
3730 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3734 DAG.setRoot(OutChain);
3737 void SelectionDAGBuilder::visitFence(const FenceInst &I) {
3738 SDLoc dl = getCurSDLoc();
3739 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
3742 Ops[1] = DAG.getConstant(I.getOrdering(), TLI->getPointerTy());
3743 Ops[2] = DAG.getConstant(I.getSynchScope(), TLI->getPointerTy());
3744 DAG.setRoot(DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops));
3747 void SelectionDAGBuilder::visitAtomicLoad(const LoadInst &I) {
3748 SDLoc dl = getCurSDLoc();
3749 AtomicOrdering Order = I.getOrdering();
3750 SynchronizationScope Scope = I.getSynchScope();
3752 SDValue InChain = getRoot();
3754 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
3755 EVT VT = TLI->getValueType(I.getType());
3757 if (I.getAlignment() < VT.getSizeInBits() / 8)
3758 report_fatal_error("Cannot generate unaligned atomic load");
3760 MachineMemOperand *MMO =
3761 DAG.getMachineFunction().
3762 getMachineMemOperand(MachinePointerInfo(I.getPointerOperand()),
3763 MachineMemOperand::MOVolatile |
3764 MachineMemOperand::MOLoad,
3766 I.getAlignment() ? I.getAlignment() :
3767 DAG.getEVTAlignment(VT));
3769 InChain = TLI->prepareVolatileOrAtomicLoad(InChain, dl, DAG);
3771 DAG.getAtomic(ISD::ATOMIC_LOAD, dl, VT, VT, InChain,
3772 getValue(I.getPointerOperand()), MMO,
3773 TLI->getInsertFencesForAtomic() ? Monotonic : Order,
3776 SDValue OutChain = L.getValue(1);
3778 if (TLI->getInsertFencesForAtomic())
3779 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3783 DAG.setRoot(OutChain);
3786 void SelectionDAGBuilder::visitAtomicStore(const StoreInst &I) {
3787 SDLoc dl = getCurSDLoc();
3789 AtomicOrdering Order = I.getOrdering();
3790 SynchronizationScope Scope = I.getSynchScope();
3792 SDValue InChain = getRoot();
3794 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
3795 EVT VT = TLI->getValueType(I.getValueOperand()->getType());
3797 if (I.getAlignment() < VT.getSizeInBits() / 8)
3798 report_fatal_error("Cannot generate unaligned atomic store");
3800 if (TLI->getInsertFencesForAtomic())
3801 InChain = InsertFenceForAtomic(InChain, Order, Scope, true, dl,
3805 DAG.getAtomic(ISD::ATOMIC_STORE, dl, VT,
3807 getValue(I.getPointerOperand()),
3808 getValue(I.getValueOperand()),
3809 I.getPointerOperand(), I.getAlignment(),
3810 TLI->getInsertFencesForAtomic() ? Monotonic : Order,
3813 if (TLI->getInsertFencesForAtomic())
3814 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3817 DAG.setRoot(OutChain);
3820 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
3822 void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I,
3823 unsigned Intrinsic) {
3824 bool HasChain = !I.doesNotAccessMemory();
3825 bool OnlyLoad = HasChain && I.onlyReadsMemory();
3827 // Build the operand list.
3828 SmallVector<SDValue, 8> Ops;
3829 if (HasChain) { // If this intrinsic has side-effects, chainify it.
3831 // We don't need to serialize loads against other loads.
3832 Ops.push_back(DAG.getRoot());
3834 Ops.push_back(getRoot());
3838 // Info is set by getTgtMemInstrinsic
3839 TargetLowering::IntrinsicInfo Info;
3840 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
3841 bool IsTgtIntrinsic = TLI->getTgtMemIntrinsic(Info, I, Intrinsic);
3843 // Add the intrinsic ID as an integer operand if it's not a target intrinsic.
3844 if (!IsTgtIntrinsic || Info.opc == ISD::INTRINSIC_VOID ||
3845 Info.opc == ISD::INTRINSIC_W_CHAIN)
3846 Ops.push_back(DAG.getTargetConstant(Intrinsic, TLI->getPointerTy()));
3848 // Add all operands of the call to the operand list.
3849 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) {
3850 SDValue Op = getValue(I.getArgOperand(i));
3854 SmallVector<EVT, 4> ValueVTs;
3855 ComputeValueVTs(*TLI, I.getType(), ValueVTs);
3858 ValueVTs.push_back(MVT::Other);
3860 SDVTList VTs = DAG.getVTList(ValueVTs);
3864 if (IsTgtIntrinsic) {
3865 // This is target intrinsic that touches memory
3866 Result = DAG.getMemIntrinsicNode(Info.opc, getCurSDLoc(),
3867 VTs, Ops, Info.memVT,
3868 MachinePointerInfo(Info.ptrVal, Info.offset),
3869 Info.align, Info.vol,
3870 Info.readMem, Info.writeMem, Info.size);
3871 } else if (!HasChain) {
3872 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurSDLoc(), VTs, Ops);
3873 } else if (!I.getType()->isVoidTy()) {
3874 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurSDLoc(), VTs, Ops);
3876 Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurSDLoc(), VTs, Ops);
3880 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1);
3882 PendingLoads.push_back(Chain);
3887 if (!I.getType()->isVoidTy()) {
3888 if (VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
3889 EVT VT = TLI->getValueType(PTy);
3890 Result = DAG.getNode(ISD::BITCAST, getCurSDLoc(), VT, Result);
3893 setValue(&I, Result);
3897 /// GetSignificand - Get the significand and build it into a floating-point
3898 /// number with exponent of 1:
3900 /// Op = (Op & 0x007fffff) | 0x3f800000;
3902 /// where Op is the hexadecimal representation of floating point value.
3904 GetSignificand(SelectionDAG &DAG, SDValue Op, SDLoc dl) {
3905 SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3906 DAG.getConstant(0x007fffff, MVT::i32));
3907 SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1,
3908 DAG.getConstant(0x3f800000, MVT::i32));
3909 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, t2);
3912 /// GetExponent - Get the exponent:
3914 /// (float)(int)(((Op & 0x7f800000) >> 23) - 127);
3916 /// where Op is the hexadecimal representation of floating point value.
3918 GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI,
3920 SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3921 DAG.getConstant(0x7f800000, MVT::i32));
3922 SDValue t1 = DAG.getNode(ISD::SRL, dl, MVT::i32, t0,
3923 DAG.getConstant(23, TLI.getPointerTy()));
3924 SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1,
3925 DAG.getConstant(127, MVT::i32));
3926 return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2);
3929 /// getF32Constant - Get 32-bit floating point constant.
3931 getF32Constant(SelectionDAG &DAG, unsigned Flt) {
3932 return DAG.getConstantFP(APFloat(APFloat::IEEEsingle, APInt(32, Flt)),
3936 /// expandExp - Lower an exp intrinsic. Handles the special sequences for
3937 /// limited-precision mode.
3938 static SDValue expandExp(SDLoc dl, SDValue Op, SelectionDAG &DAG,
3939 const TargetLowering &TLI) {
3940 if (Op.getValueType() == MVT::f32 &&
3941 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3943 // Put the exponent in the right bit position for later addition to the
3946 // #define LOG2OFe 1.4426950f
3947 // IntegerPartOfX = ((int32_t)(X * LOG2OFe));
3948 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3949 getF32Constant(DAG, 0x3fb8aa3b));
3950 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3952 // FractionalPartOfX = (X * LOG2OFe) - (float)IntegerPartOfX;
3953 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3954 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3956 // IntegerPartOfX <<= 23;
3957 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3958 DAG.getConstant(23, TLI.getPointerTy()));
3960 SDValue TwoToFracPartOfX;
3961 if (LimitFloatPrecision <= 6) {
3962 // For floating-point precision of 6:
3964 // TwoToFractionalPartOfX =
3966 // (0.735607626f + 0.252464424f * x) * x;
3968 // error 0.0144103317, which is 6 bits
3969 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3970 getF32Constant(DAG, 0x3e814304));
3971 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3972 getF32Constant(DAG, 0x3f3c50c8));
3973 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3974 TwoToFracPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3975 getF32Constant(DAG, 0x3f7f5e7e));
3976 } else if (LimitFloatPrecision <= 12) {
3977 // For floating-point precision of 12:
3979 // TwoToFractionalPartOfX =
3982 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3984 // 0.000107046256 error, which is 13 to 14 bits
3985 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3986 getF32Constant(DAG, 0x3da235e3));
3987 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3988 getF32Constant(DAG, 0x3e65b8f3));
3989 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3990 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3991 getF32Constant(DAG, 0x3f324b07));
3992 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3993 TwoToFracPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3994 getF32Constant(DAG, 0x3f7ff8fd));
3995 } else { // LimitFloatPrecision <= 18
3996 // For floating-point precision of 18:
3998 // TwoToFractionalPartOfX =
4002 // (0.554906021e-1f +
4003 // (0.961591928e-2f +
4004 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
4006 // error 2.47208000*10^(-7), which is better than 18 bits
4007 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4008 getF32Constant(DAG, 0x3924b03e));
4009 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4010 getF32Constant(DAG, 0x3ab24b87));
4011 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4012 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4013 getF32Constant(DAG, 0x3c1d8c17));
4014 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4015 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4016 getF32Constant(DAG, 0x3d634a1d));
4017 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4018 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4019 getF32Constant(DAG, 0x3e75fe14));
4020 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4021 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
4022 getF32Constant(DAG, 0x3f317234));
4023 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
4024 TwoToFracPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
4025 getF32Constant(DAG, 0x3f800000));
4028 // Add the exponent into the result in integer domain.
4029 SDValue t13 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, TwoToFracPartOfX);
4030 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4031 DAG.getNode(ISD::ADD, dl, MVT::i32,
4032 t13, IntegerPartOfX));
4035 // No special expansion.
4036 return DAG.getNode(ISD::FEXP, dl, Op.getValueType(), Op);
4039 /// expandLog - Lower a log intrinsic. Handles the special sequences for
4040 /// limited-precision mode.
4041 static SDValue expandLog(SDLoc dl, SDValue Op, SelectionDAG &DAG,
4042 const TargetLowering &TLI) {
4043 if (Op.getValueType() == MVT::f32 &&
4044 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4045 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
4047 // Scale the exponent by log(2) [0.69314718f].
4048 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
4049 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
4050 getF32Constant(DAG, 0x3f317218));
4052 // Get the significand and build it into a floating-point number with
4054 SDValue X = GetSignificand(DAG, Op1, dl);
4056 SDValue LogOfMantissa;
4057 if (LimitFloatPrecision <= 6) {
4058 // For floating-point precision of 6:
4062 // (1.4034025f - 0.23903021f * x) * x;
4064 // error 0.0034276066, which is better than 8 bits
4065 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4066 getF32Constant(DAG, 0xbe74c456));
4067 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4068 getF32Constant(DAG, 0x3fb3a2b1));
4069 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4070 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4071 getF32Constant(DAG, 0x3f949a29));
4072 } else if (LimitFloatPrecision <= 12) {
4073 // For floating-point precision of 12:
4079 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x;
4081 // error 0.000061011436, which is 14 bits
4082 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4083 getF32Constant(DAG, 0xbd67b6d6));
4084 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4085 getF32Constant(DAG, 0x3ee4f4b8));
4086 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4087 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4088 getF32Constant(DAG, 0x3fbc278b));
4089 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4090 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4091 getF32Constant(DAG, 0x40348e95));
4092 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4093 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
4094 getF32Constant(DAG, 0x3fdef31a));
4095 } else { // LimitFloatPrecision <= 18
4096 // For floating-point precision of 18:
4104 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x;
4106 // error 0.0000023660568, which is better than 18 bits
4107 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4108 getF32Constant(DAG, 0xbc91e5ac));
4109 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4110 getF32Constant(DAG, 0x3e4350aa));
4111 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4112 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4113 getF32Constant(DAG, 0x3f60d3e3));
4114 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4115 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4116 getF32Constant(DAG, 0x4011cdf0));
4117 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4118 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
4119 getF32Constant(DAG, 0x406cfd1c));
4120 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4121 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4122 getF32Constant(DAG, 0x408797cb));
4123 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4124 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
4125 getF32Constant(DAG, 0x4006dcab));
4128 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, LogOfMantissa);
4131 // No special expansion.
4132 return DAG.getNode(ISD::FLOG, dl, Op.getValueType(), Op);
4135 /// expandLog2 - Lower a log2 intrinsic. Handles the special sequences for
4136 /// limited-precision mode.
4137 static SDValue expandLog2(SDLoc dl, SDValue Op, SelectionDAG &DAG,
4138 const TargetLowering &TLI) {
4139 if (Op.getValueType() == MVT::f32 &&
4140 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4141 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
4143 // Get the exponent.
4144 SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl);
4146 // Get the significand and build it into a floating-point number with
4148 SDValue X = GetSignificand(DAG, Op1, dl);
4150 // Different possible minimax approximations of significand in
4151 // floating-point for various degrees of accuracy over [1,2].
4152 SDValue Log2ofMantissa;
4153 if (LimitFloatPrecision <= 6) {
4154 // For floating-point precision of 6:
4156 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x;
4158 // error 0.0049451742, which is more than 7 bits
4159 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4160 getF32Constant(DAG, 0xbeb08fe0));
4161 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4162 getF32Constant(DAG, 0x40019463));
4163 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4164 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4165 getF32Constant(DAG, 0x3fd6633d));
4166 } else if (LimitFloatPrecision <= 12) {
4167 // For floating-point precision of 12:
4173 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x;
4175 // error 0.0000876136000, which is better than 13 bits
4176 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4177 getF32Constant(DAG, 0xbda7262e));
4178 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4179 getF32Constant(DAG, 0x3f25280b));
4180 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4181 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4182 getF32Constant(DAG, 0x4007b923));
4183 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4184 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4185 getF32Constant(DAG, 0x40823e2f));
4186 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4187 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
4188 getF32Constant(DAG, 0x4020d29c));
4189 } else { // LimitFloatPrecision <= 18
4190 // For floating-point precision of 18:
4199 // 0.25691327e-1f * x) * x) * x) * x) * x) * x;
4201 // error 0.0000018516, which is better than 18 bits
4202 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4203 getF32Constant(DAG, 0xbcd2769e));
4204 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4205 getF32Constant(DAG, 0x3e8ce0b9));
4206 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4207 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4208 getF32Constant(DAG, 0x3fa22ae7));
4209 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4210 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4211 getF32Constant(DAG, 0x40525723));
4212 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4213 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
4214 getF32Constant(DAG, 0x40aaf200));
4215 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4216 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4217 getF32Constant(DAG, 0x40c39dad));
4218 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4219 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
4220 getF32Constant(DAG, 0x4042902c));
4223 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, Log2ofMantissa);
4226 // No special expansion.
4227 return DAG.getNode(ISD::FLOG2, dl, Op.getValueType(), Op);
4230 /// expandLog10 - Lower a log10 intrinsic. Handles the special sequences for
4231 /// limited-precision mode.
4232 static SDValue expandLog10(SDLoc dl, SDValue Op, SelectionDAG &DAG,
4233 const TargetLowering &TLI) {
4234 if (Op.getValueType() == MVT::f32 &&
4235 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4236 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
4238 // Scale the exponent by log10(2) [0.30102999f].
4239 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
4240 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
4241 getF32Constant(DAG, 0x3e9a209a));
4243 // Get the significand and build it into a floating-point number with
4245 SDValue X = GetSignificand(DAG, Op1, dl);
4247 SDValue Log10ofMantissa;
4248 if (LimitFloatPrecision <= 6) {
4249 // For floating-point precision of 6:
4251 // Log10ofMantissa =
4253 // (0.60948995f - 0.10380950f * x) * x;
4255 // error 0.0014886165, which is 6 bits
4256 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4257 getF32Constant(DAG, 0xbdd49a13));
4258 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4259 getF32Constant(DAG, 0x3f1c0789));
4260 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4261 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4262 getF32Constant(DAG, 0x3f011300));
4263 } else if (LimitFloatPrecision <= 12) {
4264 // For floating-point precision of 12:
4266 // Log10ofMantissa =
4269 // (-0.31664806f + 0.47637168e-1f * x) * x) * x;
4271 // error 0.00019228036, which is better than 12 bits
4272 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4273 getF32Constant(DAG, 0x3d431f31));
4274 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
4275 getF32Constant(DAG, 0x3ea21fb2));
4276 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4277 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4278 getF32Constant(DAG, 0x3f6ae232));
4279 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4280 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
4281 getF32Constant(DAG, 0x3f25f7c3));
4282 } else { // LimitFloatPrecision <= 18
4283 // For floating-point precision of 18:
4285 // Log10ofMantissa =
4290 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x;
4292 // error 0.0000037995730, which is better than 18 bits
4293 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4294 getF32Constant(DAG, 0x3c5d51ce));
4295 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
4296 getF32Constant(DAG, 0x3e00685a));
4297 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4298 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4299 getF32Constant(DAG, 0x3efb6798));
4300 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4301 SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
4302 getF32Constant(DAG, 0x3f88d192));
4303 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4304 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4305 getF32Constant(DAG, 0x3fc4316c));
4306 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4307 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8,
4308 getF32Constant(DAG, 0x3f57ce70));
4311 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, Log10ofMantissa);
4314 // No special expansion.
4315 return DAG.getNode(ISD::FLOG10, dl, Op.getValueType(), Op);
4318 /// expandExp2 - Lower an exp2 intrinsic. Handles the special sequences for
4319 /// limited-precision mode.
4320 static SDValue expandExp2(SDLoc dl, SDValue Op, SelectionDAG &DAG,
4321 const TargetLowering &TLI) {
4322 if (Op.getValueType() == MVT::f32 &&
4323 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4324 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Op);
4326 // FractionalPartOfX = x - (float)IntegerPartOfX;
4327 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
4328 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, Op, t1);
4330 // IntegerPartOfX <<= 23;
4331 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
4332 DAG.getConstant(23, TLI.getPointerTy()));
4334 SDValue TwoToFractionalPartOfX;
4335 if (LimitFloatPrecision <= 6) {
4336 // For floating-point precision of 6:
4338 // TwoToFractionalPartOfX =
4340 // (0.735607626f + 0.252464424f * x) * x;
4342 // error 0.0144103317, which is 6 bits
4343 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4344 getF32Constant(DAG, 0x3e814304));
4345 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4346 getF32Constant(DAG, 0x3f3c50c8));
4347 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4348 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4349 getF32Constant(DAG, 0x3f7f5e7e));
4350 } else if (LimitFloatPrecision <= 12) {
4351 // For floating-point precision of 12:
4353 // TwoToFractionalPartOfX =
4356 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
4358 // error 0.000107046256, which is 13 to 14 bits
4359 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4360 getF32Constant(DAG, 0x3da235e3));
4361 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4362 getF32Constant(DAG, 0x3e65b8f3));
4363 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4364 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4365 getF32Constant(DAG, 0x3f324b07));
4366 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4367 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4368 getF32Constant(DAG, 0x3f7ff8fd));
4369 } else { // LimitFloatPrecision <= 18
4370 // For floating-point precision of 18:
4372 // TwoToFractionalPartOfX =
4376 // (0.554906021e-1f +
4377 // (0.961591928e-2f +
4378 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
4379 // error 2.47208000*10^(-7), which is better than 18 bits
4380 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4381 getF32Constant(DAG, 0x3924b03e));
4382 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4383 getF32Constant(DAG, 0x3ab24b87));
4384 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4385 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4386 getF32Constant(DAG, 0x3c1d8c17));
4387 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4388 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4389 getF32Constant(DAG, 0x3d634a1d));
4390 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4391 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4392 getF32Constant(DAG, 0x3e75fe14));
4393 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4394 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
4395 getF32Constant(DAG, 0x3f317234));
4396 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
4397 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
4398 getF32Constant(DAG, 0x3f800000));
4401 // Add the exponent into the result in integer domain.
4402 SDValue t13 = DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4403 TwoToFractionalPartOfX);
4404 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4405 DAG.getNode(ISD::ADD, dl, MVT::i32,
4406 t13, IntegerPartOfX));
4409 // No special expansion.
4410 return DAG.getNode(ISD::FEXP2, dl, Op.getValueType(), Op);
4413 /// visitPow - Lower a pow intrinsic. Handles the special sequences for
4414 /// limited-precision mode with x == 10.0f.
4415 static SDValue expandPow(SDLoc dl, SDValue LHS, SDValue RHS,
4416 SelectionDAG &DAG, const TargetLowering &TLI) {
4417 bool IsExp10 = false;
4418 if (LHS.getValueType() == MVT::f32 && RHS.getValueType() == MVT::f32 &&
4419 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4420 if (ConstantFPSDNode *LHSC = dyn_cast<ConstantFPSDNode>(LHS)) {
4422 IsExp10 = LHSC->isExactlyValue(Ten);
4427 // Put the exponent in the right bit position for later addition to the
4430 // #define LOG2OF10 3.3219281f
4431 // IntegerPartOfX = (int32_t)(x * LOG2OF10);
4432 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, RHS,
4433 getF32Constant(DAG, 0x40549a78));
4434 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
4436 // FractionalPartOfX = x - (float)IntegerPartOfX;
4437 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
4438 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
4440 // IntegerPartOfX <<= 23;
4441 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
4442 DAG.getConstant(23, TLI.getPointerTy()));
4444 SDValue TwoToFractionalPartOfX;
4445 if (LimitFloatPrecision <= 6) {
4446 // For floating-point precision of 6:
4448 // twoToFractionalPartOfX =
4450 // (0.735607626f + 0.252464424f * x) * x;
4452 // error 0.0144103317, which is 6 bits
4453 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4454 getF32Constant(DAG, 0x3e814304));
4455 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4456 getF32Constant(DAG, 0x3f3c50c8));
4457 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4458 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4459 getF32Constant(DAG, 0x3f7f5e7e));
4460 } else if (LimitFloatPrecision <= 12) {
4461 // For floating-point precision of 12:
4463 // TwoToFractionalPartOfX =
4466 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
4468 // error 0.000107046256, which is 13 to 14 bits
4469 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4470 getF32Constant(DAG, 0x3da235e3));
4471 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4472 getF32Constant(DAG, 0x3e65b8f3));
4473 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4474 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4475 getF32Constant(DAG, 0x3f324b07));
4476 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4477 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4478 getF32Constant(DAG, 0x3f7ff8fd));
4479 } else { // LimitFloatPrecision <= 18
4480 // For floating-point precision of 18:
4482 // TwoToFractionalPartOfX =
4486 // (0.554906021e-1f +
4487 // (0.961591928e-2f +
4488 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
4489 // error 2.47208000*10^(-7), which is better than 18 bits
4490 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4491 getF32Constant(DAG, 0x3924b03e));
4492 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4493 getF32Constant(DAG, 0x3ab24b87));
4494 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4495 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4496 getF32Constant(DAG, 0x3c1d8c17));
4497 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4498 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4499 getF32Constant(DAG, 0x3d634a1d));
4500 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4501 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4502 getF32Constant(DAG, 0x3e75fe14));
4503 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4504 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
4505 getF32Constant(DAG, 0x3f317234));
4506 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
4507 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
4508 getF32Constant(DAG, 0x3f800000));
4511 SDValue t13 = DAG.getNode(ISD::BITCAST, dl,MVT::i32,TwoToFractionalPartOfX);
4512 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4513 DAG.getNode(ISD::ADD, dl, MVT::i32,
4514 t13, IntegerPartOfX));
4517 // No special expansion.
4518 return DAG.getNode(ISD::FPOW, dl, LHS.getValueType(), LHS, RHS);
4522 /// ExpandPowI - Expand a llvm.powi intrinsic.
4523 static SDValue ExpandPowI(SDLoc DL, SDValue LHS, SDValue RHS,
4524 SelectionDAG &DAG) {
4525 // If RHS is a constant, we can expand this out to a multiplication tree,
4526 // otherwise we end up lowering to a call to __powidf2 (for example). When
4527 // optimizing for size, we only want to do this if the expansion would produce
4528 // a small number of multiplies, otherwise we do the full expansion.
4529 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
4530 // Get the exponent as a positive value.
4531 unsigned Val = RHSC->getSExtValue();
4532 if ((int)Val < 0) Val = -Val;
4534 // powi(x, 0) -> 1.0
4536 return DAG.getConstantFP(1.0, LHS.getValueType());
4538 const Function *F = DAG.getMachineFunction().getFunction();
4539 if (!F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
4540 Attribute::OptimizeForSize) ||
4541 // If optimizing for size, don't insert too many multiplies. This
4542 // inserts up to 5 multiplies.
4543 CountPopulation_32(Val)+Log2_32(Val) < 7) {
4544 // We use the simple binary decomposition method to generate the multiply
4545 // sequence. There are more optimal ways to do this (for example,
4546 // powi(x,15) generates one more multiply than it should), but this has
4547 // the benefit of being both really simple and much better than a libcall.
4548 SDValue Res; // Logically starts equal to 1.0
4549 SDValue CurSquare = LHS;
4553 Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare);
4555 Res = CurSquare; // 1.0*CurSquare.
4558 CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(),
4559 CurSquare, CurSquare);
4563 // If the original was negative, invert the result, producing 1/(x*x*x).
4564 if (RHSC->getSExtValue() < 0)
4565 Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(),
4566 DAG.getConstantFP(1.0, LHS.getValueType()), Res);
4571 // Otherwise, expand to a libcall.
4572 return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS);
4575 // getTruncatedArgReg - Find underlying register used for an truncated
4577 static unsigned getTruncatedArgReg(const SDValue &N) {
4578 if (N.getOpcode() != ISD::TRUNCATE)
4581 const SDValue &Ext = N.getOperand(0);
4582 if (Ext.getOpcode() == ISD::AssertZext ||
4583 Ext.getOpcode() == ISD::AssertSext) {
4584 const SDValue &CFR = Ext.getOperand(0);
4585 if (CFR.getOpcode() == ISD::CopyFromReg)
4586 return cast<RegisterSDNode>(CFR.getOperand(1))->getReg();
4587 if (CFR.getOpcode() == ISD::TRUNCATE)
4588 return getTruncatedArgReg(CFR);
4593 /// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a function
4594 /// argument, create the corresponding DBG_VALUE machine instruction for it now.
4595 /// At the end of instruction selection, they will be inserted to the entry BB.
4597 SelectionDAGBuilder::EmitFuncArgumentDbgValue(const Value *V, MDNode *Variable,
4598 int64_t Offset, bool IsIndirect,
4600 const Argument *Arg = dyn_cast<Argument>(V);
4604 MachineFunction &MF = DAG.getMachineFunction();
4605 const TargetInstrInfo *TII = DAG.getSubtarget().getInstrInfo();
4607 // Ignore inlined function arguments here.
4608 DIVariable DV(Variable);
4609 if (DV.isInlinedFnArgument(MF.getFunction()))
4612 Optional<MachineOperand> Op;
4613 // Some arguments' frame index is recorded during argument lowering.
4614 if (int FI = FuncInfo.getArgumentFrameIndex(Arg))
4615 Op = MachineOperand::CreateFI(FI);
4617 if (!Op && N.getNode()) {
4619 if (N.getOpcode() == ISD::CopyFromReg)
4620 Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg();
4622 Reg = getTruncatedArgReg(N);
4623 if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) {
4624 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4625 unsigned PR = RegInfo.getLiveInPhysReg(Reg);
4630 Op = MachineOperand::CreateReg(Reg, false);
4634 // Check if ValueMap has reg number.
4635 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
4636 if (VMI != FuncInfo.ValueMap.end())
4637 Op = MachineOperand::CreateReg(VMI->second, false);
4640 if (!Op && N.getNode())
4641 // Check if frame index is available.
4642 if (LoadSDNode *LNode = dyn_cast<LoadSDNode>(N.getNode()))
4643 if (FrameIndexSDNode *FINode =
4644 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode()))
4645 Op = MachineOperand::CreateFI(FINode->getIndex());
4651 FuncInfo.ArgDbgValues.push_back(BuildMI(MF, getCurDebugLoc(),
4652 TII->get(TargetOpcode::DBG_VALUE),
4654 Op->getReg(), Offset, Variable));
4656 FuncInfo.ArgDbgValues.push_back(
4657 BuildMI(MF, getCurDebugLoc(), TII->get(TargetOpcode::DBG_VALUE))
4658 .addOperand(*Op).addImm(Offset).addMetadata(Variable));
4663 // VisualStudio defines setjmp as _setjmp
4664 #if defined(_MSC_VER) && defined(setjmp) && \
4665 !defined(setjmp_undefined_for_msvc)
4666 # pragma push_macro("setjmp")
4668 # define setjmp_undefined_for_msvc
4671 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
4672 /// we want to emit this as a call to a named external function, return the name
4673 /// otherwise lower it and return null.
4675 SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) {
4676 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
4677 SDLoc sdl = getCurSDLoc();
4678 DebugLoc dl = getCurDebugLoc();
4681 switch (Intrinsic) {
4683 // By default, turn this into a target intrinsic node.
4684 visitTargetIntrinsic(I, Intrinsic);
4686 case Intrinsic::vastart: visitVAStart(I); return nullptr;
4687 case Intrinsic::vaend: visitVAEnd(I); return nullptr;
4688 case Intrinsic::vacopy: visitVACopy(I); return nullptr;
4689 case Intrinsic::returnaddress:
4690 setValue(&I, DAG.getNode(ISD::RETURNADDR, sdl, TLI->getPointerTy(),
4691 getValue(I.getArgOperand(0))));
4693 case Intrinsic::frameaddress:
4694 setValue(&I, DAG.getNode(ISD::FRAMEADDR, sdl, TLI->getPointerTy(),
4695 getValue(I.getArgOperand(0))));
4697 case Intrinsic::read_register: {
4698 Value *Reg = I.getArgOperand(0);
4699 SDValue RegName = DAG.getMDNode(cast<MDNode>(Reg));
4701 TM.getSubtargetImpl()->getTargetLowering()->getValueType(I.getType());
4702 setValue(&I, DAG.getNode(ISD::READ_REGISTER, sdl, VT, RegName));
4705 case Intrinsic::write_register: {
4706 Value *Reg = I.getArgOperand(0);
4707 Value *RegValue = I.getArgOperand(1);
4708 SDValue Chain = getValue(RegValue).getOperand(0);
4709 SDValue RegName = DAG.getMDNode(cast<MDNode>(Reg));
4710 DAG.setRoot(DAG.getNode(ISD::WRITE_REGISTER, sdl, MVT::Other, Chain,
4711 RegName, getValue(RegValue)));
4714 case Intrinsic::setjmp:
4715 return &"_setjmp"[!TLI->usesUnderscoreSetJmp()];
4716 case Intrinsic::longjmp:
4717 return &"_longjmp"[!TLI->usesUnderscoreLongJmp()];
4718 case Intrinsic::memcpy: {
4719 // Assert for address < 256 since we support only user defined address
4721 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4723 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4725 "Unknown address space");
4726 SDValue Op1 = getValue(I.getArgOperand(0));
4727 SDValue Op2 = getValue(I.getArgOperand(1));
4728 SDValue Op3 = getValue(I.getArgOperand(2));
4729 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4731 Align = 1; // @llvm.memcpy defines 0 and 1 to both mean no alignment.
4732 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4733 DAG.setRoot(DAG.getMemcpy(getRoot(), sdl, Op1, Op2, Op3, Align, isVol, false,
4734 MachinePointerInfo(I.getArgOperand(0)),
4735 MachinePointerInfo(I.getArgOperand(1))));
4738 case Intrinsic::memset: {
4739 // Assert for address < 256 since we support only user defined address
4741 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4743 "Unknown address space");
4744 SDValue Op1 = getValue(I.getArgOperand(0));
4745 SDValue Op2 = getValue(I.getArgOperand(1));
4746 SDValue Op3 = getValue(I.getArgOperand(2));
4747 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4749 Align = 1; // @llvm.memset defines 0 and 1 to both mean no alignment.
4750 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4751 DAG.setRoot(DAG.getMemset(getRoot(), sdl, Op1, Op2, Op3, Align, isVol,
4752 MachinePointerInfo(I.getArgOperand(0))));
4755 case Intrinsic::memmove: {
4756 // Assert for address < 256 since we support only user defined address
4758 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4760 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4762 "Unknown address space");
4763 SDValue Op1 = getValue(I.getArgOperand(0));
4764 SDValue Op2 = getValue(I.getArgOperand(1));
4765 SDValue Op3 = getValue(I.getArgOperand(2));
4766 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4768 Align = 1; // @llvm.memmove defines 0 and 1 to both mean no alignment.
4769 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4770 DAG.setRoot(DAG.getMemmove(getRoot(), sdl, Op1, Op2, Op3, Align, isVol,
4771 MachinePointerInfo(I.getArgOperand(0)),
4772 MachinePointerInfo(I.getArgOperand(1))));
4775 case Intrinsic::dbg_declare: {
4776 const DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
4777 MDNode *Variable = DI.getVariable();
4778 const Value *Address = DI.getAddress();
4779 DIVariable DIVar(Variable);
4780 assert((!DIVar || DIVar.isVariable()) &&
4781 "Variable in DbgDeclareInst should be either null or a DIVariable.");
4782 if (!Address || !DIVar) {
4783 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4787 // Check if address has undef value.
4788 if (isa<UndefValue>(Address) ||
4789 (Address->use_empty() && !isa<Argument>(Address))) {
4790 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4794 SDValue &N = NodeMap[Address];
4795 if (!N.getNode() && isa<Argument>(Address))
4796 // Check unused arguments map.
4797 N = UnusedArgNodeMap[Address];
4800 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
4801 Address = BCI->getOperand(0);
4802 // Parameters are handled specially.
4804 (DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable ||
4805 isa<Argument>(Address));
4807 const AllocaInst *AI = dyn_cast<AllocaInst>(Address);
4809 if (isParameter && !AI) {
4810 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode());
4812 // Byval parameter. We have a frame index at this point.
4813 SDV = DAG.getFrameIndexDbgValue(Variable, FINode->getIndex(),
4814 0, dl, SDNodeOrder);
4816 // Address is an argument, so try to emit its dbg value using
4817 // virtual register info from the FuncInfo.ValueMap.
4818 EmitFuncArgumentDbgValue(Address, Variable, 0, false, N);
4822 SDV = DAG.getDbgValue(Variable, N.getNode(), N.getResNo(),
4823 true, 0, dl, SDNodeOrder);
4825 // Can't do anything with other non-AI cases yet.
4826 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4827 DEBUG(dbgs() << "non-AllocaInst issue for Address: \n\t");
4828 DEBUG(Address->dump());
4831 DAG.AddDbgValue(SDV, N.getNode(), isParameter);
4833 // If Address is an argument then try to emit its dbg value using
4834 // virtual register info from the FuncInfo.ValueMap.
4835 if (!EmitFuncArgumentDbgValue(Address, Variable, 0, false, N)) {
4836 // If variable is pinned by a alloca in dominating bb then
4837 // use StaticAllocaMap.
4838 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Address)) {
4839 if (AI->getParent() != DI.getParent()) {
4840 DenseMap<const AllocaInst*, int>::iterator SI =
4841 FuncInfo.StaticAllocaMap.find(AI);
4842 if (SI != FuncInfo.StaticAllocaMap.end()) {
4843 SDV = DAG.getFrameIndexDbgValue(Variable, SI->second,
4844 0, dl, SDNodeOrder);
4845 DAG.AddDbgValue(SDV, nullptr, false);
4850 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4855 case Intrinsic::dbg_value: {
4856 const DbgValueInst &DI = cast<DbgValueInst>(I);
4857 DIVariable DIVar(DI.getVariable());
4858 assert((!DIVar || DIVar.isVariable()) &&
4859 "Variable in DbgValueInst should be either null or a DIVariable.");
4863 MDNode *Variable = DI.getVariable();
4864 uint64_t Offset = DI.getOffset();
4865 const Value *V = DI.getValue();
4870 if (isa<ConstantInt>(V) || isa<ConstantFP>(V) || isa<UndefValue>(V)) {
4871 SDV = DAG.getConstantDbgValue(Variable, V, Offset, dl, SDNodeOrder);
4872 DAG.AddDbgValue(SDV, nullptr, false);
4874 // Do not use getValue() in here; we don't want to generate code at
4875 // this point if it hasn't been done yet.
4876 SDValue N = NodeMap[V];
4877 if (!N.getNode() && isa<Argument>(V))
4878 // Check unused arguments map.
4879 N = UnusedArgNodeMap[V];
4881 // A dbg.value for an alloca is always indirect.
4882 bool IsIndirect = isa<AllocaInst>(V) || Offset != 0;
4883 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, IsIndirect, N)) {
4884 SDV = DAG.getDbgValue(Variable, N.getNode(),
4885 N.getResNo(), IsIndirect,
4886 Offset, dl, SDNodeOrder);
4887 DAG.AddDbgValue(SDV, N.getNode(), false);
4889 } else if (!V->use_empty() ) {
4890 // Do not call getValue(V) yet, as we don't want to generate code.
4891 // Remember it for later.
4892 DanglingDebugInfo DDI(&DI, dl, SDNodeOrder);
4893 DanglingDebugInfoMap[V] = DDI;
4895 // We may expand this to cover more cases. One case where we have no
4896 // data available is an unreferenced parameter.
4897 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4901 // Build a debug info table entry.
4902 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
4903 V = BCI->getOperand(0);
4904 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
4905 // Don't handle byval struct arguments or VLAs, for example.
4907 DEBUG(dbgs() << "Dropping debug location info for:\n " << DI << "\n");
4908 DEBUG(dbgs() << " Last seen at:\n " << *V << "\n");
4911 DenseMap<const AllocaInst*, int>::iterator SI =
4912 FuncInfo.StaticAllocaMap.find(AI);
4913 if (SI == FuncInfo.StaticAllocaMap.end())
4914 return nullptr; // VLAs.
4918 case Intrinsic::eh_typeid_for: {
4919 // Find the type id for the given typeinfo.
4920 GlobalVariable *GV = ExtractTypeInfo(I.getArgOperand(0));
4921 unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(GV);
4922 Res = DAG.getConstant(TypeID, MVT::i32);
4927 case Intrinsic::eh_return_i32:
4928 case Intrinsic::eh_return_i64:
4929 DAG.getMachineFunction().getMMI().setCallsEHReturn(true);
4930 DAG.setRoot(DAG.getNode(ISD::EH_RETURN, sdl,
4933 getValue(I.getArgOperand(0)),
4934 getValue(I.getArgOperand(1))));
4936 case Intrinsic::eh_unwind_init:
4937 DAG.getMachineFunction().getMMI().setCallsUnwindInit(true);
4939 case Intrinsic::eh_dwarf_cfa: {
4940 SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getArgOperand(0)), sdl,
4941 TLI->getPointerTy());
4942 SDValue Offset = DAG.getNode(ISD::ADD, sdl,
4943 CfaArg.getValueType(),
4944 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, sdl,
4945 CfaArg.getValueType()),
4947 SDValue FA = DAG.getNode(ISD::FRAMEADDR, sdl,
4948 TLI->getPointerTy(),
4949 DAG.getConstant(0, TLI->getPointerTy()));
4950 setValue(&I, DAG.getNode(ISD::ADD, sdl, FA.getValueType(),
4954 case Intrinsic::eh_sjlj_callsite: {
4955 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4956 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(0));
4957 assert(CI && "Non-constant call site value in eh.sjlj.callsite!");
4958 assert(MMI.getCurrentCallSite() == 0 && "Overlapping call sites!");
4960 MMI.setCurrentCallSite(CI->getZExtValue());
4963 case Intrinsic::eh_sjlj_functioncontext: {
4964 // Get and store the index of the function context.
4965 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4967 cast<AllocaInst>(I.getArgOperand(0)->stripPointerCasts());
4968 int FI = FuncInfo.StaticAllocaMap[FnCtx];
4969 MFI->setFunctionContextIndex(FI);
4972 case Intrinsic::eh_sjlj_setjmp: {
4975 Ops[1] = getValue(I.getArgOperand(0));
4976 SDValue Op = DAG.getNode(ISD::EH_SJLJ_SETJMP, sdl,
4977 DAG.getVTList(MVT::i32, MVT::Other), Ops);
4978 setValue(&I, Op.getValue(0));
4979 DAG.setRoot(Op.getValue(1));
4982 case Intrinsic::eh_sjlj_longjmp: {
4983 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, sdl, MVT::Other,
4984 getRoot(), getValue(I.getArgOperand(0))));
4988 case Intrinsic::x86_mmx_pslli_w:
4989 case Intrinsic::x86_mmx_pslli_d:
4990 case Intrinsic::x86_mmx_pslli_q:
4991 case Intrinsic::x86_mmx_psrli_w:
4992 case Intrinsic::x86_mmx_psrli_d:
4993 case Intrinsic::x86_mmx_psrli_q:
4994 case Intrinsic::x86_mmx_psrai_w:
4995 case Intrinsic::x86_mmx_psrai_d: {
4996 SDValue ShAmt = getValue(I.getArgOperand(1));
4997 if (isa<ConstantSDNode>(ShAmt)) {
4998 visitTargetIntrinsic(I, Intrinsic);
5001 unsigned NewIntrinsic = 0;
5002 EVT ShAmtVT = MVT::v2i32;
5003 switch (Intrinsic) {
5004 case Intrinsic::x86_mmx_pslli_w:
5005 NewIntrinsic = Intrinsic::x86_mmx_psll_w;
5007 case Intrinsic::x86_mmx_pslli_d:
5008 NewIntrinsic = Intrinsic::x86_mmx_psll_d;
5010 case Intrinsic::x86_mmx_pslli_q:
5011 NewIntrinsic = Intrinsic::x86_mmx_psll_q;
5013 case Intrinsic::x86_mmx_psrli_w:
5014 NewIntrinsic = Intrinsic::x86_mmx_psrl_w;
5016 case Intrinsic::x86_mmx_psrli_d:
5017 NewIntrinsic = Intrinsic::x86_mmx_psrl_d;
5019 case Intrinsic::x86_mmx_psrli_q:
5020 NewIntrinsic = Intrinsic::x86_mmx_psrl_q;
5022 case Intrinsic::x86_mmx_psrai_w:
5023 NewIntrinsic = Intrinsic::x86_mmx_psra_w;
5025 case Intrinsic::x86_mmx_psrai_d:
5026 NewIntrinsic = Intrinsic::x86_mmx_psra_d;
5028 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
5031 // The vector shift intrinsics with scalars uses 32b shift amounts but
5032 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
5034 // We must do this early because v2i32 is not a legal type.
5037 ShOps[1] = DAG.getConstant(0, MVT::i32);
5038 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, sdl, ShAmtVT, ShOps);
5039 EVT DestVT = TLI->getValueType(I.getType());
5040 ShAmt = DAG.getNode(ISD::BITCAST, sdl, DestVT, ShAmt);
5041 Res = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, sdl, DestVT,
5042 DAG.getConstant(NewIntrinsic, MVT::i32),
5043 getValue(I.getArgOperand(0)), ShAmt);
5047 case Intrinsic::x86_avx_vinsertf128_pd_256:
5048 case Intrinsic::x86_avx_vinsertf128_ps_256:
5049 case Intrinsic::x86_avx_vinsertf128_si_256:
5050 case Intrinsic::x86_avx2_vinserti128: {
5051 EVT DestVT = TLI->getValueType(I.getType());
5052 EVT ElVT = TLI->getValueType(I.getArgOperand(1)->getType());
5053 uint64_t Idx = (cast<ConstantInt>(I.getArgOperand(2))->getZExtValue() & 1) *
5054 ElVT.getVectorNumElements();
5055 Res = DAG.getNode(ISD::INSERT_SUBVECTOR, sdl, DestVT,
5056 getValue(I.getArgOperand(0)),
5057 getValue(I.getArgOperand(1)),
5058 DAG.getConstant(Idx, TLI->getVectorIdxTy()));
5062 case Intrinsic::x86_avx_vextractf128_pd_256:
5063 case Intrinsic::x86_avx_vextractf128_ps_256:
5064 case Intrinsic::x86_avx_vextractf128_si_256:
5065 case Intrinsic::x86_avx2_vextracti128: {
5066 EVT DestVT = TLI->getValueType(I.getType());
5067 uint64_t Idx = (cast<ConstantInt>(I.getArgOperand(1))->getZExtValue() & 1) *
5068 DestVT.getVectorNumElements();
5069 Res = DAG.getNode(ISD::EXTRACT_SUBVECTOR, sdl, DestVT,
5070 getValue(I.getArgOperand(0)),
5071 DAG.getConstant(Idx, TLI->getVectorIdxTy()));
5075 case Intrinsic::convertff:
5076 case Intrinsic::convertfsi:
5077 case Intrinsic::convertfui:
5078 case Intrinsic::convertsif:
5079 case Intrinsic::convertuif:
5080 case Intrinsic::convertss:
5081 case Intrinsic::convertsu:
5082 case Intrinsic::convertus:
5083 case Intrinsic::convertuu: {
5084 ISD::CvtCode Code = ISD::CVT_INVALID;
5085 switch (Intrinsic) {
5086 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
5087 case Intrinsic::convertff: Code = ISD::CVT_FF; break;
5088 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break;
5089 case Intrinsic::convertfui: Code = ISD::CVT_FU; break;
5090 case Intrinsic::convertsif: Code = ISD::CVT_SF; break;
5091 case Intrinsic::convertuif: Code = ISD::CVT_UF; break;
5092 case Intrinsic::convertss: Code = ISD::CVT_SS; break;
5093 case Intrinsic::convertsu: Code = ISD::CVT_SU; break;
5094 case Intrinsic::convertus: Code = ISD::CVT_US; break;
5095 case Intrinsic::convertuu: Code = ISD::CVT_UU; break;
5097 EVT DestVT = TLI->getValueType(I.getType());
5098 const Value *Op1 = I.getArgOperand(0);
5099 Res = DAG.getConvertRndSat(DestVT, sdl, getValue(Op1),
5100 DAG.getValueType(DestVT),
5101 DAG.getValueType(getValue(Op1).getValueType()),
5102 getValue(I.getArgOperand(1)),
5103 getValue(I.getArgOperand(2)),
5108 case Intrinsic::powi:
5109 setValue(&I, ExpandPowI(sdl, getValue(I.getArgOperand(0)),
5110 getValue(I.getArgOperand(1)), DAG));
5112 case Intrinsic::log:
5113 setValue(&I, expandLog(sdl, getValue(I.getArgOperand(0)), DAG, *TLI));
5115 case Intrinsic::log2:
5116 setValue(&I, expandLog2(sdl, getValue(I.getArgOperand(0)), DAG, *TLI));
5118 case Intrinsic::log10:
5119 setValue(&I, expandLog10(sdl, getValue(I.getArgOperand(0)), DAG, *TLI));
5121 case Intrinsic::exp:
5122 setValue(&I, expandExp(sdl, getValue(I.getArgOperand(0)), DAG, *TLI));
5124 case Intrinsic::exp2:
5125 setValue(&I, expandExp2(sdl, getValue(I.getArgOperand(0)), DAG, *TLI));
5127 case Intrinsic::pow:
5128 setValue(&I, expandPow(sdl, getValue(I.getArgOperand(0)),
5129 getValue(I.getArgOperand(1)), DAG, *TLI));
5131 case Intrinsic::sqrt:
5132 case Intrinsic::fabs:
5133 case Intrinsic::sin:
5134 case Intrinsic::cos:
5135 case Intrinsic::floor:
5136 case Intrinsic::ceil:
5137 case Intrinsic::trunc:
5138 case Intrinsic::rint:
5139 case Intrinsic::nearbyint:
5140 case Intrinsic::round: {
5142 switch (Intrinsic) {
5143 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
5144 case Intrinsic::sqrt: Opcode = ISD::FSQRT; break;
5145 case Intrinsic::fabs: Opcode = ISD::FABS; break;
5146 case Intrinsic::sin: Opcode = ISD::FSIN; break;
5147 case Intrinsic::cos: Opcode = ISD::FCOS; break;
5148 case Intrinsic::floor: Opcode = ISD::FFLOOR; break;
5149 case Intrinsic::ceil: Opcode = ISD::FCEIL; break;
5150 case Intrinsic::trunc: Opcode = ISD::FTRUNC; break;
5151 case Intrinsic::rint: Opcode = ISD::FRINT; break;
5152 case Intrinsic::nearbyint: Opcode = ISD::FNEARBYINT; break;
5153 case Intrinsic::round: Opcode = ISD::FROUND; break;
5156 setValue(&I, DAG.getNode(Opcode, sdl,
5157 getValue(I.getArgOperand(0)).getValueType(),
5158 getValue(I.getArgOperand(0))));
5161 case Intrinsic::copysign:
5162 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, sdl,
5163 getValue(I.getArgOperand(0)).getValueType(),
5164 getValue(I.getArgOperand(0)),
5165 getValue(I.getArgOperand(1))));
5167 case Intrinsic::fma:
5168 setValue(&I, DAG.getNode(ISD::FMA, sdl,
5169 getValue(I.getArgOperand(0)).getValueType(),
5170 getValue(I.getArgOperand(0)),
5171 getValue(I.getArgOperand(1)),
5172 getValue(I.getArgOperand(2))));
5174 case Intrinsic::fmuladd: {
5175 EVT VT = TLI->getValueType(I.getType());
5176 if (TM.Options.AllowFPOpFusion != FPOpFusion::Strict &&
5177 TLI->isFMAFasterThanFMulAndFAdd(VT)) {
5178 setValue(&I, DAG.getNode(ISD::FMA, sdl,
5179 getValue(I.getArgOperand(0)).getValueType(),
5180 getValue(I.getArgOperand(0)),
5181 getValue(I.getArgOperand(1)),
5182 getValue(I.getArgOperand(2))));
5184 SDValue Mul = DAG.getNode(ISD::FMUL, sdl,
5185 getValue(I.getArgOperand(0)).getValueType(),
5186 getValue(I.getArgOperand(0)),
5187 getValue(I.getArgOperand(1)));
5188 SDValue Add = DAG.getNode(ISD::FADD, sdl,
5189 getValue(I.getArgOperand(0)).getValueType(),
5191 getValue(I.getArgOperand(2)));
5196 case Intrinsic::convert_to_fp16:
5197 setValue(&I, DAG.getNode(ISD::BITCAST, sdl, MVT::i16,
5198 DAG.getNode(ISD::FP_ROUND, sdl, MVT::f16,
5199 getValue(I.getArgOperand(0)),
5200 DAG.getTargetConstant(0, MVT::i32))));
5202 case Intrinsic::convert_from_fp16:
5204 DAG.getNode(ISD::FP_EXTEND, sdl, TLI->getValueType(I.getType()),
5205 DAG.getNode(ISD::BITCAST, sdl, MVT::f16,
5206 getValue(I.getArgOperand(0)))));
5208 case Intrinsic::pcmarker: {
5209 SDValue Tmp = getValue(I.getArgOperand(0));
5210 DAG.setRoot(DAG.getNode(ISD::PCMARKER, sdl, MVT::Other, getRoot(), Tmp));
5213 case Intrinsic::readcyclecounter: {
5214 SDValue Op = getRoot();
5215 Res = DAG.getNode(ISD::READCYCLECOUNTER, sdl,
5216 DAG.getVTList(MVT::i64, MVT::Other), Op);
5218 DAG.setRoot(Res.getValue(1));
5221 case Intrinsic::bswap:
5222 setValue(&I, DAG.getNode(ISD::BSWAP, sdl,
5223 getValue(I.getArgOperand(0)).getValueType(),
5224 getValue(I.getArgOperand(0))));
5226 case Intrinsic::cttz: {
5227 SDValue Arg = getValue(I.getArgOperand(0));
5228 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1));
5229 EVT Ty = Arg.getValueType();
5230 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTTZ : ISD::CTTZ_ZERO_UNDEF,
5234 case Intrinsic::ctlz: {
5235 SDValue Arg = getValue(I.getArgOperand(0));
5236 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1));
5237 EVT Ty = Arg.getValueType();
5238 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTLZ : ISD::CTLZ_ZERO_UNDEF,
5242 case Intrinsic::ctpop: {
5243 SDValue Arg = getValue(I.getArgOperand(0));
5244 EVT Ty = Arg.getValueType();
5245 setValue(&I, DAG.getNode(ISD::CTPOP, sdl, Ty, Arg));
5248 case Intrinsic::stacksave: {
5249 SDValue Op = getRoot();
5250 Res = DAG.getNode(ISD::STACKSAVE, sdl,
5251 DAG.getVTList(TLI->getPointerTy(), MVT::Other), Op);
5253 DAG.setRoot(Res.getValue(1));
5256 case Intrinsic::stackrestore: {
5257 Res = getValue(I.getArgOperand(0));
5258 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, sdl, MVT::Other, getRoot(), Res));
5261 case Intrinsic::stackprotector: {
5262 // Emit code into the DAG to store the stack guard onto the stack.
5263 MachineFunction &MF = DAG.getMachineFunction();
5264 MachineFrameInfo *MFI = MF.getFrameInfo();
5265 EVT PtrTy = TLI->getPointerTy();
5266 SDValue Src, Chain = getRoot();
5267 const Value *Ptr = cast<LoadInst>(I.getArgOperand(0))->getPointerOperand();
5268 const GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr);
5270 // See if Ptr is a bitcast. If it is, look through it and see if we can get
5271 // global variable __stack_chk_guard.
5273 if (const Operator *BC = dyn_cast<Operator>(Ptr))
5274 if (BC->getOpcode() == Instruction::BitCast)
5275 GV = dyn_cast<GlobalVariable>(BC->getOperand(0));
5277 if (GV && TLI->useLoadStackGuardNode()) {
5278 // Emit a LOAD_STACK_GUARD node.
5279 MachineSDNode *Node = DAG.getMachineNode(TargetOpcode::LOAD_STACK_GUARD,
5281 MachinePointerInfo MPInfo(GV);
5282 MachineInstr::mmo_iterator MemRefs = MF.allocateMemRefsArray(1);
5283 unsigned Flags = MachineMemOperand::MOLoad |
5284 MachineMemOperand::MOInvariant;
5285 *MemRefs = MF.getMachineMemOperand(MPInfo, Flags,
5286 PtrTy.getSizeInBits() / 8,
5287 DAG.getEVTAlignment(PtrTy));
5288 Node->setMemRefs(MemRefs, MemRefs + 1);
5290 // Copy the guard value to a virtual register so that it can be
5291 // retrieved in the epilogue.
5292 Src = SDValue(Node, 0);
5293 const TargetRegisterClass *RC =
5294 TLI->getRegClassFor(Src.getSimpleValueType());
5295 unsigned Reg = MF.getRegInfo().createVirtualRegister(RC);
5297 SPDescriptor.setGuardReg(Reg);
5298 Chain = DAG.getCopyToReg(Chain, sdl, Reg, Src);
5300 Src = getValue(I.getArgOperand(0)); // The guard's value.
5303 AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
5305 int FI = FuncInfo.StaticAllocaMap[Slot];
5306 MFI->setStackProtectorIndex(FI);
5308 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
5310 // Store the stack protector onto the stack.
5311 Res = DAG.getStore(Chain, sdl, Src, FIN,
5312 MachinePointerInfo::getFixedStack(FI),
5318 case Intrinsic::objectsize: {
5319 // If we don't know by now, we're never going to know.
5320 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1));
5322 assert(CI && "Non-constant type in __builtin_object_size?");
5324 SDValue Arg = getValue(I.getCalledValue());
5325 EVT Ty = Arg.getValueType();
5328 Res = DAG.getConstant(-1ULL, Ty);
5330 Res = DAG.getConstant(0, Ty);
5335 case Intrinsic::annotation:
5336 case Intrinsic::ptr_annotation:
5337 // Drop the intrinsic, but forward the value
5338 setValue(&I, getValue(I.getOperand(0)));
5340 case Intrinsic::assume:
5341 case Intrinsic::var_annotation:
5342 // Discard annotate attributes and assumptions
5345 case Intrinsic::init_trampoline: {
5346 const Function *F = cast<Function>(I.getArgOperand(1)->stripPointerCasts());
5350 Ops[1] = getValue(I.getArgOperand(0));
5351 Ops[2] = getValue(I.getArgOperand(1));
5352 Ops[3] = getValue(I.getArgOperand(2));
5353 Ops[4] = DAG.getSrcValue(I.getArgOperand(0));
5354 Ops[5] = DAG.getSrcValue(F);
5356 Res = DAG.getNode(ISD::INIT_TRAMPOLINE, sdl, MVT::Other, Ops);
5361 case Intrinsic::adjust_trampoline: {
5362 setValue(&I, DAG.getNode(ISD::ADJUST_TRAMPOLINE, sdl,
5363 TLI->getPointerTy(),
5364 getValue(I.getArgOperand(0))));
5367 case Intrinsic::gcroot:
5369 const Value *Alloca = I.getArgOperand(0)->stripPointerCasts();
5370 const Constant *TypeMap = cast<Constant>(I.getArgOperand(1));
5372 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode());
5373 GFI->addStackRoot(FI->getIndex(), TypeMap);
5376 case Intrinsic::gcread:
5377 case Intrinsic::gcwrite:
5378 llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!");
5379 case Intrinsic::flt_rounds:
5380 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, sdl, MVT::i32));
5383 case Intrinsic::expect: {
5384 // Just replace __builtin_expect(exp, c) with EXP.
5385 setValue(&I, getValue(I.getArgOperand(0)));
5389 case Intrinsic::debugtrap:
5390 case Intrinsic::trap: {
5391 StringRef TrapFuncName = TM.Options.getTrapFunctionName();
5392 if (TrapFuncName.empty()) {
5393 ISD::NodeType Op = (Intrinsic == Intrinsic::trap) ?
5394 ISD::TRAP : ISD::DEBUGTRAP;
5395 DAG.setRoot(DAG.getNode(Op, sdl,MVT::Other, getRoot()));
5398 TargetLowering::ArgListTy Args;
5400 TargetLowering::CallLoweringInfo CLI(DAG);
5401 CLI.setDebugLoc(sdl).setChain(getRoot())
5402 .setCallee(CallingConv::C, I.getType(),
5403 DAG.getExternalSymbol(TrapFuncName.data(), TLI->getPointerTy()),
5404 std::move(Args), 0);
5406 std::pair<SDValue, SDValue> Result = TLI->LowerCallTo(CLI);
5407 DAG.setRoot(Result.second);
5411 case Intrinsic::uadd_with_overflow:
5412 case Intrinsic::sadd_with_overflow:
5413 case Intrinsic::usub_with_overflow:
5414 case Intrinsic::ssub_with_overflow:
5415 case Intrinsic::umul_with_overflow:
5416 case Intrinsic::smul_with_overflow: {
5418 switch (Intrinsic) {
5419 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
5420 case Intrinsic::uadd_with_overflow: Op = ISD::UADDO; break;
5421 case Intrinsic::sadd_with_overflow: Op = ISD::SADDO; break;
5422 case Intrinsic::usub_with_overflow: Op = ISD::USUBO; break;
5423 case Intrinsic::ssub_with_overflow: Op = ISD::SSUBO; break;
5424 case Intrinsic::umul_with_overflow: Op = ISD::UMULO; break;
5425 case Intrinsic::smul_with_overflow: Op = ISD::SMULO; break;
5427 SDValue Op1 = getValue(I.getArgOperand(0));
5428 SDValue Op2 = getValue(I.getArgOperand(1));
5430 SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1);
5431 setValue(&I, DAG.getNode(Op, sdl, VTs, Op1, Op2));
5434 case Intrinsic::prefetch: {
5436 unsigned rw = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
5438 Ops[1] = getValue(I.getArgOperand(0));
5439 Ops[2] = getValue(I.getArgOperand(1));
5440 Ops[3] = getValue(I.getArgOperand(2));
5441 Ops[4] = getValue(I.getArgOperand(3));
5442 DAG.setRoot(DAG.getMemIntrinsicNode(ISD::PREFETCH, sdl,
5443 DAG.getVTList(MVT::Other), Ops,
5444 EVT::getIntegerVT(*Context, 8),
5445 MachinePointerInfo(I.getArgOperand(0)),
5447 false, /* volatile */
5449 rw==1)); /* write */
5452 case Intrinsic::lifetime_start:
5453 case Intrinsic::lifetime_end: {
5454 bool IsStart = (Intrinsic == Intrinsic::lifetime_start);
5455 // Stack coloring is not enabled in O0, discard region information.
5456 if (TM.getOptLevel() == CodeGenOpt::None)
5459 SmallVector<Value *, 4> Allocas;
5460 GetUnderlyingObjects(I.getArgOperand(1), Allocas, DL);
5462 for (SmallVectorImpl<Value*>::iterator Object = Allocas.begin(),
5463 E = Allocas.end(); Object != E; ++Object) {
5464 AllocaInst *LifetimeObject = dyn_cast_or_null<AllocaInst>(*Object);
5466 // Could not find an Alloca.
5467 if (!LifetimeObject)
5470 int FI = FuncInfo.StaticAllocaMap[LifetimeObject];
5474 Ops[1] = DAG.getFrameIndex(FI, TLI->getPointerTy(), true);
5475 unsigned Opcode = (IsStart ? ISD::LIFETIME_START : ISD::LIFETIME_END);
5477 Res = DAG.getNode(Opcode, sdl, MVT::Other, Ops);
5482 case Intrinsic::invariant_start:
5483 // Discard region information.
5484 setValue(&I, DAG.getUNDEF(TLI->getPointerTy()));
5486 case Intrinsic::invariant_end:
5487 // Discard region information.
5489 case Intrinsic::stackprotectorcheck: {
5490 // Do not actually emit anything for this basic block. Instead we initialize
5491 // the stack protector descriptor and export the guard variable so we can
5492 // access it in FinishBasicBlock.
5493 const BasicBlock *BB = I.getParent();
5494 SPDescriptor.initialize(BB, FuncInfo.MBBMap[BB], I);
5495 ExportFromCurrentBlock(SPDescriptor.getGuard());
5497 // Flush our exports since we are going to process a terminator.
5498 (void)getControlRoot();
5501 case Intrinsic::clear_cache:
5502 return TLI->getClearCacheBuiltinName();
5503 case Intrinsic::donothing:
5506 case Intrinsic::experimental_stackmap: {
5510 case Intrinsic::experimental_patchpoint_void:
5511 case Intrinsic::experimental_patchpoint_i64: {
5518 void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee,
5520 MachineBasicBlock *LandingPad) {
5521 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
5522 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
5523 FunctionType *FTy = cast<FunctionType>(PT->getElementType());
5524 Type *RetTy = FTy->getReturnType();
5525 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5526 MCSymbol *BeginLabel = nullptr;
5528 TargetLowering::ArgListTy Args;
5529 TargetLowering::ArgListEntry Entry;
5530 Args.reserve(CS.arg_size());
5532 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
5534 const Value *V = *i;
5537 if (V->getType()->isEmptyTy())
5540 SDValue ArgNode = getValue(V);
5541 Entry.Node = ArgNode; Entry.Ty = V->getType();
5543 // Skip the first return-type Attribute to get to params.
5544 Entry.setAttributes(&CS, i - CS.arg_begin() + 1);
5545 Args.push_back(Entry);
5549 // Insert a label before the invoke call to mark the try range. This can be
5550 // used to detect deletion of the invoke via the MachineModuleInfo.
5551 BeginLabel = MMI.getContext().CreateTempSymbol();
5553 // For SjLj, keep track of which landing pads go with which invokes
5554 // so as to maintain the ordering of pads in the LSDA.
5555 unsigned CallSiteIndex = MMI.getCurrentCallSite();
5556 if (CallSiteIndex) {
5557 MMI.setCallSiteBeginLabel(BeginLabel, CallSiteIndex);
5558 LPadToCallSiteMap[LandingPad].push_back(CallSiteIndex);
5560 // Now that the call site is handled, stop tracking it.
5561 MMI.setCurrentCallSite(0);
5564 // Both PendingLoads and PendingExports must be flushed here;
5565 // this call might not return.
5567 DAG.setRoot(DAG.getEHLabel(getCurSDLoc(), getControlRoot(), BeginLabel));
5570 // Check if target-independent constraints permit a tail call here.
5571 // Target-dependent constraints are checked within TLI->LowerCallTo.
5572 if (isTailCall && !isInTailCallPosition(CS, DAG.getTarget()))
5575 TargetLowering::CallLoweringInfo CLI(DAG);
5576 CLI.setDebugLoc(getCurSDLoc()).setChain(getRoot())
5577 .setCallee(RetTy, FTy, Callee, std::move(Args), CS).setTailCall(isTailCall);
5579 std::pair<SDValue,SDValue> Result = TLI->LowerCallTo(CLI);
5580 assert((isTailCall || Result.second.getNode()) &&
5581 "Non-null chain expected with non-tail call!");
5582 assert((Result.second.getNode() || !Result.first.getNode()) &&
5583 "Null value expected with tail call!");
5584 if (Result.first.getNode())
5585 setValue(CS.getInstruction(), Result.first);
5587 if (!Result.second.getNode()) {
5588 // As a special case, a null chain means that a tail call has been emitted
5589 // and the DAG root is already updated.
5592 // Since there's no actual continuation from this block, nothing can be
5593 // relying on us setting vregs for them.
5594 PendingExports.clear();
5596 DAG.setRoot(Result.second);
5600 // Insert a label at the end of the invoke call to mark the try range. This
5601 // can be used to detect deletion of the invoke via the MachineModuleInfo.
5602 MCSymbol *EndLabel = MMI.getContext().CreateTempSymbol();
5603 DAG.setRoot(DAG.getEHLabel(getCurSDLoc(), getRoot(), EndLabel));
5605 // Inform MachineModuleInfo of range.
5606 MMI.addInvoke(LandingPad, BeginLabel, EndLabel);
5610 /// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the
5611 /// value is equal or not-equal to zero.
5612 static bool IsOnlyUsedInZeroEqualityComparison(const Value *V) {
5613 for (const User *U : V->users()) {
5614 if (const ICmpInst *IC = dyn_cast<ICmpInst>(U))
5615 if (IC->isEquality())
5616 if (const Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
5617 if (C->isNullValue())
5619 // Unknown instruction.
5625 static SDValue getMemCmpLoad(const Value *PtrVal, MVT LoadVT,
5627 SelectionDAGBuilder &Builder) {
5629 // Check to see if this load can be trivially constant folded, e.g. if the
5630 // input is from a string literal.
5631 if (const Constant *LoadInput = dyn_cast<Constant>(PtrVal)) {
5632 // Cast pointer to the type we really want to load.
5633 LoadInput = ConstantExpr::getBitCast(const_cast<Constant *>(LoadInput),
5634 PointerType::getUnqual(LoadTy));
5636 if (const Constant *LoadCst =
5637 ConstantFoldLoadFromConstPtr(const_cast<Constant *>(LoadInput),
5639 return Builder.getValue(LoadCst);
5642 // Otherwise, we have to emit the load. If the pointer is to unfoldable but
5643 // still constant memory, the input chain can be the entry node.
5645 bool ConstantMemory = false;
5647 // Do not serialize (non-volatile) loads of constant memory with anything.
5648 if (Builder.AA->pointsToConstantMemory(PtrVal)) {
5649 Root = Builder.DAG.getEntryNode();
5650 ConstantMemory = true;
5652 // Do not serialize non-volatile loads against each other.
5653 Root = Builder.DAG.getRoot();
5656 SDValue Ptr = Builder.getValue(PtrVal);
5657 SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurSDLoc(), Root,
5658 Ptr, MachinePointerInfo(PtrVal),
5660 false /*nontemporal*/,
5661 false /*isinvariant*/, 1 /* align=1 */);
5663 if (!ConstantMemory)
5664 Builder.PendingLoads.push_back(LoadVal.getValue(1));
5668 /// processIntegerCallValue - Record the value for an instruction that
5669 /// produces an integer result, converting the type where necessary.
5670 void SelectionDAGBuilder::processIntegerCallValue(const Instruction &I,
5673 EVT VT = TM.getSubtargetImpl()->getTargetLowering()->getValueType(I.getType(),
5676 Value = DAG.getSExtOrTrunc(Value, getCurSDLoc(), VT);
5678 Value = DAG.getZExtOrTrunc(Value, getCurSDLoc(), VT);
5679 setValue(&I, Value);
5682 /// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form.
5683 /// If so, return true and lower it, otherwise return false and it will be
5684 /// lowered like a normal call.
5685 bool SelectionDAGBuilder::visitMemCmpCall(const CallInst &I) {
5686 // Verify that the prototype makes sense. int memcmp(void*,void*,size_t)
5687 if (I.getNumArgOperands() != 3)
5690 const Value *LHS = I.getArgOperand(0), *RHS = I.getArgOperand(1);
5691 if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() ||
5692 !I.getArgOperand(2)->getType()->isIntegerTy() ||
5693 !I.getType()->isIntegerTy())
5696 const Value *Size = I.getArgOperand(2);
5697 const ConstantInt *CSize = dyn_cast<ConstantInt>(Size);
5698 if (CSize && CSize->getZExtValue() == 0) {
5699 EVT CallVT = TM.getSubtargetImpl()->getTargetLowering()->getValueType(
5701 setValue(&I, DAG.getConstant(0, CallVT));
5705 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5706 std::pair<SDValue, SDValue> Res =
5707 TSI.EmitTargetCodeForMemcmp(DAG, getCurSDLoc(), DAG.getRoot(),
5708 getValue(LHS), getValue(RHS), getValue(Size),
5709 MachinePointerInfo(LHS),
5710 MachinePointerInfo(RHS));
5711 if (Res.first.getNode()) {
5712 processIntegerCallValue(I, Res.first, true);
5713 PendingLoads.push_back(Res.second);
5717 // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0
5718 // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0
5719 if (CSize && IsOnlyUsedInZeroEqualityComparison(&I)) {
5720 bool ActuallyDoIt = true;
5723 switch (CSize->getZExtValue()) {
5725 LoadVT = MVT::Other;
5727 ActuallyDoIt = false;
5731 LoadTy = Type::getInt16Ty(CSize->getContext());
5735 LoadTy = Type::getInt32Ty(CSize->getContext());
5739 LoadTy = Type::getInt64Ty(CSize->getContext());
5743 LoadVT = MVT::v4i32;
5744 LoadTy = Type::getInt32Ty(CSize->getContext());
5745 LoadTy = VectorType::get(LoadTy, 4);
5750 // This turns into unaligned loads. We only do this if the target natively
5751 // supports the MVT we'll be loading or if it is small enough (<= 4) that
5752 // we'll only produce a small number of byte loads.
5754 // Require that we can find a legal MVT, and only do this if the target
5755 // supports unaligned loads of that type. Expanding into byte loads would
5757 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
5758 if (ActuallyDoIt && CSize->getZExtValue() > 4) {
5759 unsigned DstAS = LHS->getType()->getPointerAddressSpace();
5760 unsigned SrcAS = RHS->getType()->getPointerAddressSpace();
5761 // TODO: Handle 5 byte compare as 4-byte + 1 byte.
5762 // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads.
5763 // TODO: Check alignment of src and dest ptrs.
5764 if (!TLI->isTypeLegal(LoadVT) ||
5765 !TLI->allowsMisalignedMemoryAccesses(LoadVT, SrcAS) ||
5766 !TLI->allowsMisalignedMemoryAccesses(LoadVT, DstAS))
5767 ActuallyDoIt = false;
5771 SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this);
5772 SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this);
5774 SDValue Res = DAG.getSetCC(getCurSDLoc(), MVT::i1, LHSVal, RHSVal,
5776 processIntegerCallValue(I, Res, false);
5785 /// visitMemChrCall -- See if we can lower a memchr call into an optimized
5786 /// form. If so, return true and lower it, otherwise return false and it
5787 /// will be lowered like a normal call.
5788 bool SelectionDAGBuilder::visitMemChrCall(const CallInst &I) {
5789 // Verify that the prototype makes sense. void *memchr(void *, int, size_t)
5790 if (I.getNumArgOperands() != 3)
5793 const Value *Src = I.getArgOperand(0);
5794 const Value *Char = I.getArgOperand(1);
5795 const Value *Length = I.getArgOperand(2);
5796 if (!Src->getType()->isPointerTy() ||
5797 !Char->getType()->isIntegerTy() ||
5798 !Length->getType()->isIntegerTy() ||
5799 !I.getType()->isPointerTy())
5802 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5803 std::pair<SDValue, SDValue> Res =
5804 TSI.EmitTargetCodeForMemchr(DAG, getCurSDLoc(), DAG.getRoot(),
5805 getValue(Src), getValue(Char), getValue(Length),
5806 MachinePointerInfo(Src));
5807 if (Res.first.getNode()) {
5808 setValue(&I, Res.first);
5809 PendingLoads.push_back(Res.second);
5816 /// visitStrCpyCall -- See if we can lower a strcpy or stpcpy call into an
5817 /// optimized form. If so, return true and lower it, otherwise return false
5818 /// and it will be lowered like a normal call.
5819 bool SelectionDAGBuilder::visitStrCpyCall(const CallInst &I, bool isStpcpy) {
5820 // Verify that the prototype makes sense. char *strcpy(char *, char *)
5821 if (I.getNumArgOperands() != 2)
5824 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5825 if (!Arg0->getType()->isPointerTy() ||
5826 !Arg1->getType()->isPointerTy() ||
5827 !I.getType()->isPointerTy())
5830 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5831 std::pair<SDValue, SDValue> Res =
5832 TSI.EmitTargetCodeForStrcpy(DAG, getCurSDLoc(), getRoot(),
5833 getValue(Arg0), getValue(Arg1),
5834 MachinePointerInfo(Arg0),
5835 MachinePointerInfo(Arg1), isStpcpy);
5836 if (Res.first.getNode()) {
5837 setValue(&I, Res.first);
5838 DAG.setRoot(Res.second);
5845 /// visitStrCmpCall - See if we can lower a call to strcmp in an optimized form.
5846 /// If so, return true and lower it, otherwise return false and it will be
5847 /// lowered like a normal call.
5848 bool SelectionDAGBuilder::visitStrCmpCall(const CallInst &I) {
5849 // Verify that the prototype makes sense. int strcmp(void*,void*)
5850 if (I.getNumArgOperands() != 2)
5853 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5854 if (!Arg0->getType()->isPointerTy() ||
5855 !Arg1->getType()->isPointerTy() ||
5856 !I.getType()->isIntegerTy())
5859 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5860 std::pair<SDValue, SDValue> Res =
5861 TSI.EmitTargetCodeForStrcmp(DAG, getCurSDLoc(), DAG.getRoot(),
5862 getValue(Arg0), getValue(Arg1),
5863 MachinePointerInfo(Arg0),
5864 MachinePointerInfo(Arg1));
5865 if (Res.first.getNode()) {
5866 processIntegerCallValue(I, Res.first, true);
5867 PendingLoads.push_back(Res.second);
5874 /// visitStrLenCall -- See if we can lower a strlen call into an optimized
5875 /// form. If so, return true and lower it, otherwise return false and it
5876 /// will be lowered like a normal call.
5877 bool SelectionDAGBuilder::visitStrLenCall(const CallInst &I) {
5878 // Verify that the prototype makes sense. size_t strlen(char *)
5879 if (I.getNumArgOperands() != 1)
5882 const Value *Arg0 = I.getArgOperand(0);
5883 if (!Arg0->getType()->isPointerTy() || !I.getType()->isIntegerTy())
5886 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5887 std::pair<SDValue, SDValue> Res =
5888 TSI.EmitTargetCodeForStrlen(DAG, getCurSDLoc(), DAG.getRoot(),
5889 getValue(Arg0), MachinePointerInfo(Arg0));
5890 if (Res.first.getNode()) {
5891 processIntegerCallValue(I, Res.first, false);
5892 PendingLoads.push_back(Res.second);
5899 /// visitStrNLenCall -- See if we can lower a strnlen call into an optimized
5900 /// form. If so, return true and lower it, otherwise return false and it
5901 /// will be lowered like a normal call.
5902 bool SelectionDAGBuilder::visitStrNLenCall(const CallInst &I) {
5903 // Verify that the prototype makes sense. size_t strnlen(char *, size_t)
5904 if (I.getNumArgOperands() != 2)
5907 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5908 if (!Arg0->getType()->isPointerTy() ||
5909 !Arg1->getType()->isIntegerTy() ||
5910 !I.getType()->isIntegerTy())
5913 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5914 std::pair<SDValue, SDValue> Res =
5915 TSI.EmitTargetCodeForStrnlen(DAG, getCurSDLoc(), DAG.getRoot(),
5916 getValue(Arg0), getValue(Arg1),
5917 MachinePointerInfo(Arg0));
5918 if (Res.first.getNode()) {
5919 processIntegerCallValue(I, Res.first, false);
5920 PendingLoads.push_back(Res.second);
5927 /// visitUnaryFloatCall - If a call instruction is a unary floating-point
5928 /// operation (as expected), translate it to an SDNode with the specified opcode
5929 /// and return true.
5930 bool SelectionDAGBuilder::visitUnaryFloatCall(const CallInst &I,
5932 // Sanity check that it really is a unary floating-point call.
5933 if (I.getNumArgOperands() != 1 ||
5934 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
5935 I.getType() != I.getArgOperand(0)->getType() ||
5936 !I.onlyReadsMemory())
5939 SDValue Tmp = getValue(I.getArgOperand(0));
5940 setValue(&I, DAG.getNode(Opcode, getCurSDLoc(), Tmp.getValueType(), Tmp));
5944 void SelectionDAGBuilder::visitCall(const CallInst &I) {
5945 // Handle inline assembly differently.
5946 if (isa<InlineAsm>(I.getCalledValue())) {
5951 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5952 ComputeUsesVAFloatArgument(I, &MMI);
5954 const char *RenameFn = nullptr;
5955 if (Function *F = I.getCalledFunction()) {
5956 if (F->isDeclaration()) {
5957 if (const TargetIntrinsicInfo *II = TM.getIntrinsicInfo()) {
5958 if (unsigned IID = II->getIntrinsicID(F)) {
5959 RenameFn = visitIntrinsicCall(I, IID);
5964 if (unsigned IID = F->getIntrinsicID()) {
5965 RenameFn = visitIntrinsicCall(I, IID);
5971 // Check for well-known libc/libm calls. If the function is internal, it
5972 // can't be a library call.
5974 if (!F->hasLocalLinkage() && F->hasName() &&
5975 LibInfo->getLibFunc(F->getName(), Func) &&
5976 LibInfo->hasOptimizedCodeGen(Func)) {
5979 case LibFunc::copysign:
5980 case LibFunc::copysignf:
5981 case LibFunc::copysignl:
5982 if (I.getNumArgOperands() == 2 && // Basic sanity checks.
5983 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5984 I.getType() == I.getArgOperand(0)->getType() &&
5985 I.getType() == I.getArgOperand(1)->getType() &&
5986 I.onlyReadsMemory()) {
5987 SDValue LHS = getValue(I.getArgOperand(0));
5988 SDValue RHS = getValue(I.getArgOperand(1));
5989 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurSDLoc(),
5990 LHS.getValueType(), LHS, RHS));
5995 case LibFunc::fabsf:
5996 case LibFunc::fabsl:
5997 if (visitUnaryFloatCall(I, ISD::FABS))
6003 if (visitUnaryFloatCall(I, ISD::FSIN))
6009 if (visitUnaryFloatCall(I, ISD::FCOS))
6013 case LibFunc::sqrtf:
6014 case LibFunc::sqrtl:
6015 case LibFunc::sqrt_finite:
6016 case LibFunc::sqrtf_finite:
6017 case LibFunc::sqrtl_finite:
6018 if (visitUnaryFloatCall(I, ISD::FSQRT))
6021 case LibFunc::floor:
6022 case LibFunc::floorf:
6023 case LibFunc::floorl:
6024 if (visitUnaryFloatCall(I, ISD::FFLOOR))
6027 case LibFunc::nearbyint:
6028 case LibFunc::nearbyintf:
6029 case LibFunc::nearbyintl:
6030 if (visitUnaryFloatCall(I, ISD::FNEARBYINT))
6034 case LibFunc::ceilf:
6035 case LibFunc::ceill:
6036 if (visitUnaryFloatCall(I, ISD::FCEIL))
6040 case LibFunc::rintf:
6041 case LibFunc::rintl:
6042 if (visitUnaryFloatCall(I, ISD::FRINT))
6045 case LibFunc::round:
6046 case LibFunc::roundf:
6047 case LibFunc::roundl:
6048 if (visitUnaryFloatCall(I, ISD::FROUND))
6051 case LibFunc::trunc:
6052 case LibFunc::truncf:
6053 case LibFunc::truncl:
6054 if (visitUnaryFloatCall(I, ISD::FTRUNC))
6058 case LibFunc::log2f:
6059 case LibFunc::log2l:
6060 if (visitUnaryFloatCall(I, ISD::FLOG2))
6064 case LibFunc::exp2f:
6065 case LibFunc::exp2l:
6066 if (visitUnaryFloatCall(I, ISD::FEXP2))
6069 case LibFunc::memcmp:
6070 if (visitMemCmpCall(I))
6073 case LibFunc::memchr:
6074 if (visitMemChrCall(I))
6077 case LibFunc::strcpy:
6078 if (visitStrCpyCall(I, false))
6081 case LibFunc::stpcpy:
6082 if (visitStrCpyCall(I, true))
6085 case LibFunc::strcmp:
6086 if (visitStrCmpCall(I))
6089 case LibFunc::strlen:
6090 if (visitStrLenCall(I))
6093 case LibFunc::strnlen:
6094 if (visitStrNLenCall(I))
6103 Callee = getValue(I.getCalledValue());
6105 Callee = DAG.getExternalSymbol(
6106 RenameFn, TM.getSubtargetImpl()->getTargetLowering()->getPointerTy());
6108 // Check if we can potentially perform a tail call. More detailed checking is
6109 // be done within LowerCallTo, after more information about the call is known.
6110 LowerCallTo(&I, Callee, I.isTailCall());
6115 /// AsmOperandInfo - This contains information for each constraint that we are
6117 class SDISelAsmOperandInfo : public TargetLowering::AsmOperandInfo {
6119 /// CallOperand - If this is the result output operand or a clobber
6120 /// this is null, otherwise it is the incoming operand to the CallInst.
6121 /// This gets modified as the asm is processed.
6122 SDValue CallOperand;
6124 /// AssignedRegs - If this is a register or register class operand, this
6125 /// contains the set of register corresponding to the operand.
6126 RegsForValue AssignedRegs;
6128 explicit SDISelAsmOperandInfo(const TargetLowering::AsmOperandInfo &info)
6129 : TargetLowering::AsmOperandInfo(info), CallOperand(nullptr,0) {
6132 /// getCallOperandValEVT - Return the EVT of the Value* that this operand
6133 /// corresponds to. If there is no Value* for this operand, it returns
6135 EVT getCallOperandValEVT(LLVMContext &Context,
6136 const TargetLowering &TLI,
6137 const DataLayout *DL) const {
6138 if (!CallOperandVal) return MVT::Other;
6140 if (isa<BasicBlock>(CallOperandVal))
6141 return TLI.getPointerTy();
6143 llvm::Type *OpTy = CallOperandVal->getType();
6145 // FIXME: code duplicated from TargetLowering::ParseConstraints().
6146 // If this is an indirect operand, the operand is a pointer to the
6149 llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
6151 report_fatal_error("Indirect operand for inline asm not a pointer!");
6152 OpTy = PtrTy->getElementType();
6155 // Look for vector wrapped in a struct. e.g. { <16 x i8> }.
6156 if (StructType *STy = dyn_cast<StructType>(OpTy))
6157 if (STy->getNumElements() == 1)
6158 OpTy = STy->getElementType(0);
6160 // If OpTy is not a single value, it may be a struct/union that we
6161 // can tile with integers.
6162 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
6163 unsigned BitSize = DL->getTypeSizeInBits(OpTy);
6172 OpTy = IntegerType::get(Context, BitSize);
6177 return TLI.getValueType(OpTy, true);
6181 typedef SmallVector<SDISelAsmOperandInfo,16> SDISelAsmOperandInfoVector;
6183 } // end anonymous namespace
6185 /// GetRegistersForValue - Assign registers (virtual or physical) for the
6186 /// specified operand. We prefer to assign virtual registers, to allow the
6187 /// register allocator to handle the assignment process. However, if the asm
6188 /// uses features that we can't model on machineinstrs, we have SDISel do the
6189 /// allocation. This produces generally horrible, but correct, code.
6191 /// OpInfo describes the operand.
6193 static void GetRegistersForValue(SelectionDAG &DAG,
6194 const TargetLowering &TLI,
6196 SDISelAsmOperandInfo &OpInfo) {
6197 LLVMContext &Context = *DAG.getContext();
6199 MachineFunction &MF = DAG.getMachineFunction();
6200 SmallVector<unsigned, 4> Regs;
6202 // If this is a constraint for a single physreg, or a constraint for a
6203 // register class, find it.
6204 std::pair<unsigned, const TargetRegisterClass*> PhysReg =
6205 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
6206 OpInfo.ConstraintVT);
6208 unsigned NumRegs = 1;
6209 if (OpInfo.ConstraintVT != MVT::Other) {
6210 // If this is a FP input in an integer register (or visa versa) insert a bit
6211 // cast of the input value. More generally, handle any case where the input
6212 // value disagrees with the register class we plan to stick this in.
6213 if (OpInfo.Type == InlineAsm::isInput &&
6214 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) {
6215 // Try to convert to the first EVT that the reg class contains. If the
6216 // types are identical size, use a bitcast to convert (e.g. two differing
6218 MVT RegVT = *PhysReg.second->vt_begin();
6219 if (RegVT.getSizeInBits() == OpInfo.CallOperand.getValueSizeInBits()) {
6220 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
6221 RegVT, OpInfo.CallOperand);
6222 OpInfo.ConstraintVT = RegVT;
6223 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) {
6224 // If the input is a FP value and we want it in FP registers, do a
6225 // bitcast to the corresponding integer type. This turns an f64 value
6226 // into i64, which can be passed with two i32 values on a 32-bit
6228 RegVT = MVT::getIntegerVT(OpInfo.ConstraintVT.getSizeInBits());
6229 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
6230 RegVT, OpInfo.CallOperand);
6231 OpInfo.ConstraintVT = RegVT;
6235 NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT);
6239 EVT ValueVT = OpInfo.ConstraintVT;
6241 // If this is a constraint for a specific physical register, like {r17},
6243 if (unsigned AssignedReg = PhysReg.first) {
6244 const TargetRegisterClass *RC = PhysReg.second;
6245 if (OpInfo.ConstraintVT == MVT::Other)
6246 ValueVT = *RC->vt_begin();
6248 // Get the actual register value type. This is important, because the user
6249 // may have asked for (e.g.) the AX register in i32 type. We need to
6250 // remember that AX is actually i16 to get the right extension.
6251 RegVT = *RC->vt_begin();
6253 // This is a explicit reference to a physical register.
6254 Regs.push_back(AssignedReg);
6256 // If this is an expanded reference, add the rest of the regs to Regs.
6258 TargetRegisterClass::iterator I = RC->begin();
6259 for (; *I != AssignedReg; ++I)
6260 assert(I != RC->end() && "Didn't find reg!");
6262 // Already added the first reg.
6264 for (; NumRegs; --NumRegs, ++I) {
6265 assert(I != RC->end() && "Ran out of registers to allocate!");
6270 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
6274 // Otherwise, if this was a reference to an LLVM register class, create vregs
6275 // for this reference.
6276 if (const TargetRegisterClass *RC = PhysReg.second) {
6277 RegVT = *RC->vt_begin();
6278 if (OpInfo.ConstraintVT == MVT::Other)
6281 // Create the appropriate number of virtual registers.
6282 MachineRegisterInfo &RegInfo = MF.getRegInfo();
6283 for (; NumRegs; --NumRegs)
6284 Regs.push_back(RegInfo.createVirtualRegister(RC));
6286 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
6290 // Otherwise, we couldn't allocate enough registers for this.
6293 /// visitInlineAsm - Handle a call to an InlineAsm object.
6295 void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) {
6296 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
6298 /// ConstraintOperands - Information about all of the constraints.
6299 SDISelAsmOperandInfoVector ConstraintOperands;
6301 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
6302 TargetLowering::AsmOperandInfoVector
6303 TargetConstraints = TLI->ParseConstraints(CS);
6305 bool hasMemory = false;
6307 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
6308 unsigned ResNo = 0; // ResNo - The result number of the next output.
6309 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
6310 ConstraintOperands.push_back(SDISelAsmOperandInfo(TargetConstraints[i]));
6311 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
6313 MVT OpVT = MVT::Other;
6315 // Compute the value type for each operand.
6316 switch (OpInfo.Type) {
6317 case InlineAsm::isOutput:
6318 // Indirect outputs just consume an argument.
6319 if (OpInfo.isIndirect) {
6320 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
6324 // The return value of the call is this value. As such, there is no
6325 // corresponding argument.
6326 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
6327 if (StructType *STy = dyn_cast<StructType>(CS.getType())) {
6328 OpVT = TLI->getSimpleValueType(STy->getElementType(ResNo));
6330 assert(ResNo == 0 && "Asm only has one result!");
6331 OpVT = TLI->getSimpleValueType(CS.getType());
6335 case InlineAsm::isInput:
6336 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
6338 case InlineAsm::isClobber:
6343 // If this is an input or an indirect output, process the call argument.
6344 // BasicBlocks are labels, currently appearing only in asm's.
6345 if (OpInfo.CallOperandVal) {
6346 if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) {
6347 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]);
6349 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
6352 OpVT = OpInfo.getCallOperandValEVT(*DAG.getContext(), *TLI, DL).
6356 OpInfo.ConstraintVT = OpVT;
6358 // Indirect operand accesses access memory.
6359 if (OpInfo.isIndirect)
6362 for (unsigned j = 0, ee = OpInfo.Codes.size(); j != ee; ++j) {
6363 TargetLowering::ConstraintType
6364 CType = TLI->getConstraintType(OpInfo.Codes[j]);
6365 if (CType == TargetLowering::C_Memory) {
6373 SDValue Chain, Flag;
6375 // We won't need to flush pending loads if this asm doesn't touch
6376 // memory and is nonvolatile.
6377 if (hasMemory || IA->hasSideEffects())
6380 Chain = DAG.getRoot();
6382 // Second pass over the constraints: compute which constraint option to use
6383 // and assign registers to constraints that want a specific physreg.
6384 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6385 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6387 // If this is an output operand with a matching input operand, look up the
6388 // matching input. If their types mismatch, e.g. one is an integer, the
6389 // other is floating point, or their sizes are different, flag it as an
6391 if (OpInfo.hasMatchingInput()) {
6392 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
6394 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
6395 std::pair<unsigned, const TargetRegisterClass*> MatchRC =
6396 TLI->getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
6397 OpInfo.ConstraintVT);
6398 std::pair<unsigned, const TargetRegisterClass*> InputRC =
6399 TLI->getRegForInlineAsmConstraint(Input.ConstraintCode,
6400 Input.ConstraintVT);
6401 if ((OpInfo.ConstraintVT.isInteger() !=
6402 Input.ConstraintVT.isInteger()) ||
6403 (MatchRC.second != InputRC.second)) {
6404 report_fatal_error("Unsupported asm: input constraint"
6405 " with a matching output constraint of"
6406 " incompatible type!");
6408 Input.ConstraintVT = OpInfo.ConstraintVT;
6412 // Compute the constraint code and ConstraintType to use.
6413 TLI->ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, &DAG);
6415 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
6416 OpInfo.Type == InlineAsm::isClobber)
6419 // If this is a memory input, and if the operand is not indirect, do what we
6420 // need to to provide an address for the memory input.
6421 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
6422 !OpInfo.isIndirect) {
6423 assert((OpInfo.isMultipleAlternative ||
6424 (OpInfo.Type == InlineAsm::isInput)) &&
6425 "Can only indirectify direct input operands!");
6427 // Memory operands really want the address of the value. If we don't have
6428 // an indirect input, put it in the constpool if we can, otherwise spill
6429 // it to a stack slot.
6430 // TODO: This isn't quite right. We need to handle these according to
6431 // the addressing mode that the constraint wants. Also, this may take
6432 // an additional register for the computation and we don't want that
6435 // If the operand is a float, integer, or vector constant, spill to a
6436 // constant pool entry to get its address.
6437 const Value *OpVal = OpInfo.CallOperandVal;
6438 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
6439 isa<ConstantVector>(OpVal) || isa<ConstantDataVector>(OpVal)) {
6440 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal),
6441 TLI->getPointerTy());
6443 // Otherwise, create a stack slot and emit a store to it before the
6445 Type *Ty = OpVal->getType();
6446 uint64_t TySize = TLI->getDataLayout()->getTypeAllocSize(Ty);
6447 unsigned Align = TLI->getDataLayout()->getPrefTypeAlignment(Ty);
6448 MachineFunction &MF = DAG.getMachineFunction();
6449 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
6450 SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI->getPointerTy());
6451 Chain = DAG.getStore(Chain, getCurSDLoc(),
6452 OpInfo.CallOperand, StackSlot,
6453 MachinePointerInfo::getFixedStack(SSFI),
6455 OpInfo.CallOperand = StackSlot;
6458 // There is no longer a Value* corresponding to this operand.
6459 OpInfo.CallOperandVal = nullptr;
6461 // It is now an indirect operand.
6462 OpInfo.isIndirect = true;
6465 // If this constraint is for a specific register, allocate it before
6467 if (OpInfo.ConstraintType == TargetLowering::C_Register)
6468 GetRegistersForValue(DAG, *TLI, getCurSDLoc(), OpInfo);
6471 // Second pass - Loop over all of the operands, assigning virtual or physregs
6472 // to register class operands.
6473 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6474 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6476 // C_Register operands have already been allocated, Other/Memory don't need
6478 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
6479 GetRegistersForValue(DAG, *TLI, getCurSDLoc(), OpInfo);
6482 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
6483 std::vector<SDValue> AsmNodeOperands;
6484 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain
6485 AsmNodeOperands.push_back(
6486 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(),
6487 TLI->getPointerTy()));
6489 // If we have a !srcloc metadata node associated with it, we want to attach
6490 // this to the ultimately generated inline asm machineinstr. To do this, we
6491 // pass in the third operand as this (potentially null) inline asm MDNode.
6492 const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc");
6493 AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc));
6495 // Remember the HasSideEffect, AlignStack, AsmDialect, MayLoad and MayStore
6496 // bits as operand 3.
6497 unsigned ExtraInfo = 0;
6498 if (IA->hasSideEffects())
6499 ExtraInfo |= InlineAsm::Extra_HasSideEffects;
6500 if (IA->isAlignStack())
6501 ExtraInfo |= InlineAsm::Extra_IsAlignStack;
6502 // Set the asm dialect.
6503 ExtraInfo |= IA->getDialect() * InlineAsm::Extra_AsmDialect;
6505 // Determine if this InlineAsm MayLoad or MayStore based on the constraints.
6506 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
6507 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
6509 // Compute the constraint code and ConstraintType to use.
6510 TLI->ComputeConstraintToUse(OpInfo, SDValue());
6512 // Ideally, we would only check against memory constraints. However, the
6513 // meaning of an other constraint can be target-specific and we can't easily
6514 // reason about it. Therefore, be conservative and set MayLoad/MayStore
6515 // for other constriants as well.
6516 if (OpInfo.ConstraintType == TargetLowering::C_Memory ||
6517 OpInfo.ConstraintType == TargetLowering::C_Other) {
6518 if (OpInfo.Type == InlineAsm::isInput)
6519 ExtraInfo |= InlineAsm::Extra_MayLoad;
6520 else if (OpInfo.Type == InlineAsm::isOutput)
6521 ExtraInfo |= InlineAsm::Extra_MayStore;
6522 else if (OpInfo.Type == InlineAsm::isClobber)
6523 ExtraInfo |= (InlineAsm::Extra_MayLoad | InlineAsm::Extra_MayStore);
6527 AsmNodeOperands.push_back(DAG.getTargetConstant(ExtraInfo,
6528 TLI->getPointerTy()));
6530 // Loop over all of the inputs, copying the operand values into the
6531 // appropriate registers and processing the output regs.
6532 RegsForValue RetValRegs;
6534 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
6535 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
6537 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6538 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6540 switch (OpInfo.Type) {
6541 case InlineAsm::isOutput: {
6542 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
6543 OpInfo.ConstraintType != TargetLowering::C_Register) {
6544 // Memory output, or 'other' output (e.g. 'X' constraint).
6545 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
6547 // Add information to the INLINEASM node to know about this output.
6548 unsigned OpFlags = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
6549 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlags,
6550 TLI->getPointerTy()));
6551 AsmNodeOperands.push_back(OpInfo.CallOperand);
6555 // Otherwise, this is a register or register class output.
6557 // Copy the output from the appropriate register. Find a register that
6559 if (OpInfo.AssignedRegs.Regs.empty()) {
6560 LLVMContext &Ctx = *DAG.getContext();
6561 Ctx.emitError(CS.getInstruction(),
6562 "couldn't allocate output register for constraint '" +
6563 Twine(OpInfo.ConstraintCode) + "'");
6567 // If this is an indirect operand, store through the pointer after the
6569 if (OpInfo.isIndirect) {
6570 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
6571 OpInfo.CallOperandVal));
6573 // This is the result value of the call.
6574 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
6575 // Concatenate this output onto the outputs list.
6576 RetValRegs.append(OpInfo.AssignedRegs);
6579 // Add information to the INLINEASM node to know that this register is
6582 .AddInlineAsmOperands(OpInfo.isEarlyClobber
6583 ? InlineAsm::Kind_RegDefEarlyClobber
6584 : InlineAsm::Kind_RegDef,
6585 false, 0, DAG, AsmNodeOperands);
6588 case InlineAsm::isInput: {
6589 SDValue InOperandVal = OpInfo.CallOperand;
6591 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint?
6592 // If this is required to match an output register we have already set,
6593 // just use its register.
6594 unsigned OperandNo = OpInfo.getMatchedOperand();
6596 // Scan until we find the definition we already emitted of this operand.
6597 // When we find it, create a RegsForValue operand.
6598 unsigned CurOp = InlineAsm::Op_FirstOperand;
6599 for (; OperandNo; --OperandNo) {
6600 // Advance to the next operand.
6602 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6603 assert((InlineAsm::isRegDefKind(OpFlag) ||
6604 InlineAsm::isRegDefEarlyClobberKind(OpFlag) ||
6605 InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?");
6606 CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1;
6610 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6611 if (InlineAsm::isRegDefKind(OpFlag) ||
6612 InlineAsm::isRegDefEarlyClobberKind(OpFlag)) {
6613 // Add (OpFlag&0xffff)>>3 registers to MatchedRegs.
6614 if (OpInfo.isIndirect) {
6615 // This happens on gcc/testsuite/gcc.dg/pr8788-1.c
6616 LLVMContext &Ctx = *DAG.getContext();
6617 Ctx.emitError(CS.getInstruction(), "inline asm not supported yet:"
6618 " don't know how to handle tied "
6619 "indirect register inputs");
6623 RegsForValue MatchedRegs;
6624 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
6625 MVT RegVT = AsmNodeOperands[CurOp+1].getSimpleValueType();
6626 MatchedRegs.RegVTs.push_back(RegVT);
6627 MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo();
6628 for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag);
6630 if (const TargetRegisterClass *RC = TLI->getRegClassFor(RegVT))
6631 MatchedRegs.Regs.push_back(RegInfo.createVirtualRegister(RC));
6633 LLVMContext &Ctx = *DAG.getContext();
6634 Ctx.emitError(CS.getInstruction(),
6635 "inline asm error: This value"
6636 " type register class is not natively supported!");
6640 // Use the produced MatchedRegs object to
6641 MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurSDLoc(),
6642 Chain, &Flag, CS.getInstruction());
6643 MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse,
6644 true, OpInfo.getMatchedOperand(),
6645 DAG, AsmNodeOperands);
6649 assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!");
6650 assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 &&
6651 "Unexpected number of operands");
6652 // Add information to the INLINEASM node to know about this input.
6653 // See InlineAsm.h isUseOperandTiedToDef.
6654 OpFlag = InlineAsm::getFlagWordForMatchingOp(OpFlag,
6655 OpInfo.getMatchedOperand());
6656 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlag,
6657 TLI->getPointerTy()));
6658 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]);
6662 // Treat indirect 'X' constraint as memory.
6663 if (OpInfo.ConstraintType == TargetLowering::C_Other &&
6665 OpInfo.ConstraintType = TargetLowering::C_Memory;
6667 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
6668 std::vector<SDValue> Ops;
6669 TLI->LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode,
6672 LLVMContext &Ctx = *DAG.getContext();
6673 Ctx.emitError(CS.getInstruction(),
6674 "invalid operand for inline asm constraint '" +
6675 Twine(OpInfo.ConstraintCode) + "'");
6679 // Add information to the INLINEASM node to know about this input.
6680 unsigned ResOpType =
6681 InlineAsm::getFlagWord(InlineAsm::Kind_Imm, Ops.size());
6682 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
6683 TLI->getPointerTy()));
6684 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
6688 if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
6689 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
6690 assert(InOperandVal.getValueType() == TLI->getPointerTy() &&
6691 "Memory operands expect pointer values");
6693 // Add information to the INLINEASM node to know about this input.
6694 unsigned ResOpType = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
6695 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
6696 TLI->getPointerTy()));
6697 AsmNodeOperands.push_back(InOperandVal);
6701 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
6702 OpInfo.ConstraintType == TargetLowering::C_Register) &&
6703 "Unknown constraint type!");
6705 // TODO: Support this.
6706 if (OpInfo.isIndirect) {
6707 LLVMContext &Ctx = *DAG.getContext();
6708 Ctx.emitError(CS.getInstruction(),
6709 "Don't know how to handle indirect register inputs yet "
6710 "for constraint '" +
6711 Twine(OpInfo.ConstraintCode) + "'");
6715 // Copy the input into the appropriate registers.
6716 if (OpInfo.AssignedRegs.Regs.empty()) {
6717 LLVMContext &Ctx = *DAG.getContext();
6718 Ctx.emitError(CS.getInstruction(),
6719 "couldn't allocate input reg for constraint '" +
6720 Twine(OpInfo.ConstraintCode) + "'");
6724 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, getCurSDLoc(),
6725 Chain, &Flag, CS.getInstruction());
6727 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, false, 0,
6728 DAG, AsmNodeOperands);
6731 case InlineAsm::isClobber: {
6732 // Add the clobbered value to the operand list, so that the register
6733 // allocator is aware that the physreg got clobbered.
6734 if (!OpInfo.AssignedRegs.Regs.empty())
6735 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_Clobber,
6743 // Finish up input operands. Set the input chain and add the flag last.
6744 AsmNodeOperands[InlineAsm::Op_InputChain] = Chain;
6745 if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
6747 Chain = DAG.getNode(ISD::INLINEASM, getCurSDLoc(),
6748 DAG.getVTList(MVT::Other, MVT::Glue), AsmNodeOperands);
6749 Flag = Chain.getValue(1);
6751 // If this asm returns a register value, copy the result from that register
6752 // and set it as the value of the call.
6753 if (!RetValRegs.Regs.empty()) {
6754 SDValue Val = RetValRegs.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(),
6755 Chain, &Flag, CS.getInstruction());
6757 // FIXME: Why don't we do this for inline asms with MRVs?
6758 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) {
6759 EVT ResultType = TLI->getValueType(CS.getType());
6761 // If any of the results of the inline asm is a vector, it may have the
6762 // wrong width/num elts. This can happen for register classes that can
6763 // contain multiple different value types. The preg or vreg allocated may
6764 // not have the same VT as was expected. Convert it to the right type
6765 // with bit_convert.
6766 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) {
6767 Val = DAG.getNode(ISD::BITCAST, getCurSDLoc(),
6770 } else if (ResultType != Val.getValueType() &&
6771 ResultType.isInteger() && Val.getValueType().isInteger()) {
6772 // If a result value was tied to an input value, the computed result may
6773 // have a wider width than the expected result. Extract the relevant
6775 Val = DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), ResultType, Val);
6778 assert(ResultType == Val.getValueType() && "Asm result value mismatch!");
6781 setValue(CS.getInstruction(), Val);
6782 // Don't need to use this as a chain in this case.
6783 if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty())
6787 std::vector<std::pair<SDValue, const Value *> > StoresToEmit;
6789 // Process indirect outputs, first output all of the flagged copies out of
6791 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
6792 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
6793 const Value *Ptr = IndirectStoresToEmit[i].second;
6794 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(),
6796 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
6799 // Emit the non-flagged stores from the physregs.
6800 SmallVector<SDValue, 8> OutChains;
6801 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) {
6802 SDValue Val = DAG.getStore(Chain, getCurSDLoc(),
6803 StoresToEmit[i].first,
6804 getValue(StoresToEmit[i].second),
6805 MachinePointerInfo(StoresToEmit[i].second),
6807 OutChains.push_back(Val);
6810 if (!OutChains.empty())
6811 Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other, OutChains);
6816 void SelectionDAGBuilder::visitVAStart(const CallInst &I) {
6817 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurSDLoc(),
6818 MVT::Other, getRoot(),
6819 getValue(I.getArgOperand(0)),
6820 DAG.getSrcValue(I.getArgOperand(0))));
6823 void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) {
6824 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
6825 const DataLayout &DL = *TLI->getDataLayout();
6826 SDValue V = DAG.getVAArg(TLI->getValueType(I.getType()), getCurSDLoc(),
6827 getRoot(), getValue(I.getOperand(0)),
6828 DAG.getSrcValue(I.getOperand(0)),
6829 DL.getABITypeAlignment(I.getType()));
6831 DAG.setRoot(V.getValue(1));
6834 void SelectionDAGBuilder::visitVAEnd(const CallInst &I) {
6835 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurSDLoc(),
6836 MVT::Other, getRoot(),
6837 getValue(I.getArgOperand(0)),
6838 DAG.getSrcValue(I.getArgOperand(0))));
6841 void SelectionDAGBuilder::visitVACopy(const CallInst &I) {
6842 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurSDLoc(),
6843 MVT::Other, getRoot(),
6844 getValue(I.getArgOperand(0)),
6845 getValue(I.getArgOperand(1)),
6846 DAG.getSrcValue(I.getArgOperand(0)),
6847 DAG.getSrcValue(I.getArgOperand(1))));
6850 /// \brief Lower an argument list according to the target calling convention.
6852 /// \return A tuple of <return-value, token-chain>
6854 /// This is a helper for lowering intrinsics that follow a target calling
6855 /// convention or require stack pointer adjustment. Only a subset of the
6856 /// intrinsic's operands need to participate in the calling convention.
6857 std::pair<SDValue, SDValue>
6858 SelectionDAGBuilder::LowerCallOperands(const CallInst &CI, unsigned ArgIdx,
6859 unsigned NumArgs, SDValue Callee,
6861 TargetLowering::ArgListTy Args;
6862 Args.reserve(NumArgs);
6864 // Populate the argument list.
6865 // Attributes for args start at offset 1, after the return attribute.
6866 ImmutableCallSite CS(&CI);
6867 for (unsigned ArgI = ArgIdx, ArgE = ArgIdx + NumArgs, AttrI = ArgIdx + 1;
6868 ArgI != ArgE; ++ArgI) {
6869 const Value *V = CI.getOperand(ArgI);
6871 assert(!V->getType()->isEmptyTy() && "Empty type passed to intrinsic.");
6873 TargetLowering::ArgListEntry Entry;
6874 Entry.Node = getValue(V);
6875 Entry.Ty = V->getType();
6876 Entry.setAttributes(&CS, AttrI);
6877 Args.push_back(Entry);
6880 Type *retTy = useVoidTy ? Type::getVoidTy(*DAG.getContext()) : CI.getType();
6881 TargetLowering::CallLoweringInfo CLI(DAG);
6882 CLI.setDebugLoc(getCurSDLoc()).setChain(getRoot())
6883 .setCallee(CI.getCallingConv(), retTy, Callee, std::move(Args), NumArgs)
6884 .setDiscardResult(!CI.use_empty());
6886 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
6887 return TLI->LowerCallTo(CLI);
6890 /// \brief Add a stack map intrinsic call's live variable operands to a stackmap
6891 /// or patchpoint target node's operand list.
6893 /// Constants are converted to TargetConstants purely as an optimization to
6894 /// avoid constant materialization and register allocation.
6896 /// FrameIndex operands are converted to TargetFrameIndex so that ISEL does not
6897 /// generate addess computation nodes, and so ExpandISelPseudo can convert the
6898 /// TargetFrameIndex into a DirectMemRefOp StackMap location. This avoids
6899 /// address materialization and register allocation, but may also be required
6900 /// for correctness. If a StackMap (or PatchPoint) intrinsic directly uses an
6901 /// alloca in the entry block, then the runtime may assume that the alloca's
6902 /// StackMap location can be read immediately after compilation and that the
6903 /// location is valid at any point during execution (this is similar to the
6904 /// assumption made by the llvm.gcroot intrinsic). If the alloca's location were
6905 /// only available in a register, then the runtime would need to trap when
6906 /// execution reaches the StackMap in order to read the alloca's location.
6907 static void addStackMapLiveVars(const CallInst &CI, unsigned StartIdx,
6908 SmallVectorImpl<SDValue> &Ops,
6909 SelectionDAGBuilder &Builder) {
6910 for (unsigned i = StartIdx, e = CI.getNumArgOperands(); i != e; ++i) {
6911 SDValue OpVal = Builder.getValue(CI.getArgOperand(i));
6912 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(OpVal)) {
6914 Builder.DAG.getTargetConstant(StackMaps::ConstantOp, MVT::i64));
6916 Builder.DAG.getTargetConstant(C->getSExtValue(), MVT::i64));
6917 } else if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(OpVal)) {
6918 const TargetLowering &TLI = Builder.DAG.getTargetLoweringInfo();
6920 Builder.DAG.getTargetFrameIndex(FI->getIndex(), TLI.getPointerTy()));
6922 Ops.push_back(OpVal);
6926 /// \brief Lower llvm.experimental.stackmap directly to its target opcode.
6927 void SelectionDAGBuilder::visitStackmap(const CallInst &CI) {
6928 // void @llvm.experimental.stackmap(i32 <id>, i32 <numShadowBytes>,
6929 // [live variables...])
6931 assert(CI.getType()->isVoidTy() && "Stackmap cannot return a value.");
6933 SDValue Chain, InFlag, Callee, NullPtr;
6934 SmallVector<SDValue, 32> Ops;
6936 SDLoc DL = getCurSDLoc();
6937 Callee = getValue(CI.getCalledValue());
6938 NullPtr = DAG.getIntPtrConstant(0, true);
6940 // The stackmap intrinsic only records the live variables (the arguemnts
6941 // passed to it) and emits NOPS (if requested). Unlike the patchpoint
6942 // intrinsic, this won't be lowered to a function call. This means we don't
6943 // have to worry about calling conventions and target specific lowering code.
6944 // Instead we perform the call lowering right here.
6946 // chain, flag = CALLSEQ_START(chain, 0)
6947 // chain, flag = STACKMAP(id, nbytes, ..., chain, flag)
6948 // chain, flag = CALLSEQ_END(chain, 0, 0, flag)
6950 Chain = DAG.getCALLSEQ_START(getRoot(), NullPtr, DL);
6951 InFlag = Chain.getValue(1);
6953 // Add the <id> and <numBytes> constants.
6954 SDValue IDVal = getValue(CI.getOperand(PatchPointOpers::IDPos));
6955 Ops.push_back(DAG.getTargetConstant(
6956 cast<ConstantSDNode>(IDVal)->getZExtValue(), MVT::i64));
6957 SDValue NBytesVal = getValue(CI.getOperand(PatchPointOpers::NBytesPos));
6958 Ops.push_back(DAG.getTargetConstant(
6959 cast<ConstantSDNode>(NBytesVal)->getZExtValue(), MVT::i32));
6961 // Push live variables for the stack map.
6962 addStackMapLiveVars(CI, 2, Ops, *this);
6964 // We are not pushing any register mask info here on the operands list,
6965 // because the stackmap doesn't clobber anything.
6967 // Push the chain and the glue flag.
6968 Ops.push_back(Chain);
6969 Ops.push_back(InFlag);
6971 // Create the STACKMAP node.
6972 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6973 SDNode *SM = DAG.getMachineNode(TargetOpcode::STACKMAP, DL, NodeTys, Ops);
6974 Chain = SDValue(SM, 0);
6975 InFlag = Chain.getValue(1);
6977 Chain = DAG.getCALLSEQ_END(Chain, NullPtr, NullPtr, InFlag, DL);
6979 // Stackmaps don't generate values, so nothing goes into the NodeMap.
6981 // Set the root to the target-lowered call chain.
6984 // Inform the Frame Information that we have a stackmap in this function.
6985 FuncInfo.MF->getFrameInfo()->setHasStackMap();
6988 /// \brief Lower llvm.experimental.patchpoint directly to its target opcode.
6989 void SelectionDAGBuilder::visitPatchpoint(const CallInst &CI) {
6990 // void|i64 @llvm.experimental.patchpoint.void|i64(i64 <id>,
6995 // [live variables...])
6997 CallingConv::ID CC = CI.getCallingConv();
6998 bool isAnyRegCC = CC == CallingConv::AnyReg;
6999 bool hasDef = !CI.getType()->isVoidTy();
7000 SDValue Callee = getValue(CI.getOperand(2)); // <target>
7002 // Get the real number of arguments participating in the call <numArgs>
7003 SDValue NArgVal = getValue(CI.getArgOperand(PatchPointOpers::NArgPos));
7004 unsigned NumArgs = cast<ConstantSDNode>(NArgVal)->getZExtValue();
7006 // Skip the four meta args: <id>, <numNopBytes>, <target>, <numArgs>
7007 // Intrinsics include all meta-operands up to but not including CC.
7008 unsigned NumMetaOpers = PatchPointOpers::CCPos;
7009 assert(CI.getNumArgOperands() >= NumMetaOpers + NumArgs &&
7010 "Not enough arguments provided to the patchpoint intrinsic");
7012 // For AnyRegCC the arguments are lowered later on manually.
7013 unsigned NumCallArgs = isAnyRegCC ? 0 : NumArgs;
7014 std::pair<SDValue, SDValue> Result =
7015 LowerCallOperands(CI, NumMetaOpers, NumCallArgs, Callee, isAnyRegCC);
7017 // Set the root to the target-lowered call chain.
7018 SDValue Chain = Result.second;
7021 SDNode *CallEnd = Chain.getNode();
7022 if (hasDef && (CallEnd->getOpcode() == ISD::CopyFromReg))
7023 CallEnd = CallEnd->getOperand(0).getNode();
7025 /// Get a call instruction from the call sequence chain.
7026 /// Tail calls are not allowed.
7027 assert(CallEnd->getOpcode() == ISD::CALLSEQ_END &&
7028 "Expected a callseq node.");
7029 SDNode *Call = CallEnd->getOperand(0).getNode();
7030 bool hasGlue = Call->getGluedNode();
7032 // Replace the target specific call node with the patchable intrinsic.
7033 SmallVector<SDValue, 8> Ops;
7035 // Add the <id> and <numBytes> constants.
7036 SDValue IDVal = getValue(CI.getOperand(PatchPointOpers::IDPos));
7037 Ops.push_back(DAG.getTargetConstant(
7038 cast<ConstantSDNode>(IDVal)->getZExtValue(), MVT::i64));
7039 SDValue NBytesVal = getValue(CI.getOperand(PatchPointOpers::NBytesPos));
7040 Ops.push_back(DAG.getTargetConstant(
7041 cast<ConstantSDNode>(NBytesVal)->getZExtValue(), MVT::i32));
7043 // Assume that the Callee is a constant address.
7044 // FIXME: handle function symbols in the future.
7046 DAG.getIntPtrConstant(cast<ConstantSDNode>(Callee)->getZExtValue(),
7047 /*isTarget=*/true));
7049 // Adjust <numArgs> to account for any arguments that have been passed on the
7051 // Call Node: Chain, Target, {Args}, RegMask, [Glue]
7052 unsigned NumCallRegArgs = Call->getNumOperands() - (hasGlue ? 4 : 3);
7053 NumCallRegArgs = isAnyRegCC ? NumArgs : NumCallRegArgs;
7054 Ops.push_back(DAG.getTargetConstant(NumCallRegArgs, MVT::i32));
7056 // Add the calling convention
7057 Ops.push_back(DAG.getTargetConstant((unsigned)CC, MVT::i32));
7059 // Add the arguments we omitted previously. The register allocator should
7060 // place these in any free register.
7062 for (unsigned i = NumMetaOpers, e = NumMetaOpers + NumArgs; i != e; ++i)
7063 Ops.push_back(getValue(CI.getArgOperand(i)));
7065 // Push the arguments from the call instruction up to the register mask.
7066 SDNode::op_iterator e = hasGlue ? Call->op_end()-2 : Call->op_end()-1;
7067 for (SDNode::op_iterator i = Call->op_begin()+2; i != e; ++i)
7070 // Push live variables for the stack map.
7071 addStackMapLiveVars(CI, NumMetaOpers + NumArgs, Ops, *this);
7073 // Push the register mask info.
7075 Ops.push_back(*(Call->op_end()-2));
7077 Ops.push_back(*(Call->op_end()-1));
7079 // Push the chain (this is originally the first operand of the call, but
7080 // becomes now the last or second to last operand).
7081 Ops.push_back(*(Call->op_begin()));
7083 // Push the glue flag (last operand).
7085 Ops.push_back(*(Call->op_end()-1));
7088 if (isAnyRegCC && hasDef) {
7089 // Create the return types based on the intrinsic definition
7090 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7091 SmallVector<EVT, 3> ValueVTs;
7092 ComputeValueVTs(TLI, CI.getType(), ValueVTs);
7093 assert(ValueVTs.size() == 1 && "Expected only one return value type.");
7095 // There is always a chain and a glue type at the end
7096 ValueVTs.push_back(MVT::Other);
7097 ValueVTs.push_back(MVT::Glue);
7098 NodeTys = DAG.getVTList(ValueVTs);
7100 NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
7102 // Replace the target specific call node with a PATCHPOINT node.
7103 MachineSDNode *MN = DAG.getMachineNode(TargetOpcode::PATCHPOINT,
7104 getCurSDLoc(), NodeTys, Ops);
7106 // Update the NodeMap.
7109 setValue(&CI, SDValue(MN, 0));
7111 setValue(&CI, Result.first);
7114 // Fixup the consumers of the intrinsic. The chain and glue may be used in the
7115 // call sequence. Furthermore the location of the chain and glue can change
7116 // when the AnyReg calling convention is used and the intrinsic returns a
7118 if (isAnyRegCC && hasDef) {
7119 SDValue From[] = {SDValue(Call, 0), SDValue(Call, 1)};
7120 SDValue To[] = {SDValue(MN, 1), SDValue(MN, 2)};
7121 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
7123 DAG.ReplaceAllUsesWith(Call, MN);
7124 DAG.DeleteNode(Call);
7126 // Inform the Frame Information that we have a patchpoint in this function.
7127 FuncInfo.MF->getFrameInfo()->setHasPatchPoint();
7130 /// Returns an AttributeSet representing the attributes applied to the return
7131 /// value of the given call.
7132 static AttributeSet getReturnAttrs(TargetLowering::CallLoweringInfo &CLI) {
7133 SmallVector<Attribute::AttrKind, 2> Attrs;
7135 Attrs.push_back(Attribute::SExt);
7137 Attrs.push_back(Attribute::ZExt);
7139 Attrs.push_back(Attribute::InReg);
7141 return AttributeSet::get(CLI.RetTy->getContext(), AttributeSet::ReturnIndex,
7145 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
7146 /// implementation, which just calls LowerCall.
7147 /// FIXME: When all targets are
7148 /// migrated to using LowerCall, this hook should be integrated into SDISel.
7149 std::pair<SDValue, SDValue>
7150 TargetLowering::LowerCallTo(TargetLowering::CallLoweringInfo &CLI) const {
7151 // Handle the incoming return values from the call.
7153 Type *OrigRetTy = CLI.RetTy;
7154 SmallVector<EVT, 4> RetTys;
7155 SmallVector<uint64_t, 4> Offsets;
7156 ComputeValueVTs(*this, CLI.RetTy, RetTys, &Offsets);
7158 SmallVector<ISD::OutputArg, 4> Outs;
7159 GetReturnInfo(CLI.RetTy, getReturnAttrs(CLI), Outs, *this);
7161 bool CanLowerReturn =
7162 this->CanLowerReturn(CLI.CallConv, CLI.DAG.getMachineFunction(),
7163 CLI.IsVarArg, Outs, CLI.RetTy->getContext());
7165 SDValue DemoteStackSlot;
7166 int DemoteStackIdx = -100;
7167 if (!CanLowerReturn) {
7168 // FIXME: equivalent assert?
7169 // assert(!CS.hasInAllocaArgument() &&
7170 // "sret demotion is incompatible with inalloca");
7171 uint64_t TySize = getDataLayout()->getTypeAllocSize(CLI.RetTy);
7172 unsigned Align = getDataLayout()->getPrefTypeAlignment(CLI.RetTy);
7173 MachineFunction &MF = CLI.DAG.getMachineFunction();
7174 DemoteStackIdx = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
7175 Type *StackSlotPtrType = PointerType::getUnqual(CLI.RetTy);
7177 DemoteStackSlot = CLI.DAG.getFrameIndex(DemoteStackIdx, getPointerTy());
7179 Entry.Node = DemoteStackSlot;
7180 Entry.Ty = StackSlotPtrType;
7181 Entry.isSExt = false;
7182 Entry.isZExt = false;
7183 Entry.isInReg = false;
7184 Entry.isSRet = true;
7185 Entry.isNest = false;
7186 Entry.isByVal = false;
7187 Entry.isReturned = false;
7188 Entry.Alignment = Align;
7189 CLI.getArgs().insert(CLI.getArgs().begin(), Entry);
7190 CLI.RetTy = Type::getVoidTy(CLI.RetTy->getContext());
7192 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
7194 MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
7195 unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
7196 for (unsigned i = 0; i != NumRegs; ++i) {
7197 ISD::InputArg MyFlags;
7198 MyFlags.VT = RegisterVT;
7200 MyFlags.Used = CLI.IsReturnValueUsed;
7202 MyFlags.Flags.setSExt();
7204 MyFlags.Flags.setZExt();
7206 MyFlags.Flags.setInReg();
7207 CLI.Ins.push_back(MyFlags);
7212 // Handle all of the outgoing arguments.
7214 CLI.OutVals.clear();
7215 ArgListTy &Args = CLI.getArgs();
7216 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
7217 SmallVector<EVT, 4> ValueVTs;
7218 ComputeValueVTs(*this, Args[i].Ty, ValueVTs);
7219 Type *FinalType = Args[i].Ty;
7220 if (Args[i].isByVal)
7221 FinalType = cast<PointerType>(Args[i].Ty)->getElementType();
7222 bool NeedsRegBlock = functionArgumentNeedsConsecutiveRegisters(
7223 FinalType, CLI.CallConv, CLI.IsVarArg);
7224 for (unsigned Value = 0, NumValues = ValueVTs.size(); Value != NumValues;
7226 EVT VT = ValueVTs[Value];
7227 Type *ArgTy = VT.getTypeForEVT(CLI.RetTy->getContext());
7228 SDValue Op = SDValue(Args[i].Node.getNode(),
7229 Args[i].Node.getResNo() + Value);
7230 ISD::ArgFlagsTy Flags;
7231 unsigned OriginalAlignment = getDataLayout()->getABITypeAlignment(ArgTy);
7237 if (Args[i].isInReg)
7241 if (Args[i].isByVal)
7243 if (Args[i].isInAlloca) {
7244 Flags.setInAlloca();
7245 // Set the byval flag for CCAssignFn callbacks that don't know about
7246 // inalloca. This way we can know how many bytes we should've allocated
7247 // and how many bytes a callee cleanup function will pop. If we port
7248 // inalloca to more targets, we'll have to add custom inalloca handling
7249 // in the various CC lowering callbacks.
7252 if (Args[i].isByVal || Args[i].isInAlloca) {
7253 PointerType *Ty = cast<PointerType>(Args[i].Ty);
7254 Type *ElementTy = Ty->getElementType();
7255 Flags.setByValSize(getDataLayout()->getTypeAllocSize(ElementTy));
7256 // For ByVal, alignment should come from FE. BE will guess if this
7257 // info is not there but there are cases it cannot get right.
7258 unsigned FrameAlign;
7259 if (Args[i].Alignment)
7260 FrameAlign = Args[i].Alignment;
7262 FrameAlign = getByValTypeAlignment(ElementTy);
7263 Flags.setByValAlign(FrameAlign);
7267 if (NeedsRegBlock) {
7268 Flags.setInConsecutiveRegs();
7269 if (Value == NumValues - 1)
7270 Flags.setInConsecutiveRegsLast();
7272 Flags.setOrigAlign(OriginalAlignment);
7274 MVT PartVT = getRegisterType(CLI.RetTy->getContext(), VT);
7275 unsigned NumParts = getNumRegisters(CLI.RetTy->getContext(), VT);
7276 SmallVector<SDValue, 4> Parts(NumParts);
7277 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
7280 ExtendKind = ISD::SIGN_EXTEND;
7281 else if (Args[i].isZExt)
7282 ExtendKind = ISD::ZERO_EXTEND;
7284 // Conservatively only handle 'returned' on non-vectors for now
7285 if (Args[i].isReturned && !Op.getValueType().isVector()) {
7286 assert(CLI.RetTy == Args[i].Ty && RetTys.size() == NumValues &&
7287 "unexpected use of 'returned'");
7288 // Before passing 'returned' to the target lowering code, ensure that
7289 // either the register MVT and the actual EVT are the same size or that
7290 // the return value and argument are extended in the same way; in these
7291 // cases it's safe to pass the argument register value unchanged as the
7292 // return register value (although it's at the target's option whether
7294 // TODO: allow code generation to take advantage of partially preserved
7295 // registers rather than clobbering the entire register when the
7296 // parameter extension method is not compatible with the return
7298 if ((NumParts * PartVT.getSizeInBits() == VT.getSizeInBits()) ||
7299 (ExtendKind != ISD::ANY_EXTEND &&
7300 CLI.RetSExt == Args[i].isSExt && CLI.RetZExt == Args[i].isZExt))
7301 Flags.setReturned();
7304 getCopyToParts(CLI.DAG, CLI.DL, Op, &Parts[0], NumParts, PartVT,
7305 CLI.CS ? CLI.CS->getInstruction() : nullptr, ExtendKind);
7307 for (unsigned j = 0; j != NumParts; ++j) {
7308 // if it isn't first piece, alignment must be 1
7309 ISD::OutputArg MyFlags(Flags, Parts[j].getValueType(), VT,
7310 i < CLI.NumFixedArgs,
7311 i, j*Parts[j].getValueType().getStoreSize());
7312 if (NumParts > 1 && j == 0)
7313 MyFlags.Flags.setSplit();
7315 MyFlags.Flags.setOrigAlign(1);
7317 CLI.Outs.push_back(MyFlags);
7318 CLI.OutVals.push_back(Parts[j]);
7323 SmallVector<SDValue, 4> InVals;
7324 CLI.Chain = LowerCall(CLI, InVals);
7326 // Verify that the target's LowerCall behaved as expected.
7327 assert(CLI.Chain.getNode() && CLI.Chain.getValueType() == MVT::Other &&
7328 "LowerCall didn't return a valid chain!");
7329 assert((!CLI.IsTailCall || InVals.empty()) &&
7330 "LowerCall emitted a return value for a tail call!");
7331 assert((CLI.IsTailCall || InVals.size() == CLI.Ins.size()) &&
7332 "LowerCall didn't emit the correct number of values!");
7334 // For a tail call, the return value is merely live-out and there aren't
7335 // any nodes in the DAG representing it. Return a special value to
7336 // indicate that a tail call has been emitted and no more Instructions
7337 // should be processed in the current block.
7338 if (CLI.IsTailCall) {
7339 CLI.DAG.setRoot(CLI.Chain);
7340 return std::make_pair(SDValue(), SDValue());
7343 DEBUG(for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++i) {
7344 assert(InVals[i].getNode() &&
7345 "LowerCall emitted a null value!");
7346 assert(EVT(CLI.Ins[i].VT) == InVals[i].getValueType() &&
7347 "LowerCall emitted a value with the wrong type!");
7350 SmallVector<SDValue, 4> ReturnValues;
7351 if (!CanLowerReturn) {
7352 // The instruction result is the result of loading from the
7353 // hidden sret parameter.
7354 SmallVector<EVT, 1> PVTs;
7355 Type *PtrRetTy = PointerType::getUnqual(OrigRetTy);
7357 ComputeValueVTs(*this, PtrRetTy, PVTs);
7358 assert(PVTs.size() == 1 && "Pointers should fit in one register");
7359 EVT PtrVT = PVTs[0];
7361 unsigned NumValues = RetTys.size();
7362 ReturnValues.resize(NumValues);
7363 SmallVector<SDValue, 4> Chains(NumValues);
7365 for (unsigned i = 0; i < NumValues; ++i) {
7366 SDValue Add = CLI.DAG.getNode(ISD::ADD, CLI.DL, PtrVT, DemoteStackSlot,
7367 CLI.DAG.getConstant(Offsets[i], PtrVT));
7368 SDValue L = CLI.DAG.getLoad(
7369 RetTys[i], CLI.DL, CLI.Chain, Add,
7370 MachinePointerInfo::getFixedStack(DemoteStackIdx, Offsets[i]), false,
7372 ReturnValues[i] = L;
7373 Chains[i] = L.getValue(1);
7376 CLI.Chain = CLI.DAG.getNode(ISD::TokenFactor, CLI.DL, MVT::Other, Chains);
7378 // Collect the legal value parts into potentially illegal values
7379 // that correspond to the original function's return values.
7380 ISD::NodeType AssertOp = ISD::DELETED_NODE;
7382 AssertOp = ISD::AssertSext;
7383 else if (CLI.RetZExt)
7384 AssertOp = ISD::AssertZext;
7385 unsigned CurReg = 0;
7386 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
7388 MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
7389 unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
7391 ReturnValues.push_back(getCopyFromParts(CLI.DAG, CLI.DL, &InVals[CurReg],
7392 NumRegs, RegisterVT, VT, nullptr,
7397 // For a function returning void, there is no return value. We can't create
7398 // such a node, so we just return a null return value in that case. In
7399 // that case, nothing will actually look at the value.
7400 if (ReturnValues.empty())
7401 return std::make_pair(SDValue(), CLI.Chain);
7404 SDValue Res = CLI.DAG.getNode(ISD::MERGE_VALUES, CLI.DL,
7405 CLI.DAG.getVTList(RetTys), ReturnValues);
7406 return std::make_pair(Res, CLI.Chain);
7409 void TargetLowering::LowerOperationWrapper(SDNode *N,
7410 SmallVectorImpl<SDValue> &Results,
7411 SelectionDAG &DAG) const {
7412 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
7414 Results.push_back(Res);
7417 SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
7418 llvm_unreachable("LowerOperation not implemented for this target!");
7422 SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) {
7423 SDValue Op = getNonRegisterValue(V);
7424 assert((Op.getOpcode() != ISD::CopyFromReg ||
7425 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
7426 "Copy from a reg to the same reg!");
7427 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg");
7429 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
7430 RegsForValue RFV(V->getContext(), *TLI, Reg, V->getType());
7431 SDValue Chain = DAG.getEntryNode();
7432 RFV.getCopyToRegs(Op, DAG, getCurSDLoc(), Chain, nullptr, V);
7433 PendingExports.push_back(Chain);
7436 #include "llvm/CodeGen/SelectionDAGISel.h"
7438 /// isOnlyUsedInEntryBlock - If the specified argument is only used in the
7439 /// entry block, return true. This includes arguments used by switches, since
7440 /// the switch may expand into multiple basic blocks.
7441 static bool isOnlyUsedInEntryBlock(const Argument *A, bool FastISel) {
7442 // With FastISel active, we may be splitting blocks, so force creation
7443 // of virtual registers for all non-dead arguments.
7445 return A->use_empty();
7447 const BasicBlock *Entry = A->getParent()->begin();
7448 for (const User *U : A->users())
7449 if (cast<Instruction>(U)->getParent() != Entry || isa<SwitchInst>(U))
7450 return false; // Use not in entry block.
7455 void SelectionDAGISel::LowerArguments(const Function &F) {
7456 SelectionDAG &DAG = SDB->DAG;
7457 SDLoc dl = SDB->getCurSDLoc();
7458 const TargetLowering *TLI = getTargetLowering();
7459 const DataLayout *DL = TLI->getDataLayout();
7460 SmallVector<ISD::InputArg, 16> Ins;
7462 if (!FuncInfo->CanLowerReturn) {
7463 // Put in an sret pointer parameter before all the other parameters.
7464 SmallVector<EVT, 1> ValueVTs;
7465 ComputeValueVTs(*getTargetLowering(),
7466 PointerType::getUnqual(F.getReturnType()), ValueVTs);
7468 // NOTE: Assuming that a pointer will never break down to more than one VT
7470 ISD::ArgFlagsTy Flags;
7472 MVT RegisterVT = TLI->getRegisterType(*DAG.getContext(), ValueVTs[0]);
7473 ISD::InputArg RetArg(Flags, RegisterVT, ValueVTs[0], true, 0, 0);
7474 Ins.push_back(RetArg);
7477 // Set up the incoming argument description vector.
7479 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end();
7480 I != E; ++I, ++Idx) {
7481 SmallVector<EVT, 4> ValueVTs;
7482 ComputeValueVTs(*TLI, I->getType(), ValueVTs);
7483 bool isArgValueUsed = !I->use_empty();
7484 unsigned PartBase = 0;
7485 Type *FinalType = I->getType();
7486 if (F.getAttributes().hasAttribute(Idx, Attribute::ByVal))
7487 FinalType = cast<PointerType>(FinalType)->getElementType();
7488 bool NeedsRegBlock = TLI->functionArgumentNeedsConsecutiveRegisters(
7489 FinalType, F.getCallingConv(), F.isVarArg());
7490 for (unsigned Value = 0, NumValues = ValueVTs.size();
7491 Value != NumValues; ++Value) {
7492 EVT VT = ValueVTs[Value];
7493 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
7494 ISD::ArgFlagsTy Flags;
7495 unsigned OriginalAlignment = DL->getABITypeAlignment(ArgTy);
7497 if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt))
7499 if (F.getAttributes().hasAttribute(Idx, Attribute::SExt))
7501 if (F.getAttributes().hasAttribute(Idx, Attribute::InReg))
7503 if (F.getAttributes().hasAttribute(Idx, Attribute::StructRet))
7505 if (F.getAttributes().hasAttribute(Idx, Attribute::ByVal))
7507 if (F.getAttributes().hasAttribute(Idx, Attribute::InAlloca)) {
7508 Flags.setInAlloca();
7509 // Set the byval flag for CCAssignFn callbacks that don't know about
7510 // inalloca. This way we can know how many bytes we should've allocated
7511 // and how many bytes a callee cleanup function will pop. If we port
7512 // inalloca to more targets, we'll have to add custom inalloca handling
7513 // in the various CC lowering callbacks.
7516 if (Flags.isByVal() || Flags.isInAlloca()) {
7517 PointerType *Ty = cast<PointerType>(I->getType());
7518 Type *ElementTy = Ty->getElementType();
7519 Flags.setByValSize(DL->getTypeAllocSize(ElementTy));
7520 // For ByVal, alignment should be passed from FE. BE will guess if
7521 // this info is not there but there are cases it cannot get right.
7522 unsigned FrameAlign;
7523 if (F.getParamAlignment(Idx))
7524 FrameAlign = F.getParamAlignment(Idx);
7526 FrameAlign = TLI->getByValTypeAlignment(ElementTy);
7527 Flags.setByValAlign(FrameAlign);
7529 if (F.getAttributes().hasAttribute(Idx, Attribute::Nest))
7531 if (NeedsRegBlock) {
7532 Flags.setInConsecutiveRegs();
7533 if (Value == NumValues - 1)
7534 Flags.setInConsecutiveRegsLast();
7536 Flags.setOrigAlign(OriginalAlignment);
7538 MVT RegisterVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7539 unsigned NumRegs = TLI->getNumRegisters(*CurDAG->getContext(), VT);
7540 for (unsigned i = 0; i != NumRegs; ++i) {
7541 ISD::InputArg MyFlags(Flags, RegisterVT, VT, isArgValueUsed,
7542 Idx-1, PartBase+i*RegisterVT.getStoreSize());
7543 if (NumRegs > 1 && i == 0)
7544 MyFlags.Flags.setSplit();
7545 // if it isn't first piece, alignment must be 1
7547 MyFlags.Flags.setOrigAlign(1);
7548 Ins.push_back(MyFlags);
7550 PartBase += VT.getStoreSize();
7554 // Call the target to set up the argument values.
7555 SmallVector<SDValue, 8> InVals;
7556 SDValue NewRoot = TLI->LowerFormalArguments(DAG.getRoot(), F.getCallingConv(),
7560 // Verify that the target's LowerFormalArguments behaved as expected.
7561 assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other &&
7562 "LowerFormalArguments didn't return a valid chain!");
7563 assert(InVals.size() == Ins.size() &&
7564 "LowerFormalArguments didn't emit the correct number of values!");
7566 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
7567 assert(InVals[i].getNode() &&
7568 "LowerFormalArguments emitted a null value!");
7569 assert(EVT(Ins[i].VT) == InVals[i].getValueType() &&
7570 "LowerFormalArguments emitted a value with the wrong type!");
7574 // Update the DAG with the new chain value resulting from argument lowering.
7575 DAG.setRoot(NewRoot);
7577 // Set up the argument values.
7580 if (!FuncInfo->CanLowerReturn) {
7581 // Create a virtual register for the sret pointer, and put in a copy
7582 // from the sret argument into it.
7583 SmallVector<EVT, 1> ValueVTs;
7584 ComputeValueVTs(*TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
7585 MVT VT = ValueVTs[0].getSimpleVT();
7586 MVT RegVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7587 ISD::NodeType AssertOp = ISD::DELETED_NODE;
7588 SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1,
7589 RegVT, VT, nullptr, AssertOp);
7591 MachineFunction& MF = SDB->DAG.getMachineFunction();
7592 MachineRegisterInfo& RegInfo = MF.getRegInfo();
7593 unsigned SRetReg = RegInfo.createVirtualRegister(TLI->getRegClassFor(RegVT));
7594 FuncInfo->DemoteRegister = SRetReg;
7595 NewRoot = SDB->DAG.getCopyToReg(NewRoot, SDB->getCurSDLoc(),
7597 DAG.setRoot(NewRoot);
7599 // i indexes lowered arguments. Bump it past the hidden sret argument.
7600 // Idx indexes LLVM arguments. Don't touch it.
7604 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
7606 SmallVector<SDValue, 4> ArgValues;
7607 SmallVector<EVT, 4> ValueVTs;
7608 ComputeValueVTs(*TLI, I->getType(), ValueVTs);
7609 unsigned NumValues = ValueVTs.size();
7611 // If this argument is unused then remember its value. It is used to generate
7612 // debugging information.
7613 if (I->use_empty() && NumValues) {
7614 SDB->setUnusedArgValue(I, InVals[i]);
7616 // Also remember any frame index for use in FastISel.
7617 if (FrameIndexSDNode *FI =
7618 dyn_cast<FrameIndexSDNode>(InVals[i].getNode()))
7619 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
7622 for (unsigned Val = 0; Val != NumValues; ++Val) {
7623 EVT VT = ValueVTs[Val];
7624 MVT PartVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7625 unsigned NumParts = TLI->getNumRegisters(*CurDAG->getContext(), VT);
7627 if (!I->use_empty()) {
7628 ISD::NodeType AssertOp = ISD::DELETED_NODE;
7629 if (F.getAttributes().hasAttribute(Idx, Attribute::SExt))
7630 AssertOp = ISD::AssertSext;
7631 else if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt))
7632 AssertOp = ISD::AssertZext;
7634 ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i],
7635 NumParts, PartVT, VT,
7636 nullptr, AssertOp));
7642 // We don't need to do anything else for unused arguments.
7643 if (ArgValues.empty())
7646 // Note down frame index.
7647 if (FrameIndexSDNode *FI =
7648 dyn_cast<FrameIndexSDNode>(ArgValues[0].getNode()))
7649 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
7651 SDValue Res = DAG.getMergeValues(makeArrayRef(ArgValues.data(), NumValues),
7652 SDB->getCurSDLoc());
7654 SDB->setValue(I, Res);
7655 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::BUILD_PAIR) {
7656 if (LoadSDNode *LNode =
7657 dyn_cast<LoadSDNode>(Res.getOperand(0).getNode()))
7658 if (FrameIndexSDNode *FI =
7659 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode()))
7660 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
7663 // If this argument is live outside of the entry block, insert a copy from
7664 // wherever we got it to the vreg that other BB's will reference it as.
7665 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::CopyFromReg) {
7666 // If we can, though, try to skip creating an unnecessary vreg.
7667 // FIXME: This isn't very clean... it would be nice to make this more
7668 // general. It's also subtly incompatible with the hacks FastISel
7670 unsigned Reg = cast<RegisterSDNode>(Res.getOperand(1))->getReg();
7671 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
7672 FuncInfo->ValueMap[I] = Reg;
7676 if (!isOnlyUsedInEntryBlock(I, TM.Options.EnableFastISel)) {
7677 FuncInfo->InitializeRegForValue(I);
7678 SDB->CopyToExportRegsIfNeeded(I);
7682 assert(i == InVals.size() && "Argument register count mismatch!");
7684 // Finally, if the target has anything special to do, allow it to do so.
7685 // FIXME: this should insert code into the DAG!
7686 EmitFunctionEntryCode();
7689 /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
7690 /// ensure constants are generated when needed. Remember the virtual registers
7691 /// that need to be added to the Machine PHI nodes as input. We cannot just
7692 /// directly add them, because expansion might result in multiple MBB's for one
7693 /// BB. As such, the start of the BB might correspond to a different MBB than
7697 SelectionDAGBuilder::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
7698 const TerminatorInst *TI = LLVMBB->getTerminator();
7700 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
7702 // Check successor nodes' PHI nodes that expect a constant to be available
7704 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
7705 const BasicBlock *SuccBB = TI->getSuccessor(succ);
7706 if (!isa<PHINode>(SuccBB->begin())) continue;
7707 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
7709 // If this terminator has multiple identical successors (common for
7710 // switches), only handle each succ once.
7711 if (!SuccsHandled.insert(SuccMBB)) continue;
7713 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
7715 // At this point we know that there is a 1-1 correspondence between LLVM PHI
7716 // nodes and Machine PHI nodes, but the incoming operands have not been
7718 for (BasicBlock::const_iterator I = SuccBB->begin();
7719 const PHINode *PN = dyn_cast<PHINode>(I); ++I) {
7720 // Ignore dead phi's.
7721 if (PN->use_empty()) continue;
7724 if (PN->getType()->isEmptyTy())
7728 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
7730 if (const Constant *C = dyn_cast<Constant>(PHIOp)) {
7731 unsigned &RegOut = ConstantsOut[C];
7733 RegOut = FuncInfo.CreateRegs(C->getType());
7734 CopyValueToVirtualRegister(C, RegOut);
7738 DenseMap<const Value *, unsigned>::iterator I =
7739 FuncInfo.ValueMap.find(PHIOp);
7740 if (I != FuncInfo.ValueMap.end())
7743 assert(isa<AllocaInst>(PHIOp) &&
7744 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
7745 "Didn't codegen value into a register!??");
7746 Reg = FuncInfo.CreateRegs(PHIOp->getType());
7747 CopyValueToVirtualRegister(PHIOp, Reg);
7751 // Remember that this register needs to added to the machine PHI node as
7752 // the input for this MBB.
7753 SmallVector<EVT, 4> ValueVTs;
7754 const TargetLowering *TLI = TM.getSubtargetImpl()->getTargetLowering();
7755 ComputeValueVTs(*TLI, PN->getType(), ValueVTs);
7756 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
7757 EVT VT = ValueVTs[vti];
7758 unsigned NumRegisters = TLI->getNumRegisters(*DAG.getContext(), VT);
7759 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
7760 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
7761 Reg += NumRegisters;
7766 ConstantsOut.clear();
7769 /// Add a successor MBB to ParentMBB< creating a new MachineBB for BB if SuccMBB
7772 SelectionDAGBuilder::StackProtectorDescriptor::
7773 AddSuccessorMBB(const BasicBlock *BB,
7774 MachineBasicBlock *ParentMBB,
7775 MachineBasicBlock *SuccMBB) {
7776 // If SuccBB has not been created yet, create it.
7778 MachineFunction *MF = ParentMBB->getParent();
7779 MachineFunction::iterator BBI = ParentMBB;
7780 SuccMBB = MF->CreateMachineBasicBlock(BB);
7781 MF->insert(++BBI, SuccMBB);
7783 // Add it as a successor of ParentMBB.
7784 ParentMBB->addSuccessor(SuccMBB);