1 //===-- SelectionDAGBuilder.cpp - Selection-DAG building ------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements routines for translating from LLVM IR into SelectionDAG IR.
12 //===----------------------------------------------------------------------===//
14 #include "SelectionDAGBuilder.h"
15 #include "SDNodeDbgValue.h"
16 #include "llvm/ADT/BitVector.h"
17 #include "llvm/ADT/Optional.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/Analysis/BranchProbabilityInfo.h"
22 #include "llvm/Analysis/ConstantFolding.h"
23 #include "llvm/Analysis/TargetLibraryInfo.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/Analysis/VectorUtils.h"
26 #include "llvm/CodeGen/FastISel.h"
27 #include "llvm/CodeGen/FunctionLoweringInfo.h"
28 #include "llvm/CodeGen/GCMetadata.h"
29 #include "llvm/CodeGen/GCStrategy.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/SelectionDAG.h"
37 #include "llvm/CodeGen/StackMaps.h"
38 #include "llvm/CodeGen/WinEHFuncInfo.h"
39 #include "llvm/IR/CallingConv.h"
40 #include "llvm/IR/Constants.h"
41 #include "llvm/IR/DataLayout.h"
42 #include "llvm/IR/DebugInfo.h"
43 #include "llvm/IR/DerivedTypes.h"
44 #include "llvm/IR/Function.h"
45 #include "llvm/IR/GlobalVariable.h"
46 #include "llvm/IR/InlineAsm.h"
47 #include "llvm/IR/Instructions.h"
48 #include "llvm/IR/IntrinsicInst.h"
49 #include "llvm/IR/Intrinsics.h"
50 #include "llvm/IR/LLVMContext.h"
51 #include "llvm/IR/Module.h"
52 #include "llvm/IR/Statepoint.h"
53 #include "llvm/MC/MCSymbol.h"
54 #include "llvm/Support/CommandLine.h"
55 #include "llvm/Support/Debug.h"
56 #include "llvm/Support/ErrorHandling.h"
57 #include "llvm/Support/MathExtras.h"
58 #include "llvm/Support/raw_ostream.h"
59 #include "llvm/Target/TargetFrameLowering.h"
60 #include "llvm/Target/TargetInstrInfo.h"
61 #include "llvm/Target/TargetIntrinsicInfo.h"
62 #include "llvm/Target/TargetLowering.h"
63 #include "llvm/Target/TargetOptions.h"
64 #include "llvm/Target/TargetSelectionDAGInfo.h"
65 #include "llvm/Target/TargetSubtargetInfo.h"
70 #define DEBUG_TYPE "isel"
72 /// LimitFloatPrecision - Generate low-precision inline sequences for
73 /// some float libcalls (6, 8 or 12 bits).
74 static unsigned LimitFloatPrecision;
76 static cl::opt<unsigned, true>
77 LimitFPPrecision("limit-float-precision",
78 cl::desc("Generate low-precision inline sequences "
79 "for some float libcalls"),
80 cl::location(LimitFloatPrecision),
84 EnableFMFInDAG("enable-fmf-dag", cl::init(true), cl::Hidden,
85 cl::desc("Enable fast-math-flags for DAG nodes"));
87 // Limit the width of DAG chains. This is important in general to prevent
88 // DAG-based analysis from blowing up. For example, alias analysis and
89 // load clustering may not complete in reasonable time. It is difficult to
90 // recognize and avoid this situation within each individual analysis, and
91 // future analyses are likely to have the same behavior. Limiting DAG width is
92 // the safe approach and will be especially important with global DAGs.
94 // MaxParallelChains default is arbitrarily high to avoid affecting
95 // optimization, but could be lowered to improve compile time. Any ld-ld-st-st
96 // sequence over this should have been converted to llvm.memcpy by the
97 // frontend. It easy to induce this behavior with .ll code such as:
98 // %buffer = alloca [4096 x i8]
99 // %data = load [4096 x i8]* %argPtr
100 // store [4096 x i8] %data, [4096 x i8]* %buffer
101 static const unsigned MaxParallelChains = 64;
103 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, SDLoc DL,
104 const SDValue *Parts, unsigned NumParts,
105 MVT PartVT, EVT ValueVT, const Value *V);
107 /// getCopyFromParts - Create a value that contains the specified legal parts
108 /// combined into the value they represent. If the parts combine to a type
109 /// larger then ValueVT then AssertOp can be used to specify whether the extra
110 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
111 /// (ISD::AssertSext).
112 static SDValue getCopyFromParts(SelectionDAG &DAG, SDLoc DL,
113 const SDValue *Parts,
114 unsigned NumParts, MVT PartVT, EVT ValueVT,
116 ISD::NodeType AssertOp = ISD::DELETED_NODE) {
117 if (ValueVT.isVector())
118 return getCopyFromPartsVector(DAG, DL, Parts, NumParts,
121 assert(NumParts > 0 && "No parts to assemble!");
122 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
123 SDValue Val = Parts[0];
126 // Assemble the value from multiple parts.
127 if (ValueVT.isInteger()) {
128 unsigned PartBits = PartVT.getSizeInBits();
129 unsigned ValueBits = ValueVT.getSizeInBits();
131 // Assemble the power of 2 part.
132 unsigned RoundParts = NumParts & (NumParts - 1) ?
133 1 << Log2_32(NumParts) : NumParts;
134 unsigned RoundBits = PartBits * RoundParts;
135 EVT RoundVT = RoundBits == ValueBits ?
136 ValueVT : EVT::getIntegerVT(*DAG.getContext(), RoundBits);
139 EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2);
141 if (RoundParts > 2) {
142 Lo = getCopyFromParts(DAG, DL, Parts, RoundParts / 2,
144 Hi = getCopyFromParts(DAG, DL, Parts + RoundParts / 2,
145 RoundParts / 2, PartVT, HalfVT, V);
147 Lo = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[0]);
148 Hi = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[1]);
151 if (DAG.getDataLayout().isBigEndian())
154 Val = DAG.getNode(ISD::BUILD_PAIR, DL, RoundVT, Lo, Hi);
156 if (RoundParts < NumParts) {
157 // Assemble the trailing non-power-of-2 part.
158 unsigned OddParts = NumParts - RoundParts;
159 EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits);
160 Hi = getCopyFromParts(DAG, DL,
161 Parts + RoundParts, OddParts, PartVT, OddVT, V);
163 // Combine the round and odd parts.
165 if (DAG.getDataLayout().isBigEndian())
167 EVT TotalVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
168 Hi = DAG.getNode(ISD::ANY_EXTEND, DL, TotalVT, Hi);
170 DAG.getNode(ISD::SHL, DL, TotalVT, Hi,
171 DAG.getConstant(Lo.getValueType().getSizeInBits(), DL,
172 TLI.getPointerTy(DAG.getDataLayout())));
173 Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, TotalVT, Lo);
174 Val = DAG.getNode(ISD::OR, DL, TotalVT, Lo, Hi);
176 } else if (PartVT.isFloatingPoint()) {
177 // FP split into multiple FP parts (for ppcf128)
178 assert(ValueVT == EVT(MVT::ppcf128) && PartVT == MVT::f64 &&
181 Lo = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[0]);
182 Hi = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[1]);
183 if (TLI.hasBigEndianPartOrdering(ValueVT, DAG.getDataLayout()))
185 Val = DAG.getNode(ISD::BUILD_PAIR, DL, ValueVT, Lo, Hi);
187 // FP split into integer parts (soft fp)
188 assert(ValueVT.isFloatingPoint() && PartVT.isInteger() &&
189 !PartVT.isVector() && "Unexpected split");
190 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits());
191 Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT, V);
195 // There is now one part, held in Val. Correct it to match ValueVT.
196 EVT PartEVT = Val.getValueType();
198 if (PartEVT == ValueVT)
201 if (PartEVT.isInteger() && ValueVT.isFloatingPoint() &&
202 ValueVT.bitsLT(PartEVT)) {
203 // For an FP value in an integer part, we need to truncate to the right
205 PartEVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits());
206 Val = DAG.getNode(ISD::TRUNCATE, DL, PartEVT, Val);
209 if (PartEVT.isInteger() && ValueVT.isInteger()) {
210 if (ValueVT.bitsLT(PartEVT)) {
211 // For a truncate, see if we have any information to
212 // indicate whether the truncated bits will always be
213 // zero or sign-extension.
214 if (AssertOp != ISD::DELETED_NODE)
215 Val = DAG.getNode(AssertOp, DL, PartEVT, Val,
216 DAG.getValueType(ValueVT));
217 return DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
219 return DAG.getNode(ISD::ANY_EXTEND, DL, ValueVT, Val);
222 if (PartEVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
223 // FP_ROUND's are always exact here.
224 if (ValueVT.bitsLT(Val.getValueType()))
226 ISD::FP_ROUND, DL, ValueVT, Val,
227 DAG.getTargetConstant(1, DL, TLI.getPointerTy(DAG.getDataLayout())));
229 return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val);
232 if (PartEVT.getSizeInBits() == ValueVT.getSizeInBits())
233 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
235 llvm_unreachable("Unknown mismatch!");
238 static void diagnosePossiblyInvalidConstraint(LLVMContext &Ctx, const Value *V,
239 const Twine &ErrMsg) {
240 const Instruction *I = dyn_cast_or_null<Instruction>(V);
242 return Ctx.emitError(ErrMsg);
244 const char *AsmError = ", possible invalid constraint for vector type";
245 if (const CallInst *CI = dyn_cast<CallInst>(I))
246 if (isa<InlineAsm>(CI->getCalledValue()))
247 return Ctx.emitError(I, ErrMsg + AsmError);
249 return Ctx.emitError(I, ErrMsg);
252 /// getCopyFromPartsVector - Create a value that contains the specified legal
253 /// parts combined into the value they represent. If the parts combine to a
254 /// type larger then ValueVT then AssertOp can be used to specify whether the
255 /// extra bits are known to be zero (ISD::AssertZext) or sign extended from
256 /// ValueVT (ISD::AssertSext).
257 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, SDLoc DL,
258 const SDValue *Parts, unsigned NumParts,
259 MVT PartVT, EVT ValueVT, const Value *V) {
260 assert(ValueVT.isVector() && "Not a vector value");
261 assert(NumParts > 0 && "No parts to assemble!");
262 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
263 SDValue Val = Parts[0];
265 // Handle a multi-element vector.
269 unsigned NumIntermediates;
271 TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT,
272 NumIntermediates, RegisterVT);
273 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
274 NumParts = NumRegs; // Silence a compiler warning.
275 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
276 assert(RegisterVT.getSizeInBits() ==
277 Parts[0].getSimpleValueType().getSizeInBits() &&
278 "Part type sizes don't match!");
280 // Assemble the parts into intermediate operands.
281 SmallVector<SDValue, 8> Ops(NumIntermediates);
282 if (NumIntermediates == NumParts) {
283 // If the register was not expanded, truncate or copy the value,
285 for (unsigned i = 0; i != NumParts; ++i)
286 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i], 1,
287 PartVT, IntermediateVT, V);
288 } else if (NumParts > 0) {
289 // If the intermediate type was expanded, build the intermediate
290 // operands from the parts.
291 assert(NumParts % NumIntermediates == 0 &&
292 "Must expand into a divisible number of parts!");
293 unsigned Factor = NumParts / NumIntermediates;
294 for (unsigned i = 0; i != NumIntermediates; ++i)
295 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor,
296 PartVT, IntermediateVT, V);
299 // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the
300 // intermediate operands.
301 Val = DAG.getNode(IntermediateVT.isVector() ? ISD::CONCAT_VECTORS
306 // There is now one part, held in Val. Correct it to match ValueVT.
307 EVT PartEVT = Val.getValueType();
309 if (PartEVT == ValueVT)
312 if (PartEVT.isVector()) {
313 // If the element type of the source/dest vectors are the same, but the
314 // parts vector has more elements than the value vector, then we have a
315 // vector widening case (e.g. <2 x float> -> <4 x float>). Extract the
317 if (PartEVT.getVectorElementType() == ValueVT.getVectorElementType()) {
318 assert(PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements() &&
319 "Cannot narrow, it would be a lossy transformation");
321 ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val,
322 DAG.getConstant(0, DL, TLI.getVectorIdxTy(DAG.getDataLayout())));
325 // Vector/Vector bitcast.
326 if (ValueVT.getSizeInBits() == PartEVT.getSizeInBits())
327 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
329 assert(PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements() &&
330 "Cannot handle this kind of promotion");
331 // Promoted vector extract
332 return DAG.getAnyExtOrTrunc(Val, DL, ValueVT);
336 // Trivial bitcast if the types are the same size and the destination
337 // vector type is legal.
338 if (PartEVT.getSizeInBits() == ValueVT.getSizeInBits() &&
339 TLI.isTypeLegal(ValueVT))
340 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
342 // Handle cases such as i8 -> <1 x i1>
343 if (ValueVT.getVectorNumElements() != 1) {
344 diagnosePossiblyInvalidConstraint(*DAG.getContext(), V,
345 "non-trivial scalar-to-vector conversion");
346 return DAG.getUNDEF(ValueVT);
349 if (ValueVT.getVectorNumElements() == 1 &&
350 ValueVT.getVectorElementType() != PartEVT)
351 Val = DAG.getAnyExtOrTrunc(Val, DL, ValueVT.getScalarType());
353 return DAG.getNode(ISD::BUILD_VECTOR, DL, ValueVT, Val);
356 static void getCopyToPartsVector(SelectionDAG &DAG, SDLoc dl,
357 SDValue Val, SDValue *Parts, unsigned NumParts,
358 MVT PartVT, const Value *V);
360 /// getCopyToParts - Create a series of nodes that contain the specified value
361 /// split into legal parts. If the parts contain more bits than Val, then, for
362 /// integers, ExtendKind can be used to specify how to generate the extra bits.
363 static void getCopyToParts(SelectionDAG &DAG, SDLoc DL,
364 SDValue Val, SDValue *Parts, unsigned NumParts,
365 MVT PartVT, const Value *V,
366 ISD::NodeType ExtendKind = ISD::ANY_EXTEND) {
367 EVT ValueVT = Val.getValueType();
369 // Handle the vector case separately.
370 if (ValueVT.isVector())
371 return getCopyToPartsVector(DAG, DL, Val, Parts, NumParts, PartVT, V);
373 unsigned PartBits = PartVT.getSizeInBits();
374 unsigned OrigNumParts = NumParts;
375 assert(DAG.getTargetLoweringInfo().isTypeLegal(PartVT) &&
376 "Copying to an illegal type!");
381 assert(!ValueVT.isVector() && "Vector case handled elsewhere");
382 EVT PartEVT = PartVT;
383 if (PartEVT == ValueVT) {
384 assert(NumParts == 1 && "No-op copy with multiple parts!");
389 if (NumParts * PartBits > ValueVT.getSizeInBits()) {
390 // If the parts cover more bits than the value has, promote the value.
391 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
392 assert(NumParts == 1 && "Do not know what to promote to!");
393 Val = DAG.getNode(ISD::FP_EXTEND, DL, PartVT, Val);
395 if (ValueVT.isFloatingPoint()) {
396 // FP values need to be bitcast, then extended if they are being put
397 // into a larger container.
398 ValueVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits());
399 Val = DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
401 assert((PartVT.isInteger() || PartVT == MVT::x86mmx) &&
402 ValueVT.isInteger() &&
403 "Unknown mismatch!");
404 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
405 Val = DAG.getNode(ExtendKind, DL, ValueVT, Val);
406 if (PartVT == MVT::x86mmx)
407 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
409 } else if (PartBits == ValueVT.getSizeInBits()) {
410 // Different types of the same size.
411 assert(NumParts == 1 && PartEVT != ValueVT);
412 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
413 } else if (NumParts * PartBits < ValueVT.getSizeInBits()) {
414 // If the parts cover less bits than value has, truncate the value.
415 assert((PartVT.isInteger() || PartVT == MVT::x86mmx) &&
416 ValueVT.isInteger() &&
417 "Unknown mismatch!");
418 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
419 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
420 if (PartVT == MVT::x86mmx)
421 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
424 // The value may have changed - recompute ValueVT.
425 ValueVT = Val.getValueType();
426 assert(NumParts * PartBits == ValueVT.getSizeInBits() &&
427 "Failed to tile the value with PartVT!");
430 if (PartEVT != ValueVT)
431 diagnosePossiblyInvalidConstraint(*DAG.getContext(), V,
432 "scalar-to-vector conversion failed");
438 // Expand the value into multiple parts.
439 if (NumParts & (NumParts - 1)) {
440 // The number of parts is not a power of 2. Split off and copy the tail.
441 assert(PartVT.isInteger() && ValueVT.isInteger() &&
442 "Do not know what to expand to!");
443 unsigned RoundParts = 1 << Log2_32(NumParts);
444 unsigned RoundBits = RoundParts * PartBits;
445 unsigned OddParts = NumParts - RoundParts;
446 SDValue OddVal = DAG.getNode(ISD::SRL, DL, ValueVT, Val,
447 DAG.getIntPtrConstant(RoundBits, DL));
448 getCopyToParts(DAG, DL, OddVal, Parts + RoundParts, OddParts, PartVT, V);
450 if (DAG.getDataLayout().isBigEndian())
451 // The odd parts were reversed by getCopyToParts - unreverse them.
452 std::reverse(Parts + RoundParts, Parts + NumParts);
454 NumParts = RoundParts;
455 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
456 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
459 // The number of parts is a power of 2. Repeatedly bisect the value using
461 Parts[0] = DAG.getNode(ISD::BITCAST, DL,
462 EVT::getIntegerVT(*DAG.getContext(),
463 ValueVT.getSizeInBits()),
466 for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) {
467 for (unsigned i = 0; i < NumParts; i += StepSize) {
468 unsigned ThisBits = StepSize * PartBits / 2;
469 EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits);
470 SDValue &Part0 = Parts[i];
471 SDValue &Part1 = Parts[i+StepSize/2];
473 Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
474 ThisVT, Part0, DAG.getIntPtrConstant(1, DL));
475 Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
476 ThisVT, Part0, DAG.getIntPtrConstant(0, DL));
478 if (ThisBits == PartBits && ThisVT != PartVT) {
479 Part0 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part0);
480 Part1 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part1);
485 if (DAG.getDataLayout().isBigEndian())
486 std::reverse(Parts, Parts + OrigNumParts);
490 /// getCopyToPartsVector - Create a series of nodes that contain the specified
491 /// value split into legal parts.
492 static void getCopyToPartsVector(SelectionDAG &DAG, SDLoc DL,
493 SDValue Val, SDValue *Parts, unsigned NumParts,
494 MVT PartVT, const Value *V) {
495 EVT ValueVT = Val.getValueType();
496 assert(ValueVT.isVector() && "Not a vector");
497 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
500 EVT PartEVT = PartVT;
501 if (PartEVT == ValueVT) {
503 } else if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) {
504 // Bitconvert vector->vector case.
505 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
506 } else if (PartVT.isVector() &&
507 PartEVT.getVectorElementType() == ValueVT.getVectorElementType() &&
508 PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements()) {
509 EVT ElementVT = PartVT.getVectorElementType();
510 // Vector widening case, e.g. <2 x float> -> <4 x float>. Shuffle in
512 SmallVector<SDValue, 16> Ops;
513 for (unsigned i = 0, e = ValueVT.getVectorNumElements(); i != e; ++i)
514 Ops.push_back(DAG.getNode(
515 ISD::EXTRACT_VECTOR_ELT, DL, ElementVT, Val,
516 DAG.getConstant(i, DL, TLI.getVectorIdxTy(DAG.getDataLayout()))));
518 for (unsigned i = ValueVT.getVectorNumElements(),
519 e = PartVT.getVectorNumElements(); i != e; ++i)
520 Ops.push_back(DAG.getUNDEF(ElementVT));
522 Val = DAG.getNode(ISD::BUILD_VECTOR, DL, PartVT, Ops);
524 // FIXME: Use CONCAT for 2x -> 4x.
526 //SDValue UndefElts = DAG.getUNDEF(VectorTy);
527 //Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts);
528 } else if (PartVT.isVector() &&
529 PartEVT.getVectorElementType().bitsGE(
530 ValueVT.getVectorElementType()) &&
531 PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements()) {
533 // Promoted vector extract
534 Val = DAG.getAnyExtOrTrunc(Val, DL, PartVT);
536 // Vector -> scalar conversion.
537 assert(ValueVT.getVectorNumElements() == 1 &&
538 "Only trivial vector-to-scalar conversions should get here!");
540 ISD::EXTRACT_VECTOR_ELT, DL, PartVT, Val,
541 DAG.getConstant(0, DL, TLI.getVectorIdxTy(DAG.getDataLayout())));
543 Val = DAG.getAnyExtOrTrunc(Val, DL, PartVT);
550 // Handle a multi-element vector.
553 unsigned NumIntermediates;
554 unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT,
556 NumIntermediates, RegisterVT);
557 unsigned NumElements = ValueVT.getVectorNumElements();
559 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
560 NumParts = NumRegs; // Silence a compiler warning.
561 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
563 // Split the vector into intermediate operands.
564 SmallVector<SDValue, 8> Ops(NumIntermediates);
565 for (unsigned i = 0; i != NumIntermediates; ++i) {
566 if (IntermediateVT.isVector())
568 DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, IntermediateVT, Val,
569 DAG.getConstant(i * (NumElements / NumIntermediates), DL,
570 TLI.getVectorIdxTy(DAG.getDataLayout())));
572 Ops[i] = DAG.getNode(
573 ISD::EXTRACT_VECTOR_ELT, DL, IntermediateVT, Val,
574 DAG.getConstant(i, DL, TLI.getVectorIdxTy(DAG.getDataLayout())));
577 // Split the intermediate operands into legal parts.
578 if (NumParts == NumIntermediates) {
579 // If the register was not expanded, promote or copy the value,
581 for (unsigned i = 0; i != NumParts; ++i)
582 getCopyToParts(DAG, DL, Ops[i], &Parts[i], 1, PartVT, V);
583 } else if (NumParts > 0) {
584 // If the intermediate type was expanded, split each the value into
586 assert(NumIntermediates != 0 && "division by zero");
587 assert(NumParts % NumIntermediates == 0 &&
588 "Must expand into a divisible number of parts!");
589 unsigned Factor = NumParts / NumIntermediates;
590 for (unsigned i = 0; i != NumIntermediates; ++i)
591 getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT, V);
595 RegsForValue::RegsForValue() {}
597 RegsForValue::RegsForValue(const SmallVector<unsigned, 4> ®s, MVT regvt,
599 : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {}
601 RegsForValue::RegsForValue(LLVMContext &Context, const TargetLowering &TLI,
602 const DataLayout &DL, unsigned Reg, Type *Ty) {
603 ComputeValueVTs(TLI, DL, Ty, ValueVTs);
605 for (EVT ValueVT : ValueVTs) {
606 unsigned NumRegs = TLI.getNumRegisters(Context, ValueVT);
607 MVT RegisterVT = TLI.getRegisterType(Context, ValueVT);
608 for (unsigned i = 0; i != NumRegs; ++i)
609 Regs.push_back(Reg + i);
610 RegVTs.push_back(RegisterVT);
615 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
616 /// this value and returns the result as a ValueVT value. This uses
617 /// Chain/Flag as the input and updates them for the output Chain/Flag.
618 /// If the Flag pointer is NULL, no flag is used.
619 SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
620 FunctionLoweringInfo &FuncInfo,
622 SDValue &Chain, SDValue *Flag,
623 const Value *V) const {
624 // A Value with type {} or [0 x %t] needs no registers.
625 if (ValueVTs.empty())
628 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
630 // Assemble the legal parts into the final values.
631 SmallVector<SDValue, 4> Values(ValueVTs.size());
632 SmallVector<SDValue, 8> Parts;
633 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
634 // Copy the legal parts from the registers.
635 EVT ValueVT = ValueVTs[Value];
636 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
637 MVT RegisterVT = RegVTs[Value];
639 Parts.resize(NumRegs);
640 for (unsigned i = 0; i != NumRegs; ++i) {
643 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT);
645 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag);
646 *Flag = P.getValue(2);
649 Chain = P.getValue(1);
652 // If the source register was virtual and if we know something about it,
653 // add an assert node.
654 if (!TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) ||
655 !RegisterVT.isInteger() || RegisterVT.isVector())
658 const FunctionLoweringInfo::LiveOutInfo *LOI =
659 FuncInfo.GetLiveOutRegInfo(Regs[Part+i]);
663 unsigned RegSize = RegisterVT.getSizeInBits();
664 unsigned NumSignBits = LOI->NumSignBits;
665 unsigned NumZeroBits = LOI->KnownZero.countLeadingOnes();
667 if (NumZeroBits == RegSize) {
668 // The current value is a zero.
669 // Explicitly express that as it would be easier for
670 // optimizations to kick in.
671 Parts[i] = DAG.getConstant(0, dl, RegisterVT);
675 // FIXME: We capture more information than the dag can represent. For
676 // now, just use the tightest assertzext/assertsext possible.
678 EVT FromVT(MVT::Other);
679 if (NumSignBits == RegSize)
680 isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1
681 else if (NumZeroBits >= RegSize-1)
682 isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1
683 else if (NumSignBits > RegSize-8)
684 isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8
685 else if (NumZeroBits >= RegSize-8)
686 isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8
687 else if (NumSignBits > RegSize-16)
688 isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16
689 else if (NumZeroBits >= RegSize-16)
690 isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16
691 else if (NumSignBits > RegSize-32)
692 isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32
693 else if (NumZeroBits >= RegSize-32)
694 isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32
698 // Add an assertion node.
699 assert(FromVT != MVT::Other);
700 Parts[i] = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl,
701 RegisterVT, P, DAG.getValueType(FromVT));
704 Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(),
705 NumRegs, RegisterVT, ValueVT, V);
710 return DAG.getNode(ISD::MERGE_VALUES, dl, DAG.getVTList(ValueVTs), Values);
713 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
714 /// specified value into the registers specified by this object. This uses
715 /// Chain/Flag as the input and updates them for the output Chain/Flag.
716 /// If the Flag pointer is NULL, no flag is used.
717 void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, SDLoc dl,
718 SDValue &Chain, SDValue *Flag, const Value *V,
719 ISD::NodeType PreferredExtendType) const {
720 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
721 ISD::NodeType ExtendKind = PreferredExtendType;
723 // Get the list of the values's legal parts.
724 unsigned NumRegs = Regs.size();
725 SmallVector<SDValue, 8> Parts(NumRegs);
726 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
727 EVT ValueVT = ValueVTs[Value];
728 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
729 MVT RegisterVT = RegVTs[Value];
731 if (ExtendKind == ISD::ANY_EXTEND && TLI.isZExtFree(Val, RegisterVT))
732 ExtendKind = ISD::ZERO_EXTEND;
734 getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value),
735 &Parts[Part], NumParts, RegisterVT, V, ExtendKind);
739 // Copy the parts into the registers.
740 SmallVector<SDValue, 8> Chains(NumRegs);
741 for (unsigned i = 0; i != NumRegs; ++i) {
744 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]);
746 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag);
747 *Flag = Part.getValue(1);
750 Chains[i] = Part.getValue(0);
753 if (NumRegs == 1 || Flag)
754 // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is
755 // flagged to it. That is the CopyToReg nodes and the user are considered
756 // a single scheduling unit. If we create a TokenFactor and return it as
757 // chain, then the TokenFactor is both a predecessor (operand) of the
758 // user as well as a successor (the TF operands are flagged to the user).
759 // c1, f1 = CopyToReg
760 // c2, f2 = CopyToReg
761 // c3 = TokenFactor c1, c2
764 Chain = Chains[NumRegs-1];
766 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
769 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
770 /// operand list. This adds the code marker and includes the number of
771 /// values added into it.
772 void RegsForValue::AddInlineAsmOperands(unsigned Code, bool HasMatching,
773 unsigned MatchingIdx, SDLoc dl,
775 std::vector<SDValue> &Ops) const {
776 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
778 unsigned Flag = InlineAsm::getFlagWord(Code, Regs.size());
780 Flag = InlineAsm::getFlagWordForMatchingOp(Flag, MatchingIdx);
781 else if (!Regs.empty() &&
782 TargetRegisterInfo::isVirtualRegister(Regs.front())) {
783 // Put the register class of the virtual registers in the flag word. That
784 // way, later passes can recompute register class constraints for inline
785 // assembly as well as normal instructions.
786 // Don't do this for tied operands that can use the regclass information
788 const MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
789 const TargetRegisterClass *RC = MRI.getRegClass(Regs.front());
790 Flag = InlineAsm::getFlagWordForRegClass(Flag, RC->getID());
793 SDValue Res = DAG.getTargetConstant(Flag, dl, MVT::i32);
796 unsigned SP = TLI.getStackPointerRegisterToSaveRestore();
797 for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) {
798 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVTs[Value]);
799 MVT RegisterVT = RegVTs[Value];
800 for (unsigned i = 0; i != NumRegs; ++i) {
801 assert(Reg < Regs.size() && "Mismatch in # registers expected");
802 unsigned TheReg = Regs[Reg++];
803 Ops.push_back(DAG.getRegister(TheReg, RegisterVT));
805 if (TheReg == SP && Code == InlineAsm::Kind_Clobber) {
806 // If we clobbered the stack pointer, MFI should know about it.
807 assert(DAG.getMachineFunction().getFrameInfo()->
808 hasOpaqueSPAdjustment());
814 void SelectionDAGBuilder::init(GCFunctionInfo *gfi, AliasAnalysis &aa,
815 const TargetLibraryInfo *li) {
819 DL = &DAG.getDataLayout();
820 Context = DAG.getContext();
821 LPadToCallSiteMap.clear();
824 /// clear - Clear out the current SelectionDAG and the associated
825 /// state and prepare this SelectionDAGBuilder object to be used
826 /// for a new block. This doesn't clear out information about
827 /// additional blocks that are needed to complete switch lowering
828 /// or PHI node updating; that information is cleared out as it is
830 void SelectionDAGBuilder::clear() {
832 UnusedArgNodeMap.clear();
833 PendingLoads.clear();
834 PendingExports.clear();
837 SDNodeOrder = LowestSDNodeOrder;
838 StatepointLowering.clear();
841 /// clearDanglingDebugInfo - Clear the dangling debug information
842 /// map. This function is separated from the clear so that debug
843 /// information that is dangling in a basic block can be properly
844 /// resolved in a different basic block. This allows the
845 /// SelectionDAG to resolve dangling debug information attached
847 void SelectionDAGBuilder::clearDanglingDebugInfo() {
848 DanglingDebugInfoMap.clear();
851 /// getRoot - Return the current virtual root of the Selection DAG,
852 /// flushing any PendingLoad items. This must be done before emitting
853 /// a store or any other node that may need to be ordered after any
854 /// prior load instructions.
856 SDValue SelectionDAGBuilder::getRoot() {
857 if (PendingLoads.empty())
858 return DAG.getRoot();
860 if (PendingLoads.size() == 1) {
861 SDValue Root = PendingLoads[0];
863 PendingLoads.clear();
867 // Otherwise, we have to make a token factor node.
868 SDValue Root = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
870 PendingLoads.clear();
875 /// getControlRoot - Similar to getRoot, but instead of flushing all the
876 /// PendingLoad items, flush all the PendingExports items. It is necessary
877 /// to do this before emitting a terminator instruction.
879 SDValue SelectionDAGBuilder::getControlRoot() {
880 SDValue Root = DAG.getRoot();
882 if (PendingExports.empty())
885 // Turn all of the CopyToReg chains into one factored node.
886 if (Root.getOpcode() != ISD::EntryToken) {
887 unsigned i = 0, e = PendingExports.size();
888 for (; i != e; ++i) {
889 assert(PendingExports[i].getNode()->getNumOperands() > 1);
890 if (PendingExports[i].getNode()->getOperand(0) == Root)
891 break; // Don't add the root if we already indirectly depend on it.
895 PendingExports.push_back(Root);
898 Root = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
900 PendingExports.clear();
905 void SelectionDAGBuilder::visit(const Instruction &I) {
906 // Set up outgoing PHI node register values before emitting the terminator.
907 if (isa<TerminatorInst>(&I))
908 HandlePHINodesInSuccessorBlocks(I.getParent());
914 visit(I.getOpcode(), I);
916 if (!isa<TerminatorInst>(&I) && !HasTailCall &&
917 !isStatepoint(&I)) // statepoints handle their exports internally
918 CopyToExportRegsIfNeeded(&I);
923 void SelectionDAGBuilder::visitPHI(const PHINode &) {
924 llvm_unreachable("SelectionDAGBuilder shouldn't visit PHI nodes!");
927 void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) {
928 // Note: this doesn't use InstVisitor, because it has to work with
929 // ConstantExpr's in addition to instructions.
931 default: llvm_unreachable("Unknown instruction type encountered!");
932 // Build the switch statement using the Instruction.def file.
933 #define HANDLE_INST(NUM, OPCODE, CLASS) \
934 case Instruction::OPCODE: visit##OPCODE((const CLASS&)I); break;
935 #include "llvm/IR/Instruction.def"
939 // resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V,
940 // generate the debug data structures now that we've seen its definition.
941 void SelectionDAGBuilder::resolveDanglingDebugInfo(const Value *V,
943 DanglingDebugInfo &DDI = DanglingDebugInfoMap[V];
945 const DbgValueInst *DI = DDI.getDI();
946 DebugLoc dl = DDI.getdl();
947 unsigned DbgSDNodeOrder = DDI.getSDNodeOrder();
948 DILocalVariable *Variable = DI->getVariable();
949 DIExpression *Expr = DI->getExpression();
950 assert(Variable->isValidLocationForIntrinsic(dl) &&
951 "Expected inlined-at fields to agree");
952 uint64_t Offset = DI->getOffset();
953 // A dbg.value for an alloca is always indirect.
954 bool IsIndirect = isa<AllocaInst>(V) || Offset != 0;
957 if (!EmitFuncArgumentDbgValue(V, Variable, Expr, dl, Offset, IsIndirect,
959 SDV = DAG.getDbgValue(Variable, Expr, Val.getNode(), Val.getResNo(),
960 IsIndirect, Offset, dl, DbgSDNodeOrder);
961 DAG.AddDbgValue(SDV, Val.getNode(), false);
964 DEBUG(dbgs() << "Dropping debug info for " << *DI << "\n");
965 DanglingDebugInfoMap[V] = DanglingDebugInfo();
969 /// getCopyFromRegs - If there was virtual register allocated for the value V
970 /// emit CopyFromReg of the specified type Ty. Return empty SDValue() otherwise.
971 SDValue SelectionDAGBuilder::getCopyFromRegs(const Value *V, Type *Ty) {
972 DenseMap<const Value *, unsigned>::iterator It = FuncInfo.ValueMap.find(V);
975 if (It != FuncInfo.ValueMap.end()) {
976 unsigned InReg = It->second;
977 RegsForValue RFV(*DAG.getContext(), DAG.getTargetLoweringInfo(),
978 DAG.getDataLayout(), InReg, Ty);
979 SDValue Chain = DAG.getEntryNode();
980 Result = RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, nullptr, V);
981 resolveDanglingDebugInfo(V, Result);
987 /// getValue - Return an SDValue for the given Value.
988 SDValue SelectionDAGBuilder::getValue(const Value *V) {
989 // If we already have an SDValue for this value, use it. It's important
990 // to do this first, so that we don't create a CopyFromReg if we already
991 // have a regular SDValue.
992 SDValue &N = NodeMap[V];
993 if (N.getNode()) return N;
995 // If there's a virtual register allocated and initialized for this
997 SDValue copyFromReg = getCopyFromRegs(V, V->getType());
998 if (copyFromReg.getNode()) {
1002 // Otherwise create a new SDValue and remember it.
1003 SDValue Val = getValueImpl(V);
1005 resolveDanglingDebugInfo(V, Val);
1009 // Return true if SDValue exists for the given Value
1010 bool SelectionDAGBuilder::findValue(const Value *V) const {
1011 return (NodeMap.find(V) != NodeMap.end()) ||
1012 (FuncInfo.ValueMap.find(V) != FuncInfo.ValueMap.end());
1015 /// getNonRegisterValue - Return an SDValue for the given Value, but
1016 /// don't look in FuncInfo.ValueMap for a virtual register.
1017 SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) {
1018 // If we already have an SDValue for this value, use it.
1019 SDValue &N = NodeMap[V];
1021 if (isa<ConstantSDNode>(N) || isa<ConstantFPSDNode>(N)) {
1022 // Remove the debug location from the node as the node is about to be used
1023 // in a location which may differ from the original debug location. This
1024 // is relevant to Constant and ConstantFP nodes because they can appear
1025 // as constant expressions inside PHI nodes.
1026 N->setDebugLoc(DebugLoc());
1031 // Otherwise create a new SDValue and remember it.
1032 SDValue Val = getValueImpl(V);
1034 resolveDanglingDebugInfo(V, Val);
1038 /// getValueImpl - Helper function for getValue and getNonRegisterValue.
1039 /// Create an SDValue for the given value.
1040 SDValue SelectionDAGBuilder::getValueImpl(const Value *V) {
1041 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1043 if (const Constant *C = dyn_cast<Constant>(V)) {
1044 EVT VT = TLI.getValueType(DAG.getDataLayout(), V->getType(), true);
1046 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C))
1047 return DAG.getConstant(*CI, getCurSDLoc(), VT);
1049 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
1050 return DAG.getGlobalAddress(GV, getCurSDLoc(), VT);
1052 if (isa<ConstantPointerNull>(C)) {
1053 unsigned AS = V->getType()->getPointerAddressSpace();
1054 return DAG.getConstant(0, getCurSDLoc(),
1055 TLI.getPointerTy(DAG.getDataLayout(), AS));
1058 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
1059 return DAG.getConstantFP(*CFP, getCurSDLoc(), VT);
1061 if (isa<UndefValue>(C) && !V->getType()->isAggregateType())
1062 return DAG.getUNDEF(VT);
1064 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
1065 visit(CE->getOpcode(), *CE);
1066 SDValue N1 = NodeMap[V];
1067 assert(N1.getNode() && "visit didn't populate the NodeMap!");
1071 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) {
1072 SmallVector<SDValue, 4> Constants;
1073 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end();
1075 SDNode *Val = getValue(*OI).getNode();
1076 // If the operand is an empty aggregate, there are no values.
1078 // Add each leaf value from the operand to the Constants list
1079 // to form a flattened list of all the values.
1080 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
1081 Constants.push_back(SDValue(Val, i));
1084 return DAG.getMergeValues(Constants, getCurSDLoc());
1087 if (const ConstantDataSequential *CDS =
1088 dyn_cast<ConstantDataSequential>(C)) {
1089 SmallVector<SDValue, 4> Ops;
1090 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1091 SDNode *Val = getValue(CDS->getElementAsConstant(i)).getNode();
1092 // Add each leaf value from the operand to the Constants list
1093 // to form a flattened list of all the values.
1094 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
1095 Ops.push_back(SDValue(Val, i));
1098 if (isa<ArrayType>(CDS->getType()))
1099 return DAG.getMergeValues(Ops, getCurSDLoc());
1100 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(),
1104 if (C->getType()->isStructTy() || C->getType()->isArrayTy()) {
1105 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
1106 "Unknown struct or array constant!");
1108 SmallVector<EVT, 4> ValueVTs;
1109 ComputeValueVTs(TLI, DAG.getDataLayout(), C->getType(), ValueVTs);
1110 unsigned NumElts = ValueVTs.size();
1112 return SDValue(); // empty struct
1113 SmallVector<SDValue, 4> Constants(NumElts);
1114 for (unsigned i = 0; i != NumElts; ++i) {
1115 EVT EltVT = ValueVTs[i];
1116 if (isa<UndefValue>(C))
1117 Constants[i] = DAG.getUNDEF(EltVT);
1118 else if (EltVT.isFloatingPoint())
1119 Constants[i] = DAG.getConstantFP(0, getCurSDLoc(), EltVT);
1121 Constants[i] = DAG.getConstant(0, getCurSDLoc(), EltVT);
1124 return DAG.getMergeValues(Constants, getCurSDLoc());
1127 if (const BlockAddress *BA = dyn_cast<BlockAddress>(C))
1128 return DAG.getBlockAddress(BA, VT);
1130 VectorType *VecTy = cast<VectorType>(V->getType());
1131 unsigned NumElements = VecTy->getNumElements();
1133 // Now that we know the number and type of the elements, get that number of
1134 // elements into the Ops array based on what kind of constant it is.
1135 SmallVector<SDValue, 16> Ops;
1136 if (const ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
1137 for (unsigned i = 0; i != NumElements; ++i)
1138 Ops.push_back(getValue(CV->getOperand(i)));
1140 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!");
1142 TLI.getValueType(DAG.getDataLayout(), VecTy->getElementType());
1145 if (EltVT.isFloatingPoint())
1146 Op = DAG.getConstantFP(0, getCurSDLoc(), EltVT);
1148 Op = DAG.getConstant(0, getCurSDLoc(), EltVT);
1149 Ops.assign(NumElements, Op);
1152 // Create a BUILD_VECTOR node.
1153 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(), VT, Ops);
1156 // If this is a static alloca, generate it as the frameindex instead of
1158 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
1159 DenseMap<const AllocaInst*, int>::iterator SI =
1160 FuncInfo.StaticAllocaMap.find(AI);
1161 if (SI != FuncInfo.StaticAllocaMap.end())
1162 return DAG.getFrameIndex(SI->second,
1163 TLI.getPointerTy(DAG.getDataLayout()));
1166 // If this is an instruction which fast-isel has deferred, select it now.
1167 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
1168 unsigned InReg = FuncInfo.InitializeRegForValue(Inst);
1169 RegsForValue RFV(*DAG.getContext(), TLI, DAG.getDataLayout(), InReg,
1171 SDValue Chain = DAG.getEntryNode();
1172 return RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, nullptr, V);
1175 llvm_unreachable("Can't get register for value!");
1178 void SelectionDAGBuilder::visitCatchPad(const CatchPadInst &I) {
1179 auto Pers = classifyEHPersonality(FuncInfo.Fn->getPersonalityFn());
1180 bool IsMSVCCXX = Pers == EHPersonality::MSVC_CXX;
1181 bool IsCoreCLR = Pers == EHPersonality::CoreCLR;
1182 MachineBasicBlock *CatchPadMBB = FuncInfo.MBB;
1183 // In MSVC C++ and CoreCLR, catchblocks are funclets and need prologues.
1184 if (IsMSVCCXX || IsCoreCLR)
1185 CatchPadMBB->setIsEHFuncletEntry();
1187 DAG.setRoot(DAG.getNode(ISD::CATCHPAD, getCurSDLoc(), MVT::Other, getControlRoot()));
1190 void SelectionDAGBuilder::visitCatchRet(const CatchReturnInst &I) {
1191 // Update machine-CFG edge.
1192 MachineBasicBlock *TargetMBB = FuncInfo.MBBMap[I.getSuccessor()];
1193 FuncInfo.MBB->addSuccessor(TargetMBB);
1195 auto Pers = classifyEHPersonality(FuncInfo.Fn->getPersonalityFn());
1196 bool IsSEH = isAsynchronousEHPersonality(Pers);
1198 // If this is not a fall-through branch or optimizations are switched off,
1200 if (TargetMBB != NextBlock(FuncInfo.MBB) ||
1201 TM.getOptLevel() == CodeGenOpt::None)
1202 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other,
1203 getControlRoot(), DAG.getBasicBlock(TargetMBB)));
1207 // Figure out the funclet membership for the catchret's successor.
1208 // This will be used by the FuncletLayout pass to determine how to order the
1210 WinEHFuncInfo *EHInfo = DAG.getMachineFunction().getWinEHFuncInfo();
1211 const BasicBlock *SuccessorColor = EHInfo->CatchRetSuccessorColorMap[&I];
1212 assert(SuccessorColor && "No parent funclet for catchret!");
1213 MachineBasicBlock *SuccessorColorMBB = FuncInfo.MBBMap[SuccessorColor];
1214 assert(SuccessorColorMBB && "No MBB for SuccessorColor!");
1216 // Create the terminator node.
1217 SDValue Ret = DAG.getNode(ISD::CATCHRET, getCurSDLoc(), MVT::Other,
1218 getControlRoot(), DAG.getBasicBlock(TargetMBB),
1219 DAG.getBasicBlock(SuccessorColorMBB));
1223 void SelectionDAGBuilder::visitCleanupPad(const CleanupPadInst &CPI) {
1224 // Don't emit any special code for the cleanuppad instruction. It just marks
1225 // the start of a funclet.
1226 FuncInfo.MBB->setIsEHFuncletEntry();
1227 FuncInfo.MBB->setIsCleanupFuncletEntry();
1230 /// When an invoke or a cleanupret unwinds to the next EH pad, there are
1231 /// many places it could ultimately go. In the IR, we have a single unwind
1232 /// destination, but in the machine CFG, we enumerate all the possible blocks.
1233 /// This function skips over imaginary basic blocks that hold catchswitch
1234 /// instructions, and finds all the "real" machine
1235 /// basic block destinations. As those destinations may not be successors of
1236 /// EHPadBB, here we also calculate the edge probability to those destinations.
1237 /// The passed-in Prob is the edge probability to EHPadBB.
1238 static void findUnwindDestinations(
1239 FunctionLoweringInfo &FuncInfo, const BasicBlock *EHPadBB,
1240 BranchProbability Prob,
1241 SmallVectorImpl<std::pair<MachineBasicBlock *, BranchProbability>>
1243 EHPersonality Personality =
1244 classifyEHPersonality(FuncInfo.Fn->getPersonalityFn());
1245 bool IsMSVCCXX = Personality == EHPersonality::MSVC_CXX;
1246 bool IsCoreCLR = Personality == EHPersonality::CoreCLR;
1249 const Instruction *Pad = EHPadBB->getFirstNonPHI();
1250 BasicBlock *NewEHPadBB = nullptr;
1251 if (isa<LandingPadInst>(Pad)) {
1252 // Stop on landingpads. They are not funclets.
1253 UnwindDests.emplace_back(FuncInfo.MBBMap[EHPadBB], Prob);
1255 } else if (isa<CleanupPadInst>(Pad)) {
1256 // Stop on cleanup pads. Cleanups are always funclet entries for all known
1258 UnwindDests.emplace_back(FuncInfo.MBBMap[EHPadBB], Prob);
1259 UnwindDests.back().first->setIsEHFuncletEntry();
1261 } else if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(Pad)) {
1262 // Add the catchpad handlers to the possible destinations.
1263 for (const BasicBlock *CatchPadBB : CatchSwitch->handlers()) {
1264 UnwindDests.emplace_back(FuncInfo.MBBMap[CatchPadBB], Prob);
1265 // For MSVC++ and the CLR, catchblocks are funclets and need prologues.
1266 if (IsMSVCCXX || IsCoreCLR)
1267 UnwindDests.back().first->setIsEHFuncletEntry();
1269 NewEHPadBB = CatchSwitch->getUnwindDest();
1274 BranchProbabilityInfo *BPI = FuncInfo.BPI;
1275 if (BPI && NewEHPadBB)
1276 Prob *= BPI->getEdgeProbability(EHPadBB, NewEHPadBB);
1277 EHPadBB = NewEHPadBB;
1281 void SelectionDAGBuilder::visitCleanupRet(const CleanupReturnInst &I) {
1282 // Update successor info.
1283 SmallVector<std::pair<MachineBasicBlock *, BranchProbability>, 1> UnwindDests;
1284 auto UnwindDest = I.getUnwindDest();
1285 BranchProbabilityInfo *BPI = FuncInfo.BPI;
1286 BranchProbability UnwindDestProb =
1288 ? BPI->getEdgeProbability(FuncInfo.MBB->getBasicBlock(), UnwindDest)
1289 : BranchProbability::getZero();
1290 findUnwindDestinations(FuncInfo, UnwindDest, UnwindDestProb, UnwindDests);
1291 for (auto &UnwindDest : UnwindDests) {
1292 UnwindDest.first->setIsEHPad();
1293 addSuccessorWithProb(FuncInfo.MBB, UnwindDest.first, UnwindDest.second);
1295 FuncInfo.MBB->normalizeSuccProbs();
1297 // Create the terminator node.
1299 DAG.getNode(ISD::CLEANUPRET, getCurSDLoc(), MVT::Other, getControlRoot());
1303 void SelectionDAGBuilder::visitCatchSwitch(const CatchSwitchInst &CSI) {
1304 report_fatal_error("visitCatchSwitch not yet implemented!");
1307 void SelectionDAGBuilder::visitRet(const ReturnInst &I) {
1308 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1309 auto &DL = DAG.getDataLayout();
1310 SDValue Chain = getControlRoot();
1311 SmallVector<ISD::OutputArg, 8> Outs;
1312 SmallVector<SDValue, 8> OutVals;
1314 if (!FuncInfo.CanLowerReturn) {
1315 unsigned DemoteReg = FuncInfo.DemoteRegister;
1316 const Function *F = I.getParent()->getParent();
1318 // Emit a store of the return value through the virtual register.
1319 // Leave Outs empty so that LowerReturn won't try to load return
1320 // registers the usual way.
1321 SmallVector<EVT, 1> PtrValueVTs;
1322 ComputeValueVTs(TLI, DL, PointerType::getUnqual(F->getReturnType()),
1325 SDValue RetPtr = DAG.getCopyFromReg(DAG.getEntryNode(), getCurSDLoc(),
1326 DemoteReg, PtrValueVTs[0]);
1327 SDValue RetOp = getValue(I.getOperand(0));
1329 SmallVector<EVT, 4> ValueVTs;
1330 SmallVector<uint64_t, 4> Offsets;
1331 ComputeValueVTs(TLI, DL, I.getOperand(0)->getType(), ValueVTs, &Offsets);
1332 unsigned NumValues = ValueVTs.size();
1334 SmallVector<SDValue, 4> Chains(NumValues);
1335 for (unsigned i = 0; i != NumValues; ++i) {
1336 SDValue Add = DAG.getNode(ISD::ADD, getCurSDLoc(),
1337 RetPtr.getValueType(), RetPtr,
1338 DAG.getIntPtrConstant(Offsets[i],
1341 DAG.getStore(Chain, getCurSDLoc(),
1342 SDValue(RetOp.getNode(), RetOp.getResNo() + i),
1343 // FIXME: better loc info would be nice.
1344 Add, MachinePointerInfo(), false, false, 0);
1347 Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(),
1348 MVT::Other, Chains);
1349 } else if (I.getNumOperands() != 0) {
1350 SmallVector<EVT, 4> ValueVTs;
1351 ComputeValueVTs(TLI, DL, I.getOperand(0)->getType(), ValueVTs);
1352 unsigned NumValues = ValueVTs.size();
1354 SDValue RetOp = getValue(I.getOperand(0));
1356 const Function *F = I.getParent()->getParent();
1358 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1359 if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1361 ExtendKind = ISD::SIGN_EXTEND;
1362 else if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1364 ExtendKind = ISD::ZERO_EXTEND;
1366 LLVMContext &Context = F->getContext();
1367 bool RetInReg = F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1370 for (unsigned j = 0; j != NumValues; ++j) {
1371 EVT VT = ValueVTs[j];
1373 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger())
1374 VT = TLI.getTypeForExtArgOrReturn(Context, VT, ExtendKind);
1376 unsigned NumParts = TLI.getNumRegisters(Context, VT);
1377 MVT PartVT = TLI.getRegisterType(Context, VT);
1378 SmallVector<SDValue, 4> Parts(NumParts);
1379 getCopyToParts(DAG, getCurSDLoc(),
1380 SDValue(RetOp.getNode(), RetOp.getResNo() + j),
1381 &Parts[0], NumParts, PartVT, &I, ExtendKind);
1383 // 'inreg' on function refers to return value
1384 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1388 // Propagate extension type if any
1389 if (ExtendKind == ISD::SIGN_EXTEND)
1391 else if (ExtendKind == ISD::ZERO_EXTEND)
1394 for (unsigned i = 0; i < NumParts; ++i) {
1395 Outs.push_back(ISD::OutputArg(Flags, Parts[i].getValueType(),
1396 VT, /*isfixed=*/true, 0, 0));
1397 OutVals.push_back(Parts[i]);
1403 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1404 CallingConv::ID CallConv =
1405 DAG.getMachineFunction().getFunction()->getCallingConv();
1406 Chain = DAG.getTargetLoweringInfo().LowerReturn(
1407 Chain, CallConv, isVarArg, Outs, OutVals, getCurSDLoc(), DAG);
1409 // Verify that the target's LowerReturn behaved as expected.
1410 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
1411 "LowerReturn didn't return a valid chain!");
1413 // Update the DAG with the new chain value resulting from return lowering.
1417 /// CopyToExportRegsIfNeeded - If the given value has virtual registers
1418 /// created for it, emit nodes to copy the value into the virtual
1420 void SelectionDAGBuilder::CopyToExportRegsIfNeeded(const Value *V) {
1422 if (V->getType()->isEmptyTy())
1425 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
1426 if (VMI != FuncInfo.ValueMap.end()) {
1427 assert(!V->use_empty() && "Unused value assigned virtual registers!");
1428 CopyValueToVirtualRegister(V, VMI->second);
1432 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
1433 /// the current basic block, add it to ValueMap now so that we'll get a
1435 void SelectionDAGBuilder::ExportFromCurrentBlock(const Value *V) {
1436 // No need to export constants.
1437 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
1439 // Already exported?
1440 if (FuncInfo.isExportedInst(V)) return;
1442 unsigned Reg = FuncInfo.InitializeRegForValue(V);
1443 CopyValueToVirtualRegister(V, Reg);
1446 bool SelectionDAGBuilder::isExportableFromCurrentBlock(const Value *V,
1447 const BasicBlock *FromBB) {
1448 // The operands of the setcc have to be in this block. We don't know
1449 // how to export them from some other block.
1450 if (const Instruction *VI = dyn_cast<Instruction>(V)) {
1451 // Can export from current BB.
1452 if (VI->getParent() == FromBB)
1455 // Is already exported, noop.
1456 return FuncInfo.isExportedInst(V);
1459 // If this is an argument, we can export it if the BB is the entry block or
1460 // if it is already exported.
1461 if (isa<Argument>(V)) {
1462 if (FromBB == &FromBB->getParent()->getEntryBlock())
1465 // Otherwise, can only export this if it is already exported.
1466 return FuncInfo.isExportedInst(V);
1469 // Otherwise, constants can always be exported.
1473 /// Return branch probability calculated by BranchProbabilityInfo for IR blocks.
1475 SelectionDAGBuilder::getEdgeProbability(const MachineBasicBlock *Src,
1476 const MachineBasicBlock *Dst) const {
1477 BranchProbabilityInfo *BPI = FuncInfo.BPI;
1478 const BasicBlock *SrcBB = Src->getBasicBlock();
1479 const BasicBlock *DstBB = Dst->getBasicBlock();
1481 // If BPI is not available, set the default probability as 1 / N, where N is
1482 // the number of successors.
1483 auto SuccSize = std::max<uint32_t>(
1484 std::distance(succ_begin(SrcBB), succ_end(SrcBB)), 1);
1485 return BranchProbability(1, SuccSize);
1487 return BPI->getEdgeProbability(SrcBB, DstBB);
1490 void SelectionDAGBuilder::addSuccessorWithProb(MachineBasicBlock *Src,
1491 MachineBasicBlock *Dst,
1492 BranchProbability Prob) {
1494 Src->addSuccessorWithoutProb(Dst);
1496 if (Prob.isUnknown())
1497 Prob = getEdgeProbability(Src, Dst);
1498 Src->addSuccessor(Dst, Prob);
1502 static bool InBlock(const Value *V, const BasicBlock *BB) {
1503 if (const Instruction *I = dyn_cast<Instruction>(V))
1504 return I->getParent() == BB;
1508 /// EmitBranchForMergedCondition - Helper method for FindMergedConditions.
1509 /// This function emits a branch and is used at the leaves of an OR or an
1510 /// AND operator tree.
1513 SelectionDAGBuilder::EmitBranchForMergedCondition(const Value *Cond,
1514 MachineBasicBlock *TBB,
1515 MachineBasicBlock *FBB,
1516 MachineBasicBlock *CurBB,
1517 MachineBasicBlock *SwitchBB,
1518 BranchProbability TProb,
1519 BranchProbability FProb) {
1520 const BasicBlock *BB = CurBB->getBasicBlock();
1522 // If the leaf of the tree is a comparison, merge the condition into
1524 if (const CmpInst *BOp = dyn_cast<CmpInst>(Cond)) {
1525 // The operands of the cmp have to be in this block. We don't know
1526 // how to export them from some other block. If this is the first block
1527 // of the sequence, no exporting is needed.
1528 if (CurBB == SwitchBB ||
1529 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
1530 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) {
1531 ISD::CondCode Condition;
1532 if (const ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
1533 Condition = getICmpCondCode(IC->getPredicate());
1535 const FCmpInst *FC = cast<FCmpInst>(Cond);
1536 Condition = getFCmpCondCode(FC->getPredicate());
1537 if (TM.Options.NoNaNsFPMath)
1538 Condition = getFCmpCodeWithoutNaN(Condition);
1541 CaseBlock CB(Condition, BOp->getOperand(0), BOp->getOperand(1), nullptr,
1542 TBB, FBB, CurBB, TProb, FProb);
1543 SwitchCases.push_back(CB);
1548 // Create a CaseBlock record representing this branch.
1549 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()),
1550 nullptr, TBB, FBB, CurBB, TProb, FProb);
1551 SwitchCases.push_back(CB);
1554 /// FindMergedConditions - If Cond is an expression like
1555 void SelectionDAGBuilder::FindMergedConditions(const Value *Cond,
1556 MachineBasicBlock *TBB,
1557 MachineBasicBlock *FBB,
1558 MachineBasicBlock *CurBB,
1559 MachineBasicBlock *SwitchBB,
1560 Instruction::BinaryOps Opc,
1561 BranchProbability TProb,
1562 BranchProbability FProb) {
1563 // If this node is not part of the or/and tree, emit it as a branch.
1564 const Instruction *BOp = dyn_cast<Instruction>(Cond);
1565 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
1566 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
1567 BOp->getParent() != CurBB->getBasicBlock() ||
1568 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
1569 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
1570 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB,
1575 // Create TmpBB after CurBB.
1576 MachineFunction::iterator BBI(CurBB);
1577 MachineFunction &MF = DAG.getMachineFunction();
1578 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock());
1579 CurBB->getParent()->insert(++BBI, TmpBB);
1581 if (Opc == Instruction::Or) {
1582 // Codegen X | Y as:
1591 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
1592 // The requirement is that
1593 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
1594 // = TrueProb for original BB.
1595 // Assuming the original probabilities are A and B, one choice is to set
1596 // BB1's probabilities to A/2 and A/2+B, and set TmpBB's probabilities to
1597 // A/(1+B) and 2B/(1+B). This choice assumes that
1598 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
1599 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
1600 // TmpBB, but the math is more complicated.
1602 auto NewTrueProb = TProb / 2;
1603 auto NewFalseProb = TProb / 2 + FProb;
1604 // Emit the LHS condition.
1605 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, SwitchBB, Opc,
1606 NewTrueProb, NewFalseProb);
1608 // Normalize A/2 and B to get A/(1+B) and 2B/(1+B).
1609 SmallVector<BranchProbability, 2> Probs{TProb / 2, FProb};
1610 BranchProbability::normalizeProbabilities(Probs.begin(), Probs.end());
1611 // Emit the RHS condition into TmpBB.
1612 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc,
1613 Probs[0], Probs[1]);
1615 assert(Opc == Instruction::And && "Unknown merge op!");
1616 // Codegen X & Y as:
1624 // This requires creation of TmpBB after CurBB.
1626 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
1627 // The requirement is that
1628 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
1629 // = FalseProb for original BB.
1630 // Assuming the original probabilities are A and B, one choice is to set
1631 // BB1's probabilities to A+B/2 and B/2, and set TmpBB's probabilities to
1632 // 2A/(1+A) and B/(1+A). This choice assumes that FalseProb for BB1 ==
1633 // TrueProb for BB1 * FalseProb for TmpBB.
1635 auto NewTrueProb = TProb + FProb / 2;
1636 auto NewFalseProb = FProb / 2;
1637 // Emit the LHS condition.
1638 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, SwitchBB, Opc,
1639 NewTrueProb, NewFalseProb);
1641 // Normalize A and B/2 to get 2A/(1+A) and B/(1+A).
1642 SmallVector<BranchProbability, 2> Probs{TProb, FProb / 2};
1643 BranchProbability::normalizeProbabilities(Probs.begin(), Probs.end());
1644 // Emit the RHS condition into TmpBB.
1645 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc,
1646 Probs[0], Probs[1]);
1650 /// If the set of cases should be emitted as a series of branches, return true.
1651 /// If we should emit this as a bunch of and/or'd together conditions, return
1654 SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases) {
1655 if (Cases.size() != 2) return true;
1657 // If this is two comparisons of the same values or'd or and'd together, they
1658 // will get folded into a single comparison, so don't emit two blocks.
1659 if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
1660 Cases[0].CmpRHS == Cases[1].CmpRHS) ||
1661 (Cases[0].CmpRHS == Cases[1].CmpLHS &&
1662 Cases[0].CmpLHS == Cases[1].CmpRHS)) {
1666 // Handle: (X != null) | (Y != null) --> (X|Y) != 0
1667 // Handle: (X == null) & (Y == null) --> (X|Y) == 0
1668 if (Cases[0].CmpRHS == Cases[1].CmpRHS &&
1669 Cases[0].CC == Cases[1].CC &&
1670 isa<Constant>(Cases[0].CmpRHS) &&
1671 cast<Constant>(Cases[0].CmpRHS)->isNullValue()) {
1672 if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB)
1674 if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB)
1681 void SelectionDAGBuilder::visitBr(const BranchInst &I) {
1682 MachineBasicBlock *BrMBB = FuncInfo.MBB;
1684 // Update machine-CFG edges.
1685 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
1687 if (I.isUnconditional()) {
1688 // Update machine-CFG edges.
1689 BrMBB->addSuccessor(Succ0MBB);
1691 // If this is not a fall-through branch or optimizations are switched off,
1693 if (Succ0MBB != NextBlock(BrMBB) || TM.getOptLevel() == CodeGenOpt::None)
1694 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(),
1695 MVT::Other, getControlRoot(),
1696 DAG.getBasicBlock(Succ0MBB)));
1701 // If this condition is one of the special cases we handle, do special stuff
1703 const Value *CondVal = I.getCondition();
1704 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
1706 // If this is a series of conditions that are or'd or and'd together, emit
1707 // this as a sequence of branches instead of setcc's with and/or operations.
1708 // As long as jumps are not expensive, this should improve performance.
1709 // For example, instead of something like:
1722 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
1723 Instruction::BinaryOps Opcode = BOp->getOpcode();
1724 if (!DAG.getTargetLoweringInfo().isJumpExpensive() && BOp->hasOneUse() &&
1725 !I.getMetadata(LLVMContext::MD_unpredictable) &&
1726 (Opcode == Instruction::And || Opcode == Instruction::Or)) {
1727 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB,
1729 getEdgeProbability(BrMBB, Succ0MBB),
1730 getEdgeProbability(BrMBB, Succ1MBB));
1731 // If the compares in later blocks need to use values not currently
1732 // exported from this block, export them now. This block should always
1733 // be the first entry.
1734 assert(SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!");
1736 // Allow some cases to be rejected.
1737 if (ShouldEmitAsBranches(SwitchCases)) {
1738 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1739 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1740 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1743 // Emit the branch for this block.
1744 visitSwitchCase(SwitchCases[0], BrMBB);
1745 SwitchCases.erase(SwitchCases.begin());
1749 // Okay, we decided not to do this, remove any inserted MBB's and clear
1751 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
1752 FuncInfo.MF->erase(SwitchCases[i].ThisBB);
1754 SwitchCases.clear();
1758 // Create a CaseBlock record representing this branch.
1759 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()),
1760 nullptr, Succ0MBB, Succ1MBB, BrMBB);
1762 // Use visitSwitchCase to actually insert the fast branch sequence for this
1764 visitSwitchCase(CB, BrMBB);
1767 /// visitSwitchCase - Emits the necessary code to represent a single node in
1768 /// the binary search tree resulting from lowering a switch instruction.
1769 void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB,
1770 MachineBasicBlock *SwitchBB) {
1772 SDValue CondLHS = getValue(CB.CmpLHS);
1773 SDLoc dl = getCurSDLoc();
1775 // Build the setcc now.
1777 // Fold "(X == true)" to X and "(X == false)" to !X to
1778 // handle common cases produced by branch lowering.
1779 if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) &&
1780 CB.CC == ISD::SETEQ)
1782 else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) &&
1783 CB.CC == ISD::SETEQ) {
1784 SDValue True = DAG.getConstant(1, dl, CondLHS.getValueType());
1785 Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True);
1787 Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1789 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now");
1791 const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue();
1792 const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue();
1794 SDValue CmpOp = getValue(CB.CmpMHS);
1795 EVT VT = CmpOp.getValueType();
1797 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) {
1798 Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, dl, VT),
1801 SDValue SUB = DAG.getNode(ISD::SUB, dl,
1802 VT, CmpOp, DAG.getConstant(Low, dl, VT));
1803 Cond = DAG.getSetCC(dl, MVT::i1, SUB,
1804 DAG.getConstant(High-Low, dl, VT), ISD::SETULE);
1808 // Update successor info
1809 addSuccessorWithProb(SwitchBB, CB.TrueBB, CB.TrueProb);
1810 // TrueBB and FalseBB are always different unless the incoming IR is
1811 // degenerate. This only happens when running llc on weird IR.
1812 if (CB.TrueBB != CB.FalseBB)
1813 addSuccessorWithProb(SwitchBB, CB.FalseBB, CB.FalseProb);
1814 SwitchBB->normalizeSuccProbs();
1816 // If the lhs block is the next block, invert the condition so that we can
1817 // fall through to the lhs instead of the rhs block.
1818 if (CB.TrueBB == NextBlock(SwitchBB)) {
1819 std::swap(CB.TrueBB, CB.FalseBB);
1820 SDValue True = DAG.getConstant(1, dl, Cond.getValueType());
1821 Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True);
1824 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1825 MVT::Other, getControlRoot(), Cond,
1826 DAG.getBasicBlock(CB.TrueBB));
1828 // Insert the false branch. Do this even if it's a fall through branch,
1829 // this makes it easier to do DAG optimizations which require inverting
1830 // the branch condition.
1831 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1832 DAG.getBasicBlock(CB.FalseBB));
1834 DAG.setRoot(BrCond);
1837 /// visitJumpTable - Emit JumpTable node in the current MBB
1838 void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) {
1839 // Emit the code for the jump table
1840 assert(JT.Reg != -1U && "Should lower JT Header first!");
1841 EVT PTy = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
1842 SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurSDLoc(),
1844 SDValue Table = DAG.getJumpTable(JT.JTI, PTy);
1845 SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurSDLoc(),
1846 MVT::Other, Index.getValue(1),
1848 DAG.setRoot(BrJumpTable);
1851 /// visitJumpTableHeader - This function emits necessary code to produce index
1852 /// in the JumpTable from switch case.
1853 void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT,
1854 JumpTableHeader &JTH,
1855 MachineBasicBlock *SwitchBB) {
1856 SDLoc dl = getCurSDLoc();
1858 // Subtract the lowest switch case value from the value being switched on and
1859 // conditional branch to default mbb if the result is greater than the
1860 // difference between smallest and largest cases.
1861 SDValue SwitchOp = getValue(JTH.SValue);
1862 EVT VT = SwitchOp.getValueType();
1863 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, SwitchOp,
1864 DAG.getConstant(JTH.First, dl, VT));
1866 // The SDNode we just created, which holds the value being switched on minus
1867 // the smallest case value, needs to be copied to a virtual register so it
1868 // can be used as an index into the jump table in a subsequent basic block.
1869 // This value may be smaller or larger than the target's pointer type, and
1870 // therefore require extension or truncating.
1871 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1872 SwitchOp = DAG.getZExtOrTrunc(Sub, dl, TLI.getPointerTy(DAG.getDataLayout()));
1874 unsigned JumpTableReg =
1875 FuncInfo.CreateReg(TLI.getPointerTy(DAG.getDataLayout()));
1876 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), dl,
1877 JumpTableReg, SwitchOp);
1878 JT.Reg = JumpTableReg;
1880 // Emit the range check for the jump table, and branch to the default block
1881 // for the switch statement if the value being switched on exceeds the largest
1882 // case in the switch.
1883 SDValue CMP = DAG.getSetCC(
1884 dl, TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(),
1885 Sub.getValueType()),
1886 Sub, DAG.getConstant(JTH.Last - JTH.First, dl, VT), ISD::SETUGT);
1888 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1889 MVT::Other, CopyTo, CMP,
1890 DAG.getBasicBlock(JT.Default));
1892 // Avoid emitting unnecessary branches to the next block.
1893 if (JT.MBB != NextBlock(SwitchBB))
1894 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1895 DAG.getBasicBlock(JT.MBB));
1897 DAG.setRoot(BrCond);
1900 /// Codegen a new tail for a stack protector check ParentMBB which has had its
1901 /// tail spliced into a stack protector check success bb.
1903 /// For a high level explanation of how this fits into the stack protector
1904 /// generation see the comment on the declaration of class
1905 /// StackProtectorDescriptor.
1906 void SelectionDAGBuilder::visitSPDescriptorParent(StackProtectorDescriptor &SPD,
1907 MachineBasicBlock *ParentBB) {
1909 // First create the loads to the guard/stack slot for the comparison.
1910 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1911 EVT PtrTy = TLI.getPointerTy(DAG.getDataLayout());
1913 MachineFrameInfo *MFI = ParentBB->getParent()->getFrameInfo();
1914 int FI = MFI->getStackProtectorIndex();
1916 const Value *IRGuard = SPD.getGuard();
1917 SDValue GuardPtr = getValue(IRGuard);
1918 SDValue StackSlotPtr = DAG.getFrameIndex(FI, PtrTy);
1920 unsigned Align = DL->getPrefTypeAlignment(IRGuard->getType());
1923 SDLoc dl = getCurSDLoc();
1925 // If GuardReg is set and useLoadStackGuardNode returns true, retrieve the
1926 // guard value from the virtual register holding the value. Otherwise, emit a
1927 // volatile load to retrieve the stack guard value.
1928 unsigned GuardReg = SPD.getGuardReg();
1930 if (GuardReg && TLI.useLoadStackGuardNode())
1931 Guard = DAG.getCopyFromReg(DAG.getEntryNode(), dl, GuardReg,
1934 Guard = DAG.getLoad(PtrTy, dl, DAG.getEntryNode(),
1935 GuardPtr, MachinePointerInfo(IRGuard, 0),
1936 true, false, false, Align);
1938 SDValue StackSlot = DAG.getLoad(
1939 PtrTy, dl, DAG.getEntryNode(), StackSlotPtr,
1940 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), true,
1941 false, false, Align);
1943 // Perform the comparison via a subtract/getsetcc.
1944 EVT VT = Guard.getValueType();
1945 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, Guard, StackSlot);
1947 SDValue Cmp = DAG.getSetCC(dl, TLI.getSetCCResultType(DAG.getDataLayout(),
1949 Sub.getValueType()),
1950 Sub, DAG.getConstant(0, dl, VT), ISD::SETNE);
1952 // If the sub is not 0, then we know the guard/stackslot do not equal, so
1953 // branch to failure MBB.
1954 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1955 MVT::Other, StackSlot.getOperand(0),
1956 Cmp, DAG.getBasicBlock(SPD.getFailureMBB()));
1957 // Otherwise branch to success MBB.
1958 SDValue Br = DAG.getNode(ISD::BR, dl,
1960 DAG.getBasicBlock(SPD.getSuccessMBB()));
1965 /// Codegen the failure basic block for a stack protector check.
1967 /// A failure stack protector machine basic block consists simply of a call to
1968 /// __stack_chk_fail().
1970 /// For a high level explanation of how this fits into the stack protector
1971 /// generation see the comment on the declaration of class
1972 /// StackProtectorDescriptor.
1974 SelectionDAGBuilder::visitSPDescriptorFailure(StackProtectorDescriptor &SPD) {
1975 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1977 TLI.makeLibCall(DAG, RTLIB::STACKPROTECTOR_CHECK_FAIL, MVT::isVoid,
1978 None, false, getCurSDLoc(), false, false).second;
1982 /// visitBitTestHeader - This function emits necessary code to produce value
1983 /// suitable for "bit tests"
1984 void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B,
1985 MachineBasicBlock *SwitchBB) {
1986 SDLoc dl = getCurSDLoc();
1988 // Subtract the minimum value
1989 SDValue SwitchOp = getValue(B.SValue);
1990 EVT VT = SwitchOp.getValueType();
1991 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, SwitchOp,
1992 DAG.getConstant(B.First, dl, VT));
1995 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1996 SDValue RangeCmp = DAG.getSetCC(
1997 dl, TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(),
1998 Sub.getValueType()),
1999 Sub, DAG.getConstant(B.Range, dl, VT), ISD::SETUGT);
2001 // Determine the type of the test operands.
2002 bool UsePtrType = false;
2003 if (!TLI.isTypeLegal(VT))
2006 for (unsigned i = 0, e = B.Cases.size(); i != e; ++i)
2007 if (!isUIntN(VT.getSizeInBits(), B.Cases[i].Mask)) {
2008 // Switch table case range are encoded into series of masks.
2009 // Just use pointer type, it's guaranteed to fit.
2015 VT = TLI.getPointerTy(DAG.getDataLayout());
2016 Sub = DAG.getZExtOrTrunc(Sub, dl, VT);
2019 B.RegVT = VT.getSimpleVT();
2020 B.Reg = FuncInfo.CreateReg(B.RegVT);
2021 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), dl, B.Reg, Sub);
2023 MachineBasicBlock* MBB = B.Cases[0].ThisBB;
2025 addSuccessorWithProb(SwitchBB, B.Default, B.DefaultProb);
2026 addSuccessorWithProb(SwitchBB, MBB, B.Prob);
2027 SwitchBB->normalizeSuccProbs();
2029 SDValue BrRange = DAG.getNode(ISD::BRCOND, dl,
2030 MVT::Other, CopyTo, RangeCmp,
2031 DAG.getBasicBlock(B.Default));
2033 // Avoid emitting unnecessary branches to the next block.
2034 if (MBB != NextBlock(SwitchBB))
2035 BrRange = DAG.getNode(ISD::BR, dl, MVT::Other, BrRange,
2036 DAG.getBasicBlock(MBB));
2038 DAG.setRoot(BrRange);
2041 /// visitBitTestCase - this function produces one "bit test"
2042 void SelectionDAGBuilder::visitBitTestCase(BitTestBlock &BB,
2043 MachineBasicBlock* NextMBB,
2044 BranchProbability BranchProbToNext,
2047 MachineBasicBlock *SwitchBB) {
2048 SDLoc dl = getCurSDLoc();
2050 SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), dl, Reg, VT);
2052 unsigned PopCount = countPopulation(B.Mask);
2053 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2054 if (PopCount == 1) {
2055 // Testing for a single bit; just compare the shift count with what it
2056 // would need to be to shift a 1 bit in that position.
2058 dl, TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT),
2059 ShiftOp, DAG.getConstant(countTrailingZeros(B.Mask), dl, VT),
2061 } else if (PopCount == BB.Range) {
2062 // There is only one zero bit in the range, test for it directly.
2064 dl, TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT),
2065 ShiftOp, DAG.getConstant(countTrailingOnes(B.Mask), dl, VT),
2068 // Make desired shift
2069 SDValue SwitchVal = DAG.getNode(ISD::SHL, dl, VT,
2070 DAG.getConstant(1, dl, VT), ShiftOp);
2072 // Emit bit tests and jumps
2073 SDValue AndOp = DAG.getNode(ISD::AND, dl,
2074 VT, SwitchVal, DAG.getConstant(B.Mask, dl, VT));
2076 dl, TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT),
2077 AndOp, DAG.getConstant(0, dl, VT), ISD::SETNE);
2080 // The branch probability from SwitchBB to B.TargetBB is B.ExtraProb.
2081 addSuccessorWithProb(SwitchBB, B.TargetBB, B.ExtraProb);
2082 // The branch probability from SwitchBB to NextMBB is BranchProbToNext.
2083 addSuccessorWithProb(SwitchBB, NextMBB, BranchProbToNext);
2084 // It is not guaranteed that the sum of B.ExtraProb and BranchProbToNext is
2085 // one as they are relative probabilities (and thus work more like weights),
2086 // and hence we need to normalize them to let the sum of them become one.
2087 SwitchBB->normalizeSuccProbs();
2089 SDValue BrAnd = DAG.getNode(ISD::BRCOND, dl,
2090 MVT::Other, getControlRoot(),
2091 Cmp, DAG.getBasicBlock(B.TargetBB));
2093 // Avoid emitting unnecessary branches to the next block.
2094 if (NextMBB != NextBlock(SwitchBB))
2095 BrAnd = DAG.getNode(ISD::BR, dl, MVT::Other, BrAnd,
2096 DAG.getBasicBlock(NextMBB));
2101 void SelectionDAGBuilder::visitInvoke(const InvokeInst &I) {
2102 MachineBasicBlock *InvokeMBB = FuncInfo.MBB;
2104 // Retrieve successors. Look through artificial IR level blocks like
2105 // catchswitch for successors.
2106 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
2107 const BasicBlock *EHPadBB = I.getSuccessor(1);
2109 const Value *Callee(I.getCalledValue());
2110 const Function *Fn = dyn_cast<Function>(Callee);
2111 if (isa<InlineAsm>(Callee))
2113 else if (Fn && Fn->isIntrinsic()) {
2114 switch (Fn->getIntrinsicID()) {
2116 llvm_unreachable("Cannot invoke this intrinsic");
2117 case Intrinsic::donothing:
2118 // Ignore invokes to @llvm.donothing: jump directly to the next BB.
2120 case Intrinsic::experimental_patchpoint_void:
2121 case Intrinsic::experimental_patchpoint_i64:
2122 visitPatchpoint(&I, EHPadBB);
2124 case Intrinsic::experimental_gc_statepoint:
2125 LowerStatepoint(ImmutableStatepoint(&I), EHPadBB);
2129 LowerCallTo(&I, getValue(Callee), false, EHPadBB);
2131 // If the value of the invoke is used outside of its defining block, make it
2132 // available as a virtual register.
2133 // We already took care of the exported value for the statepoint instruction
2134 // during call to the LowerStatepoint.
2135 if (!isStatepoint(I)) {
2136 CopyToExportRegsIfNeeded(&I);
2139 SmallVector<std::pair<MachineBasicBlock *, BranchProbability>, 1> UnwindDests;
2140 BranchProbabilityInfo *BPI = FuncInfo.BPI;
2141 BranchProbability EHPadBBProb =
2142 BPI ? BPI->getEdgeProbability(InvokeMBB->getBasicBlock(), EHPadBB)
2143 : BranchProbability::getZero();
2144 findUnwindDestinations(FuncInfo, EHPadBB, EHPadBBProb, UnwindDests);
2146 // Update successor info.
2147 addSuccessorWithProb(InvokeMBB, Return);
2148 for (auto &UnwindDest : UnwindDests) {
2149 UnwindDest.first->setIsEHPad();
2150 addSuccessorWithProb(InvokeMBB, UnwindDest.first, UnwindDest.second);
2152 InvokeMBB->normalizeSuccProbs();
2154 // Drop into normal successor.
2155 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(),
2156 MVT::Other, getControlRoot(),
2157 DAG.getBasicBlock(Return)));
2160 void SelectionDAGBuilder::visitResume(const ResumeInst &RI) {
2161 llvm_unreachable("SelectionDAGBuilder shouldn't visit resume instructions!");
2164 void SelectionDAGBuilder::visitLandingPad(const LandingPadInst &LP) {
2165 assert(FuncInfo.MBB->isEHPad() &&
2166 "Call to landingpad not in landing pad!");
2168 MachineBasicBlock *MBB = FuncInfo.MBB;
2169 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
2170 AddLandingPadInfo(LP, MMI, MBB);
2172 // If there aren't registers to copy the values into (e.g., during SjLj
2173 // exceptions), then don't bother to create these DAG nodes.
2174 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2175 const Constant *PersonalityFn = FuncInfo.Fn->getPersonalityFn();
2176 if (TLI.getExceptionPointerRegister(PersonalityFn) == 0 &&
2177 TLI.getExceptionSelectorRegister(PersonalityFn) == 0)
2180 // If landingpad's return type is token type, we don't create DAG nodes
2181 // for its exception pointer and selector value. The extraction of exception
2182 // pointer or selector value from token type landingpads is not currently
2184 if (LP.getType()->isTokenTy())
2187 SmallVector<EVT, 2> ValueVTs;
2188 SDLoc dl = getCurSDLoc();
2189 ComputeValueVTs(TLI, DAG.getDataLayout(), LP.getType(), ValueVTs);
2190 assert(ValueVTs.size() == 2 && "Only two-valued landingpads are supported");
2192 // Get the two live-in registers as SDValues. The physregs have already been
2193 // copied into virtual registers.
2195 if (FuncInfo.ExceptionPointerVirtReg) {
2196 Ops[0] = DAG.getZExtOrTrunc(
2197 DAG.getCopyFromReg(DAG.getEntryNode(), dl,
2198 FuncInfo.ExceptionPointerVirtReg,
2199 TLI.getPointerTy(DAG.getDataLayout())),
2202 Ops[0] = DAG.getConstant(0, dl, TLI.getPointerTy(DAG.getDataLayout()));
2204 Ops[1] = DAG.getZExtOrTrunc(
2205 DAG.getCopyFromReg(DAG.getEntryNode(), dl,
2206 FuncInfo.ExceptionSelectorVirtReg,
2207 TLI.getPointerTy(DAG.getDataLayout())),
2211 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, dl,
2212 DAG.getVTList(ValueVTs), Ops);
2216 void SelectionDAGBuilder::sortAndRangeify(CaseClusterVector &Clusters) {
2218 for (const CaseCluster &CC : Clusters)
2219 assert(CC.Low == CC.High && "Input clusters must be single-case");
2222 std::sort(Clusters.begin(), Clusters.end(),
2223 [](const CaseCluster &a, const CaseCluster &b) {
2224 return a.Low->getValue().slt(b.Low->getValue());
2227 // Merge adjacent clusters with the same destination.
2228 const unsigned N = Clusters.size();
2229 unsigned DstIndex = 0;
2230 for (unsigned SrcIndex = 0; SrcIndex < N; ++SrcIndex) {
2231 CaseCluster &CC = Clusters[SrcIndex];
2232 const ConstantInt *CaseVal = CC.Low;
2233 MachineBasicBlock *Succ = CC.MBB;
2235 if (DstIndex != 0 && Clusters[DstIndex - 1].MBB == Succ &&
2236 (CaseVal->getValue() - Clusters[DstIndex - 1].High->getValue()) == 1) {
2237 // If this case has the same successor and is a neighbour, merge it into
2238 // the previous cluster.
2239 Clusters[DstIndex - 1].High = CaseVal;
2240 Clusters[DstIndex - 1].Prob += CC.Prob;
2242 std::memmove(&Clusters[DstIndex++], &Clusters[SrcIndex],
2243 sizeof(Clusters[SrcIndex]));
2246 Clusters.resize(DstIndex);
2249 void SelectionDAGBuilder::UpdateSplitBlock(MachineBasicBlock *First,
2250 MachineBasicBlock *Last) {
2252 for (unsigned i = 0, e = JTCases.size(); i != e; ++i)
2253 if (JTCases[i].first.HeaderBB == First)
2254 JTCases[i].first.HeaderBB = Last;
2256 // Update BitTestCases.
2257 for (unsigned i = 0, e = BitTestCases.size(); i != e; ++i)
2258 if (BitTestCases[i].Parent == First)
2259 BitTestCases[i].Parent = Last;
2262 void SelectionDAGBuilder::visitIndirectBr(const IndirectBrInst &I) {
2263 MachineBasicBlock *IndirectBrMBB = FuncInfo.MBB;
2265 // Update machine-CFG edges with unique successors.
2266 SmallSet<BasicBlock*, 32> Done;
2267 for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i) {
2268 BasicBlock *BB = I.getSuccessor(i);
2269 bool Inserted = Done.insert(BB).second;
2273 MachineBasicBlock *Succ = FuncInfo.MBBMap[BB];
2274 addSuccessorWithProb(IndirectBrMBB, Succ);
2276 IndirectBrMBB->normalizeSuccProbs();
2278 DAG.setRoot(DAG.getNode(ISD::BRIND, getCurSDLoc(),
2279 MVT::Other, getControlRoot(),
2280 getValue(I.getAddress())));
2283 void SelectionDAGBuilder::visitUnreachable(const UnreachableInst &I) {
2284 if (DAG.getTarget().Options.TrapUnreachable)
2286 DAG.getNode(ISD::TRAP, getCurSDLoc(), MVT::Other, DAG.getRoot()));
2289 void SelectionDAGBuilder::visitFSub(const User &I) {
2290 // -0.0 - X --> fneg
2291 Type *Ty = I.getType();
2292 if (isa<Constant>(I.getOperand(0)) &&
2293 I.getOperand(0) == ConstantFP::getZeroValueForNegation(Ty)) {
2294 SDValue Op2 = getValue(I.getOperand(1));
2295 setValue(&I, DAG.getNode(ISD::FNEG, getCurSDLoc(),
2296 Op2.getValueType(), Op2));
2300 visitBinary(I, ISD::FSUB);
2303 void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) {
2304 SDValue Op1 = getValue(I.getOperand(0));
2305 SDValue Op2 = getValue(I.getOperand(1));
2312 if (const OverflowingBinaryOperator *OFBinOp =
2313 dyn_cast<const OverflowingBinaryOperator>(&I)) {
2314 nuw = OFBinOp->hasNoUnsignedWrap();
2315 nsw = OFBinOp->hasNoSignedWrap();
2317 if (const PossiblyExactOperator *ExactOp =
2318 dyn_cast<const PossiblyExactOperator>(&I))
2319 exact = ExactOp->isExact();
2320 if (const FPMathOperator *FPOp = dyn_cast<const FPMathOperator>(&I))
2321 FMF = FPOp->getFastMathFlags();
2324 Flags.setExact(exact);
2325 Flags.setNoSignedWrap(nsw);
2326 Flags.setNoUnsignedWrap(nuw);
2327 if (EnableFMFInDAG) {
2328 Flags.setAllowReciprocal(FMF.allowReciprocal());
2329 Flags.setNoInfs(FMF.noInfs());
2330 Flags.setNoNaNs(FMF.noNaNs());
2331 Flags.setNoSignedZeros(FMF.noSignedZeros());
2332 Flags.setUnsafeAlgebra(FMF.unsafeAlgebra());
2334 SDValue BinNodeValue = DAG.getNode(OpCode, getCurSDLoc(), Op1.getValueType(),
2336 setValue(&I, BinNodeValue);
2339 void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) {
2340 SDValue Op1 = getValue(I.getOperand(0));
2341 SDValue Op2 = getValue(I.getOperand(1));
2343 EVT ShiftTy = DAG.getTargetLoweringInfo().getShiftAmountTy(
2344 Op2.getValueType(), DAG.getDataLayout());
2346 // Coerce the shift amount to the right type if we can.
2347 if (!I.getType()->isVectorTy() && Op2.getValueType() != ShiftTy) {
2348 unsigned ShiftSize = ShiftTy.getSizeInBits();
2349 unsigned Op2Size = Op2.getValueType().getSizeInBits();
2350 SDLoc DL = getCurSDLoc();
2352 // If the operand is smaller than the shift count type, promote it.
2353 if (ShiftSize > Op2Size)
2354 Op2 = DAG.getNode(ISD::ZERO_EXTEND, DL, ShiftTy, Op2);
2356 // If the operand is larger than the shift count type but the shift
2357 // count type has enough bits to represent any shift value, truncate
2358 // it now. This is a common case and it exposes the truncate to
2359 // optimization early.
2360 else if (ShiftSize >= Log2_32_Ceil(Op2.getValueType().getSizeInBits()))
2361 Op2 = DAG.getNode(ISD::TRUNCATE, DL, ShiftTy, Op2);
2362 // Otherwise we'll need to temporarily settle for some other convenient
2363 // type. Type legalization will make adjustments once the shiftee is split.
2365 Op2 = DAG.getZExtOrTrunc(Op2, DL, MVT::i32);
2372 if (Opcode == ISD::SRL || Opcode == ISD::SRA || Opcode == ISD::SHL) {
2374 if (const OverflowingBinaryOperator *OFBinOp =
2375 dyn_cast<const OverflowingBinaryOperator>(&I)) {
2376 nuw = OFBinOp->hasNoUnsignedWrap();
2377 nsw = OFBinOp->hasNoSignedWrap();
2379 if (const PossiblyExactOperator *ExactOp =
2380 dyn_cast<const PossiblyExactOperator>(&I))
2381 exact = ExactOp->isExact();
2384 Flags.setExact(exact);
2385 Flags.setNoSignedWrap(nsw);
2386 Flags.setNoUnsignedWrap(nuw);
2387 SDValue Res = DAG.getNode(Opcode, getCurSDLoc(), Op1.getValueType(), Op1, Op2,
2392 void SelectionDAGBuilder::visitSDiv(const User &I) {
2393 SDValue Op1 = getValue(I.getOperand(0));
2394 SDValue Op2 = getValue(I.getOperand(1));
2397 Flags.setExact(isa<PossiblyExactOperator>(&I) &&
2398 cast<PossiblyExactOperator>(&I)->isExact());
2399 setValue(&I, DAG.getNode(ISD::SDIV, getCurSDLoc(), Op1.getValueType(), Op1,
2403 void SelectionDAGBuilder::visitICmp(const User &I) {
2404 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2405 if (const ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2406 predicate = IC->getPredicate();
2407 else if (const ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2408 predicate = ICmpInst::Predicate(IC->getPredicate());
2409 SDValue Op1 = getValue(I.getOperand(0));
2410 SDValue Op2 = getValue(I.getOperand(1));
2411 ISD::CondCode Opcode = getICmpCondCode(predicate);
2413 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2415 setValue(&I, DAG.getSetCC(getCurSDLoc(), DestVT, Op1, Op2, Opcode));
2418 void SelectionDAGBuilder::visitFCmp(const User &I) {
2419 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2420 if (const FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2421 predicate = FC->getPredicate();
2422 else if (const ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2423 predicate = FCmpInst::Predicate(FC->getPredicate());
2424 SDValue Op1 = getValue(I.getOperand(0));
2425 SDValue Op2 = getValue(I.getOperand(1));
2426 ISD::CondCode Condition = getFCmpCondCode(predicate);
2428 // FIXME: Fcmp instructions have fast-math-flags in IR, so we should use them.
2429 // FIXME: We should propagate the fast-math-flags to the DAG node itself for
2430 // further optimization, but currently FMF is only applicable to binary nodes.
2431 if (TM.Options.NoNaNsFPMath)
2432 Condition = getFCmpCodeWithoutNaN(Condition);
2433 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2435 setValue(&I, DAG.getSetCC(getCurSDLoc(), DestVT, Op1, Op2, Condition));
2438 void SelectionDAGBuilder::visitSelect(const User &I) {
2439 SmallVector<EVT, 4> ValueVTs;
2440 ComputeValueVTs(DAG.getTargetLoweringInfo(), DAG.getDataLayout(), I.getType(),
2442 unsigned NumValues = ValueVTs.size();
2443 if (NumValues == 0) return;
2445 SmallVector<SDValue, 4> Values(NumValues);
2446 SDValue Cond = getValue(I.getOperand(0));
2447 SDValue LHSVal = getValue(I.getOperand(1));
2448 SDValue RHSVal = getValue(I.getOperand(2));
2449 auto BaseOps = {Cond};
2450 ISD::NodeType OpCode = Cond.getValueType().isVector() ?
2451 ISD::VSELECT : ISD::SELECT;
2453 // Min/max matching is only viable if all output VTs are the same.
2454 if (std::equal(ValueVTs.begin(), ValueVTs.end(), ValueVTs.begin())) {
2455 EVT VT = ValueVTs[0];
2456 LLVMContext &Ctx = *DAG.getContext();
2457 auto &TLI = DAG.getTargetLoweringInfo();
2459 // We care about the legality of the operation after it has been type
2461 while (TLI.getTypeAction(Ctx, VT) != TargetLoweringBase::TypeLegal &&
2462 VT != TLI.getTypeToTransformTo(Ctx, VT))
2463 VT = TLI.getTypeToTransformTo(Ctx, VT);
2465 // If the vselect is legal, assume we want to leave this as a vector setcc +
2466 // vselect. Otherwise, if this is going to be scalarized, we want to see if
2467 // min/max is legal on the scalar type.
2468 bool UseScalarMinMax = VT.isVector() &&
2469 !TLI.isOperationLegalOrCustom(ISD::VSELECT, VT);
2472 auto SPR = matchSelectPattern(const_cast<User*>(&I), LHS, RHS);
2473 ISD::NodeType Opc = ISD::DELETED_NODE;
2474 switch (SPR.Flavor) {
2475 case SPF_UMAX: Opc = ISD::UMAX; break;
2476 case SPF_UMIN: Opc = ISD::UMIN; break;
2477 case SPF_SMAX: Opc = ISD::SMAX; break;
2478 case SPF_SMIN: Opc = ISD::SMIN; break;
2480 switch (SPR.NaNBehavior) {
2481 case SPNB_NA: llvm_unreachable("No NaN behavior for FP op?");
2482 case SPNB_RETURNS_NAN: Opc = ISD::FMINNAN; break;
2483 case SPNB_RETURNS_OTHER: Opc = ISD::FMINNUM; break;
2484 case SPNB_RETURNS_ANY: {
2485 if (TLI.isOperationLegalOrCustom(ISD::FMINNUM, VT))
2487 else if (TLI.isOperationLegalOrCustom(ISD::FMINNAN, VT))
2489 else if (UseScalarMinMax)
2490 Opc = TLI.isOperationLegalOrCustom(ISD::FMINNUM, VT.getScalarType()) ?
2491 ISD::FMINNUM : ISD::FMINNAN;
2497 switch (SPR.NaNBehavior) {
2498 case SPNB_NA: llvm_unreachable("No NaN behavior for FP op?");
2499 case SPNB_RETURNS_NAN: Opc = ISD::FMAXNAN; break;
2500 case SPNB_RETURNS_OTHER: Opc = ISD::FMAXNUM; break;
2501 case SPNB_RETURNS_ANY:
2503 if (TLI.isOperationLegalOrCustom(ISD::FMAXNUM, VT))
2505 else if (TLI.isOperationLegalOrCustom(ISD::FMAXNAN, VT))
2507 else if (UseScalarMinMax)
2508 Opc = TLI.isOperationLegalOrCustom(ISD::FMAXNUM, VT.getScalarType()) ?
2509 ISD::FMAXNUM : ISD::FMAXNAN;
2516 if (Opc != ISD::DELETED_NODE &&
2517 (TLI.isOperationLegalOrCustom(Opc, VT) ||
2519 TLI.isOperationLegalOrCustom(Opc, VT.getScalarType()))) &&
2520 // If the underlying comparison instruction is used by any other
2521 // instruction, the consumed instructions won't be destroyed, so it is
2522 // not profitable to convert to a min/max.
2523 cast<SelectInst>(&I)->getCondition()->hasOneUse()) {
2525 LHSVal = getValue(LHS);
2526 RHSVal = getValue(RHS);
2531 for (unsigned i = 0; i != NumValues; ++i) {
2532 SmallVector<SDValue, 3> Ops(BaseOps.begin(), BaseOps.end());
2533 Ops.push_back(SDValue(LHSVal.getNode(), LHSVal.getResNo() + i));
2534 Ops.push_back(SDValue(RHSVal.getNode(), RHSVal.getResNo() + i));
2535 Values[i] = DAG.getNode(OpCode, getCurSDLoc(),
2536 LHSVal.getNode()->getValueType(LHSVal.getResNo()+i),
2540 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
2541 DAG.getVTList(ValueVTs), Values));
2544 void SelectionDAGBuilder::visitTrunc(const User &I) {
2545 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2546 SDValue N = getValue(I.getOperand(0));
2547 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2549 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), DestVT, N));
2552 void SelectionDAGBuilder::visitZExt(const User &I) {
2553 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2554 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2555 SDValue N = getValue(I.getOperand(0));
2556 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2558 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurSDLoc(), DestVT, N));
2561 void SelectionDAGBuilder::visitSExt(const User &I) {
2562 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2563 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2564 SDValue N = getValue(I.getOperand(0));
2565 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2567 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurSDLoc(), DestVT, N));
2570 void SelectionDAGBuilder::visitFPTrunc(const User &I) {
2571 // FPTrunc is never a no-op cast, no need to check
2572 SDValue N = getValue(I.getOperand(0));
2573 SDLoc dl = getCurSDLoc();
2574 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2575 EVT DestVT = TLI.getValueType(DAG.getDataLayout(), I.getType());
2576 setValue(&I, DAG.getNode(ISD::FP_ROUND, dl, DestVT, N,
2577 DAG.getTargetConstant(
2578 0, dl, TLI.getPointerTy(DAG.getDataLayout()))));
2581 void SelectionDAGBuilder::visitFPExt(const User &I) {
2582 // FPExt is never a no-op cast, no need to check
2583 SDValue N = getValue(I.getOperand(0));
2584 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2586 setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurSDLoc(), DestVT, N));
2589 void SelectionDAGBuilder::visitFPToUI(const User &I) {
2590 // FPToUI is never a no-op cast, no need to check
2591 SDValue N = getValue(I.getOperand(0));
2592 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2594 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurSDLoc(), DestVT, N));
2597 void SelectionDAGBuilder::visitFPToSI(const User &I) {
2598 // FPToSI is never a no-op cast, no need to check
2599 SDValue N = getValue(I.getOperand(0));
2600 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2602 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurSDLoc(), DestVT, N));
2605 void SelectionDAGBuilder::visitUIToFP(const User &I) {
2606 // UIToFP is never a no-op cast, no need to check
2607 SDValue N = getValue(I.getOperand(0));
2608 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2610 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurSDLoc(), DestVT, N));
2613 void SelectionDAGBuilder::visitSIToFP(const User &I) {
2614 // SIToFP is never a no-op cast, no need to check
2615 SDValue N = getValue(I.getOperand(0));
2616 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2618 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurSDLoc(), DestVT, N));
2621 void SelectionDAGBuilder::visitPtrToInt(const User &I) {
2622 // What to do depends on the size of the integer and the size of the pointer.
2623 // We can either truncate, zero extend, or no-op, accordingly.
2624 SDValue N = getValue(I.getOperand(0));
2625 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2627 setValue(&I, DAG.getZExtOrTrunc(N, getCurSDLoc(), DestVT));
2630 void SelectionDAGBuilder::visitIntToPtr(const User &I) {
2631 // What to do depends on the size of the integer and the size of the pointer.
2632 // We can either truncate, zero extend, or no-op, accordingly.
2633 SDValue N = getValue(I.getOperand(0));
2634 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2636 setValue(&I, DAG.getZExtOrTrunc(N, getCurSDLoc(), DestVT));
2639 void SelectionDAGBuilder::visitBitCast(const User &I) {
2640 SDValue N = getValue(I.getOperand(0));
2641 SDLoc dl = getCurSDLoc();
2642 EVT DestVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
2645 // BitCast assures us that source and destination are the same size so this is
2646 // either a BITCAST or a no-op.
2647 if (DestVT != N.getValueType())
2648 setValue(&I, DAG.getNode(ISD::BITCAST, dl,
2649 DestVT, N)); // convert types.
2650 // Check if the original LLVM IR Operand was a ConstantInt, because getValue()
2651 // might fold any kind of constant expression to an integer constant and that
2652 // is not what we are looking for. Only regcognize a bitcast of a genuine
2653 // constant integer as an opaque constant.
2654 else if(ConstantInt *C = dyn_cast<ConstantInt>(I.getOperand(0)))
2655 setValue(&I, DAG.getConstant(C->getValue(), dl, DestVT, /*isTarget=*/false,
2658 setValue(&I, N); // noop cast.
2661 void SelectionDAGBuilder::visitAddrSpaceCast(const User &I) {
2662 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2663 const Value *SV = I.getOperand(0);
2664 SDValue N = getValue(SV);
2665 EVT DestVT = TLI.getValueType(DAG.getDataLayout(), I.getType());
2667 unsigned SrcAS = SV->getType()->getPointerAddressSpace();
2668 unsigned DestAS = I.getType()->getPointerAddressSpace();
2670 if (!TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
2671 N = DAG.getAddrSpaceCast(getCurSDLoc(), DestVT, N, SrcAS, DestAS);
2676 void SelectionDAGBuilder::visitInsertElement(const User &I) {
2677 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2678 SDValue InVec = getValue(I.getOperand(0));
2679 SDValue InVal = getValue(I.getOperand(1));
2680 SDValue InIdx = DAG.getSExtOrTrunc(getValue(I.getOperand(2)), getCurSDLoc(),
2681 TLI.getVectorIdxTy(DAG.getDataLayout()));
2682 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurSDLoc(),
2683 TLI.getValueType(DAG.getDataLayout(), I.getType()),
2684 InVec, InVal, InIdx));
2687 void SelectionDAGBuilder::visitExtractElement(const User &I) {
2688 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2689 SDValue InVec = getValue(I.getOperand(0));
2690 SDValue InIdx = DAG.getSExtOrTrunc(getValue(I.getOperand(1)), getCurSDLoc(),
2691 TLI.getVectorIdxTy(DAG.getDataLayout()));
2692 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurSDLoc(),
2693 TLI.getValueType(DAG.getDataLayout(), I.getType()),
2697 // Utility for visitShuffleVector - Return true if every element in Mask,
2698 // beginning from position Pos and ending in Pos+Size, falls within the
2699 // specified sequential range [L, L+Pos). or is undef.
2700 static bool isSequentialInRange(const SmallVectorImpl<int> &Mask,
2701 unsigned Pos, unsigned Size, int Low) {
2702 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
2703 if (Mask[i] >= 0 && Mask[i] != Low)
2708 void SelectionDAGBuilder::visitShuffleVector(const User &I) {
2709 SDValue Src1 = getValue(I.getOperand(0));
2710 SDValue Src2 = getValue(I.getOperand(1));
2712 SmallVector<int, 8> Mask;
2713 ShuffleVectorInst::getShuffleMask(cast<Constant>(I.getOperand(2)), Mask);
2714 unsigned MaskNumElts = Mask.size();
2716 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2717 EVT VT = TLI.getValueType(DAG.getDataLayout(), I.getType());
2718 EVT SrcVT = Src1.getValueType();
2719 unsigned SrcNumElts = SrcVT.getVectorNumElements();
2721 if (SrcNumElts == MaskNumElts) {
2722 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
2727 // Normalize the shuffle vector since mask and vector length don't match.
2728 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) {
2729 // Mask is longer than the source vectors and is a multiple of the source
2730 // vectors. We can use concatenate vector to make the mask and vectors
2732 if (SrcNumElts*2 == MaskNumElts) {
2733 // First check for Src1 in low and Src2 in high
2734 if (isSequentialInRange(Mask, 0, SrcNumElts, 0) &&
2735 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, SrcNumElts)) {
2736 // The shuffle is concatenating two vectors together.
2737 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurSDLoc(),
2741 // Then check for Src2 in low and Src1 in high
2742 if (isSequentialInRange(Mask, 0, SrcNumElts, SrcNumElts) &&
2743 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, 0)) {
2744 // The shuffle is concatenating two vectors together.
2745 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurSDLoc(),
2751 // Pad both vectors with undefs to make them the same length as the mask.
2752 unsigned NumConcat = MaskNumElts / SrcNumElts;
2753 bool Src1U = Src1.getOpcode() == ISD::UNDEF;
2754 bool Src2U = Src2.getOpcode() == ISD::UNDEF;
2755 SDValue UndefVal = DAG.getUNDEF(SrcVT);
2757 SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal);
2758 SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal);
2762 Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2763 getCurSDLoc(), VT, MOps1);
2764 Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2765 getCurSDLoc(), VT, MOps2);
2767 // Readjust mask for new input vector length.
2768 SmallVector<int, 8> MappedOps;
2769 for (unsigned i = 0; i != MaskNumElts; ++i) {
2771 if (Idx >= (int)SrcNumElts)
2772 Idx -= SrcNumElts - MaskNumElts;
2773 MappedOps.push_back(Idx);
2776 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
2781 if (SrcNumElts > MaskNumElts) {
2782 // Analyze the access pattern of the vector to see if we can extract
2783 // two subvectors and do the shuffle. The analysis is done by calculating
2784 // the range of elements the mask access on both vectors.
2785 int MinRange[2] = { static_cast<int>(SrcNumElts),
2786 static_cast<int>(SrcNumElts)};
2787 int MaxRange[2] = {-1, -1};
2789 for (unsigned i = 0; i != MaskNumElts; ++i) {
2795 if (Idx >= (int)SrcNumElts) {
2799 if (Idx > MaxRange[Input])
2800 MaxRange[Input] = Idx;
2801 if (Idx < MinRange[Input])
2802 MinRange[Input] = Idx;
2805 // Check if the access is smaller than the vector size and can we find
2806 // a reasonable extract index.
2807 int RangeUse[2] = { -1, -1 }; // 0 = Unused, 1 = Extract, -1 = Can not
2809 int StartIdx[2]; // StartIdx to extract from
2810 for (unsigned Input = 0; Input < 2; ++Input) {
2811 if (MinRange[Input] >= (int)SrcNumElts && MaxRange[Input] < 0) {
2812 RangeUse[Input] = 0; // Unused
2813 StartIdx[Input] = 0;
2817 // Find a good start index that is a multiple of the mask length. Then
2818 // see if the rest of the elements are in range.
2819 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts;
2820 if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts &&
2821 StartIdx[Input] + MaskNumElts <= SrcNumElts)
2822 RangeUse[Input] = 1; // Extract from a multiple of the mask length.
2825 if (RangeUse[0] == 0 && RangeUse[1] == 0) {
2826 setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used.
2829 if (RangeUse[0] >= 0 && RangeUse[1] >= 0) {
2830 // Extract appropriate subvector and generate a vector shuffle
2831 for (unsigned Input = 0; Input < 2; ++Input) {
2832 SDValue &Src = Input == 0 ? Src1 : Src2;
2833 if (RangeUse[Input] == 0)
2834 Src = DAG.getUNDEF(VT);
2836 SDLoc dl = getCurSDLoc();
2838 ISD::EXTRACT_SUBVECTOR, dl, VT, Src,
2839 DAG.getConstant(StartIdx[Input], dl,
2840 TLI.getVectorIdxTy(DAG.getDataLayout())));
2844 // Calculate new mask.
2845 SmallVector<int, 8> MappedOps;
2846 for (unsigned i = 0; i != MaskNumElts; ++i) {
2849 if (Idx < (int)SrcNumElts)
2852 Idx -= SrcNumElts + StartIdx[1] - MaskNumElts;
2854 MappedOps.push_back(Idx);
2857 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
2863 // We can't use either concat vectors or extract subvectors so fall back to
2864 // replacing the shuffle with extract and build vector.
2865 // to insert and build vector.
2866 EVT EltVT = VT.getVectorElementType();
2867 EVT IdxVT = TLI.getVectorIdxTy(DAG.getDataLayout());
2868 SDLoc dl = getCurSDLoc();
2869 SmallVector<SDValue,8> Ops;
2870 for (unsigned i = 0; i != MaskNumElts; ++i) {
2875 Res = DAG.getUNDEF(EltVT);
2877 SDValue &Src = Idx < (int)SrcNumElts ? Src1 : Src2;
2878 if (Idx >= (int)SrcNumElts) Idx -= SrcNumElts;
2880 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
2881 EltVT, Src, DAG.getConstant(Idx, dl, IdxVT));
2887 setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops));
2890 void SelectionDAGBuilder::visitInsertValue(const InsertValueInst &I) {
2891 const Value *Op0 = I.getOperand(0);
2892 const Value *Op1 = I.getOperand(1);
2893 Type *AggTy = I.getType();
2894 Type *ValTy = Op1->getType();
2895 bool IntoUndef = isa<UndefValue>(Op0);
2896 bool FromUndef = isa<UndefValue>(Op1);
2898 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
2900 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2901 SmallVector<EVT, 4> AggValueVTs;
2902 ComputeValueVTs(TLI, DAG.getDataLayout(), AggTy, AggValueVTs);
2903 SmallVector<EVT, 4> ValValueVTs;
2904 ComputeValueVTs(TLI, DAG.getDataLayout(), ValTy, ValValueVTs);
2906 unsigned NumAggValues = AggValueVTs.size();
2907 unsigned NumValValues = ValValueVTs.size();
2908 SmallVector<SDValue, 4> Values(NumAggValues);
2910 // Ignore an insertvalue that produces an empty object
2911 if (!NumAggValues) {
2912 setValue(&I, DAG.getUNDEF(MVT(MVT::Other)));
2916 SDValue Agg = getValue(Op0);
2918 // Copy the beginning value(s) from the original aggregate.
2919 for (; i != LinearIndex; ++i)
2920 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2921 SDValue(Agg.getNode(), Agg.getResNo() + i);
2922 // Copy values from the inserted value(s).
2924 SDValue Val = getValue(Op1);
2925 for (; i != LinearIndex + NumValValues; ++i)
2926 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2927 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex);
2929 // Copy remaining value(s) from the original aggregate.
2930 for (; i != NumAggValues; ++i)
2931 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2932 SDValue(Agg.getNode(), Agg.getResNo() + i);
2934 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
2935 DAG.getVTList(AggValueVTs), Values));
2938 void SelectionDAGBuilder::visitExtractValue(const ExtractValueInst &I) {
2939 const Value *Op0 = I.getOperand(0);
2940 Type *AggTy = Op0->getType();
2941 Type *ValTy = I.getType();
2942 bool OutOfUndef = isa<UndefValue>(Op0);
2944 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
2946 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2947 SmallVector<EVT, 4> ValValueVTs;
2948 ComputeValueVTs(TLI, DAG.getDataLayout(), ValTy, ValValueVTs);
2950 unsigned NumValValues = ValValueVTs.size();
2952 // Ignore a extractvalue that produces an empty object
2953 if (!NumValValues) {
2954 setValue(&I, DAG.getUNDEF(MVT(MVT::Other)));
2958 SmallVector<SDValue, 4> Values(NumValValues);
2960 SDValue Agg = getValue(Op0);
2961 // Copy out the selected value(s).
2962 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i)
2963 Values[i - LinearIndex] =
2965 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) :
2966 SDValue(Agg.getNode(), Agg.getResNo() + i);
2968 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
2969 DAG.getVTList(ValValueVTs), Values));
2972 void SelectionDAGBuilder::visitGetElementPtr(const User &I) {
2973 Value *Op0 = I.getOperand(0);
2974 // Note that the pointer operand may be a vector of pointers. Take the scalar
2975 // element which holds a pointer.
2976 Type *Ty = Op0->getType()->getScalarType();
2977 unsigned AS = Ty->getPointerAddressSpace();
2978 SDValue N = getValue(Op0);
2979 SDLoc dl = getCurSDLoc();
2981 // Normalize Vector GEP - all scalar operands should be converted to the
2983 unsigned VectorWidth = I.getType()->isVectorTy() ?
2984 cast<VectorType>(I.getType())->getVectorNumElements() : 0;
2986 if (VectorWidth && !N.getValueType().isVector()) {
2987 MVT VT = MVT::getVectorVT(N.getValueType().getSimpleVT(), VectorWidth);
2988 SmallVector<SDValue, 16> Ops(VectorWidth, N);
2989 N = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
2991 for (GetElementPtrInst::const_op_iterator OI = I.op_begin()+1, E = I.op_end();
2993 const Value *Idx = *OI;
2994 if (StructType *StTy = dyn_cast<StructType>(Ty)) {
2995 unsigned Field = cast<Constant>(Idx)->getUniqueInteger().getZExtValue();
2998 uint64_t Offset = DL->getStructLayout(StTy)->getElementOffset(Field);
2999 N = DAG.getNode(ISD::ADD, dl, N.getValueType(), N,
3000 DAG.getConstant(Offset, dl, N.getValueType()));
3003 Ty = StTy->getElementType(Field);
3005 Ty = cast<SequentialType>(Ty)->getElementType();
3007 DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout(), AS);
3008 unsigned PtrSize = PtrTy.getSizeInBits();
3009 APInt ElementSize(PtrSize, DL->getTypeAllocSize(Ty));
3011 // If this is a scalar constant or a splat vector of constants,
3012 // handle it quickly.
3013 const auto *CI = dyn_cast<ConstantInt>(Idx);
3014 if (!CI && isa<ConstantDataVector>(Idx) &&
3015 cast<ConstantDataVector>(Idx)->getSplatValue())
3016 CI = cast<ConstantInt>(cast<ConstantDataVector>(Idx)->getSplatValue());
3021 APInt Offs = ElementSize * CI->getValue().sextOrTrunc(PtrSize);
3022 SDValue OffsVal = VectorWidth ?
3023 DAG.getConstant(Offs, dl, MVT::getVectorVT(PtrTy, VectorWidth)) :
3024 DAG.getConstant(Offs, dl, PtrTy);
3025 N = DAG.getNode(ISD::ADD, dl, N.getValueType(), N, OffsVal);
3029 // N = N + Idx * ElementSize;
3030 SDValue IdxN = getValue(Idx);
3032 if (!IdxN.getValueType().isVector() && VectorWidth) {
3033 MVT VT = MVT::getVectorVT(IdxN.getValueType().getSimpleVT(), VectorWidth);
3034 SmallVector<SDValue, 16> Ops(VectorWidth, IdxN);
3035 IdxN = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
3037 // If the index is smaller or larger than intptr_t, truncate or extend
3039 IdxN = DAG.getSExtOrTrunc(IdxN, dl, N.getValueType());
3041 // If this is a multiply by a power of two, turn it into a shl
3042 // immediately. This is a very common case.
3043 if (ElementSize != 1) {
3044 if (ElementSize.isPowerOf2()) {
3045 unsigned Amt = ElementSize.logBase2();
3046 IdxN = DAG.getNode(ISD::SHL, dl,
3047 N.getValueType(), IdxN,
3048 DAG.getConstant(Amt, dl, IdxN.getValueType()));
3050 SDValue Scale = DAG.getConstant(ElementSize, dl, IdxN.getValueType());
3051 IdxN = DAG.getNode(ISD::MUL, dl,
3052 N.getValueType(), IdxN, Scale);
3056 N = DAG.getNode(ISD::ADD, dl,
3057 N.getValueType(), N, IdxN);
3064 void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) {
3065 // If this is a fixed sized alloca in the entry block of the function,
3066 // allocate it statically on the stack.
3067 if (FuncInfo.StaticAllocaMap.count(&I))
3068 return; // getValue will auto-populate this.
3070 SDLoc dl = getCurSDLoc();
3071 Type *Ty = I.getAllocatedType();
3072 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3073 auto &DL = DAG.getDataLayout();
3074 uint64_t TySize = DL.getTypeAllocSize(Ty);
3076 std::max((unsigned)DL.getPrefTypeAlignment(Ty), I.getAlignment());
3078 SDValue AllocSize = getValue(I.getArraySize());
3080 EVT IntPtr = TLI.getPointerTy(DAG.getDataLayout());
3081 if (AllocSize.getValueType() != IntPtr)
3082 AllocSize = DAG.getZExtOrTrunc(AllocSize, dl, IntPtr);
3084 AllocSize = DAG.getNode(ISD::MUL, dl, IntPtr,
3086 DAG.getConstant(TySize, dl, IntPtr));
3088 // Handle alignment. If the requested alignment is less than or equal to
3089 // the stack alignment, ignore it. If the size is greater than or equal to
3090 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
3091 unsigned StackAlign =
3092 DAG.getSubtarget().getFrameLowering()->getStackAlignment();
3093 if (Align <= StackAlign)
3096 // Round the size of the allocation up to the stack alignment size
3097 // by add SA-1 to the size.
3098 AllocSize = DAG.getNode(ISD::ADD, dl,
3099 AllocSize.getValueType(), AllocSize,
3100 DAG.getIntPtrConstant(StackAlign - 1, dl));
3102 // Mask out the low bits for alignment purposes.
3103 AllocSize = DAG.getNode(ISD::AND, dl,
3104 AllocSize.getValueType(), AllocSize,
3105 DAG.getIntPtrConstant(~(uint64_t)(StackAlign - 1),
3108 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align, dl) };
3109 SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other);
3110 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, dl, VTs, Ops);
3112 DAG.setRoot(DSA.getValue(1));
3114 assert(FuncInfo.MF->getFrameInfo()->hasVarSizedObjects());
3117 void SelectionDAGBuilder::visitLoad(const LoadInst &I) {
3119 return visitAtomicLoad(I);
3121 const Value *SV = I.getOperand(0);
3122 SDValue Ptr = getValue(SV);
3124 Type *Ty = I.getType();
3126 bool isVolatile = I.isVolatile();
3127 bool isNonTemporal = I.getMetadata(LLVMContext::MD_nontemporal) != nullptr;
3129 // The IR notion of invariant_load only guarantees that all *non-faulting*
3130 // invariant loads result in the same value. The MI notion of invariant load
3131 // guarantees that the load can be legally moved to any location within its
3132 // containing function. The MI notion of invariant_load is stronger than the
3133 // IR notion of invariant_load -- an MI invariant_load is an IR invariant_load
3134 // with a guarantee that the location being loaded from is dereferenceable
3135 // throughout the function's lifetime.
3137 bool isInvariant = I.getMetadata(LLVMContext::MD_invariant_load) != nullptr &&
3138 isDereferenceablePointer(SV, DAG.getDataLayout());
3139 unsigned Alignment = I.getAlignment();
3142 I.getAAMetadata(AAInfo);
3143 const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range);
3145 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3146 SmallVector<EVT, 4> ValueVTs;
3147 SmallVector<uint64_t, 4> Offsets;
3148 ComputeValueVTs(TLI, DAG.getDataLayout(), Ty, ValueVTs, &Offsets);
3149 unsigned NumValues = ValueVTs.size();
3154 bool ConstantMemory = false;
3155 if (isVolatile || NumValues > MaxParallelChains)
3156 // Serialize volatile loads with other side effects.
3158 else if (AA->pointsToConstantMemory(MemoryLocation(
3159 SV, DAG.getDataLayout().getTypeStoreSize(Ty), AAInfo))) {
3160 // Do not serialize (non-volatile) loads of constant memory with anything.
3161 Root = DAG.getEntryNode();
3162 ConstantMemory = true;
3164 // Do not serialize non-volatile loads against each other.
3165 Root = DAG.getRoot();
3168 SDLoc dl = getCurSDLoc();
3171 Root = TLI.prepareVolatileOrAtomicLoad(Root, dl, DAG);
3173 SmallVector<SDValue, 4> Values(NumValues);
3174 SmallVector<SDValue, 4> Chains(std::min(MaxParallelChains, NumValues));
3175 EVT PtrVT = Ptr.getValueType();
3176 unsigned ChainI = 0;
3177 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3178 // Serializing loads here may result in excessive register pressure, and
3179 // TokenFactor places arbitrary choke points on the scheduler. SD scheduling
3180 // could recover a bit by hoisting nodes upward in the chain by recognizing
3181 // they are side-effect free or do not alias. The optimizer should really
3182 // avoid this case by converting large object/array copies to llvm.memcpy
3183 // (MaxParallelChains should always remain as failsafe).
3184 if (ChainI == MaxParallelChains) {
3185 assert(PendingLoads.empty() && "PendingLoads must be serialized first");
3186 SDValue Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3187 makeArrayRef(Chains.data(), ChainI));
3191 SDValue A = DAG.getNode(ISD::ADD, dl,
3193 DAG.getConstant(Offsets[i], dl, PtrVT));
3194 SDValue L = DAG.getLoad(ValueVTs[i], dl, Root,
3195 A, MachinePointerInfo(SV, Offsets[i]), isVolatile,
3196 isNonTemporal, isInvariant, Alignment, AAInfo,
3200 Chains[ChainI] = L.getValue(1);
3203 if (!ConstantMemory) {
3204 SDValue Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3205 makeArrayRef(Chains.data(), ChainI));
3209 PendingLoads.push_back(Chain);
3212 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, dl,
3213 DAG.getVTList(ValueVTs), Values));
3216 void SelectionDAGBuilder::visitStore(const StoreInst &I) {
3218 return visitAtomicStore(I);
3220 const Value *SrcV = I.getOperand(0);
3221 const Value *PtrV = I.getOperand(1);
3223 SmallVector<EVT, 4> ValueVTs;
3224 SmallVector<uint64_t, 4> Offsets;
3225 ComputeValueVTs(DAG.getTargetLoweringInfo(), DAG.getDataLayout(),
3226 SrcV->getType(), ValueVTs, &Offsets);
3227 unsigned NumValues = ValueVTs.size();
3231 // Get the lowered operands. Note that we do this after
3232 // checking if NumResults is zero, because with zero results
3233 // the operands won't have values in the map.
3234 SDValue Src = getValue(SrcV);
3235 SDValue Ptr = getValue(PtrV);
3237 SDValue Root = getRoot();
3238 SmallVector<SDValue, 4> Chains(std::min(MaxParallelChains, NumValues));
3239 EVT PtrVT = Ptr.getValueType();
3240 bool isVolatile = I.isVolatile();
3241 bool isNonTemporal = I.getMetadata(LLVMContext::MD_nontemporal) != nullptr;
3242 unsigned Alignment = I.getAlignment();
3243 SDLoc dl = getCurSDLoc();
3246 I.getAAMetadata(AAInfo);
3248 unsigned ChainI = 0;
3249 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3250 // See visitLoad comments.
3251 if (ChainI == MaxParallelChains) {
3252 SDValue Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3253 makeArrayRef(Chains.data(), ChainI));
3257 SDValue Add = DAG.getNode(ISD::ADD, dl, PtrVT, Ptr,
3258 DAG.getConstant(Offsets[i], dl, PtrVT));
3259 SDValue St = DAG.getStore(Root, dl,
3260 SDValue(Src.getNode(), Src.getResNo() + i),
3261 Add, MachinePointerInfo(PtrV, Offsets[i]),
3262 isVolatile, isNonTemporal, Alignment, AAInfo);
3263 Chains[ChainI] = St;
3266 SDValue StoreNode = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3267 makeArrayRef(Chains.data(), ChainI));
3268 DAG.setRoot(StoreNode);
3271 void SelectionDAGBuilder::visitMaskedStore(const CallInst &I) {
3272 SDLoc sdl = getCurSDLoc();
3274 // llvm.masked.store.*(Src0, Ptr, alignment, Mask)
3275 Value *PtrOperand = I.getArgOperand(1);
3276 SDValue Ptr = getValue(PtrOperand);
3277 SDValue Src0 = getValue(I.getArgOperand(0));
3278 SDValue Mask = getValue(I.getArgOperand(3));
3279 EVT VT = Src0.getValueType();
3280 unsigned Alignment = (cast<ConstantInt>(I.getArgOperand(2)))->getZExtValue();
3282 Alignment = DAG.getEVTAlignment(VT);
3285 I.getAAMetadata(AAInfo);
3287 MachineMemOperand *MMO =
3288 DAG.getMachineFunction().
3289 getMachineMemOperand(MachinePointerInfo(PtrOperand),
3290 MachineMemOperand::MOStore, VT.getStoreSize(),
3292 SDValue StoreNode = DAG.getMaskedStore(getRoot(), sdl, Src0, Ptr, Mask, VT,
3294 DAG.setRoot(StoreNode);
3295 setValue(&I, StoreNode);
3298 // Get a uniform base for the Gather/Scatter intrinsic.
3299 // The first argument of the Gather/Scatter intrinsic is a vector of pointers.
3300 // We try to represent it as a base pointer + vector of indices.
3301 // Usually, the vector of pointers comes from a 'getelementptr' instruction.
3302 // The first operand of the GEP may be a single pointer or a vector of pointers
3304 // %gep.ptr = getelementptr i32, <8 x i32*> %vptr, <8 x i32> %ind
3306 // %gep.ptr = getelementptr i32, i32* %ptr, <8 x i32> %ind
3307 // %res = call <8 x i32> @llvm.masked.gather.v8i32(<8 x i32*> %gep.ptr, ..
3309 // When the first GEP operand is a single pointer - it is the uniform base we
3310 // are looking for. If first operand of the GEP is a splat vector - we
3311 // extract the spalt value and use it as a uniform base.
3312 // In all other cases the function returns 'false'.
3314 static bool getUniformBase(const Value *& Ptr, SDValue& Base, SDValue& Index,
3315 SelectionDAGBuilder* SDB) {
3317 SelectionDAG& DAG = SDB->DAG;
3318 LLVMContext &Context = *DAG.getContext();
3320 assert(Ptr->getType()->isVectorTy() && "Uexpected pointer type");
3321 const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3322 if (!GEP || GEP->getNumOperands() > 2)
3325 const Value *GEPPtr = GEP->getPointerOperand();
3326 if (!GEPPtr->getType()->isVectorTy())
3328 else if (!(Ptr = getSplatValue(GEPPtr)))
3331 Value *IndexVal = GEP->getOperand(1);
3333 // The operands of the GEP may be defined in another basic block.
3334 // In this case we'll not find nodes for the operands.
3335 if (!SDB->findValue(Ptr) || !SDB->findValue(IndexVal))
3338 Base = SDB->getValue(Ptr);
3339 Index = SDB->getValue(IndexVal);
3341 // Suppress sign extension.
3342 if (SExtInst* Sext = dyn_cast<SExtInst>(IndexVal)) {
3343 if (SDB->findValue(Sext->getOperand(0))) {
3344 IndexVal = Sext->getOperand(0);
3345 Index = SDB->getValue(IndexVal);
3348 if (!Index.getValueType().isVector()) {
3349 unsigned GEPWidth = GEP->getType()->getVectorNumElements();
3350 EVT VT = EVT::getVectorVT(Context, Index.getValueType(), GEPWidth);
3351 SmallVector<SDValue, 16> Ops(GEPWidth, Index);
3352 Index = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Index), VT, Ops);
3357 void SelectionDAGBuilder::visitMaskedScatter(const CallInst &I) {
3358 SDLoc sdl = getCurSDLoc();
3360 // llvm.masked.scatter.*(Src0, Ptrs, alignemt, Mask)
3361 const Value *Ptr = I.getArgOperand(1);
3362 SDValue Src0 = getValue(I.getArgOperand(0));
3363 SDValue Mask = getValue(I.getArgOperand(3));
3364 EVT VT = Src0.getValueType();
3365 unsigned Alignment = (cast<ConstantInt>(I.getArgOperand(2)))->getZExtValue();
3367 Alignment = DAG.getEVTAlignment(VT);
3368 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3371 I.getAAMetadata(AAInfo);
3375 const Value *BasePtr = Ptr;
3376 bool UniformBase = getUniformBase(BasePtr, Base, Index, this);
3378 const Value *MemOpBasePtr = UniformBase ? BasePtr : nullptr;
3379 MachineMemOperand *MMO = DAG.getMachineFunction().
3380 getMachineMemOperand(MachinePointerInfo(MemOpBasePtr),
3381 MachineMemOperand::MOStore, VT.getStoreSize(),
3384 Base = DAG.getTargetConstant(0, sdl, TLI.getPointerTy(DAG.getDataLayout()));
3385 Index = getValue(Ptr);
3387 SDValue Ops[] = { getRoot(), Src0, Mask, Base, Index };
3388 SDValue Scatter = DAG.getMaskedScatter(DAG.getVTList(MVT::Other), VT, sdl,
3390 DAG.setRoot(Scatter);
3391 setValue(&I, Scatter);
3394 void SelectionDAGBuilder::visitMaskedLoad(const CallInst &I) {
3395 SDLoc sdl = getCurSDLoc();
3397 // @llvm.masked.load.*(Ptr, alignment, Mask, Src0)
3398 Value *PtrOperand = I.getArgOperand(0);
3399 SDValue Ptr = getValue(PtrOperand);
3400 SDValue Src0 = getValue(I.getArgOperand(3));
3401 SDValue Mask = getValue(I.getArgOperand(2));
3403 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3404 EVT VT = TLI.getValueType(DAG.getDataLayout(), I.getType());
3405 unsigned Alignment = (cast<ConstantInt>(I.getArgOperand(1)))->getZExtValue();
3407 Alignment = DAG.getEVTAlignment(VT);
3410 I.getAAMetadata(AAInfo);
3411 const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range);
3413 SDValue InChain = DAG.getRoot();
3414 if (AA->pointsToConstantMemory(MemoryLocation(
3415 PtrOperand, DAG.getDataLayout().getTypeStoreSize(I.getType()),
3417 // Do not serialize (non-volatile) loads of constant memory with anything.
3418 InChain = DAG.getEntryNode();
3421 MachineMemOperand *MMO =
3422 DAG.getMachineFunction().
3423 getMachineMemOperand(MachinePointerInfo(PtrOperand),
3424 MachineMemOperand::MOLoad, VT.getStoreSize(),
3425 Alignment, AAInfo, Ranges);
3427 SDValue Load = DAG.getMaskedLoad(VT, sdl, InChain, Ptr, Mask, Src0, VT, MMO,
3429 SDValue OutChain = Load.getValue(1);
3430 DAG.setRoot(OutChain);
3434 void SelectionDAGBuilder::visitMaskedGather(const CallInst &I) {
3435 SDLoc sdl = getCurSDLoc();
3437 // @llvm.masked.gather.*(Ptrs, alignment, Mask, Src0)
3438 const Value *Ptr = I.getArgOperand(0);
3439 SDValue Src0 = getValue(I.getArgOperand(3));
3440 SDValue Mask = getValue(I.getArgOperand(2));
3442 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3443 EVT VT = TLI.getValueType(DAG.getDataLayout(), I.getType());
3444 unsigned Alignment = (cast<ConstantInt>(I.getArgOperand(1)))->getZExtValue();
3446 Alignment = DAG.getEVTAlignment(VT);
3449 I.getAAMetadata(AAInfo);
3450 const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range);
3452 SDValue Root = DAG.getRoot();
3455 const Value *BasePtr = Ptr;
3456 bool UniformBase = getUniformBase(BasePtr, Base, Index, this);
3457 bool ConstantMemory = false;
3459 AA->pointsToConstantMemory(MemoryLocation(
3460 BasePtr, DAG.getDataLayout().getTypeStoreSize(I.getType()),
3462 // Do not serialize (non-volatile) loads of constant memory with anything.
3463 Root = DAG.getEntryNode();
3464 ConstantMemory = true;
3467 MachineMemOperand *MMO =
3468 DAG.getMachineFunction().
3469 getMachineMemOperand(MachinePointerInfo(UniformBase ? BasePtr : nullptr),
3470 MachineMemOperand::MOLoad, VT.getStoreSize(),
3471 Alignment, AAInfo, Ranges);
3474 Base = DAG.getTargetConstant(0, sdl, TLI.getPointerTy(DAG.getDataLayout()));
3475 Index = getValue(Ptr);
3477 SDValue Ops[] = { Root, Src0, Mask, Base, Index };
3478 SDValue Gather = DAG.getMaskedGather(DAG.getVTList(VT, MVT::Other), VT, sdl,
3481 SDValue OutChain = Gather.getValue(1);
3482 if (!ConstantMemory)
3483 PendingLoads.push_back(OutChain);
3484 setValue(&I, Gather);
3487 void SelectionDAGBuilder::visitAtomicCmpXchg(const AtomicCmpXchgInst &I) {
3488 SDLoc dl = getCurSDLoc();
3489 AtomicOrdering SuccessOrder = I.getSuccessOrdering();
3490 AtomicOrdering FailureOrder = I.getFailureOrdering();
3491 SynchronizationScope Scope = I.getSynchScope();
3493 SDValue InChain = getRoot();
3495 MVT MemVT = getValue(I.getCompareOperand()).getSimpleValueType();
3496 SDVTList VTs = DAG.getVTList(MemVT, MVT::i1, MVT::Other);
3497 SDValue L = DAG.getAtomicCmpSwap(
3498 ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, dl, MemVT, VTs, InChain,
3499 getValue(I.getPointerOperand()), getValue(I.getCompareOperand()),
3500 getValue(I.getNewValOperand()), MachinePointerInfo(I.getPointerOperand()),
3501 /*Alignment=*/ 0, SuccessOrder, FailureOrder, Scope);
3503 SDValue OutChain = L.getValue(2);
3506 DAG.setRoot(OutChain);
3509 void SelectionDAGBuilder::visitAtomicRMW(const AtomicRMWInst &I) {
3510 SDLoc dl = getCurSDLoc();
3512 switch (I.getOperation()) {
3513 default: llvm_unreachable("Unknown atomicrmw operation");
3514 case AtomicRMWInst::Xchg: NT = ISD::ATOMIC_SWAP; break;
3515 case AtomicRMWInst::Add: NT = ISD::ATOMIC_LOAD_ADD; break;
3516 case AtomicRMWInst::Sub: NT = ISD::ATOMIC_LOAD_SUB; break;
3517 case AtomicRMWInst::And: NT = ISD::ATOMIC_LOAD_AND; break;
3518 case AtomicRMWInst::Nand: NT = ISD::ATOMIC_LOAD_NAND; break;
3519 case AtomicRMWInst::Or: NT = ISD::ATOMIC_LOAD_OR; break;
3520 case AtomicRMWInst::Xor: NT = ISD::ATOMIC_LOAD_XOR; break;
3521 case AtomicRMWInst::Max: NT = ISD::ATOMIC_LOAD_MAX; break;
3522 case AtomicRMWInst::Min: NT = ISD::ATOMIC_LOAD_MIN; break;
3523 case AtomicRMWInst::UMax: NT = ISD::ATOMIC_LOAD_UMAX; break;
3524 case AtomicRMWInst::UMin: NT = ISD::ATOMIC_LOAD_UMIN; break;
3526 AtomicOrdering Order = I.getOrdering();
3527 SynchronizationScope Scope = I.getSynchScope();
3529 SDValue InChain = getRoot();
3532 DAG.getAtomic(NT, dl,
3533 getValue(I.getValOperand()).getSimpleValueType(),
3535 getValue(I.getPointerOperand()),
3536 getValue(I.getValOperand()),
3537 I.getPointerOperand(),
3538 /* Alignment=*/ 0, Order, Scope);
3540 SDValue OutChain = L.getValue(1);
3543 DAG.setRoot(OutChain);
3546 void SelectionDAGBuilder::visitFence(const FenceInst &I) {
3547 SDLoc dl = getCurSDLoc();
3548 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3551 Ops[1] = DAG.getConstant(I.getOrdering(), dl,
3552 TLI.getPointerTy(DAG.getDataLayout()));
3553 Ops[2] = DAG.getConstant(I.getSynchScope(), dl,
3554 TLI.getPointerTy(DAG.getDataLayout()));
3555 DAG.setRoot(DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops));
3558 void SelectionDAGBuilder::visitAtomicLoad(const LoadInst &I) {
3559 SDLoc dl = getCurSDLoc();
3560 AtomicOrdering Order = I.getOrdering();
3561 SynchronizationScope Scope = I.getSynchScope();
3563 SDValue InChain = getRoot();
3565 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3566 EVT VT = TLI.getValueType(DAG.getDataLayout(), I.getType());
3568 if (I.getAlignment() < VT.getSizeInBits() / 8)
3569 report_fatal_error("Cannot generate unaligned atomic load");
3571 MachineMemOperand *MMO =
3572 DAG.getMachineFunction().
3573 getMachineMemOperand(MachinePointerInfo(I.getPointerOperand()),
3574 MachineMemOperand::MOVolatile |
3575 MachineMemOperand::MOLoad,
3577 I.getAlignment() ? I.getAlignment() :
3578 DAG.getEVTAlignment(VT));
3580 InChain = TLI.prepareVolatileOrAtomicLoad(InChain, dl, DAG);
3582 DAG.getAtomic(ISD::ATOMIC_LOAD, dl, VT, VT, InChain,
3583 getValue(I.getPointerOperand()), MMO,
3586 SDValue OutChain = L.getValue(1);
3589 DAG.setRoot(OutChain);
3592 void SelectionDAGBuilder::visitAtomicStore(const StoreInst &I) {
3593 SDLoc dl = getCurSDLoc();
3595 AtomicOrdering Order = I.getOrdering();
3596 SynchronizationScope Scope = I.getSynchScope();
3598 SDValue InChain = getRoot();
3600 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3602 TLI.getValueType(DAG.getDataLayout(), I.getValueOperand()->getType());
3604 if (I.getAlignment() < VT.getSizeInBits() / 8)
3605 report_fatal_error("Cannot generate unaligned atomic store");
3608 DAG.getAtomic(ISD::ATOMIC_STORE, dl, VT,
3610 getValue(I.getPointerOperand()),
3611 getValue(I.getValueOperand()),
3612 I.getPointerOperand(), I.getAlignment(),
3615 DAG.setRoot(OutChain);
3618 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
3620 void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I,
3621 unsigned Intrinsic) {
3622 bool HasChain = !I.doesNotAccessMemory();
3623 bool OnlyLoad = HasChain && I.onlyReadsMemory();
3625 // Build the operand list.
3626 SmallVector<SDValue, 8> Ops;
3627 if (HasChain) { // If this intrinsic has side-effects, chainify it.
3629 // We don't need to serialize loads against other loads.
3630 Ops.push_back(DAG.getRoot());
3632 Ops.push_back(getRoot());
3636 // Info is set by getTgtMemInstrinsic
3637 TargetLowering::IntrinsicInfo Info;
3638 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3639 bool IsTgtIntrinsic = TLI.getTgtMemIntrinsic(Info, I, Intrinsic);
3641 // Add the intrinsic ID as an integer operand if it's not a target intrinsic.
3642 if (!IsTgtIntrinsic || Info.opc == ISD::INTRINSIC_VOID ||
3643 Info.opc == ISD::INTRINSIC_W_CHAIN)
3644 Ops.push_back(DAG.getTargetConstant(Intrinsic, getCurSDLoc(),
3645 TLI.getPointerTy(DAG.getDataLayout())));
3647 // Add all operands of the call to the operand list.
3648 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) {
3649 SDValue Op = getValue(I.getArgOperand(i));
3653 SmallVector<EVT, 4> ValueVTs;
3654 ComputeValueVTs(TLI, DAG.getDataLayout(), I.getType(), ValueVTs);
3657 ValueVTs.push_back(MVT::Other);
3659 SDVTList VTs = DAG.getVTList(ValueVTs);
3663 if (IsTgtIntrinsic) {
3664 // This is target intrinsic that touches memory
3665 Result = DAG.getMemIntrinsicNode(Info.opc, getCurSDLoc(),
3666 VTs, Ops, Info.memVT,
3667 MachinePointerInfo(Info.ptrVal, Info.offset),
3668 Info.align, Info.vol,
3669 Info.readMem, Info.writeMem, Info.size);
3670 } else if (!HasChain) {
3671 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurSDLoc(), VTs, Ops);
3672 } else if (!I.getType()->isVoidTy()) {
3673 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurSDLoc(), VTs, Ops);
3675 Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurSDLoc(), VTs, Ops);
3679 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1);
3681 PendingLoads.push_back(Chain);
3686 if (!I.getType()->isVoidTy()) {
3687 if (VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
3688 EVT VT = TLI.getValueType(DAG.getDataLayout(), PTy);
3689 Result = DAG.getNode(ISD::BITCAST, getCurSDLoc(), VT, Result);
3692 setValue(&I, Result);
3696 /// GetSignificand - Get the significand and build it into a floating-point
3697 /// number with exponent of 1:
3699 /// Op = (Op & 0x007fffff) | 0x3f800000;
3701 /// where Op is the hexadecimal representation of floating point value.
3703 GetSignificand(SelectionDAG &DAG, SDValue Op, SDLoc dl) {
3704 SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3705 DAG.getConstant(0x007fffff, dl, MVT::i32));
3706 SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1,
3707 DAG.getConstant(0x3f800000, dl, MVT::i32));
3708 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, t2);
3711 /// GetExponent - Get the exponent:
3713 /// (float)(int)(((Op & 0x7f800000) >> 23) - 127);
3715 /// where Op is the hexadecimal representation of floating point value.
3717 GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI,
3719 SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3720 DAG.getConstant(0x7f800000, dl, MVT::i32));
3721 SDValue t1 = DAG.getNode(
3722 ISD::SRL, dl, MVT::i32, t0,
3723 DAG.getConstant(23, dl, TLI.getPointerTy(DAG.getDataLayout())));
3724 SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1,
3725 DAG.getConstant(127, dl, MVT::i32));
3726 return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2);
3729 /// getF32Constant - Get 32-bit floating point constant.
3731 getF32Constant(SelectionDAG &DAG, unsigned Flt, SDLoc dl) {
3732 return DAG.getConstantFP(APFloat(APFloat::IEEEsingle, APInt(32, Flt)), dl,
3736 static SDValue getLimitedPrecisionExp2(SDValue t0, SDLoc dl,
3737 SelectionDAG &DAG) {
3738 // TODO: What fast-math-flags should be set on the floating-point nodes?
3740 // IntegerPartOfX = ((int32_t)(t0);
3741 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3743 // FractionalPartOfX = t0 - (float)IntegerPartOfX;
3744 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3745 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3747 // IntegerPartOfX <<= 23;
3748 IntegerPartOfX = DAG.getNode(
3749 ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3750 DAG.getConstant(23, dl, DAG.getTargetLoweringInfo().getPointerTy(
3751 DAG.getDataLayout())));
3753 SDValue TwoToFractionalPartOfX;
3754 if (LimitFloatPrecision <= 6) {
3755 // For floating-point precision of 6:
3757 // TwoToFractionalPartOfX =
3759 // (0.735607626f + 0.252464424f * x) * x;
3761 // error 0.0144103317, which is 6 bits
3762 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3763 getF32Constant(DAG, 0x3e814304, dl));
3764 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3765 getF32Constant(DAG, 0x3f3c50c8, dl));
3766 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3767 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3768 getF32Constant(DAG, 0x3f7f5e7e, dl));
3769 } else if (LimitFloatPrecision <= 12) {
3770 // For floating-point precision of 12:
3772 // TwoToFractionalPartOfX =
3775 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3777 // error 0.000107046256, which is 13 to 14 bits
3778 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3779 getF32Constant(DAG, 0x3da235e3, dl));
3780 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3781 getF32Constant(DAG, 0x3e65b8f3, dl));
3782 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3783 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3784 getF32Constant(DAG, 0x3f324b07, dl));
3785 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3786 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3787 getF32Constant(DAG, 0x3f7ff8fd, dl));
3788 } else { // LimitFloatPrecision <= 18
3789 // For floating-point precision of 18:
3791 // TwoToFractionalPartOfX =
3795 // (0.554906021e-1f +
3796 // (0.961591928e-2f +
3797 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3798 // error 2.47208000*10^(-7), which is better than 18 bits
3799 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3800 getF32Constant(DAG, 0x3924b03e, dl));
3801 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3802 getF32Constant(DAG, 0x3ab24b87, dl));
3803 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3804 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3805 getF32Constant(DAG, 0x3c1d8c17, dl));
3806 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3807 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3808 getF32Constant(DAG, 0x3d634a1d, dl));
3809 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3810 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3811 getF32Constant(DAG, 0x3e75fe14, dl));
3812 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3813 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3814 getF32Constant(DAG, 0x3f317234, dl));
3815 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3816 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3817 getF32Constant(DAG, 0x3f800000, dl));
3820 // Add the exponent into the result in integer domain.
3821 SDValue t13 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, TwoToFractionalPartOfX);
3822 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3823 DAG.getNode(ISD::ADD, dl, MVT::i32, t13, IntegerPartOfX));
3826 /// expandExp - Lower an exp intrinsic. Handles the special sequences for
3827 /// limited-precision mode.
3828 static SDValue expandExp(SDLoc dl, SDValue Op, SelectionDAG &DAG,
3829 const TargetLowering &TLI) {
3830 if (Op.getValueType() == MVT::f32 &&
3831 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3833 // Put the exponent in the right bit position for later addition to the
3836 // #define LOG2OFe 1.4426950f
3837 // t0 = Op * LOG2OFe
3839 // TODO: What fast-math-flags should be set here?
3840 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3841 getF32Constant(DAG, 0x3fb8aa3b, dl));
3842 return getLimitedPrecisionExp2(t0, dl, DAG);
3845 // No special expansion.
3846 return DAG.getNode(ISD::FEXP, dl, Op.getValueType(), Op);
3849 /// expandLog - Lower a log intrinsic. Handles the special sequences for
3850 /// limited-precision mode.
3851 static SDValue expandLog(SDLoc dl, SDValue Op, SelectionDAG &DAG,
3852 const TargetLowering &TLI) {
3854 // TODO: What fast-math-flags should be set on the floating-point nodes?
3856 if (Op.getValueType() == MVT::f32 &&
3857 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3858 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3860 // Scale the exponent by log(2) [0.69314718f].
3861 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3862 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3863 getF32Constant(DAG, 0x3f317218, dl));
3865 // Get the significand and build it into a floating-point number with
3867 SDValue X = GetSignificand(DAG, Op1, dl);
3869 SDValue LogOfMantissa;
3870 if (LimitFloatPrecision <= 6) {
3871 // For floating-point precision of 6:
3875 // (1.4034025f - 0.23903021f * x) * x;
3877 // error 0.0034276066, which is better than 8 bits
3878 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3879 getF32Constant(DAG, 0xbe74c456, dl));
3880 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3881 getF32Constant(DAG, 0x3fb3a2b1, dl));
3882 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3883 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3884 getF32Constant(DAG, 0x3f949a29, dl));
3885 } else if (LimitFloatPrecision <= 12) {
3886 // For floating-point precision of 12:
3892 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x;
3894 // error 0.000061011436, which is 14 bits
3895 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3896 getF32Constant(DAG, 0xbd67b6d6, dl));
3897 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3898 getF32Constant(DAG, 0x3ee4f4b8, dl));
3899 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3900 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3901 getF32Constant(DAG, 0x3fbc278b, dl));
3902 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3903 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3904 getF32Constant(DAG, 0x40348e95, dl));
3905 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3906 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3907 getF32Constant(DAG, 0x3fdef31a, dl));
3908 } else { // LimitFloatPrecision <= 18
3909 // For floating-point precision of 18:
3917 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x;
3919 // error 0.0000023660568, which is better than 18 bits
3920 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3921 getF32Constant(DAG, 0xbc91e5ac, dl));
3922 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3923 getF32Constant(DAG, 0x3e4350aa, dl));
3924 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3925 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3926 getF32Constant(DAG, 0x3f60d3e3, dl));
3927 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3928 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3929 getF32Constant(DAG, 0x4011cdf0, dl));
3930 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3931 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3932 getF32Constant(DAG, 0x406cfd1c, dl));
3933 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3934 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3935 getF32Constant(DAG, 0x408797cb, dl));
3936 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3937 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3938 getF32Constant(DAG, 0x4006dcab, dl));
3941 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, LogOfMantissa);
3944 // No special expansion.
3945 return DAG.getNode(ISD::FLOG, dl, Op.getValueType(), Op);
3948 /// expandLog2 - Lower a log2 intrinsic. Handles the special sequences for
3949 /// limited-precision mode.
3950 static SDValue expandLog2(SDLoc dl, SDValue Op, SelectionDAG &DAG,
3951 const TargetLowering &TLI) {
3953 // TODO: What fast-math-flags should be set on the floating-point nodes?
3955 if (Op.getValueType() == MVT::f32 &&
3956 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3957 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3959 // Get the exponent.
3960 SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl);
3962 // Get the significand and build it into a floating-point number with
3964 SDValue X = GetSignificand(DAG, Op1, dl);
3966 // Different possible minimax approximations of significand in
3967 // floating-point for various degrees of accuracy over [1,2].
3968 SDValue Log2ofMantissa;
3969 if (LimitFloatPrecision <= 6) {
3970 // For floating-point precision of 6:
3972 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x;
3974 // error 0.0049451742, which is more than 7 bits
3975 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3976 getF32Constant(DAG, 0xbeb08fe0, dl));
3977 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3978 getF32Constant(DAG, 0x40019463, dl));
3979 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3980 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3981 getF32Constant(DAG, 0x3fd6633d, dl));
3982 } else if (LimitFloatPrecision <= 12) {
3983 // For floating-point precision of 12:
3989 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x;
3991 // error 0.0000876136000, which is better than 13 bits
3992 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3993 getF32Constant(DAG, 0xbda7262e, dl));
3994 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3995 getF32Constant(DAG, 0x3f25280b, dl));
3996 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3997 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3998 getF32Constant(DAG, 0x4007b923, dl));
3999 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4000 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4001 getF32Constant(DAG, 0x40823e2f, dl));
4002 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4003 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
4004 getF32Constant(DAG, 0x4020d29c, dl));
4005 } else { // LimitFloatPrecision <= 18
4006 // For floating-point precision of 18:
4015 // 0.25691327e-1f * x) * x) * x) * x) * x) * x;
4017 // error 0.0000018516, which is better than 18 bits
4018 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4019 getF32Constant(DAG, 0xbcd2769e, dl));
4020 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4021 getF32Constant(DAG, 0x3e8ce0b9, dl));
4022 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4023 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4024 getF32Constant(DAG, 0x3fa22ae7, dl));
4025 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4026 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4027 getF32Constant(DAG, 0x40525723, dl));
4028 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4029 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
4030 getF32Constant(DAG, 0x40aaf200, dl));
4031 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4032 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4033 getF32Constant(DAG, 0x40c39dad, dl));
4034 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4035 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
4036 getF32Constant(DAG, 0x4042902c, dl));
4039 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, Log2ofMantissa);
4042 // No special expansion.
4043 return DAG.getNode(ISD::FLOG2, dl, Op.getValueType(), Op);
4046 /// expandLog10 - Lower a log10 intrinsic. Handles the special sequences for
4047 /// limited-precision mode.
4048 static SDValue expandLog10(SDLoc dl, SDValue Op, SelectionDAG &DAG,
4049 const TargetLowering &TLI) {
4051 // TODO: What fast-math-flags should be set on the floating-point nodes?
4053 if (Op.getValueType() == MVT::f32 &&
4054 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4055 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
4057 // Scale the exponent by log10(2) [0.30102999f].
4058 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
4059 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
4060 getF32Constant(DAG, 0x3e9a209a, dl));
4062 // Get the significand and build it into a floating-point number with
4064 SDValue X = GetSignificand(DAG, Op1, dl);
4066 SDValue Log10ofMantissa;
4067 if (LimitFloatPrecision <= 6) {
4068 // For floating-point precision of 6:
4070 // Log10ofMantissa =
4072 // (0.60948995f - 0.10380950f * x) * x;
4074 // error 0.0014886165, which is 6 bits
4075 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4076 getF32Constant(DAG, 0xbdd49a13, dl));
4077 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4078 getF32Constant(DAG, 0x3f1c0789, dl));
4079 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4080 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4081 getF32Constant(DAG, 0x3f011300, dl));
4082 } else if (LimitFloatPrecision <= 12) {
4083 // For floating-point precision of 12:
4085 // Log10ofMantissa =
4088 // (-0.31664806f + 0.47637168e-1f * x) * x) * x;
4090 // error 0.00019228036, which is better than 12 bits
4091 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4092 getF32Constant(DAG, 0x3d431f31, dl));
4093 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
4094 getF32Constant(DAG, 0x3ea21fb2, dl));
4095 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4096 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4097 getF32Constant(DAG, 0x3f6ae232, dl));
4098 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4099 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
4100 getF32Constant(DAG, 0x3f25f7c3, dl));
4101 } else { // LimitFloatPrecision <= 18
4102 // For floating-point precision of 18:
4104 // Log10ofMantissa =
4109 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x;
4111 // error 0.0000037995730, which is better than 18 bits
4112 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4113 getF32Constant(DAG, 0x3c5d51ce, dl));
4114 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
4115 getF32Constant(DAG, 0x3e00685a, dl));
4116 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4117 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4118 getF32Constant(DAG, 0x3efb6798, dl));
4119 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4120 SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
4121 getF32Constant(DAG, 0x3f88d192, dl));
4122 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4123 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4124 getF32Constant(DAG, 0x3fc4316c, dl));
4125 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4126 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8,
4127 getF32Constant(DAG, 0x3f57ce70, dl));
4130 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, Log10ofMantissa);
4133 // No special expansion.
4134 return DAG.getNode(ISD::FLOG10, dl, Op.getValueType(), Op);
4137 /// expandExp2 - Lower an exp2 intrinsic. Handles the special sequences for
4138 /// limited-precision mode.
4139 static SDValue expandExp2(SDLoc dl, SDValue Op, SelectionDAG &DAG,
4140 const TargetLowering &TLI) {
4141 if (Op.getValueType() == MVT::f32 &&
4142 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18)
4143 return getLimitedPrecisionExp2(Op, dl, DAG);
4145 // No special expansion.
4146 return DAG.getNode(ISD::FEXP2, dl, Op.getValueType(), Op);
4149 /// visitPow - Lower a pow intrinsic. Handles the special sequences for
4150 /// limited-precision mode with x == 10.0f.
4151 static SDValue expandPow(SDLoc dl, SDValue LHS, SDValue RHS,
4152 SelectionDAG &DAG, const TargetLowering &TLI) {
4153 bool IsExp10 = false;
4154 if (LHS.getValueType() == MVT::f32 && RHS.getValueType() == MVT::f32 &&
4155 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4156 if (ConstantFPSDNode *LHSC = dyn_cast<ConstantFPSDNode>(LHS)) {
4158 IsExp10 = LHSC->isExactlyValue(Ten);
4162 // TODO: What fast-math-flags should be set on the FMUL node?
4164 // Put the exponent in the right bit position for later addition to the
4167 // #define LOG2OF10 3.3219281f
4168 // t0 = Op * LOG2OF10;
4169 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, RHS,
4170 getF32Constant(DAG, 0x40549a78, dl));
4171 return getLimitedPrecisionExp2(t0, dl, DAG);
4174 // No special expansion.
4175 return DAG.getNode(ISD::FPOW, dl, LHS.getValueType(), LHS, RHS);
4179 /// ExpandPowI - Expand a llvm.powi intrinsic.
4180 static SDValue ExpandPowI(SDLoc DL, SDValue LHS, SDValue RHS,
4181 SelectionDAG &DAG) {
4182 // If RHS is a constant, we can expand this out to a multiplication tree,
4183 // otherwise we end up lowering to a call to __powidf2 (for example). When
4184 // optimizing for size, we only want to do this if the expansion would produce
4185 // a small number of multiplies, otherwise we do the full expansion.
4186 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
4187 // Get the exponent as a positive value.
4188 unsigned Val = RHSC->getSExtValue();
4189 if ((int)Val < 0) Val = -Val;
4191 // powi(x, 0) -> 1.0
4193 return DAG.getConstantFP(1.0, DL, LHS.getValueType());
4195 const Function *F = DAG.getMachineFunction().getFunction();
4196 if (!F->optForSize() ||
4197 // If optimizing for size, don't insert too many multiplies.
4198 // This inserts up to 5 multiplies.
4199 countPopulation(Val) + Log2_32(Val) < 7) {
4200 // We use the simple binary decomposition method to generate the multiply
4201 // sequence. There are more optimal ways to do this (for example,
4202 // powi(x,15) generates one more multiply than it should), but this has
4203 // the benefit of being both really simple and much better than a libcall.
4204 SDValue Res; // Logically starts equal to 1.0
4205 SDValue CurSquare = LHS;
4206 // TODO: Intrinsics should have fast-math-flags that propagate to these
4211 Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare);
4213 Res = CurSquare; // 1.0*CurSquare.
4216 CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(),
4217 CurSquare, CurSquare);
4221 // If the original was negative, invert the result, producing 1/(x*x*x).
4222 if (RHSC->getSExtValue() < 0)
4223 Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(),
4224 DAG.getConstantFP(1.0, DL, LHS.getValueType()), Res);
4229 // Otherwise, expand to a libcall.
4230 return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS);
4233 // getUnderlyingArgReg - Find underlying register used for a truncated or
4234 // bitcasted argument.
4235 static unsigned getUnderlyingArgReg(const SDValue &N) {
4236 switch (N.getOpcode()) {
4237 case ISD::CopyFromReg:
4238 return cast<RegisterSDNode>(N.getOperand(1))->getReg();
4240 case ISD::AssertZext:
4241 case ISD::AssertSext:
4243 return getUnderlyingArgReg(N.getOperand(0));
4249 /// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a function
4250 /// argument, create the corresponding DBG_VALUE machine instruction for it now.
4251 /// At the end of instruction selection, they will be inserted to the entry BB.
4252 bool SelectionDAGBuilder::EmitFuncArgumentDbgValue(
4253 const Value *V, DILocalVariable *Variable, DIExpression *Expr,
4254 DILocation *DL, int64_t Offset, bool IsIndirect, const SDValue &N) {
4255 const Argument *Arg = dyn_cast<Argument>(V);
4259 MachineFunction &MF = DAG.getMachineFunction();
4260 const TargetInstrInfo *TII = DAG.getSubtarget().getInstrInfo();
4262 // Ignore inlined function arguments here.
4264 // FIXME: Should we be checking DL->inlinedAt() to determine this?
4265 if (!Variable->getScope()->getSubprogram()->describes(MF.getFunction()))
4268 Optional<MachineOperand> Op;
4269 // Some arguments' frame index is recorded during argument lowering.
4270 if (int FI = FuncInfo.getArgumentFrameIndex(Arg))
4271 Op = MachineOperand::CreateFI(FI);
4273 if (!Op && N.getNode()) {
4274 unsigned Reg = getUnderlyingArgReg(N);
4275 if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) {
4276 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4277 unsigned PR = RegInfo.getLiveInPhysReg(Reg);
4282 Op = MachineOperand::CreateReg(Reg, false);
4286 // Check if ValueMap has reg number.
4287 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
4288 if (VMI != FuncInfo.ValueMap.end())
4289 Op = MachineOperand::CreateReg(VMI->second, false);
4292 if (!Op && N.getNode())
4293 // Check if frame index is available.
4294 if (LoadSDNode *LNode = dyn_cast<LoadSDNode>(N.getNode()))
4295 if (FrameIndexSDNode *FINode =
4296 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode()))
4297 Op = MachineOperand::CreateFI(FINode->getIndex());
4302 assert(Variable->isValidLocationForIntrinsic(DL) &&
4303 "Expected inlined-at fields to agree");
4305 FuncInfo.ArgDbgValues.push_back(
4306 BuildMI(MF, DL, TII->get(TargetOpcode::DBG_VALUE), IsIndirect,
4307 Op->getReg(), Offset, Variable, Expr));
4309 FuncInfo.ArgDbgValues.push_back(
4310 BuildMI(MF, DL, TII->get(TargetOpcode::DBG_VALUE))
4313 .addMetadata(Variable)
4314 .addMetadata(Expr));
4319 // VisualStudio defines setjmp as _setjmp
4320 #if defined(_MSC_VER) && defined(setjmp) && \
4321 !defined(setjmp_undefined_for_msvc)
4322 # pragma push_macro("setjmp")
4324 # define setjmp_undefined_for_msvc
4327 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
4328 /// we want to emit this as a call to a named external function, return the name
4329 /// otherwise lower it and return null.
4331 SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) {
4332 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4333 SDLoc sdl = getCurSDLoc();
4334 DebugLoc dl = getCurDebugLoc();
4337 switch (Intrinsic) {
4339 // By default, turn this into a target intrinsic node.
4340 visitTargetIntrinsic(I, Intrinsic);
4342 case Intrinsic::vastart: visitVAStart(I); return nullptr;
4343 case Intrinsic::vaend: visitVAEnd(I); return nullptr;
4344 case Intrinsic::vacopy: visitVACopy(I); return nullptr;
4345 case Intrinsic::returnaddress:
4346 setValue(&I, DAG.getNode(ISD::RETURNADDR, sdl,
4347 TLI.getPointerTy(DAG.getDataLayout()),
4348 getValue(I.getArgOperand(0))));
4350 case Intrinsic::frameaddress:
4351 setValue(&I, DAG.getNode(ISD::FRAMEADDR, sdl,
4352 TLI.getPointerTy(DAG.getDataLayout()),
4353 getValue(I.getArgOperand(0))));
4355 case Intrinsic::read_register: {
4356 Value *Reg = I.getArgOperand(0);
4357 SDValue Chain = getRoot();
4359 DAG.getMDNode(cast<MDNode>(cast<MetadataAsValue>(Reg)->getMetadata()));
4360 EVT VT = TLI.getValueType(DAG.getDataLayout(), I.getType());
4361 Res = DAG.getNode(ISD::READ_REGISTER, sdl,
4362 DAG.getVTList(VT, MVT::Other), Chain, RegName);
4364 DAG.setRoot(Res.getValue(1));
4367 case Intrinsic::write_register: {
4368 Value *Reg = I.getArgOperand(0);
4369 Value *RegValue = I.getArgOperand(1);
4370 SDValue Chain = getRoot();
4372 DAG.getMDNode(cast<MDNode>(cast<MetadataAsValue>(Reg)->getMetadata()));
4373 DAG.setRoot(DAG.getNode(ISD::WRITE_REGISTER, sdl, MVT::Other, Chain,
4374 RegName, getValue(RegValue)));
4377 case Intrinsic::setjmp:
4378 return &"_setjmp"[!TLI.usesUnderscoreSetJmp()];
4379 case Intrinsic::longjmp:
4380 return &"_longjmp"[!TLI.usesUnderscoreLongJmp()];
4381 case Intrinsic::memcpy: {
4382 SDValue Op1 = getValue(I.getArgOperand(0));
4383 SDValue Op2 = getValue(I.getArgOperand(1));
4384 SDValue Op3 = getValue(I.getArgOperand(2));
4385 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4387 Align = 1; // @llvm.memcpy defines 0 and 1 to both mean no alignment.
4388 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4389 bool isTC = I.isTailCall() && isInTailCallPosition(&I, DAG.getTarget());
4390 SDValue MC = DAG.getMemcpy(getRoot(), sdl, Op1, Op2, Op3, Align, isVol,
4392 MachinePointerInfo(I.getArgOperand(0)),
4393 MachinePointerInfo(I.getArgOperand(1)));
4394 updateDAGForMaybeTailCall(MC);
4397 case Intrinsic::memset: {
4398 SDValue Op1 = getValue(I.getArgOperand(0));
4399 SDValue Op2 = getValue(I.getArgOperand(1));
4400 SDValue Op3 = getValue(I.getArgOperand(2));
4401 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4403 Align = 1; // @llvm.memset defines 0 and 1 to both mean no alignment.
4404 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4405 bool isTC = I.isTailCall() && isInTailCallPosition(&I, DAG.getTarget());
4406 SDValue MS = DAG.getMemset(getRoot(), sdl, Op1, Op2, Op3, Align, isVol,
4407 isTC, MachinePointerInfo(I.getArgOperand(0)));
4408 updateDAGForMaybeTailCall(MS);
4411 case Intrinsic::memmove: {
4412 SDValue Op1 = getValue(I.getArgOperand(0));
4413 SDValue Op2 = getValue(I.getArgOperand(1));
4414 SDValue Op3 = getValue(I.getArgOperand(2));
4415 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4417 Align = 1; // @llvm.memmove defines 0 and 1 to both mean no alignment.
4418 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4419 bool isTC = I.isTailCall() && isInTailCallPosition(&I, DAG.getTarget());
4420 SDValue MM = DAG.getMemmove(getRoot(), sdl, Op1, Op2, Op3, Align, isVol,
4421 isTC, MachinePointerInfo(I.getArgOperand(0)),
4422 MachinePointerInfo(I.getArgOperand(1)));
4423 updateDAGForMaybeTailCall(MM);
4426 case Intrinsic::dbg_declare: {
4427 const DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
4428 DILocalVariable *Variable = DI.getVariable();
4429 DIExpression *Expression = DI.getExpression();
4430 const Value *Address = DI.getAddress();
4431 assert(Variable && "Missing variable");
4433 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4437 // Check if address has undef value.
4438 if (isa<UndefValue>(Address) ||
4439 (Address->use_empty() && !isa<Argument>(Address))) {
4440 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4444 SDValue &N = NodeMap[Address];
4445 if (!N.getNode() && isa<Argument>(Address))
4446 // Check unused arguments map.
4447 N = UnusedArgNodeMap[Address];
4450 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
4451 Address = BCI->getOperand(0);
4452 // Parameters are handled specially.
4453 bool isParameter = Variable->isParameter() || isa<Argument>(Address);
4454 auto FINode = dyn_cast<FrameIndexSDNode>(N.getNode());
4455 if (isParameter && FINode) {
4456 // Byval parameter. We have a frame index at this point.
4457 SDV = DAG.getFrameIndexDbgValue(Variable, Expression,
4458 FINode->getIndex(), 0, dl, SDNodeOrder);
4459 } else if (isa<Argument>(Address)) {
4460 // Address is an argument, so try to emit its dbg value using
4461 // virtual register info from the FuncInfo.ValueMap.
4462 EmitFuncArgumentDbgValue(Address, Variable, Expression, dl, 0, false,
4466 SDV = DAG.getDbgValue(Variable, Expression, N.getNode(), N.getResNo(),
4467 true, 0, dl, SDNodeOrder);
4469 DAG.AddDbgValue(SDV, N.getNode(), isParameter);
4471 // If Address is an argument then try to emit its dbg value using
4472 // virtual register info from the FuncInfo.ValueMap.
4473 if (!EmitFuncArgumentDbgValue(Address, Variable, Expression, dl, 0, false,
4475 // If variable is pinned by a alloca in dominating bb then
4476 // use StaticAllocaMap.
4477 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Address)) {
4478 if (AI->getParent() != DI.getParent()) {
4479 DenseMap<const AllocaInst*, int>::iterator SI =
4480 FuncInfo.StaticAllocaMap.find(AI);
4481 if (SI != FuncInfo.StaticAllocaMap.end()) {
4482 SDV = DAG.getFrameIndexDbgValue(Variable, Expression, SI->second,
4483 0, dl, SDNodeOrder);
4484 DAG.AddDbgValue(SDV, nullptr, false);
4489 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4494 case Intrinsic::dbg_value: {
4495 const DbgValueInst &DI = cast<DbgValueInst>(I);
4496 assert(DI.getVariable() && "Missing variable");
4498 DILocalVariable *Variable = DI.getVariable();
4499 DIExpression *Expression = DI.getExpression();
4500 uint64_t Offset = DI.getOffset();
4501 const Value *V = DI.getValue();
4506 if (isa<ConstantInt>(V) || isa<ConstantFP>(V) || isa<UndefValue>(V)) {
4507 SDV = DAG.getConstantDbgValue(Variable, Expression, V, Offset, dl,
4509 DAG.AddDbgValue(SDV, nullptr, false);
4511 // Do not use getValue() in here; we don't want to generate code at
4512 // this point if it hasn't been done yet.
4513 SDValue N = NodeMap[V];
4514 if (!N.getNode() && isa<Argument>(V))
4515 // Check unused arguments map.
4516 N = UnusedArgNodeMap[V];
4518 // A dbg.value for an alloca is always indirect.
4519 bool IsIndirect = isa<AllocaInst>(V) || Offset != 0;
4520 if (!EmitFuncArgumentDbgValue(V, Variable, Expression, dl, Offset,
4522 SDV = DAG.getDbgValue(Variable, Expression, N.getNode(), N.getResNo(),
4523 IsIndirect, Offset, dl, SDNodeOrder);
4524 DAG.AddDbgValue(SDV, N.getNode(), false);
4526 } else if (!V->use_empty() ) {
4527 // Do not call getValue(V) yet, as we don't want to generate code.
4528 // Remember it for later.
4529 DanglingDebugInfo DDI(&DI, dl, SDNodeOrder);
4530 DanglingDebugInfoMap[V] = DDI;
4532 // We may expand this to cover more cases. One case where we have no
4533 // data available is an unreferenced parameter.
4534 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4538 // Build a debug info table entry.
4539 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
4540 V = BCI->getOperand(0);
4541 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
4542 // Don't handle byval struct arguments or VLAs, for example.
4544 DEBUG(dbgs() << "Dropping debug location info for:\n " << DI << "\n");
4545 DEBUG(dbgs() << " Last seen at:\n " << *V << "\n");
4548 DenseMap<const AllocaInst*, int>::iterator SI =
4549 FuncInfo.StaticAllocaMap.find(AI);
4550 if (SI == FuncInfo.StaticAllocaMap.end())
4551 return nullptr; // VLAs.
4555 case Intrinsic::eh_typeid_for: {
4556 // Find the type id for the given typeinfo.
4557 GlobalValue *GV = ExtractTypeInfo(I.getArgOperand(0));
4558 unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(GV);
4559 Res = DAG.getConstant(TypeID, sdl, MVT::i32);
4564 case Intrinsic::eh_return_i32:
4565 case Intrinsic::eh_return_i64:
4566 DAG.getMachineFunction().getMMI().setCallsEHReturn(true);
4567 DAG.setRoot(DAG.getNode(ISD::EH_RETURN, sdl,
4570 getValue(I.getArgOperand(0)),
4571 getValue(I.getArgOperand(1))));
4573 case Intrinsic::eh_unwind_init:
4574 DAG.getMachineFunction().getMMI().setCallsUnwindInit(true);
4576 case Intrinsic::eh_dwarf_cfa: {
4577 SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getArgOperand(0)), sdl,
4578 TLI.getPointerTy(DAG.getDataLayout()));
4579 SDValue Offset = DAG.getNode(ISD::ADD, sdl,
4580 CfaArg.getValueType(),
4581 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, sdl,
4582 CfaArg.getValueType()),
4584 SDValue FA = DAG.getNode(
4585 ISD::FRAMEADDR, sdl, TLI.getPointerTy(DAG.getDataLayout()),
4586 DAG.getConstant(0, sdl, TLI.getPointerTy(DAG.getDataLayout())));
4587 setValue(&I, DAG.getNode(ISD::ADD, sdl, FA.getValueType(),
4591 case Intrinsic::eh_sjlj_callsite: {
4592 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4593 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(0));
4594 assert(CI && "Non-constant call site value in eh.sjlj.callsite!");
4595 assert(MMI.getCurrentCallSite() == 0 && "Overlapping call sites!");
4597 MMI.setCurrentCallSite(CI->getZExtValue());
4600 case Intrinsic::eh_sjlj_functioncontext: {
4601 // Get and store the index of the function context.
4602 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4604 cast<AllocaInst>(I.getArgOperand(0)->stripPointerCasts());
4605 int FI = FuncInfo.StaticAllocaMap[FnCtx];
4606 MFI->setFunctionContextIndex(FI);
4609 case Intrinsic::eh_sjlj_setjmp: {
4612 Ops[1] = getValue(I.getArgOperand(0));
4613 SDValue Op = DAG.getNode(ISD::EH_SJLJ_SETJMP, sdl,
4614 DAG.getVTList(MVT::i32, MVT::Other), Ops);
4615 setValue(&I, Op.getValue(0));
4616 DAG.setRoot(Op.getValue(1));
4619 case Intrinsic::eh_sjlj_longjmp: {
4620 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, sdl, MVT::Other,
4621 getRoot(), getValue(I.getArgOperand(0))));
4624 case Intrinsic::eh_sjlj_setup_dispatch: {
4625 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_SETUP_DISPATCH, sdl, MVT::Other,
4630 case Intrinsic::masked_gather:
4631 visitMaskedGather(I);
4633 case Intrinsic::masked_load:
4636 case Intrinsic::masked_scatter:
4637 visitMaskedScatter(I);
4639 case Intrinsic::masked_store:
4640 visitMaskedStore(I);
4642 case Intrinsic::x86_mmx_pslli_w:
4643 case Intrinsic::x86_mmx_pslli_d:
4644 case Intrinsic::x86_mmx_pslli_q:
4645 case Intrinsic::x86_mmx_psrli_w:
4646 case Intrinsic::x86_mmx_psrli_d:
4647 case Intrinsic::x86_mmx_psrli_q:
4648 case Intrinsic::x86_mmx_psrai_w:
4649 case Intrinsic::x86_mmx_psrai_d: {
4650 SDValue ShAmt = getValue(I.getArgOperand(1));
4651 if (isa<ConstantSDNode>(ShAmt)) {
4652 visitTargetIntrinsic(I, Intrinsic);
4655 unsigned NewIntrinsic = 0;
4656 EVT ShAmtVT = MVT::v2i32;
4657 switch (Intrinsic) {
4658 case Intrinsic::x86_mmx_pslli_w:
4659 NewIntrinsic = Intrinsic::x86_mmx_psll_w;
4661 case Intrinsic::x86_mmx_pslli_d:
4662 NewIntrinsic = Intrinsic::x86_mmx_psll_d;
4664 case Intrinsic::x86_mmx_pslli_q:
4665 NewIntrinsic = Intrinsic::x86_mmx_psll_q;
4667 case Intrinsic::x86_mmx_psrli_w:
4668 NewIntrinsic = Intrinsic::x86_mmx_psrl_w;
4670 case Intrinsic::x86_mmx_psrli_d:
4671 NewIntrinsic = Intrinsic::x86_mmx_psrl_d;
4673 case Intrinsic::x86_mmx_psrli_q:
4674 NewIntrinsic = Intrinsic::x86_mmx_psrl_q;
4676 case Intrinsic::x86_mmx_psrai_w:
4677 NewIntrinsic = Intrinsic::x86_mmx_psra_w;
4679 case Intrinsic::x86_mmx_psrai_d:
4680 NewIntrinsic = Intrinsic::x86_mmx_psra_d;
4682 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4685 // The vector shift intrinsics with scalars uses 32b shift amounts but
4686 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
4688 // We must do this early because v2i32 is not a legal type.
4691 ShOps[1] = DAG.getConstant(0, sdl, MVT::i32);
4692 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, sdl, ShAmtVT, ShOps);
4693 EVT DestVT = TLI.getValueType(DAG.getDataLayout(), I.getType());
4694 ShAmt = DAG.getNode(ISD::BITCAST, sdl, DestVT, ShAmt);
4695 Res = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, sdl, DestVT,
4696 DAG.getConstant(NewIntrinsic, sdl, MVT::i32),
4697 getValue(I.getArgOperand(0)), ShAmt);
4701 case Intrinsic::convertff:
4702 case Intrinsic::convertfsi:
4703 case Intrinsic::convertfui:
4704 case Intrinsic::convertsif:
4705 case Intrinsic::convertuif:
4706 case Intrinsic::convertss:
4707 case Intrinsic::convertsu:
4708 case Intrinsic::convertus:
4709 case Intrinsic::convertuu: {
4710 ISD::CvtCode Code = ISD::CVT_INVALID;
4711 switch (Intrinsic) {
4712 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4713 case Intrinsic::convertff: Code = ISD::CVT_FF; break;
4714 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break;
4715 case Intrinsic::convertfui: Code = ISD::CVT_FU; break;
4716 case Intrinsic::convertsif: Code = ISD::CVT_SF; break;
4717 case Intrinsic::convertuif: Code = ISD::CVT_UF; break;
4718 case Intrinsic::convertss: Code = ISD::CVT_SS; break;
4719 case Intrinsic::convertsu: Code = ISD::CVT_SU; break;
4720 case Intrinsic::convertus: Code = ISD::CVT_US; break;
4721 case Intrinsic::convertuu: Code = ISD::CVT_UU; break;
4723 EVT DestVT = TLI.getValueType(DAG.getDataLayout(), I.getType());
4724 const Value *Op1 = I.getArgOperand(0);
4725 Res = DAG.getConvertRndSat(DestVT, sdl, getValue(Op1),
4726 DAG.getValueType(DestVT),
4727 DAG.getValueType(getValue(Op1).getValueType()),
4728 getValue(I.getArgOperand(1)),
4729 getValue(I.getArgOperand(2)),
4734 case Intrinsic::powi:
4735 setValue(&I, ExpandPowI(sdl, getValue(I.getArgOperand(0)),
4736 getValue(I.getArgOperand(1)), DAG));
4738 case Intrinsic::log:
4739 setValue(&I, expandLog(sdl, getValue(I.getArgOperand(0)), DAG, TLI));
4741 case Intrinsic::log2:
4742 setValue(&I, expandLog2(sdl, getValue(I.getArgOperand(0)), DAG, TLI));
4744 case Intrinsic::log10:
4745 setValue(&I, expandLog10(sdl, getValue(I.getArgOperand(0)), DAG, TLI));
4747 case Intrinsic::exp:
4748 setValue(&I, expandExp(sdl, getValue(I.getArgOperand(0)), DAG, TLI));
4750 case Intrinsic::exp2:
4751 setValue(&I, expandExp2(sdl, getValue(I.getArgOperand(0)), DAG, TLI));
4753 case Intrinsic::pow:
4754 setValue(&I, expandPow(sdl, getValue(I.getArgOperand(0)),
4755 getValue(I.getArgOperand(1)), DAG, TLI));
4757 case Intrinsic::sqrt:
4758 case Intrinsic::fabs:
4759 case Intrinsic::sin:
4760 case Intrinsic::cos:
4761 case Intrinsic::floor:
4762 case Intrinsic::ceil:
4763 case Intrinsic::trunc:
4764 case Intrinsic::rint:
4765 case Intrinsic::nearbyint:
4766 case Intrinsic::round: {
4768 switch (Intrinsic) {
4769 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4770 case Intrinsic::sqrt: Opcode = ISD::FSQRT; break;
4771 case Intrinsic::fabs: Opcode = ISD::FABS; break;
4772 case Intrinsic::sin: Opcode = ISD::FSIN; break;
4773 case Intrinsic::cos: Opcode = ISD::FCOS; break;
4774 case Intrinsic::floor: Opcode = ISD::FFLOOR; break;
4775 case Intrinsic::ceil: Opcode = ISD::FCEIL; break;
4776 case Intrinsic::trunc: Opcode = ISD::FTRUNC; break;
4777 case Intrinsic::rint: Opcode = ISD::FRINT; break;
4778 case Intrinsic::nearbyint: Opcode = ISD::FNEARBYINT; break;
4779 case Intrinsic::round: Opcode = ISD::FROUND; break;
4782 setValue(&I, DAG.getNode(Opcode, sdl,
4783 getValue(I.getArgOperand(0)).getValueType(),
4784 getValue(I.getArgOperand(0))));
4787 case Intrinsic::minnum:
4788 setValue(&I, DAG.getNode(ISD::FMINNUM, sdl,
4789 getValue(I.getArgOperand(0)).getValueType(),
4790 getValue(I.getArgOperand(0)),
4791 getValue(I.getArgOperand(1))));
4793 case Intrinsic::maxnum:
4794 setValue(&I, DAG.getNode(ISD::FMAXNUM, sdl,
4795 getValue(I.getArgOperand(0)).getValueType(),
4796 getValue(I.getArgOperand(0)),
4797 getValue(I.getArgOperand(1))));
4799 case Intrinsic::copysign:
4800 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, sdl,
4801 getValue(I.getArgOperand(0)).getValueType(),
4802 getValue(I.getArgOperand(0)),
4803 getValue(I.getArgOperand(1))));
4805 case Intrinsic::fma:
4806 setValue(&I, DAG.getNode(ISD::FMA, sdl,
4807 getValue(I.getArgOperand(0)).getValueType(),
4808 getValue(I.getArgOperand(0)),
4809 getValue(I.getArgOperand(1)),
4810 getValue(I.getArgOperand(2))));
4812 case Intrinsic::fmuladd: {
4813 EVT VT = TLI.getValueType(DAG.getDataLayout(), I.getType());
4814 if (TM.Options.AllowFPOpFusion != FPOpFusion::Strict &&
4815 TLI.isFMAFasterThanFMulAndFAdd(VT)) {
4816 setValue(&I, DAG.getNode(ISD::FMA, sdl,
4817 getValue(I.getArgOperand(0)).getValueType(),
4818 getValue(I.getArgOperand(0)),
4819 getValue(I.getArgOperand(1)),
4820 getValue(I.getArgOperand(2))));
4822 // TODO: Intrinsic calls should have fast-math-flags.
4823 SDValue Mul = DAG.getNode(ISD::FMUL, sdl,
4824 getValue(I.getArgOperand(0)).getValueType(),
4825 getValue(I.getArgOperand(0)),
4826 getValue(I.getArgOperand(1)));
4827 SDValue Add = DAG.getNode(ISD::FADD, sdl,
4828 getValue(I.getArgOperand(0)).getValueType(),
4830 getValue(I.getArgOperand(2)));
4835 case Intrinsic::convert_to_fp16:
4836 setValue(&I, DAG.getNode(ISD::BITCAST, sdl, MVT::i16,
4837 DAG.getNode(ISD::FP_ROUND, sdl, MVT::f16,
4838 getValue(I.getArgOperand(0)),
4839 DAG.getTargetConstant(0, sdl,
4842 case Intrinsic::convert_from_fp16:
4843 setValue(&I, DAG.getNode(ISD::FP_EXTEND, sdl,
4844 TLI.getValueType(DAG.getDataLayout(), I.getType()),
4845 DAG.getNode(ISD::BITCAST, sdl, MVT::f16,
4846 getValue(I.getArgOperand(0)))));
4848 case Intrinsic::pcmarker: {
4849 SDValue Tmp = getValue(I.getArgOperand(0));
4850 DAG.setRoot(DAG.getNode(ISD::PCMARKER, sdl, MVT::Other, getRoot(), Tmp));
4853 case Intrinsic::readcyclecounter: {
4854 SDValue Op = getRoot();
4855 Res = DAG.getNode(ISD::READCYCLECOUNTER, sdl,
4856 DAG.getVTList(MVT::i64, MVT::Other), Op);
4858 DAG.setRoot(Res.getValue(1));
4861 case Intrinsic::bitreverse:
4862 setValue(&I, DAG.getNode(ISD::BITREVERSE, sdl,
4863 getValue(I.getArgOperand(0)).getValueType(),
4864 getValue(I.getArgOperand(0))));
4866 case Intrinsic::bswap:
4867 setValue(&I, DAG.getNode(ISD::BSWAP, sdl,
4868 getValue(I.getArgOperand(0)).getValueType(),
4869 getValue(I.getArgOperand(0))));
4871 case Intrinsic::cttz: {
4872 SDValue Arg = getValue(I.getArgOperand(0));
4873 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1));
4874 EVT Ty = Arg.getValueType();
4875 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTTZ : ISD::CTTZ_ZERO_UNDEF,
4879 case Intrinsic::ctlz: {
4880 SDValue Arg = getValue(I.getArgOperand(0));
4881 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1));
4882 EVT Ty = Arg.getValueType();
4883 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTLZ : ISD::CTLZ_ZERO_UNDEF,
4887 case Intrinsic::ctpop: {
4888 SDValue Arg = getValue(I.getArgOperand(0));
4889 EVT Ty = Arg.getValueType();
4890 setValue(&I, DAG.getNode(ISD::CTPOP, sdl, Ty, Arg));
4893 case Intrinsic::stacksave: {
4894 SDValue Op = getRoot();
4896 ISD::STACKSAVE, sdl,
4897 DAG.getVTList(TLI.getPointerTy(DAG.getDataLayout()), MVT::Other), Op);
4899 DAG.setRoot(Res.getValue(1));
4902 case Intrinsic::stackrestore: {
4903 Res = getValue(I.getArgOperand(0));
4904 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, sdl, MVT::Other, getRoot(), Res));
4907 case Intrinsic::get_dynamic_area_offset: {
4908 SDValue Op = getRoot();
4909 EVT PtrTy = TLI.getPointerTy(DAG.getDataLayout());
4910 EVT ResTy = TLI.getValueType(DAG.getDataLayout(), I.getType());
4911 // Result type for @llvm.get.dynamic.area.offset should match PtrTy for
4914 report_fatal_error("Wrong result type for @llvm.get.dynamic.area.offset"
4916 Res = DAG.getNode(ISD::GET_DYNAMIC_AREA_OFFSET, sdl, DAG.getVTList(ResTy),
4922 case Intrinsic::stackprotector: {
4923 // Emit code into the DAG to store the stack guard onto the stack.
4924 MachineFunction &MF = DAG.getMachineFunction();
4925 MachineFrameInfo *MFI = MF.getFrameInfo();
4926 EVT PtrTy = TLI.getPointerTy(DAG.getDataLayout());
4927 SDValue Src, Chain = getRoot();
4928 const Value *Ptr = cast<LoadInst>(I.getArgOperand(0))->getPointerOperand();
4929 const GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr);
4931 // See if Ptr is a bitcast. If it is, look through it and see if we can get
4932 // global variable __stack_chk_guard.
4934 if (const Operator *BC = dyn_cast<Operator>(Ptr))
4935 if (BC->getOpcode() == Instruction::BitCast)
4936 GV = dyn_cast<GlobalVariable>(BC->getOperand(0));
4938 if (GV && TLI.useLoadStackGuardNode()) {
4939 // Emit a LOAD_STACK_GUARD node.
4940 MachineSDNode *Node = DAG.getMachineNode(TargetOpcode::LOAD_STACK_GUARD,
4942 MachinePointerInfo MPInfo(GV);
4943 MachineInstr::mmo_iterator MemRefs = MF.allocateMemRefsArray(1);
4944 unsigned Flags = MachineMemOperand::MOLoad |
4945 MachineMemOperand::MOInvariant;
4946 *MemRefs = MF.getMachineMemOperand(MPInfo, Flags,
4947 PtrTy.getSizeInBits() / 8,
4948 DAG.getEVTAlignment(PtrTy));
4949 Node->setMemRefs(MemRefs, MemRefs + 1);
4951 // Copy the guard value to a virtual register so that it can be
4952 // retrieved in the epilogue.
4953 Src = SDValue(Node, 0);
4954 const TargetRegisterClass *RC =
4955 TLI.getRegClassFor(Src.getSimpleValueType());
4956 unsigned Reg = MF.getRegInfo().createVirtualRegister(RC);
4958 SPDescriptor.setGuardReg(Reg);
4959 Chain = DAG.getCopyToReg(Chain, sdl, Reg, Src);
4961 Src = getValue(I.getArgOperand(0)); // The guard's value.
4964 AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
4966 int FI = FuncInfo.StaticAllocaMap[Slot];
4967 MFI->setStackProtectorIndex(FI);
4969 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
4971 // Store the stack protector onto the stack.
4972 Res = DAG.getStore(Chain, sdl, Src, FIN, MachinePointerInfo::getFixedStack(
4973 DAG.getMachineFunction(), FI),
4979 case Intrinsic::objectsize: {
4980 // If we don't know by now, we're never going to know.
4981 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1));
4983 assert(CI && "Non-constant type in __builtin_object_size?");
4985 SDValue Arg = getValue(I.getCalledValue());
4986 EVT Ty = Arg.getValueType();
4989 Res = DAG.getConstant(-1ULL, sdl, Ty);
4991 Res = DAG.getConstant(0, sdl, Ty);
4996 case Intrinsic::annotation:
4997 case Intrinsic::ptr_annotation:
4998 // Drop the intrinsic, but forward the value
4999 setValue(&I, getValue(I.getOperand(0)));
5001 case Intrinsic::assume:
5002 case Intrinsic::var_annotation:
5003 // Discard annotate attributes and assumptions
5006 case Intrinsic::init_trampoline: {
5007 const Function *F = cast<Function>(I.getArgOperand(1)->stripPointerCasts());
5011 Ops[1] = getValue(I.getArgOperand(0));
5012 Ops[2] = getValue(I.getArgOperand(1));
5013 Ops[3] = getValue(I.getArgOperand(2));
5014 Ops[4] = DAG.getSrcValue(I.getArgOperand(0));
5015 Ops[5] = DAG.getSrcValue(F);
5017 Res = DAG.getNode(ISD::INIT_TRAMPOLINE, sdl, MVT::Other, Ops);
5022 case Intrinsic::adjust_trampoline: {
5023 setValue(&I, DAG.getNode(ISD::ADJUST_TRAMPOLINE, sdl,
5024 TLI.getPointerTy(DAG.getDataLayout()),
5025 getValue(I.getArgOperand(0))));
5028 case Intrinsic::gcroot:
5030 const Value *Alloca = I.getArgOperand(0)->stripPointerCasts();
5031 const Constant *TypeMap = cast<Constant>(I.getArgOperand(1));
5033 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode());
5034 GFI->addStackRoot(FI->getIndex(), TypeMap);
5037 case Intrinsic::gcread:
5038 case Intrinsic::gcwrite:
5039 llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!");
5040 case Intrinsic::flt_rounds:
5041 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, sdl, MVT::i32));
5044 case Intrinsic::expect: {
5045 // Just replace __builtin_expect(exp, c) with EXP.
5046 setValue(&I, getValue(I.getArgOperand(0)));
5050 case Intrinsic::debugtrap:
5051 case Intrinsic::trap: {
5052 StringRef TrapFuncName =
5054 .getAttribute(AttributeSet::FunctionIndex, "trap-func-name")
5055 .getValueAsString();
5056 if (TrapFuncName.empty()) {
5057 ISD::NodeType Op = (Intrinsic == Intrinsic::trap) ?
5058 ISD::TRAP : ISD::DEBUGTRAP;
5059 DAG.setRoot(DAG.getNode(Op, sdl,MVT::Other, getRoot()));
5062 TargetLowering::ArgListTy Args;
5064 TargetLowering::CallLoweringInfo CLI(DAG);
5065 CLI.setDebugLoc(sdl).setChain(getRoot()).setCallee(
5066 CallingConv::C, I.getType(),
5067 DAG.getExternalSymbol(TrapFuncName.data(),
5068 TLI.getPointerTy(DAG.getDataLayout())),
5069 std::move(Args), 0);
5071 std::pair<SDValue, SDValue> Result = TLI.LowerCallTo(CLI);
5072 DAG.setRoot(Result.second);
5076 case Intrinsic::uadd_with_overflow:
5077 case Intrinsic::sadd_with_overflow:
5078 case Intrinsic::usub_with_overflow:
5079 case Intrinsic::ssub_with_overflow:
5080 case Intrinsic::umul_with_overflow:
5081 case Intrinsic::smul_with_overflow: {
5083 switch (Intrinsic) {
5084 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
5085 case Intrinsic::uadd_with_overflow: Op = ISD::UADDO; break;
5086 case Intrinsic::sadd_with_overflow: Op = ISD::SADDO; break;
5087 case Intrinsic::usub_with_overflow: Op = ISD::USUBO; break;
5088 case Intrinsic::ssub_with_overflow: Op = ISD::SSUBO; break;
5089 case Intrinsic::umul_with_overflow: Op = ISD::UMULO; break;
5090 case Intrinsic::smul_with_overflow: Op = ISD::SMULO; break;
5092 SDValue Op1 = getValue(I.getArgOperand(0));
5093 SDValue Op2 = getValue(I.getArgOperand(1));
5095 SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1);
5096 setValue(&I, DAG.getNode(Op, sdl, VTs, Op1, Op2));
5099 case Intrinsic::prefetch: {
5101 unsigned rw = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
5103 Ops[1] = getValue(I.getArgOperand(0));
5104 Ops[2] = getValue(I.getArgOperand(1));
5105 Ops[3] = getValue(I.getArgOperand(2));
5106 Ops[4] = getValue(I.getArgOperand(3));
5107 DAG.setRoot(DAG.getMemIntrinsicNode(ISD::PREFETCH, sdl,
5108 DAG.getVTList(MVT::Other), Ops,
5109 EVT::getIntegerVT(*Context, 8),
5110 MachinePointerInfo(I.getArgOperand(0)),
5112 false, /* volatile */
5114 rw==1)); /* write */
5117 case Intrinsic::lifetime_start:
5118 case Intrinsic::lifetime_end: {
5119 bool IsStart = (Intrinsic == Intrinsic::lifetime_start);
5120 // Stack coloring is not enabled in O0, discard region information.
5121 if (TM.getOptLevel() == CodeGenOpt::None)
5124 SmallVector<Value *, 4> Allocas;
5125 GetUnderlyingObjects(I.getArgOperand(1), Allocas, *DL);
5127 for (SmallVectorImpl<Value*>::iterator Object = Allocas.begin(),
5128 E = Allocas.end(); Object != E; ++Object) {
5129 AllocaInst *LifetimeObject = dyn_cast_or_null<AllocaInst>(*Object);
5131 // Could not find an Alloca.
5132 if (!LifetimeObject)
5135 // First check that the Alloca is static, otherwise it won't have a
5136 // valid frame index.
5137 auto SI = FuncInfo.StaticAllocaMap.find(LifetimeObject);
5138 if (SI == FuncInfo.StaticAllocaMap.end())
5141 int FI = SI->second;
5146 DAG.getFrameIndex(FI, TLI.getPointerTy(DAG.getDataLayout()), true);
5147 unsigned Opcode = (IsStart ? ISD::LIFETIME_START : ISD::LIFETIME_END);
5149 Res = DAG.getNode(Opcode, sdl, MVT::Other, Ops);
5154 case Intrinsic::invariant_start:
5155 // Discard region information.
5156 setValue(&I, DAG.getUNDEF(TLI.getPointerTy(DAG.getDataLayout())));
5158 case Intrinsic::invariant_end:
5159 // Discard region information.
5161 case Intrinsic::stackprotectorcheck: {
5162 // Do not actually emit anything for this basic block. Instead we initialize
5163 // the stack protector descriptor and export the guard variable so we can
5164 // access it in FinishBasicBlock.
5165 const BasicBlock *BB = I.getParent();
5166 SPDescriptor.initialize(BB, FuncInfo.MBBMap[BB], I);
5167 ExportFromCurrentBlock(SPDescriptor.getGuard());
5169 // Flush our exports since we are going to process a terminator.
5170 (void)getControlRoot();
5173 case Intrinsic::clear_cache:
5174 return TLI.getClearCacheBuiltinName();
5175 case Intrinsic::donothing:
5178 case Intrinsic::experimental_stackmap: {
5182 case Intrinsic::experimental_patchpoint_void:
5183 case Intrinsic::experimental_patchpoint_i64: {
5184 visitPatchpoint(&I);
5187 case Intrinsic::experimental_gc_statepoint: {
5191 case Intrinsic::experimental_gc_result_int:
5192 case Intrinsic::experimental_gc_result_float:
5193 case Intrinsic::experimental_gc_result_ptr:
5194 case Intrinsic::experimental_gc_result: {
5198 case Intrinsic::experimental_gc_relocate: {
5202 case Intrinsic::instrprof_increment:
5203 llvm_unreachable("instrprof failed to lower an increment");
5204 case Intrinsic::instrprof_value_profile:
5205 llvm_unreachable("instrprof failed to lower a value profiling call");
5206 case Intrinsic::localescape: {
5207 MachineFunction &MF = DAG.getMachineFunction();
5208 const TargetInstrInfo *TII = DAG.getSubtarget().getInstrInfo();
5210 // Directly emit some LOCAL_ESCAPE machine instrs. Label assignment emission
5211 // is the same on all targets.
5212 for (unsigned Idx = 0, E = I.getNumArgOperands(); Idx < E; ++Idx) {
5213 Value *Arg = I.getArgOperand(Idx)->stripPointerCasts();
5214 if (isa<ConstantPointerNull>(Arg))
5215 continue; // Skip null pointers. They represent a hole in index space.
5216 AllocaInst *Slot = cast<AllocaInst>(Arg);
5217 assert(FuncInfo.StaticAllocaMap.count(Slot) &&
5218 "can only escape static allocas");
5219 int FI = FuncInfo.StaticAllocaMap[Slot];
5220 MCSymbol *FrameAllocSym =
5221 MF.getMMI().getContext().getOrCreateFrameAllocSymbol(
5222 GlobalValue::getRealLinkageName(MF.getName()), Idx);
5223 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, dl,
5224 TII->get(TargetOpcode::LOCAL_ESCAPE))
5225 .addSym(FrameAllocSym)
5232 case Intrinsic::localrecover: {
5233 // i8* @llvm.localrecover(i8* %fn, i8* %fp, i32 %idx)
5234 MachineFunction &MF = DAG.getMachineFunction();
5235 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout(), 0);
5237 // Get the symbol that defines the frame offset.
5238 auto *Fn = cast<Function>(I.getArgOperand(0)->stripPointerCasts());
5239 auto *Idx = cast<ConstantInt>(I.getArgOperand(2));
5240 unsigned IdxVal = unsigned(Idx->getLimitedValue(INT_MAX));
5241 MCSymbol *FrameAllocSym =
5242 MF.getMMI().getContext().getOrCreateFrameAllocSymbol(
5243 GlobalValue::getRealLinkageName(Fn->getName()), IdxVal);
5245 // Create a MCSymbol for the label to avoid any target lowering
5246 // that would make this PC relative.
5247 SDValue OffsetSym = DAG.getMCSymbol(FrameAllocSym, PtrVT);
5249 DAG.getNode(ISD::LOCAL_RECOVER, sdl, PtrVT, OffsetSym);
5251 // Add the offset to the FP.
5252 Value *FP = I.getArgOperand(1);
5253 SDValue FPVal = getValue(FP);
5254 SDValue Add = DAG.getNode(ISD::ADD, sdl, PtrVT, FPVal, OffsetVal);
5260 case Intrinsic::eh_exceptionpointer:
5261 case Intrinsic::eh_exceptioncode: {
5262 // Get the exception pointer vreg, copy from it, and resize it to fit.
5263 const auto *CPI = cast<CatchPadInst>(I.getArgOperand(0));
5264 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
5265 const TargetRegisterClass *PtrRC = TLI.getRegClassFor(PtrVT);
5266 unsigned VReg = FuncInfo.getCatchPadExceptionPointerVReg(CPI, PtrRC);
5268 DAG.getCopyFromReg(DAG.getEntryNode(), getCurSDLoc(), VReg, PtrVT);
5269 if (Intrinsic == Intrinsic::eh_exceptioncode)
5270 N = DAG.getZExtOrTrunc(N, getCurSDLoc(), MVT::i32);
5277 std::pair<SDValue, SDValue>
5278 SelectionDAGBuilder::lowerInvokable(TargetLowering::CallLoweringInfo &CLI,
5279 const BasicBlock *EHPadBB) {
5280 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5281 MCSymbol *BeginLabel = nullptr;
5284 // Insert a label before the invoke call to mark the try range. This can be
5285 // used to detect deletion of the invoke via the MachineModuleInfo.
5286 BeginLabel = MMI.getContext().createTempSymbol();
5288 // For SjLj, keep track of which landing pads go with which invokes
5289 // so as to maintain the ordering of pads in the LSDA.
5290 unsigned CallSiteIndex = MMI.getCurrentCallSite();
5291 if (CallSiteIndex) {
5292 MMI.setCallSiteBeginLabel(BeginLabel, CallSiteIndex);
5293 LPadToCallSiteMap[FuncInfo.MBBMap[EHPadBB]].push_back(CallSiteIndex);
5295 // Now that the call site is handled, stop tracking it.
5296 MMI.setCurrentCallSite(0);
5299 // Both PendingLoads and PendingExports must be flushed here;
5300 // this call might not return.
5302 DAG.setRoot(DAG.getEHLabel(getCurSDLoc(), getControlRoot(), BeginLabel));
5304 CLI.setChain(getRoot());
5306 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5307 std::pair<SDValue, SDValue> Result = TLI.LowerCallTo(CLI);
5309 assert((CLI.IsTailCall || Result.second.getNode()) &&
5310 "Non-null chain expected with non-tail call!");
5311 assert((Result.second.getNode() || !Result.first.getNode()) &&
5312 "Null value expected with tail call!");
5314 if (!Result.second.getNode()) {
5315 // As a special case, a null chain means that a tail call has been emitted
5316 // and the DAG root is already updated.
5319 // Since there's no actual continuation from this block, nothing can be
5320 // relying on us setting vregs for them.
5321 PendingExports.clear();
5323 DAG.setRoot(Result.second);
5327 // Insert a label at the end of the invoke call to mark the try range. This
5328 // can be used to detect deletion of the invoke via the MachineModuleInfo.
5329 MCSymbol *EndLabel = MMI.getContext().createTempSymbol();
5330 DAG.setRoot(DAG.getEHLabel(getCurSDLoc(), getRoot(), EndLabel));
5332 // Inform MachineModuleInfo of range.
5333 if (MMI.hasEHFunclets()) {
5335 WinEHFuncInfo *EHInfo = DAG.getMachineFunction().getWinEHFuncInfo();
5336 EHInfo->addIPToStateRange(cast<InvokeInst>(CLI.CS->getInstruction()),
5337 BeginLabel, EndLabel);
5339 MMI.addInvoke(FuncInfo.MBBMap[EHPadBB], BeginLabel, EndLabel);
5346 void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee,
5348 const BasicBlock *EHPadBB) {
5349 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
5350 FunctionType *FTy = cast<FunctionType>(PT->getElementType());
5351 Type *RetTy = FTy->getReturnType();
5353 TargetLowering::ArgListTy Args;
5354 TargetLowering::ArgListEntry Entry;
5355 Args.reserve(CS.arg_size());
5357 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
5359 const Value *V = *i;
5362 if (V->getType()->isEmptyTy())
5365 SDValue ArgNode = getValue(V);
5366 Entry.Node = ArgNode; Entry.Ty = V->getType();
5368 // Skip the first return-type Attribute to get to params.
5369 Entry.setAttributes(&CS, i - CS.arg_begin() + 1);
5370 Args.push_back(Entry);
5372 // If we have an explicit sret argument that is an Instruction, (i.e., it
5373 // might point to function-local memory), we can't meaningfully tail-call.
5374 if (Entry.isSRet && isa<Instruction>(V))
5378 // Check if target-independent constraints permit a tail call here.
5379 // Target-dependent constraints are checked within TLI->LowerCallTo.
5380 if (isTailCall && !isInTailCallPosition(CS, DAG.getTarget()))
5383 TargetLowering::CallLoweringInfo CLI(DAG);
5384 CLI.setDebugLoc(getCurSDLoc()).setChain(getRoot())
5385 .setCallee(RetTy, FTy, Callee, std::move(Args), CS)
5386 .setTailCall(isTailCall);
5387 std::pair<SDValue, SDValue> Result = lowerInvokable(CLI, EHPadBB);
5389 if (Result.first.getNode())
5390 setValue(CS.getInstruction(), Result.first);
5393 /// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the
5394 /// value is equal or not-equal to zero.
5395 static bool IsOnlyUsedInZeroEqualityComparison(const Value *V) {
5396 for (const User *U : V->users()) {
5397 if (const ICmpInst *IC = dyn_cast<ICmpInst>(U))
5398 if (IC->isEquality())
5399 if (const Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
5400 if (C->isNullValue())
5402 // Unknown instruction.
5408 static SDValue getMemCmpLoad(const Value *PtrVal, MVT LoadVT,
5410 SelectionDAGBuilder &Builder) {
5412 // Check to see if this load can be trivially constant folded, e.g. if the
5413 // input is from a string literal.
5414 if (const Constant *LoadInput = dyn_cast<Constant>(PtrVal)) {
5415 // Cast pointer to the type we really want to load.
5416 LoadInput = ConstantExpr::getBitCast(const_cast<Constant *>(LoadInput),
5417 PointerType::getUnqual(LoadTy));
5419 if (const Constant *LoadCst = ConstantFoldLoadFromConstPtr(
5420 const_cast<Constant *>(LoadInput), *Builder.DL))
5421 return Builder.getValue(LoadCst);
5424 // Otherwise, we have to emit the load. If the pointer is to unfoldable but
5425 // still constant memory, the input chain can be the entry node.
5427 bool ConstantMemory = false;
5429 // Do not serialize (non-volatile) loads of constant memory with anything.
5430 if (Builder.AA->pointsToConstantMemory(PtrVal)) {
5431 Root = Builder.DAG.getEntryNode();
5432 ConstantMemory = true;
5434 // Do not serialize non-volatile loads against each other.
5435 Root = Builder.DAG.getRoot();
5438 SDValue Ptr = Builder.getValue(PtrVal);
5439 SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurSDLoc(), Root,
5440 Ptr, MachinePointerInfo(PtrVal),
5442 false /*nontemporal*/,
5443 false /*isinvariant*/, 1 /* align=1 */);
5445 if (!ConstantMemory)
5446 Builder.PendingLoads.push_back(LoadVal.getValue(1));
5450 /// processIntegerCallValue - Record the value for an instruction that
5451 /// produces an integer result, converting the type where necessary.
5452 void SelectionDAGBuilder::processIntegerCallValue(const Instruction &I,
5455 EVT VT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
5458 Value = DAG.getSExtOrTrunc(Value, getCurSDLoc(), VT);
5460 Value = DAG.getZExtOrTrunc(Value, getCurSDLoc(), VT);
5461 setValue(&I, Value);
5464 /// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form.
5465 /// If so, return true and lower it, otherwise return false and it will be
5466 /// lowered like a normal call.
5467 bool SelectionDAGBuilder::visitMemCmpCall(const CallInst &I) {
5468 // Verify that the prototype makes sense. int memcmp(void*,void*,size_t)
5469 if (I.getNumArgOperands() != 3)
5472 const Value *LHS = I.getArgOperand(0), *RHS = I.getArgOperand(1);
5473 if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() ||
5474 !I.getArgOperand(2)->getType()->isIntegerTy() ||
5475 !I.getType()->isIntegerTy())
5478 const Value *Size = I.getArgOperand(2);
5479 const ConstantInt *CSize = dyn_cast<ConstantInt>(Size);
5480 if (CSize && CSize->getZExtValue() == 0) {
5481 EVT CallVT = DAG.getTargetLoweringInfo().getValueType(DAG.getDataLayout(),
5483 setValue(&I, DAG.getConstant(0, getCurSDLoc(), CallVT));
5487 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5488 std::pair<SDValue, SDValue> Res =
5489 TSI.EmitTargetCodeForMemcmp(DAG, getCurSDLoc(), DAG.getRoot(),
5490 getValue(LHS), getValue(RHS), getValue(Size),
5491 MachinePointerInfo(LHS),
5492 MachinePointerInfo(RHS));
5493 if (Res.first.getNode()) {
5494 processIntegerCallValue(I, Res.first, true);
5495 PendingLoads.push_back(Res.second);
5499 // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0
5500 // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0
5501 if (CSize && IsOnlyUsedInZeroEqualityComparison(&I)) {
5502 bool ActuallyDoIt = true;
5505 switch (CSize->getZExtValue()) {
5507 LoadVT = MVT::Other;
5509 ActuallyDoIt = false;
5513 LoadTy = Type::getInt16Ty(CSize->getContext());
5517 LoadTy = Type::getInt32Ty(CSize->getContext());
5521 LoadTy = Type::getInt64Ty(CSize->getContext());
5525 LoadVT = MVT::v4i32;
5526 LoadTy = Type::getInt32Ty(CSize->getContext());
5527 LoadTy = VectorType::get(LoadTy, 4);
5532 // This turns into unaligned loads. We only do this if the target natively
5533 // supports the MVT we'll be loading or if it is small enough (<= 4) that
5534 // we'll only produce a small number of byte loads.
5536 // Require that we can find a legal MVT, and only do this if the target
5537 // supports unaligned loads of that type. Expanding into byte loads would
5539 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5540 if (ActuallyDoIt && CSize->getZExtValue() > 4) {
5541 unsigned DstAS = LHS->getType()->getPointerAddressSpace();
5542 unsigned SrcAS = RHS->getType()->getPointerAddressSpace();
5543 // TODO: Handle 5 byte compare as 4-byte + 1 byte.
5544 // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads.
5545 // TODO: Check alignment of src and dest ptrs.
5546 if (!TLI.isTypeLegal(LoadVT) ||
5547 !TLI.allowsMisalignedMemoryAccesses(LoadVT, SrcAS) ||
5548 !TLI.allowsMisalignedMemoryAccesses(LoadVT, DstAS))
5549 ActuallyDoIt = false;
5553 SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this);
5554 SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this);
5556 SDValue Res = DAG.getSetCC(getCurSDLoc(), MVT::i1, LHSVal, RHSVal,
5558 processIntegerCallValue(I, Res, false);
5567 /// visitMemChrCall -- See if we can lower a memchr call into an optimized
5568 /// form. If so, return true and lower it, otherwise return false and it
5569 /// will be lowered like a normal call.
5570 bool SelectionDAGBuilder::visitMemChrCall(const CallInst &I) {
5571 // Verify that the prototype makes sense. void *memchr(void *, int, size_t)
5572 if (I.getNumArgOperands() != 3)
5575 const Value *Src = I.getArgOperand(0);
5576 const Value *Char = I.getArgOperand(1);
5577 const Value *Length = I.getArgOperand(2);
5578 if (!Src->getType()->isPointerTy() ||
5579 !Char->getType()->isIntegerTy() ||
5580 !Length->getType()->isIntegerTy() ||
5581 !I.getType()->isPointerTy())
5584 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5585 std::pair<SDValue, SDValue> Res =
5586 TSI.EmitTargetCodeForMemchr(DAG, getCurSDLoc(), DAG.getRoot(),
5587 getValue(Src), getValue(Char), getValue(Length),
5588 MachinePointerInfo(Src));
5589 if (Res.first.getNode()) {
5590 setValue(&I, Res.first);
5591 PendingLoads.push_back(Res.second);
5598 /// visitStrCpyCall -- See if we can lower a strcpy or stpcpy call into an
5599 /// optimized form. If so, return true and lower it, otherwise return false
5600 /// and it will be lowered like a normal call.
5601 bool SelectionDAGBuilder::visitStrCpyCall(const CallInst &I, bool isStpcpy) {
5602 // Verify that the prototype makes sense. char *strcpy(char *, char *)
5603 if (I.getNumArgOperands() != 2)
5606 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5607 if (!Arg0->getType()->isPointerTy() ||
5608 !Arg1->getType()->isPointerTy() ||
5609 !I.getType()->isPointerTy())
5612 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5613 std::pair<SDValue, SDValue> Res =
5614 TSI.EmitTargetCodeForStrcpy(DAG, getCurSDLoc(), getRoot(),
5615 getValue(Arg0), getValue(Arg1),
5616 MachinePointerInfo(Arg0),
5617 MachinePointerInfo(Arg1), isStpcpy);
5618 if (Res.first.getNode()) {
5619 setValue(&I, Res.first);
5620 DAG.setRoot(Res.second);
5627 /// visitStrCmpCall - See if we can lower a call to strcmp in an optimized form.
5628 /// If so, return true and lower it, otherwise return false and it will be
5629 /// lowered like a normal call.
5630 bool SelectionDAGBuilder::visitStrCmpCall(const CallInst &I) {
5631 // Verify that the prototype makes sense. int strcmp(void*,void*)
5632 if (I.getNumArgOperands() != 2)
5635 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5636 if (!Arg0->getType()->isPointerTy() ||
5637 !Arg1->getType()->isPointerTy() ||
5638 !I.getType()->isIntegerTy())
5641 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5642 std::pair<SDValue, SDValue> Res =
5643 TSI.EmitTargetCodeForStrcmp(DAG, getCurSDLoc(), DAG.getRoot(),
5644 getValue(Arg0), getValue(Arg1),
5645 MachinePointerInfo(Arg0),
5646 MachinePointerInfo(Arg1));
5647 if (Res.first.getNode()) {
5648 processIntegerCallValue(I, Res.first, true);
5649 PendingLoads.push_back(Res.second);
5656 /// visitStrLenCall -- See if we can lower a strlen call into an optimized
5657 /// form. If so, return true and lower it, otherwise return false and it
5658 /// will be lowered like a normal call.
5659 bool SelectionDAGBuilder::visitStrLenCall(const CallInst &I) {
5660 // Verify that the prototype makes sense. size_t strlen(char *)
5661 if (I.getNumArgOperands() != 1)
5664 const Value *Arg0 = I.getArgOperand(0);
5665 if (!Arg0->getType()->isPointerTy() || !I.getType()->isIntegerTy())
5668 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5669 std::pair<SDValue, SDValue> Res =
5670 TSI.EmitTargetCodeForStrlen(DAG, getCurSDLoc(), DAG.getRoot(),
5671 getValue(Arg0), MachinePointerInfo(Arg0));
5672 if (Res.first.getNode()) {
5673 processIntegerCallValue(I, Res.first, false);
5674 PendingLoads.push_back(Res.second);
5681 /// visitStrNLenCall -- See if we can lower a strnlen call into an optimized
5682 /// form. If so, return true and lower it, otherwise return false and it
5683 /// will be lowered like a normal call.
5684 bool SelectionDAGBuilder::visitStrNLenCall(const CallInst &I) {
5685 // Verify that the prototype makes sense. size_t strnlen(char *, size_t)
5686 if (I.getNumArgOperands() != 2)
5689 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5690 if (!Arg0->getType()->isPointerTy() ||
5691 !Arg1->getType()->isIntegerTy() ||
5692 !I.getType()->isIntegerTy())
5695 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5696 std::pair<SDValue, SDValue> Res =
5697 TSI.EmitTargetCodeForStrnlen(DAG, getCurSDLoc(), DAG.getRoot(),
5698 getValue(Arg0), getValue(Arg1),
5699 MachinePointerInfo(Arg0));
5700 if (Res.first.getNode()) {
5701 processIntegerCallValue(I, Res.first, false);
5702 PendingLoads.push_back(Res.second);
5709 /// visitUnaryFloatCall - If a call instruction is a unary floating-point
5710 /// operation (as expected), translate it to an SDNode with the specified opcode
5711 /// and return true.
5712 bool SelectionDAGBuilder::visitUnaryFloatCall(const CallInst &I,
5714 // Sanity check that it really is a unary floating-point call.
5715 if (I.getNumArgOperands() != 1 ||
5716 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
5717 I.getType() != I.getArgOperand(0)->getType() ||
5718 !I.onlyReadsMemory())
5721 SDValue Tmp = getValue(I.getArgOperand(0));
5722 setValue(&I, DAG.getNode(Opcode, getCurSDLoc(), Tmp.getValueType(), Tmp));
5726 /// visitBinaryFloatCall - If a call instruction is a binary floating-point
5727 /// operation (as expected), translate it to an SDNode with the specified opcode
5728 /// and return true.
5729 bool SelectionDAGBuilder::visitBinaryFloatCall(const CallInst &I,
5731 // Sanity check that it really is a binary floating-point call.
5732 if (I.getNumArgOperands() != 2 ||
5733 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
5734 I.getType() != I.getArgOperand(0)->getType() ||
5735 I.getType() != I.getArgOperand(1)->getType() ||
5736 !I.onlyReadsMemory())
5739 SDValue Tmp0 = getValue(I.getArgOperand(0));
5740 SDValue Tmp1 = getValue(I.getArgOperand(1));
5741 EVT VT = Tmp0.getValueType();
5742 setValue(&I, DAG.getNode(Opcode, getCurSDLoc(), VT, Tmp0, Tmp1));
5746 void SelectionDAGBuilder::visitCall(const CallInst &I) {
5747 // Handle inline assembly differently.
5748 if (isa<InlineAsm>(I.getCalledValue())) {
5753 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5754 ComputeUsesVAFloatArgument(I, &MMI);
5756 const char *RenameFn = nullptr;
5757 if (Function *F = I.getCalledFunction()) {
5758 if (F->isDeclaration()) {
5759 if (const TargetIntrinsicInfo *II = TM.getIntrinsicInfo()) {
5760 if (unsigned IID = II->getIntrinsicID(F)) {
5761 RenameFn = visitIntrinsicCall(I, IID);
5766 if (Intrinsic::ID IID = F->getIntrinsicID()) {
5767 RenameFn = visitIntrinsicCall(I, IID);
5773 // Check for well-known libc/libm calls. If the function is internal, it
5774 // can't be a library call.
5776 if (!F->hasLocalLinkage() && F->hasName() &&
5777 LibInfo->getLibFunc(F->getName(), Func) &&
5778 LibInfo->hasOptimizedCodeGen(Func)) {
5781 case LibFunc::copysign:
5782 case LibFunc::copysignf:
5783 case LibFunc::copysignl:
5784 if (I.getNumArgOperands() == 2 && // Basic sanity checks.
5785 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5786 I.getType() == I.getArgOperand(0)->getType() &&
5787 I.getType() == I.getArgOperand(1)->getType() &&
5788 I.onlyReadsMemory()) {
5789 SDValue LHS = getValue(I.getArgOperand(0));
5790 SDValue RHS = getValue(I.getArgOperand(1));
5791 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurSDLoc(),
5792 LHS.getValueType(), LHS, RHS));
5797 case LibFunc::fabsf:
5798 case LibFunc::fabsl:
5799 if (visitUnaryFloatCall(I, ISD::FABS))
5803 case LibFunc::fminf:
5804 case LibFunc::fminl:
5805 if (visitBinaryFloatCall(I, ISD::FMINNUM))
5809 case LibFunc::fmaxf:
5810 case LibFunc::fmaxl:
5811 if (visitBinaryFloatCall(I, ISD::FMAXNUM))
5817 if (visitUnaryFloatCall(I, ISD::FSIN))
5823 if (visitUnaryFloatCall(I, ISD::FCOS))
5827 case LibFunc::sqrtf:
5828 case LibFunc::sqrtl:
5829 case LibFunc::sqrt_finite:
5830 case LibFunc::sqrtf_finite:
5831 case LibFunc::sqrtl_finite:
5832 if (visitUnaryFloatCall(I, ISD::FSQRT))
5835 case LibFunc::floor:
5836 case LibFunc::floorf:
5837 case LibFunc::floorl:
5838 if (visitUnaryFloatCall(I, ISD::FFLOOR))
5841 case LibFunc::nearbyint:
5842 case LibFunc::nearbyintf:
5843 case LibFunc::nearbyintl:
5844 if (visitUnaryFloatCall(I, ISD::FNEARBYINT))
5848 case LibFunc::ceilf:
5849 case LibFunc::ceill:
5850 if (visitUnaryFloatCall(I, ISD::FCEIL))
5854 case LibFunc::rintf:
5855 case LibFunc::rintl:
5856 if (visitUnaryFloatCall(I, ISD::FRINT))
5859 case LibFunc::round:
5860 case LibFunc::roundf:
5861 case LibFunc::roundl:
5862 if (visitUnaryFloatCall(I, ISD::FROUND))
5865 case LibFunc::trunc:
5866 case LibFunc::truncf:
5867 case LibFunc::truncl:
5868 if (visitUnaryFloatCall(I, ISD::FTRUNC))
5872 case LibFunc::log2f:
5873 case LibFunc::log2l:
5874 if (visitUnaryFloatCall(I, ISD::FLOG2))
5878 case LibFunc::exp2f:
5879 case LibFunc::exp2l:
5880 if (visitUnaryFloatCall(I, ISD::FEXP2))
5883 case LibFunc::memcmp:
5884 if (visitMemCmpCall(I))
5887 case LibFunc::memchr:
5888 if (visitMemChrCall(I))
5891 case LibFunc::strcpy:
5892 if (visitStrCpyCall(I, false))
5895 case LibFunc::stpcpy:
5896 if (visitStrCpyCall(I, true))
5899 case LibFunc::strcmp:
5900 if (visitStrCmpCall(I))
5903 case LibFunc::strlen:
5904 if (visitStrLenCall(I))
5907 case LibFunc::strnlen:
5908 if (visitStrNLenCall(I))
5917 Callee = getValue(I.getCalledValue());
5919 Callee = DAG.getExternalSymbol(
5921 DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout()));
5923 // Check if we can potentially perform a tail call. More detailed checking is
5924 // be done within LowerCallTo, after more information about the call is known.
5925 LowerCallTo(&I, Callee, I.isTailCall());
5930 /// AsmOperandInfo - This contains information for each constraint that we are
5932 class SDISelAsmOperandInfo : public TargetLowering::AsmOperandInfo {
5934 /// CallOperand - If this is the result output operand or a clobber
5935 /// this is null, otherwise it is the incoming operand to the CallInst.
5936 /// This gets modified as the asm is processed.
5937 SDValue CallOperand;
5939 /// AssignedRegs - If this is a register or register class operand, this
5940 /// contains the set of register corresponding to the operand.
5941 RegsForValue AssignedRegs;
5943 explicit SDISelAsmOperandInfo(const TargetLowering::AsmOperandInfo &info)
5944 : TargetLowering::AsmOperandInfo(info), CallOperand(nullptr,0) {
5947 /// getCallOperandValEVT - Return the EVT of the Value* that this operand
5948 /// corresponds to. If there is no Value* for this operand, it returns
5950 EVT getCallOperandValEVT(LLVMContext &Context, const TargetLowering &TLI,
5951 const DataLayout &DL) const {
5952 if (!CallOperandVal) return MVT::Other;
5954 if (isa<BasicBlock>(CallOperandVal))
5955 return TLI.getPointerTy(DL);
5957 llvm::Type *OpTy = CallOperandVal->getType();
5959 // FIXME: code duplicated from TargetLowering::ParseConstraints().
5960 // If this is an indirect operand, the operand is a pointer to the
5963 llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
5965 report_fatal_error("Indirect operand for inline asm not a pointer!");
5966 OpTy = PtrTy->getElementType();
5969 // Look for vector wrapped in a struct. e.g. { <16 x i8> }.
5970 if (StructType *STy = dyn_cast<StructType>(OpTy))
5971 if (STy->getNumElements() == 1)
5972 OpTy = STy->getElementType(0);
5974 // If OpTy is not a single value, it may be a struct/union that we
5975 // can tile with integers.
5976 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
5977 unsigned BitSize = DL.getTypeSizeInBits(OpTy);
5986 OpTy = IntegerType::get(Context, BitSize);
5991 return TLI.getValueType(DL, OpTy, true);
5995 typedef SmallVector<SDISelAsmOperandInfo,16> SDISelAsmOperandInfoVector;
5997 } // end anonymous namespace
5999 /// GetRegistersForValue - Assign registers (virtual or physical) for the
6000 /// specified operand. We prefer to assign virtual registers, to allow the
6001 /// register allocator to handle the assignment process. However, if the asm
6002 /// uses features that we can't model on machineinstrs, we have SDISel do the
6003 /// allocation. This produces generally horrible, but correct, code.
6005 /// OpInfo describes the operand.
6007 static void GetRegistersForValue(SelectionDAG &DAG,
6008 const TargetLowering &TLI,
6010 SDISelAsmOperandInfo &OpInfo) {
6011 LLVMContext &Context = *DAG.getContext();
6013 MachineFunction &MF = DAG.getMachineFunction();
6014 SmallVector<unsigned, 4> Regs;
6016 // If this is a constraint for a single physreg, or a constraint for a
6017 // register class, find it.
6018 std::pair<unsigned, const TargetRegisterClass *> PhysReg =
6019 TLI.getRegForInlineAsmConstraint(MF.getSubtarget().getRegisterInfo(),
6020 OpInfo.ConstraintCode,
6021 OpInfo.ConstraintVT);
6023 unsigned NumRegs = 1;
6024 if (OpInfo.ConstraintVT != MVT::Other) {
6025 // If this is a FP input in an integer register (or visa versa) insert a bit
6026 // cast of the input value. More generally, handle any case where the input
6027 // value disagrees with the register class we plan to stick this in.
6028 if (OpInfo.Type == InlineAsm::isInput &&
6029 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) {
6030 // Try to convert to the first EVT that the reg class contains. If the
6031 // types are identical size, use a bitcast to convert (e.g. two differing
6033 MVT RegVT = *PhysReg.second->vt_begin();
6034 if (RegVT.getSizeInBits() == OpInfo.CallOperand.getValueSizeInBits()) {
6035 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
6036 RegVT, OpInfo.CallOperand);
6037 OpInfo.ConstraintVT = RegVT;
6038 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) {
6039 // If the input is a FP value and we want it in FP registers, do a
6040 // bitcast to the corresponding integer type. This turns an f64 value
6041 // into i64, which can be passed with two i32 values on a 32-bit
6043 RegVT = MVT::getIntegerVT(OpInfo.ConstraintVT.getSizeInBits());
6044 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
6045 RegVT, OpInfo.CallOperand);
6046 OpInfo.ConstraintVT = RegVT;
6050 NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT);
6054 EVT ValueVT = OpInfo.ConstraintVT;
6056 // If this is a constraint for a specific physical register, like {r17},
6058 if (unsigned AssignedReg = PhysReg.first) {
6059 const TargetRegisterClass *RC = PhysReg.second;
6060 if (OpInfo.ConstraintVT == MVT::Other)
6061 ValueVT = *RC->vt_begin();
6063 // Get the actual register value type. This is important, because the user
6064 // may have asked for (e.g.) the AX register in i32 type. We need to
6065 // remember that AX is actually i16 to get the right extension.
6066 RegVT = *RC->vt_begin();
6068 // This is a explicit reference to a physical register.
6069 Regs.push_back(AssignedReg);
6071 // If this is an expanded reference, add the rest of the regs to Regs.
6073 TargetRegisterClass::iterator I = RC->begin();
6074 for (; *I != AssignedReg; ++I)
6075 assert(I != RC->end() && "Didn't find reg!");
6077 // Already added the first reg.
6079 for (; NumRegs; --NumRegs, ++I) {
6080 assert(I != RC->end() && "Ran out of registers to allocate!");
6085 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
6089 // Otherwise, if this was a reference to an LLVM register class, create vregs
6090 // for this reference.
6091 if (const TargetRegisterClass *RC = PhysReg.second) {
6092 RegVT = *RC->vt_begin();
6093 if (OpInfo.ConstraintVT == MVT::Other)
6096 // Create the appropriate number of virtual registers.
6097 MachineRegisterInfo &RegInfo = MF.getRegInfo();
6098 for (; NumRegs; --NumRegs)
6099 Regs.push_back(RegInfo.createVirtualRegister(RC));
6101 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
6105 // Otherwise, we couldn't allocate enough registers for this.
6108 /// visitInlineAsm - Handle a call to an InlineAsm object.
6110 void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) {
6111 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
6113 /// ConstraintOperands - Information about all of the constraints.
6114 SDISelAsmOperandInfoVector ConstraintOperands;
6116 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6117 TargetLowering::AsmOperandInfoVector TargetConstraints = TLI.ParseConstraints(
6118 DAG.getDataLayout(), DAG.getSubtarget().getRegisterInfo(), CS);
6120 bool hasMemory = false;
6122 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
6123 unsigned ResNo = 0; // ResNo - The result number of the next output.
6124 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
6125 ConstraintOperands.push_back(SDISelAsmOperandInfo(TargetConstraints[i]));
6126 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
6128 MVT OpVT = MVT::Other;
6130 // Compute the value type for each operand.
6131 switch (OpInfo.Type) {
6132 case InlineAsm::isOutput:
6133 // Indirect outputs just consume an argument.
6134 if (OpInfo.isIndirect) {
6135 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
6139 // The return value of the call is this value. As such, there is no
6140 // corresponding argument.
6141 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
6142 if (StructType *STy = dyn_cast<StructType>(CS.getType())) {
6143 OpVT = TLI.getSimpleValueType(DAG.getDataLayout(),
6144 STy->getElementType(ResNo));
6146 assert(ResNo == 0 && "Asm only has one result!");
6147 OpVT = TLI.getSimpleValueType(DAG.getDataLayout(), CS.getType());
6151 case InlineAsm::isInput:
6152 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
6154 case InlineAsm::isClobber:
6159 // If this is an input or an indirect output, process the call argument.
6160 // BasicBlocks are labels, currently appearing only in asm's.
6161 if (OpInfo.CallOperandVal) {
6162 if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) {
6163 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]);
6165 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
6168 OpVT = OpInfo.getCallOperandValEVT(*DAG.getContext(), TLI,
6169 DAG.getDataLayout()).getSimpleVT();
6172 OpInfo.ConstraintVT = OpVT;
6174 // Indirect operand accesses access memory.
6175 if (OpInfo.isIndirect)
6178 for (unsigned j = 0, ee = OpInfo.Codes.size(); j != ee; ++j) {
6179 TargetLowering::ConstraintType
6180 CType = TLI.getConstraintType(OpInfo.Codes[j]);
6181 if (CType == TargetLowering::C_Memory) {
6189 SDValue Chain, Flag;
6191 // We won't need to flush pending loads if this asm doesn't touch
6192 // memory and is nonvolatile.
6193 if (hasMemory || IA->hasSideEffects())
6196 Chain = DAG.getRoot();
6198 // Second pass over the constraints: compute which constraint option to use
6199 // and assign registers to constraints that want a specific physreg.
6200 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6201 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6203 // If this is an output operand with a matching input operand, look up the
6204 // matching input. If their types mismatch, e.g. one is an integer, the
6205 // other is floating point, or their sizes are different, flag it as an
6207 if (OpInfo.hasMatchingInput()) {
6208 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
6210 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
6211 const TargetRegisterInfo *TRI = DAG.getSubtarget().getRegisterInfo();
6212 std::pair<unsigned, const TargetRegisterClass *> MatchRC =
6213 TLI.getRegForInlineAsmConstraint(TRI, OpInfo.ConstraintCode,
6214 OpInfo.ConstraintVT);
6215 std::pair<unsigned, const TargetRegisterClass *> InputRC =
6216 TLI.getRegForInlineAsmConstraint(TRI, Input.ConstraintCode,
6217 Input.ConstraintVT);
6218 if ((OpInfo.ConstraintVT.isInteger() !=
6219 Input.ConstraintVT.isInteger()) ||
6220 (MatchRC.second != InputRC.second)) {
6221 report_fatal_error("Unsupported asm: input constraint"
6222 " with a matching output constraint of"
6223 " incompatible type!");
6225 Input.ConstraintVT = OpInfo.ConstraintVT;
6229 // Compute the constraint code and ConstraintType to use.
6230 TLI.ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, &DAG);
6232 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
6233 OpInfo.Type == InlineAsm::isClobber)
6236 // If this is a memory input, and if the operand is not indirect, do what we
6237 // need to to provide an address for the memory input.
6238 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
6239 !OpInfo.isIndirect) {
6240 assert((OpInfo.isMultipleAlternative ||
6241 (OpInfo.Type == InlineAsm::isInput)) &&
6242 "Can only indirectify direct input operands!");
6244 // Memory operands really want the address of the value. If we don't have
6245 // an indirect input, put it in the constpool if we can, otherwise spill
6246 // it to a stack slot.
6247 // TODO: This isn't quite right. We need to handle these according to
6248 // the addressing mode that the constraint wants. Also, this may take
6249 // an additional register for the computation and we don't want that
6252 // If the operand is a float, integer, or vector constant, spill to a
6253 // constant pool entry to get its address.
6254 const Value *OpVal = OpInfo.CallOperandVal;
6255 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
6256 isa<ConstantVector>(OpVal) || isa<ConstantDataVector>(OpVal)) {
6257 OpInfo.CallOperand = DAG.getConstantPool(
6258 cast<Constant>(OpVal), TLI.getPointerTy(DAG.getDataLayout()));
6260 // Otherwise, create a stack slot and emit a store to it before the
6262 Type *Ty = OpVal->getType();
6263 auto &DL = DAG.getDataLayout();
6264 uint64_t TySize = DL.getTypeAllocSize(Ty);
6265 unsigned Align = DL.getPrefTypeAlignment(Ty);
6266 MachineFunction &MF = DAG.getMachineFunction();
6267 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
6269 DAG.getFrameIndex(SSFI, TLI.getPointerTy(DAG.getDataLayout()));
6270 Chain = DAG.getStore(
6271 Chain, getCurSDLoc(), OpInfo.CallOperand, StackSlot,
6272 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
6274 OpInfo.CallOperand = StackSlot;
6277 // There is no longer a Value* corresponding to this operand.
6278 OpInfo.CallOperandVal = nullptr;
6280 // It is now an indirect operand.
6281 OpInfo.isIndirect = true;
6284 // If this constraint is for a specific register, allocate it before
6286 if (OpInfo.ConstraintType == TargetLowering::C_Register)
6287 GetRegistersForValue(DAG, TLI, getCurSDLoc(), OpInfo);
6290 // Second pass - Loop over all of the operands, assigning virtual or physregs
6291 // to register class operands.
6292 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6293 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6295 // C_Register operands have already been allocated, Other/Memory don't need
6297 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
6298 GetRegistersForValue(DAG, TLI, getCurSDLoc(), OpInfo);
6301 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
6302 std::vector<SDValue> AsmNodeOperands;
6303 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain
6304 AsmNodeOperands.push_back(DAG.getTargetExternalSymbol(
6305 IA->getAsmString().c_str(), TLI.getPointerTy(DAG.getDataLayout())));
6307 // If we have a !srcloc metadata node associated with it, we want to attach
6308 // this to the ultimately generated inline asm machineinstr. To do this, we
6309 // pass in the third operand as this (potentially null) inline asm MDNode.
6310 const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc");
6311 AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc));
6313 // Remember the HasSideEffect, AlignStack, AsmDialect, MayLoad and MayStore
6314 // bits as operand 3.
6315 unsigned ExtraInfo = 0;
6316 if (IA->hasSideEffects())
6317 ExtraInfo |= InlineAsm::Extra_HasSideEffects;
6318 if (IA->isAlignStack())
6319 ExtraInfo |= InlineAsm::Extra_IsAlignStack;
6320 // Set the asm dialect.
6321 ExtraInfo |= IA->getDialect() * InlineAsm::Extra_AsmDialect;
6323 // Determine if this InlineAsm MayLoad or MayStore based on the constraints.
6324 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
6325 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
6327 // Compute the constraint code and ConstraintType to use.
6328 TLI.ComputeConstraintToUse(OpInfo, SDValue());
6330 // Ideally, we would only check against memory constraints. However, the
6331 // meaning of an other constraint can be target-specific and we can't easily
6332 // reason about it. Therefore, be conservative and set MayLoad/MayStore
6333 // for other constriants as well.
6334 if (OpInfo.ConstraintType == TargetLowering::C_Memory ||
6335 OpInfo.ConstraintType == TargetLowering::C_Other) {
6336 if (OpInfo.Type == InlineAsm::isInput)
6337 ExtraInfo |= InlineAsm::Extra_MayLoad;
6338 else if (OpInfo.Type == InlineAsm::isOutput)
6339 ExtraInfo |= InlineAsm::Extra_MayStore;
6340 else if (OpInfo.Type == InlineAsm::isClobber)
6341 ExtraInfo |= (InlineAsm::Extra_MayLoad | InlineAsm::Extra_MayStore);
6345 AsmNodeOperands.push_back(DAG.getTargetConstant(
6346 ExtraInfo, getCurSDLoc(), TLI.getPointerTy(DAG.getDataLayout())));
6348 // Loop over all of the inputs, copying the operand values into the
6349 // appropriate registers and processing the output regs.
6350 RegsForValue RetValRegs;
6352 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
6353 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
6355 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6356 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6358 switch (OpInfo.Type) {
6359 case InlineAsm::isOutput: {
6360 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
6361 OpInfo.ConstraintType != TargetLowering::C_Register) {
6362 // Memory output, or 'other' output (e.g. 'X' constraint).
6363 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
6365 unsigned ConstraintID =
6366 TLI.getInlineAsmMemConstraint(OpInfo.ConstraintCode);
6367 assert(ConstraintID != InlineAsm::Constraint_Unknown &&
6368 "Failed to convert memory constraint code to constraint id.");
6370 // Add information to the INLINEASM node to know about this output.
6371 unsigned OpFlags = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
6372 OpFlags = InlineAsm::getFlagWordForMem(OpFlags, ConstraintID);
6373 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlags, getCurSDLoc(),
6375 AsmNodeOperands.push_back(OpInfo.CallOperand);
6379 // Otherwise, this is a register or register class output.
6381 // Copy the output from the appropriate register. Find a register that
6383 if (OpInfo.AssignedRegs.Regs.empty()) {
6384 LLVMContext &Ctx = *DAG.getContext();
6385 Ctx.emitError(CS.getInstruction(),
6386 "couldn't allocate output register for constraint '" +
6387 Twine(OpInfo.ConstraintCode) + "'");
6391 // If this is an indirect operand, store through the pointer after the
6393 if (OpInfo.isIndirect) {
6394 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
6395 OpInfo.CallOperandVal));
6397 // This is the result value of the call.
6398 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
6399 // Concatenate this output onto the outputs list.
6400 RetValRegs.append(OpInfo.AssignedRegs);
6403 // Add information to the INLINEASM node to know that this register is
6406 .AddInlineAsmOperands(OpInfo.isEarlyClobber
6407 ? InlineAsm::Kind_RegDefEarlyClobber
6408 : InlineAsm::Kind_RegDef,
6409 false, 0, getCurSDLoc(), DAG, AsmNodeOperands);
6412 case InlineAsm::isInput: {
6413 SDValue InOperandVal = OpInfo.CallOperand;
6415 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint?
6416 // If this is required to match an output register we have already set,
6417 // just use its register.
6418 unsigned OperandNo = OpInfo.getMatchedOperand();
6420 // Scan until we find the definition we already emitted of this operand.
6421 // When we find it, create a RegsForValue operand.
6422 unsigned CurOp = InlineAsm::Op_FirstOperand;
6423 for (; OperandNo; --OperandNo) {
6424 // Advance to the next operand.
6426 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6427 assert((InlineAsm::isRegDefKind(OpFlag) ||
6428 InlineAsm::isRegDefEarlyClobberKind(OpFlag) ||
6429 InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?");
6430 CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1;
6434 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6435 if (InlineAsm::isRegDefKind(OpFlag) ||
6436 InlineAsm::isRegDefEarlyClobberKind(OpFlag)) {
6437 // Add (OpFlag&0xffff)>>3 registers to MatchedRegs.
6438 if (OpInfo.isIndirect) {
6439 // This happens on gcc/testsuite/gcc.dg/pr8788-1.c
6440 LLVMContext &Ctx = *DAG.getContext();
6441 Ctx.emitError(CS.getInstruction(), "inline asm not supported yet:"
6442 " don't know how to handle tied "
6443 "indirect register inputs");
6447 RegsForValue MatchedRegs;
6448 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
6449 MVT RegVT = AsmNodeOperands[CurOp+1].getSimpleValueType();
6450 MatchedRegs.RegVTs.push_back(RegVT);
6451 MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo();
6452 for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag);
6454 if (const TargetRegisterClass *RC = TLI.getRegClassFor(RegVT))
6455 MatchedRegs.Regs.push_back(RegInfo.createVirtualRegister(RC));
6457 LLVMContext &Ctx = *DAG.getContext();
6458 Ctx.emitError(CS.getInstruction(),
6459 "inline asm error: This value"
6460 " type register class is not natively supported!");
6464 SDLoc dl = getCurSDLoc();
6465 // Use the produced MatchedRegs object to
6466 MatchedRegs.getCopyToRegs(InOperandVal, DAG, dl,
6467 Chain, &Flag, CS.getInstruction());
6468 MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse,
6469 true, OpInfo.getMatchedOperand(), dl,
6470 DAG, AsmNodeOperands);
6474 assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!");
6475 assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 &&
6476 "Unexpected number of operands");
6477 // Add information to the INLINEASM node to know about this input.
6478 // See InlineAsm.h isUseOperandTiedToDef.
6479 OpFlag = InlineAsm::convertMemFlagWordToMatchingFlagWord(OpFlag);
6480 OpFlag = InlineAsm::getFlagWordForMatchingOp(OpFlag,
6481 OpInfo.getMatchedOperand());
6482 AsmNodeOperands.push_back(DAG.getTargetConstant(
6483 OpFlag, getCurSDLoc(), TLI.getPointerTy(DAG.getDataLayout())));
6484 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]);
6488 // Treat indirect 'X' constraint as memory.
6489 if (OpInfo.ConstraintType == TargetLowering::C_Other &&
6491 OpInfo.ConstraintType = TargetLowering::C_Memory;
6493 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
6494 std::vector<SDValue> Ops;
6495 TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode,
6498 LLVMContext &Ctx = *DAG.getContext();
6499 Ctx.emitError(CS.getInstruction(),
6500 "invalid operand for inline asm constraint '" +
6501 Twine(OpInfo.ConstraintCode) + "'");
6505 // Add information to the INLINEASM node to know about this input.
6506 unsigned ResOpType =
6507 InlineAsm::getFlagWord(InlineAsm::Kind_Imm, Ops.size());
6508 AsmNodeOperands.push_back(DAG.getTargetConstant(
6509 ResOpType, getCurSDLoc(), TLI.getPointerTy(DAG.getDataLayout())));
6510 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
6514 if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
6515 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
6516 assert(InOperandVal.getValueType() ==
6517 TLI.getPointerTy(DAG.getDataLayout()) &&
6518 "Memory operands expect pointer values");
6520 unsigned ConstraintID =
6521 TLI.getInlineAsmMemConstraint(OpInfo.ConstraintCode);
6522 assert(ConstraintID != InlineAsm::Constraint_Unknown &&
6523 "Failed to convert memory constraint code to constraint id.");
6525 // Add information to the INLINEASM node to know about this input.
6526 unsigned ResOpType = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
6527 ResOpType = InlineAsm::getFlagWordForMem(ResOpType, ConstraintID);
6528 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
6531 AsmNodeOperands.push_back(InOperandVal);
6535 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
6536 OpInfo.ConstraintType == TargetLowering::C_Register) &&
6537 "Unknown constraint type!");
6539 // TODO: Support this.
6540 if (OpInfo.isIndirect) {
6541 LLVMContext &Ctx = *DAG.getContext();
6542 Ctx.emitError(CS.getInstruction(),
6543 "Don't know how to handle indirect register inputs yet "
6544 "for constraint '" +
6545 Twine(OpInfo.ConstraintCode) + "'");
6549 // Copy the input into the appropriate registers.
6550 if (OpInfo.AssignedRegs.Regs.empty()) {
6551 LLVMContext &Ctx = *DAG.getContext();
6552 Ctx.emitError(CS.getInstruction(),
6553 "couldn't allocate input reg for constraint '" +
6554 Twine(OpInfo.ConstraintCode) + "'");
6558 SDLoc dl = getCurSDLoc();
6560 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, dl,
6561 Chain, &Flag, CS.getInstruction());
6563 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, false, 0,
6564 dl, DAG, AsmNodeOperands);
6567 case InlineAsm::isClobber: {
6568 // Add the clobbered value to the operand list, so that the register
6569 // allocator is aware that the physreg got clobbered.
6570 if (!OpInfo.AssignedRegs.Regs.empty())
6571 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_Clobber,
6572 false, 0, getCurSDLoc(), DAG,
6579 // Finish up input operands. Set the input chain and add the flag last.
6580 AsmNodeOperands[InlineAsm::Op_InputChain] = Chain;
6581 if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
6583 Chain = DAG.getNode(ISD::INLINEASM, getCurSDLoc(),
6584 DAG.getVTList(MVT::Other, MVT::Glue), AsmNodeOperands);
6585 Flag = Chain.getValue(1);
6587 // If this asm returns a register value, copy the result from that register
6588 // and set it as the value of the call.
6589 if (!RetValRegs.Regs.empty()) {
6590 SDValue Val = RetValRegs.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(),
6591 Chain, &Flag, CS.getInstruction());
6593 // FIXME: Why don't we do this for inline asms with MRVs?
6594 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) {
6595 EVT ResultType = TLI.getValueType(DAG.getDataLayout(), CS.getType());
6597 // If any of the results of the inline asm is a vector, it may have the
6598 // wrong width/num elts. This can happen for register classes that can
6599 // contain multiple different value types. The preg or vreg allocated may
6600 // not have the same VT as was expected. Convert it to the right type
6601 // with bit_convert.
6602 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) {
6603 Val = DAG.getNode(ISD::BITCAST, getCurSDLoc(),
6606 } else if (ResultType != Val.getValueType() &&
6607 ResultType.isInteger() && Val.getValueType().isInteger()) {
6608 // If a result value was tied to an input value, the computed result may
6609 // have a wider width than the expected result. Extract the relevant
6611 Val = DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), ResultType, Val);
6614 assert(ResultType == Val.getValueType() && "Asm result value mismatch!");
6617 setValue(CS.getInstruction(), Val);
6618 // Don't need to use this as a chain in this case.
6619 if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty())
6623 std::vector<std::pair<SDValue, const Value *> > StoresToEmit;
6625 // Process indirect outputs, first output all of the flagged copies out of
6627 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
6628 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
6629 const Value *Ptr = IndirectStoresToEmit[i].second;
6630 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(),
6632 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
6635 // Emit the non-flagged stores from the physregs.
6636 SmallVector<SDValue, 8> OutChains;
6637 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) {
6638 SDValue Val = DAG.getStore(Chain, getCurSDLoc(),
6639 StoresToEmit[i].first,
6640 getValue(StoresToEmit[i].second),
6641 MachinePointerInfo(StoresToEmit[i].second),
6643 OutChains.push_back(Val);
6646 if (!OutChains.empty())
6647 Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other, OutChains);
6652 void SelectionDAGBuilder::visitVAStart(const CallInst &I) {
6653 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurSDLoc(),
6654 MVT::Other, getRoot(),
6655 getValue(I.getArgOperand(0)),
6656 DAG.getSrcValue(I.getArgOperand(0))));
6659 void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) {
6660 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6661 const DataLayout &DL = DAG.getDataLayout();
6662 SDValue V = DAG.getVAArg(TLI.getValueType(DAG.getDataLayout(), I.getType()),
6663 getCurSDLoc(), getRoot(), getValue(I.getOperand(0)),
6664 DAG.getSrcValue(I.getOperand(0)),
6665 DL.getABITypeAlignment(I.getType()));
6667 DAG.setRoot(V.getValue(1));
6670 void SelectionDAGBuilder::visitVAEnd(const CallInst &I) {
6671 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurSDLoc(),
6672 MVT::Other, getRoot(),
6673 getValue(I.getArgOperand(0)),
6674 DAG.getSrcValue(I.getArgOperand(0))));
6677 void SelectionDAGBuilder::visitVACopy(const CallInst &I) {
6678 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurSDLoc(),
6679 MVT::Other, getRoot(),
6680 getValue(I.getArgOperand(0)),
6681 getValue(I.getArgOperand(1)),
6682 DAG.getSrcValue(I.getArgOperand(0)),
6683 DAG.getSrcValue(I.getArgOperand(1))));
6686 /// \brief Lower an argument list according to the target calling convention.
6688 /// \return A tuple of <return-value, token-chain>
6690 /// This is a helper for lowering intrinsics that follow a target calling
6691 /// convention or require stack pointer adjustment. Only a subset of the
6692 /// intrinsic's operands need to participate in the calling convention.
6693 std::pair<SDValue, SDValue> SelectionDAGBuilder::lowerCallOperands(
6694 ImmutableCallSite CS, unsigned ArgIdx, unsigned NumArgs, SDValue Callee,
6695 Type *ReturnTy, const BasicBlock *EHPadBB, bool IsPatchPoint) {
6696 TargetLowering::ArgListTy Args;
6697 Args.reserve(NumArgs);
6699 // Populate the argument list.
6700 // Attributes for args start at offset 1, after the return attribute.
6701 for (unsigned ArgI = ArgIdx, ArgE = ArgIdx + NumArgs, AttrI = ArgIdx + 1;
6702 ArgI != ArgE; ++ArgI) {
6703 const Value *V = CS->getOperand(ArgI);
6705 assert(!V->getType()->isEmptyTy() && "Empty type passed to intrinsic.");
6707 TargetLowering::ArgListEntry Entry;
6708 Entry.Node = getValue(V);
6709 Entry.Ty = V->getType();
6710 Entry.setAttributes(&CS, AttrI);
6711 Args.push_back(Entry);
6714 TargetLowering::CallLoweringInfo CLI(DAG);
6715 CLI.setDebugLoc(getCurSDLoc()).setChain(getRoot())
6716 .setCallee(CS.getCallingConv(), ReturnTy, Callee, std::move(Args), NumArgs)
6717 .setDiscardResult(CS->use_empty()).setIsPatchPoint(IsPatchPoint);
6719 return lowerInvokable(CLI, EHPadBB);
6722 /// \brief Add a stack map intrinsic call's live variable operands to a stackmap
6723 /// or patchpoint target node's operand list.
6725 /// Constants are converted to TargetConstants purely as an optimization to
6726 /// avoid constant materialization and register allocation.
6728 /// FrameIndex operands are converted to TargetFrameIndex so that ISEL does not
6729 /// generate addess computation nodes, and so ExpandISelPseudo can convert the
6730 /// TargetFrameIndex into a DirectMemRefOp StackMap location. This avoids
6731 /// address materialization and register allocation, but may also be required
6732 /// for correctness. If a StackMap (or PatchPoint) intrinsic directly uses an
6733 /// alloca in the entry block, then the runtime may assume that the alloca's
6734 /// StackMap location can be read immediately after compilation and that the
6735 /// location is valid at any point during execution (this is similar to the
6736 /// assumption made by the llvm.gcroot intrinsic). If the alloca's location were
6737 /// only available in a register, then the runtime would need to trap when
6738 /// execution reaches the StackMap in order to read the alloca's location.
6739 static void addStackMapLiveVars(ImmutableCallSite CS, unsigned StartIdx,
6740 SDLoc DL, SmallVectorImpl<SDValue> &Ops,
6741 SelectionDAGBuilder &Builder) {
6742 for (unsigned i = StartIdx, e = CS.arg_size(); i != e; ++i) {
6743 SDValue OpVal = Builder.getValue(CS.getArgument(i));
6744 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(OpVal)) {
6746 Builder.DAG.getTargetConstant(StackMaps::ConstantOp, DL, MVT::i64));
6748 Builder.DAG.getTargetConstant(C->getSExtValue(), DL, MVT::i64));
6749 } else if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(OpVal)) {
6750 const TargetLowering &TLI = Builder.DAG.getTargetLoweringInfo();
6751 Ops.push_back(Builder.DAG.getTargetFrameIndex(
6752 FI->getIndex(), TLI.getPointerTy(Builder.DAG.getDataLayout())));
6754 Ops.push_back(OpVal);
6758 /// \brief Lower llvm.experimental.stackmap directly to its target opcode.
6759 void SelectionDAGBuilder::visitStackmap(const CallInst &CI) {
6760 // void @llvm.experimental.stackmap(i32 <id>, i32 <numShadowBytes>,
6761 // [live variables...])
6763 assert(CI.getType()->isVoidTy() && "Stackmap cannot return a value.");
6765 SDValue Chain, InFlag, Callee, NullPtr;
6766 SmallVector<SDValue, 32> Ops;
6768 SDLoc DL = getCurSDLoc();
6769 Callee = getValue(CI.getCalledValue());
6770 NullPtr = DAG.getIntPtrConstant(0, DL, true);
6772 // The stackmap intrinsic only records the live variables (the arguemnts
6773 // passed to it) and emits NOPS (if requested). Unlike the patchpoint
6774 // intrinsic, this won't be lowered to a function call. This means we don't
6775 // have to worry about calling conventions and target specific lowering code.
6776 // Instead we perform the call lowering right here.
6778 // chain, flag = CALLSEQ_START(chain, 0)
6779 // chain, flag = STACKMAP(id, nbytes, ..., chain, flag)
6780 // chain, flag = CALLSEQ_END(chain, 0, 0, flag)
6782 Chain = DAG.getCALLSEQ_START(getRoot(), NullPtr, DL);
6783 InFlag = Chain.getValue(1);
6785 // Add the <id> and <numBytes> constants.
6786 SDValue IDVal = getValue(CI.getOperand(PatchPointOpers::IDPos));
6787 Ops.push_back(DAG.getTargetConstant(
6788 cast<ConstantSDNode>(IDVal)->getZExtValue(), DL, MVT::i64));
6789 SDValue NBytesVal = getValue(CI.getOperand(PatchPointOpers::NBytesPos));
6790 Ops.push_back(DAG.getTargetConstant(
6791 cast<ConstantSDNode>(NBytesVal)->getZExtValue(), DL,
6794 // Push live variables for the stack map.
6795 addStackMapLiveVars(&CI, 2, DL, Ops, *this);
6797 // We are not pushing any register mask info here on the operands list,
6798 // because the stackmap doesn't clobber anything.
6800 // Push the chain and the glue flag.
6801 Ops.push_back(Chain);
6802 Ops.push_back(InFlag);
6804 // Create the STACKMAP node.
6805 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6806 SDNode *SM = DAG.getMachineNode(TargetOpcode::STACKMAP, DL, NodeTys, Ops);
6807 Chain = SDValue(SM, 0);
6808 InFlag = Chain.getValue(1);
6810 Chain = DAG.getCALLSEQ_END(Chain, NullPtr, NullPtr, InFlag, DL);
6812 // Stackmaps don't generate values, so nothing goes into the NodeMap.
6814 // Set the root to the target-lowered call chain.
6817 // Inform the Frame Information that we have a stackmap in this function.
6818 FuncInfo.MF->getFrameInfo()->setHasStackMap();
6821 /// \brief Lower llvm.experimental.patchpoint directly to its target opcode.
6822 void SelectionDAGBuilder::visitPatchpoint(ImmutableCallSite CS,
6823 const BasicBlock *EHPadBB) {
6824 // void|i64 @llvm.experimental.patchpoint.void|i64(i64 <id>,
6829 // [live variables...])
6831 CallingConv::ID CC = CS.getCallingConv();
6832 bool IsAnyRegCC = CC == CallingConv::AnyReg;
6833 bool HasDef = !CS->getType()->isVoidTy();
6834 SDLoc dl = getCurSDLoc();
6835 SDValue Callee = getValue(CS->getOperand(PatchPointOpers::TargetPos));
6837 // Handle immediate and symbolic callees.
6838 if (auto* ConstCallee = dyn_cast<ConstantSDNode>(Callee))
6839 Callee = DAG.getIntPtrConstant(ConstCallee->getZExtValue(), dl,
6841 else if (auto* SymbolicCallee = dyn_cast<GlobalAddressSDNode>(Callee))
6842 Callee = DAG.getTargetGlobalAddress(SymbolicCallee->getGlobal(),
6843 SDLoc(SymbolicCallee),
6844 SymbolicCallee->getValueType(0));
6846 // Get the real number of arguments participating in the call <numArgs>
6847 SDValue NArgVal = getValue(CS.getArgument(PatchPointOpers::NArgPos));
6848 unsigned NumArgs = cast<ConstantSDNode>(NArgVal)->getZExtValue();
6850 // Skip the four meta args: <id>, <numNopBytes>, <target>, <numArgs>
6851 // Intrinsics include all meta-operands up to but not including CC.
6852 unsigned NumMetaOpers = PatchPointOpers::CCPos;
6853 assert(CS.arg_size() >= NumMetaOpers + NumArgs &&
6854 "Not enough arguments provided to the patchpoint intrinsic");
6856 // For AnyRegCC the arguments are lowered later on manually.
6857 unsigned NumCallArgs = IsAnyRegCC ? 0 : NumArgs;
6859 IsAnyRegCC ? Type::getVoidTy(*DAG.getContext()) : CS->getType();
6860 std::pair<SDValue, SDValue> Result = lowerCallOperands(
6861 CS, NumMetaOpers, NumCallArgs, Callee, ReturnTy, EHPadBB, true);
6863 SDNode *CallEnd = Result.second.getNode();
6864 if (HasDef && (CallEnd->getOpcode() == ISD::CopyFromReg))
6865 CallEnd = CallEnd->getOperand(0).getNode();
6867 /// Get a call instruction from the call sequence chain.
6868 /// Tail calls are not allowed.
6869 assert(CallEnd->getOpcode() == ISD::CALLSEQ_END &&
6870 "Expected a callseq node.");
6871 SDNode *Call = CallEnd->getOperand(0).getNode();
6872 bool HasGlue = Call->getGluedNode();
6874 // Replace the target specific call node with the patchable intrinsic.
6875 SmallVector<SDValue, 8> Ops;
6877 // Add the <id> and <numBytes> constants.
6878 SDValue IDVal = getValue(CS->getOperand(PatchPointOpers::IDPos));
6879 Ops.push_back(DAG.getTargetConstant(
6880 cast<ConstantSDNode>(IDVal)->getZExtValue(), dl, MVT::i64));
6881 SDValue NBytesVal = getValue(CS->getOperand(PatchPointOpers::NBytesPos));
6882 Ops.push_back(DAG.getTargetConstant(
6883 cast<ConstantSDNode>(NBytesVal)->getZExtValue(), dl,
6887 Ops.push_back(Callee);
6889 // Adjust <numArgs> to account for any arguments that have been passed on the
6891 // Call Node: Chain, Target, {Args}, RegMask, [Glue]
6892 unsigned NumCallRegArgs = Call->getNumOperands() - (HasGlue ? 4 : 3);
6893 NumCallRegArgs = IsAnyRegCC ? NumArgs : NumCallRegArgs;
6894 Ops.push_back(DAG.getTargetConstant(NumCallRegArgs, dl, MVT::i32));
6896 // Add the calling convention
6897 Ops.push_back(DAG.getTargetConstant((unsigned)CC, dl, MVT::i32));
6899 // Add the arguments we omitted previously. The register allocator should
6900 // place these in any free register.
6902 for (unsigned i = NumMetaOpers, e = NumMetaOpers + NumArgs; i != e; ++i)
6903 Ops.push_back(getValue(CS.getArgument(i)));
6905 // Push the arguments from the call instruction up to the register mask.
6906 SDNode::op_iterator e = HasGlue ? Call->op_end()-2 : Call->op_end()-1;
6907 Ops.append(Call->op_begin() + 2, e);
6909 // Push live variables for the stack map.
6910 addStackMapLiveVars(CS, NumMetaOpers + NumArgs, dl, Ops, *this);
6912 // Push the register mask info.
6914 Ops.push_back(*(Call->op_end()-2));
6916 Ops.push_back(*(Call->op_end()-1));
6918 // Push the chain (this is originally the first operand of the call, but
6919 // becomes now the last or second to last operand).
6920 Ops.push_back(*(Call->op_begin()));
6922 // Push the glue flag (last operand).
6924 Ops.push_back(*(Call->op_end()-1));
6927 if (IsAnyRegCC && HasDef) {
6928 // Create the return types based on the intrinsic definition
6929 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6930 SmallVector<EVT, 3> ValueVTs;
6931 ComputeValueVTs(TLI, DAG.getDataLayout(), CS->getType(), ValueVTs);
6932 assert(ValueVTs.size() == 1 && "Expected only one return value type.");
6934 // There is always a chain and a glue type at the end
6935 ValueVTs.push_back(MVT::Other);
6936 ValueVTs.push_back(MVT::Glue);
6937 NodeTys = DAG.getVTList(ValueVTs);
6939 NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6941 // Replace the target specific call node with a PATCHPOINT node.
6942 MachineSDNode *MN = DAG.getMachineNode(TargetOpcode::PATCHPOINT,
6945 // Update the NodeMap.
6948 setValue(CS.getInstruction(), SDValue(MN, 0));
6950 setValue(CS.getInstruction(), Result.first);
6953 // Fixup the consumers of the intrinsic. The chain and glue may be used in the
6954 // call sequence. Furthermore the location of the chain and glue can change
6955 // when the AnyReg calling convention is used and the intrinsic returns a
6957 if (IsAnyRegCC && HasDef) {
6958 SDValue From[] = {SDValue(Call, 0), SDValue(Call, 1)};
6959 SDValue To[] = {SDValue(MN, 1), SDValue(MN, 2)};
6960 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
6962 DAG.ReplaceAllUsesWith(Call, MN);
6963 DAG.DeleteNode(Call);
6965 // Inform the Frame Information that we have a patchpoint in this function.
6966 FuncInfo.MF->getFrameInfo()->setHasPatchPoint();
6969 /// Returns an AttributeSet representing the attributes applied to the return
6970 /// value of the given call.
6971 static AttributeSet getReturnAttrs(TargetLowering::CallLoweringInfo &CLI) {
6972 SmallVector<Attribute::AttrKind, 2> Attrs;
6974 Attrs.push_back(Attribute::SExt);
6976 Attrs.push_back(Attribute::ZExt);
6978 Attrs.push_back(Attribute::InReg);
6980 return AttributeSet::get(CLI.RetTy->getContext(), AttributeSet::ReturnIndex,
6984 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
6985 /// implementation, which just calls LowerCall.
6986 /// FIXME: When all targets are
6987 /// migrated to using LowerCall, this hook should be integrated into SDISel.
6988 std::pair<SDValue, SDValue>
6989 TargetLowering::LowerCallTo(TargetLowering::CallLoweringInfo &CLI) const {
6990 // Handle the incoming return values from the call.
6992 Type *OrigRetTy = CLI.RetTy;
6993 SmallVector<EVT, 4> RetTys;
6994 SmallVector<uint64_t, 4> Offsets;
6995 auto &DL = CLI.DAG.getDataLayout();
6996 ComputeValueVTs(*this, DL, CLI.RetTy, RetTys, &Offsets);
6998 SmallVector<ISD::OutputArg, 4> Outs;
6999 GetReturnInfo(CLI.RetTy, getReturnAttrs(CLI), Outs, *this, DL);
7001 bool CanLowerReturn =
7002 this->CanLowerReturn(CLI.CallConv, CLI.DAG.getMachineFunction(),
7003 CLI.IsVarArg, Outs, CLI.RetTy->getContext());
7005 SDValue DemoteStackSlot;
7006 int DemoteStackIdx = -100;
7007 if (!CanLowerReturn) {
7008 // FIXME: equivalent assert?
7009 // assert(!CS.hasInAllocaArgument() &&
7010 // "sret demotion is incompatible with inalloca");
7011 uint64_t TySize = DL.getTypeAllocSize(CLI.RetTy);
7012 unsigned Align = DL.getPrefTypeAlignment(CLI.RetTy);
7013 MachineFunction &MF = CLI.DAG.getMachineFunction();
7014 DemoteStackIdx = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
7015 Type *StackSlotPtrType = PointerType::getUnqual(CLI.RetTy);
7017 DemoteStackSlot = CLI.DAG.getFrameIndex(DemoteStackIdx, getPointerTy(DL));
7019 Entry.Node = DemoteStackSlot;
7020 Entry.Ty = StackSlotPtrType;
7021 Entry.isSExt = false;
7022 Entry.isZExt = false;
7023 Entry.isInReg = false;
7024 Entry.isSRet = true;
7025 Entry.isNest = false;
7026 Entry.isByVal = false;
7027 Entry.isReturned = false;
7028 Entry.Alignment = Align;
7029 CLI.getArgs().insert(CLI.getArgs().begin(), Entry);
7030 CLI.RetTy = Type::getVoidTy(CLI.RetTy->getContext());
7032 // sret demotion isn't compatible with tail-calls, since the sret argument
7033 // points into the callers stack frame.
7034 CLI.IsTailCall = false;
7036 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
7038 MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
7039 unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
7040 for (unsigned i = 0; i != NumRegs; ++i) {
7041 ISD::InputArg MyFlags;
7042 MyFlags.VT = RegisterVT;
7044 MyFlags.Used = CLI.IsReturnValueUsed;
7046 MyFlags.Flags.setSExt();
7048 MyFlags.Flags.setZExt();
7050 MyFlags.Flags.setInReg();
7051 CLI.Ins.push_back(MyFlags);
7056 // Handle all of the outgoing arguments.
7058 CLI.OutVals.clear();
7059 ArgListTy &Args = CLI.getArgs();
7060 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
7061 SmallVector<EVT, 4> ValueVTs;
7062 ComputeValueVTs(*this, DL, Args[i].Ty, ValueVTs);
7063 Type *FinalType = Args[i].Ty;
7064 if (Args[i].isByVal)
7065 FinalType = cast<PointerType>(Args[i].Ty)->getElementType();
7066 bool NeedsRegBlock = functionArgumentNeedsConsecutiveRegisters(
7067 FinalType, CLI.CallConv, CLI.IsVarArg);
7068 for (unsigned Value = 0, NumValues = ValueVTs.size(); Value != NumValues;
7070 EVT VT = ValueVTs[Value];
7071 Type *ArgTy = VT.getTypeForEVT(CLI.RetTy->getContext());
7072 SDValue Op = SDValue(Args[i].Node.getNode(),
7073 Args[i].Node.getResNo() + Value);
7074 ISD::ArgFlagsTy Flags;
7075 unsigned OriginalAlignment = DL.getABITypeAlignment(ArgTy);
7081 if (Args[i].isInReg)
7085 if (Args[i].isByVal)
7087 if (Args[i].isInAlloca) {
7088 Flags.setInAlloca();
7089 // Set the byval flag for CCAssignFn callbacks that don't know about
7090 // inalloca. This way we can know how many bytes we should've allocated
7091 // and how many bytes a callee cleanup function will pop. If we port
7092 // inalloca to more targets, we'll have to add custom inalloca handling
7093 // in the various CC lowering callbacks.
7096 if (Args[i].isByVal || Args[i].isInAlloca) {
7097 PointerType *Ty = cast<PointerType>(Args[i].Ty);
7098 Type *ElementTy = Ty->getElementType();
7099 Flags.setByValSize(DL.getTypeAllocSize(ElementTy));
7100 // For ByVal, alignment should come from FE. BE will guess if this
7101 // info is not there but there are cases it cannot get right.
7102 unsigned FrameAlign;
7103 if (Args[i].Alignment)
7104 FrameAlign = Args[i].Alignment;
7106 FrameAlign = getByValTypeAlignment(ElementTy, DL);
7107 Flags.setByValAlign(FrameAlign);
7112 Flags.setInConsecutiveRegs();
7113 Flags.setOrigAlign(OriginalAlignment);
7115 MVT PartVT = getRegisterType(CLI.RetTy->getContext(), VT);
7116 unsigned NumParts = getNumRegisters(CLI.RetTy->getContext(), VT);
7117 SmallVector<SDValue, 4> Parts(NumParts);
7118 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
7121 ExtendKind = ISD::SIGN_EXTEND;
7122 else if (Args[i].isZExt)
7123 ExtendKind = ISD::ZERO_EXTEND;
7125 // Conservatively only handle 'returned' on non-vectors for now
7126 if (Args[i].isReturned && !Op.getValueType().isVector()) {
7127 assert(CLI.RetTy == Args[i].Ty && RetTys.size() == NumValues &&
7128 "unexpected use of 'returned'");
7129 // Before passing 'returned' to the target lowering code, ensure that
7130 // either the register MVT and the actual EVT are the same size or that
7131 // the return value and argument are extended in the same way; in these
7132 // cases it's safe to pass the argument register value unchanged as the
7133 // return register value (although it's at the target's option whether
7135 // TODO: allow code generation to take advantage of partially preserved
7136 // registers rather than clobbering the entire register when the
7137 // parameter extension method is not compatible with the return
7139 if ((NumParts * PartVT.getSizeInBits() == VT.getSizeInBits()) ||
7140 (ExtendKind != ISD::ANY_EXTEND &&
7141 CLI.RetSExt == Args[i].isSExt && CLI.RetZExt == Args[i].isZExt))
7142 Flags.setReturned();
7145 getCopyToParts(CLI.DAG, CLI.DL, Op, &Parts[0], NumParts, PartVT,
7146 CLI.CS ? CLI.CS->getInstruction() : nullptr, ExtendKind);
7148 for (unsigned j = 0; j != NumParts; ++j) {
7149 // if it isn't first piece, alignment must be 1
7150 ISD::OutputArg MyFlags(Flags, Parts[j].getValueType(), VT,
7151 i < CLI.NumFixedArgs,
7152 i, j*Parts[j].getValueType().getStoreSize());
7153 if (NumParts > 1 && j == 0)
7154 MyFlags.Flags.setSplit();
7156 MyFlags.Flags.setOrigAlign(1);
7158 CLI.Outs.push_back(MyFlags);
7159 CLI.OutVals.push_back(Parts[j]);
7162 if (NeedsRegBlock && Value == NumValues - 1)
7163 CLI.Outs[CLI.Outs.size() - 1].Flags.setInConsecutiveRegsLast();
7167 SmallVector<SDValue, 4> InVals;
7168 CLI.Chain = LowerCall(CLI, InVals);
7170 // Verify that the target's LowerCall behaved as expected.
7171 assert(CLI.Chain.getNode() && CLI.Chain.getValueType() == MVT::Other &&
7172 "LowerCall didn't return a valid chain!");
7173 assert((!CLI.IsTailCall || InVals.empty()) &&
7174 "LowerCall emitted a return value for a tail call!");
7175 assert((CLI.IsTailCall || InVals.size() == CLI.Ins.size()) &&
7176 "LowerCall didn't emit the correct number of values!");
7178 // For a tail call, the return value is merely live-out and there aren't
7179 // any nodes in the DAG representing it. Return a special value to
7180 // indicate that a tail call has been emitted and no more Instructions
7181 // should be processed in the current block.
7182 if (CLI.IsTailCall) {
7183 CLI.DAG.setRoot(CLI.Chain);
7184 return std::make_pair(SDValue(), SDValue());
7187 DEBUG(for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++i) {
7188 assert(InVals[i].getNode() &&
7189 "LowerCall emitted a null value!");
7190 assert(EVT(CLI.Ins[i].VT) == InVals[i].getValueType() &&
7191 "LowerCall emitted a value with the wrong type!");
7194 SmallVector<SDValue, 4> ReturnValues;
7195 if (!CanLowerReturn) {
7196 // The instruction result is the result of loading from the
7197 // hidden sret parameter.
7198 SmallVector<EVT, 1> PVTs;
7199 Type *PtrRetTy = PointerType::getUnqual(OrigRetTy);
7201 ComputeValueVTs(*this, DL, PtrRetTy, PVTs);
7202 assert(PVTs.size() == 1 && "Pointers should fit in one register");
7203 EVT PtrVT = PVTs[0];
7205 unsigned NumValues = RetTys.size();
7206 ReturnValues.resize(NumValues);
7207 SmallVector<SDValue, 4> Chains(NumValues);
7209 for (unsigned i = 0; i < NumValues; ++i) {
7210 SDValue Add = CLI.DAG.getNode(ISD::ADD, CLI.DL, PtrVT, DemoteStackSlot,
7211 CLI.DAG.getConstant(Offsets[i], CLI.DL,
7213 SDValue L = CLI.DAG.getLoad(
7214 RetTys[i], CLI.DL, CLI.Chain, Add,
7215 MachinePointerInfo::getFixedStack(CLI.DAG.getMachineFunction(),
7216 DemoteStackIdx, Offsets[i]),
7217 false, false, false, 1);
7218 ReturnValues[i] = L;
7219 Chains[i] = L.getValue(1);
7222 CLI.Chain = CLI.DAG.getNode(ISD::TokenFactor, CLI.DL, MVT::Other, Chains);
7224 // Collect the legal value parts into potentially illegal values
7225 // that correspond to the original function's return values.
7226 ISD::NodeType AssertOp = ISD::DELETED_NODE;
7228 AssertOp = ISD::AssertSext;
7229 else if (CLI.RetZExt)
7230 AssertOp = ISD::AssertZext;
7231 unsigned CurReg = 0;
7232 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
7234 MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
7235 unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
7237 ReturnValues.push_back(getCopyFromParts(CLI.DAG, CLI.DL, &InVals[CurReg],
7238 NumRegs, RegisterVT, VT, nullptr,
7243 // For a function returning void, there is no return value. We can't create
7244 // such a node, so we just return a null return value in that case. In
7245 // that case, nothing will actually look at the value.
7246 if (ReturnValues.empty())
7247 return std::make_pair(SDValue(), CLI.Chain);
7250 SDValue Res = CLI.DAG.getNode(ISD::MERGE_VALUES, CLI.DL,
7251 CLI.DAG.getVTList(RetTys), ReturnValues);
7252 return std::make_pair(Res, CLI.Chain);
7255 void TargetLowering::LowerOperationWrapper(SDNode *N,
7256 SmallVectorImpl<SDValue> &Results,
7257 SelectionDAG &DAG) const {
7258 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
7260 Results.push_back(Res);
7263 SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
7264 llvm_unreachable("LowerOperation not implemented for this target!");
7268 SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) {
7269 SDValue Op = getNonRegisterValue(V);
7270 assert((Op.getOpcode() != ISD::CopyFromReg ||
7271 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
7272 "Copy from a reg to the same reg!");
7273 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg");
7275 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7276 RegsForValue RFV(V->getContext(), TLI, DAG.getDataLayout(), Reg,
7278 SDValue Chain = DAG.getEntryNode();
7280 ISD::NodeType ExtendType = (FuncInfo.PreferredExtendType.find(V) ==
7281 FuncInfo.PreferredExtendType.end())
7283 : FuncInfo.PreferredExtendType[V];
7284 RFV.getCopyToRegs(Op, DAG, getCurSDLoc(), Chain, nullptr, V, ExtendType);
7285 PendingExports.push_back(Chain);
7288 #include "llvm/CodeGen/SelectionDAGISel.h"
7290 /// isOnlyUsedInEntryBlock - If the specified argument is only used in the
7291 /// entry block, return true. This includes arguments used by switches, since
7292 /// the switch may expand into multiple basic blocks.
7293 static bool isOnlyUsedInEntryBlock(const Argument *A, bool FastISel) {
7294 // With FastISel active, we may be splitting blocks, so force creation
7295 // of virtual registers for all non-dead arguments.
7297 return A->use_empty();
7299 const BasicBlock &Entry = A->getParent()->front();
7300 for (const User *U : A->users())
7301 if (cast<Instruction>(U)->getParent() != &Entry || isa<SwitchInst>(U))
7302 return false; // Use not in entry block.
7307 void SelectionDAGISel::LowerArguments(const Function &F) {
7308 SelectionDAG &DAG = SDB->DAG;
7309 SDLoc dl = SDB->getCurSDLoc();
7310 const DataLayout &DL = DAG.getDataLayout();
7311 SmallVector<ISD::InputArg, 16> Ins;
7313 if (!FuncInfo->CanLowerReturn) {
7314 // Put in an sret pointer parameter before all the other parameters.
7315 SmallVector<EVT, 1> ValueVTs;
7316 ComputeValueVTs(*TLI, DAG.getDataLayout(),
7317 PointerType::getUnqual(F.getReturnType()), ValueVTs);
7319 // NOTE: Assuming that a pointer will never break down to more than one VT
7321 ISD::ArgFlagsTy Flags;
7323 MVT RegisterVT = TLI->getRegisterType(*DAG.getContext(), ValueVTs[0]);
7324 ISD::InputArg RetArg(Flags, RegisterVT, ValueVTs[0], true,
7325 ISD::InputArg::NoArgIndex, 0);
7326 Ins.push_back(RetArg);
7329 // Set up the incoming argument description vector.
7331 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end();
7332 I != E; ++I, ++Idx) {
7333 SmallVector<EVT, 4> ValueVTs;
7334 ComputeValueVTs(*TLI, DAG.getDataLayout(), I->getType(), ValueVTs);
7335 bool isArgValueUsed = !I->use_empty();
7336 unsigned PartBase = 0;
7337 Type *FinalType = I->getType();
7338 if (F.getAttributes().hasAttribute(Idx, Attribute::ByVal))
7339 FinalType = cast<PointerType>(FinalType)->getElementType();
7340 bool NeedsRegBlock = TLI->functionArgumentNeedsConsecutiveRegisters(
7341 FinalType, F.getCallingConv(), F.isVarArg());
7342 for (unsigned Value = 0, NumValues = ValueVTs.size();
7343 Value != NumValues; ++Value) {
7344 EVT VT = ValueVTs[Value];
7345 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
7346 ISD::ArgFlagsTy Flags;
7347 unsigned OriginalAlignment = DL.getABITypeAlignment(ArgTy);
7349 if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt))
7351 if (F.getAttributes().hasAttribute(Idx, Attribute::SExt))
7353 if (F.getAttributes().hasAttribute(Idx, Attribute::InReg))
7355 if (F.getAttributes().hasAttribute(Idx, Attribute::StructRet))
7357 if (F.getAttributes().hasAttribute(Idx, Attribute::ByVal))
7359 if (F.getAttributes().hasAttribute(Idx, Attribute::InAlloca)) {
7360 Flags.setInAlloca();
7361 // Set the byval flag for CCAssignFn callbacks that don't know about
7362 // inalloca. This way we can know how many bytes we should've allocated
7363 // and how many bytes a callee cleanup function will pop. If we port
7364 // inalloca to more targets, we'll have to add custom inalloca handling
7365 // in the various CC lowering callbacks.
7368 if (Flags.isByVal() || Flags.isInAlloca()) {
7369 PointerType *Ty = cast<PointerType>(I->getType());
7370 Type *ElementTy = Ty->getElementType();
7371 Flags.setByValSize(DL.getTypeAllocSize(ElementTy));
7372 // For ByVal, alignment should be passed from FE. BE will guess if
7373 // this info is not there but there are cases it cannot get right.
7374 unsigned FrameAlign;
7375 if (F.getParamAlignment(Idx))
7376 FrameAlign = F.getParamAlignment(Idx);
7378 FrameAlign = TLI->getByValTypeAlignment(ElementTy, DL);
7379 Flags.setByValAlign(FrameAlign);
7381 if (F.getAttributes().hasAttribute(Idx, Attribute::Nest))
7384 Flags.setInConsecutiveRegs();
7385 Flags.setOrigAlign(OriginalAlignment);
7387 MVT RegisterVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7388 unsigned NumRegs = TLI->getNumRegisters(*CurDAG->getContext(), VT);
7389 for (unsigned i = 0; i != NumRegs; ++i) {
7390 ISD::InputArg MyFlags(Flags, RegisterVT, VT, isArgValueUsed,
7391 Idx-1, PartBase+i*RegisterVT.getStoreSize());
7392 if (NumRegs > 1 && i == 0)
7393 MyFlags.Flags.setSplit();
7394 // if it isn't first piece, alignment must be 1
7396 MyFlags.Flags.setOrigAlign(1);
7397 Ins.push_back(MyFlags);
7399 if (NeedsRegBlock && Value == NumValues - 1)
7400 Ins[Ins.size() - 1].Flags.setInConsecutiveRegsLast();
7401 PartBase += VT.getStoreSize();
7405 // Call the target to set up the argument values.
7406 SmallVector<SDValue, 8> InVals;
7407 SDValue NewRoot = TLI->LowerFormalArguments(
7408 DAG.getRoot(), F.getCallingConv(), F.isVarArg(), Ins, dl, DAG, InVals);
7410 // Verify that the target's LowerFormalArguments behaved as expected.
7411 assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other &&
7412 "LowerFormalArguments didn't return a valid chain!");
7413 assert(InVals.size() == Ins.size() &&
7414 "LowerFormalArguments didn't emit the correct number of values!");
7416 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
7417 assert(InVals[i].getNode() &&
7418 "LowerFormalArguments emitted a null value!");
7419 assert(EVT(Ins[i].VT) == InVals[i].getValueType() &&
7420 "LowerFormalArguments emitted a value with the wrong type!");
7424 // Update the DAG with the new chain value resulting from argument lowering.
7425 DAG.setRoot(NewRoot);
7427 // Set up the argument values.
7430 if (!FuncInfo->CanLowerReturn) {
7431 // Create a virtual register for the sret pointer, and put in a copy
7432 // from the sret argument into it.
7433 SmallVector<EVT, 1> ValueVTs;
7434 ComputeValueVTs(*TLI, DAG.getDataLayout(),
7435 PointerType::getUnqual(F.getReturnType()), ValueVTs);
7436 MVT VT = ValueVTs[0].getSimpleVT();
7437 MVT RegVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7438 ISD::NodeType AssertOp = ISD::DELETED_NODE;
7439 SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1,
7440 RegVT, VT, nullptr, AssertOp);
7442 MachineFunction& MF = SDB->DAG.getMachineFunction();
7443 MachineRegisterInfo& RegInfo = MF.getRegInfo();
7444 unsigned SRetReg = RegInfo.createVirtualRegister(TLI->getRegClassFor(RegVT));
7445 FuncInfo->DemoteRegister = SRetReg;
7447 SDB->DAG.getCopyToReg(NewRoot, SDB->getCurSDLoc(), SRetReg, ArgValue);
7448 DAG.setRoot(NewRoot);
7450 // i indexes lowered arguments. Bump it past the hidden sret argument.
7451 // Idx indexes LLVM arguments. Don't touch it.
7455 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
7457 SmallVector<SDValue, 4> ArgValues;
7458 SmallVector<EVT, 4> ValueVTs;
7459 ComputeValueVTs(*TLI, DAG.getDataLayout(), I->getType(), ValueVTs);
7460 unsigned NumValues = ValueVTs.size();
7462 // If this argument is unused then remember its value. It is used to generate
7463 // debugging information.
7464 if (I->use_empty() && NumValues) {
7465 SDB->setUnusedArgValue(&*I, InVals[i]);
7467 // Also remember any frame index for use in FastISel.
7468 if (FrameIndexSDNode *FI =
7469 dyn_cast<FrameIndexSDNode>(InVals[i].getNode()))
7470 FuncInfo->setArgumentFrameIndex(&*I, FI->getIndex());
7473 for (unsigned Val = 0; Val != NumValues; ++Val) {
7474 EVT VT = ValueVTs[Val];
7475 MVT PartVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7476 unsigned NumParts = TLI->getNumRegisters(*CurDAG->getContext(), VT);
7478 if (!I->use_empty()) {
7479 ISD::NodeType AssertOp = ISD::DELETED_NODE;
7480 if (F.getAttributes().hasAttribute(Idx, Attribute::SExt))
7481 AssertOp = ISD::AssertSext;
7482 else if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt))
7483 AssertOp = ISD::AssertZext;
7485 ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i],
7486 NumParts, PartVT, VT,
7487 nullptr, AssertOp));
7493 // We don't need to do anything else for unused arguments.
7494 if (ArgValues.empty())
7497 // Note down frame index.
7498 if (FrameIndexSDNode *FI =
7499 dyn_cast<FrameIndexSDNode>(ArgValues[0].getNode()))
7500 FuncInfo->setArgumentFrameIndex(&*I, FI->getIndex());
7502 SDValue Res = DAG.getMergeValues(makeArrayRef(ArgValues.data(), NumValues),
7503 SDB->getCurSDLoc());
7505 SDB->setValue(&*I, Res);
7506 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::BUILD_PAIR) {
7507 if (LoadSDNode *LNode =
7508 dyn_cast<LoadSDNode>(Res.getOperand(0).getNode()))
7509 if (FrameIndexSDNode *FI =
7510 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode()))
7511 FuncInfo->setArgumentFrameIndex(&*I, FI->getIndex());
7514 // If this argument is live outside of the entry block, insert a copy from
7515 // wherever we got it to the vreg that other BB's will reference it as.
7516 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::CopyFromReg) {
7517 // If we can, though, try to skip creating an unnecessary vreg.
7518 // FIXME: This isn't very clean... it would be nice to make this more
7519 // general. It's also subtly incompatible with the hacks FastISel
7521 unsigned Reg = cast<RegisterSDNode>(Res.getOperand(1))->getReg();
7522 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
7523 FuncInfo->ValueMap[&*I] = Reg;
7527 if (!isOnlyUsedInEntryBlock(&*I, TM.Options.EnableFastISel)) {
7528 FuncInfo->InitializeRegForValue(&*I);
7529 SDB->CopyToExportRegsIfNeeded(&*I);
7533 assert(i == InVals.size() && "Argument register count mismatch!");
7535 // Finally, if the target has anything special to do, allow it to do so.
7536 EmitFunctionEntryCode();
7539 /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
7540 /// ensure constants are generated when needed. Remember the virtual registers
7541 /// that need to be added to the Machine PHI nodes as input. We cannot just
7542 /// directly add them, because expansion might result in multiple MBB's for one
7543 /// BB. As such, the start of the BB might correspond to a different MBB than
7547 SelectionDAGBuilder::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
7548 const TerminatorInst *TI = LLVMBB->getTerminator();
7550 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
7552 // Check PHI nodes in successors that expect a value to be available from this
7554 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
7555 const BasicBlock *SuccBB = TI->getSuccessor(succ);
7556 if (!isa<PHINode>(SuccBB->begin())) continue;
7557 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
7559 // If this terminator has multiple identical successors (common for
7560 // switches), only handle each succ once.
7561 if (!SuccsHandled.insert(SuccMBB).second)
7564 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
7566 // At this point we know that there is a 1-1 correspondence between LLVM PHI
7567 // nodes and Machine PHI nodes, but the incoming operands have not been
7569 for (BasicBlock::const_iterator I = SuccBB->begin();
7570 const PHINode *PN = dyn_cast<PHINode>(I); ++I) {
7571 // Ignore dead phi's.
7572 if (PN->use_empty()) continue;
7575 if (PN->getType()->isEmptyTy())
7579 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
7581 if (const Constant *C = dyn_cast<Constant>(PHIOp)) {
7582 unsigned &RegOut = ConstantsOut[C];
7584 RegOut = FuncInfo.CreateRegs(C->getType());
7585 CopyValueToVirtualRegister(C, RegOut);
7589 DenseMap<const Value *, unsigned>::iterator I =
7590 FuncInfo.ValueMap.find(PHIOp);
7591 if (I != FuncInfo.ValueMap.end())
7594 assert(isa<AllocaInst>(PHIOp) &&
7595 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
7596 "Didn't codegen value into a register!??");
7597 Reg = FuncInfo.CreateRegs(PHIOp->getType());
7598 CopyValueToVirtualRegister(PHIOp, Reg);
7602 // Remember that this register needs to added to the machine PHI node as
7603 // the input for this MBB.
7604 SmallVector<EVT, 4> ValueVTs;
7605 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7606 ComputeValueVTs(TLI, DAG.getDataLayout(), PN->getType(), ValueVTs);
7607 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
7608 EVT VT = ValueVTs[vti];
7609 unsigned NumRegisters = TLI.getNumRegisters(*DAG.getContext(), VT);
7610 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
7611 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
7612 Reg += NumRegisters;
7617 ConstantsOut.clear();
7620 /// Add a successor MBB to ParentMBB< creating a new MachineBB for BB if SuccMBB
7623 SelectionDAGBuilder::StackProtectorDescriptor::
7624 AddSuccessorMBB(const BasicBlock *BB,
7625 MachineBasicBlock *ParentMBB,
7627 MachineBasicBlock *SuccMBB) {
7628 // If SuccBB has not been created yet, create it.
7630 MachineFunction *MF = ParentMBB->getParent();
7631 MachineFunction::iterator BBI(ParentMBB);
7632 SuccMBB = MF->CreateMachineBasicBlock(BB);
7633 MF->insert(++BBI, SuccMBB);
7635 // Add it as a successor of ParentMBB.
7636 ParentMBB->addSuccessor(
7637 SuccMBB, BranchProbabilityInfo::getBranchProbStackProtector(IsLikely));
7641 MachineBasicBlock *SelectionDAGBuilder::NextBlock(MachineBasicBlock *MBB) {
7642 MachineFunction::iterator I(MBB);
7643 if (++I == FuncInfo.MF->end())
7648 /// During lowering new call nodes can be created (such as memset, etc.).
7649 /// Those will become new roots of the current DAG, but complications arise
7650 /// when they are tail calls. In such cases, the call lowering will update
7651 /// the root, but the builder still needs to know that a tail call has been
7652 /// lowered in order to avoid generating an additional return.
7653 void SelectionDAGBuilder::updateDAGForMaybeTailCall(SDValue MaybeTC) {
7654 // If the node is null, we do have a tail call.
7655 if (MaybeTC.getNode() != nullptr)
7656 DAG.setRoot(MaybeTC);
7661 bool SelectionDAGBuilder::isDense(const CaseClusterVector &Clusters,
7662 unsigned *TotalCases, unsigned First,
7664 assert(Last >= First);
7665 assert(TotalCases[Last] >= TotalCases[First]);
7667 APInt LowCase = Clusters[First].Low->getValue();
7668 APInt HighCase = Clusters[Last].High->getValue();
7669 assert(LowCase.getBitWidth() == HighCase.getBitWidth());
7671 // FIXME: A range of consecutive cases has 100% density, but only requires one
7672 // comparison to lower. We should discriminate against such consecutive ranges
7675 uint64_t Diff = (HighCase - LowCase).getLimitedValue((UINT64_MAX - 1) / 100);
7676 uint64_t Range = Diff + 1;
7679 TotalCases[Last] - (First == 0 ? 0 : TotalCases[First - 1]);
7681 assert(NumCases < UINT64_MAX / 100);
7682 assert(Range >= NumCases);
7684 return NumCases * 100 >= Range * MinJumpTableDensity;
7687 static inline bool areJTsAllowed(const TargetLowering &TLI) {
7688 return TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
7689 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other);
7692 bool SelectionDAGBuilder::buildJumpTable(CaseClusterVector &Clusters,
7693 unsigned First, unsigned Last,
7694 const SwitchInst *SI,
7695 MachineBasicBlock *DefaultMBB,
7696 CaseCluster &JTCluster) {
7697 assert(First <= Last);
7699 auto Prob = BranchProbability::getZero();
7700 unsigned NumCmps = 0;
7701 std::vector<MachineBasicBlock*> Table;
7702 DenseMap<MachineBasicBlock*, BranchProbability> JTProbs;
7704 // Initialize probabilities in JTProbs.
7705 for (unsigned I = First; I <= Last; ++I)
7706 JTProbs[Clusters[I].MBB] = BranchProbability::getZero();
7708 for (unsigned I = First; I <= Last; ++I) {
7709 assert(Clusters[I].Kind == CC_Range);
7710 Prob += Clusters[I].Prob;
7711 APInt Low = Clusters[I].Low->getValue();
7712 APInt High = Clusters[I].High->getValue();
7713 NumCmps += (Low == High) ? 1 : 2;
7715 // Fill the gap between this and the previous cluster.
7716 APInt PreviousHigh = Clusters[I - 1].High->getValue();
7717 assert(PreviousHigh.slt(Low));
7718 uint64_t Gap = (Low - PreviousHigh).getLimitedValue() - 1;
7719 for (uint64_t J = 0; J < Gap; J++)
7720 Table.push_back(DefaultMBB);
7722 uint64_t ClusterSize = (High - Low).getLimitedValue() + 1;
7723 for (uint64_t J = 0; J < ClusterSize; ++J)
7724 Table.push_back(Clusters[I].MBB);
7725 JTProbs[Clusters[I].MBB] += Clusters[I].Prob;
7728 unsigned NumDests = JTProbs.size();
7729 if (isSuitableForBitTests(NumDests, NumCmps,
7730 Clusters[First].Low->getValue(),
7731 Clusters[Last].High->getValue())) {
7732 // Clusters[First..Last] should be lowered as bit tests instead.
7736 // Create the MBB that will load from and jump through the table.
7737 // Note: We create it here, but it's not inserted into the function yet.
7738 MachineFunction *CurMF = FuncInfo.MF;
7739 MachineBasicBlock *JumpTableMBB =
7740 CurMF->CreateMachineBasicBlock(SI->getParent());
7742 // Add successors. Note: use table order for determinism.
7743 SmallPtrSet<MachineBasicBlock *, 8> Done;
7744 for (MachineBasicBlock *Succ : Table) {
7745 if (Done.count(Succ))
7747 addSuccessorWithProb(JumpTableMBB, Succ, JTProbs[Succ]);
7750 JumpTableMBB->normalizeSuccProbs();
7752 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7753 unsigned JTI = CurMF->getOrCreateJumpTableInfo(TLI.getJumpTableEncoding())
7754 ->createJumpTableIndex(Table);
7756 // Set up the jump table info.
7757 JumpTable JT(-1U, JTI, JumpTableMBB, nullptr);
7758 JumpTableHeader JTH(Clusters[First].Low->getValue(),
7759 Clusters[Last].High->getValue(), SI->getCondition(),
7761 JTCases.emplace_back(std::move(JTH), std::move(JT));
7763 JTCluster = CaseCluster::jumpTable(Clusters[First].Low, Clusters[Last].High,
7764 JTCases.size() - 1, Prob);
7768 void SelectionDAGBuilder::findJumpTables(CaseClusterVector &Clusters,
7769 const SwitchInst *SI,
7770 MachineBasicBlock *DefaultMBB) {
7772 // Clusters must be non-empty, sorted, and only contain Range clusters.
7773 assert(!Clusters.empty());
7774 for (CaseCluster &C : Clusters)
7775 assert(C.Kind == CC_Range);
7776 for (unsigned i = 1, e = Clusters.size(); i < e; ++i)
7777 assert(Clusters[i - 1].High->getValue().slt(Clusters[i].Low->getValue()));
7780 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7781 if (!areJTsAllowed(TLI))
7784 const int64_t N = Clusters.size();
7785 const unsigned MinJumpTableSize = TLI.getMinimumJumpTableEntries();
7787 // TotalCases[i]: Total nbr of cases in Clusters[0..i].
7788 SmallVector<unsigned, 8> TotalCases(N);
7790 for (unsigned i = 0; i < N; ++i) {
7791 APInt Hi = Clusters[i].High->getValue();
7792 APInt Lo = Clusters[i].Low->getValue();
7793 TotalCases[i] = (Hi - Lo).getLimitedValue() + 1;
7795 TotalCases[i] += TotalCases[i - 1];
7798 if (N >= MinJumpTableSize && isDense(Clusters, &TotalCases[0], 0, N - 1)) {
7799 // Cheap case: the whole range might be suitable for jump table.
7800 CaseCluster JTCluster;
7801 if (buildJumpTable(Clusters, 0, N - 1, SI, DefaultMBB, JTCluster)) {
7802 Clusters[0] = JTCluster;
7808 // The algorithm below is not suitable for -O0.
7809 if (TM.getOptLevel() == CodeGenOpt::None)
7812 // Split Clusters into minimum number of dense partitions. The algorithm uses
7813 // the same idea as Kannan & Proebsting "Correction to 'Producing Good Code
7814 // for the Case Statement'" (1994), but builds the MinPartitions array in
7815 // reverse order to make it easier to reconstruct the partitions in ascending
7816 // order. In the choice between two optimal partitionings, it picks the one
7817 // which yields more jump tables.
7819 // MinPartitions[i] is the minimum nbr of partitions of Clusters[i..N-1].
7820 SmallVector<unsigned, 8> MinPartitions(N);
7821 // LastElement[i] is the last element of the partition starting at i.
7822 SmallVector<unsigned, 8> LastElement(N);
7823 // NumTables[i]: nbr of >= MinJumpTableSize partitions from Clusters[i..N-1].
7824 SmallVector<unsigned, 8> NumTables(N);
7826 // Base case: There is only one way to partition Clusters[N-1].
7827 MinPartitions[N - 1] = 1;
7828 LastElement[N - 1] = N - 1;
7829 assert(MinJumpTableSize > 1);
7830 NumTables[N - 1] = 0;
7832 // Note: loop indexes are signed to avoid underflow.
7833 for (int64_t i = N - 2; i >= 0; i--) {
7834 // Find optimal partitioning of Clusters[i..N-1].
7835 // Baseline: Put Clusters[i] into a partition on its own.
7836 MinPartitions[i] = MinPartitions[i + 1] + 1;
7838 NumTables[i] = NumTables[i + 1];
7840 // Search for a solution that results in fewer partitions.
7841 for (int64_t j = N - 1; j > i; j--) {
7842 // Try building a partition from Clusters[i..j].
7843 if (isDense(Clusters, &TotalCases[0], i, j)) {
7844 unsigned NumPartitions = 1 + (j == N - 1 ? 0 : MinPartitions[j + 1]);
7845 bool IsTable = j - i + 1 >= MinJumpTableSize;
7846 unsigned Tables = IsTable + (j == N - 1 ? 0 : NumTables[j + 1]);
7848 // If this j leads to fewer partitions, or same number of partitions
7849 // with more lookup tables, it is a better partitioning.
7850 if (NumPartitions < MinPartitions[i] ||
7851 (NumPartitions == MinPartitions[i] && Tables > NumTables[i])) {
7852 MinPartitions[i] = NumPartitions;
7854 NumTables[i] = Tables;
7860 // Iterate over the partitions, replacing some with jump tables in-place.
7861 unsigned DstIndex = 0;
7862 for (unsigned First = 0, Last; First < N; First = Last + 1) {
7863 Last = LastElement[First];
7864 assert(Last >= First);
7865 assert(DstIndex <= First);
7866 unsigned NumClusters = Last - First + 1;
7868 CaseCluster JTCluster;
7869 if (NumClusters >= MinJumpTableSize &&
7870 buildJumpTable(Clusters, First, Last, SI, DefaultMBB, JTCluster)) {
7871 Clusters[DstIndex++] = JTCluster;
7873 for (unsigned I = First; I <= Last; ++I)
7874 std::memmove(&Clusters[DstIndex++], &Clusters[I], sizeof(Clusters[I]));
7877 Clusters.resize(DstIndex);
7880 bool SelectionDAGBuilder::rangeFitsInWord(const APInt &Low, const APInt &High) {
7881 // FIXME: Using the pointer type doesn't seem ideal.
7882 uint64_t BW = DAG.getDataLayout().getPointerSizeInBits();
7883 uint64_t Range = (High - Low).getLimitedValue(UINT64_MAX - 1) + 1;
7887 bool SelectionDAGBuilder::isSuitableForBitTests(unsigned NumDests,
7890 const APInt &High) {
7891 // FIXME: I don't think NumCmps is the correct metric: a single case and a
7892 // range of cases both require only one branch to lower. Just looking at the
7893 // number of clusters and destinations should be enough to decide whether to
7896 // To lower a range with bit tests, the range must fit the bitwidth of a
7898 if (!rangeFitsInWord(Low, High))
7901 // Decide whether it's profitable to lower this range with bit tests. Each
7902 // destination requires a bit test and branch, and there is an overall range
7903 // check branch. For a small number of clusters, separate comparisons might be
7904 // cheaper, and for many destinations, splitting the range might be better.
7905 return (NumDests == 1 && NumCmps >= 3) ||
7906 (NumDests == 2 && NumCmps >= 5) ||
7907 (NumDests == 3 && NumCmps >= 6);
7910 bool SelectionDAGBuilder::buildBitTests(CaseClusterVector &Clusters,
7911 unsigned First, unsigned Last,
7912 const SwitchInst *SI,
7913 CaseCluster &BTCluster) {
7914 assert(First <= Last);
7918 BitVector Dests(FuncInfo.MF->getNumBlockIDs());
7919 unsigned NumCmps = 0;
7920 for (int64_t I = First; I <= Last; ++I) {
7921 assert(Clusters[I].Kind == CC_Range);
7922 Dests.set(Clusters[I].MBB->getNumber());
7923 NumCmps += (Clusters[I].Low == Clusters[I].High) ? 1 : 2;
7925 unsigned NumDests = Dests.count();
7927 APInt Low = Clusters[First].Low->getValue();
7928 APInt High = Clusters[Last].High->getValue();
7929 assert(Low.slt(High));
7931 if (!isSuitableForBitTests(NumDests, NumCmps, Low, High))
7937 const int BitWidth = DAG.getTargetLoweringInfo()
7938 .getPointerTy(DAG.getDataLayout())
7940 assert(rangeFitsInWord(Low, High) && "Case range must fit in bit mask!");
7942 // Check if the clusters cover a contiguous range such that no value in the
7943 // range will jump to the default statement.
7944 bool ContiguousRange = true;
7945 for (int64_t I = First + 1; I <= Last; ++I) {
7946 if (Clusters[I].Low->getValue() != Clusters[I - 1].High->getValue() + 1) {
7947 ContiguousRange = false;
7952 if (Low.isStrictlyPositive() && High.slt(BitWidth)) {
7953 // Optimize the case where all the case values fit in a word without having
7954 // to subtract minValue. In this case, we can optimize away the subtraction.
7955 LowBound = APInt::getNullValue(Low.getBitWidth());
7957 ContiguousRange = false;
7960 CmpRange = High - Low;
7964 auto TotalProb = BranchProbability::getZero();
7965 for (unsigned i = First; i <= Last; ++i) {
7966 // Find the CaseBits for this destination.
7968 for (j = 0; j < CBV.size(); ++j)
7969 if (CBV[j].BB == Clusters[i].MBB)
7971 if (j == CBV.size())
7973 CaseBits(0, Clusters[i].MBB, 0, BranchProbability::getZero()));
7974 CaseBits *CB = &CBV[j];
7976 // Update Mask, Bits and ExtraProb.
7977 uint64_t Lo = (Clusters[i].Low->getValue() - LowBound).getZExtValue();
7978 uint64_t Hi = (Clusters[i].High->getValue() - LowBound).getZExtValue();
7979 assert(Hi >= Lo && Hi < 64 && "Invalid bit case!");
7980 CB->Mask |= (-1ULL >> (63 - (Hi - Lo))) << Lo;
7981 CB->Bits += Hi - Lo + 1;
7982 CB->ExtraProb += Clusters[i].Prob;
7983 TotalProb += Clusters[i].Prob;
7987 std::sort(CBV.begin(), CBV.end(), [](const CaseBits &a, const CaseBits &b) {
7988 // Sort by probability first, number of bits second.
7989 if (a.ExtraProb != b.ExtraProb)
7990 return a.ExtraProb > b.ExtraProb;
7991 return a.Bits > b.Bits;
7994 for (auto &CB : CBV) {
7995 MachineBasicBlock *BitTestBB =
7996 FuncInfo.MF->CreateMachineBasicBlock(SI->getParent());
7997 BTI.push_back(BitTestCase(CB.Mask, BitTestBB, CB.BB, CB.ExtraProb));
7999 BitTestCases.emplace_back(std::move(LowBound), std::move(CmpRange),
8000 SI->getCondition(), -1U, MVT::Other, false,
8001 ContiguousRange, nullptr, nullptr, std::move(BTI),
8004 BTCluster = CaseCluster::bitTests(Clusters[First].Low, Clusters[Last].High,
8005 BitTestCases.size() - 1, TotalProb);
8009 void SelectionDAGBuilder::findBitTestClusters(CaseClusterVector &Clusters,
8010 const SwitchInst *SI) {
8011 // Partition Clusters into as few subsets as possible, where each subset has a
8012 // range that fits in a machine word and has <= 3 unique destinations.
8015 // Clusters must be sorted and contain Range or JumpTable clusters.
8016 assert(!Clusters.empty());
8017 assert(Clusters[0].Kind == CC_Range || Clusters[0].Kind == CC_JumpTable);
8018 for (const CaseCluster &C : Clusters)
8019 assert(C.Kind == CC_Range || C.Kind == CC_JumpTable);
8020 for (unsigned i = 1; i < Clusters.size(); ++i)
8021 assert(Clusters[i-1].High->getValue().slt(Clusters[i].Low->getValue()));
8024 // The algorithm below is not suitable for -O0.
8025 if (TM.getOptLevel() == CodeGenOpt::None)
8028 // If target does not have legal shift left, do not emit bit tests at all.
8029 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8030 EVT PTy = TLI.getPointerTy(DAG.getDataLayout());
8031 if (!TLI.isOperationLegal(ISD::SHL, PTy))
8034 int BitWidth = PTy.getSizeInBits();
8035 const int64_t N = Clusters.size();
8037 // MinPartitions[i] is the minimum nbr of partitions of Clusters[i..N-1].
8038 SmallVector<unsigned, 8> MinPartitions(N);
8039 // LastElement[i] is the last element of the partition starting at i.
8040 SmallVector<unsigned, 8> LastElement(N);
8042 // FIXME: This might not be the best algorithm for finding bit test clusters.
8044 // Base case: There is only one way to partition Clusters[N-1].
8045 MinPartitions[N - 1] = 1;
8046 LastElement[N - 1] = N - 1;
8048 // Note: loop indexes are signed to avoid underflow.
8049 for (int64_t i = N - 2; i >= 0; --i) {
8050 // Find optimal partitioning of Clusters[i..N-1].
8051 // Baseline: Put Clusters[i] into a partition on its own.
8052 MinPartitions[i] = MinPartitions[i + 1] + 1;
8055 // Search for a solution that results in fewer partitions.
8056 // Note: the search is limited by BitWidth, reducing time complexity.
8057 for (int64_t j = std::min(N - 1, i + BitWidth - 1); j > i; --j) {
8058 // Try building a partition from Clusters[i..j].
8061 if (!rangeFitsInWord(Clusters[i].Low->getValue(),
8062 Clusters[j].High->getValue()))
8065 // Check nbr of destinations and cluster types.
8066 // FIXME: This works, but doesn't seem very efficient.
8067 bool RangesOnly = true;
8068 BitVector Dests(FuncInfo.MF->getNumBlockIDs());
8069 for (int64_t k = i; k <= j; k++) {
8070 if (Clusters[k].Kind != CC_Range) {
8074 Dests.set(Clusters[k].MBB->getNumber());
8076 if (!RangesOnly || Dests.count() > 3)
8079 // Check if it's a better partition.
8080 unsigned NumPartitions = 1 + (j == N - 1 ? 0 : MinPartitions[j + 1]);
8081 if (NumPartitions < MinPartitions[i]) {
8082 // Found a better partition.
8083 MinPartitions[i] = NumPartitions;
8089 // Iterate over the partitions, replacing with bit-test clusters in-place.
8090 unsigned DstIndex = 0;
8091 for (unsigned First = 0, Last; First < N; First = Last + 1) {
8092 Last = LastElement[First];
8093 assert(First <= Last);
8094 assert(DstIndex <= First);
8096 CaseCluster BitTestCluster;
8097 if (buildBitTests(Clusters, First, Last, SI, BitTestCluster)) {
8098 Clusters[DstIndex++] = BitTestCluster;
8100 size_t NumClusters = Last - First + 1;
8101 std::memmove(&Clusters[DstIndex], &Clusters[First],
8102 sizeof(Clusters[0]) * NumClusters);
8103 DstIndex += NumClusters;
8106 Clusters.resize(DstIndex);
8109 void SelectionDAGBuilder::lowerWorkItem(SwitchWorkListItem W, Value *Cond,
8110 MachineBasicBlock *SwitchMBB,
8111 MachineBasicBlock *DefaultMBB) {
8112 MachineFunction *CurMF = FuncInfo.MF;
8113 MachineBasicBlock *NextMBB = nullptr;
8114 MachineFunction::iterator BBI(W.MBB);
8115 if (++BBI != FuncInfo.MF->end())
8118 unsigned Size = W.LastCluster - W.FirstCluster + 1;
8120 BranchProbabilityInfo *BPI = FuncInfo.BPI;
8122 if (Size == 2 && W.MBB == SwitchMBB) {
8123 // If any two of the cases has the same destination, and if one value
8124 // is the same as the other, but has one bit unset that the other has set,
8125 // use bit manipulation to do two compares at once. For example:
8126 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
8127 // TODO: This could be extended to merge any 2 cases in switches with 3
8129 // TODO: Handle cases where W.CaseBB != SwitchBB.
8130 CaseCluster &Small = *W.FirstCluster;
8131 CaseCluster &Big = *W.LastCluster;
8133 if (Small.Low == Small.High && Big.Low == Big.High &&
8134 Small.MBB == Big.MBB) {
8135 const APInt &SmallValue = Small.Low->getValue();
8136 const APInt &BigValue = Big.Low->getValue();
8138 // Check that there is only one bit different.
8139 APInt CommonBit = BigValue ^ SmallValue;
8140 if (CommonBit.isPowerOf2()) {
8141 SDValue CondLHS = getValue(Cond);
8142 EVT VT = CondLHS.getValueType();
8143 SDLoc DL = getCurSDLoc();
8145 SDValue Or = DAG.getNode(ISD::OR, DL, VT, CondLHS,
8146 DAG.getConstant(CommonBit, DL, VT));
8147 SDValue Cond = DAG.getSetCC(
8148 DL, MVT::i1, Or, DAG.getConstant(BigValue | SmallValue, DL, VT),
8151 // Update successor info.
8152 // Both Small and Big will jump to Small.BB, so we sum up the
8154 addSuccessorWithProb(SwitchMBB, Small.MBB, Small.Prob + Big.Prob);
8156 addSuccessorWithProb(
8157 SwitchMBB, DefaultMBB,
8158 // The default destination is the first successor in IR.
8159 BPI->getEdgeProbability(SwitchMBB->getBasicBlock(), (unsigned)0));
8161 addSuccessorWithProb(SwitchMBB, DefaultMBB);
8163 // Insert the true branch.
8165 DAG.getNode(ISD::BRCOND, DL, MVT::Other, getControlRoot(), Cond,
8166 DAG.getBasicBlock(Small.MBB));
8167 // Insert the false branch.
8168 BrCond = DAG.getNode(ISD::BR, DL, MVT::Other, BrCond,
8169 DAG.getBasicBlock(DefaultMBB));
8171 DAG.setRoot(BrCond);
8177 if (TM.getOptLevel() != CodeGenOpt::None) {
8178 // Order cases by probability so the most likely case will be checked first.
8179 std::sort(W.FirstCluster, W.LastCluster + 1,
8180 [](const CaseCluster &a, const CaseCluster &b) {
8181 return a.Prob > b.Prob;
8184 // Rearrange the case blocks so that the last one falls through if possible
8185 // without without changing the order of probabilities.
8186 for (CaseClusterIt I = W.LastCluster; I > W.FirstCluster; ) {
8188 if (I->Prob > W.LastCluster->Prob)
8190 if (I->Kind == CC_Range && I->MBB == NextMBB) {
8191 std::swap(*I, *W.LastCluster);
8197 // Compute total probability.
8198 BranchProbability DefaultProb = W.DefaultProb;
8199 BranchProbability UnhandledProbs = DefaultProb;
8200 for (CaseClusterIt I = W.FirstCluster; I <= W.LastCluster; ++I)
8201 UnhandledProbs += I->Prob;
8203 MachineBasicBlock *CurMBB = W.MBB;
8204 for (CaseClusterIt I = W.FirstCluster, E = W.LastCluster; I <= E; ++I) {
8205 MachineBasicBlock *Fallthrough;
8206 if (I == W.LastCluster) {
8207 // For the last cluster, fall through to the default destination.
8208 Fallthrough = DefaultMBB;
8210 Fallthrough = CurMF->CreateMachineBasicBlock(CurMBB->getBasicBlock());
8211 CurMF->insert(BBI, Fallthrough);
8212 // Put Cond in a virtual register to make it available from the new blocks.
8213 ExportFromCurrentBlock(Cond);
8215 UnhandledProbs -= I->Prob;
8218 case CC_JumpTable: {
8219 // FIXME: Optimize away range check based on pivot comparisons.
8220 JumpTableHeader *JTH = &JTCases[I->JTCasesIndex].first;
8221 JumpTable *JT = &JTCases[I->JTCasesIndex].second;
8223 // The jump block hasn't been inserted yet; insert it here.
8224 MachineBasicBlock *JumpMBB = JT->MBB;
8225 CurMF->insert(BBI, JumpMBB);
8227 auto JumpProb = I->Prob;
8228 auto FallthroughProb = UnhandledProbs;
8230 // If the default statement is a target of the jump table, we evenly
8231 // distribute the default probability to successors of CurMBB. Also
8232 // update the probability on the edge from JumpMBB to Fallthrough.
8233 for (MachineBasicBlock::succ_iterator SI = JumpMBB->succ_begin(),
8234 SE = JumpMBB->succ_end();
8236 if (*SI == DefaultMBB) {
8237 JumpProb += DefaultProb / 2;
8238 FallthroughProb -= DefaultProb / 2;
8239 JumpMBB->setSuccProbability(SI, DefaultProb / 2);
8240 JumpMBB->normalizeSuccProbs();
8245 addSuccessorWithProb(CurMBB, Fallthrough, FallthroughProb);
8246 addSuccessorWithProb(CurMBB, JumpMBB, JumpProb);
8247 CurMBB->normalizeSuccProbs();
8249 // The jump table header will be inserted in our current block, do the
8250 // range check, and fall through to our fallthrough block.
8251 JTH->HeaderBB = CurMBB;
8252 JT->Default = Fallthrough; // FIXME: Move Default to JumpTableHeader.
8254 // If we're in the right place, emit the jump table header right now.
8255 if (CurMBB == SwitchMBB) {
8256 visitJumpTableHeader(*JT, *JTH, SwitchMBB);
8257 JTH->Emitted = true;
8262 // FIXME: Optimize away range check based on pivot comparisons.
8263 BitTestBlock *BTB = &BitTestCases[I->BTCasesIndex];
8265 // The bit test blocks haven't been inserted yet; insert them here.
8266 for (BitTestCase &BTC : BTB->Cases)
8267 CurMF->insert(BBI, BTC.ThisBB);
8269 // Fill in fields of the BitTestBlock.
8270 BTB->Parent = CurMBB;
8271 BTB->Default = Fallthrough;
8273 BTB->DefaultProb = UnhandledProbs;
8274 // If the cases in bit test don't form a contiguous range, we evenly
8275 // distribute the probability on the edge to Fallthrough to two
8276 // successors of CurMBB.
8277 if (!BTB->ContiguousRange) {
8278 BTB->Prob += DefaultProb / 2;
8279 BTB->DefaultProb -= DefaultProb / 2;
8282 // If we're in the right place, emit the bit test header right now.
8283 if (CurMBB == SwitchMBB) {
8284 visitBitTestHeader(*BTB, SwitchMBB);
8285 BTB->Emitted = true;
8290 const Value *RHS, *LHS, *MHS;
8292 if (I->Low == I->High) {
8293 // Check Cond == I->Low.
8299 // Check I->Low <= Cond <= I->High.
8306 // The false probability is the sum of all unhandled cases.
8307 CaseBlock CB(CC, LHS, RHS, MHS, I->MBB, Fallthrough, CurMBB, I->Prob,
8310 if (CurMBB == SwitchMBB)
8311 visitSwitchCase(CB, SwitchMBB);
8313 SwitchCases.push_back(CB);
8318 CurMBB = Fallthrough;
8322 unsigned SelectionDAGBuilder::caseClusterRank(const CaseCluster &CC,
8323 CaseClusterIt First,
8324 CaseClusterIt Last) {
8325 return std::count_if(First, Last + 1, [&](const CaseCluster &X) {
8326 if (X.Prob != CC.Prob)
8327 return X.Prob > CC.Prob;
8329 // Ties are broken by comparing the case value.
8330 return X.Low->getValue().slt(CC.Low->getValue());
8334 void SelectionDAGBuilder::splitWorkItem(SwitchWorkList &WorkList,
8335 const SwitchWorkListItem &W,
8337 MachineBasicBlock *SwitchMBB) {
8338 assert(W.FirstCluster->Low->getValue().slt(W.LastCluster->Low->getValue()) &&
8339 "Clusters not sorted?");
8341 assert(W.LastCluster - W.FirstCluster + 1 >= 2 && "Too small to split!");
8343 // Balance the tree based on branch probabilities to create a near-optimal (in
8344 // terms of search time given key frequency) binary search tree. See e.g. Kurt
8345 // Mehlhorn "Nearly Optimal Binary Search Trees" (1975).
8346 CaseClusterIt LastLeft = W.FirstCluster;
8347 CaseClusterIt FirstRight = W.LastCluster;
8348 auto LeftProb = LastLeft->Prob + W.DefaultProb / 2;
8349 auto RightProb = FirstRight->Prob + W.DefaultProb / 2;
8351 // Move LastLeft and FirstRight towards each other from opposite directions to
8352 // find a partitioning of the clusters which balances the probability on both
8353 // sides. If LeftProb and RightProb are equal, alternate which side is
8354 // taken to ensure 0-probability nodes are distributed evenly.
8356 while (LastLeft + 1 < FirstRight) {
8357 if (LeftProb < RightProb || (LeftProb == RightProb && (I & 1)))
8358 LeftProb += (++LastLeft)->Prob;
8360 RightProb += (--FirstRight)->Prob;
8365 // Our binary search tree differs from a typical BST in that ours can have up
8366 // to three values in each leaf. The pivot selection above doesn't take that
8367 // into account, which means the tree might require more nodes and be less
8368 // efficient. We compensate for this here.
8370 unsigned NumLeft = LastLeft - W.FirstCluster + 1;
8371 unsigned NumRight = W.LastCluster - FirstRight + 1;
8373 if (std::min(NumLeft, NumRight) < 3 && std::max(NumLeft, NumRight) > 3) {
8374 // If one side has less than 3 clusters, and the other has more than 3,
8375 // consider taking a cluster from the other side.
8377 if (NumLeft < NumRight) {
8378 // Consider moving the first cluster on the right to the left side.
8379 CaseCluster &CC = *FirstRight;
8380 unsigned RightSideRank = caseClusterRank(CC, FirstRight, W.LastCluster);
8381 unsigned LeftSideRank = caseClusterRank(CC, W.FirstCluster, LastLeft);
8382 if (LeftSideRank <= RightSideRank) {
8383 // Moving the cluster to the left does not demote it.
8389 assert(NumRight < NumLeft);
8390 // Consider moving the last element on the left to the right side.
8391 CaseCluster &CC = *LastLeft;
8392 unsigned LeftSideRank = caseClusterRank(CC, W.FirstCluster, LastLeft);
8393 unsigned RightSideRank = caseClusterRank(CC, FirstRight, W.LastCluster);
8394 if (RightSideRank <= LeftSideRank) {
8395 // Moving the cluster to the right does not demot it.
8405 assert(LastLeft + 1 == FirstRight);
8406 assert(LastLeft >= W.FirstCluster);
8407 assert(FirstRight <= W.LastCluster);
8409 // Use the first element on the right as pivot since we will make less-than
8410 // comparisons against it.
8411 CaseClusterIt PivotCluster = FirstRight;
8412 assert(PivotCluster > W.FirstCluster);
8413 assert(PivotCluster <= W.LastCluster);
8415 CaseClusterIt FirstLeft = W.FirstCluster;
8416 CaseClusterIt LastRight = W.LastCluster;
8418 const ConstantInt *Pivot = PivotCluster->Low;
8420 // New blocks will be inserted immediately after the current one.
8421 MachineFunction::iterator BBI(W.MBB);
8424 // We will branch to the LHS if Value < Pivot. If LHS is a single cluster,
8425 // we can branch to its destination directly if it's squeezed exactly in
8426 // between the known lower bound and Pivot - 1.
8427 MachineBasicBlock *LeftMBB;
8428 if (FirstLeft == LastLeft && FirstLeft->Kind == CC_Range &&
8429 FirstLeft->Low == W.GE &&
8430 (FirstLeft->High->getValue() + 1LL) == Pivot->getValue()) {
8431 LeftMBB = FirstLeft->MBB;
8433 LeftMBB = FuncInfo.MF->CreateMachineBasicBlock(W.MBB->getBasicBlock());
8434 FuncInfo.MF->insert(BBI, LeftMBB);
8436 {LeftMBB, FirstLeft, LastLeft, W.GE, Pivot, W.DefaultProb / 2});
8437 // Put Cond in a virtual register to make it available from the new blocks.
8438 ExportFromCurrentBlock(Cond);
8441 // Similarly, we will branch to the RHS if Value >= Pivot. If RHS is a
8442 // single cluster, RHS.Low == Pivot, and we can branch to its destination
8443 // directly if RHS.High equals the current upper bound.
8444 MachineBasicBlock *RightMBB;
8445 if (FirstRight == LastRight && FirstRight->Kind == CC_Range &&
8446 W.LT && (FirstRight->High->getValue() + 1ULL) == W.LT->getValue()) {
8447 RightMBB = FirstRight->MBB;
8449 RightMBB = FuncInfo.MF->CreateMachineBasicBlock(W.MBB->getBasicBlock());
8450 FuncInfo.MF->insert(BBI, RightMBB);
8452 {RightMBB, FirstRight, LastRight, Pivot, W.LT, W.DefaultProb / 2});
8453 // Put Cond in a virtual register to make it available from the new blocks.
8454 ExportFromCurrentBlock(Cond);
8457 // Create the CaseBlock record that will be used to lower the branch.
8458 CaseBlock CB(ISD::SETLT, Cond, Pivot, nullptr, LeftMBB, RightMBB, W.MBB,
8459 LeftProb, RightProb);
8461 if (W.MBB == SwitchMBB)
8462 visitSwitchCase(CB, SwitchMBB);
8464 SwitchCases.push_back(CB);
8467 void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) {
8468 // Extract cases from the switch.
8469 BranchProbabilityInfo *BPI = FuncInfo.BPI;
8470 CaseClusterVector Clusters;
8471 Clusters.reserve(SI.getNumCases());
8472 for (auto I : SI.cases()) {
8473 MachineBasicBlock *Succ = FuncInfo.MBBMap[I.getCaseSuccessor()];
8474 const ConstantInt *CaseVal = I.getCaseValue();
8475 BranchProbability Prob =
8476 BPI ? BPI->getEdgeProbability(SI.getParent(), I.getSuccessorIndex())
8477 : BranchProbability(1, SI.getNumCases() + 1);
8478 Clusters.push_back(CaseCluster::range(CaseVal, CaseVal, Succ, Prob));
8481 MachineBasicBlock *DefaultMBB = FuncInfo.MBBMap[SI.getDefaultDest()];
8483 // Cluster adjacent cases with the same destination. We do this at all
8484 // optimization levels because it's cheap to do and will make codegen faster
8485 // if there are many clusters.
8486 sortAndRangeify(Clusters);
8488 if (TM.getOptLevel() != CodeGenOpt::None) {
8489 // Replace an unreachable default with the most popular destination.
8490 // FIXME: Exploit unreachable default more aggressively.
8491 bool UnreachableDefault =
8492 isa<UnreachableInst>(SI.getDefaultDest()->getFirstNonPHIOrDbg());
8493 if (UnreachableDefault && !Clusters.empty()) {
8494 DenseMap<const BasicBlock *, unsigned> Popularity;
8495 unsigned MaxPop = 0;
8496 const BasicBlock *MaxBB = nullptr;
8497 for (auto I : SI.cases()) {
8498 const BasicBlock *BB = I.getCaseSuccessor();
8499 if (++Popularity[BB] > MaxPop) {
8500 MaxPop = Popularity[BB];
8505 assert(MaxPop > 0 && MaxBB);
8506 DefaultMBB = FuncInfo.MBBMap[MaxBB];
8508 // Remove cases that were pointing to the destination that is now the
8510 CaseClusterVector New;
8511 New.reserve(Clusters.size());
8512 for (CaseCluster &CC : Clusters) {
8513 if (CC.MBB != DefaultMBB)
8516 Clusters = std::move(New);
8520 // If there is only the default destination, jump there directly.
8521 MachineBasicBlock *SwitchMBB = FuncInfo.MBB;
8522 if (Clusters.empty()) {
8523 SwitchMBB->addSuccessor(DefaultMBB);
8524 if (DefaultMBB != NextBlock(SwitchMBB)) {
8525 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other,
8526 getControlRoot(), DAG.getBasicBlock(DefaultMBB)));
8531 findJumpTables(Clusters, &SI, DefaultMBB);
8532 findBitTestClusters(Clusters, &SI);
8535 dbgs() << "Case clusters: ";
8536 for (const CaseCluster &C : Clusters) {
8537 if (C.Kind == CC_JumpTable) dbgs() << "JT:";
8538 if (C.Kind == CC_BitTests) dbgs() << "BT:";
8540 C.Low->getValue().print(dbgs(), true);
8541 if (C.Low != C.High) {
8543 C.High->getValue().print(dbgs(), true);
8550 assert(!Clusters.empty());
8551 SwitchWorkList WorkList;
8552 CaseClusterIt First = Clusters.begin();
8553 CaseClusterIt Last = Clusters.end() - 1;
8554 auto DefaultProb = getEdgeProbability(SwitchMBB, DefaultMBB);
8555 WorkList.push_back({SwitchMBB, First, Last, nullptr, nullptr, DefaultProb});
8557 while (!WorkList.empty()) {
8558 SwitchWorkListItem W = WorkList.back();
8559 WorkList.pop_back();
8560 unsigned NumClusters = W.LastCluster - W.FirstCluster + 1;
8562 if (NumClusters > 3 && TM.getOptLevel() != CodeGenOpt::None) {
8563 // For optimized builds, lower large range as a balanced binary tree.
8564 splitWorkItem(WorkList, W, SI.getCondition(), SwitchMBB);
8568 lowerWorkItem(W, SI.getCondition(), SwitchMBB, DefaultMBB);