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 #define DEBUG_TYPE "isel"
15 #include "SelectionDAGBuilder.h"
16 #include "SDNodeDbgValue.h"
17 #include "llvm/ADT/BitVector.h"
18 #include "llvm/ADT/Optional.h"
19 #include "llvm/ADT/SmallSet.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/Analysis/BranchProbabilityInfo.h"
22 #include "llvm/Analysis/ConstantFolding.h"
23 #include "llvm/Analysis/ValueTracking.h"
24 #include "llvm/CodeGen/Analysis.h"
25 #include "llvm/CodeGen/FastISel.h"
26 #include "llvm/CodeGen/FunctionLoweringInfo.h"
27 #include "llvm/CodeGen/GCMetadata.h"
28 #include "llvm/CodeGen/GCStrategy.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineJumpTableInfo.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/CodeGen/SelectionDAG.h"
36 #include "llvm/DebugInfo.h"
37 #include "llvm/IR/CallingConv.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/DerivedTypes.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/GlobalVariable.h"
43 #include "llvm/IR/InlineAsm.h"
44 #include "llvm/IR/Instructions.h"
45 #include "llvm/IR/IntrinsicInst.h"
46 #include "llvm/IR/Intrinsics.h"
47 #include "llvm/IR/LLVMContext.h"
48 #include "llvm/IR/Module.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/IntegersSubsetMapping.h"
53 #include "llvm/Support/MathExtras.h"
54 #include "llvm/Support/raw_ostream.h"
55 #include "llvm/Target/TargetFrameLowering.h"
56 #include "llvm/Target/TargetInstrInfo.h"
57 #include "llvm/Target/TargetIntrinsicInfo.h"
58 #include "llvm/Target/TargetLibraryInfo.h"
59 #include "llvm/Target/TargetLowering.h"
60 #include "llvm/Target/TargetOptions.h"
61 #include "llvm/Target/TargetSelectionDAGInfo.h"
65 /// LimitFloatPrecision - Generate low-precision inline sequences for
66 /// some float libcalls (6, 8 or 12 bits).
67 static unsigned LimitFloatPrecision;
69 static cl::opt<unsigned, true>
70 LimitFPPrecision("limit-float-precision",
71 cl::desc("Generate low-precision inline sequences "
72 "for some float libcalls"),
73 cl::location(LimitFloatPrecision),
76 // Limit the width of DAG chains. This is important in general to prevent
77 // prevent DAG-based analysis from blowing up. For example, alias analysis and
78 // load clustering may not complete in reasonable time. It is difficult to
79 // recognize and avoid this situation within each individual analysis, and
80 // future analyses are likely to have the same behavior. Limiting DAG width is
81 // the safe approach, and will be especially important with global DAGs.
83 // MaxParallelChains default is arbitrarily high to avoid affecting
84 // optimization, but could be lowered to improve compile time. Any ld-ld-st-st
85 // sequence over this should have been converted to llvm.memcpy by the
86 // frontend. It easy to induce this behavior with .ll code such as:
87 // %buffer = alloca [4096 x i8]
88 // %data = load [4096 x i8]* %argPtr
89 // store [4096 x i8] %data, [4096 x i8]* %buffer
90 static const unsigned MaxParallelChains = 64;
92 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, SDLoc DL,
93 const SDValue *Parts, unsigned NumParts,
94 MVT PartVT, EVT ValueVT, const Value *V);
96 /// getCopyFromParts - Create a value that contains the specified legal parts
97 /// combined into the value they represent. If the parts combine to a type
98 /// larger then ValueVT then AssertOp can be used to specify whether the extra
99 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
100 /// (ISD::AssertSext).
101 static SDValue getCopyFromParts(SelectionDAG &DAG, SDLoc DL,
102 const SDValue *Parts,
103 unsigned NumParts, MVT PartVT, EVT ValueVT,
105 ISD::NodeType AssertOp = ISD::DELETED_NODE) {
106 if (ValueVT.isVector())
107 return getCopyFromPartsVector(DAG, DL, Parts, NumParts,
110 assert(NumParts > 0 && "No parts to assemble!");
111 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
112 SDValue Val = Parts[0];
115 // Assemble the value from multiple parts.
116 if (ValueVT.isInteger()) {
117 unsigned PartBits = PartVT.getSizeInBits();
118 unsigned ValueBits = ValueVT.getSizeInBits();
120 // Assemble the power of 2 part.
121 unsigned RoundParts = NumParts & (NumParts - 1) ?
122 1 << Log2_32(NumParts) : NumParts;
123 unsigned RoundBits = PartBits * RoundParts;
124 EVT RoundVT = RoundBits == ValueBits ?
125 ValueVT : EVT::getIntegerVT(*DAG.getContext(), RoundBits);
128 EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2);
130 if (RoundParts > 2) {
131 Lo = getCopyFromParts(DAG, DL, Parts, RoundParts / 2,
133 Hi = getCopyFromParts(DAG, DL, Parts + RoundParts / 2,
134 RoundParts / 2, PartVT, HalfVT, V);
136 Lo = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[0]);
137 Hi = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[1]);
140 if (TLI.isBigEndian())
143 Val = DAG.getNode(ISD::BUILD_PAIR, DL, RoundVT, Lo, Hi);
145 if (RoundParts < NumParts) {
146 // Assemble the trailing non-power-of-2 part.
147 unsigned OddParts = NumParts - RoundParts;
148 EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits);
149 Hi = getCopyFromParts(DAG, DL,
150 Parts + RoundParts, OddParts, PartVT, OddVT, V);
152 // Combine the round and odd parts.
154 if (TLI.isBigEndian())
156 EVT TotalVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
157 Hi = DAG.getNode(ISD::ANY_EXTEND, DL, TotalVT, Hi);
158 Hi = DAG.getNode(ISD::SHL, DL, TotalVT, Hi,
159 DAG.getConstant(Lo.getValueType().getSizeInBits(),
160 TLI.getPointerTy()));
161 Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, TotalVT, Lo);
162 Val = DAG.getNode(ISD::OR, DL, TotalVT, Lo, Hi);
164 } else if (PartVT.isFloatingPoint()) {
165 // FP split into multiple FP parts (for ppcf128)
166 assert(ValueVT == EVT(MVT::ppcf128) && PartVT == MVT::f64 &&
169 Lo = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[0]);
170 Hi = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[1]);
171 if (TLI.isBigEndian())
173 Val = DAG.getNode(ISD::BUILD_PAIR, DL, ValueVT, Lo, Hi);
175 // FP split into integer parts (soft fp)
176 assert(ValueVT.isFloatingPoint() && PartVT.isInteger() &&
177 !PartVT.isVector() && "Unexpected split");
178 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits());
179 Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT, V);
183 // There is now one part, held in Val. Correct it to match ValueVT.
184 EVT PartEVT = Val.getValueType();
186 if (PartEVT == ValueVT)
189 if (PartEVT.isInteger() && ValueVT.isInteger()) {
190 if (ValueVT.bitsLT(PartEVT)) {
191 // For a truncate, see if we have any information to
192 // indicate whether the truncated bits will always be
193 // zero or sign-extension.
194 if (AssertOp != ISD::DELETED_NODE)
195 Val = DAG.getNode(AssertOp, DL, PartEVT, Val,
196 DAG.getValueType(ValueVT));
197 return DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
199 return DAG.getNode(ISD::ANY_EXTEND, DL, ValueVT, Val);
202 if (PartEVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
203 // FP_ROUND's are always exact here.
204 if (ValueVT.bitsLT(Val.getValueType()))
205 return DAG.getNode(ISD::FP_ROUND, DL, ValueVT, Val,
206 DAG.getTargetConstant(1, TLI.getPointerTy()));
208 return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val);
211 if (PartEVT.getSizeInBits() == ValueVT.getSizeInBits())
212 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
214 llvm_unreachable("Unknown mismatch!");
217 /// getCopyFromPartsVector - Create a value that contains the specified legal
218 /// parts combined into the value they represent. If the parts combine to a
219 /// type larger then ValueVT then AssertOp can be used to specify whether the
220 /// extra bits are known to be zero (ISD::AssertZext) or sign extended from
221 /// ValueVT (ISD::AssertSext).
222 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, SDLoc DL,
223 const SDValue *Parts, unsigned NumParts,
224 MVT PartVT, EVT ValueVT, const Value *V) {
225 assert(ValueVT.isVector() && "Not a vector value");
226 assert(NumParts > 0 && "No parts to assemble!");
227 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
228 SDValue Val = Parts[0];
230 // Handle a multi-element vector.
234 unsigned NumIntermediates;
236 TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT,
237 NumIntermediates, RegisterVT);
238 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
239 NumParts = NumRegs; // Silence a compiler warning.
240 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
241 assert(RegisterVT == Parts[0].getSimpleValueType() &&
242 "Part type doesn't match part!");
244 // Assemble the parts into intermediate operands.
245 SmallVector<SDValue, 8> Ops(NumIntermediates);
246 if (NumIntermediates == NumParts) {
247 // If the register was not expanded, truncate or copy the value,
249 for (unsigned i = 0; i != NumParts; ++i)
250 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i], 1,
251 PartVT, IntermediateVT, V);
252 } else if (NumParts > 0) {
253 // If the intermediate type was expanded, build the intermediate
254 // operands from the parts.
255 assert(NumParts % NumIntermediates == 0 &&
256 "Must expand into a divisible number of parts!");
257 unsigned Factor = NumParts / NumIntermediates;
258 for (unsigned i = 0; i != NumIntermediates; ++i)
259 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor,
260 PartVT, IntermediateVT, V);
263 // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the
264 // intermediate operands.
265 Val = DAG.getNode(IntermediateVT.isVector() ?
266 ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR, DL,
267 ValueVT, &Ops[0], NumIntermediates);
270 // There is now one part, held in Val. Correct it to match ValueVT.
271 EVT PartEVT = Val.getValueType();
273 if (PartEVT == ValueVT)
276 if (PartEVT.isVector()) {
277 // If the element type of the source/dest vectors are the same, but the
278 // parts vector has more elements than the value vector, then we have a
279 // vector widening case (e.g. <2 x float> -> <4 x float>). Extract the
281 if (PartEVT.getVectorElementType() == ValueVT.getVectorElementType()) {
282 assert(PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements() &&
283 "Cannot narrow, it would be a lossy transformation");
284 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val,
285 DAG.getConstant(0, TLI.getVectorIdxTy()));
288 // Vector/Vector bitcast.
289 if (ValueVT.getSizeInBits() == PartEVT.getSizeInBits())
290 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
292 assert(PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements() &&
293 "Cannot handle this kind of promotion");
294 // Promoted vector extract
295 bool Smaller = ValueVT.bitsLE(PartEVT);
296 return DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
301 // Trivial bitcast if the types are the same size and the destination
302 // vector type is legal.
303 if (PartEVT.getSizeInBits() == ValueVT.getSizeInBits() &&
304 TLI.isTypeLegal(ValueVT))
305 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
307 // Handle cases such as i8 -> <1 x i1>
308 if (ValueVT.getVectorNumElements() != 1) {
309 LLVMContext &Ctx = *DAG.getContext();
310 Twine ErrMsg("non-trivial scalar-to-vector conversion");
311 if (const Instruction *I = dyn_cast_or_null<Instruction>(V)) {
312 if (const CallInst *CI = dyn_cast<CallInst>(I))
313 if (isa<InlineAsm>(CI->getCalledValue()))
314 ErrMsg = ErrMsg + ", possible invalid constraint for vector type";
315 Ctx.emitError(I, ErrMsg);
317 Ctx.emitError(ErrMsg);
319 return DAG.getUNDEF(ValueVT);
322 if (ValueVT.getVectorNumElements() == 1 &&
323 ValueVT.getVectorElementType() != PartEVT) {
324 bool Smaller = ValueVT.bitsLE(PartEVT);
325 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
326 DL, ValueVT.getScalarType(), Val);
329 return DAG.getNode(ISD::BUILD_VECTOR, DL, ValueVT, Val);
332 static void getCopyToPartsVector(SelectionDAG &DAG, SDLoc dl,
333 SDValue Val, SDValue *Parts, unsigned NumParts,
334 MVT PartVT, const Value *V);
336 /// getCopyToParts - Create a series of nodes that contain the specified value
337 /// split into legal parts. If the parts contain more bits than Val, then, for
338 /// integers, ExtendKind can be used to specify how to generate the extra bits.
339 static void getCopyToParts(SelectionDAG &DAG, SDLoc DL,
340 SDValue Val, SDValue *Parts, unsigned NumParts,
341 MVT PartVT, const Value *V,
342 ISD::NodeType ExtendKind = ISD::ANY_EXTEND) {
343 EVT ValueVT = Val.getValueType();
345 // Handle the vector case separately.
346 if (ValueVT.isVector())
347 return getCopyToPartsVector(DAG, DL, Val, Parts, NumParts, PartVT, V);
349 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
350 unsigned PartBits = PartVT.getSizeInBits();
351 unsigned OrigNumParts = NumParts;
352 assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!");
357 assert(!ValueVT.isVector() && "Vector case handled elsewhere");
358 EVT PartEVT = PartVT;
359 if (PartEVT == ValueVT) {
360 assert(NumParts == 1 && "No-op copy with multiple parts!");
365 if (NumParts * PartBits > ValueVT.getSizeInBits()) {
366 // If the parts cover more bits than the value has, promote the value.
367 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
368 assert(NumParts == 1 && "Do not know what to promote to!");
369 Val = DAG.getNode(ISD::FP_EXTEND, DL, PartVT, Val);
371 assert((PartVT.isInteger() || PartVT == MVT::x86mmx) &&
372 ValueVT.isInteger() &&
373 "Unknown mismatch!");
374 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
375 Val = DAG.getNode(ExtendKind, DL, ValueVT, Val);
376 if (PartVT == MVT::x86mmx)
377 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
379 } else if (PartBits == ValueVT.getSizeInBits()) {
380 // Different types of the same size.
381 assert(NumParts == 1 && PartEVT != ValueVT);
382 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
383 } else if (NumParts * PartBits < ValueVT.getSizeInBits()) {
384 // If the parts cover less bits than value has, truncate the value.
385 assert((PartVT.isInteger() || PartVT == MVT::x86mmx) &&
386 ValueVT.isInteger() &&
387 "Unknown mismatch!");
388 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
389 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
390 if (PartVT == MVT::x86mmx)
391 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
394 // The value may have changed - recompute ValueVT.
395 ValueVT = Val.getValueType();
396 assert(NumParts * PartBits == ValueVT.getSizeInBits() &&
397 "Failed to tile the value with PartVT!");
400 if (PartEVT != ValueVT) {
401 LLVMContext &Ctx = *DAG.getContext();
402 Twine ErrMsg("scalar-to-vector conversion failed");
403 if (const Instruction *I = dyn_cast_or_null<Instruction>(V)) {
404 if (const CallInst *CI = dyn_cast<CallInst>(I))
405 if (isa<InlineAsm>(CI->getCalledValue()))
406 ErrMsg = ErrMsg + ", possible invalid constraint for vector type";
407 Ctx.emitError(I, ErrMsg);
409 Ctx.emitError(ErrMsg);
417 // Expand the value into multiple parts.
418 if (NumParts & (NumParts - 1)) {
419 // The number of parts is not a power of 2. Split off and copy the tail.
420 assert(PartVT.isInteger() && ValueVT.isInteger() &&
421 "Do not know what to expand to!");
422 unsigned RoundParts = 1 << Log2_32(NumParts);
423 unsigned RoundBits = RoundParts * PartBits;
424 unsigned OddParts = NumParts - RoundParts;
425 SDValue OddVal = DAG.getNode(ISD::SRL, DL, ValueVT, Val,
426 DAG.getIntPtrConstant(RoundBits));
427 getCopyToParts(DAG, DL, OddVal, Parts + RoundParts, OddParts, PartVT, V);
429 if (TLI.isBigEndian())
430 // The odd parts were reversed by getCopyToParts - unreverse them.
431 std::reverse(Parts + RoundParts, Parts + NumParts);
433 NumParts = RoundParts;
434 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
435 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
438 // The number of parts is a power of 2. Repeatedly bisect the value using
440 Parts[0] = DAG.getNode(ISD::BITCAST, DL,
441 EVT::getIntegerVT(*DAG.getContext(),
442 ValueVT.getSizeInBits()),
445 for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) {
446 for (unsigned i = 0; i < NumParts; i += StepSize) {
447 unsigned ThisBits = StepSize * PartBits / 2;
448 EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits);
449 SDValue &Part0 = Parts[i];
450 SDValue &Part1 = Parts[i+StepSize/2];
452 Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
453 ThisVT, Part0, DAG.getIntPtrConstant(1));
454 Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
455 ThisVT, Part0, DAG.getIntPtrConstant(0));
457 if (ThisBits == PartBits && ThisVT != PartVT) {
458 Part0 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part0);
459 Part1 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part1);
464 if (TLI.isBigEndian())
465 std::reverse(Parts, Parts + OrigNumParts);
469 /// getCopyToPartsVector - Create a series of nodes that contain the specified
470 /// value split into legal parts.
471 static void getCopyToPartsVector(SelectionDAG &DAG, SDLoc DL,
472 SDValue Val, SDValue *Parts, unsigned NumParts,
473 MVT PartVT, const Value *V) {
474 EVT ValueVT = Val.getValueType();
475 assert(ValueVT.isVector() && "Not a vector");
476 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
479 EVT PartEVT = PartVT;
480 if (PartEVT == ValueVT) {
482 } else if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) {
483 // Bitconvert vector->vector case.
484 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
485 } else if (PartVT.isVector() &&
486 PartEVT.getVectorElementType() == ValueVT.getVectorElementType() &&
487 PartEVT.getVectorNumElements() > ValueVT.getVectorNumElements()) {
488 EVT ElementVT = PartVT.getVectorElementType();
489 // Vector widening case, e.g. <2 x float> -> <4 x float>. Shuffle in
491 SmallVector<SDValue, 16> Ops;
492 for (unsigned i = 0, e = ValueVT.getVectorNumElements(); i != e; ++i)
493 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
494 ElementVT, Val, DAG.getConstant(i,
495 TLI.getVectorIdxTy())));
497 for (unsigned i = ValueVT.getVectorNumElements(),
498 e = PartVT.getVectorNumElements(); i != e; ++i)
499 Ops.push_back(DAG.getUNDEF(ElementVT));
501 Val = DAG.getNode(ISD::BUILD_VECTOR, DL, PartVT, &Ops[0], Ops.size());
503 // FIXME: Use CONCAT for 2x -> 4x.
505 //SDValue UndefElts = DAG.getUNDEF(VectorTy);
506 //Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts);
507 } else if (PartVT.isVector() &&
508 PartEVT.getVectorElementType().bitsGE(
509 ValueVT.getVectorElementType()) &&
510 PartEVT.getVectorNumElements() == ValueVT.getVectorNumElements()) {
512 // Promoted vector extract
513 bool Smaller = PartEVT.bitsLE(ValueVT);
514 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
517 // Vector -> scalar conversion.
518 assert(ValueVT.getVectorNumElements() == 1 &&
519 "Only trivial vector-to-scalar conversions should get here!");
520 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
521 PartVT, Val, DAG.getConstant(0, TLI.getVectorIdxTy()));
523 bool Smaller = ValueVT.bitsLE(PartVT);
524 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
532 // Handle a multi-element vector.
535 unsigned NumIntermediates;
536 unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT,
538 NumIntermediates, RegisterVT);
539 unsigned NumElements = ValueVT.getVectorNumElements();
541 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
542 NumParts = NumRegs; // Silence a compiler warning.
543 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
545 // Split the vector into intermediate operands.
546 SmallVector<SDValue, 8> Ops(NumIntermediates);
547 for (unsigned i = 0; i != NumIntermediates; ++i) {
548 if (IntermediateVT.isVector())
549 Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL,
551 DAG.getConstant(i * (NumElements / NumIntermediates),
552 TLI.getVectorIdxTy()));
554 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
556 DAG.getConstant(i, TLI.getVectorIdxTy()));
559 // Split the intermediate operands into legal parts.
560 if (NumParts == NumIntermediates) {
561 // If the register was not expanded, promote or copy the value,
563 for (unsigned i = 0; i != NumParts; ++i)
564 getCopyToParts(DAG, DL, Ops[i], &Parts[i], 1, PartVT, V);
565 } else if (NumParts > 0) {
566 // If the intermediate type was expanded, split each the value into
568 assert(NumParts % NumIntermediates == 0 &&
569 "Must expand into a divisible number of parts!");
570 unsigned Factor = NumParts / NumIntermediates;
571 for (unsigned i = 0; i != NumIntermediates; ++i)
572 getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT, V);
577 /// RegsForValue - This struct represents the registers (physical or virtual)
578 /// that a particular set of values is assigned, and the type information
579 /// about the value. The most common situation is to represent one value at a
580 /// time, but struct or array values are handled element-wise as multiple
581 /// values. The splitting of aggregates is performed recursively, so that we
582 /// never have aggregate-typed registers. The values at this point do not
583 /// necessarily have legal types, so each value may require one or more
584 /// registers of some legal type.
586 struct RegsForValue {
587 /// ValueVTs - The value types of the values, which may not be legal, and
588 /// may need be promoted or synthesized from one or more registers.
590 SmallVector<EVT, 4> ValueVTs;
592 /// RegVTs - The value types of the registers. This is the same size as
593 /// ValueVTs and it records, for each value, what the type of the assigned
594 /// register or registers are. (Individual values are never synthesized
595 /// from more than one type of register.)
597 /// With virtual registers, the contents of RegVTs is redundant with TLI's
598 /// getRegisterType member function, however when with physical registers
599 /// it is necessary to have a separate record of the types.
601 SmallVector<MVT, 4> RegVTs;
603 /// Regs - This list holds the registers assigned to the values.
604 /// Each legal or promoted value requires one register, and each
605 /// expanded value requires multiple registers.
607 SmallVector<unsigned, 4> Regs;
611 RegsForValue(const SmallVector<unsigned, 4> ®s,
612 MVT regvt, EVT valuevt)
613 : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {}
615 RegsForValue(LLVMContext &Context, const TargetLowering &tli,
616 unsigned Reg, Type *Ty) {
617 ComputeValueVTs(tli, Ty, ValueVTs);
619 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
620 EVT ValueVT = ValueVTs[Value];
621 unsigned NumRegs = tli.getNumRegisters(Context, ValueVT);
622 MVT RegisterVT = tli.getRegisterType(Context, ValueVT);
623 for (unsigned i = 0; i != NumRegs; ++i)
624 Regs.push_back(Reg + i);
625 RegVTs.push_back(RegisterVT);
630 /// areValueTypesLegal - Return true if types of all the values are legal.
631 bool areValueTypesLegal(const TargetLowering &TLI) {
632 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
633 MVT RegisterVT = RegVTs[Value];
634 if (!TLI.isTypeLegal(RegisterVT))
640 /// append - Add the specified values to this one.
641 void append(const RegsForValue &RHS) {
642 ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end());
643 RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end());
644 Regs.append(RHS.Regs.begin(), RHS.Regs.end());
647 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
648 /// this value and returns the result as a ValueVTs value. This uses
649 /// Chain/Flag as the input and updates them for the output Chain/Flag.
650 /// If the Flag pointer is NULL, no flag is used.
651 SDValue getCopyFromRegs(SelectionDAG &DAG, FunctionLoweringInfo &FuncInfo,
653 SDValue &Chain, SDValue *Flag,
654 const Value *V = 0) const;
656 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
657 /// specified value into the registers specified by this object. This uses
658 /// Chain/Flag as the input and updates them for the output Chain/Flag.
659 /// If the Flag pointer is NULL, no flag is used.
660 void getCopyToRegs(SDValue Val, SelectionDAG &DAG, SDLoc dl,
661 SDValue &Chain, SDValue *Flag, const Value *V) const;
663 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
664 /// operand list. This adds the code marker, matching input operand index
665 /// (if applicable), and includes the number of values added into it.
666 void AddInlineAsmOperands(unsigned Kind,
667 bool HasMatching, unsigned MatchingIdx,
669 std::vector<SDValue> &Ops) const;
673 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
674 /// this value and returns the result as a ValueVT value. This uses
675 /// Chain/Flag as the input and updates them for the output Chain/Flag.
676 /// If the Flag pointer is NULL, no flag is used.
677 SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
678 FunctionLoweringInfo &FuncInfo,
680 SDValue &Chain, SDValue *Flag,
681 const Value *V) const {
682 // A Value with type {} or [0 x %t] needs no registers.
683 if (ValueVTs.empty())
686 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
688 // Assemble the legal parts into the final values.
689 SmallVector<SDValue, 4> Values(ValueVTs.size());
690 SmallVector<SDValue, 8> Parts;
691 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
692 // Copy the legal parts from the registers.
693 EVT ValueVT = ValueVTs[Value];
694 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
695 MVT RegisterVT = RegVTs[Value];
697 Parts.resize(NumRegs);
698 for (unsigned i = 0; i != NumRegs; ++i) {
701 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT);
703 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag);
704 *Flag = P.getValue(2);
707 Chain = P.getValue(1);
710 // If the source register was virtual and if we know something about it,
711 // add an assert node.
712 if (!TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) ||
713 !RegisterVT.isInteger() || RegisterVT.isVector())
716 const FunctionLoweringInfo::LiveOutInfo *LOI =
717 FuncInfo.GetLiveOutRegInfo(Regs[Part+i]);
721 unsigned RegSize = RegisterVT.getSizeInBits();
722 unsigned NumSignBits = LOI->NumSignBits;
723 unsigned NumZeroBits = LOI->KnownZero.countLeadingOnes();
725 if (NumZeroBits == RegSize) {
726 // The current value is a zero.
727 // Explicitly express that as it would be easier for
728 // optimizations to kick in.
729 Parts[i] = DAG.getConstant(0, RegisterVT);
733 // FIXME: We capture more information than the dag can represent. For
734 // now, just use the tightest assertzext/assertsext possible.
736 EVT FromVT(MVT::Other);
737 if (NumSignBits == RegSize)
738 isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1
739 else if (NumZeroBits >= RegSize-1)
740 isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1
741 else if (NumSignBits > RegSize-8)
742 isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8
743 else if (NumZeroBits >= RegSize-8)
744 isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8
745 else if (NumSignBits > RegSize-16)
746 isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16
747 else if (NumZeroBits >= RegSize-16)
748 isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16
749 else if (NumSignBits > RegSize-32)
750 isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32
751 else if (NumZeroBits >= RegSize-32)
752 isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32
756 // Add an assertion node.
757 assert(FromVT != MVT::Other);
758 Parts[i] = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl,
759 RegisterVT, P, DAG.getValueType(FromVT));
762 Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(),
763 NumRegs, RegisterVT, ValueVT, V);
768 return DAG.getNode(ISD::MERGE_VALUES, dl,
769 DAG.getVTList(&ValueVTs[0], ValueVTs.size()),
770 &Values[0], ValueVTs.size());
773 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
774 /// specified value into the registers specified by this object. This uses
775 /// Chain/Flag as the input and updates them for the output Chain/Flag.
776 /// If the Flag pointer is NULL, no flag is used.
777 void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, SDLoc dl,
778 SDValue &Chain, SDValue *Flag,
779 const Value *V) const {
780 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
782 // Get the list of the values's legal parts.
783 unsigned NumRegs = Regs.size();
784 SmallVector<SDValue, 8> Parts(NumRegs);
785 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
786 EVT ValueVT = ValueVTs[Value];
787 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
788 MVT RegisterVT = RegVTs[Value];
789 ISD::NodeType ExtendKind =
790 TLI.isZExtFree(Val, RegisterVT)? ISD::ZERO_EXTEND: ISD::ANY_EXTEND;
792 getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value),
793 &Parts[Part], NumParts, RegisterVT, V, ExtendKind);
797 // Copy the parts into the registers.
798 SmallVector<SDValue, 8> Chains(NumRegs);
799 for (unsigned i = 0; i != NumRegs; ++i) {
802 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]);
804 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag);
805 *Flag = Part.getValue(1);
808 Chains[i] = Part.getValue(0);
811 if (NumRegs == 1 || Flag)
812 // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is
813 // flagged to it. That is the CopyToReg nodes and the user are considered
814 // a single scheduling unit. If we create a TokenFactor and return it as
815 // chain, then the TokenFactor is both a predecessor (operand) of the
816 // user as well as a successor (the TF operands are flagged to the user).
817 // c1, f1 = CopyToReg
818 // c2, f2 = CopyToReg
819 // c3 = TokenFactor c1, c2
822 Chain = Chains[NumRegs-1];
824 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0], NumRegs);
827 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
828 /// operand list. This adds the code marker and includes the number of
829 /// values added into it.
830 void RegsForValue::AddInlineAsmOperands(unsigned Code, bool HasMatching,
831 unsigned MatchingIdx,
833 std::vector<SDValue> &Ops) const {
834 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
836 unsigned Flag = InlineAsm::getFlagWord(Code, Regs.size());
838 Flag = InlineAsm::getFlagWordForMatchingOp(Flag, MatchingIdx);
839 else if (!Regs.empty() &&
840 TargetRegisterInfo::isVirtualRegister(Regs.front())) {
841 // Put the register class of the virtual registers in the flag word. That
842 // way, later passes can recompute register class constraints for inline
843 // assembly as well as normal instructions.
844 // Don't do this for tied operands that can use the regclass information
846 const MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
847 const TargetRegisterClass *RC = MRI.getRegClass(Regs.front());
848 Flag = InlineAsm::getFlagWordForRegClass(Flag, RC->getID());
851 SDValue Res = DAG.getTargetConstant(Flag, MVT::i32);
854 for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) {
855 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVTs[Value]);
856 MVT RegisterVT = RegVTs[Value];
857 for (unsigned i = 0; i != NumRegs; ++i) {
858 assert(Reg < Regs.size() && "Mismatch in # registers expected");
859 Ops.push_back(DAG.getRegister(Regs[Reg++], RegisterVT));
864 void SelectionDAGBuilder::init(GCFunctionInfo *gfi, AliasAnalysis &aa,
865 const TargetLibraryInfo *li) {
869 TD = DAG.getTarget().getDataLayout();
870 Context = DAG.getContext();
871 LPadToCallSiteMap.clear();
874 /// clear - Clear out the current SelectionDAG and the associated
875 /// state and prepare this SelectionDAGBuilder object to be used
876 /// for a new block. This doesn't clear out information about
877 /// additional blocks that are needed to complete switch lowering
878 /// or PHI node updating; that information is cleared out as it is
880 void SelectionDAGBuilder::clear() {
882 UnusedArgNodeMap.clear();
883 PendingLoads.clear();
884 PendingExports.clear();
889 /// clearDanglingDebugInfo - Clear the dangling debug information
890 /// map. This function is separated from the clear so that debug
891 /// information that is dangling in a basic block can be properly
892 /// resolved in a different basic block. This allows the
893 /// SelectionDAG to resolve dangling debug information attached
895 void SelectionDAGBuilder::clearDanglingDebugInfo() {
896 DanglingDebugInfoMap.clear();
899 /// getRoot - Return the current virtual root of the Selection DAG,
900 /// flushing any PendingLoad items. This must be done before emitting
901 /// a store or any other node that may need to be ordered after any
902 /// prior load instructions.
904 SDValue SelectionDAGBuilder::getRoot() {
905 if (PendingLoads.empty())
906 return DAG.getRoot();
908 if (PendingLoads.size() == 1) {
909 SDValue Root = PendingLoads[0];
911 PendingLoads.clear();
915 // Otherwise, we have to make a token factor node.
916 SDValue Root = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
917 &PendingLoads[0], PendingLoads.size());
918 PendingLoads.clear();
923 /// getControlRoot - Similar to getRoot, but instead of flushing all the
924 /// PendingLoad items, flush all the PendingExports items. It is necessary
925 /// to do this before emitting a terminator instruction.
927 SDValue SelectionDAGBuilder::getControlRoot() {
928 SDValue Root = DAG.getRoot();
930 if (PendingExports.empty())
933 // Turn all of the CopyToReg chains into one factored node.
934 if (Root.getOpcode() != ISD::EntryToken) {
935 unsigned i = 0, e = PendingExports.size();
936 for (; i != e; ++i) {
937 assert(PendingExports[i].getNode()->getNumOperands() > 1);
938 if (PendingExports[i].getNode()->getOperand(0) == Root)
939 break; // Don't add the root if we already indirectly depend on it.
943 PendingExports.push_back(Root);
946 Root = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
948 PendingExports.size());
949 PendingExports.clear();
954 void SelectionDAGBuilder::visit(const Instruction &I) {
955 // Set up outgoing PHI node register values before emitting the terminator.
956 if (isa<TerminatorInst>(&I))
957 HandlePHINodesInSuccessorBlocks(I.getParent());
963 visit(I.getOpcode(), I);
965 if (!isa<TerminatorInst>(&I) && !HasTailCall)
966 CopyToExportRegsIfNeeded(&I);
971 void SelectionDAGBuilder::visitPHI(const PHINode &) {
972 llvm_unreachable("SelectionDAGBuilder shouldn't visit PHI nodes!");
975 void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) {
976 // Note: this doesn't use InstVisitor, because it has to work with
977 // ConstantExpr's in addition to instructions.
979 default: llvm_unreachable("Unknown instruction type encountered!");
980 // Build the switch statement using the Instruction.def file.
981 #define HANDLE_INST(NUM, OPCODE, CLASS) \
982 case Instruction::OPCODE: visit##OPCODE((const CLASS&)I); break;
983 #include "llvm/IR/Instruction.def"
987 // resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V,
988 // generate the debug data structures now that we've seen its definition.
989 void SelectionDAGBuilder::resolveDanglingDebugInfo(const Value *V,
991 DanglingDebugInfo &DDI = DanglingDebugInfoMap[V];
993 const DbgValueInst *DI = DDI.getDI();
994 DebugLoc dl = DDI.getdl();
995 unsigned DbgSDNodeOrder = DDI.getSDNodeOrder();
996 MDNode *Variable = DI->getVariable();
997 uint64_t Offset = DI->getOffset();
1000 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, Val)) {
1001 SDV = DAG.getDbgValue(Variable, Val.getNode(),
1002 Val.getResNo(), Offset, dl, DbgSDNodeOrder);
1003 DAG.AddDbgValue(SDV, Val.getNode(), false);
1006 DEBUG(dbgs() << "Dropping debug info for " << *DI << "\n");
1007 DanglingDebugInfoMap[V] = DanglingDebugInfo();
1011 /// getValue - Return an SDValue for the given Value.
1012 SDValue SelectionDAGBuilder::getValue(const Value *V) {
1013 // If we already have an SDValue for this value, use it. It's important
1014 // to do this first, so that we don't create a CopyFromReg if we already
1015 // have a regular SDValue.
1016 SDValue &N = NodeMap[V];
1017 if (N.getNode()) return N;
1019 // If there's a virtual register allocated and initialized for this
1021 DenseMap<const Value *, unsigned>::iterator It = FuncInfo.ValueMap.find(V);
1022 if (It != FuncInfo.ValueMap.end()) {
1023 unsigned InReg = It->second;
1024 RegsForValue RFV(*DAG.getContext(), *TM.getTargetLowering(),
1025 InReg, V->getType());
1026 SDValue Chain = DAG.getEntryNode();
1027 N = RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, NULL, V);
1028 resolveDanglingDebugInfo(V, N);
1032 // Otherwise create a new SDValue and remember it.
1033 SDValue Val = getValueImpl(V);
1035 resolveDanglingDebugInfo(V, Val);
1039 /// getNonRegisterValue - Return an SDValue for the given Value, but
1040 /// don't look in FuncInfo.ValueMap for a virtual register.
1041 SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) {
1042 // If we already have an SDValue for this value, use it.
1043 SDValue &N = NodeMap[V];
1044 if (N.getNode()) return N;
1046 // Otherwise create a new SDValue and remember it.
1047 SDValue Val = getValueImpl(V);
1049 resolveDanglingDebugInfo(V, Val);
1053 /// getValueImpl - Helper function for getValue and getNonRegisterValue.
1054 /// Create an SDValue for the given value.
1055 SDValue SelectionDAGBuilder::getValueImpl(const Value *V) {
1056 const TargetLowering *TLI = TM.getTargetLowering();
1058 if (const Constant *C = dyn_cast<Constant>(V)) {
1059 EVT VT = TLI->getValueType(V->getType(), true);
1061 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C))
1062 return DAG.getConstant(*CI, VT);
1064 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
1065 return DAG.getGlobalAddress(GV, getCurSDLoc(), VT);
1067 if (isa<ConstantPointerNull>(C))
1068 return DAG.getConstant(0, TLI->getPointerTy());
1070 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
1071 return DAG.getConstantFP(*CFP, VT);
1073 if (isa<UndefValue>(C) && !V->getType()->isAggregateType())
1074 return DAG.getUNDEF(VT);
1076 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
1077 visit(CE->getOpcode(), *CE);
1078 SDValue N1 = NodeMap[V];
1079 assert(N1.getNode() && "visit didn't populate the NodeMap!");
1083 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) {
1084 SmallVector<SDValue, 4> Constants;
1085 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end();
1087 SDNode *Val = getValue(*OI).getNode();
1088 // If the operand is an empty aggregate, there are no values.
1090 // Add each leaf value from the operand to the Constants list
1091 // to form a flattened list of all the values.
1092 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
1093 Constants.push_back(SDValue(Val, i));
1096 return DAG.getMergeValues(&Constants[0], Constants.size(),
1100 if (const ConstantDataSequential *CDS =
1101 dyn_cast<ConstantDataSequential>(C)) {
1102 SmallVector<SDValue, 4> Ops;
1103 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1104 SDNode *Val = getValue(CDS->getElementAsConstant(i)).getNode();
1105 // Add each leaf value from the operand to the Constants list
1106 // to form a flattened list of all the values.
1107 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
1108 Ops.push_back(SDValue(Val, i));
1111 if (isa<ArrayType>(CDS->getType()))
1112 return DAG.getMergeValues(&Ops[0], Ops.size(), getCurSDLoc());
1113 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(),
1114 VT, &Ops[0], Ops.size());
1117 if (C->getType()->isStructTy() || C->getType()->isArrayTy()) {
1118 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
1119 "Unknown struct or array constant!");
1121 SmallVector<EVT, 4> ValueVTs;
1122 ComputeValueVTs(*TLI, C->getType(), ValueVTs);
1123 unsigned NumElts = ValueVTs.size();
1125 return SDValue(); // empty struct
1126 SmallVector<SDValue, 4> Constants(NumElts);
1127 for (unsigned i = 0; i != NumElts; ++i) {
1128 EVT EltVT = ValueVTs[i];
1129 if (isa<UndefValue>(C))
1130 Constants[i] = DAG.getUNDEF(EltVT);
1131 else if (EltVT.isFloatingPoint())
1132 Constants[i] = DAG.getConstantFP(0, EltVT);
1134 Constants[i] = DAG.getConstant(0, EltVT);
1137 return DAG.getMergeValues(&Constants[0], NumElts,
1141 if (const BlockAddress *BA = dyn_cast<BlockAddress>(C))
1142 return DAG.getBlockAddress(BA, VT);
1144 VectorType *VecTy = cast<VectorType>(V->getType());
1145 unsigned NumElements = VecTy->getNumElements();
1147 // Now that we know the number and type of the elements, get that number of
1148 // elements into the Ops array based on what kind of constant it is.
1149 SmallVector<SDValue, 16> Ops;
1150 if (const ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
1151 for (unsigned i = 0; i != NumElements; ++i)
1152 Ops.push_back(getValue(CV->getOperand(i)));
1154 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!");
1155 EVT EltVT = TLI->getValueType(VecTy->getElementType());
1158 if (EltVT.isFloatingPoint())
1159 Op = DAG.getConstantFP(0, EltVT);
1161 Op = DAG.getConstant(0, EltVT);
1162 Ops.assign(NumElements, Op);
1165 // Create a BUILD_VECTOR node.
1166 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(),
1167 VT, &Ops[0], Ops.size());
1170 // If this is a static alloca, generate it as the frameindex instead of
1172 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
1173 DenseMap<const AllocaInst*, int>::iterator SI =
1174 FuncInfo.StaticAllocaMap.find(AI);
1175 if (SI != FuncInfo.StaticAllocaMap.end())
1176 return DAG.getFrameIndex(SI->second, TLI->getPointerTy());
1179 // If this is an instruction which fast-isel has deferred, select it now.
1180 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
1181 unsigned InReg = FuncInfo.InitializeRegForValue(Inst);
1182 RegsForValue RFV(*DAG.getContext(), *TLI, InReg, Inst->getType());
1183 SDValue Chain = DAG.getEntryNode();
1184 return RFV.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(), Chain, NULL, V);
1187 llvm_unreachable("Can't get register for value!");
1190 void SelectionDAGBuilder::visitRet(const ReturnInst &I) {
1191 const TargetLowering *TLI = TM.getTargetLowering();
1192 SDValue Chain = getControlRoot();
1193 SmallVector<ISD::OutputArg, 8> Outs;
1194 SmallVector<SDValue, 8> OutVals;
1196 if (!FuncInfo.CanLowerReturn) {
1197 unsigned DemoteReg = FuncInfo.DemoteRegister;
1198 const Function *F = I.getParent()->getParent();
1200 // Emit a store of the return value through the virtual register.
1201 // Leave Outs empty so that LowerReturn won't try to load return
1202 // registers the usual way.
1203 SmallVector<EVT, 1> PtrValueVTs;
1204 ComputeValueVTs(*TLI, PointerType::getUnqual(F->getReturnType()),
1207 SDValue RetPtr = DAG.getRegister(DemoteReg, PtrValueVTs[0]);
1208 SDValue RetOp = getValue(I.getOperand(0));
1210 SmallVector<EVT, 4> ValueVTs;
1211 SmallVector<uint64_t, 4> Offsets;
1212 ComputeValueVTs(*TLI, I.getOperand(0)->getType(), ValueVTs, &Offsets);
1213 unsigned NumValues = ValueVTs.size();
1215 SmallVector<SDValue, 4> Chains(NumValues);
1216 for (unsigned i = 0; i != NumValues; ++i) {
1217 SDValue Add = DAG.getNode(ISD::ADD, getCurSDLoc(),
1218 RetPtr.getValueType(), RetPtr,
1219 DAG.getIntPtrConstant(Offsets[i]));
1221 DAG.getStore(Chain, getCurSDLoc(),
1222 SDValue(RetOp.getNode(), RetOp.getResNo() + i),
1223 // FIXME: better loc info would be nice.
1224 Add, MachinePointerInfo(), false, false, 0);
1227 Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(),
1228 MVT::Other, &Chains[0], NumValues);
1229 } else if (I.getNumOperands() != 0) {
1230 SmallVector<EVT, 4> ValueVTs;
1231 ComputeValueVTs(*TLI, I.getOperand(0)->getType(), ValueVTs);
1232 unsigned NumValues = ValueVTs.size();
1234 SDValue RetOp = getValue(I.getOperand(0));
1235 for (unsigned j = 0, f = NumValues; j != f; ++j) {
1236 EVT VT = ValueVTs[j];
1238 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1240 const Function *F = I.getParent()->getParent();
1241 if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1243 ExtendKind = ISD::SIGN_EXTEND;
1244 else if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1246 ExtendKind = ISD::ZERO_EXTEND;
1248 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger())
1249 VT = TLI->getTypeForExtArgOrReturn(VT.getSimpleVT(), ExtendKind);
1251 unsigned NumParts = TLI->getNumRegisters(*DAG.getContext(), VT);
1252 MVT PartVT = TLI->getRegisterType(*DAG.getContext(), VT);
1253 SmallVector<SDValue, 4> Parts(NumParts);
1254 getCopyToParts(DAG, getCurSDLoc(),
1255 SDValue(RetOp.getNode(), RetOp.getResNo() + j),
1256 &Parts[0], NumParts, PartVT, &I, ExtendKind);
1258 // 'inreg' on function refers to return value
1259 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1260 if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
1264 // Propagate extension type if any
1265 if (ExtendKind == ISD::SIGN_EXTEND)
1267 else if (ExtendKind == ISD::ZERO_EXTEND)
1270 for (unsigned i = 0; i < NumParts; ++i) {
1271 Outs.push_back(ISD::OutputArg(Flags, Parts[i].getValueType(),
1272 /*isfixed=*/true, 0, 0));
1273 OutVals.push_back(Parts[i]);
1279 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1280 CallingConv::ID CallConv =
1281 DAG.getMachineFunction().getFunction()->getCallingConv();
1282 Chain = TM.getTargetLowering()->LowerReturn(Chain, CallConv, isVarArg,
1283 Outs, OutVals, getCurSDLoc(),
1286 // Verify that the target's LowerReturn behaved as expected.
1287 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
1288 "LowerReturn didn't return a valid chain!");
1290 // Update the DAG with the new chain value resulting from return lowering.
1294 /// CopyToExportRegsIfNeeded - If the given value has virtual registers
1295 /// created for it, emit nodes to copy the value into the virtual
1297 void SelectionDAGBuilder::CopyToExportRegsIfNeeded(const Value *V) {
1299 if (V->getType()->isEmptyTy())
1302 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
1303 if (VMI != FuncInfo.ValueMap.end()) {
1304 assert(!V->use_empty() && "Unused value assigned virtual registers!");
1305 CopyValueToVirtualRegister(V, VMI->second);
1309 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
1310 /// the current basic block, add it to ValueMap now so that we'll get a
1312 void SelectionDAGBuilder::ExportFromCurrentBlock(const Value *V) {
1313 // No need to export constants.
1314 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
1316 // Already exported?
1317 if (FuncInfo.isExportedInst(V)) return;
1319 unsigned Reg = FuncInfo.InitializeRegForValue(V);
1320 CopyValueToVirtualRegister(V, Reg);
1323 bool SelectionDAGBuilder::isExportableFromCurrentBlock(const Value *V,
1324 const BasicBlock *FromBB) {
1325 // The operands of the setcc have to be in this block. We don't know
1326 // how to export them from some other block.
1327 if (const Instruction *VI = dyn_cast<Instruction>(V)) {
1328 // Can export from current BB.
1329 if (VI->getParent() == FromBB)
1332 // Is already exported, noop.
1333 return FuncInfo.isExportedInst(V);
1336 // If this is an argument, we can export it if the BB is the entry block or
1337 // if it is already exported.
1338 if (isa<Argument>(V)) {
1339 if (FromBB == &FromBB->getParent()->getEntryBlock())
1342 // Otherwise, can only export this if it is already exported.
1343 return FuncInfo.isExportedInst(V);
1346 // Otherwise, constants can always be exported.
1350 /// Return branch probability calculated by BranchProbabilityInfo for IR blocks.
1351 uint32_t SelectionDAGBuilder::getEdgeWeight(const MachineBasicBlock *Src,
1352 const MachineBasicBlock *Dst) const {
1353 BranchProbabilityInfo *BPI = FuncInfo.BPI;
1356 const BasicBlock *SrcBB = Src->getBasicBlock();
1357 const BasicBlock *DstBB = Dst->getBasicBlock();
1358 return BPI->getEdgeWeight(SrcBB, DstBB);
1361 void SelectionDAGBuilder::
1362 addSuccessorWithWeight(MachineBasicBlock *Src, MachineBasicBlock *Dst,
1363 uint32_t Weight /* = 0 */) {
1365 Weight = getEdgeWeight(Src, Dst);
1366 Src->addSuccessor(Dst, Weight);
1370 static bool InBlock(const Value *V, const BasicBlock *BB) {
1371 if (const Instruction *I = dyn_cast<Instruction>(V))
1372 return I->getParent() == BB;
1376 /// EmitBranchForMergedCondition - Helper method for FindMergedConditions.
1377 /// This function emits a branch and is used at the leaves of an OR or an
1378 /// AND operator tree.
1381 SelectionDAGBuilder::EmitBranchForMergedCondition(const Value *Cond,
1382 MachineBasicBlock *TBB,
1383 MachineBasicBlock *FBB,
1384 MachineBasicBlock *CurBB,
1385 MachineBasicBlock *SwitchBB) {
1386 const BasicBlock *BB = CurBB->getBasicBlock();
1388 // If the leaf of the tree is a comparison, merge the condition into
1390 if (const CmpInst *BOp = dyn_cast<CmpInst>(Cond)) {
1391 // The operands of the cmp have to be in this block. We don't know
1392 // how to export them from some other block. If this is the first block
1393 // of the sequence, no exporting is needed.
1394 if (CurBB == SwitchBB ||
1395 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
1396 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) {
1397 ISD::CondCode Condition;
1398 if (const ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
1399 Condition = getICmpCondCode(IC->getPredicate());
1400 } else if (const FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) {
1401 Condition = getFCmpCondCode(FC->getPredicate());
1402 if (TM.Options.NoNaNsFPMath)
1403 Condition = getFCmpCodeWithoutNaN(Condition);
1405 Condition = ISD::SETEQ; // silence warning.
1406 llvm_unreachable("Unknown compare instruction");
1409 CaseBlock CB(Condition, BOp->getOperand(0),
1410 BOp->getOperand(1), NULL, TBB, FBB, CurBB);
1411 SwitchCases.push_back(CB);
1416 // Create a CaseBlock record representing this branch.
1417 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()),
1418 NULL, TBB, FBB, CurBB);
1419 SwitchCases.push_back(CB);
1422 /// FindMergedConditions - If Cond is an expression like
1423 void SelectionDAGBuilder::FindMergedConditions(const Value *Cond,
1424 MachineBasicBlock *TBB,
1425 MachineBasicBlock *FBB,
1426 MachineBasicBlock *CurBB,
1427 MachineBasicBlock *SwitchBB,
1429 // If this node is not part of the or/and tree, emit it as a branch.
1430 const Instruction *BOp = dyn_cast<Instruction>(Cond);
1431 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
1432 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
1433 BOp->getParent() != CurBB->getBasicBlock() ||
1434 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
1435 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
1436 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB);
1440 // Create TmpBB after CurBB.
1441 MachineFunction::iterator BBI = CurBB;
1442 MachineFunction &MF = DAG.getMachineFunction();
1443 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock());
1444 CurBB->getParent()->insert(++BBI, TmpBB);
1446 if (Opc == Instruction::Or) {
1447 // Codegen X | Y as:
1455 // Emit the LHS condition.
1456 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, SwitchBB, Opc);
1458 // Emit the RHS condition into TmpBB.
1459 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
1461 assert(Opc == Instruction::And && "Unknown merge op!");
1462 // Codegen X & Y as:
1469 // This requires creation of TmpBB after CurBB.
1471 // Emit the LHS condition.
1472 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, SwitchBB, Opc);
1474 // Emit the RHS condition into TmpBB.
1475 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
1479 /// If the set of cases should be emitted as a series of branches, return true.
1480 /// If we should emit this as a bunch of and/or'd together conditions, return
1483 SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases) {
1484 if (Cases.size() != 2) return true;
1486 // If this is two comparisons of the same values or'd or and'd together, they
1487 // will get folded into a single comparison, so don't emit two blocks.
1488 if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
1489 Cases[0].CmpRHS == Cases[1].CmpRHS) ||
1490 (Cases[0].CmpRHS == Cases[1].CmpLHS &&
1491 Cases[0].CmpLHS == Cases[1].CmpRHS)) {
1495 // Handle: (X != null) | (Y != null) --> (X|Y) != 0
1496 // Handle: (X == null) & (Y == null) --> (X|Y) == 0
1497 if (Cases[0].CmpRHS == Cases[1].CmpRHS &&
1498 Cases[0].CC == Cases[1].CC &&
1499 isa<Constant>(Cases[0].CmpRHS) &&
1500 cast<Constant>(Cases[0].CmpRHS)->isNullValue()) {
1501 if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB)
1503 if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB)
1510 void SelectionDAGBuilder::visitBr(const BranchInst &I) {
1511 MachineBasicBlock *BrMBB = FuncInfo.MBB;
1513 // Update machine-CFG edges.
1514 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
1516 // Figure out which block is immediately after the current one.
1517 MachineBasicBlock *NextBlock = 0;
1518 MachineFunction::iterator BBI = BrMBB;
1519 if (++BBI != FuncInfo.MF->end())
1522 if (I.isUnconditional()) {
1523 // Update machine-CFG edges.
1524 BrMBB->addSuccessor(Succ0MBB);
1526 // If this is not a fall-through branch, emit the branch.
1527 if (Succ0MBB != NextBlock)
1528 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(),
1529 MVT::Other, getControlRoot(),
1530 DAG.getBasicBlock(Succ0MBB)));
1535 // If this condition is one of the special cases we handle, do special stuff
1537 const Value *CondVal = I.getCondition();
1538 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
1540 // If this is a series of conditions that are or'd or and'd together, emit
1541 // this as a sequence of branches instead of setcc's with and/or operations.
1542 // As long as jumps are not expensive, this should improve performance.
1543 // For example, instead of something like:
1556 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
1557 if (!TM.getTargetLowering()->isJumpExpensive() &&
1559 (BOp->getOpcode() == Instruction::And ||
1560 BOp->getOpcode() == Instruction::Or)) {
1561 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB,
1563 // If the compares in later blocks need to use values not currently
1564 // exported from this block, export them now. This block should always
1565 // be the first entry.
1566 assert(SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!");
1568 // Allow some cases to be rejected.
1569 if (ShouldEmitAsBranches(SwitchCases)) {
1570 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1571 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1572 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1575 // Emit the branch for this block.
1576 visitSwitchCase(SwitchCases[0], BrMBB);
1577 SwitchCases.erase(SwitchCases.begin());
1581 // Okay, we decided not to do this, remove any inserted MBB's and clear
1583 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
1584 FuncInfo.MF->erase(SwitchCases[i].ThisBB);
1586 SwitchCases.clear();
1590 // Create a CaseBlock record representing this branch.
1591 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()),
1592 NULL, Succ0MBB, Succ1MBB, BrMBB);
1594 // Use visitSwitchCase to actually insert the fast branch sequence for this
1596 visitSwitchCase(CB, BrMBB);
1599 /// visitSwitchCase - Emits the necessary code to represent a single node in
1600 /// the binary search tree resulting from lowering a switch instruction.
1601 void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB,
1602 MachineBasicBlock *SwitchBB) {
1604 SDValue CondLHS = getValue(CB.CmpLHS);
1605 SDLoc dl = getCurSDLoc();
1607 // Build the setcc now.
1608 if (CB.CmpMHS == NULL) {
1609 // Fold "(X == true)" to X and "(X == false)" to !X to
1610 // handle common cases produced by branch lowering.
1611 if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) &&
1612 CB.CC == ISD::SETEQ)
1614 else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) &&
1615 CB.CC == ISD::SETEQ) {
1616 SDValue True = DAG.getConstant(1, CondLHS.getValueType());
1617 Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True);
1619 Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1621 assert(CB.CC == ISD::SETCC_INVALID &&
1622 "Condition is undefined for to-the-range belonging check.");
1624 const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue();
1625 const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue();
1627 SDValue CmpOp = getValue(CB.CmpMHS);
1628 EVT VT = CmpOp.getValueType();
1630 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(false)) {
1631 Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, VT),
1634 SDValue SUB = DAG.getNode(ISD::SUB, dl,
1635 VT, CmpOp, DAG.getConstant(Low, VT));
1636 Cond = DAG.getSetCC(dl, MVT::i1, SUB,
1637 DAG.getConstant(High-Low, VT), ISD::SETULE);
1641 // Update successor info
1642 addSuccessorWithWeight(SwitchBB, CB.TrueBB, CB.TrueWeight);
1643 // TrueBB and FalseBB are always different unless the incoming IR is
1644 // degenerate. This only happens when running llc on weird IR.
1645 if (CB.TrueBB != CB.FalseBB)
1646 addSuccessorWithWeight(SwitchBB, CB.FalseBB, CB.FalseWeight);
1648 // Set NextBlock to be the MBB immediately after the current one, if any.
1649 // This is used to avoid emitting unnecessary branches to the next block.
1650 MachineBasicBlock *NextBlock = 0;
1651 MachineFunction::iterator BBI = SwitchBB;
1652 if (++BBI != FuncInfo.MF->end())
1655 // If the lhs block is the next block, invert the condition so that we can
1656 // fall through to the lhs instead of the rhs block.
1657 if (CB.TrueBB == NextBlock) {
1658 std::swap(CB.TrueBB, CB.FalseBB);
1659 SDValue True = DAG.getConstant(1, Cond.getValueType());
1660 Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True);
1663 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1664 MVT::Other, getControlRoot(), Cond,
1665 DAG.getBasicBlock(CB.TrueBB));
1667 // Insert the false branch. Do this even if it's a fall through branch,
1668 // this makes it easier to do DAG optimizations which require inverting
1669 // the branch condition.
1670 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1671 DAG.getBasicBlock(CB.FalseBB));
1673 DAG.setRoot(BrCond);
1676 /// visitJumpTable - Emit JumpTable node in the current MBB
1677 void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) {
1678 // Emit the code for the jump table
1679 assert(JT.Reg != -1U && "Should lower JT Header first!");
1680 EVT PTy = TM.getTargetLowering()->getPointerTy();
1681 SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurSDLoc(),
1683 SDValue Table = DAG.getJumpTable(JT.JTI, PTy);
1684 SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurSDLoc(),
1685 MVT::Other, Index.getValue(1),
1687 DAG.setRoot(BrJumpTable);
1690 /// visitJumpTableHeader - This function emits necessary code to produce index
1691 /// in the JumpTable from switch case.
1692 void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT,
1693 JumpTableHeader &JTH,
1694 MachineBasicBlock *SwitchBB) {
1695 // Subtract the lowest switch case value from the value being switched on and
1696 // conditional branch to default mbb if the result is greater than the
1697 // difference between smallest and largest cases.
1698 SDValue SwitchOp = getValue(JTH.SValue);
1699 EVT VT = SwitchOp.getValueType();
1700 SDValue Sub = DAG.getNode(ISD::SUB, getCurSDLoc(), VT, SwitchOp,
1701 DAG.getConstant(JTH.First, VT));
1703 // The SDNode we just created, which holds the value being switched on minus
1704 // the smallest case value, needs to be copied to a virtual register so it
1705 // can be used as an index into the jump table in a subsequent basic block.
1706 // This value may be smaller or larger than the target's pointer type, and
1707 // therefore require extension or truncating.
1708 const TargetLowering *TLI = TM.getTargetLowering();
1709 SwitchOp = DAG.getZExtOrTrunc(Sub, getCurSDLoc(), TLI->getPointerTy());
1711 unsigned JumpTableReg = FuncInfo.CreateReg(TLI->getPointerTy());
1712 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurSDLoc(),
1713 JumpTableReg, SwitchOp);
1714 JT.Reg = JumpTableReg;
1716 // Emit the range check for the jump table, and branch to the default block
1717 // for the switch statement if the value being switched on exceeds the largest
1718 // case in the switch.
1719 SDValue CMP = DAG.getSetCC(getCurSDLoc(),
1720 TLI->getSetCCResultType(*DAG.getContext(),
1721 Sub.getValueType()),
1723 DAG.getConstant(JTH.Last - JTH.First,VT),
1726 // Set NextBlock to be the MBB immediately after the current one, if any.
1727 // This is used to avoid emitting unnecessary branches to the next block.
1728 MachineBasicBlock *NextBlock = 0;
1729 MachineFunction::iterator BBI = SwitchBB;
1731 if (++BBI != FuncInfo.MF->end())
1734 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurSDLoc(),
1735 MVT::Other, CopyTo, CMP,
1736 DAG.getBasicBlock(JT.Default));
1738 if (JT.MBB != NextBlock)
1739 BrCond = DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, BrCond,
1740 DAG.getBasicBlock(JT.MBB));
1742 DAG.setRoot(BrCond);
1745 /// Codegen a new tail for a stack protector check ParentMBB which has had its
1746 /// tail spliced into a stack protector check success bb.
1748 /// For a high level explanation of how this fits into the stack protector
1749 /// generation see the comment on the declaration of class
1750 /// StackProtectorDescriptor.
1751 void SelectionDAGBuilder::visitSPDescriptorParent(StackProtectorDescriptor &SPD,
1752 MachineBasicBlock *ParentBB) {
1754 // First create the loads to the guard/stack slot for the comparison.
1755 const TargetLowering *TLI = TM.getTargetLowering();
1756 EVT PtrTy = TLI->getPointerTy();
1758 MachineFrameInfo *MFI = ParentBB->getParent()->getFrameInfo();
1759 int FI = MFI->getStackProtectorIndex();
1761 const Value *IRGuard = SPD.getGuard();
1762 SDValue GuardPtr = getValue(IRGuard);
1763 SDValue StackSlotPtr = DAG.getFrameIndex(FI, PtrTy);
1766 TLI->getDataLayout()->getPrefTypeAlignment(IRGuard->getType());
1767 SDValue Guard = DAG.getLoad(PtrTy, getCurSDLoc(), DAG.getEntryNode(),
1768 GuardPtr, MachinePointerInfo(IRGuard, 0),
1769 true, false, false, Align);
1771 SDValue StackSlot = DAG.getLoad(PtrTy, getCurSDLoc(), DAG.getEntryNode(),
1773 MachinePointerInfo::getFixedStack(FI),
1774 true, false, false, Align);
1776 // Perform the comparison via a subtract/getsetcc.
1777 EVT VT = Guard.getValueType();
1778 SDValue Sub = DAG.getNode(ISD::SUB, getCurSDLoc(), VT, Guard, StackSlot);
1780 SDValue Cmp = DAG.getSetCC(getCurSDLoc(),
1781 TLI->getSetCCResultType(*DAG.getContext(),
1782 Sub.getValueType()),
1783 Sub, DAG.getConstant(0, VT),
1786 // If the sub is not 0, then we know the guard/stackslot do not equal, so
1787 // branch to failure MBB.
1788 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurSDLoc(),
1789 MVT::Other, StackSlot.getOperand(0),
1790 Cmp, DAG.getBasicBlock(SPD.getFailureMBB()));
1791 // Otherwise branch to success MBB.
1792 SDValue Br = DAG.getNode(ISD::BR, getCurSDLoc(),
1794 DAG.getBasicBlock(SPD.getSuccessMBB()));
1799 /// Codegen the failure basic block for a stack protector check.
1801 /// A failure stack protector machine basic block consists simply of a call to
1802 /// __stack_chk_fail().
1804 /// For a high level explanation of how this fits into the stack protector
1805 /// generation see the comment on the declaration of class
1806 /// StackProtectorDescriptor.
1808 SelectionDAGBuilder::visitSPDescriptorFailure(StackProtectorDescriptor &SPD) {
1809 const TargetLowering *TLI = TM.getTargetLowering();
1810 SDValue Chain = TLI->makeLibCall(DAG, RTLIB::STACKPROTECTOR_CHECK_FAIL,
1811 MVT::isVoid, 0, 0, false, getCurSDLoc(),
1812 false, false).second;
1816 /// visitBitTestHeader - This function emits necessary code to produce value
1817 /// suitable for "bit tests"
1818 void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B,
1819 MachineBasicBlock *SwitchBB) {
1820 // Subtract the minimum value
1821 SDValue SwitchOp = getValue(B.SValue);
1822 EVT VT = SwitchOp.getValueType();
1823 SDValue Sub = DAG.getNode(ISD::SUB, getCurSDLoc(), VT, SwitchOp,
1824 DAG.getConstant(B.First, VT));
1827 const TargetLowering *TLI = TM.getTargetLowering();
1828 SDValue RangeCmp = DAG.getSetCC(getCurSDLoc(),
1829 TLI->getSetCCResultType(*DAG.getContext(),
1830 Sub.getValueType()),
1831 Sub, DAG.getConstant(B.Range, VT),
1834 // Determine the type of the test operands.
1835 bool UsePtrType = false;
1836 if (!TLI->isTypeLegal(VT))
1839 for (unsigned i = 0, e = B.Cases.size(); i != e; ++i)
1840 if (!isUIntN(VT.getSizeInBits(), B.Cases[i].Mask)) {
1841 // Switch table case range are encoded into series of masks.
1842 // Just use pointer type, it's guaranteed to fit.
1848 VT = TLI->getPointerTy();
1849 Sub = DAG.getZExtOrTrunc(Sub, getCurSDLoc(), VT);
1852 B.RegVT = VT.getSimpleVT();
1853 B.Reg = FuncInfo.CreateReg(B.RegVT);
1854 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurSDLoc(),
1857 // Set NextBlock to be the MBB immediately after the current one, if any.
1858 // This is used to avoid emitting unnecessary branches to the next block.
1859 MachineBasicBlock *NextBlock = 0;
1860 MachineFunction::iterator BBI = SwitchBB;
1861 if (++BBI != FuncInfo.MF->end())
1864 MachineBasicBlock* MBB = B.Cases[0].ThisBB;
1866 addSuccessorWithWeight(SwitchBB, B.Default);
1867 addSuccessorWithWeight(SwitchBB, MBB);
1869 SDValue BrRange = DAG.getNode(ISD::BRCOND, getCurSDLoc(),
1870 MVT::Other, CopyTo, RangeCmp,
1871 DAG.getBasicBlock(B.Default));
1873 if (MBB != NextBlock)
1874 BrRange = DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, CopyTo,
1875 DAG.getBasicBlock(MBB));
1877 DAG.setRoot(BrRange);
1880 /// visitBitTestCase - this function produces one "bit test"
1881 void SelectionDAGBuilder::visitBitTestCase(BitTestBlock &BB,
1882 MachineBasicBlock* NextMBB,
1883 uint32_t BranchWeightToNext,
1886 MachineBasicBlock *SwitchBB) {
1888 SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurSDLoc(),
1891 unsigned PopCount = CountPopulation_64(B.Mask);
1892 const TargetLowering *TLI = TM.getTargetLowering();
1893 if (PopCount == 1) {
1894 // Testing for a single bit; just compare the shift count with what it
1895 // would need to be to shift a 1 bit in that position.
1896 Cmp = DAG.getSetCC(getCurSDLoc(),
1897 TLI->getSetCCResultType(*DAG.getContext(), VT),
1899 DAG.getConstant(countTrailingZeros(B.Mask), VT),
1901 } else if (PopCount == BB.Range) {
1902 // There is only one zero bit in the range, test for it directly.
1903 Cmp = DAG.getSetCC(getCurSDLoc(),
1904 TLI->getSetCCResultType(*DAG.getContext(), VT),
1906 DAG.getConstant(CountTrailingOnes_64(B.Mask), VT),
1909 // Make desired shift
1910 SDValue SwitchVal = DAG.getNode(ISD::SHL, getCurSDLoc(), VT,
1911 DAG.getConstant(1, VT), ShiftOp);
1913 // Emit bit tests and jumps
1914 SDValue AndOp = DAG.getNode(ISD::AND, getCurSDLoc(),
1915 VT, SwitchVal, DAG.getConstant(B.Mask, VT));
1916 Cmp = DAG.getSetCC(getCurSDLoc(),
1917 TLI->getSetCCResultType(*DAG.getContext(), VT),
1918 AndOp, DAG.getConstant(0, VT),
1922 // The branch weight from SwitchBB to B.TargetBB is B.ExtraWeight.
1923 addSuccessorWithWeight(SwitchBB, B.TargetBB, B.ExtraWeight);
1924 // The branch weight from SwitchBB to NextMBB is BranchWeightToNext.
1925 addSuccessorWithWeight(SwitchBB, NextMBB, BranchWeightToNext);
1927 SDValue BrAnd = DAG.getNode(ISD::BRCOND, getCurSDLoc(),
1928 MVT::Other, getControlRoot(),
1929 Cmp, DAG.getBasicBlock(B.TargetBB));
1931 // Set NextBlock to be the MBB immediately after the current one, if any.
1932 // This is used to avoid emitting unnecessary branches to the next block.
1933 MachineBasicBlock *NextBlock = 0;
1934 MachineFunction::iterator BBI = SwitchBB;
1935 if (++BBI != FuncInfo.MF->end())
1938 if (NextMBB != NextBlock)
1939 BrAnd = DAG.getNode(ISD::BR, getCurSDLoc(), MVT::Other, BrAnd,
1940 DAG.getBasicBlock(NextMBB));
1945 void SelectionDAGBuilder::visitInvoke(const InvokeInst &I) {
1946 MachineBasicBlock *InvokeMBB = FuncInfo.MBB;
1948 // Retrieve successors.
1949 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
1950 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)];
1952 const Value *Callee(I.getCalledValue());
1953 const Function *Fn = dyn_cast<Function>(Callee);
1954 if (isa<InlineAsm>(Callee))
1956 else if (Fn && Fn->isIntrinsic()) {
1957 assert(Fn->getIntrinsicID() == Intrinsic::donothing);
1958 // Ignore invokes to @llvm.donothing: jump directly to the next BB.
1960 LowerCallTo(&I, getValue(Callee), false, LandingPad);
1962 // If the value of the invoke is used outside of its defining block, make it
1963 // available as a virtual register.
1964 CopyToExportRegsIfNeeded(&I);
1966 // Update successor info
1967 addSuccessorWithWeight(InvokeMBB, Return);
1968 addSuccessorWithWeight(InvokeMBB, LandingPad);
1970 // Drop into normal successor.
1971 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(),
1972 MVT::Other, getControlRoot(),
1973 DAG.getBasicBlock(Return)));
1976 void SelectionDAGBuilder::visitResume(const ResumeInst &RI) {
1977 llvm_unreachable("SelectionDAGBuilder shouldn't visit resume instructions!");
1980 void SelectionDAGBuilder::visitLandingPad(const LandingPadInst &LP) {
1981 assert(FuncInfo.MBB->isLandingPad() &&
1982 "Call to landingpad not in landing pad!");
1984 MachineBasicBlock *MBB = FuncInfo.MBB;
1985 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
1986 AddLandingPadInfo(LP, MMI, MBB);
1988 // If there aren't registers to copy the values into (e.g., during SjLj
1989 // exceptions), then don't bother to create these DAG nodes.
1990 const TargetLowering *TLI = TM.getTargetLowering();
1991 if (TLI->getExceptionPointerRegister() == 0 &&
1992 TLI->getExceptionSelectorRegister() == 0)
1995 SmallVector<EVT, 2> ValueVTs;
1996 ComputeValueVTs(*TLI, LP.getType(), ValueVTs);
1997 assert(ValueVTs.size() == 2 && "Only two-valued landingpads are supported");
1999 // Get the two live-in registers as SDValues. The physregs have already been
2000 // copied into virtual registers.
2002 Ops[0] = DAG.getZExtOrTrunc(
2003 DAG.getCopyFromReg(DAG.getEntryNode(), getCurSDLoc(),
2004 FuncInfo.ExceptionPointerVirtReg, TLI->getPointerTy()),
2005 getCurSDLoc(), ValueVTs[0]);
2006 Ops[1] = DAG.getZExtOrTrunc(
2007 DAG.getCopyFromReg(DAG.getEntryNode(), getCurSDLoc(),
2008 FuncInfo.ExceptionSelectorVirtReg, TLI->getPointerTy()),
2009 getCurSDLoc(), ValueVTs[1]);
2012 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
2013 DAG.getVTList(&ValueVTs[0], ValueVTs.size()),
2018 /// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for
2019 /// small case ranges).
2020 bool SelectionDAGBuilder::handleSmallSwitchRange(CaseRec& CR,
2021 CaseRecVector& WorkList,
2023 MachineBasicBlock *Default,
2024 MachineBasicBlock *SwitchBB) {
2025 // Size is the number of Cases represented by this range.
2026 size_t Size = CR.Range.second - CR.Range.first;
2030 // Get the MachineFunction which holds the current MBB. This is used when
2031 // inserting any additional MBBs necessary to represent the switch.
2032 MachineFunction *CurMF = FuncInfo.MF;
2034 // Figure out which block is immediately after the current one.
2035 MachineBasicBlock *NextBlock = 0;
2036 MachineFunction::iterator BBI = CR.CaseBB;
2038 if (++BBI != FuncInfo.MF->end())
2041 BranchProbabilityInfo *BPI = FuncInfo.BPI;
2042 // If any two of the cases has the same destination, and if one value
2043 // is the same as the other, but has one bit unset that the other has set,
2044 // use bit manipulation to do two compares at once. For example:
2045 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
2046 // TODO: This could be extended to merge any 2 cases in switches with 3 cases.
2047 // TODO: Handle cases where CR.CaseBB != SwitchBB.
2048 if (Size == 2 && CR.CaseBB == SwitchBB) {
2049 Case &Small = *CR.Range.first;
2050 Case &Big = *(CR.Range.second-1);
2052 if (Small.Low == Small.High && Big.Low == Big.High && Small.BB == Big.BB) {
2053 const APInt& SmallValue = cast<ConstantInt>(Small.Low)->getValue();
2054 const APInt& BigValue = cast<ConstantInt>(Big.Low)->getValue();
2056 // Check that there is only one bit different.
2057 if (BigValue.countPopulation() == SmallValue.countPopulation() + 1 &&
2058 (SmallValue | BigValue) == BigValue) {
2059 // Isolate the common bit.
2060 APInt CommonBit = BigValue & ~SmallValue;
2061 assert((SmallValue | CommonBit) == BigValue &&
2062 CommonBit.countPopulation() == 1 && "Not a common bit?");
2064 SDValue CondLHS = getValue(SV);
2065 EVT VT = CondLHS.getValueType();
2066 SDLoc DL = getCurSDLoc();
2068 SDValue Or = DAG.getNode(ISD::OR, DL, VT, CondLHS,
2069 DAG.getConstant(CommonBit, VT));
2070 SDValue Cond = DAG.getSetCC(DL, MVT::i1,
2071 Or, DAG.getConstant(BigValue, VT),
2074 // Update successor info.
2075 // Both Small and Big will jump to Small.BB, so we sum up the weights.
2076 addSuccessorWithWeight(SwitchBB, Small.BB,
2077 Small.ExtraWeight + Big.ExtraWeight);
2078 addSuccessorWithWeight(SwitchBB, Default,
2079 // The default destination is the first successor in IR.
2080 BPI ? BPI->getEdgeWeight(SwitchBB->getBasicBlock(), (unsigned)0) : 0);
2082 // Insert the true branch.
2083 SDValue BrCond = DAG.getNode(ISD::BRCOND, DL, MVT::Other,
2084 getControlRoot(), Cond,
2085 DAG.getBasicBlock(Small.BB));
2087 // Insert the false branch.
2088 BrCond = DAG.getNode(ISD::BR, DL, MVT::Other, BrCond,
2089 DAG.getBasicBlock(Default));
2091 DAG.setRoot(BrCond);
2097 // Order cases by weight so the most likely case will be checked first.
2098 uint32_t UnhandledWeights = 0;
2100 for (CaseItr I = CR.Range.first, IE = CR.Range.second; I != IE; ++I) {
2101 uint32_t IWeight = I->ExtraWeight;
2102 UnhandledWeights += IWeight;
2103 for (CaseItr J = CR.Range.first; J < I; ++J) {
2104 uint32_t JWeight = J->ExtraWeight;
2105 if (IWeight > JWeight)
2110 // Rearrange the case blocks so that the last one falls through if possible.
2111 Case &BackCase = *(CR.Range.second-1);
2113 NextBlock && Default != NextBlock && BackCase.BB != NextBlock) {
2114 // The last case block won't fall through into 'NextBlock' if we emit the
2115 // branches in this order. See if rearranging a case value would help.
2116 // We start at the bottom as it's the case with the least weight.
2117 for (Case *I = &*(CR.Range.second-2), *E = &*CR.Range.first-1; I != E; --I)
2118 if (I->BB == NextBlock) {
2119 std::swap(*I, BackCase);
2124 // Create a CaseBlock record representing a conditional branch to
2125 // the Case's target mbb if the value being switched on SV is equal
2127 MachineBasicBlock *CurBlock = CR.CaseBB;
2128 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
2129 MachineBasicBlock *FallThrough;
2131 FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock());
2132 CurMF->insert(BBI, FallThrough);
2134 // Put SV in a virtual register to make it available from the new blocks.
2135 ExportFromCurrentBlock(SV);
2137 // If the last case doesn't match, go to the default block.
2138 FallThrough = Default;
2141 const Value *RHS, *LHS, *MHS;
2143 if (I->High == I->Low) {
2144 // This is just small small case range :) containing exactly 1 case
2146 LHS = SV; RHS = I->High; MHS = NULL;
2148 CC = ISD::SETCC_INVALID;
2149 LHS = I->Low; MHS = SV; RHS = I->High;
2152 // The false weight should be sum of all un-handled cases.
2153 UnhandledWeights -= I->ExtraWeight;
2154 CaseBlock CB(CC, LHS, RHS, MHS, /* truebb */ I->BB, /* falsebb */ FallThrough,
2156 /* trueweight */ I->ExtraWeight,
2157 /* falseweight */ UnhandledWeights);
2159 // If emitting the first comparison, just call visitSwitchCase to emit the
2160 // code into the current block. Otherwise, push the CaseBlock onto the
2161 // vector to be later processed by SDISel, and insert the node's MBB
2162 // before the next MBB.
2163 if (CurBlock == SwitchBB)
2164 visitSwitchCase(CB, SwitchBB);
2166 SwitchCases.push_back(CB);
2168 CurBlock = FallThrough;
2174 static inline bool areJTsAllowed(const TargetLowering &TLI) {
2175 return TLI.supportJumpTables() &&
2176 (TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
2177 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other));
2180 static APInt ComputeRange(const APInt &First, const APInt &Last) {
2181 uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1;
2182 APInt LastExt = Last.zext(BitWidth), FirstExt = First.zext(BitWidth);
2183 return (LastExt - FirstExt + 1ULL);
2186 /// handleJTSwitchCase - Emit jumptable for current switch case range
2187 bool SelectionDAGBuilder::handleJTSwitchCase(CaseRec &CR,
2188 CaseRecVector &WorkList,
2190 MachineBasicBlock *Default,
2191 MachineBasicBlock *SwitchBB) {
2192 Case& FrontCase = *CR.Range.first;
2193 Case& BackCase = *(CR.Range.second-1);
2195 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
2196 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
2198 APInt TSize(First.getBitWidth(), 0);
2199 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I)
2202 const TargetLowering *TLI = TM.getTargetLowering();
2203 if (!areJTsAllowed(*TLI) || TSize.ult(TLI->getMinimumJumpTableEntries()))
2206 APInt Range = ComputeRange(First, Last);
2207 // The density is TSize / Range. Require at least 40%.
2208 // It should not be possible for IntTSize to saturate for sane code, but make
2209 // sure we handle Range saturation correctly.
2210 uint64_t IntRange = Range.getLimitedValue(UINT64_MAX/10);
2211 uint64_t IntTSize = TSize.getLimitedValue(UINT64_MAX/10);
2212 if (IntTSize * 10 < IntRange * 4)
2215 DEBUG(dbgs() << "Lowering jump table\n"
2216 << "First entry: " << First << ". Last entry: " << Last << '\n'
2217 << "Range: " << Range << ". Size: " << TSize << ".\n\n");
2219 // Get the MachineFunction which holds the current MBB. This is used when
2220 // inserting any additional MBBs necessary to represent the switch.
2221 MachineFunction *CurMF = FuncInfo.MF;
2223 // Figure out which block is immediately after the current one.
2224 MachineFunction::iterator BBI = CR.CaseBB;
2227 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2229 // Create a new basic block to hold the code for loading the address
2230 // of the jump table, and jumping to it. Update successor information;
2231 // we will either branch to the default case for the switch, or the jump
2233 MachineBasicBlock *JumpTableBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2234 CurMF->insert(BBI, JumpTableBB);
2236 addSuccessorWithWeight(CR.CaseBB, Default);
2237 addSuccessorWithWeight(CR.CaseBB, JumpTableBB);
2239 // Build a vector of destination BBs, corresponding to each target
2240 // of the jump table. If the value of the jump table slot corresponds to
2241 // a case statement, push the case's BB onto the vector, otherwise, push
2243 std::vector<MachineBasicBlock*> DestBBs;
2245 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) {
2246 const APInt &Low = cast<ConstantInt>(I->Low)->getValue();
2247 const APInt &High = cast<ConstantInt>(I->High)->getValue();
2249 if (Low.ule(TEI) && TEI.ule(High)) {
2250 DestBBs.push_back(I->BB);
2254 DestBBs.push_back(Default);
2258 // Calculate weight for each unique destination in CR.
2259 DenseMap<MachineBasicBlock*, uint32_t> DestWeights;
2261 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
2262 DenseMap<MachineBasicBlock*, uint32_t>::iterator Itr =
2263 DestWeights.find(I->BB);
2264 if (Itr != DestWeights.end())
2265 Itr->second += I->ExtraWeight;
2267 DestWeights[I->BB] = I->ExtraWeight;
2270 // Update successor info. Add one edge to each unique successor.
2271 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs());
2272 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(),
2273 E = DestBBs.end(); I != E; ++I) {
2274 if (!SuccsHandled[(*I)->getNumber()]) {
2275 SuccsHandled[(*I)->getNumber()] = true;
2276 DenseMap<MachineBasicBlock*, uint32_t>::iterator Itr =
2277 DestWeights.find(*I);
2278 addSuccessorWithWeight(JumpTableBB, *I,
2279 Itr != DestWeights.end() ? Itr->second : 0);
2283 // Create a jump table index for this jump table.
2284 unsigned JTEncoding = TLI->getJumpTableEncoding();
2285 unsigned JTI = CurMF->getOrCreateJumpTableInfo(JTEncoding)
2286 ->createJumpTableIndex(DestBBs);
2288 // Set the jump table information so that we can codegen it as a second
2289 // MachineBasicBlock
2290 JumpTable JT(-1U, JTI, JumpTableBB, Default);
2291 JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == SwitchBB));
2292 if (CR.CaseBB == SwitchBB)
2293 visitJumpTableHeader(JT, JTH, SwitchBB);
2295 JTCases.push_back(JumpTableBlock(JTH, JT));
2299 /// handleBTSplitSwitchCase - emit comparison and split binary search tree into
2301 bool SelectionDAGBuilder::handleBTSplitSwitchCase(CaseRec& CR,
2302 CaseRecVector& WorkList,
2304 MachineBasicBlock* Default,
2305 MachineBasicBlock* SwitchBB) {
2306 // Get the MachineFunction which holds the current MBB. This is used when
2307 // inserting any additional MBBs necessary to represent the switch.
2308 MachineFunction *CurMF = FuncInfo.MF;
2310 // Figure out which block is immediately after the current one.
2311 MachineFunction::iterator BBI = CR.CaseBB;
2314 Case& FrontCase = *CR.Range.first;
2315 Case& BackCase = *(CR.Range.second-1);
2316 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2318 // Size is the number of Cases represented by this range.
2319 unsigned Size = CR.Range.second - CR.Range.first;
2321 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
2322 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
2324 CaseItr Pivot = CR.Range.first + Size/2;
2326 // Select optimal pivot, maximizing sum density of LHS and RHS. This will
2327 // (heuristically) allow us to emit JumpTable's later.
2328 APInt TSize(First.getBitWidth(), 0);
2329 for (CaseItr I = CR.Range.first, E = CR.Range.second;
2333 APInt LSize = FrontCase.size();
2334 APInt RSize = TSize-LSize;
2335 DEBUG(dbgs() << "Selecting best pivot: \n"
2336 << "First: " << First << ", Last: " << Last <<'\n'
2337 << "LSize: " << LSize << ", RSize: " << RSize << '\n');
2338 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second;
2340 const APInt &LEnd = cast<ConstantInt>(I->High)->getValue();
2341 const APInt &RBegin = cast<ConstantInt>(J->Low)->getValue();
2342 APInt Range = ComputeRange(LEnd, RBegin);
2343 assert((Range - 2ULL).isNonNegative() &&
2344 "Invalid case distance");
2345 // Use volatile double here to avoid excess precision issues on some hosts,
2346 // e.g. that use 80-bit X87 registers.
2347 volatile double LDensity =
2348 (double)LSize.roundToDouble() /
2349 (LEnd - First + 1ULL).roundToDouble();
2350 volatile double RDensity =
2351 (double)RSize.roundToDouble() /
2352 (Last - RBegin + 1ULL).roundToDouble();
2353 double Metric = Range.logBase2()*(LDensity+RDensity);
2354 // Should always split in some non-trivial place
2355 DEBUG(dbgs() <<"=>Step\n"
2356 << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n'
2357 << "LDensity: " << LDensity
2358 << ", RDensity: " << RDensity << '\n'
2359 << "Metric: " << Metric << '\n');
2360 if (FMetric < Metric) {
2363 DEBUG(dbgs() << "Current metric set to: " << FMetric << '\n');
2370 const TargetLowering *TLI = TM.getTargetLowering();
2371 if (areJTsAllowed(*TLI)) {
2372 // If our case is dense we *really* should handle it earlier!
2373 assert((FMetric > 0) && "Should handle dense range earlier!");
2375 Pivot = CR.Range.first + Size/2;
2378 CaseRange LHSR(CR.Range.first, Pivot);
2379 CaseRange RHSR(Pivot, CR.Range.second);
2380 const Constant *C = Pivot->Low;
2381 MachineBasicBlock *FalseBB = 0, *TrueBB = 0;
2383 // We know that we branch to the LHS if the Value being switched on is
2384 // less than the Pivot value, C. We use this to optimize our binary
2385 // tree a bit, by recognizing that if SV is greater than or equal to the
2386 // LHS's Case Value, and that Case Value is exactly one less than the
2387 // Pivot's Value, then we can branch directly to the LHS's Target,
2388 // rather than creating a leaf node for it.
2389 if ((LHSR.second - LHSR.first) == 1 &&
2390 LHSR.first->High == CR.GE &&
2391 cast<ConstantInt>(C)->getValue() ==
2392 (cast<ConstantInt>(CR.GE)->getValue() + 1LL)) {
2393 TrueBB = LHSR.first->BB;
2395 TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2396 CurMF->insert(BBI, TrueBB);
2397 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR));
2399 // Put SV in a virtual register to make it available from the new blocks.
2400 ExportFromCurrentBlock(SV);
2403 // Similar to the optimization above, if the Value being switched on is
2404 // known to be less than the Constant CR.LT, and the current Case Value
2405 // is CR.LT - 1, then we can branch directly to the target block for
2406 // the current Case Value, rather than emitting a RHS leaf node for it.
2407 if ((RHSR.second - RHSR.first) == 1 && CR.LT &&
2408 cast<ConstantInt>(RHSR.first->Low)->getValue() ==
2409 (cast<ConstantInt>(CR.LT)->getValue() - 1LL)) {
2410 FalseBB = RHSR.first->BB;
2412 FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2413 CurMF->insert(BBI, FalseBB);
2414 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR));
2416 // Put SV in a virtual register to make it available from the new blocks.
2417 ExportFromCurrentBlock(SV);
2420 // Create a CaseBlock record representing a conditional branch to
2421 // the LHS node if the value being switched on SV is less than C.
2422 // Otherwise, branch to LHS.
2423 CaseBlock CB(ISD::SETULT, SV, C, NULL, TrueBB, FalseBB, CR.CaseBB);
2425 if (CR.CaseBB == SwitchBB)
2426 visitSwitchCase(CB, SwitchBB);
2428 SwitchCases.push_back(CB);
2433 /// handleBitTestsSwitchCase - if current case range has few destination and
2434 /// range span less, than machine word bitwidth, encode case range into series
2435 /// of masks and emit bit tests with these masks.
2436 bool SelectionDAGBuilder::handleBitTestsSwitchCase(CaseRec& CR,
2437 CaseRecVector& WorkList,
2439 MachineBasicBlock* Default,
2440 MachineBasicBlock* SwitchBB) {
2441 const TargetLowering *TLI = TM.getTargetLowering();
2442 EVT PTy = TLI->getPointerTy();
2443 unsigned IntPtrBits = PTy.getSizeInBits();
2445 Case& FrontCase = *CR.Range.first;
2446 Case& BackCase = *(CR.Range.second-1);
2448 // Get the MachineFunction which holds the current MBB. This is used when
2449 // inserting any additional MBBs necessary to represent the switch.
2450 MachineFunction *CurMF = FuncInfo.MF;
2452 // If target does not have legal shift left, do not emit bit tests at all.
2453 if (!TLI->isOperationLegal(ISD::SHL, TLI->getPointerTy()))
2457 for (CaseItr I = CR.Range.first, E = CR.Range.second;
2459 // Single case counts one, case range - two.
2460 numCmps += (I->Low == I->High ? 1 : 2);
2463 // Count unique destinations
2464 SmallSet<MachineBasicBlock*, 4> Dests;
2465 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2466 Dests.insert(I->BB);
2467 if (Dests.size() > 3)
2468 // Don't bother the code below, if there are too much unique destinations
2471 DEBUG(dbgs() << "Total number of unique destinations: "
2472 << Dests.size() << '\n'
2473 << "Total number of comparisons: " << numCmps << '\n');
2475 // Compute span of values.
2476 const APInt& minValue = cast<ConstantInt>(FrontCase.Low)->getValue();
2477 const APInt& maxValue = cast<ConstantInt>(BackCase.High)->getValue();
2478 APInt cmpRange = maxValue - minValue;
2480 DEBUG(dbgs() << "Compare range: " << cmpRange << '\n'
2481 << "Low bound: " << minValue << '\n'
2482 << "High bound: " << maxValue << '\n');
2484 if (cmpRange.uge(IntPtrBits) ||
2485 (!(Dests.size() == 1 && numCmps >= 3) &&
2486 !(Dests.size() == 2 && numCmps >= 5) &&
2487 !(Dests.size() >= 3 && numCmps >= 6)))
2490 DEBUG(dbgs() << "Emitting bit tests\n");
2491 APInt lowBound = APInt::getNullValue(cmpRange.getBitWidth());
2493 // Optimize the case where all the case values fit in a
2494 // word without having to subtract minValue. In this case,
2495 // we can optimize away the subtraction.
2496 if (maxValue.ult(IntPtrBits)) {
2497 cmpRange = maxValue;
2499 lowBound = minValue;
2502 CaseBitsVector CasesBits;
2503 unsigned i, count = 0;
2505 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2506 MachineBasicBlock* Dest = I->BB;
2507 for (i = 0; i < count; ++i)
2508 if (Dest == CasesBits[i].BB)
2512 assert((count < 3) && "Too much destinations to test!");
2513 CasesBits.push_back(CaseBits(0, Dest, 0, 0/*Weight*/));
2517 const APInt& lowValue = cast<ConstantInt>(I->Low)->getValue();
2518 const APInt& highValue = cast<ConstantInt>(I->High)->getValue();
2520 uint64_t lo = (lowValue - lowBound).getZExtValue();
2521 uint64_t hi = (highValue - lowBound).getZExtValue();
2522 CasesBits[i].ExtraWeight += I->ExtraWeight;
2524 for (uint64_t j = lo; j <= hi; j++) {
2525 CasesBits[i].Mask |= 1ULL << j;
2526 CasesBits[i].Bits++;
2530 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp());
2534 // Figure out which block is immediately after the current one.
2535 MachineFunction::iterator BBI = CR.CaseBB;
2538 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2540 DEBUG(dbgs() << "Cases:\n");
2541 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) {
2542 DEBUG(dbgs() << "Mask: " << CasesBits[i].Mask
2543 << ", Bits: " << CasesBits[i].Bits
2544 << ", BB: " << CasesBits[i].BB << '\n');
2546 MachineBasicBlock *CaseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2547 CurMF->insert(BBI, CaseBB);
2548 BTC.push_back(BitTestCase(CasesBits[i].Mask,
2550 CasesBits[i].BB, CasesBits[i].ExtraWeight));
2552 // Put SV in a virtual register to make it available from the new blocks.
2553 ExportFromCurrentBlock(SV);
2556 BitTestBlock BTB(lowBound, cmpRange, SV,
2557 -1U, MVT::Other, (CR.CaseBB == SwitchBB),
2558 CR.CaseBB, Default, BTC);
2560 if (CR.CaseBB == SwitchBB)
2561 visitBitTestHeader(BTB, SwitchBB);
2563 BitTestCases.push_back(BTB);
2568 /// Clusterify - Transform simple list of Cases into list of CaseRange's
2569 size_t SelectionDAGBuilder::Clusterify(CaseVector& Cases,
2570 const SwitchInst& SI) {
2572 /// Use a shorter form of declaration, and also
2573 /// show the we want to use CRSBuilder as Clusterifier.
2574 typedef IntegersSubsetMapping<MachineBasicBlock> Clusterifier;
2576 Clusterifier TheClusterifier;
2578 BranchProbabilityInfo *BPI = FuncInfo.BPI;
2579 // Start with "simple" cases
2580 for (SwitchInst::ConstCaseIt i = SI.case_begin(), e = SI.case_end();
2582 const BasicBlock *SuccBB = i.getCaseSuccessor();
2583 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SuccBB];
2585 TheClusterifier.add(i.getCaseValueEx(), SMBB,
2586 BPI ? BPI->getEdgeWeight(SI.getParent(), i.getSuccessorIndex()) : 0);
2589 TheClusterifier.optimize();
2592 for (Clusterifier::RangeIterator i = TheClusterifier.begin(),
2593 e = TheClusterifier.end(); i != e; ++i, ++numCmps) {
2594 Clusterifier::Cluster &C = *i;
2595 // Update edge weight for the cluster.
2596 unsigned W = C.first.Weight;
2598 // FIXME: Currently work with ConstantInt based numbers.
2599 // Changing it to APInt based is a pretty heavy for this commit.
2600 Cases.push_back(Case(C.first.getLow().toConstantInt(),
2601 C.first.getHigh().toConstantInt(), C.second, W));
2603 if (C.first.getLow() != C.first.getHigh())
2604 // A range counts double, since it requires two compares.
2611 void SelectionDAGBuilder::UpdateSplitBlock(MachineBasicBlock *First,
2612 MachineBasicBlock *Last) {
2614 for (unsigned i = 0, e = JTCases.size(); i != e; ++i)
2615 if (JTCases[i].first.HeaderBB == First)
2616 JTCases[i].first.HeaderBB = Last;
2618 // Update BitTestCases.
2619 for (unsigned i = 0, e = BitTestCases.size(); i != e; ++i)
2620 if (BitTestCases[i].Parent == First)
2621 BitTestCases[i].Parent = Last;
2624 void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) {
2625 MachineBasicBlock *SwitchMBB = FuncInfo.MBB;
2627 // Figure out which block is immediately after the current one.
2628 MachineBasicBlock *NextBlock = 0;
2629 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()];
2631 // If there is only the default destination, branch to it if it is not the
2632 // next basic block. Otherwise, just fall through.
2633 if (!SI.getNumCases()) {
2634 // Update machine-CFG edges.
2636 // If this is not a fall-through branch, emit the branch.
2637 SwitchMBB->addSuccessor(Default);
2638 if (Default != NextBlock)
2639 DAG.setRoot(DAG.getNode(ISD::BR, getCurSDLoc(),
2640 MVT::Other, getControlRoot(),
2641 DAG.getBasicBlock(Default)));
2646 // If there are any non-default case statements, create a vector of Cases
2647 // representing each one, and sort the vector so that we can efficiently
2648 // create a binary search tree from them.
2650 size_t numCmps = Clusterify(Cases, SI);
2651 DEBUG(dbgs() << "Clusterify finished. Total clusters: " << Cases.size()
2652 << ". Total compares: " << numCmps << '\n');
2655 // Get the Value to be switched on and default basic blocks, which will be
2656 // inserted into CaseBlock records, representing basic blocks in the binary
2658 const Value *SV = SI.getCondition();
2660 // Push the initial CaseRec onto the worklist
2661 CaseRecVector WorkList;
2662 WorkList.push_back(CaseRec(SwitchMBB,0,0,
2663 CaseRange(Cases.begin(),Cases.end())));
2665 while (!WorkList.empty()) {
2666 // Grab a record representing a case range to process off the worklist
2667 CaseRec CR = WorkList.back();
2668 WorkList.pop_back();
2670 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2673 // If the range has few cases (two or less) emit a series of specific
2675 if (handleSmallSwitchRange(CR, WorkList, SV, Default, SwitchMBB))
2678 // If the switch has more than N blocks, and is at least 40% dense, and the
2679 // target supports indirect branches, then emit a jump table rather than
2680 // lowering the switch to a binary tree of conditional branches.
2681 // N defaults to 4 and is controlled via TLS.getMinimumJumpTableEntries().
2682 if (handleJTSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2685 // Emit binary tree. We need to pick a pivot, and push left and right ranges
2686 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call.
2687 handleBTSplitSwitchCase(CR, WorkList, SV, Default, SwitchMBB);
2691 void SelectionDAGBuilder::visitIndirectBr(const IndirectBrInst &I) {
2692 MachineBasicBlock *IndirectBrMBB = FuncInfo.MBB;
2694 // Update machine-CFG edges with unique successors.
2695 SmallSet<BasicBlock*, 32> Done;
2696 for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i) {
2697 BasicBlock *BB = I.getSuccessor(i);
2698 bool Inserted = Done.insert(BB);
2702 MachineBasicBlock *Succ = FuncInfo.MBBMap[BB];
2703 addSuccessorWithWeight(IndirectBrMBB, Succ);
2706 DAG.setRoot(DAG.getNode(ISD::BRIND, getCurSDLoc(),
2707 MVT::Other, getControlRoot(),
2708 getValue(I.getAddress())));
2711 void SelectionDAGBuilder::visitFSub(const User &I) {
2712 // -0.0 - X --> fneg
2713 Type *Ty = I.getType();
2714 if (isa<Constant>(I.getOperand(0)) &&
2715 I.getOperand(0) == ConstantFP::getZeroValueForNegation(Ty)) {
2716 SDValue Op2 = getValue(I.getOperand(1));
2717 setValue(&I, DAG.getNode(ISD::FNEG, getCurSDLoc(),
2718 Op2.getValueType(), Op2));
2722 visitBinary(I, ISD::FSUB);
2725 void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) {
2726 SDValue Op1 = getValue(I.getOperand(0));
2727 SDValue Op2 = getValue(I.getOperand(1));
2728 setValue(&I, DAG.getNode(OpCode, getCurSDLoc(),
2729 Op1.getValueType(), Op1, Op2));
2732 void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) {
2733 SDValue Op1 = getValue(I.getOperand(0));
2734 SDValue Op2 = getValue(I.getOperand(1));
2736 EVT ShiftTy = TM.getTargetLowering()->getShiftAmountTy(Op2.getValueType());
2738 // Coerce the shift amount to the right type if we can.
2739 if (!I.getType()->isVectorTy() && Op2.getValueType() != ShiftTy) {
2740 unsigned ShiftSize = ShiftTy.getSizeInBits();
2741 unsigned Op2Size = Op2.getValueType().getSizeInBits();
2742 SDLoc DL = getCurSDLoc();
2744 // If the operand is smaller than the shift count type, promote it.
2745 if (ShiftSize > Op2Size)
2746 Op2 = DAG.getNode(ISD::ZERO_EXTEND, DL, ShiftTy, Op2);
2748 // If the operand is larger than the shift count type but the shift
2749 // count type has enough bits to represent any shift value, truncate
2750 // it now. This is a common case and it exposes the truncate to
2751 // optimization early.
2752 else if (ShiftSize >= Log2_32_Ceil(Op2.getValueType().getSizeInBits()))
2753 Op2 = DAG.getNode(ISD::TRUNCATE, DL, ShiftTy, Op2);
2754 // Otherwise we'll need to temporarily settle for some other convenient
2755 // type. Type legalization will make adjustments once the shiftee is split.
2757 Op2 = DAG.getZExtOrTrunc(Op2, DL, MVT::i32);
2760 setValue(&I, DAG.getNode(Opcode, getCurSDLoc(),
2761 Op1.getValueType(), Op1, Op2));
2764 void SelectionDAGBuilder::visitSDiv(const User &I) {
2765 SDValue Op1 = getValue(I.getOperand(0));
2766 SDValue Op2 = getValue(I.getOperand(1));
2768 // Turn exact SDivs into multiplications.
2769 // FIXME: This should be in DAGCombiner, but it doesn't have access to the
2771 if (isa<BinaryOperator>(&I) && cast<BinaryOperator>(&I)->isExact() &&
2772 !isa<ConstantSDNode>(Op1) &&
2773 isa<ConstantSDNode>(Op2) && !cast<ConstantSDNode>(Op2)->isNullValue())
2774 setValue(&I, TM.getTargetLowering()->BuildExactSDIV(Op1, Op2,
2775 getCurSDLoc(), DAG));
2777 setValue(&I, DAG.getNode(ISD::SDIV, getCurSDLoc(), Op1.getValueType(),
2781 void SelectionDAGBuilder::visitICmp(const User &I) {
2782 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2783 if (const ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2784 predicate = IC->getPredicate();
2785 else if (const ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2786 predicate = ICmpInst::Predicate(IC->getPredicate());
2787 SDValue Op1 = getValue(I.getOperand(0));
2788 SDValue Op2 = getValue(I.getOperand(1));
2789 ISD::CondCode Opcode = getICmpCondCode(predicate);
2791 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2792 setValue(&I, DAG.getSetCC(getCurSDLoc(), DestVT, Op1, Op2, Opcode));
2795 void SelectionDAGBuilder::visitFCmp(const User &I) {
2796 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2797 if (const FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2798 predicate = FC->getPredicate();
2799 else if (const ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2800 predicate = FCmpInst::Predicate(FC->getPredicate());
2801 SDValue Op1 = getValue(I.getOperand(0));
2802 SDValue Op2 = getValue(I.getOperand(1));
2803 ISD::CondCode Condition = getFCmpCondCode(predicate);
2804 if (TM.Options.NoNaNsFPMath)
2805 Condition = getFCmpCodeWithoutNaN(Condition);
2806 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2807 setValue(&I, DAG.getSetCC(getCurSDLoc(), DestVT, Op1, Op2, Condition));
2810 void SelectionDAGBuilder::visitSelect(const User &I) {
2811 SmallVector<EVT, 4> ValueVTs;
2812 ComputeValueVTs(*TM.getTargetLowering(), I.getType(), ValueVTs);
2813 unsigned NumValues = ValueVTs.size();
2814 if (NumValues == 0) return;
2816 SmallVector<SDValue, 4> Values(NumValues);
2817 SDValue Cond = getValue(I.getOperand(0));
2818 SDValue TrueVal = getValue(I.getOperand(1));
2819 SDValue FalseVal = getValue(I.getOperand(2));
2820 ISD::NodeType OpCode = Cond.getValueType().isVector() ?
2821 ISD::VSELECT : ISD::SELECT;
2823 for (unsigned i = 0; i != NumValues; ++i)
2824 Values[i] = DAG.getNode(OpCode, getCurSDLoc(),
2825 TrueVal.getNode()->getValueType(TrueVal.getResNo()+i),
2827 SDValue(TrueVal.getNode(),
2828 TrueVal.getResNo() + i),
2829 SDValue(FalseVal.getNode(),
2830 FalseVal.getResNo() + i));
2832 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
2833 DAG.getVTList(&ValueVTs[0], NumValues),
2834 &Values[0], NumValues));
2837 void SelectionDAGBuilder::visitTrunc(const User &I) {
2838 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2839 SDValue N = getValue(I.getOperand(0));
2840 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2841 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), DestVT, N));
2844 void SelectionDAGBuilder::visitZExt(const User &I) {
2845 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2846 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2847 SDValue N = getValue(I.getOperand(0));
2848 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2849 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurSDLoc(), DestVT, N));
2852 void SelectionDAGBuilder::visitSExt(const User &I) {
2853 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2854 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2855 SDValue N = getValue(I.getOperand(0));
2856 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2857 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurSDLoc(), DestVT, N));
2860 void SelectionDAGBuilder::visitFPTrunc(const User &I) {
2861 // FPTrunc is never a no-op cast, no need to check
2862 SDValue N = getValue(I.getOperand(0));
2863 const TargetLowering *TLI = TM.getTargetLowering();
2864 EVT DestVT = TLI->getValueType(I.getType());
2865 setValue(&I, DAG.getNode(ISD::FP_ROUND, getCurSDLoc(),
2867 DAG.getTargetConstant(0, TLI->getPointerTy())));
2870 void SelectionDAGBuilder::visitFPExt(const User &I) {
2871 // FPExt is never a no-op cast, no need to check
2872 SDValue N = getValue(I.getOperand(0));
2873 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2874 setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurSDLoc(), DestVT, N));
2877 void SelectionDAGBuilder::visitFPToUI(const User &I) {
2878 // FPToUI is never a no-op cast, no need to check
2879 SDValue N = getValue(I.getOperand(0));
2880 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2881 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurSDLoc(), DestVT, N));
2884 void SelectionDAGBuilder::visitFPToSI(const User &I) {
2885 // FPToSI is never a no-op cast, no need to check
2886 SDValue N = getValue(I.getOperand(0));
2887 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2888 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurSDLoc(), DestVT, N));
2891 void SelectionDAGBuilder::visitUIToFP(const User &I) {
2892 // UIToFP is never a no-op cast, no need to check
2893 SDValue N = getValue(I.getOperand(0));
2894 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2895 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurSDLoc(), DestVT, N));
2898 void SelectionDAGBuilder::visitSIToFP(const User &I) {
2899 // SIToFP is never a no-op cast, no need to check
2900 SDValue N = getValue(I.getOperand(0));
2901 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2902 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurSDLoc(), DestVT, N));
2905 void SelectionDAGBuilder::visitPtrToInt(const User &I) {
2906 // What to do depends on the size of the integer and the size of the pointer.
2907 // We can either truncate, zero extend, or no-op, accordingly.
2908 SDValue N = getValue(I.getOperand(0));
2909 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2910 setValue(&I, DAG.getZExtOrTrunc(N, getCurSDLoc(), DestVT));
2913 void SelectionDAGBuilder::visitIntToPtr(const User &I) {
2914 // What to do depends on the size of the integer and the size of the pointer.
2915 // We can either truncate, zero extend, or no-op, accordingly.
2916 SDValue N = getValue(I.getOperand(0));
2917 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2918 setValue(&I, DAG.getZExtOrTrunc(N, getCurSDLoc(), DestVT));
2921 void SelectionDAGBuilder::visitBitCast(const User &I) {
2922 SDValue N = getValue(I.getOperand(0));
2923 EVT DestVT = TM.getTargetLowering()->getValueType(I.getType());
2925 // BitCast assures us that source and destination are the same size so this is
2926 // either a BITCAST or a no-op.
2927 if (DestVT != N.getValueType())
2928 setValue(&I, DAG.getNode(ISD::BITCAST, getCurSDLoc(),
2929 DestVT, N)); // convert types.
2931 setValue(&I, N); // noop cast.
2934 void SelectionDAGBuilder::visitInsertElement(const User &I) {
2935 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2936 SDValue InVec = getValue(I.getOperand(0));
2937 SDValue InVal = getValue(I.getOperand(1));
2938 SDValue InIdx = DAG.getSExtOrTrunc(getValue(I.getOperand(2)),
2939 getCurSDLoc(), TLI.getVectorIdxTy());
2940 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurSDLoc(),
2941 TM.getTargetLowering()->getValueType(I.getType()),
2942 InVec, InVal, InIdx));
2945 void SelectionDAGBuilder::visitExtractElement(const User &I) {
2946 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2947 SDValue InVec = getValue(I.getOperand(0));
2948 SDValue InIdx = DAG.getSExtOrTrunc(getValue(I.getOperand(1)),
2949 getCurSDLoc(), TLI.getVectorIdxTy());
2950 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurSDLoc(),
2951 TM.getTargetLowering()->getValueType(I.getType()),
2955 // Utility for visitShuffleVector - Return true if every element in Mask,
2956 // beginning from position Pos and ending in Pos+Size, falls within the
2957 // specified sequential range [L, L+Pos). or is undef.
2958 static bool isSequentialInRange(const SmallVectorImpl<int> &Mask,
2959 unsigned Pos, unsigned Size, int Low) {
2960 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
2961 if (Mask[i] >= 0 && Mask[i] != Low)
2966 void SelectionDAGBuilder::visitShuffleVector(const User &I) {
2967 SDValue Src1 = getValue(I.getOperand(0));
2968 SDValue Src2 = getValue(I.getOperand(1));
2970 SmallVector<int, 8> Mask;
2971 ShuffleVectorInst::getShuffleMask(cast<Constant>(I.getOperand(2)), Mask);
2972 unsigned MaskNumElts = Mask.size();
2974 const TargetLowering *TLI = TM.getTargetLowering();
2975 EVT VT = TLI->getValueType(I.getType());
2976 EVT SrcVT = Src1.getValueType();
2977 unsigned SrcNumElts = SrcVT.getVectorNumElements();
2979 if (SrcNumElts == MaskNumElts) {
2980 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
2985 // Normalize the shuffle vector since mask and vector length don't match.
2986 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) {
2987 // Mask is longer than the source vectors and is a multiple of the source
2988 // vectors. We can use concatenate vector to make the mask and vectors
2990 if (SrcNumElts*2 == MaskNumElts) {
2991 // First check for Src1 in low and Src2 in high
2992 if (isSequentialInRange(Mask, 0, SrcNumElts, 0) &&
2993 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, SrcNumElts)) {
2994 // The shuffle is concatenating two vectors together.
2995 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurSDLoc(),
2999 // Then check for Src2 in low and Src1 in high
3000 if (isSequentialInRange(Mask, 0, SrcNumElts, SrcNumElts) &&
3001 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, 0)) {
3002 // The shuffle is concatenating two vectors together.
3003 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurSDLoc(),
3009 // Pad both vectors with undefs to make them the same length as the mask.
3010 unsigned NumConcat = MaskNumElts / SrcNumElts;
3011 bool Src1U = Src1.getOpcode() == ISD::UNDEF;
3012 bool Src2U = Src2.getOpcode() == ISD::UNDEF;
3013 SDValue UndefVal = DAG.getUNDEF(SrcVT);
3015 SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal);
3016 SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal);
3020 Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
3022 &MOps1[0], NumConcat);
3023 Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
3025 &MOps2[0], NumConcat);
3027 // Readjust mask for new input vector length.
3028 SmallVector<int, 8> MappedOps;
3029 for (unsigned i = 0; i != MaskNumElts; ++i) {
3031 if (Idx >= (int)SrcNumElts)
3032 Idx -= SrcNumElts - MaskNumElts;
3033 MappedOps.push_back(Idx);
3036 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
3041 if (SrcNumElts > MaskNumElts) {
3042 // Analyze the access pattern of the vector to see if we can extract
3043 // two subvectors and do the shuffle. The analysis is done by calculating
3044 // the range of elements the mask access on both vectors.
3045 int MinRange[2] = { static_cast<int>(SrcNumElts),
3046 static_cast<int>(SrcNumElts)};
3047 int MaxRange[2] = {-1, -1};
3049 for (unsigned i = 0; i != MaskNumElts; ++i) {
3055 if (Idx >= (int)SrcNumElts) {
3059 if (Idx > MaxRange[Input])
3060 MaxRange[Input] = Idx;
3061 if (Idx < MinRange[Input])
3062 MinRange[Input] = Idx;
3065 // Check if the access is smaller than the vector size and can we find
3066 // a reasonable extract index.
3067 int RangeUse[2] = { -1, -1 }; // 0 = Unused, 1 = Extract, -1 = Can not
3069 int StartIdx[2]; // StartIdx to extract from
3070 for (unsigned Input = 0; Input < 2; ++Input) {
3071 if (MinRange[Input] >= (int)SrcNumElts && MaxRange[Input] < 0) {
3072 RangeUse[Input] = 0; // Unused
3073 StartIdx[Input] = 0;
3077 // Find a good start index that is a multiple of the mask length. Then
3078 // see if the rest of the elements are in range.
3079 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts;
3080 if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts &&
3081 StartIdx[Input] + MaskNumElts <= SrcNumElts)
3082 RangeUse[Input] = 1; // Extract from a multiple of the mask length.
3085 if (RangeUse[0] == 0 && RangeUse[1] == 0) {
3086 setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used.
3089 if (RangeUse[0] >= 0 && RangeUse[1] >= 0) {
3090 // Extract appropriate subvector and generate a vector shuffle
3091 for (unsigned Input = 0; Input < 2; ++Input) {
3092 SDValue &Src = Input == 0 ? Src1 : Src2;
3093 if (RangeUse[Input] == 0)
3094 Src = DAG.getUNDEF(VT);
3096 Src = DAG.getNode(ISD::EXTRACT_SUBVECTOR, getCurSDLoc(), VT,
3097 Src, DAG.getConstant(StartIdx[Input],
3098 TLI->getVectorIdxTy()));
3101 // Calculate new mask.
3102 SmallVector<int, 8> MappedOps;
3103 for (unsigned i = 0; i != MaskNumElts; ++i) {
3106 if (Idx < (int)SrcNumElts)
3109 Idx -= SrcNumElts + StartIdx[1] - MaskNumElts;
3111 MappedOps.push_back(Idx);
3114 setValue(&I, DAG.getVectorShuffle(VT, getCurSDLoc(), Src1, Src2,
3120 // We can't use either concat vectors or extract subvectors so fall back to
3121 // replacing the shuffle with extract and build vector.
3122 // to insert and build vector.
3123 EVT EltVT = VT.getVectorElementType();
3124 EVT IdxVT = TLI->getVectorIdxTy();
3125 SmallVector<SDValue,8> Ops;
3126 for (unsigned i = 0; i != MaskNumElts; ++i) {
3131 Res = DAG.getUNDEF(EltVT);
3133 SDValue &Src = Idx < (int)SrcNumElts ? Src1 : Src2;
3134 if (Idx >= (int)SrcNumElts) Idx -= SrcNumElts;
3136 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurSDLoc(),
3137 EltVT, Src, DAG.getConstant(Idx, IdxVT));
3143 setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, getCurSDLoc(),
3144 VT, &Ops[0], Ops.size()));
3147 void SelectionDAGBuilder::visitInsertValue(const InsertValueInst &I) {
3148 const Value *Op0 = I.getOperand(0);
3149 const Value *Op1 = I.getOperand(1);
3150 Type *AggTy = I.getType();
3151 Type *ValTy = Op1->getType();
3152 bool IntoUndef = isa<UndefValue>(Op0);
3153 bool FromUndef = isa<UndefValue>(Op1);
3155 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
3157 const TargetLowering *TLI = TM.getTargetLowering();
3158 SmallVector<EVT, 4> AggValueVTs;
3159 ComputeValueVTs(*TLI, AggTy, AggValueVTs);
3160 SmallVector<EVT, 4> ValValueVTs;
3161 ComputeValueVTs(*TLI, ValTy, ValValueVTs);
3163 unsigned NumAggValues = AggValueVTs.size();
3164 unsigned NumValValues = ValValueVTs.size();
3165 SmallVector<SDValue, 4> Values(NumAggValues);
3167 SDValue Agg = getValue(Op0);
3169 // Copy the beginning value(s) from the original aggregate.
3170 for (; i != LinearIndex; ++i)
3171 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
3172 SDValue(Agg.getNode(), Agg.getResNo() + i);
3173 // Copy values from the inserted value(s).
3175 SDValue Val = getValue(Op1);
3176 for (; i != LinearIndex + NumValValues; ++i)
3177 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) :
3178 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex);
3180 // Copy remaining value(s) from the original aggregate.
3181 for (; i != NumAggValues; ++i)
3182 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
3183 SDValue(Agg.getNode(), Agg.getResNo() + i);
3185 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
3186 DAG.getVTList(&AggValueVTs[0], NumAggValues),
3187 &Values[0], NumAggValues));
3190 void SelectionDAGBuilder::visitExtractValue(const ExtractValueInst &I) {
3191 const Value *Op0 = I.getOperand(0);
3192 Type *AggTy = Op0->getType();
3193 Type *ValTy = I.getType();
3194 bool OutOfUndef = isa<UndefValue>(Op0);
3196 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
3198 const TargetLowering *TLI = TM.getTargetLowering();
3199 SmallVector<EVT, 4> ValValueVTs;
3200 ComputeValueVTs(*TLI, ValTy, ValValueVTs);
3202 unsigned NumValValues = ValValueVTs.size();
3204 // Ignore a extractvalue that produces an empty object
3205 if (!NumValValues) {
3206 setValue(&I, DAG.getUNDEF(MVT(MVT::Other)));
3210 SmallVector<SDValue, 4> Values(NumValValues);
3212 SDValue Agg = getValue(Op0);
3213 // Copy out the selected value(s).
3214 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i)
3215 Values[i - LinearIndex] =
3217 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) :
3218 SDValue(Agg.getNode(), Agg.getResNo() + i);
3220 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
3221 DAG.getVTList(&ValValueVTs[0], NumValValues),
3222 &Values[0], NumValValues));
3225 void SelectionDAGBuilder::visitGetElementPtr(const User &I) {
3226 SDValue N = getValue(I.getOperand(0));
3227 // Note that the pointer operand may be a vector of pointers. Take the scalar
3228 // element which holds a pointer.
3229 Type *Ty = I.getOperand(0)->getType()->getScalarType();
3231 for (GetElementPtrInst::const_op_iterator OI = I.op_begin()+1, E = I.op_end();
3233 const Value *Idx = *OI;
3234 if (StructType *StTy = dyn_cast<StructType>(Ty)) {
3235 unsigned Field = cast<Constant>(Idx)->getUniqueInteger().getZExtValue();
3238 uint64_t Offset = TD->getStructLayout(StTy)->getElementOffset(Field);
3239 N = DAG.getNode(ISD::ADD, getCurSDLoc(), N.getValueType(), N,
3240 DAG.getConstant(Offset, N.getValueType()));
3243 Ty = StTy->getElementType(Field);
3245 Ty = cast<SequentialType>(Ty)->getElementType();
3247 // If this is a constant subscript, handle it quickly.
3248 const TargetLowering *TLI = TM.getTargetLowering();
3249 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
3250 if (CI->isZero()) continue;
3252 TD->getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
3254 EVT PTy = TLI->getPointerTy();
3255 unsigned PtrBits = PTy.getSizeInBits();
3257 OffsVal = DAG.getNode(ISD::TRUNCATE, getCurSDLoc(),
3258 TLI->getPointerTy(),
3259 DAG.getConstant(Offs, MVT::i64));
3261 OffsVal = DAG.getIntPtrConstant(Offs);
3263 N = DAG.getNode(ISD::ADD, getCurSDLoc(), N.getValueType(), N,
3268 // N = N + Idx * ElementSize;
3269 APInt ElementSize = APInt(TLI->getPointerTy().getSizeInBits(),
3270 TD->getTypeAllocSize(Ty));
3271 SDValue IdxN = getValue(Idx);
3273 // If the index is smaller or larger than intptr_t, truncate or extend
3275 IdxN = DAG.getSExtOrTrunc(IdxN, getCurSDLoc(), N.getValueType());
3277 // If this is a multiply by a power of two, turn it into a shl
3278 // immediately. This is a very common case.
3279 if (ElementSize != 1) {
3280 if (ElementSize.isPowerOf2()) {
3281 unsigned Amt = ElementSize.logBase2();
3282 IdxN = DAG.getNode(ISD::SHL, getCurSDLoc(),
3283 N.getValueType(), IdxN,
3284 DAG.getConstant(Amt, IdxN.getValueType()));
3286 SDValue Scale = DAG.getConstant(ElementSize, IdxN.getValueType());
3287 IdxN = DAG.getNode(ISD::MUL, getCurSDLoc(),
3288 N.getValueType(), IdxN, Scale);
3292 N = DAG.getNode(ISD::ADD, getCurSDLoc(),
3293 N.getValueType(), N, IdxN);
3300 void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) {
3301 // If this is a fixed sized alloca in the entry block of the function,
3302 // allocate it statically on the stack.
3303 if (FuncInfo.StaticAllocaMap.count(&I))
3304 return; // getValue will auto-populate this.
3306 Type *Ty = I.getAllocatedType();
3307 const TargetLowering *TLI = TM.getTargetLowering();
3308 uint64_t TySize = TLI->getDataLayout()->getTypeAllocSize(Ty);
3310 std::max((unsigned)TLI->getDataLayout()->getPrefTypeAlignment(Ty),
3313 SDValue AllocSize = getValue(I.getArraySize());
3315 EVT IntPtr = TLI->getPointerTy();
3316 if (AllocSize.getValueType() != IntPtr)
3317 AllocSize = DAG.getZExtOrTrunc(AllocSize, getCurSDLoc(), IntPtr);
3319 AllocSize = DAG.getNode(ISD::MUL, getCurSDLoc(), IntPtr,
3321 DAG.getConstant(TySize, IntPtr));
3323 // Handle alignment. If the requested alignment is less than or equal to
3324 // the stack alignment, ignore it. If the size is greater than or equal to
3325 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
3326 unsigned StackAlign = TM.getFrameLowering()->getStackAlignment();
3327 if (Align <= StackAlign)
3330 // Round the size of the allocation up to the stack alignment size
3331 // by add SA-1 to the size.
3332 AllocSize = DAG.getNode(ISD::ADD, getCurSDLoc(),
3333 AllocSize.getValueType(), AllocSize,
3334 DAG.getIntPtrConstant(StackAlign-1));
3336 // Mask out the low bits for alignment purposes.
3337 AllocSize = DAG.getNode(ISD::AND, getCurSDLoc(),
3338 AllocSize.getValueType(), AllocSize,
3339 DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1)));
3341 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) };
3342 SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other);
3343 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, getCurSDLoc(),
3346 DAG.setRoot(DSA.getValue(1));
3348 // Inform the Frame Information that we have just allocated a variable-sized
3350 FuncInfo.MF->getFrameInfo()->CreateVariableSizedObject(Align ? Align : 1);
3353 void SelectionDAGBuilder::visitLoad(const LoadInst &I) {
3355 return visitAtomicLoad(I);
3357 const Value *SV = I.getOperand(0);
3358 SDValue Ptr = getValue(SV);
3360 Type *Ty = I.getType();
3362 bool isVolatile = I.isVolatile();
3363 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
3364 bool isInvariant = I.getMetadata("invariant.load") != 0;
3365 unsigned Alignment = I.getAlignment();
3366 const MDNode *TBAAInfo = I.getMetadata(LLVMContext::MD_tbaa);
3367 const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range);
3369 SmallVector<EVT, 4> ValueVTs;
3370 SmallVector<uint64_t, 4> Offsets;
3371 ComputeValueVTs(*TM.getTargetLowering(), Ty, ValueVTs, &Offsets);
3372 unsigned NumValues = ValueVTs.size();
3377 bool ConstantMemory = false;
3378 if (I.isVolatile() || NumValues > MaxParallelChains)
3379 // Serialize volatile loads with other side effects.
3381 else if (AA->pointsToConstantMemory(
3382 AliasAnalysis::Location(SV, AA->getTypeStoreSize(Ty), TBAAInfo))) {
3383 // Do not serialize (non-volatile) loads of constant memory with anything.
3384 Root = DAG.getEntryNode();
3385 ConstantMemory = true;
3387 // Do not serialize non-volatile loads against each other.
3388 Root = DAG.getRoot();
3391 SmallVector<SDValue, 4> Values(NumValues);
3392 SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains),
3394 EVT PtrVT = Ptr.getValueType();
3395 unsigned ChainI = 0;
3396 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3397 // Serializing loads here may result in excessive register pressure, and
3398 // TokenFactor places arbitrary choke points on the scheduler. SD scheduling
3399 // could recover a bit by hoisting nodes upward in the chain by recognizing
3400 // they are side-effect free or do not alias. The optimizer should really
3401 // avoid this case by converting large object/array copies to llvm.memcpy
3402 // (MaxParallelChains should always remain as failsafe).
3403 if (ChainI == MaxParallelChains) {
3404 assert(PendingLoads.empty() && "PendingLoads must be serialized first");
3405 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(),
3406 MVT::Other, &Chains[0], ChainI);
3410 SDValue A = DAG.getNode(ISD::ADD, getCurSDLoc(),
3412 DAG.getConstant(Offsets[i], PtrVT));
3413 SDValue L = DAG.getLoad(ValueVTs[i], getCurSDLoc(), Root,
3414 A, MachinePointerInfo(SV, Offsets[i]), isVolatile,
3415 isNonTemporal, isInvariant, Alignment, TBAAInfo,
3419 Chains[ChainI] = L.getValue(1);
3422 if (!ConstantMemory) {
3423 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(),
3424 MVT::Other, &Chains[0], ChainI);
3428 PendingLoads.push_back(Chain);
3431 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
3432 DAG.getVTList(&ValueVTs[0], NumValues),
3433 &Values[0], NumValues));
3436 void SelectionDAGBuilder::visitStore(const StoreInst &I) {
3438 return visitAtomicStore(I);
3440 const Value *SrcV = I.getOperand(0);
3441 const Value *PtrV = I.getOperand(1);
3443 SmallVector<EVT, 4> ValueVTs;
3444 SmallVector<uint64_t, 4> Offsets;
3445 ComputeValueVTs(*TM.getTargetLowering(), SrcV->getType(), ValueVTs, &Offsets);
3446 unsigned NumValues = ValueVTs.size();
3450 // Get the lowered operands. Note that we do this after
3451 // checking if NumResults is zero, because with zero results
3452 // the operands won't have values in the map.
3453 SDValue Src = getValue(SrcV);
3454 SDValue Ptr = getValue(PtrV);
3456 SDValue Root = getRoot();
3457 SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains),
3459 EVT PtrVT = Ptr.getValueType();
3460 bool isVolatile = I.isVolatile();
3461 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
3462 unsigned Alignment = I.getAlignment();
3463 const MDNode *TBAAInfo = I.getMetadata(LLVMContext::MD_tbaa);
3465 unsigned ChainI = 0;
3466 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3467 // See visitLoad comments.
3468 if (ChainI == MaxParallelChains) {
3469 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(),
3470 MVT::Other, &Chains[0], ChainI);
3474 SDValue Add = DAG.getNode(ISD::ADD, getCurSDLoc(), PtrVT, Ptr,
3475 DAG.getConstant(Offsets[i], PtrVT));
3476 SDValue St = DAG.getStore(Root, getCurSDLoc(),
3477 SDValue(Src.getNode(), Src.getResNo() + i),
3478 Add, MachinePointerInfo(PtrV, Offsets[i]),
3479 isVolatile, isNonTemporal, Alignment, TBAAInfo);
3480 Chains[ChainI] = St;
3483 SDValue StoreNode = DAG.getNode(ISD::TokenFactor, getCurSDLoc(),
3484 MVT::Other, &Chains[0], ChainI);
3485 DAG.setRoot(StoreNode);
3488 static SDValue InsertFenceForAtomic(SDValue Chain, AtomicOrdering Order,
3489 SynchronizationScope Scope,
3490 bool Before, SDLoc dl,
3492 const TargetLowering &TLI) {
3493 // Fence, if necessary
3495 if (Order == AcquireRelease || Order == SequentiallyConsistent)
3497 else if (Order == Acquire || Order == Monotonic)
3500 if (Order == AcquireRelease)
3502 else if (Order == Release || Order == Monotonic)
3507 Ops[1] = DAG.getConstant(Order, TLI.getPointerTy());
3508 Ops[2] = DAG.getConstant(Scope, TLI.getPointerTy());
3509 return DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops, 3);
3512 void SelectionDAGBuilder::visitAtomicCmpXchg(const AtomicCmpXchgInst &I) {
3513 SDLoc dl = getCurSDLoc();
3514 AtomicOrdering Order = I.getOrdering();
3515 SynchronizationScope Scope = I.getSynchScope();
3517 SDValue InChain = getRoot();
3519 const TargetLowering *TLI = TM.getTargetLowering();
3520 if (TLI->getInsertFencesForAtomic())
3521 InChain = InsertFenceForAtomic(InChain, Order, Scope, true, dl,
3525 DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, dl,
3526 getValue(I.getCompareOperand()).getSimpleValueType(),
3528 getValue(I.getPointerOperand()),
3529 getValue(I.getCompareOperand()),
3530 getValue(I.getNewValOperand()),
3531 MachinePointerInfo(I.getPointerOperand()), 0 /* Alignment */,
3532 TLI->getInsertFencesForAtomic() ? Monotonic : Order,
3535 SDValue OutChain = L.getValue(1);
3537 if (TLI->getInsertFencesForAtomic())
3538 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3542 DAG.setRoot(OutChain);
3545 void SelectionDAGBuilder::visitAtomicRMW(const AtomicRMWInst &I) {
3546 SDLoc dl = getCurSDLoc();
3548 switch (I.getOperation()) {
3549 default: llvm_unreachable("Unknown atomicrmw operation");
3550 case AtomicRMWInst::Xchg: NT = ISD::ATOMIC_SWAP; break;
3551 case AtomicRMWInst::Add: NT = ISD::ATOMIC_LOAD_ADD; break;
3552 case AtomicRMWInst::Sub: NT = ISD::ATOMIC_LOAD_SUB; break;
3553 case AtomicRMWInst::And: NT = ISD::ATOMIC_LOAD_AND; break;
3554 case AtomicRMWInst::Nand: NT = ISD::ATOMIC_LOAD_NAND; break;
3555 case AtomicRMWInst::Or: NT = ISD::ATOMIC_LOAD_OR; break;
3556 case AtomicRMWInst::Xor: NT = ISD::ATOMIC_LOAD_XOR; break;
3557 case AtomicRMWInst::Max: NT = ISD::ATOMIC_LOAD_MAX; break;
3558 case AtomicRMWInst::Min: NT = ISD::ATOMIC_LOAD_MIN; break;
3559 case AtomicRMWInst::UMax: NT = ISD::ATOMIC_LOAD_UMAX; break;
3560 case AtomicRMWInst::UMin: NT = ISD::ATOMIC_LOAD_UMIN; break;
3562 AtomicOrdering Order = I.getOrdering();
3563 SynchronizationScope Scope = I.getSynchScope();
3565 SDValue InChain = getRoot();
3567 const TargetLowering *TLI = TM.getTargetLowering();
3568 if (TLI->getInsertFencesForAtomic())
3569 InChain = InsertFenceForAtomic(InChain, Order, Scope, true, dl,
3573 DAG.getAtomic(NT, dl,
3574 getValue(I.getValOperand()).getSimpleValueType(),
3576 getValue(I.getPointerOperand()),
3577 getValue(I.getValOperand()),
3578 I.getPointerOperand(), 0 /* Alignment */,
3579 TLI->getInsertFencesForAtomic() ? Monotonic : Order,
3582 SDValue OutChain = L.getValue(1);
3584 if (TLI->getInsertFencesForAtomic())
3585 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3589 DAG.setRoot(OutChain);
3592 void SelectionDAGBuilder::visitFence(const FenceInst &I) {
3593 SDLoc dl = getCurSDLoc();
3594 const TargetLowering *TLI = TM.getTargetLowering();
3597 Ops[1] = DAG.getConstant(I.getOrdering(), TLI->getPointerTy());
3598 Ops[2] = DAG.getConstant(I.getSynchScope(), TLI->getPointerTy());
3599 DAG.setRoot(DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops, 3));
3602 void SelectionDAGBuilder::visitAtomicLoad(const LoadInst &I) {
3603 SDLoc dl = getCurSDLoc();
3604 AtomicOrdering Order = I.getOrdering();
3605 SynchronizationScope Scope = I.getSynchScope();
3607 SDValue InChain = getRoot();
3609 const TargetLowering *TLI = TM.getTargetLowering();
3610 EVT VT = TLI->getValueType(I.getType());
3612 if (I.getAlignment() < VT.getSizeInBits() / 8)
3613 report_fatal_error("Cannot generate unaligned atomic load");
3616 DAG.getAtomic(ISD::ATOMIC_LOAD, dl, VT, VT, InChain,
3617 getValue(I.getPointerOperand()),
3618 I.getPointerOperand(), I.getAlignment(),
3619 TLI->getInsertFencesForAtomic() ? Monotonic : Order,
3622 SDValue OutChain = L.getValue(1);
3624 if (TLI->getInsertFencesForAtomic())
3625 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3629 DAG.setRoot(OutChain);
3632 void SelectionDAGBuilder::visitAtomicStore(const StoreInst &I) {
3633 SDLoc dl = getCurSDLoc();
3635 AtomicOrdering Order = I.getOrdering();
3636 SynchronizationScope Scope = I.getSynchScope();
3638 SDValue InChain = getRoot();
3640 const TargetLowering *TLI = TM.getTargetLowering();
3641 EVT VT = TLI->getValueType(I.getValueOperand()->getType());
3643 if (I.getAlignment() < VT.getSizeInBits() / 8)
3644 report_fatal_error("Cannot generate unaligned atomic store");
3646 if (TLI->getInsertFencesForAtomic())
3647 InChain = InsertFenceForAtomic(InChain, Order, Scope, true, dl,
3651 DAG.getAtomic(ISD::ATOMIC_STORE, dl, VT,
3653 getValue(I.getPointerOperand()),
3654 getValue(I.getValueOperand()),
3655 I.getPointerOperand(), I.getAlignment(),
3656 TLI->getInsertFencesForAtomic() ? Monotonic : Order,
3659 if (TLI->getInsertFencesForAtomic())
3660 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3663 DAG.setRoot(OutChain);
3666 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
3668 void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I,
3669 unsigned Intrinsic) {
3670 bool HasChain = !I.doesNotAccessMemory();
3671 bool OnlyLoad = HasChain && I.onlyReadsMemory();
3673 // Build the operand list.
3674 SmallVector<SDValue, 8> Ops;
3675 if (HasChain) { // If this intrinsic has side-effects, chainify it.
3677 // We don't need to serialize loads against other loads.
3678 Ops.push_back(DAG.getRoot());
3680 Ops.push_back(getRoot());
3684 // Info is set by getTgtMemInstrinsic
3685 TargetLowering::IntrinsicInfo Info;
3686 const TargetLowering *TLI = TM.getTargetLowering();
3687 bool IsTgtIntrinsic = TLI->getTgtMemIntrinsic(Info, I, Intrinsic);
3689 // Add the intrinsic ID as an integer operand if it's not a target intrinsic.
3690 if (!IsTgtIntrinsic || Info.opc == ISD::INTRINSIC_VOID ||
3691 Info.opc == ISD::INTRINSIC_W_CHAIN)
3692 Ops.push_back(DAG.getTargetConstant(Intrinsic, TLI->getPointerTy()));
3694 // Add all operands of the call to the operand list.
3695 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) {
3696 SDValue Op = getValue(I.getArgOperand(i));
3700 SmallVector<EVT, 4> ValueVTs;
3701 ComputeValueVTs(*TLI, I.getType(), ValueVTs);
3704 ValueVTs.push_back(MVT::Other);
3706 SDVTList VTs = DAG.getVTList(ValueVTs.data(), ValueVTs.size());
3710 if (IsTgtIntrinsic) {
3711 // This is target intrinsic that touches memory
3712 Result = DAG.getMemIntrinsicNode(Info.opc, getCurSDLoc(),
3713 VTs, &Ops[0], Ops.size(),
3715 MachinePointerInfo(Info.ptrVal, Info.offset),
3716 Info.align, Info.vol,
3717 Info.readMem, Info.writeMem);
3718 } else if (!HasChain) {
3719 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurSDLoc(),
3720 VTs, &Ops[0], Ops.size());
3721 } else if (!I.getType()->isVoidTy()) {
3722 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurSDLoc(),
3723 VTs, &Ops[0], Ops.size());
3725 Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurSDLoc(),
3726 VTs, &Ops[0], Ops.size());
3730 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1);
3732 PendingLoads.push_back(Chain);
3737 if (!I.getType()->isVoidTy()) {
3738 if (VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
3739 EVT VT = TLI->getValueType(PTy);
3740 Result = DAG.getNode(ISD::BITCAST, getCurSDLoc(), VT, Result);
3743 setValue(&I, Result);
3747 /// GetSignificand - Get the significand and build it into a floating-point
3748 /// number with exponent of 1:
3750 /// Op = (Op & 0x007fffff) | 0x3f800000;
3752 /// where Op is the hexadecimal representation of floating point value.
3754 GetSignificand(SelectionDAG &DAG, SDValue Op, SDLoc dl) {
3755 SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3756 DAG.getConstant(0x007fffff, MVT::i32));
3757 SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1,
3758 DAG.getConstant(0x3f800000, MVT::i32));
3759 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, t2);
3762 /// GetExponent - Get the exponent:
3764 /// (float)(int)(((Op & 0x7f800000) >> 23) - 127);
3766 /// where Op is the hexadecimal representation of floating point value.
3768 GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI,
3770 SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3771 DAG.getConstant(0x7f800000, MVT::i32));
3772 SDValue t1 = DAG.getNode(ISD::SRL, dl, MVT::i32, t0,
3773 DAG.getConstant(23, TLI.getPointerTy()));
3774 SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1,
3775 DAG.getConstant(127, MVT::i32));
3776 return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2);
3779 /// getF32Constant - Get 32-bit floating point constant.
3781 getF32Constant(SelectionDAG &DAG, unsigned Flt) {
3782 return DAG.getConstantFP(APFloat(APFloat::IEEEsingle, APInt(32, Flt)),
3786 /// expandExp - Lower an exp intrinsic. Handles the special sequences for
3787 /// limited-precision mode.
3788 static SDValue expandExp(SDLoc dl, SDValue Op, SelectionDAG &DAG,
3789 const TargetLowering &TLI) {
3790 if (Op.getValueType() == MVT::f32 &&
3791 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3793 // Put the exponent in the right bit position for later addition to the
3796 // #define LOG2OFe 1.4426950f
3797 // IntegerPartOfX = ((int32_t)(X * LOG2OFe));
3798 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3799 getF32Constant(DAG, 0x3fb8aa3b));
3800 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3802 // FractionalPartOfX = (X * LOG2OFe) - (float)IntegerPartOfX;
3803 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3804 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3806 // IntegerPartOfX <<= 23;
3807 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3808 DAG.getConstant(23, TLI.getPointerTy()));
3810 SDValue TwoToFracPartOfX;
3811 if (LimitFloatPrecision <= 6) {
3812 // For floating-point precision of 6:
3814 // TwoToFractionalPartOfX =
3816 // (0.735607626f + 0.252464424f * x) * x;
3818 // error 0.0144103317, which is 6 bits
3819 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3820 getF32Constant(DAG, 0x3e814304));
3821 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3822 getF32Constant(DAG, 0x3f3c50c8));
3823 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3824 TwoToFracPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3825 getF32Constant(DAG, 0x3f7f5e7e));
3826 } else if (LimitFloatPrecision <= 12) {
3827 // For floating-point precision of 12:
3829 // TwoToFractionalPartOfX =
3832 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3834 // 0.000107046256 error, which is 13 to 14 bits
3835 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3836 getF32Constant(DAG, 0x3da235e3));
3837 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3838 getF32Constant(DAG, 0x3e65b8f3));
3839 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3840 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3841 getF32Constant(DAG, 0x3f324b07));
3842 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3843 TwoToFracPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3844 getF32Constant(DAG, 0x3f7ff8fd));
3845 } else { // LimitFloatPrecision <= 18
3846 // For floating-point precision of 18:
3848 // TwoToFractionalPartOfX =
3852 // (0.554906021e-1f +
3853 // (0.961591928e-2f +
3854 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3856 // error 2.47208000*10^(-7), which is better than 18 bits
3857 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3858 getF32Constant(DAG, 0x3924b03e));
3859 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3860 getF32Constant(DAG, 0x3ab24b87));
3861 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3862 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3863 getF32Constant(DAG, 0x3c1d8c17));
3864 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3865 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3866 getF32Constant(DAG, 0x3d634a1d));
3867 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3868 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3869 getF32Constant(DAG, 0x3e75fe14));
3870 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3871 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3872 getF32Constant(DAG, 0x3f317234));
3873 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3874 TwoToFracPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3875 getF32Constant(DAG, 0x3f800000));
3878 // Add the exponent into the result in integer domain.
3879 SDValue t13 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, TwoToFracPartOfX);
3880 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3881 DAG.getNode(ISD::ADD, dl, MVT::i32,
3882 t13, IntegerPartOfX));
3885 // No special expansion.
3886 return DAG.getNode(ISD::FEXP, dl, Op.getValueType(), Op);
3889 /// expandLog - Lower a log intrinsic. Handles the special sequences for
3890 /// limited-precision mode.
3891 static SDValue expandLog(SDLoc dl, SDValue Op, SelectionDAG &DAG,
3892 const TargetLowering &TLI) {
3893 if (Op.getValueType() == MVT::f32 &&
3894 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3895 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3897 // Scale the exponent by log(2) [0.69314718f].
3898 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3899 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3900 getF32Constant(DAG, 0x3f317218));
3902 // Get the significand and build it into a floating-point number with
3904 SDValue X = GetSignificand(DAG, Op1, dl);
3906 SDValue LogOfMantissa;
3907 if (LimitFloatPrecision <= 6) {
3908 // For floating-point precision of 6:
3912 // (1.4034025f - 0.23903021f * x) * x;
3914 // error 0.0034276066, which is better than 8 bits
3915 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3916 getF32Constant(DAG, 0xbe74c456));
3917 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3918 getF32Constant(DAG, 0x3fb3a2b1));
3919 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3920 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3921 getF32Constant(DAG, 0x3f949a29));
3922 } else if (LimitFloatPrecision <= 12) {
3923 // For floating-point precision of 12:
3929 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x;
3931 // error 0.000061011436, which is 14 bits
3932 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3933 getF32Constant(DAG, 0xbd67b6d6));
3934 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3935 getF32Constant(DAG, 0x3ee4f4b8));
3936 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3937 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3938 getF32Constant(DAG, 0x3fbc278b));
3939 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3940 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3941 getF32Constant(DAG, 0x40348e95));
3942 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3943 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3944 getF32Constant(DAG, 0x3fdef31a));
3945 } else { // LimitFloatPrecision <= 18
3946 // For floating-point precision of 18:
3954 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x;
3956 // error 0.0000023660568, which is better than 18 bits
3957 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3958 getF32Constant(DAG, 0xbc91e5ac));
3959 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3960 getF32Constant(DAG, 0x3e4350aa));
3961 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3962 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3963 getF32Constant(DAG, 0x3f60d3e3));
3964 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3965 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3966 getF32Constant(DAG, 0x4011cdf0));
3967 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3968 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3969 getF32Constant(DAG, 0x406cfd1c));
3970 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3971 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3972 getF32Constant(DAG, 0x408797cb));
3973 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3974 LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3975 getF32Constant(DAG, 0x4006dcab));
3978 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, LogOfMantissa);
3981 // No special expansion.
3982 return DAG.getNode(ISD::FLOG, dl, Op.getValueType(), Op);
3985 /// expandLog2 - Lower a log2 intrinsic. Handles the special sequences for
3986 /// limited-precision mode.
3987 static SDValue expandLog2(SDLoc dl, SDValue Op, SelectionDAG &DAG,
3988 const TargetLowering &TLI) {
3989 if (Op.getValueType() == MVT::f32 &&
3990 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3991 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3993 // Get the exponent.
3994 SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl);
3996 // Get the significand and build it into a floating-point number with
3998 SDValue X = GetSignificand(DAG, Op1, dl);
4000 // Different possible minimax approximations of significand in
4001 // floating-point for various degrees of accuracy over [1,2].
4002 SDValue Log2ofMantissa;
4003 if (LimitFloatPrecision <= 6) {
4004 // For floating-point precision of 6:
4006 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x;
4008 // error 0.0049451742, which is more than 7 bits
4009 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4010 getF32Constant(DAG, 0xbeb08fe0));
4011 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4012 getF32Constant(DAG, 0x40019463));
4013 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4014 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4015 getF32Constant(DAG, 0x3fd6633d));
4016 } else if (LimitFloatPrecision <= 12) {
4017 // For floating-point precision of 12:
4023 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x;
4025 // error 0.0000876136000, which is better than 13 bits
4026 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4027 getF32Constant(DAG, 0xbda7262e));
4028 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4029 getF32Constant(DAG, 0x3f25280b));
4030 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4031 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4032 getF32Constant(DAG, 0x4007b923));
4033 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4034 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4035 getF32Constant(DAG, 0x40823e2f));
4036 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4037 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
4038 getF32Constant(DAG, 0x4020d29c));
4039 } else { // LimitFloatPrecision <= 18
4040 // For floating-point precision of 18:
4049 // 0.25691327e-1f * x) * x) * x) * x) * x) * x;
4051 // error 0.0000018516, which is better than 18 bits
4052 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4053 getF32Constant(DAG, 0xbcd2769e));
4054 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4055 getF32Constant(DAG, 0x3e8ce0b9));
4056 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4057 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4058 getF32Constant(DAG, 0x3fa22ae7));
4059 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4060 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4061 getF32Constant(DAG, 0x40525723));
4062 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4063 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
4064 getF32Constant(DAG, 0x40aaf200));
4065 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4066 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4067 getF32Constant(DAG, 0x40c39dad));
4068 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4069 Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
4070 getF32Constant(DAG, 0x4042902c));
4073 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, Log2ofMantissa);
4076 // No special expansion.
4077 return DAG.getNode(ISD::FLOG2, dl, Op.getValueType(), Op);
4080 /// expandLog10 - Lower a log10 intrinsic. Handles the special sequences for
4081 /// limited-precision mode.
4082 static SDValue expandLog10(SDLoc dl, SDValue Op, SelectionDAG &DAG,
4083 const TargetLowering &TLI) {
4084 if (Op.getValueType() == MVT::f32 &&
4085 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4086 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
4088 // Scale the exponent by log10(2) [0.30102999f].
4089 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
4090 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
4091 getF32Constant(DAG, 0x3e9a209a));
4093 // Get the significand and build it into a floating-point number with
4095 SDValue X = GetSignificand(DAG, Op1, dl);
4097 SDValue Log10ofMantissa;
4098 if (LimitFloatPrecision <= 6) {
4099 // For floating-point precision of 6:
4101 // Log10ofMantissa =
4103 // (0.60948995f - 0.10380950f * x) * x;
4105 // error 0.0014886165, which is 6 bits
4106 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4107 getF32Constant(DAG, 0xbdd49a13));
4108 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4109 getF32Constant(DAG, 0x3f1c0789));
4110 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4111 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4112 getF32Constant(DAG, 0x3f011300));
4113 } else if (LimitFloatPrecision <= 12) {
4114 // For floating-point precision of 12:
4116 // Log10ofMantissa =
4119 // (-0.31664806f + 0.47637168e-1f * x) * x) * x;
4121 // error 0.00019228036, which is better than 12 bits
4122 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4123 getF32Constant(DAG, 0x3d431f31));
4124 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
4125 getF32Constant(DAG, 0x3ea21fb2));
4126 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4127 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4128 getF32Constant(DAG, 0x3f6ae232));
4129 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4130 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
4131 getF32Constant(DAG, 0x3f25f7c3));
4132 } else { // LimitFloatPrecision <= 18
4133 // For floating-point precision of 18:
4135 // Log10ofMantissa =
4140 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x;
4142 // error 0.0000037995730, which is better than 18 bits
4143 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4144 getF32Constant(DAG, 0x3c5d51ce));
4145 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
4146 getF32Constant(DAG, 0x3e00685a));
4147 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4148 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4149 getF32Constant(DAG, 0x3efb6798));
4150 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4151 SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
4152 getF32Constant(DAG, 0x3f88d192));
4153 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4154 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4155 getF32Constant(DAG, 0x3fc4316c));
4156 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4157 Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8,
4158 getF32Constant(DAG, 0x3f57ce70));
4161 return DAG.getNode(ISD::FADD, dl, MVT::f32, LogOfExponent, Log10ofMantissa);
4164 // No special expansion.
4165 return DAG.getNode(ISD::FLOG10, dl, Op.getValueType(), Op);
4168 /// expandExp2 - Lower an exp2 intrinsic. Handles the special sequences for
4169 /// limited-precision mode.
4170 static SDValue expandExp2(SDLoc dl, SDValue Op, SelectionDAG &DAG,
4171 const TargetLowering &TLI) {
4172 if (Op.getValueType() == MVT::f32 &&
4173 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4174 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Op);
4176 // FractionalPartOfX = x - (float)IntegerPartOfX;
4177 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
4178 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, Op, t1);
4180 // IntegerPartOfX <<= 23;
4181 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
4182 DAG.getConstant(23, TLI.getPointerTy()));
4184 SDValue TwoToFractionalPartOfX;
4185 if (LimitFloatPrecision <= 6) {
4186 // For floating-point precision of 6:
4188 // TwoToFractionalPartOfX =
4190 // (0.735607626f + 0.252464424f * x) * x;
4192 // error 0.0144103317, which is 6 bits
4193 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4194 getF32Constant(DAG, 0x3e814304));
4195 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4196 getF32Constant(DAG, 0x3f3c50c8));
4197 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4198 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4199 getF32Constant(DAG, 0x3f7f5e7e));
4200 } else if (LimitFloatPrecision <= 12) {
4201 // For floating-point precision of 12:
4203 // TwoToFractionalPartOfX =
4206 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
4208 // error 0.000107046256, which is 13 to 14 bits
4209 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4210 getF32Constant(DAG, 0x3da235e3));
4211 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4212 getF32Constant(DAG, 0x3e65b8f3));
4213 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4214 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4215 getF32Constant(DAG, 0x3f324b07));
4216 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4217 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4218 getF32Constant(DAG, 0x3f7ff8fd));
4219 } else { // LimitFloatPrecision <= 18
4220 // For floating-point precision of 18:
4222 // TwoToFractionalPartOfX =
4226 // (0.554906021e-1f +
4227 // (0.961591928e-2f +
4228 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
4229 // error 2.47208000*10^(-7), which is better than 18 bits
4230 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4231 getF32Constant(DAG, 0x3924b03e));
4232 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4233 getF32Constant(DAG, 0x3ab24b87));
4234 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4235 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4236 getF32Constant(DAG, 0x3c1d8c17));
4237 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4238 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4239 getF32Constant(DAG, 0x3d634a1d));
4240 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4241 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4242 getF32Constant(DAG, 0x3e75fe14));
4243 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4244 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
4245 getF32Constant(DAG, 0x3f317234));
4246 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
4247 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
4248 getF32Constant(DAG, 0x3f800000));
4251 // Add the exponent into the result in integer domain.
4252 SDValue t13 = DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4253 TwoToFractionalPartOfX);
4254 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4255 DAG.getNode(ISD::ADD, dl, MVT::i32,
4256 t13, IntegerPartOfX));
4259 // No special expansion.
4260 return DAG.getNode(ISD::FEXP2, dl, Op.getValueType(), Op);
4263 /// visitPow - Lower a pow intrinsic. Handles the special sequences for
4264 /// limited-precision mode with x == 10.0f.
4265 static SDValue expandPow(SDLoc dl, SDValue LHS, SDValue RHS,
4266 SelectionDAG &DAG, const TargetLowering &TLI) {
4267 bool IsExp10 = false;
4268 if (LHS.getValueType() == MVT::f32 && LHS.getValueType() == MVT::f32 &&
4269 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4270 if (ConstantFPSDNode *LHSC = dyn_cast<ConstantFPSDNode>(LHS)) {
4272 IsExp10 = LHSC->isExactlyValue(Ten);
4277 // Put the exponent in the right bit position for later addition to the
4280 // #define LOG2OF10 3.3219281f
4281 // IntegerPartOfX = (int32_t)(x * LOG2OF10);
4282 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, RHS,
4283 getF32Constant(DAG, 0x40549a78));
4284 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
4286 // FractionalPartOfX = x - (float)IntegerPartOfX;
4287 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
4288 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
4290 // IntegerPartOfX <<= 23;
4291 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
4292 DAG.getConstant(23, TLI.getPointerTy()));
4294 SDValue TwoToFractionalPartOfX;
4295 if (LimitFloatPrecision <= 6) {
4296 // For floating-point precision of 6:
4298 // twoToFractionalPartOfX =
4300 // (0.735607626f + 0.252464424f * x) * x;
4302 // error 0.0144103317, which is 6 bits
4303 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4304 getF32Constant(DAG, 0x3e814304));
4305 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4306 getF32Constant(DAG, 0x3f3c50c8));
4307 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4308 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4309 getF32Constant(DAG, 0x3f7f5e7e));
4310 } else if (LimitFloatPrecision <= 12) {
4311 // For floating-point precision of 12:
4313 // TwoToFractionalPartOfX =
4316 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
4318 // error 0.000107046256, which is 13 to 14 bits
4319 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4320 getF32Constant(DAG, 0x3da235e3));
4321 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4322 getF32Constant(DAG, 0x3e65b8f3));
4323 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4324 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4325 getF32Constant(DAG, 0x3f324b07));
4326 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4327 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4328 getF32Constant(DAG, 0x3f7ff8fd));
4329 } else { // LimitFloatPrecision <= 18
4330 // For floating-point precision of 18:
4332 // TwoToFractionalPartOfX =
4336 // (0.554906021e-1f +
4337 // (0.961591928e-2f +
4338 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
4339 // error 2.47208000*10^(-7), which is better than 18 bits
4340 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4341 getF32Constant(DAG, 0x3924b03e));
4342 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4343 getF32Constant(DAG, 0x3ab24b87));
4344 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4345 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4346 getF32Constant(DAG, 0x3c1d8c17));
4347 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4348 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4349 getF32Constant(DAG, 0x3d634a1d));
4350 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4351 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4352 getF32Constant(DAG, 0x3e75fe14));
4353 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4354 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
4355 getF32Constant(DAG, 0x3f317234));
4356 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
4357 TwoToFractionalPartOfX = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
4358 getF32Constant(DAG, 0x3f800000));
4361 SDValue t13 = DAG.getNode(ISD::BITCAST, dl,MVT::i32,TwoToFractionalPartOfX);
4362 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4363 DAG.getNode(ISD::ADD, dl, MVT::i32,
4364 t13, IntegerPartOfX));
4367 // No special expansion.
4368 return DAG.getNode(ISD::FPOW, dl, LHS.getValueType(), LHS, RHS);
4372 /// ExpandPowI - Expand a llvm.powi intrinsic.
4373 static SDValue ExpandPowI(SDLoc DL, SDValue LHS, SDValue RHS,
4374 SelectionDAG &DAG) {
4375 // If RHS is a constant, we can expand this out to a multiplication tree,
4376 // otherwise we end up lowering to a call to __powidf2 (for example). When
4377 // optimizing for size, we only want to do this if the expansion would produce
4378 // a small number of multiplies, otherwise we do the full expansion.
4379 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
4380 // Get the exponent as a positive value.
4381 unsigned Val = RHSC->getSExtValue();
4382 if ((int)Val < 0) Val = -Val;
4384 // powi(x, 0) -> 1.0
4386 return DAG.getConstantFP(1.0, LHS.getValueType());
4388 const Function *F = DAG.getMachineFunction().getFunction();
4389 if (!F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
4390 Attribute::OptimizeForSize) ||
4391 // If optimizing for size, don't insert too many multiplies. This
4392 // inserts up to 5 multiplies.
4393 CountPopulation_32(Val)+Log2_32(Val) < 7) {
4394 // We use the simple binary decomposition method to generate the multiply
4395 // sequence. There are more optimal ways to do this (for example,
4396 // powi(x,15) generates one more multiply than it should), but this has
4397 // the benefit of being both really simple and much better than a libcall.
4398 SDValue Res; // Logically starts equal to 1.0
4399 SDValue CurSquare = LHS;
4403 Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare);
4405 Res = CurSquare; // 1.0*CurSquare.
4408 CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(),
4409 CurSquare, CurSquare);
4413 // If the original was negative, invert the result, producing 1/(x*x*x).
4414 if (RHSC->getSExtValue() < 0)
4415 Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(),
4416 DAG.getConstantFP(1.0, LHS.getValueType()), Res);
4421 // Otherwise, expand to a libcall.
4422 return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS);
4425 // getTruncatedArgReg - Find underlying register used for an truncated
4427 static unsigned getTruncatedArgReg(const SDValue &N) {
4428 if (N.getOpcode() != ISD::TRUNCATE)
4431 const SDValue &Ext = N.getOperand(0);
4432 if (Ext.getOpcode() == ISD::AssertZext ||
4433 Ext.getOpcode() == ISD::AssertSext) {
4434 const SDValue &CFR = Ext.getOperand(0);
4435 if (CFR.getOpcode() == ISD::CopyFromReg)
4436 return cast<RegisterSDNode>(CFR.getOperand(1))->getReg();
4437 if (CFR.getOpcode() == ISD::TRUNCATE)
4438 return getTruncatedArgReg(CFR);
4443 /// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a function
4444 /// argument, create the corresponding DBG_VALUE machine instruction for it now.
4445 /// At the end of instruction selection, they will be inserted to the entry BB.
4447 SelectionDAGBuilder::EmitFuncArgumentDbgValue(const Value *V, MDNode *Variable,
4450 const Argument *Arg = dyn_cast<Argument>(V);
4454 MachineFunction &MF = DAG.getMachineFunction();
4455 const TargetInstrInfo *TII = DAG.getTarget().getInstrInfo();
4457 // Ignore inlined function arguments here.
4458 DIVariable DV(Variable);
4459 if (DV.isInlinedFnArgument(MF.getFunction()))
4462 Optional<MachineOperand> Op;
4463 // Some arguments' frame index is recorded during argument lowering.
4464 if (int FI = FuncInfo.getArgumentFrameIndex(Arg))
4465 Op = MachineOperand::CreateFI(FI);
4467 if (!Op && N.getNode()) {
4469 if (N.getOpcode() == ISD::CopyFromReg)
4470 Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg();
4472 Reg = getTruncatedArgReg(N);
4473 if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) {
4474 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4475 unsigned PR = RegInfo.getLiveInPhysReg(Reg);
4480 Op = MachineOperand::CreateReg(Reg, false);
4484 // Check if ValueMap has reg number.
4485 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
4486 if (VMI != FuncInfo.ValueMap.end())
4487 Op = MachineOperand::CreateReg(VMI->second, false);
4490 if (!Op && N.getNode())
4491 // Check if frame index is available.
4492 if (LoadSDNode *LNode = dyn_cast<LoadSDNode>(N.getNode()))
4493 if (FrameIndexSDNode *FINode =
4494 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode()))
4495 Op = MachineOperand::CreateFI(FINode->getIndex());
4500 // FIXME: This does not handle register-indirect values at offset 0.
4501 bool IsIndirect = Offset != 0;
4503 FuncInfo.ArgDbgValues.push_back(BuildMI(MF, getCurDebugLoc(),
4504 TII->get(TargetOpcode::DBG_VALUE),
4506 Op->getReg(), Offset, Variable));
4508 FuncInfo.ArgDbgValues.push_back(
4509 BuildMI(MF, getCurDebugLoc(), TII->get(TargetOpcode::DBG_VALUE))
4510 .addOperand(*Op).addImm(Offset).addMetadata(Variable));
4515 // VisualStudio defines setjmp as _setjmp
4516 #if defined(_MSC_VER) && defined(setjmp) && \
4517 !defined(setjmp_undefined_for_msvc)
4518 # pragma push_macro("setjmp")
4520 # define setjmp_undefined_for_msvc
4523 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
4524 /// we want to emit this as a call to a named external function, return the name
4525 /// otherwise lower it and return null.
4527 SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) {
4528 const TargetLowering *TLI = TM.getTargetLowering();
4529 SDLoc sdl = getCurSDLoc();
4530 DebugLoc dl = getCurDebugLoc();
4533 switch (Intrinsic) {
4535 // By default, turn this into a target intrinsic node.
4536 visitTargetIntrinsic(I, Intrinsic);
4538 case Intrinsic::vastart: visitVAStart(I); return 0;
4539 case Intrinsic::vaend: visitVAEnd(I); return 0;
4540 case Intrinsic::vacopy: visitVACopy(I); return 0;
4541 case Intrinsic::returnaddress:
4542 setValue(&I, DAG.getNode(ISD::RETURNADDR, sdl, TLI->getPointerTy(),
4543 getValue(I.getArgOperand(0))));
4545 case Intrinsic::frameaddress:
4546 setValue(&I, DAG.getNode(ISD::FRAMEADDR, sdl, TLI->getPointerTy(),
4547 getValue(I.getArgOperand(0))));
4549 case Intrinsic::setjmp:
4550 return &"_setjmp"[!TLI->usesUnderscoreSetJmp()];
4551 case Intrinsic::longjmp:
4552 return &"_longjmp"[!TLI->usesUnderscoreLongJmp()];
4553 case Intrinsic::memcpy: {
4554 // Assert for address < 256 since we support only user defined address
4556 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4558 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4560 "Unknown address space");
4561 SDValue Op1 = getValue(I.getArgOperand(0));
4562 SDValue Op2 = getValue(I.getArgOperand(1));
4563 SDValue Op3 = getValue(I.getArgOperand(2));
4564 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4566 Align = 1; // @llvm.memcpy defines 0 and 1 to both mean no alignment.
4567 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4568 DAG.setRoot(DAG.getMemcpy(getRoot(), sdl, Op1, Op2, Op3, Align, isVol, false,
4569 MachinePointerInfo(I.getArgOperand(0)),
4570 MachinePointerInfo(I.getArgOperand(1))));
4573 case Intrinsic::memset: {
4574 // Assert for address < 256 since we support only user defined address
4576 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4578 "Unknown address space");
4579 SDValue Op1 = getValue(I.getArgOperand(0));
4580 SDValue Op2 = getValue(I.getArgOperand(1));
4581 SDValue Op3 = getValue(I.getArgOperand(2));
4582 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4584 Align = 1; // @llvm.memset defines 0 and 1 to both mean no alignment.
4585 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4586 DAG.setRoot(DAG.getMemset(getRoot(), sdl, Op1, Op2, Op3, Align, isVol,
4587 MachinePointerInfo(I.getArgOperand(0))));
4590 case Intrinsic::memmove: {
4591 // Assert for address < 256 since we support only user defined address
4593 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4595 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4597 "Unknown address space");
4598 SDValue Op1 = getValue(I.getArgOperand(0));
4599 SDValue Op2 = getValue(I.getArgOperand(1));
4600 SDValue Op3 = getValue(I.getArgOperand(2));
4601 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4603 Align = 1; // @llvm.memmove defines 0 and 1 to both mean no alignment.
4604 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4605 DAG.setRoot(DAG.getMemmove(getRoot(), sdl, Op1, Op2, Op3, Align, isVol,
4606 MachinePointerInfo(I.getArgOperand(0)),
4607 MachinePointerInfo(I.getArgOperand(1))));
4610 case Intrinsic::dbg_declare: {
4611 const DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
4612 MDNode *Variable = DI.getVariable();
4613 const Value *Address = DI.getAddress();
4614 DIVariable DIVar(Variable);
4615 assert((!DIVar || DIVar.isVariable()) &&
4616 "Variable in DbgDeclareInst should be either null or a DIVariable.");
4617 if (!Address || !DIVar) {
4618 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4622 // Check if address has undef value.
4623 if (isa<UndefValue>(Address) ||
4624 (Address->use_empty() && !isa<Argument>(Address))) {
4625 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4629 SDValue &N = NodeMap[Address];
4630 if (!N.getNode() && isa<Argument>(Address))
4631 // Check unused arguments map.
4632 N = UnusedArgNodeMap[Address];
4635 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
4636 Address = BCI->getOperand(0);
4637 // Parameters are handled specially.
4639 (DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable ||
4640 isa<Argument>(Address));
4642 const AllocaInst *AI = dyn_cast<AllocaInst>(Address);
4644 if (isParameter && !AI) {
4645 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode());
4647 // Byval parameter. We have a frame index at this point.
4648 SDV = DAG.getDbgValue(Variable, FINode->getIndex(),
4649 0, dl, SDNodeOrder);
4651 // Address is an argument, so try to emit its dbg value using
4652 // virtual register info from the FuncInfo.ValueMap.
4653 EmitFuncArgumentDbgValue(Address, Variable, 0, N);
4657 SDV = DAG.getDbgValue(Variable, N.getNode(), N.getResNo(),
4658 0, dl, SDNodeOrder);
4660 // Can't do anything with other non-AI cases yet.
4661 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4662 DEBUG(dbgs() << "non-AllocaInst issue for Address: \n\t");
4663 DEBUG(Address->dump());
4666 DAG.AddDbgValue(SDV, N.getNode(), isParameter);
4668 // If Address is an argument then try to emit its dbg value using
4669 // virtual register info from the FuncInfo.ValueMap.
4670 if (!EmitFuncArgumentDbgValue(Address, Variable, 0, N)) {
4671 // If variable is pinned by a alloca in dominating bb then
4672 // use StaticAllocaMap.
4673 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Address)) {
4674 if (AI->getParent() != DI.getParent()) {
4675 DenseMap<const AllocaInst*, int>::iterator SI =
4676 FuncInfo.StaticAllocaMap.find(AI);
4677 if (SI != FuncInfo.StaticAllocaMap.end()) {
4678 SDV = DAG.getDbgValue(Variable, SI->second,
4679 0, dl, SDNodeOrder);
4680 DAG.AddDbgValue(SDV, 0, false);
4685 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4690 case Intrinsic::dbg_value: {
4691 const DbgValueInst &DI = cast<DbgValueInst>(I);
4692 DIVariable DIVar(DI.getVariable());
4693 assert((!DIVar || DIVar.isVariable()) &&
4694 "Variable in DbgValueInst should be either null or a DIVariable.");
4698 MDNode *Variable = DI.getVariable();
4699 uint64_t Offset = DI.getOffset();
4700 const Value *V = DI.getValue();
4705 if (isa<ConstantInt>(V) || isa<ConstantFP>(V) || isa<UndefValue>(V)) {
4706 SDV = DAG.getDbgValue(Variable, V, Offset, dl, SDNodeOrder);
4707 DAG.AddDbgValue(SDV, 0, false);
4709 // Do not use getValue() in here; we don't want to generate code at
4710 // this point if it hasn't been done yet.
4711 SDValue N = NodeMap[V];
4712 if (!N.getNode() && isa<Argument>(V))
4713 // Check unused arguments map.
4714 N = UnusedArgNodeMap[V];
4716 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, N)) {
4717 SDV = DAG.getDbgValue(Variable, N.getNode(),
4718 N.getResNo(), Offset, dl, SDNodeOrder);
4719 DAG.AddDbgValue(SDV, N.getNode(), false);
4721 } else if (!V->use_empty() ) {
4722 // Do not call getValue(V) yet, as we don't want to generate code.
4723 // Remember it for later.
4724 DanglingDebugInfo DDI(&DI, dl, SDNodeOrder);
4725 DanglingDebugInfoMap[V] = DDI;
4727 // We may expand this to cover more cases. One case where we have no
4728 // data available is an unreferenced parameter.
4729 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4733 // Build a debug info table entry.
4734 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
4735 V = BCI->getOperand(0);
4736 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
4737 // Don't handle byval struct arguments or VLAs, for example.
4739 DEBUG(dbgs() << "Dropping debug location info for:\n " << DI << "\n");
4740 DEBUG(dbgs() << " Last seen at:\n " << *V << "\n");
4743 DenseMap<const AllocaInst*, int>::iterator SI =
4744 FuncInfo.StaticAllocaMap.find(AI);
4745 if (SI == FuncInfo.StaticAllocaMap.end())
4747 int FI = SI->second;
4749 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4750 if (!DI.getDebugLoc().isUnknown() && MMI.hasDebugInfo())
4751 MMI.setVariableDbgInfo(Variable, FI, DI.getDebugLoc());
4755 case Intrinsic::eh_typeid_for: {
4756 // Find the type id for the given typeinfo.
4757 GlobalVariable *GV = ExtractTypeInfo(I.getArgOperand(0));
4758 unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(GV);
4759 Res = DAG.getConstant(TypeID, MVT::i32);
4764 case Intrinsic::eh_return_i32:
4765 case Intrinsic::eh_return_i64:
4766 DAG.getMachineFunction().getMMI().setCallsEHReturn(true);
4767 DAG.setRoot(DAG.getNode(ISD::EH_RETURN, sdl,
4770 getValue(I.getArgOperand(0)),
4771 getValue(I.getArgOperand(1))));
4773 case Intrinsic::eh_unwind_init:
4774 DAG.getMachineFunction().getMMI().setCallsUnwindInit(true);
4776 case Intrinsic::eh_dwarf_cfa: {
4777 SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getArgOperand(0)), sdl,
4778 TLI->getPointerTy());
4779 SDValue Offset = DAG.getNode(ISD::ADD, sdl,
4780 TLI->getPointerTy(),
4781 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, sdl,
4782 TLI->getPointerTy()),
4784 SDValue FA = DAG.getNode(ISD::FRAMEADDR, sdl,
4785 TLI->getPointerTy(),
4786 DAG.getConstant(0, TLI->getPointerTy()));
4787 setValue(&I, DAG.getNode(ISD::ADD, sdl, TLI->getPointerTy(),
4791 case Intrinsic::eh_sjlj_callsite: {
4792 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4793 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(0));
4794 assert(CI && "Non-constant call site value in eh.sjlj.callsite!");
4795 assert(MMI.getCurrentCallSite() == 0 && "Overlapping call sites!");
4797 MMI.setCurrentCallSite(CI->getZExtValue());
4800 case Intrinsic::eh_sjlj_functioncontext: {
4801 // Get and store the index of the function context.
4802 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4804 cast<AllocaInst>(I.getArgOperand(0)->stripPointerCasts());
4805 int FI = FuncInfo.StaticAllocaMap[FnCtx];
4806 MFI->setFunctionContextIndex(FI);
4809 case Intrinsic::eh_sjlj_setjmp: {
4812 Ops[1] = getValue(I.getArgOperand(0));
4813 SDValue Op = DAG.getNode(ISD::EH_SJLJ_SETJMP, sdl,
4814 DAG.getVTList(MVT::i32, MVT::Other),
4816 setValue(&I, Op.getValue(0));
4817 DAG.setRoot(Op.getValue(1));
4820 case Intrinsic::eh_sjlj_longjmp: {
4821 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, sdl, MVT::Other,
4822 getRoot(), getValue(I.getArgOperand(0))));
4826 case Intrinsic::x86_mmx_pslli_w:
4827 case Intrinsic::x86_mmx_pslli_d:
4828 case Intrinsic::x86_mmx_pslli_q:
4829 case Intrinsic::x86_mmx_psrli_w:
4830 case Intrinsic::x86_mmx_psrli_d:
4831 case Intrinsic::x86_mmx_psrli_q:
4832 case Intrinsic::x86_mmx_psrai_w:
4833 case Intrinsic::x86_mmx_psrai_d: {
4834 SDValue ShAmt = getValue(I.getArgOperand(1));
4835 if (isa<ConstantSDNode>(ShAmt)) {
4836 visitTargetIntrinsic(I, Intrinsic);
4839 unsigned NewIntrinsic = 0;
4840 EVT ShAmtVT = MVT::v2i32;
4841 switch (Intrinsic) {
4842 case Intrinsic::x86_mmx_pslli_w:
4843 NewIntrinsic = Intrinsic::x86_mmx_psll_w;
4845 case Intrinsic::x86_mmx_pslli_d:
4846 NewIntrinsic = Intrinsic::x86_mmx_psll_d;
4848 case Intrinsic::x86_mmx_pslli_q:
4849 NewIntrinsic = Intrinsic::x86_mmx_psll_q;
4851 case Intrinsic::x86_mmx_psrli_w:
4852 NewIntrinsic = Intrinsic::x86_mmx_psrl_w;
4854 case Intrinsic::x86_mmx_psrli_d:
4855 NewIntrinsic = Intrinsic::x86_mmx_psrl_d;
4857 case Intrinsic::x86_mmx_psrli_q:
4858 NewIntrinsic = Intrinsic::x86_mmx_psrl_q;
4860 case Intrinsic::x86_mmx_psrai_w:
4861 NewIntrinsic = Intrinsic::x86_mmx_psra_w;
4863 case Intrinsic::x86_mmx_psrai_d:
4864 NewIntrinsic = Intrinsic::x86_mmx_psra_d;
4866 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4869 // The vector shift intrinsics with scalars uses 32b shift amounts but
4870 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
4872 // We must do this early because v2i32 is not a legal type.
4875 ShOps[1] = DAG.getConstant(0, MVT::i32);
4876 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, sdl, ShAmtVT, &ShOps[0], 2);
4877 EVT DestVT = TLI->getValueType(I.getType());
4878 ShAmt = DAG.getNode(ISD::BITCAST, sdl, DestVT, ShAmt);
4879 Res = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, sdl, DestVT,
4880 DAG.getConstant(NewIntrinsic, MVT::i32),
4881 getValue(I.getArgOperand(0)), ShAmt);
4885 case Intrinsic::x86_avx_vinsertf128_pd_256:
4886 case Intrinsic::x86_avx_vinsertf128_ps_256:
4887 case Intrinsic::x86_avx_vinsertf128_si_256:
4888 case Intrinsic::x86_avx2_vinserti128: {
4889 EVT DestVT = TLI->getValueType(I.getType());
4890 EVT ElVT = TLI->getValueType(I.getArgOperand(1)->getType());
4891 uint64_t Idx = (cast<ConstantInt>(I.getArgOperand(2))->getZExtValue() & 1) *
4892 ElVT.getVectorNumElements();
4893 Res = DAG.getNode(ISD::INSERT_SUBVECTOR, sdl, DestVT,
4894 getValue(I.getArgOperand(0)),
4895 getValue(I.getArgOperand(1)),
4896 DAG.getConstant(Idx, TLI->getVectorIdxTy()));
4900 case Intrinsic::x86_avx_vextractf128_pd_256:
4901 case Intrinsic::x86_avx_vextractf128_ps_256:
4902 case Intrinsic::x86_avx_vextractf128_si_256:
4903 case Intrinsic::x86_avx2_vextracti128: {
4904 EVT DestVT = TLI->getValueType(I.getType());
4905 uint64_t Idx = (cast<ConstantInt>(I.getArgOperand(1))->getZExtValue() & 1) *
4906 DestVT.getVectorNumElements();
4907 Res = DAG.getNode(ISD::EXTRACT_SUBVECTOR, sdl, DestVT,
4908 getValue(I.getArgOperand(0)),
4909 DAG.getConstant(Idx, TLI->getVectorIdxTy()));
4913 case Intrinsic::convertff:
4914 case Intrinsic::convertfsi:
4915 case Intrinsic::convertfui:
4916 case Intrinsic::convertsif:
4917 case Intrinsic::convertuif:
4918 case Intrinsic::convertss:
4919 case Intrinsic::convertsu:
4920 case Intrinsic::convertus:
4921 case Intrinsic::convertuu: {
4922 ISD::CvtCode Code = ISD::CVT_INVALID;
4923 switch (Intrinsic) {
4924 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4925 case Intrinsic::convertff: Code = ISD::CVT_FF; break;
4926 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break;
4927 case Intrinsic::convertfui: Code = ISD::CVT_FU; break;
4928 case Intrinsic::convertsif: Code = ISD::CVT_SF; break;
4929 case Intrinsic::convertuif: Code = ISD::CVT_UF; break;
4930 case Intrinsic::convertss: Code = ISD::CVT_SS; break;
4931 case Intrinsic::convertsu: Code = ISD::CVT_SU; break;
4932 case Intrinsic::convertus: Code = ISD::CVT_US; break;
4933 case Intrinsic::convertuu: Code = ISD::CVT_UU; break;
4935 EVT DestVT = TLI->getValueType(I.getType());
4936 const Value *Op1 = I.getArgOperand(0);
4937 Res = DAG.getConvertRndSat(DestVT, sdl, getValue(Op1),
4938 DAG.getValueType(DestVT),
4939 DAG.getValueType(getValue(Op1).getValueType()),
4940 getValue(I.getArgOperand(1)),
4941 getValue(I.getArgOperand(2)),
4946 case Intrinsic::powi:
4947 setValue(&I, ExpandPowI(sdl, getValue(I.getArgOperand(0)),
4948 getValue(I.getArgOperand(1)), DAG));
4950 case Intrinsic::log:
4951 setValue(&I, expandLog(sdl, getValue(I.getArgOperand(0)), DAG, *TLI));
4953 case Intrinsic::log2:
4954 setValue(&I, expandLog2(sdl, getValue(I.getArgOperand(0)), DAG, *TLI));
4956 case Intrinsic::log10:
4957 setValue(&I, expandLog10(sdl, getValue(I.getArgOperand(0)), DAG, *TLI));
4959 case Intrinsic::exp:
4960 setValue(&I, expandExp(sdl, getValue(I.getArgOperand(0)), DAG, *TLI));
4962 case Intrinsic::exp2:
4963 setValue(&I, expandExp2(sdl, getValue(I.getArgOperand(0)), DAG, *TLI));
4965 case Intrinsic::pow:
4966 setValue(&I, expandPow(sdl, getValue(I.getArgOperand(0)),
4967 getValue(I.getArgOperand(1)), DAG, *TLI));
4969 case Intrinsic::sqrt:
4970 case Intrinsic::fabs:
4971 case Intrinsic::sin:
4972 case Intrinsic::cos:
4973 case Intrinsic::floor:
4974 case Intrinsic::ceil:
4975 case Intrinsic::trunc:
4976 case Intrinsic::rint:
4977 case Intrinsic::nearbyint:
4978 case Intrinsic::round: {
4980 switch (Intrinsic) {
4981 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4982 case Intrinsic::sqrt: Opcode = ISD::FSQRT; break;
4983 case Intrinsic::fabs: Opcode = ISD::FABS; break;
4984 case Intrinsic::sin: Opcode = ISD::FSIN; break;
4985 case Intrinsic::cos: Opcode = ISD::FCOS; break;
4986 case Intrinsic::floor: Opcode = ISD::FFLOOR; break;
4987 case Intrinsic::ceil: Opcode = ISD::FCEIL; break;
4988 case Intrinsic::trunc: Opcode = ISD::FTRUNC; break;
4989 case Intrinsic::rint: Opcode = ISD::FRINT; break;
4990 case Intrinsic::nearbyint: Opcode = ISD::FNEARBYINT; break;
4991 case Intrinsic::round: Opcode = ISD::FROUND; break;
4994 setValue(&I, DAG.getNode(Opcode, sdl,
4995 getValue(I.getArgOperand(0)).getValueType(),
4996 getValue(I.getArgOperand(0))));
4999 case Intrinsic::copysign:
5000 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, sdl,
5001 getValue(I.getArgOperand(0)).getValueType(),
5002 getValue(I.getArgOperand(0)),
5003 getValue(I.getArgOperand(1))));
5005 case Intrinsic::fma:
5006 setValue(&I, DAG.getNode(ISD::FMA, sdl,
5007 getValue(I.getArgOperand(0)).getValueType(),
5008 getValue(I.getArgOperand(0)),
5009 getValue(I.getArgOperand(1)),
5010 getValue(I.getArgOperand(2))));
5012 case Intrinsic::fmuladd: {
5013 EVT VT = TLI->getValueType(I.getType());
5014 if (TM.Options.AllowFPOpFusion != FPOpFusion::Strict &&
5015 TLI->isFMAFasterThanFMulAndFAdd(VT)) {
5016 setValue(&I, DAG.getNode(ISD::FMA, sdl,
5017 getValue(I.getArgOperand(0)).getValueType(),
5018 getValue(I.getArgOperand(0)),
5019 getValue(I.getArgOperand(1)),
5020 getValue(I.getArgOperand(2))));
5022 SDValue Mul = DAG.getNode(ISD::FMUL, sdl,
5023 getValue(I.getArgOperand(0)).getValueType(),
5024 getValue(I.getArgOperand(0)),
5025 getValue(I.getArgOperand(1)));
5026 SDValue Add = DAG.getNode(ISD::FADD, sdl,
5027 getValue(I.getArgOperand(0)).getValueType(),
5029 getValue(I.getArgOperand(2)));
5034 case Intrinsic::convert_to_fp16:
5035 setValue(&I, DAG.getNode(ISD::FP32_TO_FP16, sdl,
5036 MVT::i16, getValue(I.getArgOperand(0))));
5038 case Intrinsic::convert_from_fp16:
5039 setValue(&I, DAG.getNode(ISD::FP16_TO_FP32, sdl,
5040 MVT::f32, getValue(I.getArgOperand(0))));
5042 case Intrinsic::pcmarker: {
5043 SDValue Tmp = getValue(I.getArgOperand(0));
5044 DAG.setRoot(DAG.getNode(ISD::PCMARKER, sdl, MVT::Other, getRoot(), Tmp));
5047 case Intrinsic::readcyclecounter: {
5048 SDValue Op = getRoot();
5049 Res = DAG.getNode(ISD::READCYCLECOUNTER, sdl,
5050 DAG.getVTList(MVT::i64, MVT::Other),
5053 DAG.setRoot(Res.getValue(1));
5056 case Intrinsic::bswap:
5057 setValue(&I, DAG.getNode(ISD::BSWAP, sdl,
5058 getValue(I.getArgOperand(0)).getValueType(),
5059 getValue(I.getArgOperand(0))));
5061 case Intrinsic::cttz: {
5062 SDValue Arg = getValue(I.getArgOperand(0));
5063 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1));
5064 EVT Ty = Arg.getValueType();
5065 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTTZ : ISD::CTTZ_ZERO_UNDEF,
5069 case Intrinsic::ctlz: {
5070 SDValue Arg = getValue(I.getArgOperand(0));
5071 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1));
5072 EVT Ty = Arg.getValueType();
5073 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTLZ : ISD::CTLZ_ZERO_UNDEF,
5077 case Intrinsic::ctpop: {
5078 SDValue Arg = getValue(I.getArgOperand(0));
5079 EVT Ty = Arg.getValueType();
5080 setValue(&I, DAG.getNode(ISD::CTPOP, sdl, Ty, Arg));
5083 case Intrinsic::stacksave: {
5084 SDValue Op = getRoot();
5085 Res = DAG.getNode(ISD::STACKSAVE, sdl,
5086 DAG.getVTList(TLI->getPointerTy(), MVT::Other), &Op, 1);
5088 DAG.setRoot(Res.getValue(1));
5091 case Intrinsic::stackrestore: {
5092 Res = getValue(I.getArgOperand(0));
5093 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, sdl, MVT::Other, getRoot(), Res));
5096 case Intrinsic::stackprotector: {
5097 // Emit code into the DAG to store the stack guard onto the stack.
5098 MachineFunction &MF = DAG.getMachineFunction();
5099 MachineFrameInfo *MFI = MF.getFrameInfo();
5100 EVT PtrTy = TLI->getPointerTy();
5102 SDValue Src = getValue(I.getArgOperand(0)); // The guard's value.
5103 AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
5105 int FI = FuncInfo.StaticAllocaMap[Slot];
5106 MFI->setStackProtectorIndex(FI);
5108 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
5110 // Store the stack protector onto the stack.
5111 Res = DAG.getStore(getRoot(), sdl, Src, FIN,
5112 MachinePointerInfo::getFixedStack(FI),
5118 case Intrinsic::objectsize: {
5119 // If we don't know by now, we're never going to know.
5120 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1));
5122 assert(CI && "Non-constant type in __builtin_object_size?");
5124 SDValue Arg = getValue(I.getCalledValue());
5125 EVT Ty = Arg.getValueType();
5128 Res = DAG.getConstant(-1ULL, Ty);
5130 Res = DAG.getConstant(0, Ty);
5135 case Intrinsic::annotation:
5136 case Intrinsic::ptr_annotation:
5137 // Drop the intrinsic, but forward the value
5138 setValue(&I, getValue(I.getOperand(0)));
5140 case Intrinsic::var_annotation:
5141 // Discard annotate attributes
5144 case Intrinsic::init_trampoline: {
5145 const Function *F = cast<Function>(I.getArgOperand(1)->stripPointerCasts());
5149 Ops[1] = getValue(I.getArgOperand(0));
5150 Ops[2] = getValue(I.getArgOperand(1));
5151 Ops[3] = getValue(I.getArgOperand(2));
5152 Ops[4] = DAG.getSrcValue(I.getArgOperand(0));
5153 Ops[5] = DAG.getSrcValue(F);
5155 Res = DAG.getNode(ISD::INIT_TRAMPOLINE, sdl, MVT::Other, Ops, 6);
5160 case Intrinsic::adjust_trampoline: {
5161 setValue(&I, DAG.getNode(ISD::ADJUST_TRAMPOLINE, sdl,
5162 TLI->getPointerTy(),
5163 getValue(I.getArgOperand(0))));
5166 case Intrinsic::gcroot:
5168 const Value *Alloca = I.getArgOperand(0)->stripPointerCasts();
5169 const Constant *TypeMap = cast<Constant>(I.getArgOperand(1));
5171 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode());
5172 GFI->addStackRoot(FI->getIndex(), TypeMap);
5175 case Intrinsic::gcread:
5176 case Intrinsic::gcwrite:
5177 llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!");
5178 case Intrinsic::flt_rounds:
5179 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, sdl, MVT::i32));
5182 case Intrinsic::expect: {
5183 // Just replace __builtin_expect(exp, c) with EXP.
5184 setValue(&I, getValue(I.getArgOperand(0)));
5188 case Intrinsic::debugtrap:
5189 case Intrinsic::trap: {
5190 StringRef TrapFuncName = TM.Options.getTrapFunctionName();
5191 if (TrapFuncName.empty()) {
5192 ISD::NodeType Op = (Intrinsic == Intrinsic::trap) ?
5193 ISD::TRAP : ISD::DEBUGTRAP;
5194 DAG.setRoot(DAG.getNode(Op, sdl,MVT::Other, getRoot()));
5197 TargetLowering::ArgListTy Args;
5199 CallLoweringInfo CLI(getRoot(), I.getType(),
5200 false, false, false, false, 0, CallingConv::C,
5201 /*isTailCall=*/false,
5202 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
5203 DAG.getExternalSymbol(TrapFuncName.data(),
5204 TLI->getPointerTy()),
5206 std::pair<SDValue, SDValue> Result = TLI->LowerCallTo(CLI);
5207 DAG.setRoot(Result.second);
5211 case Intrinsic::uadd_with_overflow:
5212 case Intrinsic::sadd_with_overflow:
5213 case Intrinsic::usub_with_overflow:
5214 case Intrinsic::ssub_with_overflow:
5215 case Intrinsic::umul_with_overflow:
5216 case Intrinsic::smul_with_overflow: {
5218 switch (Intrinsic) {
5219 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
5220 case Intrinsic::uadd_with_overflow: Op = ISD::UADDO; break;
5221 case Intrinsic::sadd_with_overflow: Op = ISD::SADDO; break;
5222 case Intrinsic::usub_with_overflow: Op = ISD::USUBO; break;
5223 case Intrinsic::ssub_with_overflow: Op = ISD::SSUBO; break;
5224 case Intrinsic::umul_with_overflow: Op = ISD::UMULO; break;
5225 case Intrinsic::smul_with_overflow: Op = ISD::SMULO; break;
5227 SDValue Op1 = getValue(I.getArgOperand(0));
5228 SDValue Op2 = getValue(I.getArgOperand(1));
5230 SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1);
5231 setValue(&I, DAG.getNode(Op, sdl, VTs, Op1, Op2));
5234 case Intrinsic::prefetch: {
5236 unsigned rw = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
5238 Ops[1] = getValue(I.getArgOperand(0));
5239 Ops[2] = getValue(I.getArgOperand(1));
5240 Ops[3] = getValue(I.getArgOperand(2));
5241 Ops[4] = getValue(I.getArgOperand(3));
5242 DAG.setRoot(DAG.getMemIntrinsicNode(ISD::PREFETCH, sdl,
5243 DAG.getVTList(MVT::Other),
5245 EVT::getIntegerVT(*Context, 8),
5246 MachinePointerInfo(I.getArgOperand(0)),
5248 false, /* volatile */
5250 rw==1)); /* write */
5253 case Intrinsic::lifetime_start:
5254 case Intrinsic::lifetime_end: {
5255 bool IsStart = (Intrinsic == Intrinsic::lifetime_start);
5256 // Stack coloring is not enabled in O0, discard region information.
5257 if (TM.getOptLevel() == CodeGenOpt::None)
5260 SmallVector<Value *, 4> Allocas;
5261 GetUnderlyingObjects(I.getArgOperand(1), Allocas, TD);
5263 for (SmallVectorImpl<Value*>::iterator Object = Allocas.begin(),
5264 E = Allocas.end(); Object != E; ++Object) {
5265 AllocaInst *LifetimeObject = dyn_cast_or_null<AllocaInst>(*Object);
5267 // Could not find an Alloca.
5268 if (!LifetimeObject)
5271 int FI = FuncInfo.StaticAllocaMap[LifetimeObject];
5275 Ops[1] = DAG.getFrameIndex(FI, TLI->getPointerTy(), true);
5276 unsigned Opcode = (IsStart ? ISD::LIFETIME_START : ISD::LIFETIME_END);
5278 Res = DAG.getNode(Opcode, sdl, MVT::Other, Ops, 2);
5283 case Intrinsic::invariant_start:
5284 // Discard region information.
5285 setValue(&I, DAG.getUNDEF(TLI->getPointerTy()));
5287 case Intrinsic::invariant_end:
5288 // Discard region information.
5290 case Intrinsic::stackprotectorcheck: {
5291 // Do not actually emit anything for this basic block. Instead we initialize
5292 // the stack protector descriptor and export the guard variable so we can
5293 // access it in FinishBasicBlock.
5294 const BasicBlock *BB = I.getParent();
5295 SPDescriptor.initialize(BB, FuncInfo.MBBMap[BB], I);
5296 ExportFromCurrentBlock(SPDescriptor.getGuard());
5298 // Flush our exports since we are going to process a terminator.
5299 (void)getControlRoot();
5302 case Intrinsic::donothing:
5308 void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee,
5310 MachineBasicBlock *LandingPad) {
5311 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
5312 FunctionType *FTy = cast<FunctionType>(PT->getElementType());
5313 Type *RetTy = FTy->getReturnType();
5314 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5315 MCSymbol *BeginLabel = 0;
5317 TargetLowering::ArgListTy Args;
5318 TargetLowering::ArgListEntry Entry;
5319 Args.reserve(CS.arg_size());
5321 // Check whether the function can return without sret-demotion.
5322 SmallVector<ISD::OutputArg, 4> Outs;
5323 const TargetLowering *TLI = TM.getTargetLowering();
5324 GetReturnInfo(RetTy, CS.getAttributes(), Outs, *TLI);
5326 bool CanLowerReturn = TLI->CanLowerReturn(CS.getCallingConv(),
5327 DAG.getMachineFunction(),
5328 FTy->isVarArg(), Outs,
5331 SDValue DemoteStackSlot;
5332 int DemoteStackIdx = -100;
5334 if (!CanLowerReturn) {
5335 uint64_t TySize = TLI->getDataLayout()->getTypeAllocSize(
5336 FTy->getReturnType());
5337 unsigned Align = TLI->getDataLayout()->getPrefTypeAlignment(
5338 FTy->getReturnType());
5339 MachineFunction &MF = DAG.getMachineFunction();
5340 DemoteStackIdx = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
5341 Type *StackSlotPtrType = PointerType::getUnqual(FTy->getReturnType());
5343 DemoteStackSlot = DAG.getFrameIndex(DemoteStackIdx, TLI->getPointerTy());
5344 Entry.Node = DemoteStackSlot;
5345 Entry.Ty = StackSlotPtrType;
5346 Entry.isSExt = false;
5347 Entry.isZExt = false;
5348 Entry.isInReg = false;
5349 Entry.isSRet = true;
5350 Entry.isNest = false;
5351 Entry.isByVal = false;
5352 Entry.isReturned = false;
5353 Entry.Alignment = Align;
5354 Args.push_back(Entry);
5355 RetTy = Type::getVoidTy(FTy->getContext());
5358 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
5360 const Value *V = *i;
5363 if (V->getType()->isEmptyTy())
5366 SDValue ArgNode = getValue(V);
5367 Entry.Node = ArgNode; Entry.Ty = V->getType();
5369 unsigned attrInd = i - CS.arg_begin() + 1;
5370 Entry.isSExt = CS.paramHasAttr(attrInd, Attribute::SExt);
5371 Entry.isZExt = CS.paramHasAttr(attrInd, Attribute::ZExt);
5372 Entry.isInReg = CS.paramHasAttr(attrInd, Attribute::InReg);
5373 Entry.isSRet = CS.paramHasAttr(attrInd, Attribute::StructRet);
5374 Entry.isNest = CS.paramHasAttr(attrInd, Attribute::Nest);
5375 Entry.isByVal = CS.paramHasAttr(attrInd, Attribute::ByVal);
5376 Entry.isReturned = CS.paramHasAttr(attrInd, Attribute::Returned);
5377 Entry.Alignment = CS.getParamAlignment(attrInd);
5378 Args.push_back(Entry);
5382 // Insert a label before the invoke call to mark the try range. This can be
5383 // used to detect deletion of the invoke via the MachineModuleInfo.
5384 BeginLabel = MMI.getContext().CreateTempSymbol();
5386 // For SjLj, keep track of which landing pads go with which invokes
5387 // so as to maintain the ordering of pads in the LSDA.
5388 unsigned CallSiteIndex = MMI.getCurrentCallSite();
5389 if (CallSiteIndex) {
5390 MMI.setCallSiteBeginLabel(BeginLabel, CallSiteIndex);
5391 LPadToCallSiteMap[LandingPad].push_back(CallSiteIndex);
5393 // Now that the call site is handled, stop tracking it.
5394 MMI.setCurrentCallSite(0);
5397 // Both PendingLoads and PendingExports must be flushed here;
5398 // this call might not return.
5400 DAG.setRoot(DAG.getEHLabel(getCurSDLoc(), getControlRoot(), BeginLabel));
5403 // Check if target-independent constraints permit a tail call here.
5404 // Target-dependent constraints are checked within TLI->LowerCallTo.
5405 if (isTailCall && !isInTailCallPosition(CS, *TLI))
5409 CallLoweringInfo CLI(getRoot(), RetTy, FTy, isTailCall, Callee, Args, DAG,
5411 std::pair<SDValue,SDValue> Result = TLI->LowerCallTo(CLI);
5412 assert((isTailCall || Result.second.getNode()) &&
5413 "Non-null chain expected with non-tail call!");
5414 assert((Result.second.getNode() || !Result.first.getNode()) &&
5415 "Null value expected with tail call!");
5416 if (Result.first.getNode()) {
5417 setValue(CS.getInstruction(), Result.first);
5418 } else if (!CanLowerReturn && Result.second.getNode()) {
5419 // The instruction result is the result of loading from the
5420 // hidden sret parameter.
5421 SmallVector<EVT, 1> PVTs;
5422 Type *PtrRetTy = PointerType::getUnqual(FTy->getReturnType());
5424 ComputeValueVTs(*TLI, PtrRetTy, PVTs);
5425 assert(PVTs.size() == 1 && "Pointers should fit in one register");
5426 EVT PtrVT = PVTs[0];
5428 SmallVector<EVT, 4> RetTys;
5429 SmallVector<uint64_t, 4> Offsets;
5430 RetTy = FTy->getReturnType();
5431 ComputeValueVTs(*TLI, RetTy, RetTys, &Offsets);
5433 unsigned NumValues = RetTys.size();
5434 SmallVector<SDValue, 4> Values(NumValues);
5435 SmallVector<SDValue, 4> Chains(NumValues);
5437 for (unsigned i = 0; i < NumValues; ++i) {
5438 SDValue Add = DAG.getNode(ISD::ADD, getCurSDLoc(), PtrVT,
5440 DAG.getConstant(Offsets[i], PtrVT));
5441 SDValue L = DAG.getLoad(RetTys[i], getCurSDLoc(), Result.second, Add,
5442 MachinePointerInfo::getFixedStack(DemoteStackIdx, Offsets[i]),
5443 false, false, false, 1);
5445 Chains[i] = L.getValue(1);
5448 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(),
5449 MVT::Other, &Chains[0], NumValues);
5450 PendingLoads.push_back(Chain);
5452 setValue(CS.getInstruction(),
5453 DAG.getNode(ISD::MERGE_VALUES, getCurSDLoc(),
5454 DAG.getVTList(&RetTys[0], RetTys.size()),
5455 &Values[0], Values.size()));
5458 if (!Result.second.getNode()) {
5459 // As a special case, a null chain means that a tail call has been emitted and
5460 // the DAG root is already updated.
5463 // Since there's no actual continuation from this block, nothing can be
5464 // relying on us setting vregs for them.
5465 PendingExports.clear();
5467 DAG.setRoot(Result.second);
5471 // Insert a label at the end of the invoke call to mark the try range. This
5472 // can be used to detect deletion of the invoke via the MachineModuleInfo.
5473 MCSymbol *EndLabel = MMI.getContext().CreateTempSymbol();
5474 DAG.setRoot(DAG.getEHLabel(getCurSDLoc(), getRoot(), EndLabel));
5476 // Inform MachineModuleInfo of range.
5477 MMI.addInvoke(LandingPad, BeginLabel, EndLabel);
5481 /// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the
5482 /// value is equal or not-equal to zero.
5483 static bool IsOnlyUsedInZeroEqualityComparison(const Value *V) {
5484 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end();
5486 if (const ICmpInst *IC = dyn_cast<ICmpInst>(*UI))
5487 if (IC->isEquality())
5488 if (const Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
5489 if (C->isNullValue())
5491 // Unknown instruction.
5497 static SDValue getMemCmpLoad(const Value *PtrVal, MVT LoadVT,
5499 SelectionDAGBuilder &Builder) {
5501 // Check to see if this load can be trivially constant folded, e.g. if the
5502 // input is from a string literal.
5503 if (const Constant *LoadInput = dyn_cast<Constant>(PtrVal)) {
5504 // Cast pointer to the type we really want to load.
5505 LoadInput = ConstantExpr::getBitCast(const_cast<Constant *>(LoadInput),
5506 PointerType::getUnqual(LoadTy));
5508 if (const Constant *LoadCst =
5509 ConstantFoldLoadFromConstPtr(const_cast<Constant *>(LoadInput),
5511 return Builder.getValue(LoadCst);
5514 // Otherwise, we have to emit the load. If the pointer is to unfoldable but
5515 // still constant memory, the input chain can be the entry node.
5517 bool ConstantMemory = false;
5519 // Do not serialize (non-volatile) loads of constant memory with anything.
5520 if (Builder.AA->pointsToConstantMemory(PtrVal)) {
5521 Root = Builder.DAG.getEntryNode();
5522 ConstantMemory = true;
5524 // Do not serialize non-volatile loads against each other.
5525 Root = Builder.DAG.getRoot();
5528 SDValue Ptr = Builder.getValue(PtrVal);
5529 SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurSDLoc(), Root,
5530 Ptr, MachinePointerInfo(PtrVal),
5532 false /*nontemporal*/,
5533 false /*isinvariant*/, 1 /* align=1 */);
5535 if (!ConstantMemory)
5536 Builder.PendingLoads.push_back(LoadVal.getValue(1));
5540 /// processIntegerCallValue - Record the value for an instruction that
5541 /// produces an integer result, converting the type where necessary.
5542 void SelectionDAGBuilder::processIntegerCallValue(const Instruction &I,
5545 EVT VT = TM.getTargetLowering()->getValueType(I.getType(), true);
5547 Value = DAG.getSExtOrTrunc(Value, getCurSDLoc(), VT);
5549 Value = DAG.getZExtOrTrunc(Value, getCurSDLoc(), VT);
5550 setValue(&I, Value);
5553 /// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form.
5554 /// If so, return true and lower it, otherwise return false and it will be
5555 /// lowered like a normal call.
5556 bool SelectionDAGBuilder::visitMemCmpCall(const CallInst &I) {
5557 // Verify that the prototype makes sense. int memcmp(void*,void*,size_t)
5558 if (I.getNumArgOperands() != 3)
5561 const Value *LHS = I.getArgOperand(0), *RHS = I.getArgOperand(1);
5562 if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() ||
5563 !I.getArgOperand(2)->getType()->isIntegerTy() ||
5564 !I.getType()->isIntegerTy())
5567 const Value *Size = I.getArgOperand(2);
5568 const ConstantInt *CSize = dyn_cast<ConstantInt>(Size);
5569 if (CSize && CSize->getZExtValue() == 0) {
5570 EVT CallVT = TM.getTargetLowering()->getValueType(I.getType(), true);
5571 setValue(&I, DAG.getConstant(0, CallVT));
5575 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5576 std::pair<SDValue, SDValue> Res =
5577 TSI.EmitTargetCodeForMemcmp(DAG, getCurSDLoc(), DAG.getRoot(),
5578 getValue(LHS), getValue(RHS), getValue(Size),
5579 MachinePointerInfo(LHS),
5580 MachinePointerInfo(RHS));
5581 if (Res.first.getNode()) {
5582 processIntegerCallValue(I, Res.first, true);
5583 PendingLoads.push_back(Res.second);
5587 // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0
5588 // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0
5589 if (CSize && IsOnlyUsedInZeroEqualityComparison(&I)) {
5590 bool ActuallyDoIt = true;
5593 switch (CSize->getZExtValue()) {
5595 LoadVT = MVT::Other;
5597 ActuallyDoIt = false;
5601 LoadTy = Type::getInt16Ty(CSize->getContext());
5605 LoadTy = Type::getInt32Ty(CSize->getContext());
5609 LoadTy = Type::getInt64Ty(CSize->getContext());
5613 LoadVT = MVT::v4i32;
5614 LoadTy = Type::getInt32Ty(CSize->getContext());
5615 LoadTy = VectorType::get(LoadTy, 4);
5620 // This turns into unaligned loads. We only do this if the target natively
5621 // supports the MVT we'll be loading or if it is small enough (<= 4) that
5622 // we'll only produce a small number of byte loads.
5624 // Require that we can find a legal MVT, and only do this if the target
5625 // supports unaligned loads of that type. Expanding into byte loads would
5627 const TargetLowering *TLI = TM.getTargetLowering();
5628 if (ActuallyDoIt && CSize->getZExtValue() > 4) {
5629 // TODO: Handle 5 byte compare as 4-byte + 1 byte.
5630 // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads.
5631 if (!TLI->isTypeLegal(LoadVT) ||!TLI->allowsUnalignedMemoryAccesses(LoadVT))
5632 ActuallyDoIt = false;
5636 SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this);
5637 SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this);
5639 SDValue Res = DAG.getSetCC(getCurSDLoc(), MVT::i1, LHSVal, RHSVal,
5641 processIntegerCallValue(I, Res, false);
5650 /// visitMemChrCall -- See if we can lower a memchr call into an optimized
5651 /// form. If so, return true and lower it, otherwise return false and it
5652 /// will be lowered like a normal call.
5653 bool SelectionDAGBuilder::visitMemChrCall(const CallInst &I) {
5654 // Verify that the prototype makes sense. void *memchr(void *, int, size_t)
5655 if (I.getNumArgOperands() != 3)
5658 const Value *Src = I.getArgOperand(0);
5659 const Value *Char = I.getArgOperand(1);
5660 const Value *Length = I.getArgOperand(2);
5661 if (!Src->getType()->isPointerTy() ||
5662 !Char->getType()->isIntegerTy() ||
5663 !Length->getType()->isIntegerTy() ||
5664 !I.getType()->isPointerTy())
5667 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5668 std::pair<SDValue, SDValue> Res =
5669 TSI.EmitTargetCodeForMemchr(DAG, getCurSDLoc(), DAG.getRoot(),
5670 getValue(Src), getValue(Char), getValue(Length),
5671 MachinePointerInfo(Src));
5672 if (Res.first.getNode()) {
5673 setValue(&I, Res.first);
5674 PendingLoads.push_back(Res.second);
5681 /// visitStrCpyCall -- See if we can lower a strcpy or stpcpy call into an
5682 /// optimized form. If so, return true and lower it, otherwise return false
5683 /// and it will be lowered like a normal call.
5684 bool SelectionDAGBuilder::visitStrCpyCall(const CallInst &I, bool isStpcpy) {
5685 // Verify that the prototype makes sense. char *strcpy(char *, char *)
5686 if (I.getNumArgOperands() != 2)
5689 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5690 if (!Arg0->getType()->isPointerTy() ||
5691 !Arg1->getType()->isPointerTy() ||
5692 !I.getType()->isPointerTy())
5695 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5696 std::pair<SDValue, SDValue> Res =
5697 TSI.EmitTargetCodeForStrcpy(DAG, getCurSDLoc(), getRoot(),
5698 getValue(Arg0), getValue(Arg1),
5699 MachinePointerInfo(Arg0),
5700 MachinePointerInfo(Arg1), isStpcpy);
5701 if (Res.first.getNode()) {
5702 setValue(&I, Res.first);
5703 DAG.setRoot(Res.second);
5710 /// visitStrCmpCall - See if we can lower a call to strcmp in an optimized form.
5711 /// If so, return true and lower it, otherwise return false and it will be
5712 /// lowered like a normal call.
5713 bool SelectionDAGBuilder::visitStrCmpCall(const CallInst &I) {
5714 // Verify that the prototype makes sense. int strcmp(void*,void*)
5715 if (I.getNumArgOperands() != 2)
5718 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5719 if (!Arg0->getType()->isPointerTy() ||
5720 !Arg1->getType()->isPointerTy() ||
5721 !I.getType()->isIntegerTy())
5724 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5725 std::pair<SDValue, SDValue> Res =
5726 TSI.EmitTargetCodeForStrcmp(DAG, getCurSDLoc(), DAG.getRoot(),
5727 getValue(Arg0), getValue(Arg1),
5728 MachinePointerInfo(Arg0),
5729 MachinePointerInfo(Arg1));
5730 if (Res.first.getNode()) {
5731 processIntegerCallValue(I, Res.first, true);
5732 PendingLoads.push_back(Res.second);
5739 /// visitStrLenCall -- See if we can lower a strlen call into an optimized
5740 /// form. If so, return true and lower it, otherwise return false and it
5741 /// will be lowered like a normal call.
5742 bool SelectionDAGBuilder::visitStrLenCall(const CallInst &I) {
5743 // Verify that the prototype makes sense. size_t strlen(char *)
5744 if (I.getNumArgOperands() != 1)
5747 const Value *Arg0 = I.getArgOperand(0);
5748 if (!Arg0->getType()->isPointerTy() || !I.getType()->isIntegerTy())
5751 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5752 std::pair<SDValue, SDValue> Res =
5753 TSI.EmitTargetCodeForStrlen(DAG, getCurSDLoc(), DAG.getRoot(),
5754 getValue(Arg0), MachinePointerInfo(Arg0));
5755 if (Res.first.getNode()) {
5756 processIntegerCallValue(I, Res.first, false);
5757 PendingLoads.push_back(Res.second);
5764 /// visitStrNLenCall -- See if we can lower a strnlen call into an optimized
5765 /// form. If so, return true and lower it, otherwise return false and it
5766 /// will be lowered like a normal call.
5767 bool SelectionDAGBuilder::visitStrNLenCall(const CallInst &I) {
5768 // Verify that the prototype makes sense. size_t strnlen(char *, size_t)
5769 if (I.getNumArgOperands() != 2)
5772 const Value *Arg0 = I.getArgOperand(0), *Arg1 = I.getArgOperand(1);
5773 if (!Arg0->getType()->isPointerTy() ||
5774 !Arg1->getType()->isIntegerTy() ||
5775 !I.getType()->isIntegerTy())
5778 const TargetSelectionDAGInfo &TSI = DAG.getSelectionDAGInfo();
5779 std::pair<SDValue, SDValue> Res =
5780 TSI.EmitTargetCodeForStrnlen(DAG, getCurSDLoc(), DAG.getRoot(),
5781 getValue(Arg0), getValue(Arg1),
5782 MachinePointerInfo(Arg0));
5783 if (Res.first.getNode()) {
5784 processIntegerCallValue(I, Res.first, false);
5785 PendingLoads.push_back(Res.second);
5792 /// visitUnaryFloatCall - If a call instruction is a unary floating-point
5793 /// operation (as expected), translate it to an SDNode with the specified opcode
5794 /// and return true.
5795 bool SelectionDAGBuilder::visitUnaryFloatCall(const CallInst &I,
5797 // Sanity check that it really is a unary floating-point call.
5798 if (I.getNumArgOperands() != 1 ||
5799 !I.getArgOperand(0)->getType()->isFloatingPointTy() ||
5800 I.getType() != I.getArgOperand(0)->getType() ||
5801 !I.onlyReadsMemory())
5804 SDValue Tmp = getValue(I.getArgOperand(0));
5805 setValue(&I, DAG.getNode(Opcode, getCurSDLoc(), Tmp.getValueType(), Tmp));
5809 void SelectionDAGBuilder::visitCall(const CallInst &I) {
5810 // Handle inline assembly differently.
5811 if (isa<InlineAsm>(I.getCalledValue())) {
5816 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5817 ComputeUsesVAFloatArgument(I, &MMI);
5819 const char *RenameFn = 0;
5820 if (Function *F = I.getCalledFunction()) {
5821 if (F->isDeclaration()) {
5822 if (const TargetIntrinsicInfo *II = TM.getIntrinsicInfo()) {
5823 if (unsigned IID = II->getIntrinsicID(F)) {
5824 RenameFn = visitIntrinsicCall(I, IID);
5829 if (unsigned IID = F->getIntrinsicID()) {
5830 RenameFn = visitIntrinsicCall(I, IID);
5836 // Check for well-known libc/libm calls. If the function is internal, it
5837 // can't be a library call.
5839 if (!F->hasLocalLinkage() && F->hasName() &&
5840 LibInfo->getLibFunc(F->getName(), Func) &&
5841 LibInfo->hasOptimizedCodeGen(Func)) {
5844 case LibFunc::copysign:
5845 case LibFunc::copysignf:
5846 case LibFunc::copysignl:
5847 if (I.getNumArgOperands() == 2 && // Basic sanity checks.
5848 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5849 I.getType() == I.getArgOperand(0)->getType() &&
5850 I.getType() == I.getArgOperand(1)->getType() &&
5851 I.onlyReadsMemory()) {
5852 SDValue LHS = getValue(I.getArgOperand(0));
5853 SDValue RHS = getValue(I.getArgOperand(1));
5854 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurSDLoc(),
5855 LHS.getValueType(), LHS, RHS));
5860 case LibFunc::fabsf:
5861 case LibFunc::fabsl:
5862 if (visitUnaryFloatCall(I, ISD::FABS))
5868 if (visitUnaryFloatCall(I, ISD::FSIN))
5874 if (visitUnaryFloatCall(I, ISD::FCOS))
5878 case LibFunc::sqrtf:
5879 case LibFunc::sqrtl:
5880 case LibFunc::sqrt_finite:
5881 case LibFunc::sqrtf_finite:
5882 case LibFunc::sqrtl_finite:
5883 if (visitUnaryFloatCall(I, ISD::FSQRT))
5886 case LibFunc::floor:
5887 case LibFunc::floorf:
5888 case LibFunc::floorl:
5889 if (visitUnaryFloatCall(I, ISD::FFLOOR))
5892 case LibFunc::nearbyint:
5893 case LibFunc::nearbyintf:
5894 case LibFunc::nearbyintl:
5895 if (visitUnaryFloatCall(I, ISD::FNEARBYINT))
5899 case LibFunc::ceilf:
5900 case LibFunc::ceill:
5901 if (visitUnaryFloatCall(I, ISD::FCEIL))
5905 case LibFunc::rintf:
5906 case LibFunc::rintl:
5907 if (visitUnaryFloatCall(I, ISD::FRINT))
5910 case LibFunc::round:
5911 case LibFunc::roundf:
5912 case LibFunc::roundl:
5913 if (visitUnaryFloatCall(I, ISD::FROUND))
5916 case LibFunc::trunc:
5917 case LibFunc::truncf:
5918 case LibFunc::truncl:
5919 if (visitUnaryFloatCall(I, ISD::FTRUNC))
5923 case LibFunc::log2f:
5924 case LibFunc::log2l:
5925 if (visitUnaryFloatCall(I, ISD::FLOG2))
5929 case LibFunc::exp2f:
5930 case LibFunc::exp2l:
5931 if (visitUnaryFloatCall(I, ISD::FEXP2))
5934 case LibFunc::memcmp:
5935 if (visitMemCmpCall(I))
5938 case LibFunc::memchr:
5939 if (visitMemChrCall(I))
5942 case LibFunc::strcpy:
5943 if (visitStrCpyCall(I, false))
5946 case LibFunc::stpcpy:
5947 if (visitStrCpyCall(I, true))
5950 case LibFunc::strcmp:
5951 if (visitStrCmpCall(I))
5954 case LibFunc::strlen:
5955 if (visitStrLenCall(I))
5958 case LibFunc::strnlen:
5959 if (visitStrNLenCall(I))
5968 Callee = getValue(I.getCalledValue());
5970 Callee = DAG.getExternalSymbol(RenameFn,
5971 TM.getTargetLowering()->getPointerTy());
5973 // Check if we can potentially perform a tail call. More detailed checking is
5974 // be done within LowerCallTo, after more information about the call is known.
5975 LowerCallTo(&I, Callee, I.isTailCall());
5980 /// AsmOperandInfo - This contains information for each constraint that we are
5982 class SDISelAsmOperandInfo : public TargetLowering::AsmOperandInfo {
5984 /// CallOperand - If this is the result output operand or a clobber
5985 /// this is null, otherwise it is the incoming operand to the CallInst.
5986 /// This gets modified as the asm is processed.
5987 SDValue CallOperand;
5989 /// AssignedRegs - If this is a register or register class operand, this
5990 /// contains the set of register corresponding to the operand.
5991 RegsForValue AssignedRegs;
5993 explicit SDISelAsmOperandInfo(const TargetLowering::AsmOperandInfo &info)
5994 : TargetLowering::AsmOperandInfo(info), CallOperand(0,0) {
5997 /// getCallOperandValEVT - Return the EVT of the Value* that this operand
5998 /// corresponds to. If there is no Value* for this operand, it returns
6000 EVT getCallOperandValEVT(LLVMContext &Context,
6001 const TargetLowering &TLI,
6002 const DataLayout *TD) const {
6003 if (CallOperandVal == 0) return MVT::Other;
6005 if (isa<BasicBlock>(CallOperandVal))
6006 return TLI.getPointerTy();
6008 llvm::Type *OpTy = CallOperandVal->getType();
6010 // FIXME: code duplicated from TargetLowering::ParseConstraints().
6011 // If this is an indirect operand, the operand is a pointer to the
6014 llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
6016 report_fatal_error("Indirect operand for inline asm not a pointer!");
6017 OpTy = PtrTy->getElementType();
6020 // Look for vector wrapped in a struct. e.g. { <16 x i8> }.
6021 if (StructType *STy = dyn_cast<StructType>(OpTy))
6022 if (STy->getNumElements() == 1)
6023 OpTy = STy->getElementType(0);
6025 // If OpTy is not a single value, it may be a struct/union that we
6026 // can tile with integers.
6027 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
6028 unsigned BitSize = TD->getTypeSizeInBits(OpTy);
6037 OpTy = IntegerType::get(Context, BitSize);
6042 return TLI.getValueType(OpTy, true);
6046 typedef SmallVector<SDISelAsmOperandInfo,16> SDISelAsmOperandInfoVector;
6048 } // end anonymous namespace
6050 /// GetRegistersForValue - Assign registers (virtual or physical) for the
6051 /// specified operand. We prefer to assign virtual registers, to allow the
6052 /// register allocator to handle the assignment process. However, if the asm
6053 /// uses features that we can't model on machineinstrs, we have SDISel do the
6054 /// allocation. This produces generally horrible, but correct, code.
6056 /// OpInfo describes the operand.
6058 static void GetRegistersForValue(SelectionDAG &DAG,
6059 const TargetLowering &TLI,
6061 SDISelAsmOperandInfo &OpInfo) {
6062 LLVMContext &Context = *DAG.getContext();
6064 MachineFunction &MF = DAG.getMachineFunction();
6065 SmallVector<unsigned, 4> Regs;
6067 // If this is a constraint for a single physreg, or a constraint for a
6068 // register class, find it.
6069 std::pair<unsigned, const TargetRegisterClass*> PhysReg =
6070 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
6071 OpInfo.ConstraintVT);
6073 unsigned NumRegs = 1;
6074 if (OpInfo.ConstraintVT != MVT::Other) {
6075 // If this is a FP input in an integer register (or visa versa) insert a bit
6076 // cast of the input value. More generally, handle any case where the input
6077 // value disagrees with the register class we plan to stick this in.
6078 if (OpInfo.Type == InlineAsm::isInput &&
6079 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) {
6080 // Try to convert to the first EVT that the reg class contains. If the
6081 // types are identical size, use a bitcast to convert (e.g. two differing
6083 MVT RegVT = *PhysReg.second->vt_begin();
6084 if (RegVT.getSizeInBits() == OpInfo.ConstraintVT.getSizeInBits()) {
6085 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
6086 RegVT, OpInfo.CallOperand);
6087 OpInfo.ConstraintVT = RegVT;
6088 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) {
6089 // If the input is a FP value and we want it in FP registers, do a
6090 // bitcast to the corresponding integer type. This turns an f64 value
6091 // into i64, which can be passed with two i32 values on a 32-bit
6093 RegVT = MVT::getIntegerVT(OpInfo.ConstraintVT.getSizeInBits());
6094 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
6095 RegVT, OpInfo.CallOperand);
6096 OpInfo.ConstraintVT = RegVT;
6100 NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT);
6104 EVT ValueVT = OpInfo.ConstraintVT;
6106 // If this is a constraint for a specific physical register, like {r17},
6108 if (unsigned AssignedReg = PhysReg.first) {
6109 const TargetRegisterClass *RC = PhysReg.second;
6110 if (OpInfo.ConstraintVT == MVT::Other)
6111 ValueVT = *RC->vt_begin();
6113 // Get the actual register value type. This is important, because the user
6114 // may have asked for (e.g.) the AX register in i32 type. We need to
6115 // remember that AX is actually i16 to get the right extension.
6116 RegVT = *RC->vt_begin();
6118 // This is a explicit reference to a physical register.
6119 Regs.push_back(AssignedReg);
6121 // If this is an expanded reference, add the rest of the regs to Regs.
6123 TargetRegisterClass::iterator I = RC->begin();
6124 for (; *I != AssignedReg; ++I)
6125 assert(I != RC->end() && "Didn't find reg!");
6127 // Already added the first reg.
6129 for (; NumRegs; --NumRegs, ++I) {
6130 assert(I != RC->end() && "Ran out of registers to allocate!");
6135 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
6139 // Otherwise, if this was a reference to an LLVM register class, create vregs
6140 // for this reference.
6141 if (const TargetRegisterClass *RC = PhysReg.second) {
6142 RegVT = *RC->vt_begin();
6143 if (OpInfo.ConstraintVT == MVT::Other)
6146 // Create the appropriate number of virtual registers.
6147 MachineRegisterInfo &RegInfo = MF.getRegInfo();
6148 for (; NumRegs; --NumRegs)
6149 Regs.push_back(RegInfo.createVirtualRegister(RC));
6151 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
6155 // Otherwise, we couldn't allocate enough registers for this.
6158 /// visitInlineAsm - Handle a call to an InlineAsm object.
6160 void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) {
6161 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
6163 /// ConstraintOperands - Information about all of the constraints.
6164 SDISelAsmOperandInfoVector ConstraintOperands;
6166 const TargetLowering *TLI = TM.getTargetLowering();
6167 TargetLowering::AsmOperandInfoVector
6168 TargetConstraints = TLI->ParseConstraints(CS);
6170 bool hasMemory = false;
6172 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
6173 unsigned ResNo = 0; // ResNo - The result number of the next output.
6174 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
6175 ConstraintOperands.push_back(SDISelAsmOperandInfo(TargetConstraints[i]));
6176 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
6178 MVT OpVT = MVT::Other;
6180 // Compute the value type for each operand.
6181 switch (OpInfo.Type) {
6182 case InlineAsm::isOutput:
6183 // Indirect outputs just consume an argument.
6184 if (OpInfo.isIndirect) {
6185 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
6189 // The return value of the call is this value. As such, there is no
6190 // corresponding argument.
6191 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
6192 if (StructType *STy = dyn_cast<StructType>(CS.getType())) {
6193 OpVT = TLI->getSimpleValueType(STy->getElementType(ResNo));
6195 assert(ResNo == 0 && "Asm only has one result!");
6196 OpVT = TLI->getSimpleValueType(CS.getType());
6200 case InlineAsm::isInput:
6201 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
6203 case InlineAsm::isClobber:
6208 // If this is an input or an indirect output, process the call argument.
6209 // BasicBlocks are labels, currently appearing only in asm's.
6210 if (OpInfo.CallOperandVal) {
6211 if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) {
6212 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]);
6214 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
6217 OpVT = OpInfo.getCallOperandValEVT(*DAG.getContext(), *TLI, TD).
6221 OpInfo.ConstraintVT = OpVT;
6223 // Indirect operand accesses access memory.
6224 if (OpInfo.isIndirect)
6227 for (unsigned j = 0, ee = OpInfo.Codes.size(); j != ee; ++j) {
6228 TargetLowering::ConstraintType
6229 CType = TLI->getConstraintType(OpInfo.Codes[j]);
6230 if (CType == TargetLowering::C_Memory) {
6238 SDValue Chain, Flag;
6240 // We won't need to flush pending loads if this asm doesn't touch
6241 // memory and is nonvolatile.
6242 if (hasMemory || IA->hasSideEffects())
6245 Chain = DAG.getRoot();
6247 // Second pass over the constraints: compute which constraint option to use
6248 // and assign registers to constraints that want a specific physreg.
6249 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6250 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6252 // If this is an output operand with a matching input operand, look up the
6253 // matching input. If their types mismatch, e.g. one is an integer, the
6254 // other is floating point, or their sizes are different, flag it as an
6256 if (OpInfo.hasMatchingInput()) {
6257 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
6259 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
6260 std::pair<unsigned, const TargetRegisterClass*> MatchRC =
6261 TLI->getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
6262 OpInfo.ConstraintVT);
6263 std::pair<unsigned, const TargetRegisterClass*> InputRC =
6264 TLI->getRegForInlineAsmConstraint(Input.ConstraintCode,
6265 Input.ConstraintVT);
6266 if ((OpInfo.ConstraintVT.isInteger() !=
6267 Input.ConstraintVT.isInteger()) ||
6268 (MatchRC.second != InputRC.second)) {
6269 report_fatal_error("Unsupported asm: input constraint"
6270 " with a matching output constraint of"
6271 " incompatible type!");
6273 Input.ConstraintVT = OpInfo.ConstraintVT;
6277 // Compute the constraint code and ConstraintType to use.
6278 TLI->ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, &DAG);
6280 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
6281 OpInfo.Type == InlineAsm::isClobber)
6284 // If this is a memory input, and if the operand is not indirect, do what we
6285 // need to to provide an address for the memory input.
6286 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
6287 !OpInfo.isIndirect) {
6288 assert((OpInfo.isMultipleAlternative ||
6289 (OpInfo.Type == InlineAsm::isInput)) &&
6290 "Can only indirectify direct input operands!");
6292 // Memory operands really want the address of the value. If we don't have
6293 // an indirect input, put it in the constpool if we can, otherwise spill
6294 // it to a stack slot.
6295 // TODO: This isn't quite right. We need to handle these according to
6296 // the addressing mode that the constraint wants. Also, this may take
6297 // an additional register for the computation and we don't want that
6300 // If the operand is a float, integer, or vector constant, spill to a
6301 // constant pool entry to get its address.
6302 const Value *OpVal = OpInfo.CallOperandVal;
6303 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
6304 isa<ConstantVector>(OpVal) || isa<ConstantDataVector>(OpVal)) {
6305 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal),
6306 TLI->getPointerTy());
6308 // Otherwise, create a stack slot and emit a store to it before the
6310 Type *Ty = OpVal->getType();
6311 uint64_t TySize = TLI->getDataLayout()->getTypeAllocSize(Ty);
6312 unsigned Align = TLI->getDataLayout()->getPrefTypeAlignment(Ty);
6313 MachineFunction &MF = DAG.getMachineFunction();
6314 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
6315 SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI->getPointerTy());
6316 Chain = DAG.getStore(Chain, getCurSDLoc(),
6317 OpInfo.CallOperand, StackSlot,
6318 MachinePointerInfo::getFixedStack(SSFI),
6320 OpInfo.CallOperand = StackSlot;
6323 // There is no longer a Value* corresponding to this operand.
6324 OpInfo.CallOperandVal = 0;
6326 // It is now an indirect operand.
6327 OpInfo.isIndirect = true;
6330 // If this constraint is for a specific register, allocate it before
6332 if (OpInfo.ConstraintType == TargetLowering::C_Register)
6333 GetRegistersForValue(DAG, *TLI, getCurSDLoc(), OpInfo);
6336 // Second pass - Loop over all of the operands, assigning virtual or physregs
6337 // to register class operands.
6338 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6339 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6341 // C_Register operands have already been allocated, Other/Memory don't need
6343 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
6344 GetRegistersForValue(DAG, *TLI, getCurSDLoc(), OpInfo);
6347 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
6348 std::vector<SDValue> AsmNodeOperands;
6349 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain
6350 AsmNodeOperands.push_back(
6351 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(),
6352 TLI->getPointerTy()));
6354 // If we have a !srcloc metadata node associated with it, we want to attach
6355 // this to the ultimately generated inline asm machineinstr. To do this, we
6356 // pass in the third operand as this (potentially null) inline asm MDNode.
6357 const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc");
6358 AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc));
6360 // Remember the HasSideEffect, AlignStack, AsmDialect, MayLoad and MayStore
6361 // bits as operand 3.
6362 unsigned ExtraInfo = 0;
6363 if (IA->hasSideEffects())
6364 ExtraInfo |= InlineAsm::Extra_HasSideEffects;
6365 if (IA->isAlignStack())
6366 ExtraInfo |= InlineAsm::Extra_IsAlignStack;
6367 // Set the asm dialect.
6368 ExtraInfo |= IA->getDialect() * InlineAsm::Extra_AsmDialect;
6370 // Determine if this InlineAsm MayLoad or MayStore based on the constraints.
6371 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
6372 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
6374 // Compute the constraint code and ConstraintType to use.
6375 TLI->ComputeConstraintToUse(OpInfo, SDValue());
6377 // Ideally, we would only check against memory constraints. However, the
6378 // meaning of an other constraint can be target-specific and we can't easily
6379 // reason about it. Therefore, be conservative and set MayLoad/MayStore
6380 // for other constriants as well.
6381 if (OpInfo.ConstraintType == TargetLowering::C_Memory ||
6382 OpInfo.ConstraintType == TargetLowering::C_Other) {
6383 if (OpInfo.Type == InlineAsm::isInput)
6384 ExtraInfo |= InlineAsm::Extra_MayLoad;
6385 else if (OpInfo.Type == InlineAsm::isOutput)
6386 ExtraInfo |= InlineAsm::Extra_MayStore;
6387 else if (OpInfo.Type == InlineAsm::isClobber)
6388 ExtraInfo |= (InlineAsm::Extra_MayLoad | InlineAsm::Extra_MayStore);
6392 AsmNodeOperands.push_back(DAG.getTargetConstant(ExtraInfo,
6393 TLI->getPointerTy()));
6395 // Loop over all of the inputs, copying the operand values into the
6396 // appropriate registers and processing the output regs.
6397 RegsForValue RetValRegs;
6399 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
6400 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
6402 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6403 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6405 switch (OpInfo.Type) {
6406 case InlineAsm::isOutput: {
6407 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
6408 OpInfo.ConstraintType != TargetLowering::C_Register) {
6409 // Memory output, or 'other' output (e.g. 'X' constraint).
6410 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
6412 // Add information to the INLINEASM node to know about this output.
6413 unsigned OpFlags = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
6414 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlags,
6415 TLI->getPointerTy()));
6416 AsmNodeOperands.push_back(OpInfo.CallOperand);
6420 // Otherwise, this is a register or register class output.
6422 // Copy the output from the appropriate register. Find a register that
6424 if (OpInfo.AssignedRegs.Regs.empty()) {
6425 LLVMContext &Ctx = *DAG.getContext();
6426 Ctx.emitError(CS.getInstruction(),
6427 "couldn't allocate output register for constraint '" +
6428 Twine(OpInfo.ConstraintCode) + "'");
6432 // If this is an indirect operand, store through the pointer after the
6434 if (OpInfo.isIndirect) {
6435 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
6436 OpInfo.CallOperandVal));
6438 // This is the result value of the call.
6439 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
6440 // Concatenate this output onto the outputs list.
6441 RetValRegs.append(OpInfo.AssignedRegs);
6444 // Add information to the INLINEASM node to know that this register is
6447 .AddInlineAsmOperands(OpInfo.isEarlyClobber
6448 ? InlineAsm::Kind_RegDefEarlyClobber
6449 : InlineAsm::Kind_RegDef,
6450 false, 0, DAG, AsmNodeOperands);
6453 case InlineAsm::isInput: {
6454 SDValue InOperandVal = OpInfo.CallOperand;
6456 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint?
6457 // If this is required to match an output register we have already set,
6458 // just use its register.
6459 unsigned OperandNo = OpInfo.getMatchedOperand();
6461 // Scan until we find the definition we already emitted of this operand.
6462 // When we find it, create a RegsForValue operand.
6463 unsigned CurOp = InlineAsm::Op_FirstOperand;
6464 for (; OperandNo; --OperandNo) {
6465 // Advance to the next operand.
6467 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6468 assert((InlineAsm::isRegDefKind(OpFlag) ||
6469 InlineAsm::isRegDefEarlyClobberKind(OpFlag) ||
6470 InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?");
6471 CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1;
6475 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6476 if (InlineAsm::isRegDefKind(OpFlag) ||
6477 InlineAsm::isRegDefEarlyClobberKind(OpFlag)) {
6478 // Add (OpFlag&0xffff)>>3 registers to MatchedRegs.
6479 if (OpInfo.isIndirect) {
6480 // This happens on gcc/testsuite/gcc.dg/pr8788-1.c
6481 LLVMContext &Ctx = *DAG.getContext();
6482 Ctx.emitError(CS.getInstruction(), "inline asm not supported yet:"
6483 " don't know how to handle tied "
6484 "indirect register inputs");
6488 RegsForValue MatchedRegs;
6489 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
6490 MVT RegVT = AsmNodeOperands[CurOp+1].getSimpleValueType();
6491 MatchedRegs.RegVTs.push_back(RegVT);
6492 MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo();
6493 for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag);
6495 if (const TargetRegisterClass *RC = TLI->getRegClassFor(RegVT))
6496 MatchedRegs.Regs.push_back(RegInfo.createVirtualRegister(RC));
6498 LLVMContext &Ctx = *DAG.getContext();
6499 Ctx.emitError(CS.getInstruction(),
6500 "inline asm error: This value"
6501 " type register class is not natively supported!");
6505 // Use the produced MatchedRegs object to
6506 MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurSDLoc(),
6507 Chain, &Flag, CS.getInstruction());
6508 MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse,
6509 true, OpInfo.getMatchedOperand(),
6510 DAG, AsmNodeOperands);
6514 assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!");
6515 assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 &&
6516 "Unexpected number of operands");
6517 // Add information to the INLINEASM node to know about this input.
6518 // See InlineAsm.h isUseOperandTiedToDef.
6519 OpFlag = InlineAsm::getFlagWordForMatchingOp(OpFlag,
6520 OpInfo.getMatchedOperand());
6521 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlag,
6522 TLI->getPointerTy()));
6523 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]);
6527 // Treat indirect 'X' constraint as memory.
6528 if (OpInfo.ConstraintType == TargetLowering::C_Other &&
6530 OpInfo.ConstraintType = TargetLowering::C_Memory;
6532 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
6533 std::vector<SDValue> Ops;
6534 TLI->LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode,
6537 LLVMContext &Ctx = *DAG.getContext();
6538 Ctx.emitError(CS.getInstruction(),
6539 "invalid operand for inline asm constraint '" +
6540 Twine(OpInfo.ConstraintCode) + "'");
6544 // Add information to the INLINEASM node to know about this input.
6545 unsigned ResOpType =
6546 InlineAsm::getFlagWord(InlineAsm::Kind_Imm, Ops.size());
6547 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
6548 TLI->getPointerTy()));
6549 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
6553 if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
6554 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
6555 assert(InOperandVal.getValueType() == TLI->getPointerTy() &&
6556 "Memory operands expect pointer values");
6558 // Add information to the INLINEASM node to know about this input.
6559 unsigned ResOpType = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
6560 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
6561 TLI->getPointerTy()));
6562 AsmNodeOperands.push_back(InOperandVal);
6566 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
6567 OpInfo.ConstraintType == TargetLowering::C_Register) &&
6568 "Unknown constraint type!");
6570 // TODO: Support this.
6571 if (OpInfo.isIndirect) {
6572 LLVMContext &Ctx = *DAG.getContext();
6573 Ctx.emitError(CS.getInstruction(),
6574 "Don't know how to handle indirect register inputs yet "
6575 "for constraint '" +
6576 Twine(OpInfo.ConstraintCode) + "'");
6580 // Copy the input into the appropriate registers.
6581 if (OpInfo.AssignedRegs.Regs.empty()) {
6582 LLVMContext &Ctx = *DAG.getContext();
6583 Ctx.emitError(CS.getInstruction(),
6584 "couldn't allocate input reg for constraint '" +
6585 Twine(OpInfo.ConstraintCode) + "'");
6589 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, getCurSDLoc(),
6590 Chain, &Flag, CS.getInstruction());
6592 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, false, 0,
6593 DAG, AsmNodeOperands);
6596 case InlineAsm::isClobber: {
6597 // Add the clobbered value to the operand list, so that the register
6598 // allocator is aware that the physreg got clobbered.
6599 if (!OpInfo.AssignedRegs.Regs.empty())
6600 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_Clobber,
6608 // Finish up input operands. Set the input chain and add the flag last.
6609 AsmNodeOperands[InlineAsm::Op_InputChain] = Chain;
6610 if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
6612 Chain = DAG.getNode(ISD::INLINEASM, getCurSDLoc(),
6613 DAG.getVTList(MVT::Other, MVT::Glue),
6614 &AsmNodeOperands[0], AsmNodeOperands.size());
6615 Flag = Chain.getValue(1);
6617 // If this asm returns a register value, copy the result from that register
6618 // and set it as the value of the call.
6619 if (!RetValRegs.Regs.empty()) {
6620 SDValue Val = RetValRegs.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(),
6621 Chain, &Flag, CS.getInstruction());
6623 // FIXME: Why don't we do this for inline asms with MRVs?
6624 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) {
6625 EVT ResultType = TLI->getValueType(CS.getType());
6627 // If any of the results of the inline asm is a vector, it may have the
6628 // wrong width/num elts. This can happen for register classes that can
6629 // contain multiple different value types. The preg or vreg allocated may
6630 // not have the same VT as was expected. Convert it to the right type
6631 // with bit_convert.
6632 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) {
6633 Val = DAG.getNode(ISD::BITCAST, getCurSDLoc(),
6636 } else if (ResultType != Val.getValueType() &&
6637 ResultType.isInteger() && Val.getValueType().isInteger()) {
6638 // If a result value was tied to an input value, the computed result may
6639 // have a wider width than the expected result. Extract the relevant
6641 Val = DAG.getNode(ISD::TRUNCATE, getCurSDLoc(), ResultType, Val);
6644 assert(ResultType == Val.getValueType() && "Asm result value mismatch!");
6647 setValue(CS.getInstruction(), Val);
6648 // Don't need to use this as a chain in this case.
6649 if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty())
6653 std::vector<std::pair<SDValue, const Value *> > StoresToEmit;
6655 // Process indirect outputs, first output all of the flagged copies out of
6657 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
6658 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
6659 const Value *Ptr = IndirectStoresToEmit[i].second;
6660 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, FuncInfo, getCurSDLoc(),
6662 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
6665 // Emit the non-flagged stores from the physregs.
6666 SmallVector<SDValue, 8> OutChains;
6667 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) {
6668 SDValue Val = DAG.getStore(Chain, getCurSDLoc(),
6669 StoresToEmit[i].first,
6670 getValue(StoresToEmit[i].second),
6671 MachinePointerInfo(StoresToEmit[i].second),
6673 OutChains.push_back(Val);
6676 if (!OutChains.empty())
6677 Chain = DAG.getNode(ISD::TokenFactor, getCurSDLoc(), MVT::Other,
6678 &OutChains[0], OutChains.size());
6683 void SelectionDAGBuilder::visitVAStart(const CallInst &I) {
6684 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurSDLoc(),
6685 MVT::Other, getRoot(),
6686 getValue(I.getArgOperand(0)),
6687 DAG.getSrcValue(I.getArgOperand(0))));
6690 void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) {
6691 const TargetLowering *TLI = TM.getTargetLowering();
6692 const DataLayout &TD = *TLI->getDataLayout();
6693 SDValue V = DAG.getVAArg(TLI->getValueType(I.getType()), getCurSDLoc(),
6694 getRoot(), getValue(I.getOperand(0)),
6695 DAG.getSrcValue(I.getOperand(0)),
6696 TD.getABITypeAlignment(I.getType()));
6698 DAG.setRoot(V.getValue(1));
6701 void SelectionDAGBuilder::visitVAEnd(const CallInst &I) {
6702 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurSDLoc(),
6703 MVT::Other, getRoot(),
6704 getValue(I.getArgOperand(0)),
6705 DAG.getSrcValue(I.getArgOperand(0))));
6708 void SelectionDAGBuilder::visitVACopy(const CallInst &I) {
6709 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurSDLoc(),
6710 MVT::Other, getRoot(),
6711 getValue(I.getArgOperand(0)),
6712 getValue(I.getArgOperand(1)),
6713 DAG.getSrcValue(I.getArgOperand(0)),
6714 DAG.getSrcValue(I.getArgOperand(1))));
6717 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
6718 /// implementation, which just calls LowerCall.
6719 /// FIXME: When all targets are
6720 /// migrated to using LowerCall, this hook should be integrated into SDISel.
6721 std::pair<SDValue, SDValue>
6722 TargetLowering::LowerCallTo(TargetLowering::CallLoweringInfo &CLI) const {
6723 // Handle the incoming return values from the call.
6725 SmallVector<EVT, 4> RetTys;
6726 ComputeValueVTs(*this, CLI.RetTy, RetTys);
6727 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
6729 MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
6730 unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
6731 for (unsigned i = 0; i != NumRegs; ++i) {
6732 ISD::InputArg MyFlags;
6733 MyFlags.VT = RegisterVT;
6734 MyFlags.Used = CLI.IsReturnValueUsed;
6736 MyFlags.Flags.setSExt();
6738 MyFlags.Flags.setZExt();
6740 MyFlags.Flags.setInReg();
6741 CLI.Ins.push_back(MyFlags);
6745 // Handle all of the outgoing arguments.
6747 CLI.OutVals.clear();
6748 ArgListTy &Args = CLI.Args;
6749 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
6750 SmallVector<EVT, 4> ValueVTs;
6751 ComputeValueVTs(*this, Args[i].Ty, ValueVTs);
6752 for (unsigned Value = 0, NumValues = ValueVTs.size();
6753 Value != NumValues; ++Value) {
6754 EVT VT = ValueVTs[Value];
6755 Type *ArgTy = VT.getTypeForEVT(CLI.RetTy->getContext());
6756 SDValue Op = SDValue(Args[i].Node.getNode(),
6757 Args[i].Node.getResNo() + Value);
6758 ISD::ArgFlagsTy Flags;
6759 unsigned OriginalAlignment =
6760 getDataLayout()->getABITypeAlignment(ArgTy);
6766 if (Args[i].isInReg)
6770 if (Args[i].isByVal) {
6772 PointerType *Ty = cast<PointerType>(Args[i].Ty);
6773 Type *ElementTy = Ty->getElementType();
6774 Flags.setByValSize(getDataLayout()->getTypeAllocSize(ElementTy));
6775 // For ByVal, alignment should come from FE. BE will guess if this
6776 // info is not there but there are cases it cannot get right.
6777 unsigned FrameAlign;
6778 if (Args[i].Alignment)
6779 FrameAlign = Args[i].Alignment;
6781 FrameAlign = getByValTypeAlignment(ElementTy);
6782 Flags.setByValAlign(FrameAlign);
6786 Flags.setOrigAlign(OriginalAlignment);
6788 MVT PartVT = getRegisterType(CLI.RetTy->getContext(), VT);
6789 unsigned NumParts = getNumRegisters(CLI.RetTy->getContext(), VT);
6790 SmallVector<SDValue, 4> Parts(NumParts);
6791 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
6794 ExtendKind = ISD::SIGN_EXTEND;
6795 else if (Args[i].isZExt)
6796 ExtendKind = ISD::ZERO_EXTEND;
6798 // Conservatively only handle 'returned' on non-vectors for now
6799 if (Args[i].isReturned && !Op.getValueType().isVector()) {
6800 assert(CLI.RetTy == Args[i].Ty && RetTys.size() == NumValues &&
6801 "unexpected use of 'returned'");
6802 // Before passing 'returned' to the target lowering code, ensure that
6803 // either the register MVT and the actual EVT are the same size or that
6804 // the return value and argument are extended in the same way; in these
6805 // cases it's safe to pass the argument register value unchanged as the
6806 // return register value (although it's at the target's option whether
6808 // TODO: allow code generation to take advantage of partially preserved
6809 // registers rather than clobbering the entire register when the
6810 // parameter extension method is not compatible with the return
6812 if ((NumParts * PartVT.getSizeInBits() == VT.getSizeInBits()) ||
6813 (ExtendKind != ISD::ANY_EXTEND &&
6814 CLI.RetSExt == Args[i].isSExt && CLI.RetZExt == Args[i].isZExt))
6815 Flags.setReturned();
6818 getCopyToParts(CLI.DAG, CLI.DL, Op, &Parts[0], NumParts,
6819 PartVT, CLI.CS ? CLI.CS->getInstruction() : 0, ExtendKind);
6821 for (unsigned j = 0; j != NumParts; ++j) {
6822 // if it isn't first piece, alignment must be 1
6823 ISD::OutputArg MyFlags(Flags, Parts[j].getValueType(),
6824 i < CLI.NumFixedArgs,
6825 i, j*Parts[j].getValueType().getStoreSize());
6826 if (NumParts > 1 && j == 0)
6827 MyFlags.Flags.setSplit();
6829 MyFlags.Flags.setOrigAlign(1);
6831 CLI.Outs.push_back(MyFlags);
6832 CLI.OutVals.push_back(Parts[j]);
6837 SmallVector<SDValue, 4> InVals;
6838 CLI.Chain = LowerCall(CLI, InVals);
6840 // Verify that the target's LowerCall behaved as expected.
6841 assert(CLI.Chain.getNode() && CLI.Chain.getValueType() == MVT::Other &&
6842 "LowerCall didn't return a valid chain!");
6843 assert((!CLI.IsTailCall || InVals.empty()) &&
6844 "LowerCall emitted a return value for a tail call!");
6845 assert((CLI.IsTailCall || InVals.size() == CLI.Ins.size()) &&
6846 "LowerCall didn't emit the correct number of values!");
6848 // For a tail call, the return value is merely live-out and there aren't
6849 // any nodes in the DAG representing it. Return a special value to
6850 // indicate that a tail call has been emitted and no more Instructions
6851 // should be processed in the current block.
6852 if (CLI.IsTailCall) {
6853 CLI.DAG.setRoot(CLI.Chain);
6854 return std::make_pair(SDValue(), SDValue());
6857 DEBUG(for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++i) {
6858 assert(InVals[i].getNode() &&
6859 "LowerCall emitted a null value!");
6860 assert(EVT(CLI.Ins[i].VT) == InVals[i].getValueType() &&
6861 "LowerCall emitted a value with the wrong type!");
6864 // Collect the legal value parts into potentially illegal values
6865 // that correspond to the original function's return values.
6866 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6868 AssertOp = ISD::AssertSext;
6869 else if (CLI.RetZExt)
6870 AssertOp = ISD::AssertZext;
6871 SmallVector<SDValue, 4> ReturnValues;
6872 unsigned CurReg = 0;
6873 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
6875 MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
6876 unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
6878 ReturnValues.push_back(getCopyFromParts(CLI.DAG, CLI.DL, &InVals[CurReg],
6879 NumRegs, RegisterVT, VT, NULL,
6884 // For a function returning void, there is no return value. We can't create
6885 // such a node, so we just return a null return value in that case. In
6886 // that case, nothing will actually look at the value.
6887 if (ReturnValues.empty())
6888 return std::make_pair(SDValue(), CLI.Chain);
6890 SDValue Res = CLI.DAG.getNode(ISD::MERGE_VALUES, CLI.DL,
6891 CLI.DAG.getVTList(&RetTys[0], RetTys.size()),
6892 &ReturnValues[0], ReturnValues.size());
6893 return std::make_pair(Res, CLI.Chain);
6896 void TargetLowering::LowerOperationWrapper(SDNode *N,
6897 SmallVectorImpl<SDValue> &Results,
6898 SelectionDAG &DAG) const {
6899 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
6901 Results.push_back(Res);
6904 SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6905 llvm_unreachable("LowerOperation not implemented for this target!");
6909 SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) {
6910 SDValue Op = getNonRegisterValue(V);
6911 assert((Op.getOpcode() != ISD::CopyFromReg ||
6912 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
6913 "Copy from a reg to the same reg!");
6914 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg");
6916 const TargetLowering *TLI = TM.getTargetLowering();
6917 RegsForValue RFV(V->getContext(), *TLI, Reg, V->getType());
6918 SDValue Chain = DAG.getEntryNode();
6919 RFV.getCopyToRegs(Op, DAG, getCurSDLoc(), Chain, 0, V);
6920 PendingExports.push_back(Chain);
6923 #include "llvm/CodeGen/SelectionDAGISel.h"
6925 /// isOnlyUsedInEntryBlock - If the specified argument is only used in the
6926 /// entry block, return true. This includes arguments used by switches, since
6927 /// the switch may expand into multiple basic blocks.
6928 static bool isOnlyUsedInEntryBlock(const Argument *A, bool FastISel) {
6929 // With FastISel active, we may be splitting blocks, so force creation
6930 // of virtual registers for all non-dead arguments.
6932 return A->use_empty();
6934 const BasicBlock *Entry = A->getParent()->begin();
6935 for (Value::const_use_iterator UI = A->use_begin(), E = A->use_end();
6937 const User *U = *UI;
6938 if (cast<Instruction>(U)->getParent() != Entry || isa<SwitchInst>(U))
6939 return false; // Use not in entry block.
6944 void SelectionDAGISel::LowerArguments(const Function &F) {
6945 SelectionDAG &DAG = SDB->DAG;
6946 SDLoc dl = SDB->getCurSDLoc();
6947 const TargetLowering *TLI = getTargetLowering();
6948 const DataLayout *TD = TLI->getDataLayout();
6949 SmallVector<ISD::InputArg, 16> Ins;
6951 if (!FuncInfo->CanLowerReturn) {
6952 // Put in an sret pointer parameter before all the other parameters.
6953 SmallVector<EVT, 1> ValueVTs;
6954 ComputeValueVTs(*getTargetLowering(),
6955 PointerType::getUnqual(F.getReturnType()), ValueVTs);
6957 // NOTE: Assuming that a pointer will never break down to more than one VT
6959 ISD::ArgFlagsTy Flags;
6961 MVT RegisterVT = TLI->getRegisterType(*DAG.getContext(), ValueVTs[0]);
6962 ISD::InputArg RetArg(Flags, RegisterVT, true, 0, 0);
6963 Ins.push_back(RetArg);
6966 // Set up the incoming argument description vector.
6968 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end();
6969 I != E; ++I, ++Idx) {
6970 SmallVector<EVT, 4> ValueVTs;
6971 ComputeValueVTs(*TLI, I->getType(), ValueVTs);
6972 bool isArgValueUsed = !I->use_empty();
6973 for (unsigned Value = 0, NumValues = ValueVTs.size();
6974 Value != NumValues; ++Value) {
6975 EVT VT = ValueVTs[Value];
6976 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
6977 ISD::ArgFlagsTy Flags;
6978 unsigned OriginalAlignment =
6979 TD->getABITypeAlignment(ArgTy);
6981 if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt))
6983 if (F.getAttributes().hasAttribute(Idx, Attribute::SExt))
6985 if (F.getAttributes().hasAttribute(Idx, Attribute::InReg))
6987 if (F.getAttributes().hasAttribute(Idx, Attribute::StructRet))
6989 if (F.getAttributes().hasAttribute(Idx, Attribute::ByVal)) {
6991 PointerType *Ty = cast<PointerType>(I->getType());
6992 Type *ElementTy = Ty->getElementType();
6993 Flags.setByValSize(TD->getTypeAllocSize(ElementTy));
6994 // For ByVal, alignment should be passed from FE. BE will guess if
6995 // this info is not there but there are cases it cannot get right.
6996 unsigned FrameAlign;
6997 if (F.getParamAlignment(Idx))
6998 FrameAlign = F.getParamAlignment(Idx);
7000 FrameAlign = TLI->getByValTypeAlignment(ElementTy);
7001 Flags.setByValAlign(FrameAlign);
7003 if (F.getAttributes().hasAttribute(Idx, Attribute::Nest))
7005 Flags.setOrigAlign(OriginalAlignment);
7007 MVT RegisterVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7008 unsigned NumRegs = TLI->getNumRegisters(*CurDAG->getContext(), VT);
7009 for (unsigned i = 0; i != NumRegs; ++i) {
7010 ISD::InputArg MyFlags(Flags, RegisterVT, isArgValueUsed,
7011 Idx-1, i*RegisterVT.getStoreSize());
7012 if (NumRegs > 1 && i == 0)
7013 MyFlags.Flags.setSplit();
7014 // if it isn't first piece, alignment must be 1
7016 MyFlags.Flags.setOrigAlign(1);
7017 Ins.push_back(MyFlags);
7022 // Call the target to set up the argument values.
7023 SmallVector<SDValue, 8> InVals;
7024 SDValue NewRoot = TLI->LowerFormalArguments(DAG.getRoot(), F.getCallingConv(),
7028 // Verify that the target's LowerFormalArguments behaved as expected.
7029 assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other &&
7030 "LowerFormalArguments didn't return a valid chain!");
7031 assert(InVals.size() == Ins.size() &&
7032 "LowerFormalArguments didn't emit the correct number of values!");
7034 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
7035 assert(InVals[i].getNode() &&
7036 "LowerFormalArguments emitted a null value!");
7037 assert(EVT(Ins[i].VT) == InVals[i].getValueType() &&
7038 "LowerFormalArguments emitted a value with the wrong type!");
7042 // Update the DAG with the new chain value resulting from argument lowering.
7043 DAG.setRoot(NewRoot);
7045 // Set up the argument values.
7048 if (!FuncInfo->CanLowerReturn) {
7049 // Create a virtual register for the sret pointer, and put in a copy
7050 // from the sret argument into it.
7051 SmallVector<EVT, 1> ValueVTs;
7052 ComputeValueVTs(*TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
7053 MVT VT = ValueVTs[0].getSimpleVT();
7054 MVT RegVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7055 ISD::NodeType AssertOp = ISD::DELETED_NODE;
7056 SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1,
7057 RegVT, VT, NULL, AssertOp);
7059 MachineFunction& MF = SDB->DAG.getMachineFunction();
7060 MachineRegisterInfo& RegInfo = MF.getRegInfo();
7061 unsigned SRetReg = RegInfo.createVirtualRegister(TLI->getRegClassFor(RegVT));
7062 FuncInfo->DemoteRegister = SRetReg;
7063 NewRoot = SDB->DAG.getCopyToReg(NewRoot, SDB->getCurSDLoc(),
7065 DAG.setRoot(NewRoot);
7067 // i indexes lowered arguments. Bump it past the hidden sret argument.
7068 // Idx indexes LLVM arguments. Don't touch it.
7072 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
7074 SmallVector<SDValue, 4> ArgValues;
7075 SmallVector<EVT, 4> ValueVTs;
7076 ComputeValueVTs(*TLI, I->getType(), ValueVTs);
7077 unsigned NumValues = ValueVTs.size();
7079 // If this argument is unused then remember its value. It is used to generate
7080 // debugging information.
7081 if (I->use_empty() && NumValues) {
7082 SDB->setUnusedArgValue(I, InVals[i]);
7084 // Also remember any frame index for use in FastISel.
7085 if (FrameIndexSDNode *FI =
7086 dyn_cast<FrameIndexSDNode>(InVals[i].getNode()))
7087 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
7090 for (unsigned Val = 0; Val != NumValues; ++Val) {
7091 EVT VT = ValueVTs[Val];
7092 MVT PartVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
7093 unsigned NumParts = TLI->getNumRegisters(*CurDAG->getContext(), VT);
7095 if (!I->use_empty()) {
7096 ISD::NodeType AssertOp = ISD::DELETED_NODE;
7097 if (F.getAttributes().hasAttribute(Idx, Attribute::SExt))
7098 AssertOp = ISD::AssertSext;
7099 else if (F.getAttributes().hasAttribute(Idx, Attribute::ZExt))
7100 AssertOp = ISD::AssertZext;
7102 ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i],
7103 NumParts, PartVT, VT,
7110 // We don't need to do anything else for unused arguments.
7111 if (ArgValues.empty())
7114 // Note down frame index.
7115 if (FrameIndexSDNode *FI =
7116 dyn_cast<FrameIndexSDNode>(ArgValues[0].getNode()))
7117 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
7119 SDValue Res = DAG.getMergeValues(&ArgValues[0], NumValues,
7120 SDB->getCurSDLoc());
7122 SDB->setValue(I, Res);
7123 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::BUILD_PAIR) {
7124 if (LoadSDNode *LNode =
7125 dyn_cast<LoadSDNode>(Res.getOperand(0).getNode()))
7126 if (FrameIndexSDNode *FI =
7127 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode()))
7128 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
7131 // If this argument is live outside of the entry block, insert a copy from
7132 // wherever we got it to the vreg that other BB's will reference it as.
7133 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::CopyFromReg) {
7134 // If we can, though, try to skip creating an unnecessary vreg.
7135 // FIXME: This isn't very clean... it would be nice to make this more
7136 // general. It's also subtly incompatible with the hacks FastISel
7138 unsigned Reg = cast<RegisterSDNode>(Res.getOperand(1))->getReg();
7139 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
7140 FuncInfo->ValueMap[I] = Reg;
7144 if (!isOnlyUsedInEntryBlock(I, TM.Options.EnableFastISel)) {
7145 FuncInfo->InitializeRegForValue(I);
7146 SDB->CopyToExportRegsIfNeeded(I);
7150 assert(i == InVals.size() && "Argument register count mismatch!");
7152 // Finally, if the target has anything special to do, allow it to do so.
7153 // FIXME: this should insert code into the DAG!
7154 EmitFunctionEntryCode();
7157 /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
7158 /// ensure constants are generated when needed. Remember the virtual registers
7159 /// that need to be added to the Machine PHI nodes as input. We cannot just
7160 /// directly add them, because expansion might result in multiple MBB's for one
7161 /// BB. As such, the start of the BB might correspond to a different MBB than
7165 SelectionDAGBuilder::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
7166 const TerminatorInst *TI = LLVMBB->getTerminator();
7168 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
7170 // Check successor nodes' PHI nodes that expect a constant to be available
7172 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
7173 const BasicBlock *SuccBB = TI->getSuccessor(succ);
7174 if (!isa<PHINode>(SuccBB->begin())) continue;
7175 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
7177 // If this terminator has multiple identical successors (common for
7178 // switches), only handle each succ once.
7179 if (!SuccsHandled.insert(SuccMBB)) continue;
7181 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
7183 // At this point we know that there is a 1-1 correspondence between LLVM PHI
7184 // nodes and Machine PHI nodes, but the incoming operands have not been
7186 for (BasicBlock::const_iterator I = SuccBB->begin();
7187 const PHINode *PN = dyn_cast<PHINode>(I); ++I) {
7188 // Ignore dead phi's.
7189 if (PN->use_empty()) continue;
7192 if (PN->getType()->isEmptyTy())
7196 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
7198 if (const Constant *C = dyn_cast<Constant>(PHIOp)) {
7199 unsigned &RegOut = ConstantsOut[C];
7201 RegOut = FuncInfo.CreateRegs(C->getType());
7202 CopyValueToVirtualRegister(C, RegOut);
7206 DenseMap<const Value *, unsigned>::iterator I =
7207 FuncInfo.ValueMap.find(PHIOp);
7208 if (I != FuncInfo.ValueMap.end())
7211 assert(isa<AllocaInst>(PHIOp) &&
7212 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
7213 "Didn't codegen value into a register!??");
7214 Reg = FuncInfo.CreateRegs(PHIOp->getType());
7215 CopyValueToVirtualRegister(PHIOp, Reg);
7219 // Remember that this register needs to added to the machine PHI node as
7220 // the input for this MBB.
7221 SmallVector<EVT, 4> ValueVTs;
7222 const TargetLowering *TLI = TM.getTargetLowering();
7223 ComputeValueVTs(*TLI, PN->getType(), ValueVTs);
7224 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
7225 EVT VT = ValueVTs[vti];
7226 unsigned NumRegisters = TLI->getNumRegisters(*DAG.getContext(), VT);
7227 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
7228 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
7229 Reg += NumRegisters;
7234 ConstantsOut.clear();
7237 /// Add a successor MBB to ParentMBB< creating a new MachineBB for BB if SuccMBB
7240 SelectionDAGBuilder::StackProtectorDescriptor::
7241 AddSuccessorMBB(const BasicBlock *BB,
7242 MachineBasicBlock *ParentMBB,
7243 MachineBasicBlock *SuccMBB) {
7244 // If SuccBB has not been created yet, create it.
7246 MachineFunction *MF = ParentMBB->getParent();
7247 MachineFunction::iterator BBI = ParentMBB;
7248 SuccMBB = MF->CreateMachineBasicBlock(BB);
7249 MF->insert(++BBI, SuccMBB);
7251 // Add it as a successor of ParentMBB.
7252 ParentMBB->addSuccessor(SuccMBB);