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 "SDNodeDbgValue.h"
16 #include "SelectionDAGBuilder.h"
17 #include "llvm/ADT/BitVector.h"
18 #include "llvm/ADT/PostOrderIterator.h"
19 #include "llvm/ADT/SmallSet.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/Analysis/ConstantFolding.h"
22 #include "llvm/Constants.h"
23 #include "llvm/CallingConv.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/Function.h"
26 #include "llvm/GlobalVariable.h"
27 #include "llvm/InlineAsm.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/Intrinsics.h"
30 #include "llvm/IntrinsicInst.h"
31 #include "llvm/LLVMContext.h"
32 #include "llvm/Module.h"
33 #include "llvm/CodeGen/Analysis.h"
34 #include "llvm/CodeGen/FastISel.h"
35 #include "llvm/CodeGen/FunctionLoweringInfo.h"
36 #include "llvm/CodeGen/GCStrategy.h"
37 #include "llvm/CodeGen/GCMetadata.h"
38 #include "llvm/CodeGen/MachineFunction.h"
39 #include "llvm/CodeGen/MachineFrameInfo.h"
40 #include "llvm/CodeGen/MachineInstrBuilder.h"
41 #include "llvm/CodeGen/MachineJumpTableInfo.h"
42 #include "llvm/CodeGen/MachineModuleInfo.h"
43 #include "llvm/CodeGen/MachineRegisterInfo.h"
44 #include "llvm/CodeGen/PseudoSourceValue.h"
45 #include "llvm/CodeGen/SelectionDAG.h"
46 #include "llvm/Analysis/DebugInfo.h"
47 #include "llvm/Target/TargetRegisterInfo.h"
48 #include "llvm/Target/TargetData.h"
49 #include "llvm/Target/TargetFrameInfo.h"
50 #include "llvm/Target/TargetInstrInfo.h"
51 #include "llvm/Target/TargetIntrinsicInfo.h"
52 #include "llvm/Target/TargetLowering.h"
53 #include "llvm/Target/TargetOptions.h"
54 #include "llvm/Support/Compiler.h"
55 #include "llvm/Support/CommandLine.h"
56 #include "llvm/Support/Debug.h"
57 #include "llvm/Support/ErrorHandling.h"
58 #include "llvm/Support/MathExtras.h"
59 #include "llvm/Support/raw_ostream.h"
63 /// LimitFloatPrecision - Generate low-precision inline sequences for
64 /// some float libcalls (6, 8 or 12 bits).
65 static unsigned LimitFloatPrecision;
67 static cl::opt<unsigned, true>
68 LimitFPPrecision("limit-float-precision",
69 cl::desc("Generate low-precision inline sequences "
70 "for some float libcalls"),
71 cl::location(LimitFloatPrecision),
74 // Limit the width of DAG chains. This is important in general to prevent
75 // prevent DAG-based analysis from blowing up. For example, alias analysis and
76 // load clustering may not complete in reasonable time. It is difficult to
77 // recognize and avoid this situation within each individual analysis, and
78 // future analyses are likely to have the same behavior. Limiting DAG width is
79 // the safe approach, and will be especially important with global DAGs.
81 // MaxParallelChains default is arbitrarily high to avoid affecting
82 // optimization, but could be lowered to improve compile time. Any ld-ld-st-st
83 // sequence over this should have been converted to llvm.memcpy by the
84 // frontend. It easy to induce this behavior with .ll code such as:
85 // %buffer = alloca [4096 x i8]
86 // %data = load [4096 x i8]* %argPtr
87 // store [4096 x i8] %data, [4096 x i8]* %buffer
88 static cl::opt<unsigned>
89 MaxParallelChains("dag-chain-limit", cl::desc("Max parallel isel dag chains"),
90 cl::init(64), cl::Hidden);
92 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, DebugLoc DL,
93 const SDValue *Parts, unsigned NumParts,
94 EVT PartVT, EVT ValueVT);
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, DebugLoc DL,
102 const SDValue *Parts,
103 unsigned NumParts, EVT PartVT, EVT ValueVT,
104 ISD::NodeType AssertOp = ISD::DELETED_NODE) {
105 if (ValueVT.isVector())
106 return getCopyFromPartsVector(DAG, DL, Parts, NumParts, PartVT, ValueVT);
108 assert(NumParts > 0 && "No parts to assemble!");
109 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
110 SDValue Val = Parts[0];
113 // Assemble the value from multiple parts.
114 if (ValueVT.isInteger()) {
115 unsigned PartBits = PartVT.getSizeInBits();
116 unsigned ValueBits = ValueVT.getSizeInBits();
118 // Assemble the power of 2 part.
119 unsigned RoundParts = NumParts & (NumParts - 1) ?
120 1 << Log2_32(NumParts) : NumParts;
121 unsigned RoundBits = PartBits * RoundParts;
122 EVT RoundVT = RoundBits == ValueBits ?
123 ValueVT : EVT::getIntegerVT(*DAG.getContext(), RoundBits);
126 EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2);
128 if (RoundParts > 2) {
129 Lo = getCopyFromParts(DAG, DL, Parts, RoundParts / 2,
131 Hi = getCopyFromParts(DAG, DL, Parts + RoundParts / 2,
132 RoundParts / 2, PartVT, HalfVT);
134 Lo = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[0]);
135 Hi = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[1]);
138 if (TLI.isBigEndian())
141 Val = DAG.getNode(ISD::BUILD_PAIR, DL, RoundVT, Lo, Hi);
143 if (RoundParts < NumParts) {
144 // Assemble the trailing non-power-of-2 part.
145 unsigned OddParts = NumParts - RoundParts;
146 EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits);
147 Hi = getCopyFromParts(DAG, DL,
148 Parts + RoundParts, OddParts, PartVT, OddVT);
150 // Combine the round and odd parts.
152 if (TLI.isBigEndian())
154 EVT TotalVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
155 Hi = DAG.getNode(ISD::ANY_EXTEND, DL, TotalVT, Hi);
156 Hi = DAG.getNode(ISD::SHL, DL, TotalVT, Hi,
157 DAG.getConstant(Lo.getValueType().getSizeInBits(),
158 TLI.getPointerTy()));
159 Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, TotalVT, Lo);
160 Val = DAG.getNode(ISD::OR, DL, TotalVT, Lo, Hi);
162 } else if (PartVT.isFloatingPoint()) {
163 // FP split into multiple FP parts (for ppcf128)
164 assert(ValueVT == EVT(MVT::ppcf128) && PartVT == EVT(MVT::f64) &&
167 Lo = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[0]);
168 Hi = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[1]);
169 if (TLI.isBigEndian())
171 Val = DAG.getNode(ISD::BUILD_PAIR, DL, ValueVT, Lo, Hi);
173 // FP split into integer parts (soft fp)
174 assert(ValueVT.isFloatingPoint() && PartVT.isInteger() &&
175 !PartVT.isVector() && "Unexpected split");
176 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits());
177 Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT);
181 // There is now one part, held in Val. Correct it to match ValueVT.
182 PartVT = Val.getValueType();
184 if (PartVT == ValueVT)
187 if (PartVT.isInteger() && ValueVT.isInteger()) {
188 if (ValueVT.bitsLT(PartVT)) {
189 // For a truncate, see if we have any information to
190 // indicate whether the truncated bits will always be
191 // zero or sign-extension.
192 if (AssertOp != ISD::DELETED_NODE)
193 Val = DAG.getNode(AssertOp, DL, PartVT, Val,
194 DAG.getValueType(ValueVT));
195 return DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
197 return DAG.getNode(ISD::ANY_EXTEND, DL, ValueVT, Val);
200 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
201 // FP_ROUND's are always exact here.
202 if (ValueVT.bitsLT(Val.getValueType()))
203 return DAG.getNode(ISD::FP_ROUND, DL, ValueVT, Val,
204 DAG.getIntPtrConstant(1));
206 return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val);
209 if (PartVT.getSizeInBits() == ValueVT.getSizeInBits())
210 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
212 llvm_unreachable("Unknown mismatch!");
216 /// getCopyFromParts - Create a value that contains the specified legal parts
217 /// combined into the value they represent. If the parts combine to a type
218 /// larger then ValueVT then AssertOp can be used to specify whether the extra
219 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
220 /// (ISD::AssertSext).
221 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, DebugLoc DL,
222 const SDValue *Parts, unsigned NumParts,
223 EVT PartVT, EVT ValueVT) {
224 assert(ValueVT.isVector() && "Not a vector value");
225 assert(NumParts > 0 && "No parts to assemble!");
226 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
227 SDValue Val = Parts[0];
229 // Handle a multi-element vector.
231 EVT IntermediateVT, RegisterVT;
232 unsigned NumIntermediates;
234 TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT,
235 NumIntermediates, RegisterVT);
236 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
237 NumParts = NumRegs; // Silence a compiler warning.
238 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
239 assert(RegisterVT == Parts[0].getValueType() &&
240 "Part type doesn't match part!");
242 // Assemble the parts into intermediate operands.
243 SmallVector<SDValue, 8> Ops(NumIntermediates);
244 if (NumIntermediates == NumParts) {
245 // If the register was not expanded, truncate or copy the value,
247 for (unsigned i = 0; i != NumParts; ++i)
248 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i], 1,
249 PartVT, IntermediateVT);
250 } else if (NumParts > 0) {
251 // If the intermediate type was expanded, build the intermediate
252 // operands from the parts.
253 assert(NumParts % NumIntermediates == 0 &&
254 "Must expand into a divisible number of parts!");
255 unsigned Factor = NumParts / NumIntermediates;
256 for (unsigned i = 0; i != NumIntermediates; ++i)
257 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor,
258 PartVT, IntermediateVT);
261 // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the
262 // intermediate operands.
263 Val = DAG.getNode(IntermediateVT.isVector() ?
264 ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR, DL,
265 ValueVT, &Ops[0], NumIntermediates);
268 // There is now one part, held in Val. Correct it to match ValueVT.
269 PartVT = Val.getValueType();
271 if (PartVT == ValueVT)
274 if (PartVT.isVector()) {
275 // If the element type of the source/dest vectors are the same, but the
276 // parts vector has more elements than the value vector, then we have a
277 // vector widening case (e.g. <2 x float> -> <4 x float>). Extract the
279 if (PartVT.getVectorElementType() == ValueVT.getVectorElementType()) {
280 assert(PartVT.getVectorNumElements() > ValueVT.getVectorNumElements() &&
281 "Cannot narrow, it would be a lossy transformation");
282 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val,
283 DAG.getIntPtrConstant(0));
286 // Vector/Vector bitcast.
287 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
290 assert(ValueVT.getVectorElementType() == PartVT &&
291 ValueVT.getVectorNumElements() == 1 &&
292 "Only trivial scalar-to-vector conversions should get here!");
293 return DAG.getNode(ISD::BUILD_VECTOR, DL, ValueVT, Val);
299 static void getCopyToPartsVector(SelectionDAG &DAG, DebugLoc dl,
300 SDValue Val, SDValue *Parts, unsigned NumParts,
303 /// getCopyToParts - Create a series of nodes that contain the specified value
304 /// split into legal parts. If the parts contain more bits than Val, then, for
305 /// integers, ExtendKind can be used to specify how to generate the extra bits.
306 static void getCopyToParts(SelectionDAG &DAG, DebugLoc DL,
307 SDValue Val, SDValue *Parts, unsigned NumParts,
309 ISD::NodeType ExtendKind = ISD::ANY_EXTEND) {
310 EVT ValueVT = Val.getValueType();
312 // Handle the vector case separately.
313 if (ValueVT.isVector())
314 return getCopyToPartsVector(DAG, DL, Val, Parts, NumParts, PartVT);
316 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
317 unsigned PartBits = PartVT.getSizeInBits();
318 unsigned OrigNumParts = NumParts;
319 assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!");
324 assert(!ValueVT.isVector() && "Vector case handled elsewhere");
325 if (PartVT == ValueVT) {
326 assert(NumParts == 1 && "No-op copy with multiple parts!");
331 if (NumParts * PartBits > ValueVT.getSizeInBits()) {
332 // If the parts cover more bits than the value has, promote the value.
333 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
334 assert(NumParts == 1 && "Do not know what to promote to!");
335 Val = DAG.getNode(ISD::FP_EXTEND, DL, PartVT, Val);
337 assert(PartVT.isInteger() && ValueVT.isInteger() &&
338 "Unknown mismatch!");
339 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
340 Val = DAG.getNode(ExtendKind, DL, ValueVT, Val);
342 } else if (PartBits == ValueVT.getSizeInBits()) {
343 // Different types of the same size.
344 assert(NumParts == 1 && PartVT != ValueVT);
345 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
346 } else if (NumParts * PartBits < ValueVT.getSizeInBits()) {
347 // If the parts cover less bits than value has, truncate the value.
348 assert(PartVT.isInteger() && ValueVT.isInteger() &&
349 "Unknown mismatch!");
350 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
351 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
354 // The value may have changed - recompute ValueVT.
355 ValueVT = Val.getValueType();
356 assert(NumParts * PartBits == ValueVT.getSizeInBits() &&
357 "Failed to tile the value with PartVT!");
360 assert(PartVT == ValueVT && "Type conversion failed!");
365 // Expand the value into multiple parts.
366 if (NumParts & (NumParts - 1)) {
367 // The number of parts is not a power of 2. Split off and copy the tail.
368 assert(PartVT.isInteger() && ValueVT.isInteger() &&
369 "Do not know what to expand to!");
370 unsigned RoundParts = 1 << Log2_32(NumParts);
371 unsigned RoundBits = RoundParts * PartBits;
372 unsigned OddParts = NumParts - RoundParts;
373 SDValue OddVal = DAG.getNode(ISD::SRL, DL, ValueVT, Val,
374 DAG.getIntPtrConstant(RoundBits));
375 getCopyToParts(DAG, DL, OddVal, Parts + RoundParts, OddParts, PartVT);
377 if (TLI.isBigEndian())
378 // The odd parts were reversed by getCopyToParts - unreverse them.
379 std::reverse(Parts + RoundParts, Parts + NumParts);
381 NumParts = RoundParts;
382 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
383 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
386 // The number of parts is a power of 2. Repeatedly bisect the value using
388 Parts[0] = DAG.getNode(ISD::BITCAST, DL,
389 EVT::getIntegerVT(*DAG.getContext(),
390 ValueVT.getSizeInBits()),
393 for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) {
394 for (unsigned i = 0; i < NumParts; i += StepSize) {
395 unsigned ThisBits = StepSize * PartBits / 2;
396 EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits);
397 SDValue &Part0 = Parts[i];
398 SDValue &Part1 = Parts[i+StepSize/2];
400 Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
401 ThisVT, Part0, DAG.getIntPtrConstant(1));
402 Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
403 ThisVT, Part0, DAG.getIntPtrConstant(0));
405 if (ThisBits == PartBits && ThisVT != PartVT) {
406 Part0 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part0);
407 Part1 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part1);
412 if (TLI.isBigEndian())
413 std::reverse(Parts, Parts + OrigNumParts);
417 /// getCopyToPartsVector - Create a series of nodes that contain the specified
418 /// value split into legal parts.
419 static void getCopyToPartsVector(SelectionDAG &DAG, DebugLoc DL,
420 SDValue Val, SDValue *Parts, unsigned NumParts,
422 EVT ValueVT = Val.getValueType();
423 assert(ValueVT.isVector() && "Not a vector");
424 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
427 if (PartVT == ValueVT) {
429 } else if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) {
430 // Bitconvert vector->vector case.
431 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
432 } else if (PartVT.isVector() &&
433 PartVT.getVectorElementType() == ValueVT.getVectorElementType()&&
434 PartVT.getVectorNumElements() > ValueVT.getVectorNumElements()) {
435 EVT ElementVT = PartVT.getVectorElementType();
436 // Vector widening case, e.g. <2 x float> -> <4 x float>. Shuffle in
438 SmallVector<SDValue, 16> Ops;
439 for (unsigned i = 0, e = ValueVT.getVectorNumElements(); i != e; ++i)
440 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
441 ElementVT, Val, DAG.getIntPtrConstant(i)));
443 for (unsigned i = ValueVT.getVectorNumElements(),
444 e = PartVT.getVectorNumElements(); i != e; ++i)
445 Ops.push_back(DAG.getUNDEF(ElementVT));
447 Val = DAG.getNode(ISD::BUILD_VECTOR, DL, PartVT, &Ops[0], Ops.size());
449 // FIXME: Use CONCAT for 2x -> 4x.
451 //SDValue UndefElts = DAG.getUNDEF(VectorTy);
452 //Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts);
454 // Vector -> scalar conversion.
455 assert(ValueVT.getVectorElementType() == PartVT &&
456 ValueVT.getVectorNumElements() == 1 &&
457 "Only trivial vector-to-scalar conversions should get here!");
458 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
459 PartVT, Val, DAG.getIntPtrConstant(0));
466 // Handle a multi-element vector.
467 EVT IntermediateVT, RegisterVT;
468 unsigned NumIntermediates;
469 unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT,
471 NumIntermediates, RegisterVT);
472 unsigned NumElements = ValueVT.getVectorNumElements();
474 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
475 NumParts = NumRegs; // Silence a compiler warning.
476 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
478 // Split the vector into intermediate operands.
479 SmallVector<SDValue, 8> Ops(NumIntermediates);
480 for (unsigned i = 0; i != NumIntermediates; ++i) {
481 if (IntermediateVT.isVector())
482 Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL,
484 DAG.getIntPtrConstant(i * (NumElements / NumIntermediates)));
486 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
487 IntermediateVT, Val, DAG.getIntPtrConstant(i));
490 // Split the intermediate operands into legal parts.
491 if (NumParts == NumIntermediates) {
492 // If the register was not expanded, promote or copy the value,
494 for (unsigned i = 0; i != NumParts; ++i)
495 getCopyToParts(DAG, DL, Ops[i], &Parts[i], 1, PartVT);
496 } else if (NumParts > 0) {
497 // If the intermediate type was expanded, split each the value into
499 assert(NumParts % NumIntermediates == 0 &&
500 "Must expand into a divisible number of parts!");
501 unsigned Factor = NumParts / NumIntermediates;
502 for (unsigned i = 0; i != NumIntermediates; ++i)
503 getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT);
511 /// RegsForValue - This struct represents the registers (physical or virtual)
512 /// that a particular set of values is assigned, and the type information
513 /// about the value. The most common situation is to represent one value at a
514 /// time, but struct or array values are handled element-wise as multiple
515 /// values. The splitting of aggregates is performed recursively, so that we
516 /// never have aggregate-typed registers. The values at this point do not
517 /// necessarily have legal types, so each value may require one or more
518 /// registers of some legal type.
520 struct RegsForValue {
521 /// ValueVTs - The value types of the values, which may not be legal, and
522 /// may need be promoted or synthesized from one or more registers.
524 SmallVector<EVT, 4> ValueVTs;
526 /// RegVTs - The value types of the registers. This is the same size as
527 /// ValueVTs and it records, for each value, what the type of the assigned
528 /// register or registers are. (Individual values are never synthesized
529 /// from more than one type of register.)
531 /// With virtual registers, the contents of RegVTs is redundant with TLI's
532 /// getRegisterType member function, however when with physical registers
533 /// it is necessary to have a separate record of the types.
535 SmallVector<EVT, 4> RegVTs;
537 /// Regs - This list holds the registers assigned to the values.
538 /// Each legal or promoted value requires one register, and each
539 /// expanded value requires multiple registers.
541 SmallVector<unsigned, 4> Regs;
545 RegsForValue(const SmallVector<unsigned, 4> ®s,
546 EVT regvt, EVT valuevt)
547 : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {}
549 RegsForValue(LLVMContext &Context, const TargetLowering &tli,
550 unsigned Reg, const Type *Ty) {
551 ComputeValueVTs(tli, Ty, ValueVTs);
553 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
554 EVT ValueVT = ValueVTs[Value];
555 unsigned NumRegs = tli.getNumRegisters(Context, ValueVT);
556 EVT RegisterVT = tli.getRegisterType(Context, ValueVT);
557 for (unsigned i = 0; i != NumRegs; ++i)
558 Regs.push_back(Reg + i);
559 RegVTs.push_back(RegisterVT);
564 /// areValueTypesLegal - Return true if types of all the values are legal.
565 bool areValueTypesLegal(const TargetLowering &TLI) {
566 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
567 EVT RegisterVT = RegVTs[Value];
568 if (!TLI.isTypeLegal(RegisterVT))
574 /// append - Add the specified values to this one.
575 void append(const RegsForValue &RHS) {
576 ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end());
577 RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end());
578 Regs.append(RHS.Regs.begin(), RHS.Regs.end());
581 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
582 /// this value and returns the result as a ValueVTs value. This uses
583 /// Chain/Flag as the input and updates them for the output Chain/Flag.
584 /// If the Flag pointer is NULL, no flag is used.
585 SDValue getCopyFromRegs(SelectionDAG &DAG, FunctionLoweringInfo &FuncInfo,
587 SDValue &Chain, SDValue *Flag) const;
589 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
590 /// specified value into the registers specified by this object. This uses
591 /// Chain/Flag as the input and updates them for the output Chain/Flag.
592 /// If the Flag pointer is NULL, no flag is used.
593 void getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
594 SDValue &Chain, SDValue *Flag) const;
596 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
597 /// operand list. This adds the code marker, matching input operand index
598 /// (if applicable), and includes the number of values added into it.
599 void AddInlineAsmOperands(unsigned Kind,
600 bool HasMatching, unsigned MatchingIdx,
602 std::vector<SDValue> &Ops) const;
606 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
607 /// this value and returns the result as a ValueVT value. This uses
608 /// Chain/Flag as the input and updates them for the output Chain/Flag.
609 /// If the Flag pointer is NULL, no flag is used.
610 SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
611 FunctionLoweringInfo &FuncInfo,
613 SDValue &Chain, SDValue *Flag) const {
614 // A Value with type {} or [0 x %t] needs no registers.
615 if (ValueVTs.empty())
618 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
620 // Assemble the legal parts into the final values.
621 SmallVector<SDValue, 4> Values(ValueVTs.size());
622 SmallVector<SDValue, 8> Parts;
623 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
624 // Copy the legal parts from the registers.
625 EVT ValueVT = ValueVTs[Value];
626 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
627 EVT RegisterVT = RegVTs[Value];
629 Parts.resize(NumRegs);
630 for (unsigned i = 0; i != NumRegs; ++i) {
633 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT);
635 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag);
636 *Flag = P.getValue(2);
639 Chain = P.getValue(1);
641 // If the source register was virtual and if we know something about it,
642 // add an assert node.
643 if (TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) &&
644 RegisterVT.isInteger() && !RegisterVT.isVector()) {
645 unsigned SlotNo = Regs[Part+i]-TargetRegisterInfo::FirstVirtualRegister;
646 if (FuncInfo.LiveOutRegInfo.size() > SlotNo) {
647 const FunctionLoweringInfo::LiveOutInfo &LOI =
648 FuncInfo.LiveOutRegInfo[SlotNo];
650 unsigned RegSize = RegisterVT.getSizeInBits();
651 unsigned NumSignBits = LOI.NumSignBits;
652 unsigned NumZeroBits = LOI.KnownZero.countLeadingOnes();
654 // FIXME: We capture more information than the dag can represent. For
655 // now, just use the tightest assertzext/assertsext possible.
657 EVT FromVT(MVT::Other);
658 if (NumSignBits == RegSize)
659 isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1
660 else if (NumZeroBits >= RegSize-1)
661 isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1
662 else if (NumSignBits > RegSize-8)
663 isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8
664 else if (NumZeroBits >= RegSize-8)
665 isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8
666 else if (NumSignBits > RegSize-16)
667 isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16
668 else if (NumZeroBits >= RegSize-16)
669 isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16
670 else if (NumSignBits > RegSize-32)
671 isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32
672 else if (NumZeroBits >= RegSize-32)
673 isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32
675 if (FromVT != MVT::Other)
676 P = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl,
677 RegisterVT, P, DAG.getValueType(FromVT));
684 Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(),
685 NumRegs, RegisterVT, ValueVT);
690 return DAG.getNode(ISD::MERGE_VALUES, dl,
691 DAG.getVTList(&ValueVTs[0], ValueVTs.size()),
692 &Values[0], ValueVTs.size());
695 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
696 /// specified value into the registers specified by this object. This uses
697 /// Chain/Flag as the input and updates them for the output Chain/Flag.
698 /// If the Flag pointer is NULL, no flag is used.
699 void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
700 SDValue &Chain, SDValue *Flag) const {
701 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
703 // Get the list of the values's legal parts.
704 unsigned NumRegs = Regs.size();
705 SmallVector<SDValue, 8> Parts(NumRegs);
706 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
707 EVT ValueVT = ValueVTs[Value];
708 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
709 EVT RegisterVT = RegVTs[Value];
711 getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value),
712 &Parts[Part], NumParts, RegisterVT);
716 // Copy the parts into the registers.
717 SmallVector<SDValue, 8> Chains(NumRegs);
718 for (unsigned i = 0; i != NumRegs; ++i) {
721 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]);
723 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag);
724 *Flag = Part.getValue(1);
727 Chains[i] = Part.getValue(0);
730 if (NumRegs == 1 || Flag)
731 // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is
732 // flagged to it. That is the CopyToReg nodes and the user are considered
733 // a single scheduling unit. If we create a TokenFactor and return it as
734 // chain, then the TokenFactor is both a predecessor (operand) of the
735 // user as well as a successor (the TF operands are flagged to the user).
736 // c1, f1 = CopyToReg
737 // c2, f2 = CopyToReg
738 // c3 = TokenFactor c1, c2
741 Chain = Chains[NumRegs-1];
743 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0], NumRegs);
746 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
747 /// operand list. This adds the code marker and includes the number of
748 /// values added into it.
749 void RegsForValue::AddInlineAsmOperands(unsigned Code, bool HasMatching,
750 unsigned MatchingIdx,
752 std::vector<SDValue> &Ops) const {
753 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
755 unsigned Flag = InlineAsm::getFlagWord(Code, Regs.size());
757 Flag = InlineAsm::getFlagWordForMatchingOp(Flag, MatchingIdx);
758 SDValue Res = DAG.getTargetConstant(Flag, MVT::i32);
761 for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) {
762 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVTs[Value]);
763 EVT RegisterVT = RegVTs[Value];
764 for (unsigned i = 0; i != NumRegs; ++i) {
765 assert(Reg < Regs.size() && "Mismatch in # registers expected");
766 Ops.push_back(DAG.getRegister(Regs[Reg++], RegisterVT));
771 void SelectionDAGBuilder::init(GCFunctionInfo *gfi, AliasAnalysis &aa) {
774 TD = DAG.getTarget().getTargetData();
777 /// clear - Clear out the current SelectionDAG and the associated
778 /// state and prepare this SelectionDAGBuilder object to be used
779 /// for a new block. This doesn't clear out information about
780 /// additional blocks that are needed to complete switch lowering
781 /// or PHI node updating; that information is cleared out as it is
783 void SelectionDAGBuilder::clear() {
785 UnusedArgNodeMap.clear();
786 PendingLoads.clear();
787 PendingExports.clear();
788 DanglingDebugInfoMap.clear();
789 CurDebugLoc = DebugLoc();
793 /// getRoot - Return the current virtual root of the Selection DAG,
794 /// flushing any PendingLoad items. This must be done before emitting
795 /// a store or any other node that may need to be ordered after any
796 /// prior load instructions.
798 SDValue SelectionDAGBuilder::getRoot() {
799 if (PendingLoads.empty())
800 return DAG.getRoot();
802 if (PendingLoads.size() == 1) {
803 SDValue Root = PendingLoads[0];
805 PendingLoads.clear();
809 // Otherwise, we have to make a token factor node.
810 SDValue Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
811 &PendingLoads[0], PendingLoads.size());
812 PendingLoads.clear();
817 /// getControlRoot - Similar to getRoot, but instead of flushing all the
818 /// PendingLoad items, flush all the PendingExports items. It is necessary
819 /// to do this before emitting a terminator instruction.
821 SDValue SelectionDAGBuilder::getControlRoot() {
822 SDValue Root = DAG.getRoot();
824 if (PendingExports.empty())
827 // Turn all of the CopyToReg chains into one factored node.
828 if (Root.getOpcode() != ISD::EntryToken) {
829 unsigned i = 0, e = PendingExports.size();
830 for (; i != e; ++i) {
831 assert(PendingExports[i].getNode()->getNumOperands() > 1);
832 if (PendingExports[i].getNode()->getOperand(0) == Root)
833 break; // Don't add the root if we already indirectly depend on it.
837 PendingExports.push_back(Root);
840 Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
842 PendingExports.size());
843 PendingExports.clear();
848 void SelectionDAGBuilder::AssignOrderingToNode(const SDNode *Node) {
849 if (DAG.GetOrdering(Node) != 0) return; // Already has ordering.
850 DAG.AssignOrdering(Node, SDNodeOrder);
852 for (unsigned I = 0, E = Node->getNumOperands(); I != E; ++I)
853 AssignOrderingToNode(Node->getOperand(I).getNode());
856 void SelectionDAGBuilder::visit(const Instruction &I) {
857 // Set up outgoing PHI node register values before emitting the terminator.
858 if (isa<TerminatorInst>(&I))
859 HandlePHINodesInSuccessorBlocks(I.getParent());
861 CurDebugLoc = I.getDebugLoc();
863 visit(I.getOpcode(), I);
865 if (!isa<TerminatorInst>(&I) && !HasTailCall)
866 CopyToExportRegsIfNeeded(&I);
868 CurDebugLoc = DebugLoc();
871 void SelectionDAGBuilder::visitPHI(const PHINode &) {
872 llvm_unreachable("SelectionDAGBuilder shouldn't visit PHI nodes!");
875 void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) {
876 // Note: this doesn't use InstVisitor, because it has to work with
877 // ConstantExpr's in addition to instructions.
879 default: llvm_unreachable("Unknown instruction type encountered!");
880 // Build the switch statement using the Instruction.def file.
881 #define HANDLE_INST(NUM, OPCODE, CLASS) \
882 case Instruction::OPCODE: visit##OPCODE((CLASS&)I); break;
883 #include "llvm/Instruction.def"
886 // Assign the ordering to the freshly created DAG nodes.
887 if (NodeMap.count(&I)) {
889 AssignOrderingToNode(getValue(&I).getNode());
893 // resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V,
894 // generate the debug data structures now that we've seen its definition.
895 void SelectionDAGBuilder::resolveDanglingDebugInfo(const Value *V,
897 DanglingDebugInfo &DDI = DanglingDebugInfoMap[V];
899 const DbgValueInst *DI = DDI.getDI();
900 DebugLoc dl = DDI.getdl();
901 unsigned DbgSDNodeOrder = DDI.getSDNodeOrder();
902 MDNode *Variable = DI->getVariable();
903 uint64_t Offset = DI->getOffset();
906 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, Val)) {
907 SDV = DAG.getDbgValue(Variable, Val.getNode(),
908 Val.getResNo(), Offset, dl, DbgSDNodeOrder);
909 DAG.AddDbgValue(SDV, Val.getNode(), false);
912 SDV = DAG.getDbgValue(Variable, UndefValue::get(V->getType()),
913 Offset, dl, SDNodeOrder);
914 DAG.AddDbgValue(SDV, 0, false);
916 DanglingDebugInfoMap[V] = DanglingDebugInfo();
920 // getValue - Return an SDValue for the given Value.
921 SDValue SelectionDAGBuilder::getValue(const Value *V) {
922 // If we already have an SDValue for this value, use it. It's important
923 // to do this first, so that we don't create a CopyFromReg if we already
924 // have a regular SDValue.
925 SDValue &N = NodeMap[V];
926 if (N.getNode()) return N;
928 // If there's a virtual register allocated and initialized for this
930 DenseMap<const Value *, unsigned>::iterator It = FuncInfo.ValueMap.find(V);
931 if (It != FuncInfo.ValueMap.end()) {
932 unsigned InReg = It->second;
933 RegsForValue RFV(*DAG.getContext(), TLI, InReg, V->getType());
934 SDValue Chain = DAG.getEntryNode();
935 return N = RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain,NULL);
938 // Otherwise create a new SDValue and remember it.
939 SDValue Val = getValueImpl(V);
941 resolveDanglingDebugInfo(V, Val);
945 /// getNonRegisterValue - Return an SDValue for the given Value, but
946 /// don't look in FuncInfo.ValueMap for a virtual register.
947 SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) {
948 // If we already have an SDValue for this value, use it.
949 SDValue &N = NodeMap[V];
950 if (N.getNode()) return N;
952 // Otherwise create a new SDValue and remember it.
953 SDValue Val = getValueImpl(V);
955 resolveDanglingDebugInfo(V, Val);
959 /// getValueImpl - Helper function for getValue and getNonRegisterValue.
960 /// Create an SDValue for the given value.
961 SDValue SelectionDAGBuilder::getValueImpl(const Value *V) {
962 if (const Constant *C = dyn_cast<Constant>(V)) {
963 EVT VT = TLI.getValueType(V->getType(), true);
965 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C))
966 return DAG.getConstant(*CI, VT);
968 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
969 return DAG.getGlobalAddress(GV, getCurDebugLoc(), VT);
971 if (isa<ConstantPointerNull>(C))
972 return DAG.getConstant(0, TLI.getPointerTy());
974 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
975 return DAG.getConstantFP(*CFP, VT);
977 if (isa<UndefValue>(C) && !V->getType()->isAggregateType())
978 return DAG.getUNDEF(VT);
980 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
981 visit(CE->getOpcode(), *CE);
982 SDValue N1 = NodeMap[V];
983 assert(N1.getNode() && "visit didn't populate the NodeMap!");
987 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) {
988 SmallVector<SDValue, 4> Constants;
989 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end();
991 SDNode *Val = getValue(*OI).getNode();
992 // If the operand is an empty aggregate, there are no values.
994 // Add each leaf value from the operand to the Constants list
995 // to form a flattened list of all the values.
996 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
997 Constants.push_back(SDValue(Val, i));
1000 return DAG.getMergeValues(&Constants[0], Constants.size(),
1004 if (C->getType()->isStructTy() || C->getType()->isArrayTy()) {
1005 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
1006 "Unknown struct or array constant!");
1008 SmallVector<EVT, 4> ValueVTs;
1009 ComputeValueVTs(TLI, C->getType(), ValueVTs);
1010 unsigned NumElts = ValueVTs.size();
1012 return SDValue(); // empty struct
1013 SmallVector<SDValue, 4> Constants(NumElts);
1014 for (unsigned i = 0; i != NumElts; ++i) {
1015 EVT EltVT = ValueVTs[i];
1016 if (isa<UndefValue>(C))
1017 Constants[i] = DAG.getUNDEF(EltVT);
1018 else if (EltVT.isFloatingPoint())
1019 Constants[i] = DAG.getConstantFP(0, EltVT);
1021 Constants[i] = DAG.getConstant(0, EltVT);
1024 return DAG.getMergeValues(&Constants[0], NumElts,
1028 if (const BlockAddress *BA = dyn_cast<BlockAddress>(C))
1029 return DAG.getBlockAddress(BA, VT);
1031 const VectorType *VecTy = cast<VectorType>(V->getType());
1032 unsigned NumElements = VecTy->getNumElements();
1034 // Now that we know the number and type of the elements, get that number of
1035 // elements into the Ops array based on what kind of constant it is.
1036 SmallVector<SDValue, 16> Ops;
1037 if (const ConstantVector *CP = dyn_cast<ConstantVector>(C)) {
1038 for (unsigned i = 0; i != NumElements; ++i)
1039 Ops.push_back(getValue(CP->getOperand(i)));
1041 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!");
1042 EVT EltVT = TLI.getValueType(VecTy->getElementType());
1045 if (EltVT.isFloatingPoint())
1046 Op = DAG.getConstantFP(0, EltVT);
1048 Op = DAG.getConstant(0, EltVT);
1049 Ops.assign(NumElements, Op);
1052 // Create a BUILD_VECTOR node.
1053 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
1054 VT, &Ops[0], Ops.size());
1057 // If this is a static alloca, generate it as the frameindex instead of
1059 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
1060 DenseMap<const AllocaInst*, int>::iterator SI =
1061 FuncInfo.StaticAllocaMap.find(AI);
1062 if (SI != FuncInfo.StaticAllocaMap.end())
1063 return DAG.getFrameIndex(SI->second, TLI.getPointerTy());
1066 // If this is an instruction which fast-isel has deferred, select it now.
1067 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
1068 unsigned InReg = FuncInfo.InitializeRegForValue(Inst);
1069 RegsForValue RFV(*DAG.getContext(), TLI, InReg, Inst->getType());
1070 SDValue Chain = DAG.getEntryNode();
1071 return RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain, NULL);
1074 llvm_unreachable("Can't get register for value!");
1078 void SelectionDAGBuilder::visitRet(const ReturnInst &I) {
1079 SDValue Chain = getControlRoot();
1080 SmallVector<ISD::OutputArg, 8> Outs;
1081 SmallVector<SDValue, 8> OutVals;
1083 if (!FuncInfo.CanLowerReturn) {
1084 unsigned DemoteReg = FuncInfo.DemoteRegister;
1085 const Function *F = I.getParent()->getParent();
1087 // Emit a store of the return value through the virtual register.
1088 // Leave Outs empty so that LowerReturn won't try to load return
1089 // registers the usual way.
1090 SmallVector<EVT, 1> PtrValueVTs;
1091 ComputeValueVTs(TLI, PointerType::getUnqual(F->getReturnType()),
1094 SDValue RetPtr = DAG.getRegister(DemoteReg, PtrValueVTs[0]);
1095 SDValue RetOp = getValue(I.getOperand(0));
1097 SmallVector<EVT, 4> ValueVTs;
1098 SmallVector<uint64_t, 4> Offsets;
1099 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs, &Offsets);
1100 unsigned NumValues = ValueVTs.size();
1102 SmallVector<SDValue, 4> Chains(NumValues);
1103 for (unsigned i = 0; i != NumValues; ++i) {
1104 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(),
1105 RetPtr.getValueType(), RetPtr,
1106 DAG.getIntPtrConstant(Offsets[i]));
1108 DAG.getStore(Chain, getCurDebugLoc(),
1109 SDValue(RetOp.getNode(), RetOp.getResNo() + i),
1110 // FIXME: better loc info would be nice.
1111 Add, MachinePointerInfo(), false, false, 0);
1114 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
1115 MVT::Other, &Chains[0], NumValues);
1116 } else if (I.getNumOperands() != 0) {
1117 SmallVector<EVT, 4> ValueVTs;
1118 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs);
1119 unsigned NumValues = ValueVTs.size();
1121 SDValue RetOp = getValue(I.getOperand(0));
1122 for (unsigned j = 0, f = NumValues; j != f; ++j) {
1123 EVT VT = ValueVTs[j];
1125 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1127 const Function *F = I.getParent()->getParent();
1128 if (F->paramHasAttr(0, Attribute::SExt))
1129 ExtendKind = ISD::SIGN_EXTEND;
1130 else if (F->paramHasAttr(0, Attribute::ZExt))
1131 ExtendKind = ISD::ZERO_EXTEND;
1133 // FIXME: C calling convention requires the return type to be promoted
1134 // to at least 32-bit. But this is not necessary for non-C calling
1135 // conventions. The frontend should mark functions whose return values
1136 // require promoting with signext or zeroext attributes.
1137 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
1138 EVT MinVT = TLI.getRegisterType(*DAG.getContext(), MVT::i32);
1139 if (VT.bitsLT(MinVT))
1143 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), VT);
1144 EVT PartVT = TLI.getRegisterType(*DAG.getContext(), VT);
1145 SmallVector<SDValue, 4> Parts(NumParts);
1146 getCopyToParts(DAG, getCurDebugLoc(),
1147 SDValue(RetOp.getNode(), RetOp.getResNo() + j),
1148 &Parts[0], NumParts, PartVT, ExtendKind);
1150 // 'inreg' on function refers to return value
1151 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1152 if (F->paramHasAttr(0, Attribute::InReg))
1155 // Propagate extension type if any
1156 if (F->paramHasAttr(0, Attribute::SExt))
1158 else if (F->paramHasAttr(0, Attribute::ZExt))
1161 for (unsigned i = 0; i < NumParts; ++i) {
1162 Outs.push_back(ISD::OutputArg(Flags, Parts[i].getValueType(),
1164 OutVals.push_back(Parts[i]);
1170 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1171 CallingConv::ID CallConv =
1172 DAG.getMachineFunction().getFunction()->getCallingConv();
1173 Chain = TLI.LowerReturn(Chain, CallConv, isVarArg,
1174 Outs, OutVals, getCurDebugLoc(), DAG);
1176 // Verify that the target's LowerReturn behaved as expected.
1177 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
1178 "LowerReturn didn't return a valid chain!");
1180 // Update the DAG with the new chain value resulting from return lowering.
1184 /// CopyToExportRegsIfNeeded - If the given value has virtual registers
1185 /// created for it, emit nodes to copy the value into the virtual
1187 void SelectionDAGBuilder::CopyToExportRegsIfNeeded(const Value *V) {
1188 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
1189 if (VMI != FuncInfo.ValueMap.end()) {
1190 assert(!V->use_empty() && "Unused value assigned virtual registers!");
1191 CopyValueToVirtualRegister(V, VMI->second);
1195 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
1196 /// the current basic block, add it to ValueMap now so that we'll get a
1198 void SelectionDAGBuilder::ExportFromCurrentBlock(const Value *V) {
1199 // No need to export constants.
1200 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
1202 // Already exported?
1203 if (FuncInfo.isExportedInst(V)) return;
1205 unsigned Reg = FuncInfo.InitializeRegForValue(V);
1206 CopyValueToVirtualRegister(V, Reg);
1209 bool SelectionDAGBuilder::isExportableFromCurrentBlock(const Value *V,
1210 const BasicBlock *FromBB) {
1211 // The operands of the setcc have to be in this block. We don't know
1212 // how to export them from some other block.
1213 if (const Instruction *VI = dyn_cast<Instruction>(V)) {
1214 // Can export from current BB.
1215 if (VI->getParent() == FromBB)
1218 // Is already exported, noop.
1219 return FuncInfo.isExportedInst(V);
1222 // If this is an argument, we can export it if the BB is the entry block or
1223 // if it is already exported.
1224 if (isa<Argument>(V)) {
1225 if (FromBB == &FromBB->getParent()->getEntryBlock())
1228 // Otherwise, can only export this if it is already exported.
1229 return FuncInfo.isExportedInst(V);
1232 // Otherwise, constants can always be exported.
1236 static bool InBlock(const Value *V, const BasicBlock *BB) {
1237 if (const Instruction *I = dyn_cast<Instruction>(V))
1238 return I->getParent() == BB;
1242 /// EmitBranchForMergedCondition - Helper method for FindMergedConditions.
1243 /// This function emits a branch and is used at the leaves of an OR or an
1244 /// AND operator tree.
1247 SelectionDAGBuilder::EmitBranchForMergedCondition(const Value *Cond,
1248 MachineBasicBlock *TBB,
1249 MachineBasicBlock *FBB,
1250 MachineBasicBlock *CurBB,
1251 MachineBasicBlock *SwitchBB) {
1252 const BasicBlock *BB = CurBB->getBasicBlock();
1254 // If the leaf of the tree is a comparison, merge the condition into
1256 if (const CmpInst *BOp = dyn_cast<CmpInst>(Cond)) {
1257 // The operands of the cmp have to be in this block. We don't know
1258 // how to export them from some other block. If this is the first block
1259 // of the sequence, no exporting is needed.
1260 if (CurBB == SwitchBB ||
1261 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
1262 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) {
1263 ISD::CondCode Condition;
1264 if (const ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
1265 Condition = getICmpCondCode(IC->getPredicate());
1266 } else if (const FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) {
1267 Condition = getFCmpCondCode(FC->getPredicate());
1269 Condition = ISD::SETEQ; // silence warning.
1270 llvm_unreachable("Unknown compare instruction");
1273 CaseBlock CB(Condition, BOp->getOperand(0),
1274 BOp->getOperand(1), NULL, TBB, FBB, CurBB);
1275 SwitchCases.push_back(CB);
1280 // Create a CaseBlock record representing this branch.
1281 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()),
1282 NULL, TBB, FBB, CurBB);
1283 SwitchCases.push_back(CB);
1286 /// FindMergedConditions - If Cond is an expression like
1287 void SelectionDAGBuilder::FindMergedConditions(const Value *Cond,
1288 MachineBasicBlock *TBB,
1289 MachineBasicBlock *FBB,
1290 MachineBasicBlock *CurBB,
1291 MachineBasicBlock *SwitchBB,
1293 // If this node is not part of the or/and tree, emit it as a branch.
1294 const Instruction *BOp = dyn_cast<Instruction>(Cond);
1295 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
1296 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
1297 BOp->getParent() != CurBB->getBasicBlock() ||
1298 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
1299 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
1300 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB);
1304 // Create TmpBB after CurBB.
1305 MachineFunction::iterator BBI = CurBB;
1306 MachineFunction &MF = DAG.getMachineFunction();
1307 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock());
1308 CurBB->getParent()->insert(++BBI, TmpBB);
1310 if (Opc == Instruction::Or) {
1311 // Codegen X | Y as:
1319 // Emit the LHS condition.
1320 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, SwitchBB, Opc);
1322 // Emit the RHS condition into TmpBB.
1323 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
1325 assert(Opc == Instruction::And && "Unknown merge op!");
1326 // Codegen X & Y as:
1333 // This requires creation of TmpBB after CurBB.
1335 // Emit the LHS condition.
1336 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, SwitchBB, Opc);
1338 // Emit the RHS condition into TmpBB.
1339 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
1343 /// If the set of cases should be emitted as a series of branches, return true.
1344 /// If we should emit this as a bunch of and/or'd together conditions, return
1347 SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases){
1348 if (Cases.size() != 2) return true;
1350 // If this is two comparisons of the same values or'd or and'd together, they
1351 // will get folded into a single comparison, so don't emit two blocks.
1352 if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
1353 Cases[0].CmpRHS == Cases[1].CmpRHS) ||
1354 (Cases[0].CmpRHS == Cases[1].CmpLHS &&
1355 Cases[0].CmpLHS == Cases[1].CmpRHS)) {
1359 // Handle: (X != null) | (Y != null) --> (X|Y) != 0
1360 // Handle: (X == null) & (Y == null) --> (X|Y) == 0
1361 if (Cases[0].CmpRHS == Cases[1].CmpRHS &&
1362 Cases[0].CC == Cases[1].CC &&
1363 isa<Constant>(Cases[0].CmpRHS) &&
1364 cast<Constant>(Cases[0].CmpRHS)->isNullValue()) {
1365 if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB)
1367 if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB)
1374 void SelectionDAGBuilder::visitBr(const BranchInst &I) {
1375 MachineBasicBlock *BrMBB = FuncInfo.MBB;
1377 // Update machine-CFG edges.
1378 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
1380 // Figure out which block is immediately after the current one.
1381 MachineBasicBlock *NextBlock = 0;
1382 MachineFunction::iterator BBI = BrMBB;
1383 if (++BBI != FuncInfo.MF->end())
1386 if (I.isUnconditional()) {
1387 // Update machine-CFG edges.
1388 BrMBB->addSuccessor(Succ0MBB);
1390 // If this is not a fall-through branch, emit the branch.
1391 if (Succ0MBB != NextBlock)
1392 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
1393 MVT::Other, getControlRoot(),
1394 DAG.getBasicBlock(Succ0MBB)));
1399 // If this condition is one of the special cases we handle, do special stuff
1401 const Value *CondVal = I.getCondition();
1402 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
1404 // If this is a series of conditions that are or'd or and'd together, emit
1405 // this as a sequence of branches instead of setcc's with and/or operations.
1406 // As long as jumps are not expensive, this should improve performance.
1407 // For example, instead of something like:
1420 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
1421 if (!TLI.isJumpExpensive() &&
1423 (BOp->getOpcode() == Instruction::And ||
1424 BOp->getOpcode() == Instruction::Or)) {
1425 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB,
1427 // If the compares in later blocks need to use values not currently
1428 // exported from this block, export them now. This block should always
1429 // be the first entry.
1430 assert(SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!");
1432 // Allow some cases to be rejected.
1433 if (ShouldEmitAsBranches(SwitchCases)) {
1434 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1435 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1436 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1439 // Emit the branch for this block.
1440 visitSwitchCase(SwitchCases[0], BrMBB);
1441 SwitchCases.erase(SwitchCases.begin());
1445 // Okay, we decided not to do this, remove any inserted MBB's and clear
1447 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
1448 FuncInfo.MF->erase(SwitchCases[i].ThisBB);
1450 SwitchCases.clear();
1454 // Create a CaseBlock record representing this branch.
1455 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()),
1456 NULL, Succ0MBB, Succ1MBB, BrMBB);
1458 // Use visitSwitchCase to actually insert the fast branch sequence for this
1460 visitSwitchCase(CB, BrMBB);
1463 /// visitSwitchCase - Emits the necessary code to represent a single node in
1464 /// the binary search tree resulting from lowering a switch instruction.
1465 void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB,
1466 MachineBasicBlock *SwitchBB) {
1468 SDValue CondLHS = getValue(CB.CmpLHS);
1469 DebugLoc dl = getCurDebugLoc();
1471 // Build the setcc now.
1472 if (CB.CmpMHS == NULL) {
1473 // Fold "(X == true)" to X and "(X == false)" to !X to
1474 // handle common cases produced by branch lowering.
1475 if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) &&
1476 CB.CC == ISD::SETEQ)
1478 else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) &&
1479 CB.CC == ISD::SETEQ) {
1480 SDValue True = DAG.getConstant(1, CondLHS.getValueType());
1481 Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True);
1483 Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1485 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now");
1487 const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue();
1488 const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue();
1490 SDValue CmpOp = getValue(CB.CmpMHS);
1491 EVT VT = CmpOp.getValueType();
1493 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) {
1494 Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, VT),
1497 SDValue SUB = DAG.getNode(ISD::SUB, dl,
1498 VT, CmpOp, DAG.getConstant(Low, VT));
1499 Cond = DAG.getSetCC(dl, MVT::i1, SUB,
1500 DAG.getConstant(High-Low, VT), ISD::SETULE);
1504 // Update successor info
1505 SwitchBB->addSuccessor(CB.TrueBB);
1506 SwitchBB->addSuccessor(CB.FalseBB);
1508 // Set NextBlock to be the MBB immediately after the current one, if any.
1509 // This is used to avoid emitting unnecessary branches to the next block.
1510 MachineBasicBlock *NextBlock = 0;
1511 MachineFunction::iterator BBI = SwitchBB;
1512 if (++BBI != FuncInfo.MF->end())
1515 // If the lhs block is the next block, invert the condition so that we can
1516 // fall through to the lhs instead of the rhs block.
1517 if (CB.TrueBB == NextBlock) {
1518 std::swap(CB.TrueBB, CB.FalseBB);
1519 SDValue True = DAG.getConstant(1, Cond.getValueType());
1520 Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True);
1523 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1524 MVT::Other, getControlRoot(), Cond,
1525 DAG.getBasicBlock(CB.TrueBB));
1527 // Insert the false branch. Do this even if it's a fall through branch,
1528 // this makes it easier to do DAG optimizations which require inverting
1529 // the branch condition.
1530 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1531 DAG.getBasicBlock(CB.FalseBB));
1533 DAG.setRoot(BrCond);
1536 /// visitJumpTable - Emit JumpTable node in the current MBB
1537 void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) {
1538 // Emit the code for the jump table
1539 assert(JT.Reg != -1U && "Should lower JT Header first!");
1540 EVT PTy = TLI.getPointerTy();
1541 SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(),
1543 SDValue Table = DAG.getJumpTable(JT.JTI, PTy);
1544 SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurDebugLoc(),
1545 MVT::Other, Index.getValue(1),
1547 DAG.setRoot(BrJumpTable);
1550 /// visitJumpTableHeader - This function emits necessary code to produce index
1551 /// in the JumpTable from switch case.
1552 void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT,
1553 JumpTableHeader &JTH,
1554 MachineBasicBlock *SwitchBB) {
1555 // Subtract the lowest switch case value from the value being switched on and
1556 // conditional branch to default mbb if the result is greater than the
1557 // difference between smallest and largest cases.
1558 SDValue SwitchOp = getValue(JTH.SValue);
1559 EVT VT = SwitchOp.getValueType();
1560 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1561 DAG.getConstant(JTH.First, VT));
1563 // The SDNode we just created, which holds the value being switched on minus
1564 // the smallest case value, needs to be copied to a virtual register so it
1565 // can be used as an index into the jump table in a subsequent basic block.
1566 // This value may be smaller or larger than the target's pointer type, and
1567 // therefore require extension or truncating.
1568 SwitchOp = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(), TLI.getPointerTy());
1570 unsigned JumpTableReg = FuncInfo.CreateReg(TLI.getPointerTy());
1571 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1572 JumpTableReg, SwitchOp);
1573 JT.Reg = JumpTableReg;
1575 // Emit the range check for the jump table, and branch to the default block
1576 // for the switch statement if the value being switched on exceeds the largest
1577 // case in the switch.
1578 SDValue CMP = DAG.getSetCC(getCurDebugLoc(),
1579 TLI.getSetCCResultType(Sub.getValueType()), Sub,
1580 DAG.getConstant(JTH.Last-JTH.First,VT),
1583 // Set NextBlock to be the MBB immediately after the current one, if any.
1584 // This is used to avoid emitting unnecessary branches to the next block.
1585 MachineBasicBlock *NextBlock = 0;
1586 MachineFunction::iterator BBI = SwitchBB;
1588 if (++BBI != FuncInfo.MF->end())
1591 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1592 MVT::Other, CopyTo, CMP,
1593 DAG.getBasicBlock(JT.Default));
1595 if (JT.MBB != NextBlock)
1596 BrCond = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrCond,
1597 DAG.getBasicBlock(JT.MBB));
1599 DAG.setRoot(BrCond);
1602 /// visitBitTestHeader - This function emits necessary code to produce value
1603 /// suitable for "bit tests"
1604 void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B,
1605 MachineBasicBlock *SwitchBB) {
1606 // Subtract the minimum value
1607 SDValue SwitchOp = getValue(B.SValue);
1608 EVT VT = SwitchOp.getValueType();
1609 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1610 DAG.getConstant(B.First, VT));
1613 SDValue RangeCmp = DAG.getSetCC(getCurDebugLoc(),
1614 TLI.getSetCCResultType(Sub.getValueType()),
1615 Sub, DAG.getConstant(B.Range, VT),
1618 SDValue ShiftOp = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(),
1619 TLI.getPointerTy());
1621 B.Reg = FuncInfo.CreateReg(TLI.getPointerTy());
1622 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1625 // Set NextBlock to be the MBB immediately after the current one, if any.
1626 // This is used to avoid emitting unnecessary branches to the next block.
1627 MachineBasicBlock *NextBlock = 0;
1628 MachineFunction::iterator BBI = SwitchBB;
1629 if (++BBI != FuncInfo.MF->end())
1632 MachineBasicBlock* MBB = B.Cases[0].ThisBB;
1634 SwitchBB->addSuccessor(B.Default);
1635 SwitchBB->addSuccessor(MBB);
1637 SDValue BrRange = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1638 MVT::Other, CopyTo, RangeCmp,
1639 DAG.getBasicBlock(B.Default));
1641 if (MBB != NextBlock)
1642 BrRange = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, CopyTo,
1643 DAG.getBasicBlock(MBB));
1645 DAG.setRoot(BrRange);
1648 /// visitBitTestCase - this function produces one "bit test"
1649 void SelectionDAGBuilder::visitBitTestCase(MachineBasicBlock* NextMBB,
1652 MachineBasicBlock *SwitchBB) {
1653 SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(), Reg,
1654 TLI.getPointerTy());
1656 if (CountPopulation_64(B.Mask) == 1) {
1657 // Testing for a single bit; just compare the shift count with what it
1658 // would need to be to shift a 1 bit in that position.
1659 Cmp = DAG.getSetCC(getCurDebugLoc(),
1660 TLI.getSetCCResultType(ShiftOp.getValueType()),
1662 DAG.getConstant(CountTrailingZeros_64(B.Mask),
1663 TLI.getPointerTy()),
1666 // Make desired shift
1667 SDValue SwitchVal = DAG.getNode(ISD::SHL, getCurDebugLoc(),
1669 DAG.getConstant(1, TLI.getPointerTy()),
1672 // Emit bit tests and jumps
1673 SDValue AndOp = DAG.getNode(ISD::AND, getCurDebugLoc(),
1674 TLI.getPointerTy(), SwitchVal,
1675 DAG.getConstant(B.Mask, TLI.getPointerTy()));
1676 Cmp = DAG.getSetCC(getCurDebugLoc(),
1677 TLI.getSetCCResultType(AndOp.getValueType()),
1678 AndOp, DAG.getConstant(0, TLI.getPointerTy()),
1682 SwitchBB->addSuccessor(B.TargetBB);
1683 SwitchBB->addSuccessor(NextMBB);
1685 SDValue BrAnd = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1686 MVT::Other, getControlRoot(),
1687 Cmp, DAG.getBasicBlock(B.TargetBB));
1689 // Set NextBlock to be the MBB immediately after the current one, if any.
1690 // This is used to avoid emitting unnecessary branches to the next block.
1691 MachineBasicBlock *NextBlock = 0;
1692 MachineFunction::iterator BBI = SwitchBB;
1693 if (++BBI != FuncInfo.MF->end())
1696 if (NextMBB != NextBlock)
1697 BrAnd = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrAnd,
1698 DAG.getBasicBlock(NextMBB));
1703 void SelectionDAGBuilder::visitInvoke(const InvokeInst &I) {
1704 MachineBasicBlock *InvokeMBB = FuncInfo.MBB;
1706 // Retrieve successors.
1707 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
1708 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)];
1710 const Value *Callee(I.getCalledValue());
1711 if (isa<InlineAsm>(Callee))
1714 LowerCallTo(&I, getValue(Callee), false, LandingPad);
1716 // If the value of the invoke is used outside of its defining block, make it
1717 // available as a virtual register.
1718 CopyToExportRegsIfNeeded(&I);
1720 // Update successor info
1721 InvokeMBB->addSuccessor(Return);
1722 InvokeMBB->addSuccessor(LandingPad);
1724 // Drop into normal successor.
1725 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
1726 MVT::Other, getControlRoot(),
1727 DAG.getBasicBlock(Return)));
1730 void SelectionDAGBuilder::visitUnwind(const UnwindInst &I) {
1733 /// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for
1734 /// small case ranges).
1735 bool SelectionDAGBuilder::handleSmallSwitchRange(CaseRec& CR,
1736 CaseRecVector& WorkList,
1738 MachineBasicBlock *Default,
1739 MachineBasicBlock *SwitchBB) {
1740 Case& BackCase = *(CR.Range.second-1);
1742 // Size is the number of Cases represented by this range.
1743 size_t Size = CR.Range.second - CR.Range.first;
1747 // Get the MachineFunction which holds the current MBB. This is used when
1748 // inserting any additional MBBs necessary to represent the switch.
1749 MachineFunction *CurMF = FuncInfo.MF;
1751 // Figure out which block is immediately after the current one.
1752 MachineBasicBlock *NextBlock = 0;
1753 MachineFunction::iterator BBI = CR.CaseBB;
1755 if (++BBI != FuncInfo.MF->end())
1758 // If any two of the cases has the same destination, and if one value
1759 // is the same as the other, but has one bit unset that the other has set,
1760 // use bit manipulation to do two compares at once. For example:
1761 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
1762 // TODO: This could be extended to merge any 2 cases in switches with 3 cases.
1763 // TODO: Handle cases where CR.CaseBB != SwitchBB.
1764 if (Size == 2 && CR.CaseBB == SwitchBB) {
1765 Case &Small = *CR.Range.first;
1766 Case &Big = *(CR.Range.second-1);
1768 if (Small.Low == Small.High && Big.Low == Big.High && Small.BB == Big.BB) {
1769 const APInt& SmallValue = cast<ConstantInt>(Small.Low)->getValue();
1770 const APInt& BigValue = cast<ConstantInt>(Big.Low)->getValue();
1772 // Check that there is only one bit different.
1773 if (BigValue.countPopulation() == SmallValue.countPopulation() + 1 &&
1774 (SmallValue | BigValue) == BigValue) {
1775 // Isolate the common bit.
1776 APInt CommonBit = BigValue & ~SmallValue;
1777 assert((SmallValue | CommonBit) == BigValue &&
1778 CommonBit.countPopulation() == 1 && "Not a common bit?");
1780 SDValue CondLHS = getValue(SV);
1781 EVT VT = CondLHS.getValueType();
1782 DebugLoc DL = getCurDebugLoc();
1784 SDValue Or = DAG.getNode(ISD::OR, DL, VT, CondLHS,
1785 DAG.getConstant(CommonBit, VT));
1786 SDValue Cond = DAG.getSetCC(DL, MVT::i1,
1787 Or, DAG.getConstant(BigValue, VT),
1790 // Update successor info.
1791 SwitchBB->addSuccessor(Small.BB);
1792 SwitchBB->addSuccessor(Default);
1794 // Insert the true branch.
1795 SDValue BrCond = DAG.getNode(ISD::BRCOND, DL, MVT::Other,
1796 getControlRoot(), Cond,
1797 DAG.getBasicBlock(Small.BB));
1799 // Insert the false branch.
1800 BrCond = DAG.getNode(ISD::BR, DL, MVT::Other, BrCond,
1801 DAG.getBasicBlock(Default));
1803 DAG.setRoot(BrCond);
1809 // Rearrange the case blocks so that the last one falls through if possible.
1810 if (NextBlock && Default != NextBlock && BackCase.BB != NextBlock) {
1811 // The last case block won't fall through into 'NextBlock' if we emit the
1812 // branches in this order. See if rearranging a case value would help.
1813 for (CaseItr I = CR.Range.first, E = CR.Range.second-1; I != E; ++I) {
1814 if (I->BB == NextBlock) {
1815 std::swap(*I, BackCase);
1821 // Create a CaseBlock record representing a conditional branch to
1822 // the Case's target mbb if the value being switched on SV is equal
1824 MachineBasicBlock *CurBlock = CR.CaseBB;
1825 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
1826 MachineBasicBlock *FallThrough;
1828 FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock());
1829 CurMF->insert(BBI, FallThrough);
1831 // Put SV in a virtual register to make it available from the new blocks.
1832 ExportFromCurrentBlock(SV);
1834 // If the last case doesn't match, go to the default block.
1835 FallThrough = Default;
1838 const Value *RHS, *LHS, *MHS;
1840 if (I->High == I->Low) {
1841 // This is just small small case range :) containing exactly 1 case
1843 LHS = SV; RHS = I->High; MHS = NULL;
1846 LHS = I->Low; MHS = SV; RHS = I->High;
1848 CaseBlock CB(CC, LHS, RHS, MHS, I->BB, FallThrough, CurBlock);
1850 // If emitting the first comparison, just call visitSwitchCase to emit the
1851 // code into the current block. Otherwise, push the CaseBlock onto the
1852 // vector to be later processed by SDISel, and insert the node's MBB
1853 // before the next MBB.
1854 if (CurBlock == SwitchBB)
1855 visitSwitchCase(CB, SwitchBB);
1857 SwitchCases.push_back(CB);
1859 CurBlock = FallThrough;
1865 static inline bool areJTsAllowed(const TargetLowering &TLI) {
1866 return !DisableJumpTables &&
1867 (TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
1868 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other));
1871 static APInt ComputeRange(const APInt &First, const APInt &Last) {
1872 APInt LastExt(Last), FirstExt(First);
1873 uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1;
1874 LastExt.sext(BitWidth); FirstExt.sext(BitWidth);
1875 return (LastExt - FirstExt + 1ULL);
1878 /// handleJTSwitchCase - Emit jumptable for current switch case range
1879 bool SelectionDAGBuilder::handleJTSwitchCase(CaseRec& CR,
1880 CaseRecVector& WorkList,
1882 MachineBasicBlock* Default,
1883 MachineBasicBlock *SwitchBB) {
1884 Case& FrontCase = *CR.Range.first;
1885 Case& BackCase = *(CR.Range.second-1);
1887 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
1888 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
1890 APInt TSize(First.getBitWidth(), 0);
1891 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1895 if (!areJTsAllowed(TLI) || TSize.ult(4))
1898 APInt Range = ComputeRange(First, Last);
1899 double Density = TSize.roundToDouble() / Range.roundToDouble();
1903 DEBUG(dbgs() << "Lowering jump table\n"
1904 << "First entry: " << First << ". Last entry: " << Last << '\n'
1905 << "Range: " << Range
1906 << "Size: " << TSize << ". Density: " << Density << "\n\n");
1908 // Get the MachineFunction which holds the current MBB. This is used when
1909 // inserting any additional MBBs necessary to represent the switch.
1910 MachineFunction *CurMF = FuncInfo.MF;
1912 // Figure out which block is immediately after the current one.
1913 MachineFunction::iterator BBI = CR.CaseBB;
1916 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1918 // Create a new basic block to hold the code for loading the address
1919 // of the jump table, and jumping to it. Update successor information;
1920 // we will either branch to the default case for the switch, or the jump
1922 MachineBasicBlock *JumpTableBB = CurMF->CreateMachineBasicBlock(LLVMBB);
1923 CurMF->insert(BBI, JumpTableBB);
1924 CR.CaseBB->addSuccessor(Default);
1925 CR.CaseBB->addSuccessor(JumpTableBB);
1927 // Build a vector of destination BBs, corresponding to each target
1928 // of the jump table. If the value of the jump table slot corresponds to
1929 // a case statement, push the case's BB onto the vector, otherwise, push
1931 std::vector<MachineBasicBlock*> DestBBs;
1933 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) {
1934 const APInt &Low = cast<ConstantInt>(I->Low)->getValue();
1935 const APInt &High = cast<ConstantInt>(I->High)->getValue();
1937 if (Low.sle(TEI) && TEI.sle(High)) {
1938 DestBBs.push_back(I->BB);
1942 DestBBs.push_back(Default);
1946 // Update successor info. Add one edge to each unique successor.
1947 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs());
1948 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(),
1949 E = DestBBs.end(); I != E; ++I) {
1950 if (!SuccsHandled[(*I)->getNumber()]) {
1951 SuccsHandled[(*I)->getNumber()] = true;
1952 JumpTableBB->addSuccessor(*I);
1956 // Create a jump table index for this jump table.
1957 unsigned JTEncoding = TLI.getJumpTableEncoding();
1958 unsigned JTI = CurMF->getOrCreateJumpTableInfo(JTEncoding)
1959 ->createJumpTableIndex(DestBBs);
1961 // Set the jump table information so that we can codegen it as a second
1962 // MachineBasicBlock
1963 JumpTable JT(-1U, JTI, JumpTableBB, Default);
1964 JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == SwitchBB));
1965 if (CR.CaseBB == SwitchBB)
1966 visitJumpTableHeader(JT, JTH, SwitchBB);
1968 JTCases.push_back(JumpTableBlock(JTH, JT));
1973 /// handleBTSplitSwitchCase - emit comparison and split binary search tree into
1975 bool SelectionDAGBuilder::handleBTSplitSwitchCase(CaseRec& CR,
1976 CaseRecVector& WorkList,
1978 MachineBasicBlock *Default,
1979 MachineBasicBlock *SwitchBB) {
1980 // Get the MachineFunction which holds the current MBB. This is used when
1981 // inserting any additional MBBs necessary to represent the switch.
1982 MachineFunction *CurMF = FuncInfo.MF;
1984 // Figure out which block is immediately after the current one.
1985 MachineFunction::iterator BBI = CR.CaseBB;
1988 Case& FrontCase = *CR.Range.first;
1989 Case& BackCase = *(CR.Range.second-1);
1990 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1992 // Size is the number of Cases represented by this range.
1993 unsigned Size = CR.Range.second - CR.Range.first;
1995 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
1996 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
1998 CaseItr Pivot = CR.Range.first + Size/2;
2000 // Select optimal pivot, maximizing sum density of LHS and RHS. This will
2001 // (heuristically) allow us to emit JumpTable's later.
2002 APInt TSize(First.getBitWidth(), 0);
2003 for (CaseItr I = CR.Range.first, E = CR.Range.second;
2007 APInt LSize = FrontCase.size();
2008 APInt RSize = TSize-LSize;
2009 DEBUG(dbgs() << "Selecting best pivot: \n"
2010 << "First: " << First << ", Last: " << Last <<'\n'
2011 << "LSize: " << LSize << ", RSize: " << RSize << '\n');
2012 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second;
2014 const APInt &LEnd = cast<ConstantInt>(I->High)->getValue();
2015 const APInt &RBegin = cast<ConstantInt>(J->Low)->getValue();
2016 APInt Range = ComputeRange(LEnd, RBegin);
2017 assert((Range - 2ULL).isNonNegative() &&
2018 "Invalid case distance");
2019 double LDensity = (double)LSize.roundToDouble() /
2020 (LEnd - First + 1ULL).roundToDouble();
2021 double RDensity = (double)RSize.roundToDouble() /
2022 (Last - RBegin + 1ULL).roundToDouble();
2023 double Metric = Range.logBase2()*(LDensity+RDensity);
2024 // Should always split in some non-trivial place
2025 DEBUG(dbgs() <<"=>Step\n"
2026 << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n'
2027 << "LDensity: " << LDensity
2028 << ", RDensity: " << RDensity << '\n'
2029 << "Metric: " << Metric << '\n');
2030 if (FMetric < Metric) {
2033 DEBUG(dbgs() << "Current metric set to: " << FMetric << '\n');
2039 if (areJTsAllowed(TLI)) {
2040 // If our case is dense we *really* should handle it earlier!
2041 assert((FMetric > 0) && "Should handle dense range earlier!");
2043 Pivot = CR.Range.first + Size/2;
2046 CaseRange LHSR(CR.Range.first, Pivot);
2047 CaseRange RHSR(Pivot, CR.Range.second);
2048 Constant *C = Pivot->Low;
2049 MachineBasicBlock *FalseBB = 0, *TrueBB = 0;
2051 // We know that we branch to the LHS if the Value being switched on is
2052 // less than the Pivot value, C. We use this to optimize our binary
2053 // tree a bit, by recognizing that if SV is greater than or equal to the
2054 // LHS's Case Value, and that Case Value is exactly one less than the
2055 // Pivot's Value, then we can branch directly to the LHS's Target,
2056 // rather than creating a leaf node for it.
2057 if ((LHSR.second - LHSR.first) == 1 &&
2058 LHSR.first->High == CR.GE &&
2059 cast<ConstantInt>(C)->getValue() ==
2060 (cast<ConstantInt>(CR.GE)->getValue() + 1LL)) {
2061 TrueBB = LHSR.first->BB;
2063 TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2064 CurMF->insert(BBI, TrueBB);
2065 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR));
2067 // Put SV in a virtual register to make it available from the new blocks.
2068 ExportFromCurrentBlock(SV);
2071 // Similar to the optimization above, if the Value being switched on is
2072 // known to be less than the Constant CR.LT, and the current Case Value
2073 // is CR.LT - 1, then we can branch directly to the target block for
2074 // the current Case Value, rather than emitting a RHS leaf node for it.
2075 if ((RHSR.second - RHSR.first) == 1 && CR.LT &&
2076 cast<ConstantInt>(RHSR.first->Low)->getValue() ==
2077 (cast<ConstantInt>(CR.LT)->getValue() - 1LL)) {
2078 FalseBB = RHSR.first->BB;
2080 FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2081 CurMF->insert(BBI, FalseBB);
2082 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR));
2084 // Put SV in a virtual register to make it available from the new blocks.
2085 ExportFromCurrentBlock(SV);
2088 // Create a CaseBlock record representing a conditional branch to
2089 // the LHS node if the value being switched on SV is less than C.
2090 // Otherwise, branch to LHS.
2091 CaseBlock CB(ISD::SETLT, SV, C, NULL, TrueBB, FalseBB, CR.CaseBB);
2093 if (CR.CaseBB == SwitchBB)
2094 visitSwitchCase(CB, SwitchBB);
2096 SwitchCases.push_back(CB);
2101 /// handleBitTestsSwitchCase - if current case range has few destination and
2102 /// range span less, than machine word bitwidth, encode case range into series
2103 /// of masks and emit bit tests with these masks.
2104 bool SelectionDAGBuilder::handleBitTestsSwitchCase(CaseRec& CR,
2105 CaseRecVector& WorkList,
2107 MachineBasicBlock* Default,
2108 MachineBasicBlock *SwitchBB){
2109 EVT PTy = TLI.getPointerTy();
2110 unsigned IntPtrBits = PTy.getSizeInBits();
2112 Case& FrontCase = *CR.Range.first;
2113 Case& BackCase = *(CR.Range.second-1);
2115 // Get the MachineFunction which holds the current MBB. This is used when
2116 // inserting any additional MBBs necessary to represent the switch.
2117 MachineFunction *CurMF = FuncInfo.MF;
2119 // If target does not have legal shift left, do not emit bit tests at all.
2120 if (!TLI.isOperationLegal(ISD::SHL, TLI.getPointerTy()))
2124 for (CaseItr I = CR.Range.first, E = CR.Range.second;
2126 // Single case counts one, case range - two.
2127 numCmps += (I->Low == I->High ? 1 : 2);
2130 // Count unique destinations
2131 SmallSet<MachineBasicBlock*, 4> Dests;
2132 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2133 Dests.insert(I->BB);
2134 if (Dests.size() > 3)
2135 // Don't bother the code below, if there are too much unique destinations
2138 DEBUG(dbgs() << "Total number of unique destinations: "
2139 << Dests.size() << '\n'
2140 << "Total number of comparisons: " << numCmps << '\n');
2142 // Compute span of values.
2143 const APInt& minValue = cast<ConstantInt>(FrontCase.Low)->getValue();
2144 const APInt& maxValue = cast<ConstantInt>(BackCase.High)->getValue();
2145 APInt cmpRange = maxValue - minValue;
2147 DEBUG(dbgs() << "Compare range: " << cmpRange << '\n'
2148 << "Low bound: " << minValue << '\n'
2149 << "High bound: " << maxValue << '\n');
2151 if (cmpRange.uge(IntPtrBits) ||
2152 (!(Dests.size() == 1 && numCmps >= 3) &&
2153 !(Dests.size() == 2 && numCmps >= 5) &&
2154 !(Dests.size() >= 3 && numCmps >= 6)))
2157 DEBUG(dbgs() << "Emitting bit tests\n");
2158 APInt lowBound = APInt::getNullValue(cmpRange.getBitWidth());
2160 // Optimize the case where all the case values fit in a
2161 // word without having to subtract minValue. In this case,
2162 // we can optimize away the subtraction.
2163 if (minValue.isNonNegative() && maxValue.slt(IntPtrBits)) {
2164 cmpRange = maxValue;
2166 lowBound = minValue;
2169 CaseBitsVector CasesBits;
2170 unsigned i, count = 0;
2172 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2173 MachineBasicBlock* Dest = I->BB;
2174 for (i = 0; i < count; ++i)
2175 if (Dest == CasesBits[i].BB)
2179 assert((count < 3) && "Too much destinations to test!");
2180 CasesBits.push_back(CaseBits(0, Dest, 0));
2184 const APInt& lowValue = cast<ConstantInt>(I->Low)->getValue();
2185 const APInt& highValue = cast<ConstantInt>(I->High)->getValue();
2187 uint64_t lo = (lowValue - lowBound).getZExtValue();
2188 uint64_t hi = (highValue - lowBound).getZExtValue();
2190 for (uint64_t j = lo; j <= hi; j++) {
2191 CasesBits[i].Mask |= 1ULL << j;
2192 CasesBits[i].Bits++;
2196 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp());
2200 // Figure out which block is immediately after the current one.
2201 MachineFunction::iterator BBI = CR.CaseBB;
2204 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2206 DEBUG(dbgs() << "Cases:\n");
2207 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) {
2208 DEBUG(dbgs() << "Mask: " << CasesBits[i].Mask
2209 << ", Bits: " << CasesBits[i].Bits
2210 << ", BB: " << CasesBits[i].BB << '\n');
2212 MachineBasicBlock *CaseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2213 CurMF->insert(BBI, CaseBB);
2214 BTC.push_back(BitTestCase(CasesBits[i].Mask,
2218 // Put SV in a virtual register to make it available from the new blocks.
2219 ExportFromCurrentBlock(SV);
2222 BitTestBlock BTB(lowBound, cmpRange, SV,
2223 -1U, (CR.CaseBB == SwitchBB),
2224 CR.CaseBB, Default, BTC);
2226 if (CR.CaseBB == SwitchBB)
2227 visitBitTestHeader(BTB, SwitchBB);
2229 BitTestCases.push_back(BTB);
2234 /// Clusterify - Transform simple list of Cases into list of CaseRange's
2235 size_t SelectionDAGBuilder::Clusterify(CaseVector& Cases,
2236 const SwitchInst& SI) {
2239 // Start with "simple" cases
2240 for (size_t i = 1; i < SI.getNumSuccessors(); ++i) {
2241 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SI.getSuccessor(i)];
2242 Cases.push_back(Case(SI.getSuccessorValue(i),
2243 SI.getSuccessorValue(i),
2246 std::sort(Cases.begin(), Cases.end(), CaseCmp());
2248 // Merge case into clusters
2249 if (Cases.size() >= 2)
2250 // Must recompute end() each iteration because it may be
2251 // invalidated by erase if we hold on to it
2252 for (CaseItr I = Cases.begin(), J = ++(Cases.begin()); J != Cases.end(); ) {
2253 const APInt& nextValue = cast<ConstantInt>(J->Low)->getValue();
2254 const APInt& currentValue = cast<ConstantInt>(I->High)->getValue();
2255 MachineBasicBlock* nextBB = J->BB;
2256 MachineBasicBlock* currentBB = I->BB;
2258 // If the two neighboring cases go to the same destination, merge them
2259 // into a single case.
2260 if ((nextValue - currentValue == 1) && (currentBB == nextBB)) {
2268 for (CaseItr I=Cases.begin(), E=Cases.end(); I!=E; ++I, ++numCmps) {
2269 if (I->Low != I->High)
2270 // A range counts double, since it requires two compares.
2277 void SelectionDAGBuilder::UpdateSplitBlock(MachineBasicBlock *First,
2278 MachineBasicBlock *Last) {
2280 for (unsigned i = 0, e = JTCases.size(); i != e; ++i)
2281 if (JTCases[i].first.HeaderBB == First)
2282 JTCases[i].first.HeaderBB = Last;
2284 // Update BitTestCases.
2285 for (unsigned i = 0, e = BitTestCases.size(); i != e; ++i)
2286 if (BitTestCases[i].Parent == First)
2287 BitTestCases[i].Parent = Last;
2290 void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) {
2291 MachineBasicBlock *SwitchMBB = FuncInfo.MBB;
2293 // Figure out which block is immediately after the current one.
2294 MachineBasicBlock *NextBlock = 0;
2295 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()];
2297 // If there is only the default destination, branch to it if it is not the
2298 // next basic block. Otherwise, just fall through.
2299 if (SI.getNumOperands() == 2) {
2300 // Update machine-CFG edges.
2302 // If this is not a fall-through branch, emit the branch.
2303 SwitchMBB->addSuccessor(Default);
2304 if (Default != NextBlock)
2305 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
2306 MVT::Other, getControlRoot(),
2307 DAG.getBasicBlock(Default)));
2312 // If there are any non-default case statements, create a vector of Cases
2313 // representing each one, and sort the vector so that we can efficiently
2314 // create a binary search tree from them.
2316 size_t numCmps = Clusterify(Cases, SI);
2317 DEBUG(dbgs() << "Clusterify finished. Total clusters: " << Cases.size()
2318 << ". Total compares: " << numCmps << '\n');
2321 // Get the Value to be switched on and default basic blocks, which will be
2322 // inserted into CaseBlock records, representing basic blocks in the binary
2324 const Value *SV = SI.getOperand(0);
2326 // Push the initial CaseRec onto the worklist
2327 CaseRecVector WorkList;
2328 WorkList.push_back(CaseRec(SwitchMBB,0,0,
2329 CaseRange(Cases.begin(),Cases.end())));
2331 while (!WorkList.empty()) {
2332 // Grab a record representing a case range to process off the worklist
2333 CaseRec CR = WorkList.back();
2334 WorkList.pop_back();
2336 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2339 // If the range has few cases (two or less) emit a series of specific
2341 if (handleSmallSwitchRange(CR, WorkList, SV, Default, SwitchMBB))
2344 // If the switch has more than 5 blocks, and at least 40% dense, and the
2345 // target supports indirect branches, then emit a jump table rather than
2346 // lowering the switch to a binary tree of conditional branches.
2347 if (handleJTSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2350 // Emit binary tree. We need to pick a pivot, and push left and right ranges
2351 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call.
2352 handleBTSplitSwitchCase(CR, WorkList, SV, Default, SwitchMBB);
2356 void SelectionDAGBuilder::visitIndirectBr(const IndirectBrInst &I) {
2357 MachineBasicBlock *IndirectBrMBB = FuncInfo.MBB;
2359 // Update machine-CFG edges with unique successors.
2360 SmallVector<BasicBlock*, 32> succs;
2361 succs.reserve(I.getNumSuccessors());
2362 for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i)
2363 succs.push_back(I.getSuccessor(i));
2364 array_pod_sort(succs.begin(), succs.end());
2365 succs.erase(std::unique(succs.begin(), succs.end()), succs.end());
2366 for (unsigned i = 0, e = succs.size(); i != e; ++i)
2367 IndirectBrMBB->addSuccessor(FuncInfo.MBBMap[succs[i]]);
2369 DAG.setRoot(DAG.getNode(ISD::BRIND, getCurDebugLoc(),
2370 MVT::Other, getControlRoot(),
2371 getValue(I.getAddress())));
2374 void SelectionDAGBuilder::visitFSub(const User &I) {
2375 // -0.0 - X --> fneg
2376 const Type *Ty = I.getType();
2377 if (Ty->isVectorTy()) {
2378 if (ConstantVector *CV = dyn_cast<ConstantVector>(I.getOperand(0))) {
2379 const VectorType *DestTy = cast<VectorType>(I.getType());
2380 const Type *ElTy = DestTy->getElementType();
2381 unsigned VL = DestTy->getNumElements();
2382 std::vector<Constant*> NZ(VL, ConstantFP::getNegativeZero(ElTy));
2383 Constant *CNZ = ConstantVector::get(&NZ[0], NZ.size());
2385 SDValue Op2 = getValue(I.getOperand(1));
2386 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(),
2387 Op2.getValueType(), Op2));
2393 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I.getOperand(0)))
2394 if (CFP->isExactlyValue(ConstantFP::getNegativeZero(Ty)->getValueAPF())) {
2395 SDValue Op2 = getValue(I.getOperand(1));
2396 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(),
2397 Op2.getValueType(), Op2));
2401 visitBinary(I, ISD::FSUB);
2404 void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) {
2405 SDValue Op1 = getValue(I.getOperand(0));
2406 SDValue Op2 = getValue(I.getOperand(1));
2407 setValue(&I, DAG.getNode(OpCode, getCurDebugLoc(),
2408 Op1.getValueType(), Op1, Op2));
2411 void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) {
2412 SDValue Op1 = getValue(I.getOperand(0));
2413 SDValue Op2 = getValue(I.getOperand(1));
2414 if (!I.getType()->isVectorTy() &&
2415 Op2.getValueType() != TLI.getShiftAmountTy()) {
2416 // If the operand is smaller than the shift count type, promote it.
2417 EVT PTy = TLI.getPointerTy();
2418 EVT STy = TLI.getShiftAmountTy();
2419 if (STy.bitsGT(Op2.getValueType()))
2420 Op2 = DAG.getNode(ISD::ANY_EXTEND, getCurDebugLoc(),
2421 TLI.getShiftAmountTy(), Op2);
2422 // If the operand is larger than the shift count type but the shift
2423 // count type has enough bits to represent any shift value, truncate
2424 // it now. This is a common case and it exposes the truncate to
2425 // optimization early.
2426 else if (STy.getSizeInBits() >=
2427 Log2_32_Ceil(Op2.getValueType().getSizeInBits()))
2428 Op2 = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2429 TLI.getShiftAmountTy(), Op2);
2430 // Otherwise we'll need to temporarily settle for some other
2431 // convenient type; type legalization will make adjustments as
2433 else if (PTy.bitsLT(Op2.getValueType()))
2434 Op2 = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2435 TLI.getPointerTy(), Op2);
2436 else if (PTy.bitsGT(Op2.getValueType()))
2437 Op2 = DAG.getNode(ISD::ANY_EXTEND, getCurDebugLoc(),
2438 TLI.getPointerTy(), Op2);
2441 setValue(&I, DAG.getNode(Opcode, getCurDebugLoc(),
2442 Op1.getValueType(), Op1, Op2));
2445 void SelectionDAGBuilder::visitICmp(const User &I) {
2446 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2447 if (const ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2448 predicate = IC->getPredicate();
2449 else if (const ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2450 predicate = ICmpInst::Predicate(IC->getPredicate());
2451 SDValue Op1 = getValue(I.getOperand(0));
2452 SDValue Op2 = getValue(I.getOperand(1));
2453 ISD::CondCode Opcode = getICmpCondCode(predicate);
2455 EVT DestVT = TLI.getValueType(I.getType());
2456 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Opcode));
2459 void SelectionDAGBuilder::visitFCmp(const User &I) {
2460 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2461 if (const FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2462 predicate = FC->getPredicate();
2463 else if (const ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2464 predicate = FCmpInst::Predicate(FC->getPredicate());
2465 SDValue Op1 = getValue(I.getOperand(0));
2466 SDValue Op2 = getValue(I.getOperand(1));
2467 ISD::CondCode Condition = getFCmpCondCode(predicate);
2468 EVT DestVT = TLI.getValueType(I.getType());
2469 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Condition));
2472 void SelectionDAGBuilder::visitSelect(const User &I) {
2473 SmallVector<EVT, 4> ValueVTs;
2474 ComputeValueVTs(TLI, I.getType(), ValueVTs);
2475 unsigned NumValues = ValueVTs.size();
2476 if (NumValues == 0) return;
2478 SmallVector<SDValue, 4> Values(NumValues);
2479 SDValue Cond = getValue(I.getOperand(0));
2480 SDValue TrueVal = getValue(I.getOperand(1));
2481 SDValue FalseVal = getValue(I.getOperand(2));
2483 for (unsigned i = 0; i != NumValues; ++i)
2484 Values[i] = DAG.getNode(ISD::SELECT, getCurDebugLoc(),
2485 TrueVal.getNode()->getValueType(TrueVal.getResNo()+i),
2487 SDValue(TrueVal.getNode(),
2488 TrueVal.getResNo() + i),
2489 SDValue(FalseVal.getNode(),
2490 FalseVal.getResNo() + i));
2492 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2493 DAG.getVTList(&ValueVTs[0], NumValues),
2494 &Values[0], NumValues));
2497 void SelectionDAGBuilder::visitTrunc(const User &I) {
2498 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2499 SDValue N = getValue(I.getOperand(0));
2500 EVT DestVT = TLI.getValueType(I.getType());
2501 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), DestVT, N));
2504 void SelectionDAGBuilder::visitZExt(const User &I) {
2505 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2506 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2507 SDValue N = getValue(I.getOperand(0));
2508 EVT DestVT = TLI.getValueType(I.getType());
2509 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(), DestVT, N));
2512 void SelectionDAGBuilder::visitSExt(const User &I) {
2513 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2514 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2515 SDValue N = getValue(I.getOperand(0));
2516 EVT DestVT = TLI.getValueType(I.getType());
2517 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurDebugLoc(), DestVT, N));
2520 void SelectionDAGBuilder::visitFPTrunc(const User &I) {
2521 // FPTrunc is never a no-op cast, no need to check
2522 SDValue N = getValue(I.getOperand(0));
2523 EVT DestVT = TLI.getValueType(I.getType());
2524 setValue(&I, DAG.getNode(ISD::FP_ROUND, getCurDebugLoc(),
2525 DestVT, N, DAG.getIntPtrConstant(0)));
2528 void SelectionDAGBuilder::visitFPExt(const User &I){
2529 // FPTrunc is never a no-op cast, no need to check
2530 SDValue N = getValue(I.getOperand(0));
2531 EVT DestVT = TLI.getValueType(I.getType());
2532 setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurDebugLoc(), DestVT, N));
2535 void SelectionDAGBuilder::visitFPToUI(const User &I) {
2536 // FPToUI is never a no-op cast, no need to check
2537 SDValue N = getValue(I.getOperand(0));
2538 EVT DestVT = TLI.getValueType(I.getType());
2539 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurDebugLoc(), DestVT, N));
2542 void SelectionDAGBuilder::visitFPToSI(const User &I) {
2543 // FPToSI is never a no-op cast, no need to check
2544 SDValue N = getValue(I.getOperand(0));
2545 EVT DestVT = TLI.getValueType(I.getType());
2546 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurDebugLoc(), DestVT, N));
2549 void SelectionDAGBuilder::visitUIToFP(const User &I) {
2550 // UIToFP is never a no-op cast, no need to check
2551 SDValue N = getValue(I.getOperand(0));
2552 EVT DestVT = TLI.getValueType(I.getType());
2553 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurDebugLoc(), DestVT, N));
2556 void SelectionDAGBuilder::visitSIToFP(const User &I){
2557 // SIToFP is never a no-op cast, no need to check
2558 SDValue N = getValue(I.getOperand(0));
2559 EVT DestVT = TLI.getValueType(I.getType());
2560 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurDebugLoc(), DestVT, N));
2563 void SelectionDAGBuilder::visitPtrToInt(const User &I) {
2564 // What to do depends on the size of the integer and the size of the pointer.
2565 // We can either truncate, zero extend, or no-op, accordingly.
2566 SDValue N = getValue(I.getOperand(0));
2567 EVT DestVT = TLI.getValueType(I.getType());
2568 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT));
2571 void SelectionDAGBuilder::visitIntToPtr(const User &I) {
2572 // What to do depends on the size of the integer and the size of the pointer.
2573 // We can either truncate, zero extend, or no-op, accordingly.
2574 SDValue N = getValue(I.getOperand(0));
2575 EVT DestVT = TLI.getValueType(I.getType());
2576 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT));
2579 void SelectionDAGBuilder::visitBitCast(const User &I) {
2580 SDValue N = getValue(I.getOperand(0));
2581 EVT DestVT = TLI.getValueType(I.getType());
2583 // BitCast assures us that source and destination are the same size so this is
2584 // either a BITCAST or a no-op.
2585 if (DestVT != N.getValueType())
2586 setValue(&I, DAG.getNode(ISD::BITCAST, getCurDebugLoc(),
2587 DestVT, N)); // convert types.
2589 setValue(&I, N); // noop cast.
2592 void SelectionDAGBuilder::visitInsertElement(const User &I) {
2593 SDValue InVec = getValue(I.getOperand(0));
2594 SDValue InVal = getValue(I.getOperand(1));
2595 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2597 getValue(I.getOperand(2)));
2598 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurDebugLoc(),
2599 TLI.getValueType(I.getType()),
2600 InVec, InVal, InIdx));
2603 void SelectionDAGBuilder::visitExtractElement(const User &I) {
2604 SDValue InVec = getValue(I.getOperand(0));
2605 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2607 getValue(I.getOperand(1)));
2608 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2609 TLI.getValueType(I.getType()), InVec, InIdx));
2612 // Utility for visitShuffleVector - Returns true if the mask is mask starting
2613 // from SIndx and increasing to the element length (undefs are allowed).
2614 static bool SequentialMask(SmallVectorImpl<int> &Mask, unsigned SIndx) {
2615 unsigned MaskNumElts = Mask.size();
2616 for (unsigned i = 0; i != MaskNumElts; ++i)
2617 if ((Mask[i] >= 0) && (Mask[i] != (int)(i + SIndx)))
2622 void SelectionDAGBuilder::visitShuffleVector(const User &I) {
2623 SmallVector<int, 8> Mask;
2624 SDValue Src1 = getValue(I.getOperand(0));
2625 SDValue Src2 = getValue(I.getOperand(1));
2627 // Convert the ConstantVector mask operand into an array of ints, with -1
2628 // representing undef values.
2629 SmallVector<Constant*, 8> MaskElts;
2630 cast<Constant>(I.getOperand(2))->getVectorElements(MaskElts);
2631 unsigned MaskNumElts = MaskElts.size();
2632 for (unsigned i = 0; i != MaskNumElts; ++i) {
2633 if (isa<UndefValue>(MaskElts[i]))
2636 Mask.push_back(cast<ConstantInt>(MaskElts[i])->getSExtValue());
2639 EVT VT = TLI.getValueType(I.getType());
2640 EVT SrcVT = Src1.getValueType();
2641 unsigned SrcNumElts = SrcVT.getVectorNumElements();
2643 if (SrcNumElts == MaskNumElts) {
2644 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2649 // Normalize the shuffle vector since mask and vector length don't match.
2650 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) {
2651 // Mask is longer than the source vectors and is a multiple of the source
2652 // vectors. We can use concatenate vector to make the mask and vectors
2654 if (SrcNumElts*2 == MaskNumElts && SequentialMask(Mask, 0)) {
2655 // The shuffle is concatenating two vectors together.
2656 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurDebugLoc(),
2661 // Pad both vectors with undefs to make them the same length as the mask.
2662 unsigned NumConcat = MaskNumElts / SrcNumElts;
2663 bool Src1U = Src1.getOpcode() == ISD::UNDEF;
2664 bool Src2U = Src2.getOpcode() == ISD::UNDEF;
2665 SDValue UndefVal = DAG.getUNDEF(SrcVT);
2667 SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal);
2668 SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal);
2672 Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2673 getCurDebugLoc(), VT,
2674 &MOps1[0], NumConcat);
2675 Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2676 getCurDebugLoc(), VT,
2677 &MOps2[0], NumConcat);
2679 // Readjust mask for new input vector length.
2680 SmallVector<int, 8> MappedOps;
2681 for (unsigned i = 0; i != MaskNumElts; ++i) {
2683 if (Idx < (int)SrcNumElts)
2684 MappedOps.push_back(Idx);
2686 MappedOps.push_back(Idx + MaskNumElts - SrcNumElts);
2689 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2694 if (SrcNumElts > MaskNumElts) {
2695 // Analyze the access pattern of the vector to see if we can extract
2696 // two subvectors and do the shuffle. The analysis is done by calculating
2697 // the range of elements the mask access on both vectors.
2698 int MinRange[2] = { SrcNumElts+1, SrcNumElts+1};
2699 int MaxRange[2] = {-1, -1};
2701 for (unsigned i = 0; i != MaskNumElts; ++i) {
2707 if (Idx >= (int)SrcNumElts) {
2711 if (Idx > MaxRange[Input])
2712 MaxRange[Input] = Idx;
2713 if (Idx < MinRange[Input])
2714 MinRange[Input] = Idx;
2717 // Check if the access is smaller than the vector size and can we find
2718 // a reasonable extract index.
2719 int RangeUse[2] = { 2, 2 }; // 0 = Unused, 1 = Extract, 2 = Can not
2721 int StartIdx[2]; // StartIdx to extract from
2722 for (int Input=0; Input < 2; ++Input) {
2723 if (MinRange[Input] == (int)(SrcNumElts+1) && MaxRange[Input] == -1) {
2724 RangeUse[Input] = 0; // Unused
2725 StartIdx[Input] = 0;
2726 } else if (MaxRange[Input] - MinRange[Input] < (int)MaskNumElts) {
2727 // Fits within range but we should see if we can find a good
2728 // start index that is a multiple of the mask length.
2729 if (MaxRange[Input] < (int)MaskNumElts) {
2730 RangeUse[Input] = 1; // Extract from beginning of the vector
2731 StartIdx[Input] = 0;
2733 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts;
2734 if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts &&
2735 StartIdx[Input] + MaskNumElts < SrcNumElts)
2736 RangeUse[Input] = 1; // Extract from a multiple of the mask length.
2741 if (RangeUse[0] == 0 && RangeUse[1] == 0) {
2742 setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used.
2745 else if (RangeUse[0] < 2 && RangeUse[1] < 2) {
2746 // Extract appropriate subvector and generate a vector shuffle
2747 for (int Input=0; Input < 2; ++Input) {
2748 SDValue &Src = Input == 0 ? Src1 : Src2;
2749 if (RangeUse[Input] == 0)
2750 Src = DAG.getUNDEF(VT);
2752 Src = DAG.getNode(ISD::EXTRACT_SUBVECTOR, getCurDebugLoc(), VT,
2753 Src, DAG.getIntPtrConstant(StartIdx[Input]));
2756 // Calculate new mask.
2757 SmallVector<int, 8> MappedOps;
2758 for (unsigned i = 0; i != MaskNumElts; ++i) {
2761 MappedOps.push_back(Idx);
2762 else if (Idx < (int)SrcNumElts)
2763 MappedOps.push_back(Idx - StartIdx[0]);
2765 MappedOps.push_back(Idx - SrcNumElts - StartIdx[1] + MaskNumElts);
2768 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2774 // We can't use either concat vectors or extract subvectors so fall back to
2775 // replacing the shuffle with extract and build vector.
2776 // to insert and build vector.
2777 EVT EltVT = VT.getVectorElementType();
2778 EVT PtrVT = TLI.getPointerTy();
2779 SmallVector<SDValue,8> Ops;
2780 for (unsigned i = 0; i != MaskNumElts; ++i) {
2782 Ops.push_back(DAG.getUNDEF(EltVT));
2787 if (Idx < (int)SrcNumElts)
2788 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2789 EltVT, Src1, DAG.getConstant(Idx, PtrVT));
2791 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2793 DAG.getConstant(Idx - SrcNumElts, PtrVT));
2799 setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
2800 VT, &Ops[0], Ops.size()));
2803 void SelectionDAGBuilder::visitInsertValue(const InsertValueInst &I) {
2804 const Value *Op0 = I.getOperand(0);
2805 const Value *Op1 = I.getOperand(1);
2806 const Type *AggTy = I.getType();
2807 const Type *ValTy = Op1->getType();
2808 bool IntoUndef = isa<UndefValue>(Op0);
2809 bool FromUndef = isa<UndefValue>(Op1);
2811 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.idx_begin(), I.idx_end());
2813 SmallVector<EVT, 4> AggValueVTs;
2814 ComputeValueVTs(TLI, AggTy, AggValueVTs);
2815 SmallVector<EVT, 4> ValValueVTs;
2816 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2818 unsigned NumAggValues = AggValueVTs.size();
2819 unsigned NumValValues = ValValueVTs.size();
2820 SmallVector<SDValue, 4> Values(NumAggValues);
2822 SDValue Agg = getValue(Op0);
2823 SDValue Val = getValue(Op1);
2825 // Copy the beginning value(s) from the original aggregate.
2826 for (; i != LinearIndex; ++i)
2827 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2828 SDValue(Agg.getNode(), Agg.getResNo() + i);
2829 // Copy values from the inserted value(s).
2830 for (; i != LinearIndex + NumValValues; ++i)
2831 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2832 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex);
2833 // Copy remaining value(s) from the original aggregate.
2834 for (; i != NumAggValues; ++i)
2835 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2836 SDValue(Agg.getNode(), Agg.getResNo() + i);
2838 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2839 DAG.getVTList(&AggValueVTs[0], NumAggValues),
2840 &Values[0], NumAggValues));
2843 void SelectionDAGBuilder::visitExtractValue(const ExtractValueInst &I) {
2844 const Value *Op0 = I.getOperand(0);
2845 const Type *AggTy = Op0->getType();
2846 const Type *ValTy = I.getType();
2847 bool OutOfUndef = isa<UndefValue>(Op0);
2849 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.idx_begin(), I.idx_end());
2851 SmallVector<EVT, 4> ValValueVTs;
2852 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2854 unsigned NumValValues = ValValueVTs.size();
2855 SmallVector<SDValue, 4> Values(NumValValues);
2857 SDValue Agg = getValue(Op0);
2858 // Copy out the selected value(s).
2859 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i)
2860 Values[i - LinearIndex] =
2862 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) :
2863 SDValue(Agg.getNode(), Agg.getResNo() + i);
2865 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2866 DAG.getVTList(&ValValueVTs[0], NumValValues),
2867 &Values[0], NumValValues));
2870 void SelectionDAGBuilder::visitGetElementPtr(const User &I) {
2871 SDValue N = getValue(I.getOperand(0));
2872 const Type *Ty = I.getOperand(0)->getType();
2874 for (GetElementPtrInst::const_op_iterator OI = I.op_begin()+1, E = I.op_end();
2876 const Value *Idx = *OI;
2877 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
2878 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
2881 uint64_t Offset = TD->getStructLayout(StTy)->getElementOffset(Field);
2882 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
2883 DAG.getIntPtrConstant(Offset));
2886 Ty = StTy->getElementType(Field);
2888 Ty = cast<SequentialType>(Ty)->getElementType();
2890 // If this is a constant subscript, handle it quickly.
2891 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
2892 if (CI->isZero()) continue;
2894 TD->getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
2896 EVT PTy = TLI.getPointerTy();
2897 unsigned PtrBits = PTy.getSizeInBits();
2899 OffsVal = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2901 DAG.getConstant(Offs, MVT::i64));
2903 OffsVal = DAG.getIntPtrConstant(Offs);
2905 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
2910 // N = N + Idx * ElementSize;
2911 APInt ElementSize = APInt(TLI.getPointerTy().getSizeInBits(),
2912 TD->getTypeAllocSize(Ty));
2913 SDValue IdxN = getValue(Idx);
2915 // If the index is smaller or larger than intptr_t, truncate or extend
2917 IdxN = DAG.getSExtOrTrunc(IdxN, getCurDebugLoc(), N.getValueType());
2919 // If this is a multiply by a power of two, turn it into a shl
2920 // immediately. This is a very common case.
2921 if (ElementSize != 1) {
2922 if (ElementSize.isPowerOf2()) {
2923 unsigned Amt = ElementSize.logBase2();
2924 IdxN = DAG.getNode(ISD::SHL, getCurDebugLoc(),
2925 N.getValueType(), IdxN,
2926 DAG.getConstant(Amt, TLI.getPointerTy()));
2928 SDValue Scale = DAG.getConstant(ElementSize, TLI.getPointerTy());
2929 IdxN = DAG.getNode(ISD::MUL, getCurDebugLoc(),
2930 N.getValueType(), IdxN, Scale);
2934 N = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2935 N.getValueType(), N, IdxN);
2942 void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) {
2943 // If this is a fixed sized alloca in the entry block of the function,
2944 // allocate it statically on the stack.
2945 if (FuncInfo.StaticAllocaMap.count(&I))
2946 return; // getValue will auto-populate this.
2948 const Type *Ty = I.getAllocatedType();
2949 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
2951 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty),
2954 SDValue AllocSize = getValue(I.getArraySize());
2956 EVT IntPtr = TLI.getPointerTy();
2957 if (AllocSize.getValueType() != IntPtr)
2958 AllocSize = DAG.getZExtOrTrunc(AllocSize, getCurDebugLoc(), IntPtr);
2960 AllocSize = DAG.getNode(ISD::MUL, getCurDebugLoc(), IntPtr,
2962 DAG.getConstant(TySize, IntPtr));
2964 // Handle alignment. If the requested alignment is less than or equal to
2965 // the stack alignment, ignore it. If the size is greater than or equal to
2966 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
2967 unsigned StackAlign = TM.getFrameInfo()->getStackAlignment();
2968 if (Align <= StackAlign)
2971 // Round the size of the allocation up to the stack alignment size
2972 // by add SA-1 to the size.
2973 AllocSize = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2974 AllocSize.getValueType(), AllocSize,
2975 DAG.getIntPtrConstant(StackAlign-1));
2977 // Mask out the low bits for alignment purposes.
2978 AllocSize = DAG.getNode(ISD::AND, getCurDebugLoc(),
2979 AllocSize.getValueType(), AllocSize,
2980 DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1)));
2982 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) };
2983 SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other);
2984 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, getCurDebugLoc(),
2987 DAG.setRoot(DSA.getValue(1));
2989 // Inform the Frame Information that we have just allocated a variable-sized
2991 FuncInfo.MF->getFrameInfo()->CreateVariableSizedObject(Align ? Align : 1);
2994 void SelectionDAGBuilder::visitLoad(const LoadInst &I) {
2995 const Value *SV = I.getOperand(0);
2996 SDValue Ptr = getValue(SV);
2998 const Type *Ty = I.getType();
3000 bool isVolatile = I.isVolatile();
3001 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
3002 unsigned Alignment = I.getAlignment();
3003 const MDNode *TBAAInfo = I.getMetadata(LLVMContext::MD_tbaa);
3005 SmallVector<EVT, 4> ValueVTs;
3006 SmallVector<uint64_t, 4> Offsets;
3007 ComputeValueVTs(TLI, Ty, ValueVTs, &Offsets);
3008 unsigned NumValues = ValueVTs.size();
3013 bool ConstantMemory = false;
3014 if (I.isVolatile() || NumValues > MaxParallelChains)
3015 // Serialize volatile loads with other side effects.
3017 else if (AA->pointsToConstantMemory(
3018 AliasAnalysis::Location(SV, AA->getTypeStoreSize(Ty), TBAAInfo))) {
3019 // Do not serialize (non-volatile) loads of constant memory with anything.
3020 Root = DAG.getEntryNode();
3021 ConstantMemory = true;
3023 // Do not serialize non-volatile loads against each other.
3024 Root = DAG.getRoot();
3027 SmallVector<SDValue, 4> Values(NumValues);
3028 SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains),
3030 EVT PtrVT = Ptr.getValueType();
3031 unsigned ChainI = 0;
3032 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3033 // Serializing loads here may result in excessive register pressure, and
3034 // TokenFactor places arbitrary choke points on the scheduler. SD scheduling
3035 // could recover a bit by hoisting nodes upward in the chain by recognizing
3036 // they are side-effect free or do not alias. The optimizer should really
3037 // avoid this case by converting large object/array copies to llvm.memcpy
3038 // (MaxParallelChains should always remain as failsafe).
3039 if (ChainI == MaxParallelChains) {
3040 assert(PendingLoads.empty() && "PendingLoads must be serialized first");
3041 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3042 MVT::Other, &Chains[0], ChainI);
3046 SDValue A = DAG.getNode(ISD::ADD, getCurDebugLoc(),
3048 DAG.getConstant(Offsets[i], PtrVT));
3049 SDValue L = DAG.getLoad(ValueVTs[i], getCurDebugLoc(), Root,
3050 A, MachinePointerInfo(SV, Offsets[i]), isVolatile,
3051 isNonTemporal, Alignment, TBAAInfo);
3054 Chains[ChainI] = L.getValue(1);
3057 if (!ConstantMemory) {
3058 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3059 MVT::Other, &Chains[0], ChainI);
3063 PendingLoads.push_back(Chain);
3066 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
3067 DAG.getVTList(&ValueVTs[0], NumValues),
3068 &Values[0], NumValues));
3071 void SelectionDAGBuilder::visitStore(const StoreInst &I) {
3072 const Value *SrcV = I.getOperand(0);
3073 const Value *PtrV = I.getOperand(1);
3075 SmallVector<EVT, 4> ValueVTs;
3076 SmallVector<uint64_t, 4> Offsets;
3077 ComputeValueVTs(TLI, SrcV->getType(), ValueVTs, &Offsets);
3078 unsigned NumValues = ValueVTs.size();
3082 // Get the lowered operands. Note that we do this after
3083 // checking if NumResults is zero, because with zero results
3084 // the operands won't have values in the map.
3085 SDValue Src = getValue(SrcV);
3086 SDValue Ptr = getValue(PtrV);
3088 SDValue Root = getRoot();
3089 SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains),
3091 EVT PtrVT = Ptr.getValueType();
3092 bool isVolatile = I.isVolatile();
3093 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
3094 unsigned Alignment = I.getAlignment();
3095 const MDNode *TBAAInfo = I.getMetadata(LLVMContext::MD_tbaa);
3097 unsigned ChainI = 0;
3098 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3099 // See visitLoad comments.
3100 if (ChainI == MaxParallelChains) {
3101 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3102 MVT::Other, &Chains[0], ChainI);
3106 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT, Ptr,
3107 DAG.getConstant(Offsets[i], PtrVT));
3108 SDValue St = DAG.getStore(Root, getCurDebugLoc(),
3109 SDValue(Src.getNode(), Src.getResNo() + i),
3110 Add, MachinePointerInfo(PtrV, Offsets[i]),
3111 isVolatile, isNonTemporal, Alignment, TBAAInfo);
3112 Chains[ChainI] = St;
3115 SDValue StoreNode = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3116 MVT::Other, &Chains[0], ChainI);
3118 AssignOrderingToNode(StoreNode.getNode());
3119 DAG.setRoot(StoreNode);
3122 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
3124 void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I,
3125 unsigned Intrinsic) {
3126 bool HasChain = !I.doesNotAccessMemory();
3127 bool OnlyLoad = HasChain && I.onlyReadsMemory();
3129 // Build the operand list.
3130 SmallVector<SDValue, 8> Ops;
3131 if (HasChain) { // If this intrinsic has side-effects, chainify it.
3133 // We don't need to serialize loads against other loads.
3134 Ops.push_back(DAG.getRoot());
3136 Ops.push_back(getRoot());
3140 // Info is set by getTgtMemInstrinsic
3141 TargetLowering::IntrinsicInfo Info;
3142 bool IsTgtIntrinsic = TLI.getTgtMemIntrinsic(Info, I, Intrinsic);
3144 // Add the intrinsic ID as an integer operand if it's not a target intrinsic.
3145 if (!IsTgtIntrinsic || Info.opc == ISD::INTRINSIC_VOID ||
3146 Info.opc == ISD::INTRINSIC_W_CHAIN)
3147 Ops.push_back(DAG.getConstant(Intrinsic, TLI.getPointerTy()));
3149 // Add all operands of the call to the operand list.
3150 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) {
3151 SDValue Op = getValue(I.getArgOperand(i));
3152 assert(TLI.isTypeLegal(Op.getValueType()) &&
3153 "Intrinsic uses a non-legal type?");
3157 SmallVector<EVT, 4> ValueVTs;
3158 ComputeValueVTs(TLI, I.getType(), ValueVTs);
3160 for (unsigned Val = 0, E = ValueVTs.size(); Val != E; ++Val) {
3161 assert(TLI.isTypeLegal(ValueVTs[Val]) &&
3162 "Intrinsic uses a non-legal type?");
3167 ValueVTs.push_back(MVT::Other);
3169 SDVTList VTs = DAG.getVTList(ValueVTs.data(), ValueVTs.size());
3173 if (IsTgtIntrinsic) {
3174 // This is target intrinsic that touches memory
3175 Result = DAG.getMemIntrinsicNode(Info.opc, getCurDebugLoc(),
3176 VTs, &Ops[0], Ops.size(),
3178 MachinePointerInfo(Info.ptrVal, Info.offset),
3179 Info.align, Info.vol,
3180 Info.readMem, Info.writeMem);
3181 } else if (!HasChain) {
3182 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurDebugLoc(),
3183 VTs, &Ops[0], Ops.size());
3184 } else if (!I.getType()->isVoidTy()) {
3185 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurDebugLoc(),
3186 VTs, &Ops[0], Ops.size());
3188 Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurDebugLoc(),
3189 VTs, &Ops[0], Ops.size());
3193 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1);
3195 PendingLoads.push_back(Chain);
3200 if (!I.getType()->isVoidTy()) {
3201 if (const VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
3202 EVT VT = TLI.getValueType(PTy);
3203 Result = DAG.getNode(ISD::BITCAST, getCurDebugLoc(), VT, Result);
3206 setValue(&I, Result);
3210 /// GetSignificand - Get the significand and build it into a floating-point
3211 /// number with exponent of 1:
3213 /// Op = (Op & 0x007fffff) | 0x3f800000;
3215 /// where Op is the hexidecimal representation of floating point value.
3217 GetSignificand(SelectionDAG &DAG, SDValue Op, DebugLoc dl) {
3218 SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3219 DAG.getConstant(0x007fffff, MVT::i32));
3220 SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1,
3221 DAG.getConstant(0x3f800000, MVT::i32));
3222 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, t2);
3225 /// GetExponent - Get the exponent:
3227 /// (float)(int)(((Op & 0x7f800000) >> 23) - 127);
3229 /// where Op is the hexidecimal representation of floating point value.
3231 GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI,
3233 SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3234 DAG.getConstant(0x7f800000, MVT::i32));
3235 SDValue t1 = DAG.getNode(ISD::SRL, dl, MVT::i32, t0,
3236 DAG.getConstant(23, TLI.getPointerTy()));
3237 SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1,
3238 DAG.getConstant(127, MVT::i32));
3239 return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2);
3242 /// getF32Constant - Get 32-bit floating point constant.
3244 getF32Constant(SelectionDAG &DAG, unsigned Flt) {
3245 return DAG.getConstantFP(APFloat(APInt(32, Flt)), MVT::f32);
3248 /// Inlined utility function to implement binary input atomic intrinsics for
3249 /// visitIntrinsicCall: I is a call instruction
3250 /// Op is the associated NodeType for I
3252 SelectionDAGBuilder::implVisitBinaryAtomic(const CallInst& I,
3254 SDValue Root = getRoot();
3256 DAG.getAtomic(Op, getCurDebugLoc(),
3257 getValue(I.getArgOperand(1)).getValueType().getSimpleVT(),
3259 getValue(I.getArgOperand(0)),
3260 getValue(I.getArgOperand(1)),
3261 I.getArgOperand(0));
3263 DAG.setRoot(L.getValue(1));
3267 // implVisitAluOverflow - Lower arithmetic overflow instrinsics.
3269 SelectionDAGBuilder::implVisitAluOverflow(const CallInst &I, ISD::NodeType Op) {
3270 SDValue Op1 = getValue(I.getArgOperand(0));
3271 SDValue Op2 = getValue(I.getArgOperand(1));
3273 SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1);
3274 setValue(&I, DAG.getNode(Op, getCurDebugLoc(), VTs, Op1, Op2));
3278 /// visitExp - Lower an exp intrinsic. Handles the special sequences for
3279 /// limited-precision mode.
3281 SelectionDAGBuilder::visitExp(const CallInst &I) {
3283 DebugLoc dl = getCurDebugLoc();
3285 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3286 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3287 SDValue Op = getValue(I.getArgOperand(0));
3289 // Put the exponent in the right bit position for later addition to the
3292 // #define LOG2OFe 1.4426950f
3293 // IntegerPartOfX = ((int32_t)(X * LOG2OFe));
3294 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3295 getF32Constant(DAG, 0x3fb8aa3b));
3296 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3298 // FractionalPartOfX = (X * LOG2OFe) - (float)IntegerPartOfX;
3299 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3300 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3302 // IntegerPartOfX <<= 23;
3303 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3304 DAG.getConstant(23, TLI.getPointerTy()));
3306 if (LimitFloatPrecision <= 6) {
3307 // For floating-point precision of 6:
3309 // TwoToFractionalPartOfX =
3311 // (0.735607626f + 0.252464424f * x) * x;
3313 // error 0.0144103317, which is 6 bits
3314 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3315 getF32Constant(DAG, 0x3e814304));
3316 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3317 getF32Constant(DAG, 0x3f3c50c8));
3318 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3319 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3320 getF32Constant(DAG, 0x3f7f5e7e));
3321 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BITCAST, dl,MVT::i32, t5);
3323 // Add the exponent into the result in integer domain.
3324 SDValue t6 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3325 TwoToFracPartOfX, IntegerPartOfX);
3327 result = DAG.getNode(ISD::BITCAST, dl, MVT::f32, t6);
3328 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3329 // For floating-point precision of 12:
3331 // TwoToFractionalPartOfX =
3334 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3336 // 0.000107046256 error, which is 13 to 14 bits
3337 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3338 getF32Constant(DAG, 0x3da235e3));
3339 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3340 getF32Constant(DAG, 0x3e65b8f3));
3341 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3342 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3343 getF32Constant(DAG, 0x3f324b07));
3344 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3345 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3346 getF32Constant(DAG, 0x3f7ff8fd));
3347 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BITCAST, dl,MVT::i32, t7);
3349 // Add the exponent into the result in integer domain.
3350 SDValue t8 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3351 TwoToFracPartOfX, IntegerPartOfX);
3353 result = DAG.getNode(ISD::BITCAST, dl, MVT::f32, t8);
3354 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3355 // For floating-point precision of 18:
3357 // TwoToFractionalPartOfX =
3361 // (0.554906021e-1f +
3362 // (0.961591928e-2f +
3363 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3365 // error 2.47208000*10^(-7), which is better than 18 bits
3366 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3367 getF32Constant(DAG, 0x3924b03e));
3368 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3369 getF32Constant(DAG, 0x3ab24b87));
3370 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3371 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3372 getF32Constant(DAG, 0x3c1d8c17));
3373 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3374 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3375 getF32Constant(DAG, 0x3d634a1d));
3376 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3377 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3378 getF32Constant(DAG, 0x3e75fe14));
3379 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3380 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3381 getF32Constant(DAG, 0x3f317234));
3382 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3383 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3384 getF32Constant(DAG, 0x3f800000));
3385 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BITCAST, dl,
3388 // Add the exponent into the result in integer domain.
3389 SDValue t14 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3390 TwoToFracPartOfX, IntegerPartOfX);
3392 result = DAG.getNode(ISD::BITCAST, dl, MVT::f32, t14);
3395 // No special expansion.
3396 result = DAG.getNode(ISD::FEXP, dl,
3397 getValue(I.getArgOperand(0)).getValueType(),
3398 getValue(I.getArgOperand(0)));
3401 setValue(&I, result);
3404 /// visitLog - Lower a log intrinsic. Handles the special sequences for
3405 /// limited-precision mode.
3407 SelectionDAGBuilder::visitLog(const CallInst &I) {
3409 DebugLoc dl = getCurDebugLoc();
3411 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3412 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3413 SDValue Op = getValue(I.getArgOperand(0));
3414 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3416 // Scale the exponent by log(2) [0.69314718f].
3417 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3418 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3419 getF32Constant(DAG, 0x3f317218));
3421 // Get the significand and build it into a floating-point number with
3423 SDValue X = GetSignificand(DAG, Op1, dl);
3425 if (LimitFloatPrecision <= 6) {
3426 // For floating-point precision of 6:
3430 // (1.4034025f - 0.23903021f * x) * x;
3432 // error 0.0034276066, which is better than 8 bits
3433 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3434 getF32Constant(DAG, 0xbe74c456));
3435 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3436 getF32Constant(DAG, 0x3fb3a2b1));
3437 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3438 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3439 getF32Constant(DAG, 0x3f949a29));
3441 result = DAG.getNode(ISD::FADD, dl,
3442 MVT::f32, LogOfExponent, LogOfMantissa);
3443 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3444 // For floating-point precision of 12:
3450 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x;
3452 // error 0.000061011436, which is 14 bits
3453 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3454 getF32Constant(DAG, 0xbd67b6d6));
3455 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3456 getF32Constant(DAG, 0x3ee4f4b8));
3457 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3458 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3459 getF32Constant(DAG, 0x3fbc278b));
3460 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3461 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3462 getF32Constant(DAG, 0x40348e95));
3463 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3464 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3465 getF32Constant(DAG, 0x3fdef31a));
3467 result = DAG.getNode(ISD::FADD, dl,
3468 MVT::f32, LogOfExponent, LogOfMantissa);
3469 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3470 // For floating-point precision of 18:
3478 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x;
3480 // error 0.0000023660568, which is better than 18 bits
3481 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3482 getF32Constant(DAG, 0xbc91e5ac));
3483 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3484 getF32Constant(DAG, 0x3e4350aa));
3485 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3486 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3487 getF32Constant(DAG, 0x3f60d3e3));
3488 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3489 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3490 getF32Constant(DAG, 0x4011cdf0));
3491 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3492 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3493 getF32Constant(DAG, 0x406cfd1c));
3494 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3495 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3496 getF32Constant(DAG, 0x408797cb));
3497 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3498 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3499 getF32Constant(DAG, 0x4006dcab));
3501 result = DAG.getNode(ISD::FADD, dl,
3502 MVT::f32, LogOfExponent, LogOfMantissa);
3505 // No special expansion.
3506 result = DAG.getNode(ISD::FLOG, dl,
3507 getValue(I.getArgOperand(0)).getValueType(),
3508 getValue(I.getArgOperand(0)));
3511 setValue(&I, result);
3514 /// visitLog2 - Lower a log2 intrinsic. Handles the special sequences for
3515 /// limited-precision mode.
3517 SelectionDAGBuilder::visitLog2(const CallInst &I) {
3519 DebugLoc dl = getCurDebugLoc();
3521 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3522 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3523 SDValue Op = getValue(I.getArgOperand(0));
3524 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3526 // Get the exponent.
3527 SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl);
3529 // Get the significand and build it into a floating-point number with
3531 SDValue X = GetSignificand(DAG, Op1, dl);
3533 // Different possible minimax approximations of significand in
3534 // floating-point for various degrees of accuracy over [1,2].
3535 if (LimitFloatPrecision <= 6) {
3536 // For floating-point precision of 6:
3538 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x;
3540 // error 0.0049451742, which is more than 7 bits
3541 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3542 getF32Constant(DAG, 0xbeb08fe0));
3543 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3544 getF32Constant(DAG, 0x40019463));
3545 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3546 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3547 getF32Constant(DAG, 0x3fd6633d));
3549 result = DAG.getNode(ISD::FADD, dl,
3550 MVT::f32, LogOfExponent, Log2ofMantissa);
3551 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3552 // For floating-point precision of 12:
3558 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x;
3560 // error 0.0000876136000, which is better than 13 bits
3561 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3562 getF32Constant(DAG, 0xbda7262e));
3563 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3564 getF32Constant(DAG, 0x3f25280b));
3565 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3566 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3567 getF32Constant(DAG, 0x4007b923));
3568 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3569 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3570 getF32Constant(DAG, 0x40823e2f));
3571 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3572 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3573 getF32Constant(DAG, 0x4020d29c));
3575 result = DAG.getNode(ISD::FADD, dl,
3576 MVT::f32, LogOfExponent, Log2ofMantissa);
3577 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3578 // For floating-point precision of 18:
3587 // 0.25691327e-1f * x) * x) * x) * x) * x) * x;
3589 // error 0.0000018516, which is better than 18 bits
3590 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3591 getF32Constant(DAG, 0xbcd2769e));
3592 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3593 getF32Constant(DAG, 0x3e8ce0b9));
3594 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3595 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3596 getF32Constant(DAG, 0x3fa22ae7));
3597 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3598 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3599 getF32Constant(DAG, 0x40525723));
3600 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3601 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3602 getF32Constant(DAG, 0x40aaf200));
3603 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3604 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3605 getF32Constant(DAG, 0x40c39dad));
3606 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3607 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3608 getF32Constant(DAG, 0x4042902c));
3610 result = DAG.getNode(ISD::FADD, dl,
3611 MVT::f32, LogOfExponent, Log2ofMantissa);
3614 // No special expansion.
3615 result = DAG.getNode(ISD::FLOG2, dl,
3616 getValue(I.getArgOperand(0)).getValueType(),
3617 getValue(I.getArgOperand(0)));
3620 setValue(&I, result);
3623 /// visitLog10 - Lower a log10 intrinsic. Handles the special sequences for
3624 /// limited-precision mode.
3626 SelectionDAGBuilder::visitLog10(const CallInst &I) {
3628 DebugLoc dl = getCurDebugLoc();
3630 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3631 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3632 SDValue Op = getValue(I.getArgOperand(0));
3633 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3635 // Scale the exponent by log10(2) [0.30102999f].
3636 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3637 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3638 getF32Constant(DAG, 0x3e9a209a));
3640 // Get the significand and build it into a floating-point number with
3642 SDValue X = GetSignificand(DAG, Op1, dl);
3644 if (LimitFloatPrecision <= 6) {
3645 // For floating-point precision of 6:
3647 // Log10ofMantissa =
3649 // (0.60948995f - 0.10380950f * x) * x;
3651 // error 0.0014886165, which is 6 bits
3652 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3653 getF32Constant(DAG, 0xbdd49a13));
3654 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3655 getF32Constant(DAG, 0x3f1c0789));
3656 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3657 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3658 getF32Constant(DAG, 0x3f011300));
3660 result = DAG.getNode(ISD::FADD, dl,
3661 MVT::f32, LogOfExponent, Log10ofMantissa);
3662 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3663 // For floating-point precision of 12:
3665 // Log10ofMantissa =
3668 // (-0.31664806f + 0.47637168e-1f * x) * x) * x;
3670 // error 0.00019228036, which is better than 12 bits
3671 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3672 getF32Constant(DAG, 0x3d431f31));
3673 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
3674 getF32Constant(DAG, 0x3ea21fb2));
3675 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3676 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3677 getF32Constant(DAG, 0x3f6ae232));
3678 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3679 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
3680 getF32Constant(DAG, 0x3f25f7c3));
3682 result = DAG.getNode(ISD::FADD, dl,
3683 MVT::f32, LogOfExponent, Log10ofMantissa);
3684 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3685 // For floating-point precision of 18:
3687 // Log10ofMantissa =
3692 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x;
3694 // error 0.0000037995730, which is better than 18 bits
3695 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3696 getF32Constant(DAG, 0x3c5d51ce));
3697 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
3698 getF32Constant(DAG, 0x3e00685a));
3699 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3700 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3701 getF32Constant(DAG, 0x3efb6798));
3702 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3703 SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
3704 getF32Constant(DAG, 0x3f88d192));
3705 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3706 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3707 getF32Constant(DAG, 0x3fc4316c));
3708 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3709 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8,
3710 getF32Constant(DAG, 0x3f57ce70));
3712 result = DAG.getNode(ISD::FADD, dl,
3713 MVT::f32, LogOfExponent, Log10ofMantissa);
3716 // No special expansion.
3717 result = DAG.getNode(ISD::FLOG10, dl,
3718 getValue(I.getArgOperand(0)).getValueType(),
3719 getValue(I.getArgOperand(0)));
3722 setValue(&I, result);
3725 /// visitExp2 - Lower an exp2 intrinsic. Handles the special sequences for
3726 /// limited-precision mode.
3728 SelectionDAGBuilder::visitExp2(const CallInst &I) {
3730 DebugLoc dl = getCurDebugLoc();
3732 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3733 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3734 SDValue Op = getValue(I.getArgOperand(0));
3736 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Op);
3738 // FractionalPartOfX = x - (float)IntegerPartOfX;
3739 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3740 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, Op, t1);
3742 // IntegerPartOfX <<= 23;
3743 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3744 DAG.getConstant(23, TLI.getPointerTy()));
3746 if (LimitFloatPrecision <= 6) {
3747 // For floating-point precision of 6:
3749 // TwoToFractionalPartOfX =
3751 // (0.735607626f + 0.252464424f * x) * x;
3753 // error 0.0144103317, which is 6 bits
3754 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3755 getF32Constant(DAG, 0x3e814304));
3756 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3757 getF32Constant(DAG, 0x3f3c50c8));
3758 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3759 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3760 getF32Constant(DAG, 0x3f7f5e7e));
3761 SDValue t6 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t5);
3762 SDValue TwoToFractionalPartOfX =
3763 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
3765 result = DAG.getNode(ISD::BITCAST, dl,
3766 MVT::f32, TwoToFractionalPartOfX);
3767 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3768 // For floating-point precision of 12:
3770 // TwoToFractionalPartOfX =
3773 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3775 // error 0.000107046256, which is 13 to 14 bits
3776 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3777 getF32Constant(DAG, 0x3da235e3));
3778 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3779 getF32Constant(DAG, 0x3e65b8f3));
3780 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3781 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3782 getF32Constant(DAG, 0x3f324b07));
3783 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3784 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3785 getF32Constant(DAG, 0x3f7ff8fd));
3786 SDValue t8 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t7);
3787 SDValue TwoToFractionalPartOfX =
3788 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
3790 result = DAG.getNode(ISD::BITCAST, dl,
3791 MVT::f32, TwoToFractionalPartOfX);
3792 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3793 // For floating-point precision of 18:
3795 // TwoToFractionalPartOfX =
3799 // (0.554906021e-1f +
3800 // (0.961591928e-2f +
3801 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3802 // error 2.47208000*10^(-7), which is better than 18 bits
3803 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3804 getF32Constant(DAG, 0x3924b03e));
3805 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3806 getF32Constant(DAG, 0x3ab24b87));
3807 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3808 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3809 getF32Constant(DAG, 0x3c1d8c17));
3810 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3811 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3812 getF32Constant(DAG, 0x3d634a1d));
3813 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3814 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3815 getF32Constant(DAG, 0x3e75fe14));
3816 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3817 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3818 getF32Constant(DAG, 0x3f317234));
3819 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3820 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3821 getF32Constant(DAG, 0x3f800000));
3822 SDValue t14 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t13);
3823 SDValue TwoToFractionalPartOfX =
3824 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
3826 result = DAG.getNode(ISD::BITCAST, dl,
3827 MVT::f32, TwoToFractionalPartOfX);
3830 // No special expansion.
3831 result = DAG.getNode(ISD::FEXP2, dl,
3832 getValue(I.getArgOperand(0)).getValueType(),
3833 getValue(I.getArgOperand(0)));
3836 setValue(&I, result);
3839 /// visitPow - Lower a pow intrinsic. Handles the special sequences for
3840 /// limited-precision mode with x == 10.0f.
3842 SelectionDAGBuilder::visitPow(const CallInst &I) {
3844 const Value *Val = I.getArgOperand(0);
3845 DebugLoc dl = getCurDebugLoc();
3846 bool IsExp10 = false;
3848 if (getValue(Val).getValueType() == MVT::f32 &&
3849 getValue(I.getArgOperand(1)).getValueType() == MVT::f32 &&
3850 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3851 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(Val))) {
3852 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
3854 IsExp10 = CFP->getValueAPF().bitwiseIsEqual(Ten);
3859 if (IsExp10 && LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3860 SDValue Op = getValue(I.getArgOperand(1));
3862 // Put the exponent in the right bit position for later addition to the
3865 // #define LOG2OF10 3.3219281f
3866 // IntegerPartOfX = (int32_t)(x * LOG2OF10);
3867 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3868 getF32Constant(DAG, 0x40549a78));
3869 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3871 // FractionalPartOfX = x - (float)IntegerPartOfX;
3872 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3873 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3875 // IntegerPartOfX <<= 23;
3876 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3877 DAG.getConstant(23, TLI.getPointerTy()));
3879 if (LimitFloatPrecision <= 6) {
3880 // For floating-point precision of 6:
3882 // twoToFractionalPartOfX =
3884 // (0.735607626f + 0.252464424f * x) * x;
3886 // error 0.0144103317, which is 6 bits
3887 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3888 getF32Constant(DAG, 0x3e814304));
3889 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3890 getF32Constant(DAG, 0x3f3c50c8));
3891 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3892 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3893 getF32Constant(DAG, 0x3f7f5e7e));
3894 SDValue t6 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t5);
3895 SDValue TwoToFractionalPartOfX =
3896 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
3898 result = DAG.getNode(ISD::BITCAST, dl,
3899 MVT::f32, TwoToFractionalPartOfX);
3900 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3901 // For floating-point precision of 12:
3903 // TwoToFractionalPartOfX =
3906 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3908 // error 0.000107046256, which is 13 to 14 bits
3909 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3910 getF32Constant(DAG, 0x3da235e3));
3911 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3912 getF32Constant(DAG, 0x3e65b8f3));
3913 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3914 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3915 getF32Constant(DAG, 0x3f324b07));
3916 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3917 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3918 getF32Constant(DAG, 0x3f7ff8fd));
3919 SDValue t8 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t7);
3920 SDValue TwoToFractionalPartOfX =
3921 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
3923 result = DAG.getNode(ISD::BITCAST, dl,
3924 MVT::f32, TwoToFractionalPartOfX);
3925 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3926 // For floating-point precision of 18:
3928 // TwoToFractionalPartOfX =
3932 // (0.554906021e-1f +
3933 // (0.961591928e-2f +
3934 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3935 // error 2.47208000*10^(-7), which is better than 18 bits
3936 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3937 getF32Constant(DAG, 0x3924b03e));
3938 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3939 getF32Constant(DAG, 0x3ab24b87));
3940 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3941 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3942 getF32Constant(DAG, 0x3c1d8c17));
3943 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3944 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3945 getF32Constant(DAG, 0x3d634a1d));
3946 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3947 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3948 getF32Constant(DAG, 0x3e75fe14));
3949 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3950 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3951 getF32Constant(DAG, 0x3f317234));
3952 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3953 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3954 getF32Constant(DAG, 0x3f800000));
3955 SDValue t14 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t13);
3956 SDValue TwoToFractionalPartOfX =
3957 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
3959 result = DAG.getNode(ISD::BITCAST, dl,
3960 MVT::f32, TwoToFractionalPartOfX);
3963 // No special expansion.
3964 result = DAG.getNode(ISD::FPOW, dl,
3965 getValue(I.getArgOperand(0)).getValueType(),
3966 getValue(I.getArgOperand(0)),
3967 getValue(I.getArgOperand(1)));
3970 setValue(&I, result);
3974 /// ExpandPowI - Expand a llvm.powi intrinsic.
3975 static SDValue ExpandPowI(DebugLoc DL, SDValue LHS, SDValue RHS,
3976 SelectionDAG &DAG) {
3977 // If RHS is a constant, we can expand this out to a multiplication tree,
3978 // otherwise we end up lowering to a call to __powidf2 (for example). When
3979 // optimizing for size, we only want to do this if the expansion would produce
3980 // a small number of multiplies, otherwise we do the full expansion.
3981 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3982 // Get the exponent as a positive value.
3983 unsigned Val = RHSC->getSExtValue();
3984 if ((int)Val < 0) Val = -Val;
3986 // powi(x, 0) -> 1.0
3988 return DAG.getConstantFP(1.0, LHS.getValueType());
3990 const Function *F = DAG.getMachineFunction().getFunction();
3991 if (!F->hasFnAttr(Attribute::OptimizeForSize) ||
3992 // If optimizing for size, don't insert too many multiplies. This
3993 // inserts up to 5 multiplies.
3994 CountPopulation_32(Val)+Log2_32(Val) < 7) {
3995 // We use the simple binary decomposition method to generate the multiply
3996 // sequence. There are more optimal ways to do this (for example,
3997 // powi(x,15) generates one more multiply than it should), but this has
3998 // the benefit of being both really simple and much better than a libcall.
3999 SDValue Res; // Logically starts equal to 1.0
4000 SDValue CurSquare = LHS;
4004 Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare);
4006 Res = CurSquare; // 1.0*CurSquare.
4009 CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(),
4010 CurSquare, CurSquare);
4014 // If the original was negative, invert the result, producing 1/(x*x*x).
4015 if (RHSC->getSExtValue() < 0)
4016 Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(),
4017 DAG.getConstantFP(1.0, LHS.getValueType()), Res);
4022 // Otherwise, expand to a libcall.
4023 return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS);
4026 /// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a function
4027 /// argument, create the corresponding DBG_VALUE machine instruction for it now.
4028 /// At the end of instruction selection, they will be inserted to the entry BB.
4030 SelectionDAGBuilder::EmitFuncArgumentDbgValue(const Value *V, MDNode *Variable,
4033 const Argument *Arg = dyn_cast<Argument>(V);
4037 MachineFunction &MF = DAG.getMachineFunction();
4038 const TargetInstrInfo *TII = DAG.getTarget().getInstrInfo();
4039 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
4041 // Ignore inlined function arguments here.
4042 DIVariable DV(Variable);
4043 if (DV.isInlinedFnArgument(MF.getFunction()))
4046 MachineBasicBlock *MBB = FuncInfo.MBB;
4047 if (MBB != &MF.front())
4051 if (Arg->hasByValAttr()) {
4052 // Byval arguments' frame index is recorded during argument lowering.
4053 // Use this info directly.
4054 Reg = TRI->getFrameRegister(MF);
4055 Offset = FuncInfo.getByValArgumentFrameIndex(Arg);
4056 // If byval argument ofset is not recorded then ignore this.
4061 if (N.getNode() && N.getOpcode() == ISD::CopyFromReg) {
4062 Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg();
4063 if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) {
4064 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4065 unsigned PR = RegInfo.getLiveInPhysReg(Reg);
4072 // Check if ValueMap has reg number.
4073 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
4074 if (VMI != FuncInfo.ValueMap.end())
4078 if (!Reg && N.getNode()) {
4079 // Check if frame index is available.
4080 if (LoadSDNode *LNode = dyn_cast<LoadSDNode>(N.getNode()))
4081 if (FrameIndexSDNode *FINode =
4082 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode())) {
4083 Reg = TRI->getFrameRegister(MF);
4084 Offset = FINode->getIndex();
4091 MachineInstrBuilder MIB = BuildMI(MF, getCurDebugLoc(),
4092 TII->get(TargetOpcode::DBG_VALUE))
4093 .addReg(Reg, RegState::Debug).addImm(Offset).addMetadata(Variable);
4094 FuncInfo.ArgDbgValues.push_back(&*MIB);
4098 // VisualStudio defines setjmp as _setjmp
4099 #if defined(_MSC_VER) && defined(setjmp) && \
4100 !defined(setjmp_undefined_for_msvc)
4101 # pragma push_macro("setjmp")
4103 # define setjmp_undefined_for_msvc
4106 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
4107 /// we want to emit this as a call to a named external function, return the name
4108 /// otherwise lower it and return null.
4110 SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) {
4111 DebugLoc dl = getCurDebugLoc();
4114 switch (Intrinsic) {
4116 // By default, turn this into a target intrinsic node.
4117 visitTargetIntrinsic(I, Intrinsic);
4119 case Intrinsic::vastart: visitVAStart(I); return 0;
4120 case Intrinsic::vaend: visitVAEnd(I); return 0;
4121 case Intrinsic::vacopy: visitVACopy(I); return 0;
4122 case Intrinsic::returnaddress:
4123 setValue(&I, DAG.getNode(ISD::RETURNADDR, dl, TLI.getPointerTy(),
4124 getValue(I.getArgOperand(0))));
4126 case Intrinsic::frameaddress:
4127 setValue(&I, DAG.getNode(ISD::FRAMEADDR, dl, TLI.getPointerTy(),
4128 getValue(I.getArgOperand(0))));
4130 case Intrinsic::setjmp:
4131 return "_setjmp"+!TLI.usesUnderscoreSetJmp();
4132 case Intrinsic::longjmp:
4133 return "_longjmp"+!TLI.usesUnderscoreLongJmp();
4134 case Intrinsic::memcpy: {
4135 // Assert for address < 256 since we support only user defined address
4137 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4139 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4141 "Unknown address space");
4142 SDValue Op1 = getValue(I.getArgOperand(0));
4143 SDValue Op2 = getValue(I.getArgOperand(1));
4144 SDValue Op3 = getValue(I.getArgOperand(2));
4145 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4146 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4147 DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, isVol, false,
4148 MachinePointerInfo(I.getArgOperand(0)),
4149 MachinePointerInfo(I.getArgOperand(1))));
4152 case Intrinsic::memset: {
4153 // Assert for address < 256 since we support only user defined address
4155 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4157 "Unknown address space");
4158 SDValue Op1 = getValue(I.getArgOperand(0));
4159 SDValue Op2 = getValue(I.getArgOperand(1));
4160 SDValue Op3 = getValue(I.getArgOperand(2));
4161 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4162 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4163 DAG.setRoot(DAG.getMemset(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4164 MachinePointerInfo(I.getArgOperand(0))));
4167 case Intrinsic::memmove: {
4168 // Assert for address < 256 since we support only user defined address
4170 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4172 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4174 "Unknown address space");
4175 SDValue Op1 = getValue(I.getArgOperand(0));
4176 SDValue Op2 = getValue(I.getArgOperand(1));
4177 SDValue Op3 = getValue(I.getArgOperand(2));
4178 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4179 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4180 DAG.setRoot(DAG.getMemmove(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4181 MachinePointerInfo(I.getArgOperand(0)),
4182 MachinePointerInfo(I.getArgOperand(1))));
4185 case Intrinsic::dbg_declare: {
4186 const DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
4187 MDNode *Variable = DI.getVariable();
4188 const Value *Address = DI.getAddress();
4189 if (!Address || !DIVariable(DI.getVariable()).Verify())
4192 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder
4193 // but do not always have a corresponding SDNode built. The SDNodeOrder
4194 // absolute, but not relative, values are different depending on whether
4195 // debug info exists.
4198 // Check if address has undef value.
4199 if (isa<UndefValue>(Address) ||
4200 (Address->use_empty() && !isa<Argument>(Address))) {
4202 DAG.getDbgValue(Variable, UndefValue::get(Address->getType()),
4203 0, dl, SDNodeOrder);
4204 DAG.AddDbgValue(SDV, 0, false);
4208 SDValue &N = NodeMap[Address];
4209 if (!N.getNode() && isa<Argument>(Address))
4210 // Check unused arguments map.
4211 N = UnusedArgNodeMap[Address];
4214 // Parameters are handled specially.
4216 DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable;
4217 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
4218 Address = BCI->getOperand(0);
4219 const AllocaInst *AI = dyn_cast<AllocaInst>(Address);
4221 if (isParameter && !AI) {
4222 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode());
4224 // Byval parameter. We have a frame index at this point.
4225 SDV = DAG.getDbgValue(Variable, FINode->getIndex(),
4226 0, dl, SDNodeOrder);
4228 // Can't do anything with other non-AI cases yet. This might be a
4229 // parameter of a callee function that got inlined, for example.
4232 SDV = DAG.getDbgValue(Variable, N.getNode(), N.getResNo(),
4233 0, dl, SDNodeOrder);
4235 // Can't do anything with other non-AI cases yet.
4237 DAG.AddDbgValue(SDV, N.getNode(), isParameter);
4239 // If Address is an argument then try to emit its dbg value using
4240 // virtual register info from the FuncInfo.ValueMap.
4241 if (!EmitFuncArgumentDbgValue(Address, Variable, 0, N)) {
4242 // If variable is pinned by a alloca in dominating bb then
4243 // use StaticAllocaMap.
4244 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Address)) {
4245 if (AI->getParent() != DI.getParent()) {
4246 DenseMap<const AllocaInst*, int>::iterator SI =
4247 FuncInfo.StaticAllocaMap.find(AI);
4248 if (SI != FuncInfo.StaticAllocaMap.end()) {
4249 SDV = DAG.getDbgValue(Variable, SI->second,
4250 0, dl, SDNodeOrder);
4251 DAG.AddDbgValue(SDV, 0, false);
4256 // Otherwise add undef to help track missing debug info.
4257 SDV = DAG.getDbgValue(Variable, UndefValue::get(Address->getType()),
4258 0, dl, SDNodeOrder);
4259 DAG.AddDbgValue(SDV, 0, false);
4264 case Intrinsic::dbg_value: {
4265 const DbgValueInst &DI = cast<DbgValueInst>(I);
4266 if (!DIVariable(DI.getVariable()).Verify())
4269 MDNode *Variable = DI.getVariable();
4270 uint64_t Offset = DI.getOffset();
4271 const Value *V = DI.getValue();
4275 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder
4276 // but do not always have a corresponding SDNode built. The SDNodeOrder
4277 // absolute, but not relative, values are different depending on whether
4278 // debug info exists.
4281 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) {
4282 SDV = DAG.getDbgValue(Variable, V, Offset, dl, SDNodeOrder);
4283 DAG.AddDbgValue(SDV, 0, false);
4285 // Do not use getValue() in here; we don't want to generate code at
4286 // this point if it hasn't been done yet.
4287 SDValue N = NodeMap[V];
4288 if (!N.getNode() && isa<Argument>(V))
4289 // Check unused arguments map.
4290 N = UnusedArgNodeMap[V];
4292 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, N)) {
4293 SDV = DAG.getDbgValue(Variable, N.getNode(),
4294 N.getResNo(), Offset, dl, SDNodeOrder);
4295 DAG.AddDbgValue(SDV, N.getNode(), false);
4297 } else if (isa<PHINode>(V) && !V->use_empty() ) {
4298 // Do not call getValue(V) yet, as we don't want to generate code.
4299 // Remember it for later.
4300 DanglingDebugInfo DDI(&DI, dl, SDNodeOrder);
4301 DanglingDebugInfoMap[V] = DDI;
4303 // We may expand this to cover more cases. One case where we have no
4304 // data available is an unreferenced parameter; we need this fallback.
4305 SDV = DAG.getDbgValue(Variable, UndefValue::get(V->getType()),
4306 Offset, dl, SDNodeOrder);
4307 DAG.AddDbgValue(SDV, 0, false);
4311 // Build a debug info table entry.
4312 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
4313 V = BCI->getOperand(0);
4314 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
4315 // Don't handle byval struct arguments or VLAs, for example.
4318 DenseMap<const AllocaInst*, int>::iterator SI =
4319 FuncInfo.StaticAllocaMap.find(AI);
4320 if (SI == FuncInfo.StaticAllocaMap.end())
4322 int FI = SI->second;
4324 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4325 if (!DI.getDebugLoc().isUnknown() && MMI.hasDebugInfo())
4326 MMI.setVariableDbgInfo(Variable, FI, DI.getDebugLoc());
4329 case Intrinsic::eh_exception: {
4330 // Insert the EXCEPTIONADDR instruction.
4331 assert(FuncInfo.MBB->isLandingPad() &&
4332 "Call to eh.exception not in landing pad!");
4333 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
4335 Ops[0] = DAG.getRoot();
4336 SDValue Op = DAG.getNode(ISD::EXCEPTIONADDR, dl, VTs, Ops, 1);
4338 DAG.setRoot(Op.getValue(1));
4342 case Intrinsic::eh_selector: {
4343 MachineBasicBlock *CallMBB = FuncInfo.MBB;
4344 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4345 if (CallMBB->isLandingPad())
4346 AddCatchInfo(I, &MMI, CallMBB);
4349 FuncInfo.CatchInfoLost.insert(&I);
4351 // FIXME: Mark exception selector register as live in. Hack for PR1508.
4352 unsigned Reg = TLI.getExceptionSelectorRegister();
4353 if (Reg) FuncInfo.MBB->addLiveIn(Reg);
4356 // Insert the EHSELECTION instruction.
4357 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
4359 Ops[0] = getValue(I.getArgOperand(0));
4361 SDValue Op = DAG.getNode(ISD::EHSELECTION, dl, VTs, Ops, 2);
4362 DAG.setRoot(Op.getValue(1));
4363 setValue(&I, DAG.getSExtOrTrunc(Op, dl, MVT::i32));
4367 case Intrinsic::eh_typeid_for: {
4368 // Find the type id for the given typeinfo.
4369 GlobalVariable *GV = ExtractTypeInfo(I.getArgOperand(0));
4370 unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(GV);
4371 Res = DAG.getConstant(TypeID, MVT::i32);
4376 case Intrinsic::eh_return_i32:
4377 case Intrinsic::eh_return_i64:
4378 DAG.getMachineFunction().getMMI().setCallsEHReturn(true);
4379 DAG.setRoot(DAG.getNode(ISD::EH_RETURN, dl,
4382 getValue(I.getArgOperand(0)),
4383 getValue(I.getArgOperand(1))));
4385 case Intrinsic::eh_unwind_init:
4386 DAG.getMachineFunction().getMMI().setCallsUnwindInit(true);
4388 case Intrinsic::eh_dwarf_cfa: {
4389 SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getArgOperand(0)), dl,
4390 TLI.getPointerTy());
4391 SDValue Offset = DAG.getNode(ISD::ADD, dl,
4393 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, dl,
4394 TLI.getPointerTy()),
4396 SDValue FA = DAG.getNode(ISD::FRAMEADDR, dl,
4398 DAG.getConstant(0, TLI.getPointerTy()));
4399 setValue(&I, DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
4403 case Intrinsic::eh_sjlj_callsite: {
4404 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4405 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(0));
4406 assert(CI && "Non-constant call site value in eh.sjlj.callsite!");
4407 assert(MMI.getCurrentCallSite() == 0 && "Overlapping call sites!");
4409 MMI.setCurrentCallSite(CI->getZExtValue());
4412 case Intrinsic::eh_sjlj_setjmp: {
4413 setValue(&I, DAG.getNode(ISD::EH_SJLJ_SETJMP, dl, MVT::i32, getRoot(),
4414 getValue(I.getArgOperand(0))));
4417 case Intrinsic::eh_sjlj_longjmp: {
4418 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, dl, MVT::Other,
4419 getRoot(), getValue(I.getArgOperand(0))));
4422 case Intrinsic::eh_sjlj_dispatch_setup: {
4423 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_DISPATCHSETUP, dl, MVT::Other,
4424 getRoot(), getValue(I.getArgOperand(0))));
4428 case Intrinsic::x86_mmx_pslli_w:
4429 case Intrinsic::x86_mmx_pslli_d:
4430 case Intrinsic::x86_mmx_pslli_q:
4431 case Intrinsic::x86_mmx_psrli_w:
4432 case Intrinsic::x86_mmx_psrli_d:
4433 case Intrinsic::x86_mmx_psrli_q:
4434 case Intrinsic::x86_mmx_psrai_w:
4435 case Intrinsic::x86_mmx_psrai_d: {
4436 SDValue ShAmt = getValue(I.getArgOperand(1));
4437 if (isa<ConstantSDNode>(ShAmt)) {
4438 visitTargetIntrinsic(I, Intrinsic);
4441 unsigned NewIntrinsic = 0;
4442 EVT ShAmtVT = MVT::v2i32;
4443 switch (Intrinsic) {
4444 case Intrinsic::x86_mmx_pslli_w:
4445 NewIntrinsic = Intrinsic::x86_mmx_psll_w;
4447 case Intrinsic::x86_mmx_pslli_d:
4448 NewIntrinsic = Intrinsic::x86_mmx_psll_d;
4450 case Intrinsic::x86_mmx_pslli_q:
4451 NewIntrinsic = Intrinsic::x86_mmx_psll_q;
4453 case Intrinsic::x86_mmx_psrli_w:
4454 NewIntrinsic = Intrinsic::x86_mmx_psrl_w;
4456 case Intrinsic::x86_mmx_psrli_d:
4457 NewIntrinsic = Intrinsic::x86_mmx_psrl_d;
4459 case Intrinsic::x86_mmx_psrli_q:
4460 NewIntrinsic = Intrinsic::x86_mmx_psrl_q;
4462 case Intrinsic::x86_mmx_psrai_w:
4463 NewIntrinsic = Intrinsic::x86_mmx_psra_w;
4465 case Intrinsic::x86_mmx_psrai_d:
4466 NewIntrinsic = Intrinsic::x86_mmx_psra_d;
4468 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4471 // The vector shift intrinsics with scalars uses 32b shift amounts but
4472 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
4474 // We must do this early because v2i32 is not a legal type.
4475 DebugLoc dl = getCurDebugLoc();
4478 ShOps[1] = DAG.getConstant(0, MVT::i32);
4479 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
4480 EVT DestVT = TLI.getValueType(I.getType());
4481 ShAmt = DAG.getNode(ISD::BITCAST, dl, DestVT, ShAmt);
4482 Res = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4483 DAG.getConstant(NewIntrinsic, MVT::i32),
4484 getValue(I.getArgOperand(0)), ShAmt);
4488 case Intrinsic::convertff:
4489 case Intrinsic::convertfsi:
4490 case Intrinsic::convertfui:
4491 case Intrinsic::convertsif:
4492 case Intrinsic::convertuif:
4493 case Intrinsic::convertss:
4494 case Intrinsic::convertsu:
4495 case Intrinsic::convertus:
4496 case Intrinsic::convertuu: {
4497 ISD::CvtCode Code = ISD::CVT_INVALID;
4498 switch (Intrinsic) {
4499 case Intrinsic::convertff: Code = ISD::CVT_FF; break;
4500 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break;
4501 case Intrinsic::convertfui: Code = ISD::CVT_FU; break;
4502 case Intrinsic::convertsif: Code = ISD::CVT_SF; break;
4503 case Intrinsic::convertuif: Code = ISD::CVT_UF; break;
4504 case Intrinsic::convertss: Code = ISD::CVT_SS; break;
4505 case Intrinsic::convertsu: Code = ISD::CVT_SU; break;
4506 case Intrinsic::convertus: Code = ISD::CVT_US; break;
4507 case Intrinsic::convertuu: Code = ISD::CVT_UU; break;
4509 EVT DestVT = TLI.getValueType(I.getType());
4510 const Value *Op1 = I.getArgOperand(0);
4511 Res = DAG.getConvertRndSat(DestVT, getCurDebugLoc(), getValue(Op1),
4512 DAG.getValueType(DestVT),
4513 DAG.getValueType(getValue(Op1).getValueType()),
4514 getValue(I.getArgOperand(1)),
4515 getValue(I.getArgOperand(2)),
4520 case Intrinsic::sqrt:
4521 setValue(&I, DAG.getNode(ISD::FSQRT, dl,
4522 getValue(I.getArgOperand(0)).getValueType(),
4523 getValue(I.getArgOperand(0))));
4525 case Intrinsic::powi:
4526 setValue(&I, ExpandPowI(dl, getValue(I.getArgOperand(0)),
4527 getValue(I.getArgOperand(1)), DAG));
4529 case Intrinsic::sin:
4530 setValue(&I, DAG.getNode(ISD::FSIN, dl,
4531 getValue(I.getArgOperand(0)).getValueType(),
4532 getValue(I.getArgOperand(0))));
4534 case Intrinsic::cos:
4535 setValue(&I, DAG.getNode(ISD::FCOS, dl,
4536 getValue(I.getArgOperand(0)).getValueType(),
4537 getValue(I.getArgOperand(0))));
4539 case Intrinsic::log:
4542 case Intrinsic::log2:
4545 case Intrinsic::log10:
4548 case Intrinsic::exp:
4551 case Intrinsic::exp2:
4554 case Intrinsic::pow:
4557 case Intrinsic::convert_to_fp16:
4558 setValue(&I, DAG.getNode(ISD::FP32_TO_FP16, dl,
4559 MVT::i16, getValue(I.getArgOperand(0))));
4561 case Intrinsic::convert_from_fp16:
4562 setValue(&I, DAG.getNode(ISD::FP16_TO_FP32, dl,
4563 MVT::f32, getValue(I.getArgOperand(0))));
4565 case Intrinsic::pcmarker: {
4566 SDValue Tmp = getValue(I.getArgOperand(0));
4567 DAG.setRoot(DAG.getNode(ISD::PCMARKER, dl, MVT::Other, getRoot(), Tmp));
4570 case Intrinsic::readcyclecounter: {
4571 SDValue Op = getRoot();
4572 Res = DAG.getNode(ISD::READCYCLECOUNTER, dl,
4573 DAG.getVTList(MVT::i64, MVT::Other),
4576 DAG.setRoot(Res.getValue(1));
4579 case Intrinsic::bswap:
4580 setValue(&I, DAG.getNode(ISD::BSWAP, dl,
4581 getValue(I.getArgOperand(0)).getValueType(),
4582 getValue(I.getArgOperand(0))));
4584 case Intrinsic::cttz: {
4585 SDValue Arg = getValue(I.getArgOperand(0));
4586 EVT Ty = Arg.getValueType();
4587 setValue(&I, DAG.getNode(ISD::CTTZ, dl, Ty, Arg));
4590 case Intrinsic::ctlz: {
4591 SDValue Arg = getValue(I.getArgOperand(0));
4592 EVT Ty = Arg.getValueType();
4593 setValue(&I, DAG.getNode(ISD::CTLZ, dl, Ty, Arg));
4596 case Intrinsic::ctpop: {
4597 SDValue Arg = getValue(I.getArgOperand(0));
4598 EVT Ty = Arg.getValueType();
4599 setValue(&I, DAG.getNode(ISD::CTPOP, dl, Ty, Arg));
4602 case Intrinsic::stacksave: {
4603 SDValue Op = getRoot();
4604 Res = DAG.getNode(ISD::STACKSAVE, dl,
4605 DAG.getVTList(TLI.getPointerTy(), MVT::Other), &Op, 1);
4607 DAG.setRoot(Res.getValue(1));
4610 case Intrinsic::stackrestore: {
4611 Res = getValue(I.getArgOperand(0));
4612 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, dl, MVT::Other, getRoot(), Res));
4615 case Intrinsic::stackprotector: {
4616 // Emit code into the DAG to store the stack guard onto the stack.
4617 MachineFunction &MF = DAG.getMachineFunction();
4618 MachineFrameInfo *MFI = MF.getFrameInfo();
4619 EVT PtrTy = TLI.getPointerTy();
4621 SDValue Src = getValue(I.getArgOperand(0)); // The guard's value.
4622 AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
4624 int FI = FuncInfo.StaticAllocaMap[Slot];
4625 MFI->setStackProtectorIndex(FI);
4627 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
4629 // Store the stack protector onto the stack.
4630 Res = DAG.getStore(getRoot(), getCurDebugLoc(), Src, FIN,
4631 MachinePointerInfo::getFixedStack(FI),
4637 case Intrinsic::objectsize: {
4638 // If we don't know by now, we're never going to know.
4639 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1));
4641 assert(CI && "Non-constant type in __builtin_object_size?");
4643 SDValue Arg = getValue(I.getCalledValue());
4644 EVT Ty = Arg.getValueType();
4647 Res = DAG.getConstant(-1ULL, Ty);
4649 Res = DAG.getConstant(0, Ty);
4654 case Intrinsic::var_annotation:
4655 // Discard annotate attributes
4658 case Intrinsic::init_trampoline: {
4659 const Function *F = cast<Function>(I.getArgOperand(1)->stripPointerCasts());
4663 Ops[1] = getValue(I.getArgOperand(0));
4664 Ops[2] = getValue(I.getArgOperand(1));
4665 Ops[3] = getValue(I.getArgOperand(2));
4666 Ops[4] = DAG.getSrcValue(I.getArgOperand(0));
4667 Ops[5] = DAG.getSrcValue(F);
4669 Res = DAG.getNode(ISD::TRAMPOLINE, dl,
4670 DAG.getVTList(TLI.getPointerTy(), MVT::Other),
4674 DAG.setRoot(Res.getValue(1));
4677 case Intrinsic::gcroot:
4679 const Value *Alloca = I.getArgOperand(0);
4680 const Constant *TypeMap = cast<Constant>(I.getArgOperand(1));
4682 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode());
4683 GFI->addStackRoot(FI->getIndex(), TypeMap);
4686 case Intrinsic::gcread:
4687 case Intrinsic::gcwrite:
4688 llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!");
4690 case Intrinsic::flt_rounds:
4691 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, dl, MVT::i32));
4693 case Intrinsic::trap:
4694 DAG.setRoot(DAG.getNode(ISD::TRAP, dl,MVT::Other, getRoot()));
4696 case Intrinsic::uadd_with_overflow:
4697 return implVisitAluOverflow(I, ISD::UADDO);
4698 case Intrinsic::sadd_with_overflow:
4699 return implVisitAluOverflow(I, ISD::SADDO);
4700 case Intrinsic::usub_with_overflow:
4701 return implVisitAluOverflow(I, ISD::USUBO);
4702 case Intrinsic::ssub_with_overflow:
4703 return implVisitAluOverflow(I, ISD::SSUBO);
4704 case Intrinsic::umul_with_overflow:
4705 return implVisitAluOverflow(I, ISD::UMULO);
4706 case Intrinsic::smul_with_overflow:
4707 return implVisitAluOverflow(I, ISD::SMULO);
4709 case Intrinsic::prefetch: {
4711 unsigned rw = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
4713 Ops[1] = getValue(I.getArgOperand(0));
4714 Ops[2] = getValue(I.getArgOperand(1));
4715 Ops[3] = getValue(I.getArgOperand(2));
4716 DAG.setRoot(DAG.getMemIntrinsicNode(ISD::PREFETCH, dl,
4717 DAG.getVTList(MVT::Other),
4719 EVT::getIntegerVT(*Context, 8),
4720 MachinePointerInfo(I.getArgOperand(0)),
4722 false, /* volatile */
4724 rw==1)); /* write */
4727 case Intrinsic::memory_barrier: {
4730 for (int x = 1; x < 6; ++x)
4731 Ops[x] = getValue(I.getArgOperand(x - 1));
4733 DAG.setRoot(DAG.getNode(ISD::MEMBARRIER, dl, MVT::Other, &Ops[0], 6));
4736 case Intrinsic::atomic_cmp_swap: {
4737 SDValue Root = getRoot();
4739 DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, getCurDebugLoc(),
4740 getValue(I.getArgOperand(1)).getValueType().getSimpleVT(),
4742 getValue(I.getArgOperand(0)),
4743 getValue(I.getArgOperand(1)),
4744 getValue(I.getArgOperand(2)),
4745 MachinePointerInfo(I.getArgOperand(0)));
4747 DAG.setRoot(L.getValue(1));
4750 case Intrinsic::atomic_load_add:
4751 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_ADD);
4752 case Intrinsic::atomic_load_sub:
4753 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_SUB);
4754 case Intrinsic::atomic_load_or:
4755 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_OR);
4756 case Intrinsic::atomic_load_xor:
4757 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_XOR);
4758 case Intrinsic::atomic_load_and:
4759 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_AND);
4760 case Intrinsic::atomic_load_nand:
4761 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_NAND);
4762 case Intrinsic::atomic_load_max:
4763 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MAX);
4764 case Intrinsic::atomic_load_min:
4765 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MIN);
4766 case Intrinsic::atomic_load_umin:
4767 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMIN);
4768 case Intrinsic::atomic_load_umax:
4769 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMAX);
4770 case Intrinsic::atomic_swap:
4771 return implVisitBinaryAtomic(I, ISD::ATOMIC_SWAP);
4773 case Intrinsic::invariant_start:
4774 case Intrinsic::lifetime_start:
4775 // Discard region information.
4776 setValue(&I, DAG.getUNDEF(TLI.getPointerTy()));
4778 case Intrinsic::invariant_end:
4779 case Intrinsic::lifetime_end:
4780 // Discard region information.
4785 void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee,
4787 MachineBasicBlock *LandingPad) {
4788 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
4789 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
4790 const Type *RetTy = FTy->getReturnType();
4791 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4792 MCSymbol *BeginLabel = 0;
4794 TargetLowering::ArgListTy Args;
4795 TargetLowering::ArgListEntry Entry;
4796 Args.reserve(CS.arg_size());
4798 // Check whether the function can return without sret-demotion.
4799 SmallVector<ISD::OutputArg, 4> Outs;
4800 SmallVector<uint64_t, 4> Offsets;
4801 GetReturnInfo(RetTy, CS.getAttributes().getRetAttributes(),
4802 Outs, TLI, &Offsets);
4804 bool CanLowerReturn = TLI.CanLowerReturn(CS.getCallingConv(),
4805 FTy->isVarArg(), Outs, FTy->getContext());
4807 SDValue DemoteStackSlot;
4808 int DemoteStackIdx = -100;
4810 if (!CanLowerReturn) {
4811 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(
4812 FTy->getReturnType());
4813 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(
4814 FTy->getReturnType());
4815 MachineFunction &MF = DAG.getMachineFunction();
4816 DemoteStackIdx = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
4817 const Type *StackSlotPtrType = PointerType::getUnqual(FTy->getReturnType());
4819 DemoteStackSlot = DAG.getFrameIndex(DemoteStackIdx, TLI.getPointerTy());
4820 Entry.Node = DemoteStackSlot;
4821 Entry.Ty = StackSlotPtrType;
4822 Entry.isSExt = false;
4823 Entry.isZExt = false;
4824 Entry.isInReg = false;
4825 Entry.isSRet = true;
4826 Entry.isNest = false;
4827 Entry.isByVal = false;
4828 Entry.Alignment = Align;
4829 Args.push_back(Entry);
4830 RetTy = Type::getVoidTy(FTy->getContext());
4833 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
4835 SDValue ArgNode = getValue(*i);
4836 Entry.Node = ArgNode; Entry.Ty = (*i)->getType();
4838 unsigned attrInd = i - CS.arg_begin() + 1;
4839 Entry.isSExt = CS.paramHasAttr(attrInd, Attribute::SExt);
4840 Entry.isZExt = CS.paramHasAttr(attrInd, Attribute::ZExt);
4841 Entry.isInReg = CS.paramHasAttr(attrInd, Attribute::InReg);
4842 Entry.isSRet = CS.paramHasAttr(attrInd, Attribute::StructRet);
4843 Entry.isNest = CS.paramHasAttr(attrInd, Attribute::Nest);
4844 Entry.isByVal = CS.paramHasAttr(attrInd, Attribute::ByVal);
4845 Entry.Alignment = CS.getParamAlignment(attrInd);
4846 Args.push_back(Entry);
4850 // Insert a label before the invoke call to mark the try range. This can be
4851 // used to detect deletion of the invoke via the MachineModuleInfo.
4852 BeginLabel = MMI.getContext().CreateTempSymbol();
4854 // For SjLj, keep track of which landing pads go with which invokes
4855 // so as to maintain the ordering of pads in the LSDA.
4856 unsigned CallSiteIndex = MMI.getCurrentCallSite();
4857 if (CallSiteIndex) {
4858 MMI.setCallSiteBeginLabel(BeginLabel, CallSiteIndex);
4859 // Now that the call site is handled, stop tracking it.
4860 MMI.setCurrentCallSite(0);
4863 // Both PendingLoads and PendingExports must be flushed here;
4864 // this call might not return.
4866 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getControlRoot(), BeginLabel));
4869 // Check if target-independent constraints permit a tail call here.
4870 // Target-dependent constraints are checked within TLI.LowerCallTo.
4872 !isInTailCallPosition(CS, CS.getAttributes().getRetAttributes(), TLI))
4875 // If there's a possibility that fast-isel has already selected some amount
4876 // of the current basic block, don't emit a tail call.
4877 if (isTailCall && EnableFastISel)
4880 std::pair<SDValue,SDValue> Result =
4881 TLI.LowerCallTo(getRoot(), RetTy,
4882 CS.paramHasAttr(0, Attribute::SExt),
4883 CS.paramHasAttr(0, Attribute::ZExt), FTy->isVarArg(),
4884 CS.paramHasAttr(0, Attribute::InReg), FTy->getNumParams(),
4885 CS.getCallingConv(),
4887 !CS.getInstruction()->use_empty(),
4888 Callee, Args, DAG, getCurDebugLoc());
4889 assert((isTailCall || Result.second.getNode()) &&
4890 "Non-null chain expected with non-tail call!");
4891 assert((Result.second.getNode() || !Result.first.getNode()) &&
4892 "Null value expected with tail call!");
4893 if (Result.first.getNode()) {
4894 setValue(CS.getInstruction(), Result.first);
4895 } else if (!CanLowerReturn && Result.second.getNode()) {
4896 // The instruction result is the result of loading from the
4897 // hidden sret parameter.
4898 SmallVector<EVT, 1> PVTs;
4899 const Type *PtrRetTy = PointerType::getUnqual(FTy->getReturnType());
4901 ComputeValueVTs(TLI, PtrRetTy, PVTs);
4902 assert(PVTs.size() == 1 && "Pointers should fit in one register");
4903 EVT PtrVT = PVTs[0];
4904 unsigned NumValues = Outs.size();
4905 SmallVector<SDValue, 4> Values(NumValues);
4906 SmallVector<SDValue, 4> Chains(NumValues);
4908 for (unsigned i = 0; i < NumValues; ++i) {
4909 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT,
4911 DAG.getConstant(Offsets[i], PtrVT));
4912 SDValue L = DAG.getLoad(Outs[i].VT, getCurDebugLoc(), Result.second,
4914 MachinePointerInfo::getFixedStack(DemoteStackIdx, Offsets[i]),
4917 Chains[i] = L.getValue(1);
4920 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
4921 MVT::Other, &Chains[0], NumValues);
4922 PendingLoads.push_back(Chain);
4924 // Collect the legal value parts into potentially illegal values
4925 // that correspond to the original function's return values.
4926 SmallVector<EVT, 4> RetTys;
4927 RetTy = FTy->getReturnType();
4928 ComputeValueVTs(TLI, RetTy, RetTys);
4929 ISD::NodeType AssertOp = ISD::DELETED_NODE;
4930 SmallVector<SDValue, 4> ReturnValues;
4931 unsigned CurReg = 0;
4932 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
4934 EVT RegisterVT = TLI.getRegisterType(RetTy->getContext(), VT);
4935 unsigned NumRegs = TLI.getNumRegisters(RetTy->getContext(), VT);
4937 SDValue ReturnValue =
4938 getCopyFromParts(DAG, getCurDebugLoc(), &Values[CurReg], NumRegs,
4939 RegisterVT, VT, AssertOp);
4940 ReturnValues.push_back(ReturnValue);
4944 setValue(CS.getInstruction(),
4945 DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
4946 DAG.getVTList(&RetTys[0], RetTys.size()),
4947 &ReturnValues[0], ReturnValues.size()));
4951 // As a special case, a null chain means that a tail call has been emitted and
4952 // the DAG root is already updated.
4953 if (Result.second.getNode())
4954 DAG.setRoot(Result.second);
4959 // Insert a label at the end of the invoke call to mark the try range. This
4960 // can be used to detect deletion of the invoke via the MachineModuleInfo.
4961 MCSymbol *EndLabel = MMI.getContext().CreateTempSymbol();
4962 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getRoot(), EndLabel));
4964 // Inform MachineModuleInfo of range.
4965 MMI.addInvoke(LandingPad, BeginLabel, EndLabel);
4969 /// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the
4970 /// value is equal or not-equal to zero.
4971 static bool IsOnlyUsedInZeroEqualityComparison(const Value *V) {
4972 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end();
4974 if (const ICmpInst *IC = dyn_cast<ICmpInst>(*UI))
4975 if (IC->isEquality())
4976 if (const Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
4977 if (C->isNullValue())
4979 // Unknown instruction.
4985 static SDValue getMemCmpLoad(const Value *PtrVal, MVT LoadVT,
4987 SelectionDAGBuilder &Builder) {
4989 // Check to see if this load can be trivially constant folded, e.g. if the
4990 // input is from a string literal.
4991 if (const Constant *LoadInput = dyn_cast<Constant>(PtrVal)) {
4992 // Cast pointer to the type we really want to load.
4993 LoadInput = ConstantExpr::getBitCast(const_cast<Constant *>(LoadInput),
4994 PointerType::getUnqual(LoadTy));
4996 if (const Constant *LoadCst =
4997 ConstantFoldLoadFromConstPtr(const_cast<Constant *>(LoadInput),
4999 return Builder.getValue(LoadCst);
5002 // Otherwise, we have to emit the load. If the pointer is to unfoldable but
5003 // still constant memory, the input chain can be the entry node.
5005 bool ConstantMemory = false;
5007 // Do not serialize (non-volatile) loads of constant memory with anything.
5008 if (Builder.AA->pointsToConstantMemory(PtrVal)) {
5009 Root = Builder.DAG.getEntryNode();
5010 ConstantMemory = true;
5012 // Do not serialize non-volatile loads against each other.
5013 Root = Builder.DAG.getRoot();
5016 SDValue Ptr = Builder.getValue(PtrVal);
5017 SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurDebugLoc(), Root,
5018 Ptr, MachinePointerInfo(PtrVal),
5020 false /*nontemporal*/, 1 /* align=1 */);
5022 if (!ConstantMemory)
5023 Builder.PendingLoads.push_back(LoadVal.getValue(1));
5028 /// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form.
5029 /// If so, return true and lower it, otherwise return false and it will be
5030 /// lowered like a normal call.
5031 bool SelectionDAGBuilder::visitMemCmpCall(const CallInst &I) {
5032 // Verify that the prototype makes sense. int memcmp(void*,void*,size_t)
5033 if (I.getNumArgOperands() != 3)
5036 const Value *LHS = I.getArgOperand(0), *RHS = I.getArgOperand(1);
5037 if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() ||
5038 !I.getArgOperand(2)->getType()->isIntegerTy() ||
5039 !I.getType()->isIntegerTy())
5042 const ConstantInt *Size = dyn_cast<ConstantInt>(I.getArgOperand(2));
5044 // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0
5045 // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0
5046 if (Size && IsOnlyUsedInZeroEqualityComparison(&I)) {
5047 bool ActuallyDoIt = true;
5050 switch (Size->getZExtValue()) {
5052 LoadVT = MVT::Other;
5054 ActuallyDoIt = false;
5058 LoadTy = Type::getInt16Ty(Size->getContext());
5062 LoadTy = Type::getInt32Ty(Size->getContext());
5066 LoadTy = Type::getInt64Ty(Size->getContext());
5070 LoadVT = MVT::v4i32;
5071 LoadTy = Type::getInt32Ty(Size->getContext());
5072 LoadTy = VectorType::get(LoadTy, 4);
5077 // This turns into unaligned loads. We only do this if the target natively
5078 // supports the MVT we'll be loading or if it is small enough (<= 4) that
5079 // we'll only produce a small number of byte loads.
5081 // Require that we can find a legal MVT, and only do this if the target
5082 // supports unaligned loads of that type. Expanding into byte loads would
5084 if (ActuallyDoIt && Size->getZExtValue() > 4) {
5085 // TODO: Handle 5 byte compare as 4-byte + 1 byte.
5086 // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads.
5087 if (!TLI.isTypeLegal(LoadVT) ||!TLI.allowsUnalignedMemoryAccesses(LoadVT))
5088 ActuallyDoIt = false;
5092 SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this);
5093 SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this);
5095 SDValue Res = DAG.getSetCC(getCurDebugLoc(), MVT::i1, LHSVal, RHSVal,
5097 EVT CallVT = TLI.getValueType(I.getType(), true);
5098 setValue(&I, DAG.getZExtOrTrunc(Res, getCurDebugLoc(), CallVT));
5108 void SelectionDAGBuilder::visitCall(const CallInst &I) {
5109 // Handle inline assembly differently.
5110 if (isa<InlineAsm>(I.getCalledValue())) {
5115 // See if any floating point values are being passed to this function. This is
5116 // used to emit an undefined reference to fltused on Windows.
5117 const FunctionType *FT =
5118 cast<FunctionType>(I.getCalledValue()->getType()->getContainedType(0));
5119 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5120 if (FT->isVarArg() &&
5121 !MMI.callsExternalVAFunctionWithFloatingPointArguments()) {
5122 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) {
5123 const Type* T = I.getArgOperand(i)->getType();
5124 for (po_iterator<const Type*> i = po_begin(T), e = po_end(T);
5126 if (!i->isFloatingPointTy()) continue;
5127 MMI.setCallsExternalVAFunctionWithFloatingPointArguments(true);
5133 const char *RenameFn = 0;
5134 if (Function *F = I.getCalledFunction()) {
5135 if (F->isDeclaration()) {
5136 if (const TargetIntrinsicInfo *II = TM.getIntrinsicInfo()) {
5137 if (unsigned IID = II->getIntrinsicID(F)) {
5138 RenameFn = visitIntrinsicCall(I, IID);
5143 if (unsigned IID = F->getIntrinsicID()) {
5144 RenameFn = visitIntrinsicCall(I, IID);
5150 // Check for well-known libc/libm calls. If the function is internal, it
5151 // can't be a library call.
5152 if (!F->hasLocalLinkage() && F->hasName()) {
5153 StringRef Name = F->getName();
5154 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl") {
5155 if (I.getNumArgOperands() == 2 && // Basic sanity checks.
5156 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5157 I.getType() == I.getArgOperand(0)->getType() &&
5158 I.getType() == I.getArgOperand(1)->getType()) {
5159 SDValue LHS = getValue(I.getArgOperand(0));
5160 SDValue RHS = getValue(I.getArgOperand(1));
5161 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurDebugLoc(),
5162 LHS.getValueType(), LHS, RHS));
5165 } else if (Name == "fabs" || Name == "fabsf" || Name == "fabsl") {
5166 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5167 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5168 I.getType() == I.getArgOperand(0)->getType()) {
5169 SDValue Tmp = getValue(I.getArgOperand(0));
5170 setValue(&I, DAG.getNode(ISD::FABS, getCurDebugLoc(),
5171 Tmp.getValueType(), Tmp));
5174 } else if (Name == "sin" || Name == "sinf" || Name == "sinl") {
5175 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5176 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5177 I.getType() == I.getArgOperand(0)->getType() &&
5178 I.onlyReadsMemory()) {
5179 SDValue Tmp = getValue(I.getArgOperand(0));
5180 setValue(&I, DAG.getNode(ISD::FSIN, getCurDebugLoc(),
5181 Tmp.getValueType(), Tmp));
5184 } else if (Name == "cos" || Name == "cosf" || Name == "cosl") {
5185 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5186 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5187 I.getType() == I.getArgOperand(0)->getType() &&
5188 I.onlyReadsMemory()) {
5189 SDValue Tmp = getValue(I.getArgOperand(0));
5190 setValue(&I, DAG.getNode(ISD::FCOS, getCurDebugLoc(),
5191 Tmp.getValueType(), Tmp));
5194 } else if (Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl") {
5195 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5196 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5197 I.getType() == I.getArgOperand(0)->getType() &&
5198 I.onlyReadsMemory()) {
5199 SDValue Tmp = getValue(I.getArgOperand(0));
5200 setValue(&I, DAG.getNode(ISD::FSQRT, getCurDebugLoc(),
5201 Tmp.getValueType(), Tmp));
5204 } else if (Name == "memcmp") {
5205 if (visitMemCmpCall(I))
5213 Callee = getValue(I.getCalledValue());
5215 Callee = DAG.getExternalSymbol(RenameFn, TLI.getPointerTy());
5217 // Check if we can potentially perform a tail call. More detailed checking is
5218 // be done within LowerCallTo, after more information about the call is known.
5219 LowerCallTo(&I, Callee, I.isTailCall());
5224 /// AsmOperandInfo - This contains information for each constraint that we are
5226 class LLVM_LIBRARY_VISIBILITY SDISelAsmOperandInfo :
5227 public TargetLowering::AsmOperandInfo {
5229 /// CallOperand - If this is the result output operand or a clobber
5230 /// this is null, otherwise it is the incoming operand to the CallInst.
5231 /// This gets modified as the asm is processed.
5232 SDValue CallOperand;
5234 /// AssignedRegs - If this is a register or register class operand, this
5235 /// contains the set of register corresponding to the operand.
5236 RegsForValue AssignedRegs;
5238 explicit SDISelAsmOperandInfo(const TargetLowering::AsmOperandInfo &info)
5239 : TargetLowering::AsmOperandInfo(info), CallOperand(0,0) {
5242 /// MarkAllocatedRegs - Once AssignedRegs is set, mark the assigned registers
5243 /// busy in OutputRegs/InputRegs.
5244 void MarkAllocatedRegs(bool isOutReg, bool isInReg,
5245 std::set<unsigned> &OutputRegs,
5246 std::set<unsigned> &InputRegs,
5247 const TargetRegisterInfo &TRI) const {
5249 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
5250 MarkRegAndAliases(AssignedRegs.Regs[i], OutputRegs, TRI);
5253 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
5254 MarkRegAndAliases(AssignedRegs.Regs[i], InputRegs, TRI);
5258 /// getCallOperandValEVT - Return the EVT of the Value* that this operand
5259 /// corresponds to. If there is no Value* for this operand, it returns
5261 EVT getCallOperandValEVT(LLVMContext &Context,
5262 const TargetLowering &TLI,
5263 const TargetData *TD) const {
5264 if (CallOperandVal == 0) return MVT::Other;
5266 if (isa<BasicBlock>(CallOperandVal))
5267 return TLI.getPointerTy();
5269 const llvm::Type *OpTy = CallOperandVal->getType();
5271 // If this is an indirect operand, the operand is a pointer to the
5274 const llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
5276 report_fatal_error("Indirect operand for inline asm not a pointer!");
5277 OpTy = PtrTy->getElementType();
5280 // If OpTy is not a single value, it may be a struct/union that we
5281 // can tile with integers.
5282 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
5283 unsigned BitSize = TD->getTypeSizeInBits(OpTy);
5292 OpTy = IntegerType::get(Context, BitSize);
5297 return TLI.getValueType(OpTy, true);
5301 /// MarkRegAndAliases - Mark the specified register and all aliases in the
5303 static void MarkRegAndAliases(unsigned Reg, std::set<unsigned> &Regs,
5304 const TargetRegisterInfo &TRI) {
5305 assert(TargetRegisterInfo::isPhysicalRegister(Reg) && "Isn't a physreg");
5307 if (const unsigned *Aliases = TRI.getAliasSet(Reg))
5308 for (; *Aliases; ++Aliases)
5309 Regs.insert(*Aliases);
5313 typedef SmallVector<SDISelAsmOperandInfo,16> SDISelAsmOperandInfoVector;
5315 } // end llvm namespace.
5317 /// isAllocatableRegister - If the specified register is safe to allocate,
5318 /// i.e. it isn't a stack pointer or some other special register, return the
5319 /// register class for the register. Otherwise, return null.
5320 static const TargetRegisterClass *
5321 isAllocatableRegister(unsigned Reg, MachineFunction &MF,
5322 const TargetLowering &TLI,
5323 const TargetRegisterInfo *TRI) {
5324 EVT FoundVT = MVT::Other;
5325 const TargetRegisterClass *FoundRC = 0;
5326 for (TargetRegisterInfo::regclass_iterator RCI = TRI->regclass_begin(),
5327 E = TRI->regclass_end(); RCI != E; ++RCI) {
5328 EVT ThisVT = MVT::Other;
5330 const TargetRegisterClass *RC = *RCI;
5331 // If none of the value types for this register class are valid, we
5332 // can't use it. For example, 64-bit reg classes on 32-bit targets.
5333 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
5335 if (TLI.isTypeLegal(*I)) {
5336 // If we have already found this register in a different register class,
5337 // choose the one with the largest VT specified. For example, on
5338 // PowerPC, we favor f64 register classes over f32.
5339 if (FoundVT == MVT::Other || FoundVT.bitsLT(*I)) {
5346 if (ThisVT == MVT::Other) continue;
5348 // NOTE: This isn't ideal. In particular, this might allocate the
5349 // frame pointer in functions that need it (due to them not being taken
5350 // out of allocation, because a variable sized allocation hasn't been seen
5351 // yet). This is a slight code pessimization, but should still work.
5352 for (TargetRegisterClass::iterator I = RC->allocation_order_begin(MF),
5353 E = RC->allocation_order_end(MF); I != E; ++I)
5355 // We found a matching register class. Keep looking at others in case
5356 // we find one with larger registers that this physreg is also in.
5365 /// GetRegistersForValue - Assign registers (virtual or physical) for the
5366 /// specified operand. We prefer to assign virtual registers, to allow the
5367 /// register allocator to handle the assignment process. However, if the asm
5368 /// uses features that we can't model on machineinstrs, we have SDISel do the
5369 /// allocation. This produces generally horrible, but correct, code.
5371 /// OpInfo describes the operand.
5372 /// Input and OutputRegs are the set of already allocated physical registers.
5374 void SelectionDAGBuilder::
5375 GetRegistersForValue(SDISelAsmOperandInfo &OpInfo,
5376 std::set<unsigned> &OutputRegs,
5377 std::set<unsigned> &InputRegs) {
5378 LLVMContext &Context = FuncInfo.Fn->getContext();
5380 // Compute whether this value requires an input register, an output register,
5382 bool isOutReg = false;
5383 bool isInReg = false;
5384 switch (OpInfo.Type) {
5385 case InlineAsm::isOutput:
5388 // If there is an input constraint that matches this, we need to reserve
5389 // the input register so no other inputs allocate to it.
5390 isInReg = OpInfo.hasMatchingInput();
5392 case InlineAsm::isInput:
5396 case InlineAsm::isClobber:
5403 MachineFunction &MF = DAG.getMachineFunction();
5404 SmallVector<unsigned, 4> Regs;
5406 // If this is a constraint for a single physreg, or a constraint for a
5407 // register class, find it.
5408 std::pair<unsigned, const TargetRegisterClass*> PhysReg =
5409 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
5410 OpInfo.ConstraintVT);
5412 unsigned NumRegs = 1;
5413 if (OpInfo.ConstraintVT != MVT::Other) {
5414 // If this is a FP input in an integer register (or visa versa) insert a bit
5415 // cast of the input value. More generally, handle any case where the input
5416 // value disagrees with the register class we plan to stick this in.
5417 if (OpInfo.Type == InlineAsm::isInput &&
5418 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) {
5419 // Try to convert to the first EVT that the reg class contains. If the
5420 // types are identical size, use a bitcast to convert (e.g. two differing
5422 EVT RegVT = *PhysReg.second->vt_begin();
5423 if (RegVT.getSizeInBits() == OpInfo.ConstraintVT.getSizeInBits()) {
5424 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, getCurDebugLoc(),
5425 RegVT, OpInfo.CallOperand);
5426 OpInfo.ConstraintVT = RegVT;
5427 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) {
5428 // If the input is a FP value and we want it in FP registers, do a
5429 // bitcast to the corresponding integer type. This turns an f64 value
5430 // into i64, which can be passed with two i32 values on a 32-bit
5432 RegVT = EVT::getIntegerVT(Context,
5433 OpInfo.ConstraintVT.getSizeInBits());
5434 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, getCurDebugLoc(),
5435 RegVT, OpInfo.CallOperand);
5436 OpInfo.ConstraintVT = RegVT;
5440 NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT);
5444 EVT ValueVT = OpInfo.ConstraintVT;
5446 // If this is a constraint for a specific physical register, like {r17},
5448 if (unsigned AssignedReg = PhysReg.first) {
5449 const TargetRegisterClass *RC = PhysReg.second;
5450 if (OpInfo.ConstraintVT == MVT::Other)
5451 ValueVT = *RC->vt_begin();
5453 // Get the actual register value type. This is important, because the user
5454 // may have asked for (e.g.) the AX register in i32 type. We need to
5455 // remember that AX is actually i16 to get the right extension.
5456 RegVT = *RC->vt_begin();
5458 // This is a explicit reference to a physical register.
5459 Regs.push_back(AssignedReg);
5461 // If this is an expanded reference, add the rest of the regs to Regs.
5463 TargetRegisterClass::iterator I = RC->begin();
5464 for (; *I != AssignedReg; ++I)
5465 assert(I != RC->end() && "Didn't find reg!");
5467 // Already added the first reg.
5469 for (; NumRegs; --NumRegs, ++I) {
5470 assert(I != RC->end() && "Ran out of registers to allocate!");
5475 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5476 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
5477 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
5481 // Otherwise, if this was a reference to an LLVM register class, create vregs
5482 // for this reference.
5483 if (const TargetRegisterClass *RC = PhysReg.second) {
5484 RegVT = *RC->vt_begin();
5485 if (OpInfo.ConstraintVT == MVT::Other)
5488 // Create the appropriate number of virtual registers.
5489 MachineRegisterInfo &RegInfo = MF.getRegInfo();
5490 for (; NumRegs; --NumRegs)
5491 Regs.push_back(RegInfo.createVirtualRegister(RC));
5493 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5497 // This is a reference to a register class that doesn't directly correspond
5498 // to an LLVM register class. Allocate NumRegs consecutive, available,
5499 // registers from the class.
5500 std::vector<unsigned> RegClassRegs
5501 = TLI.getRegClassForInlineAsmConstraint(OpInfo.ConstraintCode,
5502 OpInfo.ConstraintVT);
5504 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
5505 unsigned NumAllocated = 0;
5506 for (unsigned i = 0, e = RegClassRegs.size(); i != e; ++i) {
5507 unsigned Reg = RegClassRegs[i];
5508 // See if this register is available.
5509 if ((isOutReg && OutputRegs.count(Reg)) || // Already used.
5510 (isInReg && InputRegs.count(Reg))) { // Already used.
5511 // Make sure we find consecutive registers.
5516 // Check to see if this register is allocatable (i.e. don't give out the
5518 const TargetRegisterClass *RC = isAllocatableRegister(Reg, MF, TLI, TRI);
5519 if (!RC) { // Couldn't allocate this register.
5520 // Reset NumAllocated to make sure we return consecutive registers.
5525 // Okay, this register is good, we can use it.
5528 // If we allocated enough consecutive registers, succeed.
5529 if (NumAllocated == NumRegs) {
5530 unsigned RegStart = (i-NumAllocated)+1;
5531 unsigned RegEnd = i+1;
5532 // Mark all of the allocated registers used.
5533 for (unsigned i = RegStart; i != RegEnd; ++i)
5534 Regs.push_back(RegClassRegs[i]);
5536 OpInfo.AssignedRegs = RegsForValue(Regs, *RC->vt_begin(),
5537 OpInfo.ConstraintVT);
5538 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
5543 // Otherwise, we couldn't allocate enough registers for this.
5546 /// visitInlineAsm - Handle a call to an InlineAsm object.
5548 void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) {
5549 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
5551 /// ConstraintOperands - Information about all of the constraints.
5552 SDISelAsmOperandInfoVector ConstraintOperands;
5554 std::set<unsigned> OutputRegs, InputRegs;
5556 TargetLowering::AsmOperandInfoVector TargetConstraints = TLI.ParseConstraints(CS);
5557 bool hasMemory = false;
5559 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
5560 unsigned ResNo = 0; // ResNo - The result number of the next output.
5561 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
5562 ConstraintOperands.push_back(SDISelAsmOperandInfo(TargetConstraints[i]));
5563 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
5565 EVT OpVT = MVT::Other;
5567 // Compute the value type for each operand.
5568 switch (OpInfo.Type) {
5569 case InlineAsm::isOutput:
5570 // Indirect outputs just consume an argument.
5571 if (OpInfo.isIndirect) {
5572 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5576 // The return value of the call is this value. As such, there is no
5577 // corresponding argument.
5578 assert(!CS.getType()->isVoidTy() &&
5580 if (const StructType *STy = dyn_cast<StructType>(CS.getType())) {
5581 OpVT = TLI.getValueType(STy->getElementType(ResNo));
5583 assert(ResNo == 0 && "Asm only has one result!");
5584 OpVT = TLI.getValueType(CS.getType());
5588 case InlineAsm::isInput:
5589 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5591 case InlineAsm::isClobber:
5596 // If this is an input or an indirect output, process the call argument.
5597 // BasicBlocks are labels, currently appearing only in asm's.
5598 if (OpInfo.CallOperandVal) {
5599 if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) {
5600 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]);
5602 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
5605 OpVT = OpInfo.getCallOperandValEVT(*DAG.getContext(), TLI, TD);
5608 OpInfo.ConstraintVT = OpVT;
5610 // Indirect operand accesses access memory.
5611 if (OpInfo.isIndirect)
5614 for (unsigned j = 0, ee = OpInfo.Codes.size(); j != ee; ++j) {
5615 TargetLowering::ConstraintType CType = TLI.getConstraintType(OpInfo.Codes[j]);
5616 if (CType == TargetLowering::C_Memory) {
5624 SDValue Chain, Flag;
5626 // We won't need to flush pending loads if this asm doesn't touch
5627 // memory and is nonvolatile.
5628 if (hasMemory || IA->hasSideEffects())
5631 Chain = DAG.getRoot();
5633 // Second pass over the constraints: compute which constraint option to use
5634 // and assign registers to constraints that want a specific physreg.
5635 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5636 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5638 // If this is an output operand with a matching input operand, look up the
5639 // matching input. If their types mismatch, e.g. one is an integer, the
5640 // other is floating point, or their sizes are different, flag it as an
5642 if (OpInfo.hasMatchingInput()) {
5643 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
5645 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
5646 if ((OpInfo.ConstraintVT.isInteger() !=
5647 Input.ConstraintVT.isInteger()) ||
5648 (OpInfo.ConstraintVT.getSizeInBits() !=
5649 Input.ConstraintVT.getSizeInBits())) {
5650 report_fatal_error("Unsupported asm: input constraint"
5651 " with a matching output constraint of"
5652 " incompatible type!");
5654 Input.ConstraintVT = OpInfo.ConstraintVT;
5658 // Compute the constraint code and ConstraintType to use.
5659 TLI.ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, &DAG);
5661 // If this is a memory input, and if the operand is not indirect, do what we
5662 // need to to provide an address for the memory input.
5663 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5664 !OpInfo.isIndirect) {
5665 assert((OpInfo.isMultipleAlternative || (OpInfo.Type == InlineAsm::isInput)) &&
5666 "Can only indirectify direct input operands!");
5668 // Memory operands really want the address of the value. If we don't have
5669 // an indirect input, put it in the constpool if we can, otherwise spill
5670 // it to a stack slot.
5672 // If the operand is a float, integer, or vector constant, spill to a
5673 // constant pool entry to get its address.
5674 const Value *OpVal = OpInfo.CallOperandVal;
5675 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
5676 isa<ConstantVector>(OpVal)) {
5677 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal),
5678 TLI.getPointerTy());
5680 // Otherwise, create a stack slot and emit a store to it before the
5682 const Type *Ty = OpVal->getType();
5683 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
5684 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(Ty);
5685 MachineFunction &MF = DAG.getMachineFunction();
5686 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
5687 SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
5688 Chain = DAG.getStore(Chain, getCurDebugLoc(),
5689 OpInfo.CallOperand, StackSlot,
5690 MachinePointerInfo::getFixedStack(SSFI),
5692 OpInfo.CallOperand = StackSlot;
5695 // There is no longer a Value* corresponding to this operand.
5696 OpInfo.CallOperandVal = 0;
5698 // It is now an indirect operand.
5699 OpInfo.isIndirect = true;
5702 // If this constraint is for a specific register, allocate it before
5704 if (OpInfo.ConstraintType == TargetLowering::C_Register)
5705 GetRegistersForValue(OpInfo, OutputRegs, InputRegs);
5708 // Second pass - Loop over all of the operands, assigning virtual or physregs
5709 // to register class operands.
5710 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5711 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5713 // C_Register operands have already been allocated, Other/Memory don't need
5715 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
5716 GetRegistersForValue(OpInfo, OutputRegs, InputRegs);
5719 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
5720 std::vector<SDValue> AsmNodeOperands;
5721 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain
5722 AsmNodeOperands.push_back(
5723 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(),
5724 TLI.getPointerTy()));
5726 // If we have a !srcloc metadata node associated with it, we want to attach
5727 // this to the ultimately generated inline asm machineinstr. To do this, we
5728 // pass in the third operand as this (potentially null) inline asm MDNode.
5729 const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc");
5730 AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc));
5732 // Remember the AlignStack bit as operand 3.
5733 AsmNodeOperands.push_back(DAG.getTargetConstant(IA->isAlignStack() ? 1 : 0,
5736 // Loop over all of the inputs, copying the operand values into the
5737 // appropriate registers and processing the output regs.
5738 RegsForValue RetValRegs;
5740 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
5741 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
5743 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5744 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5746 switch (OpInfo.Type) {
5747 case InlineAsm::isOutput: {
5748 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
5749 OpInfo.ConstraintType != TargetLowering::C_Register) {
5750 // Memory output, or 'other' output (e.g. 'X' constraint).
5751 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
5753 // Add information to the INLINEASM node to know about this output.
5754 unsigned OpFlags = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
5755 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlags,
5756 TLI.getPointerTy()));
5757 AsmNodeOperands.push_back(OpInfo.CallOperand);
5761 // Otherwise, this is a register or register class output.
5763 // Copy the output from the appropriate register. Find a register that
5765 if (OpInfo.AssignedRegs.Regs.empty())
5766 report_fatal_error("Couldn't allocate output reg for constraint '" +
5767 Twine(OpInfo.ConstraintCode) + "'!");
5769 // If this is an indirect operand, store through the pointer after the
5771 if (OpInfo.isIndirect) {
5772 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
5773 OpInfo.CallOperandVal));
5775 // This is the result value of the call.
5776 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
5777 // Concatenate this output onto the outputs list.
5778 RetValRegs.append(OpInfo.AssignedRegs);
5781 // Add information to the INLINEASM node to know that this register is
5783 OpInfo.AssignedRegs.AddInlineAsmOperands(OpInfo.isEarlyClobber ?
5784 InlineAsm::Kind_RegDefEarlyClobber :
5785 InlineAsm::Kind_RegDef,
5792 case InlineAsm::isInput: {
5793 SDValue InOperandVal = OpInfo.CallOperand;
5795 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint?
5796 // If this is required to match an output register we have already set,
5797 // just use its register.
5798 unsigned OperandNo = OpInfo.getMatchedOperand();
5800 // Scan until we find the definition we already emitted of this operand.
5801 // When we find it, create a RegsForValue operand.
5802 unsigned CurOp = InlineAsm::Op_FirstOperand;
5803 for (; OperandNo; --OperandNo) {
5804 // Advance to the next operand.
5806 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
5807 assert((InlineAsm::isRegDefKind(OpFlag) ||
5808 InlineAsm::isRegDefEarlyClobberKind(OpFlag) ||
5809 InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?");
5810 CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1;
5814 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
5815 if (InlineAsm::isRegDefKind(OpFlag) ||
5816 InlineAsm::isRegDefEarlyClobberKind(OpFlag)) {
5817 // Add (OpFlag&0xffff)>>3 registers to MatchedRegs.
5818 if (OpInfo.isIndirect) {
5819 // This happens on gcc/testsuite/gcc.dg/pr8788-1.c
5820 LLVMContext &Ctx = *DAG.getContext();
5821 Ctx.emitError(CS.getInstruction(), "inline asm not supported yet:"
5822 " don't know how to handle tied "
5823 "indirect register inputs");
5826 RegsForValue MatchedRegs;
5827 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
5828 EVT RegVT = AsmNodeOperands[CurOp+1].getValueType();
5829 MatchedRegs.RegVTs.push_back(RegVT);
5830 MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo();
5831 for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag);
5833 MatchedRegs.Regs.push_back
5834 (RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT)));
5836 // Use the produced MatchedRegs object to
5837 MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
5839 MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse,
5840 true, OpInfo.getMatchedOperand(),
5841 DAG, AsmNodeOperands);
5845 assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!");
5846 assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 &&
5847 "Unexpected number of operands");
5848 // Add information to the INLINEASM node to know about this input.
5849 // See InlineAsm.h isUseOperandTiedToDef.
5850 OpFlag = InlineAsm::getFlagWordForMatchingOp(OpFlag,
5851 OpInfo.getMatchedOperand());
5852 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlag,
5853 TLI.getPointerTy()));
5854 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]);
5858 // Treat indirect 'X' constraint as memory.
5859 if (OpInfo.ConstraintType == TargetLowering::C_Other &&
5861 OpInfo.ConstraintType = TargetLowering::C_Memory;
5863 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
5864 std::vector<SDValue> Ops;
5865 TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode[0],
5868 report_fatal_error("Invalid operand for inline asm constraint '" +
5869 Twine(OpInfo.ConstraintCode) + "'!");
5871 // Add information to the INLINEASM node to know about this input.
5872 unsigned ResOpType =
5873 InlineAsm::getFlagWord(InlineAsm::Kind_Imm, Ops.size());
5874 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
5875 TLI.getPointerTy()));
5876 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
5880 if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
5881 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
5882 assert(InOperandVal.getValueType() == TLI.getPointerTy() &&
5883 "Memory operands expect pointer values");
5885 // Add information to the INLINEASM node to know about this input.
5886 unsigned ResOpType = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
5887 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
5888 TLI.getPointerTy()));
5889 AsmNodeOperands.push_back(InOperandVal);
5893 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
5894 OpInfo.ConstraintType == TargetLowering::C_Register) &&
5895 "Unknown constraint type!");
5896 assert(!OpInfo.isIndirect &&
5897 "Don't know how to handle indirect register inputs yet!");
5899 // Copy the input into the appropriate registers.
5900 if (OpInfo.AssignedRegs.Regs.empty() ||
5901 !OpInfo.AssignedRegs.areValueTypesLegal(TLI))
5902 report_fatal_error("Couldn't allocate input reg for constraint '" +
5903 Twine(OpInfo.ConstraintCode) + "'!");
5905 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
5908 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, false, 0,
5909 DAG, AsmNodeOperands);
5912 case InlineAsm::isClobber: {
5913 // Add the clobbered value to the operand list, so that the register
5914 // allocator is aware that the physreg got clobbered.
5915 if (!OpInfo.AssignedRegs.Regs.empty())
5916 OpInfo.AssignedRegs.AddInlineAsmOperands(
5917 InlineAsm::Kind_RegDefEarlyClobber,
5925 // Finish up input operands. Set the input chain and add the flag last.
5926 AsmNodeOperands[InlineAsm::Op_InputChain] = Chain;
5927 if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
5929 Chain = DAG.getNode(ISD::INLINEASM, getCurDebugLoc(),
5930 DAG.getVTList(MVT::Other, MVT::Flag),
5931 &AsmNodeOperands[0], AsmNodeOperands.size());
5932 Flag = Chain.getValue(1);
5934 // If this asm returns a register value, copy the result from that register
5935 // and set it as the value of the call.
5936 if (!RetValRegs.Regs.empty()) {
5937 SDValue Val = RetValRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(),
5940 // FIXME: Why don't we do this for inline asms with MRVs?
5941 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) {
5942 EVT ResultType = TLI.getValueType(CS.getType());
5944 // If any of the results of the inline asm is a vector, it may have the
5945 // wrong width/num elts. This can happen for register classes that can
5946 // contain multiple different value types. The preg or vreg allocated may
5947 // not have the same VT as was expected. Convert it to the right type
5948 // with bit_convert.
5949 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) {
5950 Val = DAG.getNode(ISD::BITCAST, getCurDebugLoc(),
5953 } else if (ResultType != Val.getValueType() &&
5954 ResultType.isInteger() && Val.getValueType().isInteger()) {
5955 // If a result value was tied to an input value, the computed result may
5956 // have a wider width than the expected result. Extract the relevant
5958 Val = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), ResultType, Val);
5961 assert(ResultType == Val.getValueType() && "Asm result value mismatch!");
5964 setValue(CS.getInstruction(), Val);
5965 // Don't need to use this as a chain in this case.
5966 if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty())
5970 std::vector<std::pair<SDValue, const Value *> > StoresToEmit;
5972 // Process indirect outputs, first output all of the flagged copies out of
5974 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
5975 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
5976 const Value *Ptr = IndirectStoresToEmit[i].second;
5977 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(),
5979 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
5982 // Emit the non-flagged stores from the physregs.
5983 SmallVector<SDValue, 8> OutChains;
5984 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) {
5985 SDValue Val = DAG.getStore(Chain, getCurDebugLoc(),
5986 StoresToEmit[i].first,
5987 getValue(StoresToEmit[i].second),
5988 MachinePointerInfo(StoresToEmit[i].second),
5990 OutChains.push_back(Val);
5993 if (!OutChains.empty())
5994 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
5995 &OutChains[0], OutChains.size());
6000 void SelectionDAGBuilder::visitVAStart(const CallInst &I) {
6001 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurDebugLoc(),
6002 MVT::Other, getRoot(),
6003 getValue(I.getArgOperand(0)),
6004 DAG.getSrcValue(I.getArgOperand(0))));
6007 void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) {
6008 const TargetData &TD = *TLI.getTargetData();
6009 SDValue V = DAG.getVAArg(TLI.getValueType(I.getType()), getCurDebugLoc(),
6010 getRoot(), getValue(I.getOperand(0)),
6011 DAG.getSrcValue(I.getOperand(0)),
6012 TD.getABITypeAlignment(I.getType()));
6014 DAG.setRoot(V.getValue(1));
6017 void SelectionDAGBuilder::visitVAEnd(const CallInst &I) {
6018 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurDebugLoc(),
6019 MVT::Other, getRoot(),
6020 getValue(I.getArgOperand(0)),
6021 DAG.getSrcValue(I.getArgOperand(0))));
6024 void SelectionDAGBuilder::visitVACopy(const CallInst &I) {
6025 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurDebugLoc(),
6026 MVT::Other, getRoot(),
6027 getValue(I.getArgOperand(0)),
6028 getValue(I.getArgOperand(1)),
6029 DAG.getSrcValue(I.getArgOperand(0)),
6030 DAG.getSrcValue(I.getArgOperand(1))));
6033 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
6034 /// implementation, which just calls LowerCall.
6035 /// FIXME: When all targets are
6036 /// migrated to using LowerCall, this hook should be integrated into SDISel.
6037 std::pair<SDValue, SDValue>
6038 TargetLowering::LowerCallTo(SDValue Chain, const Type *RetTy,
6039 bool RetSExt, bool RetZExt, bool isVarArg,
6040 bool isInreg, unsigned NumFixedArgs,
6041 CallingConv::ID CallConv, bool isTailCall,
6042 bool isReturnValueUsed,
6044 ArgListTy &Args, SelectionDAG &DAG,
6045 DebugLoc dl) const {
6046 // Handle all of the outgoing arguments.
6047 SmallVector<ISD::OutputArg, 32> Outs;
6048 SmallVector<SDValue, 32> OutVals;
6049 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
6050 SmallVector<EVT, 4> ValueVTs;
6051 ComputeValueVTs(*this, Args[i].Ty, ValueVTs);
6052 for (unsigned Value = 0, NumValues = ValueVTs.size();
6053 Value != NumValues; ++Value) {
6054 EVT VT = ValueVTs[Value];
6055 const Type *ArgTy = VT.getTypeForEVT(RetTy->getContext());
6056 SDValue Op = SDValue(Args[i].Node.getNode(),
6057 Args[i].Node.getResNo() + Value);
6058 ISD::ArgFlagsTy Flags;
6059 unsigned OriginalAlignment =
6060 getTargetData()->getABITypeAlignment(ArgTy);
6066 if (Args[i].isInReg)
6070 if (Args[i].isByVal) {
6072 const PointerType *Ty = cast<PointerType>(Args[i].Ty);
6073 const Type *ElementTy = Ty->getElementType();
6074 unsigned FrameAlign = getByValTypeAlignment(ElementTy);
6075 unsigned FrameSize = getTargetData()->getTypeAllocSize(ElementTy);
6076 // For ByVal, alignment should come from FE. BE will guess if this
6077 // info is not there but there are cases it cannot get right.
6078 if (Args[i].Alignment)
6079 FrameAlign = Args[i].Alignment;
6080 Flags.setByValAlign(FrameAlign);
6081 Flags.setByValSize(FrameSize);
6085 Flags.setOrigAlign(OriginalAlignment);
6087 EVT PartVT = getRegisterType(RetTy->getContext(), VT);
6088 unsigned NumParts = getNumRegisters(RetTy->getContext(), VT);
6089 SmallVector<SDValue, 4> Parts(NumParts);
6090 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
6093 ExtendKind = ISD::SIGN_EXTEND;
6094 else if (Args[i].isZExt)
6095 ExtendKind = ISD::ZERO_EXTEND;
6097 getCopyToParts(DAG, dl, Op, &Parts[0], NumParts,
6098 PartVT, ExtendKind);
6100 for (unsigned j = 0; j != NumParts; ++j) {
6101 // if it isn't first piece, alignment must be 1
6102 ISD::OutputArg MyFlags(Flags, Parts[j].getValueType(),
6104 if (NumParts > 1 && j == 0)
6105 MyFlags.Flags.setSplit();
6107 MyFlags.Flags.setOrigAlign(1);
6109 Outs.push_back(MyFlags);
6110 OutVals.push_back(Parts[j]);
6115 // Handle the incoming return values from the call.
6116 SmallVector<ISD::InputArg, 32> Ins;
6117 SmallVector<EVT, 4> RetTys;
6118 ComputeValueVTs(*this, RetTy, RetTys);
6119 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
6121 EVT RegisterVT = getRegisterType(RetTy->getContext(), VT);
6122 unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT);
6123 for (unsigned i = 0; i != NumRegs; ++i) {
6124 ISD::InputArg MyFlags;
6125 MyFlags.VT = RegisterVT.getSimpleVT();
6126 MyFlags.Used = isReturnValueUsed;
6128 MyFlags.Flags.setSExt();
6130 MyFlags.Flags.setZExt();
6132 MyFlags.Flags.setInReg();
6133 Ins.push_back(MyFlags);
6137 SmallVector<SDValue, 4> InVals;
6138 Chain = LowerCall(Chain, Callee, CallConv, isVarArg, isTailCall,
6139 Outs, OutVals, Ins, dl, DAG, InVals);
6141 // Verify that the target's LowerCall behaved as expected.
6142 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
6143 "LowerCall didn't return a valid chain!");
6144 assert((!isTailCall || InVals.empty()) &&
6145 "LowerCall emitted a return value for a tail call!");
6146 assert((isTailCall || InVals.size() == Ins.size()) &&
6147 "LowerCall didn't emit the correct number of values!");
6149 // For a tail call, the return value is merely live-out and there aren't
6150 // any nodes in the DAG representing it. Return a special value to
6151 // indicate that a tail call has been emitted and no more Instructions
6152 // should be processed in the current block.
6155 return std::make_pair(SDValue(), SDValue());
6158 DEBUG(for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
6159 assert(InVals[i].getNode() &&
6160 "LowerCall emitted a null value!");
6161 assert(EVT(Ins[i].VT) == InVals[i].getValueType() &&
6162 "LowerCall emitted a value with the wrong type!");
6165 // Collect the legal value parts into potentially illegal values
6166 // that correspond to the original function's return values.
6167 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6169 AssertOp = ISD::AssertSext;
6171 AssertOp = ISD::AssertZext;
6172 SmallVector<SDValue, 4> ReturnValues;
6173 unsigned CurReg = 0;
6174 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
6176 EVT RegisterVT = getRegisterType(RetTy->getContext(), VT);
6177 unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT);
6179 ReturnValues.push_back(getCopyFromParts(DAG, dl, &InVals[CurReg],
6180 NumRegs, RegisterVT, VT,
6185 // For a function returning void, there is no return value. We can't create
6186 // such a node, so we just return a null return value in that case. In
6187 // that case, nothing will actualy look at the value.
6188 if (ReturnValues.empty())
6189 return std::make_pair(SDValue(), Chain);
6191 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, dl,
6192 DAG.getVTList(&RetTys[0], RetTys.size()),
6193 &ReturnValues[0], ReturnValues.size());
6194 return std::make_pair(Res, Chain);
6197 void TargetLowering::LowerOperationWrapper(SDNode *N,
6198 SmallVectorImpl<SDValue> &Results,
6199 SelectionDAG &DAG) const {
6200 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
6202 Results.push_back(Res);
6205 SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6206 llvm_unreachable("LowerOperation not implemented for this target!");
6211 SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) {
6212 SDValue Op = getNonRegisterValue(V);
6213 assert((Op.getOpcode() != ISD::CopyFromReg ||
6214 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
6215 "Copy from a reg to the same reg!");
6216 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg");
6218 RegsForValue RFV(V->getContext(), TLI, Reg, V->getType());
6219 SDValue Chain = DAG.getEntryNode();
6220 RFV.getCopyToRegs(Op, DAG, getCurDebugLoc(), Chain, 0);
6221 PendingExports.push_back(Chain);
6224 #include "llvm/CodeGen/SelectionDAGISel.h"
6226 void SelectionDAGISel::LowerArguments(const BasicBlock *LLVMBB) {
6227 // If this is the entry block, emit arguments.
6228 const Function &F = *LLVMBB->getParent();
6229 SelectionDAG &DAG = SDB->DAG;
6230 DebugLoc dl = SDB->getCurDebugLoc();
6231 const TargetData *TD = TLI.getTargetData();
6232 SmallVector<ISD::InputArg, 16> Ins;
6234 // Check whether the function can return without sret-demotion.
6235 SmallVector<ISD::OutputArg, 4> Outs;
6236 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
6239 if (!FuncInfo->CanLowerReturn) {
6240 // Put in an sret pointer parameter before all the other parameters.
6241 SmallVector<EVT, 1> ValueVTs;
6242 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
6244 // NOTE: Assuming that a pointer will never break down to more than one VT
6246 ISD::ArgFlagsTy Flags;
6248 EVT RegisterVT = TLI.getRegisterType(*DAG.getContext(), ValueVTs[0]);
6249 ISD::InputArg RetArg(Flags, RegisterVT, true);
6250 Ins.push_back(RetArg);
6253 // Set up the incoming argument description vector.
6255 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end();
6256 I != E; ++I, ++Idx) {
6257 SmallVector<EVT, 4> ValueVTs;
6258 ComputeValueVTs(TLI, I->getType(), ValueVTs);
6259 bool isArgValueUsed = !I->use_empty();
6260 for (unsigned Value = 0, NumValues = ValueVTs.size();
6261 Value != NumValues; ++Value) {
6262 EVT VT = ValueVTs[Value];
6263 const Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
6264 ISD::ArgFlagsTy Flags;
6265 unsigned OriginalAlignment =
6266 TD->getABITypeAlignment(ArgTy);
6268 if (F.paramHasAttr(Idx, Attribute::ZExt))
6270 if (F.paramHasAttr(Idx, Attribute::SExt))
6272 if (F.paramHasAttr(Idx, Attribute::InReg))
6274 if (F.paramHasAttr(Idx, Attribute::StructRet))
6276 if (F.paramHasAttr(Idx, Attribute::ByVal)) {
6278 const PointerType *Ty = cast<PointerType>(I->getType());
6279 const Type *ElementTy = Ty->getElementType();
6280 unsigned FrameAlign = TLI.getByValTypeAlignment(ElementTy);
6281 unsigned FrameSize = TD->getTypeAllocSize(ElementTy);
6282 // For ByVal, alignment should be passed from FE. BE will guess if
6283 // this info is not there but there are cases it cannot get right.
6284 if (F.getParamAlignment(Idx))
6285 FrameAlign = F.getParamAlignment(Idx);
6286 Flags.setByValAlign(FrameAlign);
6287 Flags.setByValSize(FrameSize);
6289 if (F.paramHasAttr(Idx, Attribute::Nest))
6291 Flags.setOrigAlign(OriginalAlignment);
6293 EVT RegisterVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6294 unsigned NumRegs = TLI.getNumRegisters(*CurDAG->getContext(), VT);
6295 for (unsigned i = 0; i != NumRegs; ++i) {
6296 ISD::InputArg MyFlags(Flags, RegisterVT, isArgValueUsed);
6297 if (NumRegs > 1 && i == 0)
6298 MyFlags.Flags.setSplit();
6299 // if it isn't first piece, alignment must be 1
6301 MyFlags.Flags.setOrigAlign(1);
6302 Ins.push_back(MyFlags);
6307 // Call the target to set up the argument values.
6308 SmallVector<SDValue, 8> InVals;
6309 SDValue NewRoot = TLI.LowerFormalArguments(DAG.getRoot(), F.getCallingConv(),
6313 // Verify that the target's LowerFormalArguments behaved as expected.
6314 assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other &&
6315 "LowerFormalArguments didn't return a valid chain!");
6316 assert(InVals.size() == Ins.size() &&
6317 "LowerFormalArguments didn't emit the correct number of values!");
6319 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
6320 assert(InVals[i].getNode() &&
6321 "LowerFormalArguments emitted a null value!");
6322 assert(EVT(Ins[i].VT) == InVals[i].getValueType() &&
6323 "LowerFormalArguments emitted a value with the wrong type!");
6327 // Update the DAG with the new chain value resulting from argument lowering.
6328 DAG.setRoot(NewRoot);
6330 // Set up the argument values.
6333 if (!FuncInfo->CanLowerReturn) {
6334 // Create a virtual register for the sret pointer, and put in a copy
6335 // from the sret argument into it.
6336 SmallVector<EVT, 1> ValueVTs;
6337 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
6338 EVT VT = ValueVTs[0];
6339 EVT RegVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6340 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6341 SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1,
6342 RegVT, VT, AssertOp);
6344 MachineFunction& MF = SDB->DAG.getMachineFunction();
6345 MachineRegisterInfo& RegInfo = MF.getRegInfo();
6346 unsigned SRetReg = RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT));
6347 FuncInfo->DemoteRegister = SRetReg;
6348 NewRoot = SDB->DAG.getCopyToReg(NewRoot, SDB->getCurDebugLoc(),
6350 DAG.setRoot(NewRoot);
6352 // i indexes lowered arguments. Bump it past the hidden sret argument.
6353 // Idx indexes LLVM arguments. Don't touch it.
6357 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
6359 SmallVector<SDValue, 4> ArgValues;
6360 SmallVector<EVT, 4> ValueVTs;
6361 ComputeValueVTs(TLI, I->getType(), ValueVTs);
6362 unsigned NumValues = ValueVTs.size();
6364 // If this argument is unused then remember its value. It is used to generate
6365 // debugging information.
6366 if (I->use_empty() && NumValues)
6367 SDB->setUnusedArgValue(I, InVals[i]);
6369 for (unsigned Value = 0; Value != NumValues; ++Value) {
6370 EVT VT = ValueVTs[Value];
6371 EVT PartVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6372 unsigned NumParts = TLI.getNumRegisters(*CurDAG->getContext(), VT);
6374 if (!I->use_empty()) {
6375 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6376 if (F.paramHasAttr(Idx, Attribute::SExt))
6377 AssertOp = ISD::AssertSext;
6378 else if (F.paramHasAttr(Idx, Attribute::ZExt))
6379 AssertOp = ISD::AssertZext;
6381 ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i],
6382 NumParts, PartVT, VT,
6389 // Note down frame index for byval arguments.
6390 if (I->hasByValAttr() && !ArgValues.empty())
6391 if (FrameIndexSDNode *FI =
6392 dyn_cast<FrameIndexSDNode>(ArgValues[0].getNode()))
6393 FuncInfo->setByValArgumentFrameIndex(I, FI->getIndex());
6395 if (!I->use_empty()) {
6397 if (!ArgValues.empty())
6398 Res = DAG.getMergeValues(&ArgValues[0], NumValues,
6399 SDB->getCurDebugLoc());
6400 SDB->setValue(I, Res);
6402 // If this argument is live outside of the entry block, insert a copy from
6403 // whereever we got it to the vreg that other BB's will reference it as.
6404 SDB->CopyToExportRegsIfNeeded(I);
6408 assert(i == InVals.size() && "Argument register count mismatch!");
6410 // Finally, if the target has anything special to do, allow it to do so.
6411 // FIXME: this should insert code into the DAG!
6412 EmitFunctionEntryCode();
6415 /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
6416 /// ensure constants are generated when needed. Remember the virtual registers
6417 /// that need to be added to the Machine PHI nodes as input. We cannot just
6418 /// directly add them, because expansion might result in multiple MBB's for one
6419 /// BB. As such, the start of the BB might correspond to a different MBB than
6423 SelectionDAGBuilder::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
6424 const TerminatorInst *TI = LLVMBB->getTerminator();
6426 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
6428 // Check successor nodes' PHI nodes that expect a constant to be available
6430 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
6431 const BasicBlock *SuccBB = TI->getSuccessor(succ);
6432 if (!isa<PHINode>(SuccBB->begin())) continue;
6433 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
6435 // If this terminator has multiple identical successors (common for
6436 // switches), only handle each succ once.
6437 if (!SuccsHandled.insert(SuccMBB)) continue;
6439 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
6441 // At this point we know that there is a 1-1 correspondence between LLVM PHI
6442 // nodes and Machine PHI nodes, but the incoming operands have not been
6444 for (BasicBlock::const_iterator I = SuccBB->begin();
6445 const PHINode *PN = dyn_cast<PHINode>(I); ++I) {
6446 // Ignore dead phi's.
6447 if (PN->use_empty()) continue;
6450 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
6452 if (const Constant *C = dyn_cast<Constant>(PHIOp)) {
6453 unsigned &RegOut = ConstantsOut[C];
6455 RegOut = FuncInfo.CreateRegs(C->getType());
6456 CopyValueToVirtualRegister(C, RegOut);
6460 DenseMap<const Value *, unsigned>::iterator I =
6461 FuncInfo.ValueMap.find(PHIOp);
6462 if (I != FuncInfo.ValueMap.end())
6465 assert(isa<AllocaInst>(PHIOp) &&
6466 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
6467 "Didn't codegen value into a register!??");
6468 Reg = FuncInfo.CreateRegs(PHIOp->getType());
6469 CopyValueToVirtualRegister(PHIOp, Reg);
6473 // Remember that this register needs to added to the machine PHI node as
6474 // the input for this MBB.
6475 SmallVector<EVT, 4> ValueVTs;
6476 ComputeValueVTs(TLI, PN->getType(), ValueVTs);
6477 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
6478 EVT VT = ValueVTs[vti];
6479 unsigned NumRegisters = TLI.getNumRegisters(*DAG.getContext(), VT);
6480 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
6481 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
6482 Reg += NumRegisters;
6486 ConstantsOut.clear();