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/SmallSet.h"
19 #include "llvm/Analysis/AliasAnalysis.h"
20 #include "llvm/Analysis/ConstantFolding.h"
21 #include "llvm/Constants.h"
22 #include "llvm/CallingConv.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/Function.h"
25 #include "llvm/GlobalVariable.h"
26 #include "llvm/InlineAsm.h"
27 #include "llvm/Instructions.h"
28 #include "llvm/Intrinsics.h"
29 #include "llvm/IntrinsicInst.h"
30 #include "llvm/LLVMContext.h"
31 #include "llvm/Module.h"
32 #include "llvm/CodeGen/Analysis.h"
33 #include "llvm/CodeGen/FastISel.h"
34 #include "llvm/CodeGen/FunctionLoweringInfo.h"
35 #include "llvm/CodeGen/GCStrategy.h"
36 #include "llvm/CodeGen/GCMetadata.h"
37 #include "llvm/CodeGen/MachineFunction.h"
38 #include "llvm/CodeGen/MachineFrameInfo.h"
39 #include "llvm/CodeGen/MachineInstrBuilder.h"
40 #include "llvm/CodeGen/MachineJumpTableInfo.h"
41 #include "llvm/CodeGen/MachineModuleInfo.h"
42 #include "llvm/CodeGen/MachineRegisterInfo.h"
43 #include "llvm/CodeGen/PseudoSourceValue.h"
44 #include "llvm/CodeGen/SelectionDAG.h"
45 #include "llvm/Analysis/DebugInfo.h"
46 #include "llvm/Target/TargetRegisterInfo.h"
47 #include "llvm/Target/TargetData.h"
48 #include "llvm/Target/TargetFrameInfo.h"
49 #include "llvm/Target/TargetInstrInfo.h"
50 #include "llvm/Target/TargetIntrinsicInfo.h"
51 #include "llvm/Target/TargetLowering.h"
52 #include "llvm/Target/TargetOptions.h"
53 #include "llvm/Support/Compiler.h"
54 #include "llvm/Support/CommandLine.h"
55 #include "llvm/Support/Debug.h"
56 #include "llvm/Support/ErrorHandling.h"
57 #include "llvm/Support/MathExtras.h"
58 #include "llvm/Support/raw_ostream.h"
62 /// LimitFloatPrecision - Generate low-precision inline sequences for
63 /// some float libcalls (6, 8 or 12 bits).
64 static unsigned LimitFloatPrecision;
66 static cl::opt<unsigned, true>
67 LimitFPPrecision("limit-float-precision",
68 cl::desc("Generate low-precision inline sequences "
69 "for some float libcalls"),
70 cl::location(LimitFloatPrecision),
73 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, DebugLoc DL,
74 const SDValue *Parts, unsigned NumParts,
75 EVT PartVT, EVT ValueVT);
77 /// getCopyFromParts - Create a value that contains the specified legal parts
78 /// combined into the value they represent. If the parts combine to a type
79 /// larger then ValueVT then AssertOp can be used to specify whether the extra
80 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
81 /// (ISD::AssertSext).
82 static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc DL,
84 unsigned NumParts, EVT PartVT, EVT ValueVT,
85 ISD::NodeType AssertOp = ISD::DELETED_NODE) {
86 if (ValueVT.isVector())
87 return getCopyFromPartsVector(DAG, DL, Parts, NumParts, PartVT, ValueVT);
89 assert(NumParts > 0 && "No parts to assemble!");
90 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
91 SDValue Val = Parts[0];
94 // Assemble the value from multiple parts.
95 if (ValueVT.isInteger()) {
96 unsigned PartBits = PartVT.getSizeInBits();
97 unsigned ValueBits = ValueVT.getSizeInBits();
99 // Assemble the power of 2 part.
100 unsigned RoundParts = NumParts & (NumParts - 1) ?
101 1 << Log2_32(NumParts) : NumParts;
102 unsigned RoundBits = PartBits * RoundParts;
103 EVT RoundVT = RoundBits == ValueBits ?
104 ValueVT : EVT::getIntegerVT(*DAG.getContext(), RoundBits);
107 EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2);
109 if (RoundParts > 2) {
110 Lo = getCopyFromParts(DAG, DL, Parts, RoundParts / 2,
112 Hi = getCopyFromParts(DAG, DL, Parts + RoundParts / 2,
113 RoundParts / 2, PartVT, HalfVT);
115 Lo = DAG.getNode(ISD::BIT_CONVERT, DL, HalfVT, Parts[0]);
116 Hi = DAG.getNode(ISD::BIT_CONVERT, DL, HalfVT, Parts[1]);
119 if (TLI.isBigEndian())
122 Val = DAG.getNode(ISD::BUILD_PAIR, DL, RoundVT, Lo, Hi);
124 if (RoundParts < NumParts) {
125 // Assemble the trailing non-power-of-2 part.
126 unsigned OddParts = NumParts - RoundParts;
127 EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits);
128 Hi = getCopyFromParts(DAG, DL,
129 Parts + RoundParts, OddParts, PartVT, OddVT);
131 // Combine the round and odd parts.
133 if (TLI.isBigEndian())
135 EVT TotalVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
136 Hi = DAG.getNode(ISD::ANY_EXTEND, DL, TotalVT, Hi);
137 Hi = DAG.getNode(ISD::SHL, DL, TotalVT, Hi,
138 DAG.getConstant(Lo.getValueType().getSizeInBits(),
139 TLI.getPointerTy()));
140 Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, TotalVT, Lo);
141 Val = DAG.getNode(ISD::OR, DL, TotalVT, Lo, Hi);
143 } else if (PartVT.isFloatingPoint()) {
144 // FP split into multiple FP parts (for ppcf128)
145 assert(ValueVT == EVT(MVT::ppcf128) && PartVT == EVT(MVT::f64) &&
148 Lo = DAG.getNode(ISD::BIT_CONVERT, DL, EVT(MVT::f64), Parts[0]);
149 Hi = DAG.getNode(ISD::BIT_CONVERT, DL, EVT(MVT::f64), Parts[1]);
150 if (TLI.isBigEndian())
152 Val = DAG.getNode(ISD::BUILD_PAIR, DL, ValueVT, Lo, Hi);
154 // FP split into integer parts (soft fp)
155 assert(ValueVT.isFloatingPoint() && PartVT.isInteger() &&
156 !PartVT.isVector() && "Unexpected split");
157 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits());
158 Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT);
162 // There is now one part, held in Val. Correct it to match ValueVT.
163 PartVT = Val.getValueType();
165 if (PartVT == ValueVT)
168 if (PartVT.isInteger() && ValueVT.isInteger()) {
169 if (ValueVT.bitsLT(PartVT)) {
170 // For a truncate, see if we have any information to
171 // indicate whether the truncated bits will always be
172 // zero or sign-extension.
173 if (AssertOp != ISD::DELETED_NODE)
174 Val = DAG.getNode(AssertOp, DL, PartVT, Val,
175 DAG.getValueType(ValueVT));
176 return DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
178 return DAG.getNode(ISD::ANY_EXTEND, DL, ValueVT, Val);
181 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
182 // FP_ROUND's are always exact here.
183 if (ValueVT.bitsLT(Val.getValueType()))
184 return DAG.getNode(ISD::FP_ROUND, DL, ValueVT, Val,
185 DAG.getIntPtrConstant(1));
187 return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val);
190 if (PartVT.getSizeInBits() == ValueVT.getSizeInBits())
191 return DAG.getNode(ISD::BIT_CONVERT, DL, ValueVT, Val);
193 llvm_unreachable("Unknown mismatch!");
197 /// getCopyFromParts - Create a value that contains the specified legal parts
198 /// combined into the value they represent. If the parts combine to a type
199 /// larger then ValueVT then AssertOp can be used to specify whether the extra
200 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
201 /// (ISD::AssertSext).
202 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, DebugLoc DL,
203 const SDValue *Parts, unsigned NumParts,
204 EVT PartVT, EVT ValueVT) {
205 assert(ValueVT.isVector() && "Not a vector value");
206 assert(NumParts > 0 && "No parts to assemble!");
207 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
208 SDValue Val = Parts[0];
210 // Handle a multi-element vector.
212 EVT IntermediateVT, RegisterVT;
213 unsigned NumIntermediates;
215 TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT,
216 NumIntermediates, RegisterVT);
217 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
218 NumParts = NumRegs; // Silence a compiler warning.
219 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
220 assert(RegisterVT == Parts[0].getValueType() &&
221 "Part type doesn't match part!");
223 // Assemble the parts into intermediate operands.
224 SmallVector<SDValue, 8> Ops(NumIntermediates);
225 if (NumIntermediates == NumParts) {
226 // If the register was not expanded, truncate or copy the value,
228 for (unsigned i = 0; i != NumParts; ++i)
229 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i], 1,
230 PartVT, IntermediateVT);
231 } else if (NumParts > 0) {
232 // If the intermediate type was expanded, build the intermediate
233 // operands from the parts.
234 assert(NumParts % NumIntermediates == 0 &&
235 "Must expand into a divisible number of parts!");
236 unsigned Factor = NumParts / NumIntermediates;
237 for (unsigned i = 0; i != NumIntermediates; ++i)
238 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor,
239 PartVT, IntermediateVT);
242 // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the
243 // intermediate operands.
244 Val = DAG.getNode(IntermediateVT.isVector() ?
245 ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR, DL,
246 ValueVT, &Ops[0], NumIntermediates);
249 // There is now one part, held in Val. Correct it to match ValueVT.
250 PartVT = Val.getValueType();
252 if (PartVT == ValueVT)
255 if (PartVT.isVector()) {
256 // If the element type of the source/dest vectors are the same, but the
257 // parts vector has more elements than the value vector, then we have a
258 // vector widening case (e.g. <2 x float> -> <4 x float>). Extract the
260 if (PartVT.getVectorElementType() == ValueVT.getVectorElementType()) {
261 assert(PartVT.getVectorNumElements() > ValueVT.getVectorNumElements() &&
262 "Cannot narrow, it would be a lossy transformation");
263 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val,
264 DAG.getIntPtrConstant(0));
267 // Vector/Vector bitcast.
268 return DAG.getNode(ISD::BIT_CONVERT, DL, ValueVT, Val);
271 assert(ValueVT.getVectorElementType() == PartVT &&
272 ValueVT.getVectorNumElements() == 1 &&
273 "Only trivial scalar-to-vector conversions should get here!");
274 return DAG.getNode(ISD::BUILD_VECTOR, DL, ValueVT, Val);
280 static void getCopyToPartsVector(SelectionDAG &DAG, DebugLoc dl,
281 SDValue Val, SDValue *Parts, unsigned NumParts,
284 /// getCopyToParts - Create a series of nodes that contain the specified value
285 /// split into legal parts. If the parts contain more bits than Val, then, for
286 /// integers, ExtendKind can be used to specify how to generate the extra bits.
287 static void getCopyToParts(SelectionDAG &DAG, DebugLoc DL,
288 SDValue Val, SDValue *Parts, unsigned NumParts,
290 ISD::NodeType ExtendKind = ISD::ANY_EXTEND) {
291 EVT ValueVT = Val.getValueType();
293 // Handle the vector case separately.
294 if (ValueVT.isVector())
295 return getCopyToPartsVector(DAG, DL, Val, Parts, NumParts, PartVT);
297 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
298 unsigned PartBits = PartVT.getSizeInBits();
299 unsigned OrigNumParts = NumParts;
300 assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!");
305 assert(!ValueVT.isVector() && "Vector case handled elsewhere");
306 if (PartVT == ValueVT) {
307 assert(NumParts == 1 && "No-op copy with multiple parts!");
312 if (NumParts * PartBits > ValueVT.getSizeInBits()) {
313 // If the parts cover more bits than the value has, promote the value.
314 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
315 assert(NumParts == 1 && "Do not know what to promote to!");
316 Val = DAG.getNode(ISD::FP_EXTEND, DL, PartVT, Val);
318 assert(PartVT.isInteger() && ValueVT.isInteger() &&
319 "Unknown mismatch!");
320 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
321 Val = DAG.getNode(ExtendKind, DL, ValueVT, Val);
323 } else if (PartBits == ValueVT.getSizeInBits()) {
324 // Different types of the same size.
325 assert(NumParts == 1 && PartVT != ValueVT);
326 Val = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Val);
327 } else if (NumParts * PartBits < ValueVT.getSizeInBits()) {
328 // If the parts cover less bits than value has, truncate the value.
329 assert(PartVT.isInteger() && ValueVT.isInteger() &&
330 "Unknown mismatch!");
331 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
332 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
335 // The value may have changed - recompute ValueVT.
336 ValueVT = Val.getValueType();
337 assert(NumParts * PartBits == ValueVT.getSizeInBits() &&
338 "Failed to tile the value with PartVT!");
341 assert(PartVT == ValueVT && "Type conversion failed!");
346 // Expand the value into multiple parts.
347 if (NumParts & (NumParts - 1)) {
348 // The number of parts is not a power of 2. Split off and copy the tail.
349 assert(PartVT.isInteger() && ValueVT.isInteger() &&
350 "Do not know what to expand to!");
351 unsigned RoundParts = 1 << Log2_32(NumParts);
352 unsigned RoundBits = RoundParts * PartBits;
353 unsigned OddParts = NumParts - RoundParts;
354 SDValue OddVal = DAG.getNode(ISD::SRL, DL, ValueVT, Val,
355 DAG.getIntPtrConstant(RoundBits));
356 getCopyToParts(DAG, DL, OddVal, Parts + RoundParts, OddParts, PartVT);
358 if (TLI.isBigEndian())
359 // The odd parts were reversed by getCopyToParts - unreverse them.
360 std::reverse(Parts + RoundParts, Parts + NumParts);
362 NumParts = RoundParts;
363 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
364 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
367 // The number of parts is a power of 2. Repeatedly bisect the value using
369 Parts[0] = DAG.getNode(ISD::BIT_CONVERT, DL,
370 EVT::getIntegerVT(*DAG.getContext(),
371 ValueVT.getSizeInBits()),
374 for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) {
375 for (unsigned i = 0; i < NumParts; i += StepSize) {
376 unsigned ThisBits = StepSize * PartBits / 2;
377 EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits);
378 SDValue &Part0 = Parts[i];
379 SDValue &Part1 = Parts[i+StepSize/2];
381 Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
382 ThisVT, Part0, DAG.getIntPtrConstant(1));
383 Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
384 ThisVT, Part0, DAG.getIntPtrConstant(0));
386 if (ThisBits == PartBits && ThisVT != PartVT) {
387 Part0 = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Part0);
388 Part1 = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Part1);
393 if (TLI.isBigEndian())
394 std::reverse(Parts, Parts + OrigNumParts);
398 /// getCopyToPartsVector - Create a series of nodes that contain the specified
399 /// value split into legal parts.
400 static void getCopyToPartsVector(SelectionDAG &DAG, DebugLoc DL,
401 SDValue Val, SDValue *Parts, unsigned NumParts,
403 EVT ValueVT = Val.getValueType();
404 assert(ValueVT.isVector() && "Not a vector");
405 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
408 if (PartVT == ValueVT) {
410 } else if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) {
411 // Bitconvert vector->vector case.
412 Val = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Val);
413 } else if (PartVT.isVector() &&
414 PartVT.getVectorElementType() == ValueVT.getVectorElementType()&&
415 PartVT.getVectorNumElements() > ValueVT.getVectorNumElements()) {
416 EVT ElementVT = PartVT.getVectorElementType();
417 // Vector widening case, e.g. <2 x float> -> <4 x float>. Shuffle in
419 SmallVector<SDValue, 16> Ops;
420 for (unsigned i = 0, e = ValueVT.getVectorNumElements(); i != e; ++i)
421 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
422 ElementVT, Val, DAG.getIntPtrConstant(i)));
424 for (unsigned i = ValueVT.getVectorNumElements(),
425 e = PartVT.getVectorNumElements(); i != e; ++i)
426 Ops.push_back(DAG.getUNDEF(ElementVT));
428 Val = DAG.getNode(ISD::BUILD_VECTOR, DL, PartVT, &Ops[0], Ops.size());
430 // FIXME: Use CONCAT for 2x -> 4x.
432 //SDValue UndefElts = DAG.getUNDEF(VectorTy);
433 //Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts);
435 // Vector -> scalar conversion.
436 assert(ValueVT.getVectorElementType() == PartVT &&
437 ValueVT.getVectorNumElements() == 1 &&
438 "Only trivial vector-to-scalar conversions should get here!");
439 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
440 PartVT, Val, DAG.getIntPtrConstant(0));
447 // Handle a multi-element vector.
448 EVT IntermediateVT, RegisterVT;
449 unsigned NumIntermediates;
450 unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT,
452 NumIntermediates, RegisterVT);
453 unsigned NumElements = ValueVT.getVectorNumElements();
455 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
456 NumParts = NumRegs; // Silence a compiler warning.
457 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
459 // Split the vector into intermediate operands.
460 SmallVector<SDValue, 8> Ops(NumIntermediates);
461 for (unsigned i = 0; i != NumIntermediates; ++i) {
462 if (IntermediateVT.isVector())
463 Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL,
465 DAG.getIntPtrConstant(i * (NumElements / NumIntermediates)));
467 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
468 IntermediateVT, Val, DAG.getIntPtrConstant(i));
471 // Split the intermediate operands into legal parts.
472 if (NumParts == NumIntermediates) {
473 // If the register was not expanded, promote or copy the value,
475 for (unsigned i = 0; i != NumParts; ++i)
476 getCopyToParts(DAG, DL, Ops[i], &Parts[i], 1, PartVT);
477 } else if (NumParts > 0) {
478 // If the intermediate type was expanded, split each the value into
480 assert(NumParts % NumIntermediates == 0 &&
481 "Must expand into a divisible number of parts!");
482 unsigned Factor = NumParts / NumIntermediates;
483 for (unsigned i = 0; i != NumIntermediates; ++i)
484 getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT);
492 /// RegsForValue - This struct represents the registers (physical or virtual)
493 /// that a particular set of values is assigned, and the type information
494 /// about the value. The most common situation is to represent one value at a
495 /// time, but struct or array values are handled element-wise as multiple
496 /// values. The splitting of aggregates is performed recursively, so that we
497 /// never have aggregate-typed registers. The values at this point do not
498 /// necessarily have legal types, so each value may require one or more
499 /// registers of some legal type.
501 struct RegsForValue {
502 /// ValueVTs - The value types of the values, which may not be legal, and
503 /// may need be promoted or synthesized from one or more registers.
505 SmallVector<EVT, 4> ValueVTs;
507 /// RegVTs - The value types of the registers. This is the same size as
508 /// ValueVTs and it records, for each value, what the type of the assigned
509 /// register or registers are. (Individual values are never synthesized
510 /// from more than one type of register.)
512 /// With virtual registers, the contents of RegVTs is redundant with TLI's
513 /// getRegisterType member function, however when with physical registers
514 /// it is necessary to have a separate record of the types.
516 SmallVector<EVT, 4> RegVTs;
518 /// Regs - This list holds the registers assigned to the values.
519 /// Each legal or promoted value requires one register, and each
520 /// expanded value requires multiple registers.
522 SmallVector<unsigned, 4> Regs;
526 RegsForValue(const SmallVector<unsigned, 4> ®s,
527 EVT regvt, EVT valuevt)
528 : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {}
530 RegsForValue(LLVMContext &Context, const TargetLowering &tli,
531 unsigned Reg, const Type *Ty) {
532 ComputeValueVTs(tli, Ty, ValueVTs);
534 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
535 EVT ValueVT = ValueVTs[Value];
536 unsigned NumRegs = tli.getNumRegisters(Context, ValueVT);
537 EVT RegisterVT = tli.getRegisterType(Context, ValueVT);
538 for (unsigned i = 0; i != NumRegs; ++i)
539 Regs.push_back(Reg + i);
540 RegVTs.push_back(RegisterVT);
545 /// areValueTypesLegal - Return true if types of all the values are legal.
546 bool areValueTypesLegal(const TargetLowering &TLI) {
547 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
548 EVT RegisterVT = RegVTs[Value];
549 if (!TLI.isTypeLegal(RegisterVT))
555 /// append - Add the specified values to this one.
556 void append(const RegsForValue &RHS) {
557 ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end());
558 RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end());
559 Regs.append(RHS.Regs.begin(), RHS.Regs.end());
562 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
563 /// this value and returns the result as a ValueVTs value. This uses
564 /// Chain/Flag as the input and updates them for the output Chain/Flag.
565 /// If the Flag pointer is NULL, no flag is used.
566 SDValue getCopyFromRegs(SelectionDAG &DAG, FunctionLoweringInfo &FuncInfo,
568 SDValue &Chain, SDValue *Flag) const;
570 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
571 /// specified value into the registers specified by this object. This uses
572 /// Chain/Flag as the input and updates them for the output Chain/Flag.
573 /// If the Flag pointer is NULL, no flag is used.
574 void getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
575 SDValue &Chain, SDValue *Flag) const;
577 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
578 /// operand list. This adds the code marker, matching input operand index
579 /// (if applicable), and includes the number of values added into it.
580 void AddInlineAsmOperands(unsigned Kind,
581 bool HasMatching, unsigned MatchingIdx,
583 std::vector<SDValue> &Ops) const;
587 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
588 /// this value and returns the result as a ValueVT value. This uses
589 /// Chain/Flag as the input and updates them for the output Chain/Flag.
590 /// If the Flag pointer is NULL, no flag is used.
591 SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
592 FunctionLoweringInfo &FuncInfo,
594 SDValue &Chain, SDValue *Flag) const {
595 // A Value with type {} or [0 x %t] needs no registers.
596 if (ValueVTs.empty())
599 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
601 // Assemble the legal parts into the final values.
602 SmallVector<SDValue, 4> Values(ValueVTs.size());
603 SmallVector<SDValue, 8> Parts;
604 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
605 // Copy the legal parts from the registers.
606 EVT ValueVT = ValueVTs[Value];
607 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
608 EVT RegisterVT = RegVTs[Value];
610 Parts.resize(NumRegs);
611 for (unsigned i = 0; i != NumRegs; ++i) {
614 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT);
616 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag);
617 *Flag = P.getValue(2);
620 Chain = P.getValue(1);
622 // If the source register was virtual and if we know something about it,
623 // add an assert node.
624 if (TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) &&
625 RegisterVT.isInteger() && !RegisterVT.isVector()) {
626 unsigned SlotNo = Regs[Part+i]-TargetRegisterInfo::FirstVirtualRegister;
627 if (FuncInfo.LiveOutRegInfo.size() > SlotNo) {
628 const FunctionLoweringInfo::LiveOutInfo &LOI =
629 FuncInfo.LiveOutRegInfo[SlotNo];
631 unsigned RegSize = RegisterVT.getSizeInBits();
632 unsigned NumSignBits = LOI.NumSignBits;
633 unsigned NumZeroBits = LOI.KnownZero.countLeadingOnes();
635 // FIXME: We capture more information than the dag can represent. For
636 // now, just use the tightest assertzext/assertsext possible.
638 EVT FromVT(MVT::Other);
639 if (NumSignBits == RegSize)
640 isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1
641 else if (NumZeroBits >= RegSize-1)
642 isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1
643 else if (NumSignBits > RegSize-8)
644 isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8
645 else if (NumZeroBits >= RegSize-8)
646 isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8
647 else if (NumSignBits > RegSize-16)
648 isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16
649 else if (NumZeroBits >= RegSize-16)
650 isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16
651 else if (NumSignBits > RegSize-32)
652 isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32
653 else if (NumZeroBits >= RegSize-32)
654 isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32
656 if (FromVT != MVT::Other)
657 P = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl,
658 RegisterVT, P, DAG.getValueType(FromVT));
665 Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(),
666 NumRegs, RegisterVT, ValueVT);
671 return DAG.getNode(ISD::MERGE_VALUES, dl,
672 DAG.getVTList(&ValueVTs[0], ValueVTs.size()),
673 &Values[0], ValueVTs.size());
676 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
677 /// specified value into the registers specified by this object. This uses
678 /// Chain/Flag as the input and updates them for the output Chain/Flag.
679 /// If the Flag pointer is NULL, no flag is used.
680 void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
681 SDValue &Chain, SDValue *Flag) const {
682 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
684 // Get the list of the values's legal parts.
685 unsigned NumRegs = Regs.size();
686 SmallVector<SDValue, 8> Parts(NumRegs);
687 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
688 EVT ValueVT = ValueVTs[Value];
689 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
690 EVT RegisterVT = RegVTs[Value];
692 getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value),
693 &Parts[Part], NumParts, RegisterVT);
697 // Copy the parts into the registers.
698 SmallVector<SDValue, 8> Chains(NumRegs);
699 for (unsigned i = 0; i != NumRegs; ++i) {
702 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]);
704 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag);
705 *Flag = Part.getValue(1);
708 Chains[i] = Part.getValue(0);
711 if (NumRegs == 1 || Flag)
712 // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is
713 // flagged to it. That is the CopyToReg nodes and the user are considered
714 // a single scheduling unit. If we create a TokenFactor and return it as
715 // chain, then the TokenFactor is both a predecessor (operand) of the
716 // user as well as a successor (the TF operands are flagged to the user).
717 // c1, f1 = CopyToReg
718 // c2, f2 = CopyToReg
719 // c3 = TokenFactor c1, c2
722 Chain = Chains[NumRegs-1];
724 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0], NumRegs);
727 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
728 /// operand list. This adds the code marker and includes the number of
729 /// values added into it.
730 void RegsForValue::AddInlineAsmOperands(unsigned Code, bool HasMatching,
731 unsigned MatchingIdx,
733 std::vector<SDValue> &Ops) const {
734 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
736 unsigned Flag = InlineAsm::getFlagWord(Code, Regs.size());
738 Flag = InlineAsm::getFlagWordForMatchingOp(Flag, MatchingIdx);
739 SDValue Res = DAG.getTargetConstant(Flag, MVT::i32);
742 for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) {
743 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVTs[Value]);
744 EVT RegisterVT = RegVTs[Value];
745 for (unsigned i = 0; i != NumRegs; ++i) {
746 assert(Reg < Regs.size() && "Mismatch in # registers expected");
747 Ops.push_back(DAG.getRegister(Regs[Reg++], RegisterVT));
752 void SelectionDAGBuilder::init(GCFunctionInfo *gfi, AliasAnalysis &aa) {
755 TD = DAG.getTarget().getTargetData();
758 /// clear - Clear out the current SelectionDAG and the associated
759 /// state and prepare this SelectionDAGBuilder object to be used
760 /// for a new block. This doesn't clear out information about
761 /// additional blocks that are needed to complete switch lowering
762 /// or PHI node updating; that information is cleared out as it is
764 void SelectionDAGBuilder::clear() {
766 UnusedArgNodeMap.clear();
767 PendingLoads.clear();
768 PendingExports.clear();
769 DanglingDebugInfoMap.clear();
770 CurDebugLoc = DebugLoc();
774 /// getRoot - Return the current virtual root of the Selection DAG,
775 /// flushing any PendingLoad items. This must be done before emitting
776 /// a store or any other node that may need to be ordered after any
777 /// prior load instructions.
779 SDValue SelectionDAGBuilder::getRoot() {
780 if (PendingLoads.empty())
781 return DAG.getRoot();
783 if (PendingLoads.size() == 1) {
784 SDValue Root = PendingLoads[0];
786 PendingLoads.clear();
790 // Otherwise, we have to make a token factor node.
791 SDValue Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
792 &PendingLoads[0], PendingLoads.size());
793 PendingLoads.clear();
798 /// getControlRoot - Similar to getRoot, but instead of flushing all the
799 /// PendingLoad items, flush all the PendingExports items. It is necessary
800 /// to do this before emitting a terminator instruction.
802 SDValue SelectionDAGBuilder::getControlRoot() {
803 SDValue Root = DAG.getRoot();
805 if (PendingExports.empty())
808 // Turn all of the CopyToReg chains into one factored node.
809 if (Root.getOpcode() != ISD::EntryToken) {
810 unsigned i = 0, e = PendingExports.size();
811 for (; i != e; ++i) {
812 assert(PendingExports[i].getNode()->getNumOperands() > 1);
813 if (PendingExports[i].getNode()->getOperand(0) == Root)
814 break; // Don't add the root if we already indirectly depend on it.
818 PendingExports.push_back(Root);
821 Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
823 PendingExports.size());
824 PendingExports.clear();
829 void SelectionDAGBuilder::AssignOrderingToNode(const SDNode *Node) {
830 if (DAG.GetOrdering(Node) != 0) return; // Already has ordering.
831 DAG.AssignOrdering(Node, SDNodeOrder);
833 for (unsigned I = 0, E = Node->getNumOperands(); I != E; ++I)
834 AssignOrderingToNode(Node->getOperand(I).getNode());
837 void SelectionDAGBuilder::visit(const Instruction &I) {
838 // Set up outgoing PHI node register values before emitting the terminator.
839 if (isa<TerminatorInst>(&I))
840 HandlePHINodesInSuccessorBlocks(I.getParent());
842 CurDebugLoc = I.getDebugLoc();
844 visit(I.getOpcode(), I);
846 if (!isa<TerminatorInst>(&I) && !HasTailCall)
847 CopyToExportRegsIfNeeded(&I);
849 CurDebugLoc = DebugLoc();
852 void SelectionDAGBuilder::visitPHI(const PHINode &) {
853 llvm_unreachable("SelectionDAGBuilder shouldn't visit PHI nodes!");
856 void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) {
857 // Note: this doesn't use InstVisitor, because it has to work with
858 // ConstantExpr's in addition to instructions.
860 default: llvm_unreachable("Unknown instruction type encountered!");
861 // Build the switch statement using the Instruction.def file.
862 #define HANDLE_INST(NUM, OPCODE, CLASS) \
863 case Instruction::OPCODE: visit##OPCODE((CLASS&)I); break;
864 #include "llvm/Instruction.def"
867 // Assign the ordering to the freshly created DAG nodes.
868 if (NodeMap.count(&I)) {
870 AssignOrderingToNode(getValue(&I).getNode());
874 // resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V,
875 // generate the debug data structures now that we've seen its definition.
876 void SelectionDAGBuilder::resolveDanglingDebugInfo(const Value *V,
878 DanglingDebugInfo &DDI = DanglingDebugInfoMap[V];
879 MDNode *Variable = NULL;
882 if (const DbgValueInst *DI = dyn_cast_or_null<DbgValueInst>(DDI.getDI())) {
883 Variable = DI->getVariable();
884 Offset = DI->getOffset();
885 } else if (const DbgDeclareInst *DI =
886 dyn_cast_or_null<DbgDeclareInst>(DDI.getDI()))
887 Variable = DI->getVariable();
889 assert (DDI.getDI() == NULL && "Invalid debug info intrinsic!");
894 DebugLoc dl = DDI.getdl();
895 unsigned DbgSDNodeOrder = DDI.getSDNodeOrder();
898 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, Val)) {
899 SDV = DAG.getDbgValue(Variable, Val.getNode(),
900 Val.getResNo(), Offset, dl, DbgSDNodeOrder);
901 DAG.AddDbgValue(SDV, Val.getNode(), false);
904 SDV = DAG.getDbgValue(Variable, UndefValue::get(V->getType()),
905 Offset, dl, SDNodeOrder);
906 DAG.AddDbgValue(SDV, 0, false);
908 DanglingDebugInfoMap[V] = DanglingDebugInfo();
912 // getValue - Return an SDValue for the given Value.
913 SDValue SelectionDAGBuilder::getValue(const Value *V) {
914 // If we already have an SDValue for this value, use it. It's important
915 // to do this first, so that we don't create a CopyFromReg if we already
916 // have a regular SDValue.
917 SDValue &N = NodeMap[V];
918 if (N.getNode()) return N;
920 // If there's a virtual register allocated and initialized for this
922 DenseMap<const Value *, unsigned>::iterator It = FuncInfo.ValueMap.find(V);
923 if (It != FuncInfo.ValueMap.end()) {
924 unsigned InReg = It->second;
925 RegsForValue RFV(*DAG.getContext(), TLI, InReg, V->getType());
926 SDValue Chain = DAG.getEntryNode();
927 return N = RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain,NULL);
930 // Otherwise create a new SDValue and remember it.
931 SDValue Val = getValueImpl(V);
933 resolveDanglingDebugInfo(V, Val);
937 /// getNonRegisterValue - Return an SDValue for the given Value, but
938 /// don't look in FuncInfo.ValueMap for a virtual register.
939 SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) {
940 // If we already have an SDValue for this value, use it.
941 SDValue &N = NodeMap[V];
942 if (N.getNode()) return N;
944 // Otherwise create a new SDValue and remember it.
945 SDValue Val = getValueImpl(V);
947 resolveDanglingDebugInfo(V, Val);
951 /// getValueImpl - Helper function for getValue and getNonRegisterValue.
952 /// Create an SDValue for the given value.
953 SDValue SelectionDAGBuilder::getValueImpl(const Value *V) {
954 if (const Constant *C = dyn_cast<Constant>(V)) {
955 EVT VT = TLI.getValueType(V->getType(), true);
957 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C))
958 return DAG.getConstant(*CI, VT);
960 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
961 return DAG.getGlobalAddress(GV, getCurDebugLoc(), VT);
963 if (isa<ConstantPointerNull>(C))
964 return DAG.getConstant(0, TLI.getPointerTy());
966 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
967 return DAG.getConstantFP(*CFP, VT);
969 if (isa<UndefValue>(C) && !V->getType()->isAggregateType())
970 return DAG.getUNDEF(VT);
972 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
973 visit(CE->getOpcode(), *CE);
974 SDValue N1 = NodeMap[V];
975 assert(N1.getNode() && "visit didn't populate the NodeMap!");
979 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) {
980 SmallVector<SDValue, 4> Constants;
981 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end();
983 SDNode *Val = getValue(*OI).getNode();
984 // If the operand is an empty aggregate, there are no values.
986 // Add each leaf value from the operand to the Constants list
987 // to form a flattened list of all the values.
988 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
989 Constants.push_back(SDValue(Val, i));
992 return DAG.getMergeValues(&Constants[0], Constants.size(),
996 if (C->getType()->isStructTy() || C->getType()->isArrayTy()) {
997 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
998 "Unknown struct or array constant!");
1000 SmallVector<EVT, 4> ValueVTs;
1001 ComputeValueVTs(TLI, C->getType(), ValueVTs);
1002 unsigned NumElts = ValueVTs.size();
1004 return SDValue(); // empty struct
1005 SmallVector<SDValue, 4> Constants(NumElts);
1006 for (unsigned i = 0; i != NumElts; ++i) {
1007 EVT EltVT = ValueVTs[i];
1008 if (isa<UndefValue>(C))
1009 Constants[i] = DAG.getUNDEF(EltVT);
1010 else if (EltVT.isFloatingPoint())
1011 Constants[i] = DAG.getConstantFP(0, EltVT);
1013 Constants[i] = DAG.getConstant(0, EltVT);
1016 return DAG.getMergeValues(&Constants[0], NumElts,
1020 if (const BlockAddress *BA = dyn_cast<BlockAddress>(C))
1021 return DAG.getBlockAddress(BA, VT);
1023 const VectorType *VecTy = cast<VectorType>(V->getType());
1024 unsigned NumElements = VecTy->getNumElements();
1026 // Now that we know the number and type of the elements, get that number of
1027 // elements into the Ops array based on what kind of constant it is.
1028 SmallVector<SDValue, 16> Ops;
1029 if (const ConstantVector *CP = dyn_cast<ConstantVector>(C)) {
1030 for (unsigned i = 0; i != NumElements; ++i)
1031 Ops.push_back(getValue(CP->getOperand(i)));
1033 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!");
1034 EVT EltVT = TLI.getValueType(VecTy->getElementType());
1037 if (EltVT.isFloatingPoint())
1038 Op = DAG.getConstantFP(0, EltVT);
1040 Op = DAG.getConstant(0, EltVT);
1041 Ops.assign(NumElements, Op);
1044 // Create a BUILD_VECTOR node.
1045 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
1046 VT, &Ops[0], Ops.size());
1049 // If this is a static alloca, generate it as the frameindex instead of
1051 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
1052 DenseMap<const AllocaInst*, int>::iterator SI =
1053 FuncInfo.StaticAllocaMap.find(AI);
1054 if (SI != FuncInfo.StaticAllocaMap.end())
1055 return DAG.getFrameIndex(SI->second, TLI.getPointerTy());
1058 // If this is an instruction which fast-isel has deferred, select it now.
1059 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
1060 unsigned InReg = FuncInfo.InitializeRegForValue(Inst);
1061 RegsForValue RFV(*DAG.getContext(), TLI, InReg, Inst->getType());
1062 SDValue Chain = DAG.getEntryNode();
1063 return RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain, NULL);
1066 llvm_unreachable("Can't get register for value!");
1070 void SelectionDAGBuilder::visitRet(const ReturnInst &I) {
1071 SDValue Chain = getControlRoot();
1072 SmallVector<ISD::OutputArg, 8> Outs;
1073 SmallVector<SDValue, 8> OutVals;
1075 if (!FuncInfo.CanLowerReturn) {
1076 unsigned DemoteReg = FuncInfo.DemoteRegister;
1077 const Function *F = I.getParent()->getParent();
1079 // Emit a store of the return value through the virtual register.
1080 // Leave Outs empty so that LowerReturn won't try to load return
1081 // registers the usual way.
1082 SmallVector<EVT, 1> PtrValueVTs;
1083 ComputeValueVTs(TLI, PointerType::getUnqual(F->getReturnType()),
1086 SDValue RetPtr = DAG.getRegister(DemoteReg, PtrValueVTs[0]);
1087 SDValue RetOp = getValue(I.getOperand(0));
1089 SmallVector<EVT, 4> ValueVTs;
1090 SmallVector<uint64_t, 4> Offsets;
1091 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs, &Offsets);
1092 unsigned NumValues = ValueVTs.size();
1094 SmallVector<SDValue, 4> Chains(NumValues);
1095 for (unsigned i = 0; i != NumValues; ++i) {
1096 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(),
1097 RetPtr.getValueType(), RetPtr,
1098 DAG.getIntPtrConstant(Offsets[i]));
1100 DAG.getStore(Chain, getCurDebugLoc(),
1101 SDValue(RetOp.getNode(), RetOp.getResNo() + i),
1102 Add, NULL, Offsets[i], false, false, 0);
1105 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
1106 MVT::Other, &Chains[0], NumValues);
1107 } else if (I.getNumOperands() != 0) {
1108 SmallVector<EVT, 4> ValueVTs;
1109 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs);
1110 unsigned NumValues = ValueVTs.size();
1112 SDValue RetOp = getValue(I.getOperand(0));
1113 for (unsigned j = 0, f = NumValues; j != f; ++j) {
1114 EVT VT = ValueVTs[j];
1116 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1118 const Function *F = I.getParent()->getParent();
1119 if (F->paramHasAttr(0, Attribute::SExt))
1120 ExtendKind = ISD::SIGN_EXTEND;
1121 else if (F->paramHasAttr(0, Attribute::ZExt))
1122 ExtendKind = ISD::ZERO_EXTEND;
1124 // FIXME: C calling convention requires the return type to be promoted
1125 // to at least 32-bit. But this is not necessary for non-C calling
1126 // conventions. The frontend should mark functions whose return values
1127 // require promoting with signext or zeroext attributes.
1128 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
1129 EVT MinVT = TLI.getRegisterType(*DAG.getContext(), MVT::i32);
1130 if (VT.bitsLT(MinVT))
1134 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), VT);
1135 EVT PartVT = TLI.getRegisterType(*DAG.getContext(), VT);
1136 SmallVector<SDValue, 4> Parts(NumParts);
1137 getCopyToParts(DAG, getCurDebugLoc(),
1138 SDValue(RetOp.getNode(), RetOp.getResNo() + j),
1139 &Parts[0], NumParts, PartVT, ExtendKind);
1141 // 'inreg' on function refers to return value
1142 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1143 if (F->paramHasAttr(0, Attribute::InReg))
1146 // Propagate extension type if any
1147 if (F->paramHasAttr(0, Attribute::SExt))
1149 else if (F->paramHasAttr(0, Attribute::ZExt))
1152 for (unsigned i = 0; i < NumParts; ++i) {
1153 Outs.push_back(ISD::OutputArg(Flags, Parts[i].getValueType(),
1155 OutVals.push_back(Parts[i]);
1161 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1162 CallingConv::ID CallConv =
1163 DAG.getMachineFunction().getFunction()->getCallingConv();
1164 Chain = TLI.LowerReturn(Chain, CallConv, isVarArg,
1165 Outs, OutVals, getCurDebugLoc(), DAG);
1167 // Verify that the target's LowerReturn behaved as expected.
1168 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
1169 "LowerReturn didn't return a valid chain!");
1171 // Update the DAG with the new chain value resulting from return lowering.
1175 /// CopyToExportRegsIfNeeded - If the given value has virtual registers
1176 /// created for it, emit nodes to copy the value into the virtual
1178 void SelectionDAGBuilder::CopyToExportRegsIfNeeded(const Value *V) {
1179 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
1180 if (VMI != FuncInfo.ValueMap.end()) {
1181 assert(!V->use_empty() && "Unused value assigned virtual registers!");
1182 CopyValueToVirtualRegister(V, VMI->second);
1186 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
1187 /// the current basic block, add it to ValueMap now so that we'll get a
1189 void SelectionDAGBuilder::ExportFromCurrentBlock(const Value *V) {
1190 // No need to export constants.
1191 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
1193 // Already exported?
1194 if (FuncInfo.isExportedInst(V)) return;
1196 unsigned Reg = FuncInfo.InitializeRegForValue(V);
1197 CopyValueToVirtualRegister(V, Reg);
1200 bool SelectionDAGBuilder::isExportableFromCurrentBlock(const Value *V,
1201 const BasicBlock *FromBB) {
1202 // The operands of the setcc have to be in this block. We don't know
1203 // how to export them from some other block.
1204 if (const Instruction *VI = dyn_cast<Instruction>(V)) {
1205 // Can export from current BB.
1206 if (VI->getParent() == FromBB)
1209 // Is already exported, noop.
1210 return FuncInfo.isExportedInst(V);
1213 // If this is an argument, we can export it if the BB is the entry block or
1214 // if it is already exported.
1215 if (isa<Argument>(V)) {
1216 if (FromBB == &FromBB->getParent()->getEntryBlock())
1219 // Otherwise, can only export this if it is already exported.
1220 return FuncInfo.isExportedInst(V);
1223 // Otherwise, constants can always be exported.
1227 static bool InBlock(const Value *V, const BasicBlock *BB) {
1228 if (const Instruction *I = dyn_cast<Instruction>(V))
1229 return I->getParent() == BB;
1233 /// EmitBranchForMergedCondition - Helper method for FindMergedConditions.
1234 /// This function emits a branch and is used at the leaves of an OR or an
1235 /// AND operator tree.
1238 SelectionDAGBuilder::EmitBranchForMergedCondition(const Value *Cond,
1239 MachineBasicBlock *TBB,
1240 MachineBasicBlock *FBB,
1241 MachineBasicBlock *CurBB,
1242 MachineBasicBlock *SwitchBB) {
1243 const BasicBlock *BB = CurBB->getBasicBlock();
1245 // If the leaf of the tree is a comparison, merge the condition into
1247 if (const CmpInst *BOp = dyn_cast<CmpInst>(Cond)) {
1248 // The operands of the cmp have to be in this block. We don't know
1249 // how to export them from some other block. If this is the first block
1250 // of the sequence, no exporting is needed.
1251 if (CurBB == SwitchBB ||
1252 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
1253 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) {
1254 ISD::CondCode Condition;
1255 if (const ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
1256 Condition = getICmpCondCode(IC->getPredicate());
1257 } else if (const FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) {
1258 Condition = getFCmpCondCode(FC->getPredicate());
1260 Condition = ISD::SETEQ; // silence warning.
1261 llvm_unreachable("Unknown compare instruction");
1264 CaseBlock CB(Condition, BOp->getOperand(0),
1265 BOp->getOperand(1), NULL, TBB, FBB, CurBB);
1266 SwitchCases.push_back(CB);
1271 // Create a CaseBlock record representing this branch.
1272 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()),
1273 NULL, TBB, FBB, CurBB);
1274 SwitchCases.push_back(CB);
1277 /// FindMergedConditions - If Cond is an expression like
1278 void SelectionDAGBuilder::FindMergedConditions(const Value *Cond,
1279 MachineBasicBlock *TBB,
1280 MachineBasicBlock *FBB,
1281 MachineBasicBlock *CurBB,
1282 MachineBasicBlock *SwitchBB,
1284 // If this node is not part of the or/and tree, emit it as a branch.
1285 const Instruction *BOp = dyn_cast<Instruction>(Cond);
1286 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
1287 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
1288 BOp->getParent() != CurBB->getBasicBlock() ||
1289 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
1290 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
1291 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB);
1295 // Create TmpBB after CurBB.
1296 MachineFunction::iterator BBI = CurBB;
1297 MachineFunction &MF = DAG.getMachineFunction();
1298 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock());
1299 CurBB->getParent()->insert(++BBI, TmpBB);
1301 if (Opc == Instruction::Or) {
1302 // Codegen X | Y as:
1310 // Emit the LHS condition.
1311 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, SwitchBB, Opc);
1313 // Emit the RHS condition into TmpBB.
1314 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
1316 assert(Opc == Instruction::And && "Unknown merge op!");
1317 // Codegen X & Y as:
1324 // This requires creation of TmpBB after CurBB.
1326 // Emit the LHS condition.
1327 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, SwitchBB, Opc);
1329 // Emit the RHS condition into TmpBB.
1330 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
1334 /// If the set of cases should be emitted as a series of branches, return true.
1335 /// If we should emit this as a bunch of and/or'd together conditions, return
1338 SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases){
1339 if (Cases.size() != 2) return true;
1341 // If this is two comparisons of the same values or'd or and'd together, they
1342 // will get folded into a single comparison, so don't emit two blocks.
1343 if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
1344 Cases[0].CmpRHS == Cases[1].CmpRHS) ||
1345 (Cases[0].CmpRHS == Cases[1].CmpLHS &&
1346 Cases[0].CmpLHS == Cases[1].CmpRHS)) {
1350 // Handle: (X != null) | (Y != null) --> (X|Y) != 0
1351 // Handle: (X == null) & (Y == null) --> (X|Y) == 0
1352 if (Cases[0].CmpRHS == Cases[1].CmpRHS &&
1353 Cases[0].CC == Cases[1].CC &&
1354 isa<Constant>(Cases[0].CmpRHS) &&
1355 cast<Constant>(Cases[0].CmpRHS)->isNullValue()) {
1356 if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB)
1358 if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB)
1365 void SelectionDAGBuilder::visitBr(const BranchInst &I) {
1366 MachineBasicBlock *BrMBB = FuncInfo.MBB;
1368 // Update machine-CFG edges.
1369 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
1371 // Figure out which block is immediately after the current one.
1372 MachineBasicBlock *NextBlock = 0;
1373 MachineFunction::iterator BBI = BrMBB;
1374 if (++BBI != FuncInfo.MF->end())
1377 if (I.isUnconditional()) {
1378 // Update machine-CFG edges.
1379 BrMBB->addSuccessor(Succ0MBB);
1381 // If this is not a fall-through branch, emit the branch.
1382 if (Succ0MBB != NextBlock)
1383 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
1384 MVT::Other, getControlRoot(),
1385 DAG.getBasicBlock(Succ0MBB)));
1390 // If this condition is one of the special cases we handle, do special stuff
1392 const Value *CondVal = I.getCondition();
1393 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
1395 // If this is a series of conditions that are or'd or and'd together, emit
1396 // this as a sequence of branches instead of setcc's with and/or operations.
1397 // For example, instead of something like:
1410 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
1411 if (BOp->hasOneUse() &&
1412 (BOp->getOpcode() == Instruction::And ||
1413 BOp->getOpcode() == Instruction::Or)) {
1414 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB,
1416 // If the compares in later blocks need to use values not currently
1417 // exported from this block, export them now. This block should always
1418 // be the first entry.
1419 assert(SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!");
1421 // Allow some cases to be rejected.
1422 if (ShouldEmitAsBranches(SwitchCases)) {
1423 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1424 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1425 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1428 // Emit the branch for this block.
1429 visitSwitchCase(SwitchCases[0], BrMBB);
1430 SwitchCases.erase(SwitchCases.begin());
1434 // Okay, we decided not to do this, remove any inserted MBB's and clear
1436 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
1437 FuncInfo.MF->erase(SwitchCases[i].ThisBB);
1439 SwitchCases.clear();
1443 // Create a CaseBlock record representing this branch.
1444 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()),
1445 NULL, Succ0MBB, Succ1MBB, BrMBB);
1447 // Use visitSwitchCase to actually insert the fast branch sequence for this
1449 visitSwitchCase(CB, BrMBB);
1452 /// visitSwitchCase - Emits the necessary code to represent a single node in
1453 /// the binary search tree resulting from lowering a switch instruction.
1454 void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB,
1455 MachineBasicBlock *SwitchBB) {
1457 SDValue CondLHS = getValue(CB.CmpLHS);
1458 DebugLoc dl = getCurDebugLoc();
1460 // Build the setcc now.
1461 if (CB.CmpMHS == NULL) {
1462 // Fold "(X == true)" to X and "(X == false)" to !X to
1463 // handle common cases produced by branch lowering.
1464 if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) &&
1465 CB.CC == ISD::SETEQ)
1467 else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) &&
1468 CB.CC == ISD::SETEQ) {
1469 SDValue True = DAG.getConstant(1, CondLHS.getValueType());
1470 Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True);
1472 Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1474 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now");
1476 const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue();
1477 const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue();
1479 SDValue CmpOp = getValue(CB.CmpMHS);
1480 EVT VT = CmpOp.getValueType();
1482 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) {
1483 Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, VT),
1486 SDValue SUB = DAG.getNode(ISD::SUB, dl,
1487 VT, CmpOp, DAG.getConstant(Low, VT));
1488 Cond = DAG.getSetCC(dl, MVT::i1, SUB,
1489 DAG.getConstant(High-Low, VT), ISD::SETULE);
1493 // Update successor info
1494 SwitchBB->addSuccessor(CB.TrueBB);
1495 SwitchBB->addSuccessor(CB.FalseBB);
1497 // Set NextBlock to be the MBB immediately after the current one, if any.
1498 // This is used to avoid emitting unnecessary branches to the next block.
1499 MachineBasicBlock *NextBlock = 0;
1500 MachineFunction::iterator BBI = SwitchBB;
1501 if (++BBI != FuncInfo.MF->end())
1504 // If the lhs block is the next block, invert the condition so that we can
1505 // fall through to the lhs instead of the rhs block.
1506 if (CB.TrueBB == NextBlock) {
1507 std::swap(CB.TrueBB, CB.FalseBB);
1508 SDValue True = DAG.getConstant(1, Cond.getValueType());
1509 Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True);
1512 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1513 MVT::Other, getControlRoot(), Cond,
1514 DAG.getBasicBlock(CB.TrueBB));
1516 // Insert the false branch.
1517 if (CB.FalseBB != NextBlock)
1518 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1519 DAG.getBasicBlock(CB.FalseBB));
1521 DAG.setRoot(BrCond);
1524 /// visitJumpTable - Emit JumpTable node in the current MBB
1525 void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) {
1526 // Emit the code for the jump table
1527 assert(JT.Reg != -1U && "Should lower JT Header first!");
1528 EVT PTy = TLI.getPointerTy();
1529 SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(),
1531 SDValue Table = DAG.getJumpTable(JT.JTI, PTy);
1532 SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurDebugLoc(),
1533 MVT::Other, Index.getValue(1),
1535 DAG.setRoot(BrJumpTable);
1538 /// visitJumpTableHeader - This function emits necessary code to produce index
1539 /// in the JumpTable from switch case.
1540 void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT,
1541 JumpTableHeader &JTH,
1542 MachineBasicBlock *SwitchBB) {
1543 // Subtract the lowest switch case value from the value being switched on and
1544 // conditional branch to default mbb if the result is greater than the
1545 // difference between smallest and largest cases.
1546 SDValue SwitchOp = getValue(JTH.SValue);
1547 EVT VT = SwitchOp.getValueType();
1548 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1549 DAG.getConstant(JTH.First, VT));
1551 // The SDNode we just created, which holds the value being switched on minus
1552 // the smallest case value, needs to be copied to a virtual register so it
1553 // can be used as an index into the jump table in a subsequent basic block.
1554 // This value may be smaller or larger than the target's pointer type, and
1555 // therefore require extension or truncating.
1556 SwitchOp = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(), TLI.getPointerTy());
1558 unsigned JumpTableReg = FuncInfo.CreateReg(TLI.getPointerTy());
1559 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1560 JumpTableReg, SwitchOp);
1561 JT.Reg = JumpTableReg;
1563 // Emit the range check for the jump table, and branch to the default block
1564 // for the switch statement if the value being switched on exceeds the largest
1565 // case in the switch.
1566 SDValue CMP = DAG.getSetCC(getCurDebugLoc(),
1567 TLI.getSetCCResultType(Sub.getValueType()), Sub,
1568 DAG.getConstant(JTH.Last-JTH.First,VT),
1571 // Set NextBlock to be the MBB immediately after the current one, if any.
1572 // This is used to avoid emitting unnecessary branches to the next block.
1573 MachineBasicBlock *NextBlock = 0;
1574 MachineFunction::iterator BBI = SwitchBB;
1576 if (++BBI != FuncInfo.MF->end())
1579 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1580 MVT::Other, CopyTo, CMP,
1581 DAG.getBasicBlock(JT.Default));
1583 if (JT.MBB != NextBlock)
1584 BrCond = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrCond,
1585 DAG.getBasicBlock(JT.MBB));
1587 DAG.setRoot(BrCond);
1590 /// visitBitTestHeader - This function emits necessary code to produce value
1591 /// suitable for "bit tests"
1592 void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B,
1593 MachineBasicBlock *SwitchBB) {
1594 // Subtract the minimum value
1595 SDValue SwitchOp = getValue(B.SValue);
1596 EVT VT = SwitchOp.getValueType();
1597 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1598 DAG.getConstant(B.First, VT));
1601 SDValue RangeCmp = DAG.getSetCC(getCurDebugLoc(),
1602 TLI.getSetCCResultType(Sub.getValueType()),
1603 Sub, DAG.getConstant(B.Range, VT),
1606 SDValue ShiftOp = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(),
1607 TLI.getPointerTy());
1609 B.Reg = FuncInfo.CreateReg(TLI.getPointerTy());
1610 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1613 // Set NextBlock to be the MBB immediately after the current one, if any.
1614 // This is used to avoid emitting unnecessary branches to the next block.
1615 MachineBasicBlock *NextBlock = 0;
1616 MachineFunction::iterator BBI = SwitchBB;
1617 if (++BBI != FuncInfo.MF->end())
1620 MachineBasicBlock* MBB = B.Cases[0].ThisBB;
1622 SwitchBB->addSuccessor(B.Default);
1623 SwitchBB->addSuccessor(MBB);
1625 SDValue BrRange = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1626 MVT::Other, CopyTo, RangeCmp,
1627 DAG.getBasicBlock(B.Default));
1629 if (MBB != NextBlock)
1630 BrRange = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, CopyTo,
1631 DAG.getBasicBlock(MBB));
1633 DAG.setRoot(BrRange);
1636 /// visitBitTestCase - this function produces one "bit test"
1637 void SelectionDAGBuilder::visitBitTestCase(MachineBasicBlock* NextMBB,
1640 MachineBasicBlock *SwitchBB) {
1641 SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(), Reg,
1642 TLI.getPointerTy());
1644 if (CountPopulation_64(B.Mask) == 1) {
1645 // Testing for a single bit; just compare the shift count with what it
1646 // would need to be to shift a 1 bit in that position.
1647 Cmp = DAG.getSetCC(getCurDebugLoc(),
1648 TLI.getSetCCResultType(ShiftOp.getValueType()),
1650 DAG.getConstant(CountTrailingZeros_64(B.Mask),
1651 TLI.getPointerTy()),
1654 // Make desired shift
1655 SDValue SwitchVal = DAG.getNode(ISD::SHL, getCurDebugLoc(),
1657 DAG.getConstant(1, TLI.getPointerTy()),
1660 // Emit bit tests and jumps
1661 SDValue AndOp = DAG.getNode(ISD::AND, getCurDebugLoc(),
1662 TLI.getPointerTy(), SwitchVal,
1663 DAG.getConstant(B.Mask, TLI.getPointerTy()));
1664 Cmp = DAG.getSetCC(getCurDebugLoc(),
1665 TLI.getSetCCResultType(AndOp.getValueType()),
1666 AndOp, DAG.getConstant(0, TLI.getPointerTy()),
1670 SwitchBB->addSuccessor(B.TargetBB);
1671 SwitchBB->addSuccessor(NextMBB);
1673 SDValue BrAnd = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1674 MVT::Other, getControlRoot(),
1675 Cmp, DAG.getBasicBlock(B.TargetBB));
1677 // Set NextBlock to be the MBB immediately after the current one, if any.
1678 // This is used to avoid emitting unnecessary branches to the next block.
1679 MachineBasicBlock *NextBlock = 0;
1680 MachineFunction::iterator BBI = SwitchBB;
1681 if (++BBI != FuncInfo.MF->end())
1684 if (NextMBB != NextBlock)
1685 BrAnd = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrAnd,
1686 DAG.getBasicBlock(NextMBB));
1691 void SelectionDAGBuilder::visitInvoke(const InvokeInst &I) {
1692 MachineBasicBlock *InvokeMBB = FuncInfo.MBB;
1694 // Retrieve successors.
1695 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
1696 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)];
1698 const Value *Callee(I.getCalledValue());
1699 if (isa<InlineAsm>(Callee))
1702 LowerCallTo(&I, getValue(Callee), false, LandingPad);
1704 // If the value of the invoke is used outside of its defining block, make it
1705 // available as a virtual register.
1706 CopyToExportRegsIfNeeded(&I);
1708 // Update successor info
1709 InvokeMBB->addSuccessor(Return);
1710 InvokeMBB->addSuccessor(LandingPad);
1712 // Drop into normal successor.
1713 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
1714 MVT::Other, getControlRoot(),
1715 DAG.getBasicBlock(Return)));
1718 void SelectionDAGBuilder::visitUnwind(const UnwindInst &I) {
1721 /// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for
1722 /// small case ranges).
1723 bool SelectionDAGBuilder::handleSmallSwitchRange(CaseRec& CR,
1724 CaseRecVector& WorkList,
1726 MachineBasicBlock *Default,
1727 MachineBasicBlock *SwitchBB) {
1728 Case& BackCase = *(CR.Range.second-1);
1730 // Size is the number of Cases represented by this range.
1731 size_t Size = CR.Range.second - CR.Range.first;
1735 // Get the MachineFunction which holds the current MBB. This is used when
1736 // inserting any additional MBBs necessary to represent the switch.
1737 MachineFunction *CurMF = FuncInfo.MF;
1739 // Figure out which block is immediately after the current one.
1740 MachineBasicBlock *NextBlock = 0;
1741 MachineFunction::iterator BBI = CR.CaseBB;
1743 if (++BBI != FuncInfo.MF->end())
1746 // TODO: If any two of the cases has the same destination, and if one value
1747 // is the same as the other, but has one bit unset that the other has set,
1748 // use bit manipulation to do two compares at once. For example:
1749 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
1751 // Rearrange the case blocks so that the last one falls through if possible.
1752 if (NextBlock && Default != NextBlock && BackCase.BB != NextBlock) {
1753 // The last case block won't fall through into 'NextBlock' if we emit the
1754 // branches in this order. See if rearranging a case value would help.
1755 for (CaseItr I = CR.Range.first, E = CR.Range.second-1; I != E; ++I) {
1756 if (I->BB == NextBlock) {
1757 std::swap(*I, BackCase);
1763 // Create a CaseBlock record representing a conditional branch to
1764 // the Case's target mbb if the value being switched on SV is equal
1766 MachineBasicBlock *CurBlock = CR.CaseBB;
1767 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
1768 MachineBasicBlock *FallThrough;
1770 FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock());
1771 CurMF->insert(BBI, FallThrough);
1773 // Put SV in a virtual register to make it available from the new blocks.
1774 ExportFromCurrentBlock(SV);
1776 // If the last case doesn't match, go to the default block.
1777 FallThrough = Default;
1780 const Value *RHS, *LHS, *MHS;
1782 if (I->High == I->Low) {
1783 // This is just small small case range :) containing exactly 1 case
1785 LHS = SV; RHS = I->High; MHS = NULL;
1788 LHS = I->Low; MHS = SV; RHS = I->High;
1790 CaseBlock CB(CC, LHS, RHS, MHS, I->BB, FallThrough, CurBlock);
1792 // If emitting the first comparison, just call visitSwitchCase to emit the
1793 // code into the current block. Otherwise, push the CaseBlock onto the
1794 // vector to be later processed by SDISel, and insert the node's MBB
1795 // before the next MBB.
1796 if (CurBlock == SwitchBB)
1797 visitSwitchCase(CB, SwitchBB);
1799 SwitchCases.push_back(CB);
1801 CurBlock = FallThrough;
1807 static inline bool areJTsAllowed(const TargetLowering &TLI) {
1808 return !DisableJumpTables &&
1809 (TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
1810 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other));
1813 static APInt ComputeRange(const APInt &First, const APInt &Last) {
1814 APInt LastExt(Last), FirstExt(First);
1815 uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1;
1816 LastExt.sext(BitWidth); FirstExt.sext(BitWidth);
1817 return (LastExt - FirstExt + 1ULL);
1820 /// handleJTSwitchCase - Emit jumptable for current switch case range
1821 bool SelectionDAGBuilder::handleJTSwitchCase(CaseRec& CR,
1822 CaseRecVector& WorkList,
1824 MachineBasicBlock* Default,
1825 MachineBasicBlock *SwitchBB) {
1826 Case& FrontCase = *CR.Range.first;
1827 Case& BackCase = *(CR.Range.second-1);
1829 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
1830 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
1832 APInt TSize(First.getBitWidth(), 0);
1833 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1837 if (!areJTsAllowed(TLI) || TSize.ult(4))
1840 APInt Range = ComputeRange(First, Last);
1841 double Density = TSize.roundToDouble() / Range.roundToDouble();
1845 DEBUG(dbgs() << "Lowering jump table\n"
1846 << "First entry: " << First << ". Last entry: " << Last << '\n'
1847 << "Range: " << Range
1848 << "Size: " << TSize << ". Density: " << Density << "\n\n");
1850 // Get the MachineFunction which holds the current MBB. This is used when
1851 // inserting any additional MBBs necessary to represent the switch.
1852 MachineFunction *CurMF = FuncInfo.MF;
1854 // Figure out which block is immediately after the current one.
1855 MachineFunction::iterator BBI = CR.CaseBB;
1858 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1860 // Create a new basic block to hold the code for loading the address
1861 // of the jump table, and jumping to it. Update successor information;
1862 // we will either branch to the default case for the switch, or the jump
1864 MachineBasicBlock *JumpTableBB = CurMF->CreateMachineBasicBlock(LLVMBB);
1865 CurMF->insert(BBI, JumpTableBB);
1866 CR.CaseBB->addSuccessor(Default);
1867 CR.CaseBB->addSuccessor(JumpTableBB);
1869 // Build a vector of destination BBs, corresponding to each target
1870 // of the jump table. If the value of the jump table slot corresponds to
1871 // a case statement, push the case's BB onto the vector, otherwise, push
1873 std::vector<MachineBasicBlock*> DestBBs;
1875 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) {
1876 const APInt &Low = cast<ConstantInt>(I->Low)->getValue();
1877 const APInt &High = cast<ConstantInt>(I->High)->getValue();
1879 if (Low.sle(TEI) && TEI.sle(High)) {
1880 DestBBs.push_back(I->BB);
1884 DestBBs.push_back(Default);
1888 // Update successor info. Add one edge to each unique successor.
1889 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs());
1890 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(),
1891 E = DestBBs.end(); I != E; ++I) {
1892 if (!SuccsHandled[(*I)->getNumber()]) {
1893 SuccsHandled[(*I)->getNumber()] = true;
1894 JumpTableBB->addSuccessor(*I);
1898 // Create a jump table index for this jump table.
1899 unsigned JTEncoding = TLI.getJumpTableEncoding();
1900 unsigned JTI = CurMF->getOrCreateJumpTableInfo(JTEncoding)
1901 ->createJumpTableIndex(DestBBs);
1903 // Set the jump table information so that we can codegen it as a second
1904 // MachineBasicBlock
1905 JumpTable JT(-1U, JTI, JumpTableBB, Default);
1906 JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == SwitchBB));
1907 if (CR.CaseBB == SwitchBB)
1908 visitJumpTableHeader(JT, JTH, SwitchBB);
1910 JTCases.push_back(JumpTableBlock(JTH, JT));
1915 /// handleBTSplitSwitchCase - emit comparison and split binary search tree into
1917 bool SelectionDAGBuilder::handleBTSplitSwitchCase(CaseRec& CR,
1918 CaseRecVector& WorkList,
1920 MachineBasicBlock *Default,
1921 MachineBasicBlock *SwitchBB) {
1922 // Get the MachineFunction which holds the current MBB. This is used when
1923 // inserting any additional MBBs necessary to represent the switch.
1924 MachineFunction *CurMF = FuncInfo.MF;
1926 // Figure out which block is immediately after the current one.
1927 MachineFunction::iterator BBI = CR.CaseBB;
1930 Case& FrontCase = *CR.Range.first;
1931 Case& BackCase = *(CR.Range.second-1);
1932 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1934 // Size is the number of Cases represented by this range.
1935 unsigned Size = CR.Range.second - CR.Range.first;
1937 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
1938 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
1940 CaseItr Pivot = CR.Range.first + Size/2;
1942 // Select optimal pivot, maximizing sum density of LHS and RHS. This will
1943 // (heuristically) allow us to emit JumpTable's later.
1944 APInt TSize(First.getBitWidth(), 0);
1945 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1949 APInt LSize = FrontCase.size();
1950 APInt RSize = TSize-LSize;
1951 DEBUG(dbgs() << "Selecting best pivot: \n"
1952 << "First: " << First << ", Last: " << Last <<'\n'
1953 << "LSize: " << LSize << ", RSize: " << RSize << '\n');
1954 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second;
1956 const APInt &LEnd = cast<ConstantInt>(I->High)->getValue();
1957 const APInt &RBegin = cast<ConstantInt>(J->Low)->getValue();
1958 APInt Range = ComputeRange(LEnd, RBegin);
1959 assert((Range - 2ULL).isNonNegative() &&
1960 "Invalid case distance");
1961 double LDensity = (double)LSize.roundToDouble() /
1962 (LEnd - First + 1ULL).roundToDouble();
1963 double RDensity = (double)RSize.roundToDouble() /
1964 (Last - RBegin + 1ULL).roundToDouble();
1965 double Metric = Range.logBase2()*(LDensity+RDensity);
1966 // Should always split in some non-trivial place
1967 DEBUG(dbgs() <<"=>Step\n"
1968 << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n'
1969 << "LDensity: " << LDensity
1970 << ", RDensity: " << RDensity << '\n'
1971 << "Metric: " << Metric << '\n');
1972 if (FMetric < Metric) {
1975 DEBUG(dbgs() << "Current metric set to: " << FMetric << '\n');
1981 if (areJTsAllowed(TLI)) {
1982 // If our case is dense we *really* should handle it earlier!
1983 assert((FMetric > 0) && "Should handle dense range earlier!");
1985 Pivot = CR.Range.first + Size/2;
1988 CaseRange LHSR(CR.Range.first, Pivot);
1989 CaseRange RHSR(Pivot, CR.Range.second);
1990 Constant *C = Pivot->Low;
1991 MachineBasicBlock *FalseBB = 0, *TrueBB = 0;
1993 // We know that we branch to the LHS if the Value being switched on is
1994 // less than the Pivot value, C. We use this to optimize our binary
1995 // tree a bit, by recognizing that if SV is greater than or equal to the
1996 // LHS's Case Value, and that Case Value is exactly one less than the
1997 // Pivot's Value, then we can branch directly to the LHS's Target,
1998 // rather than creating a leaf node for it.
1999 if ((LHSR.second - LHSR.first) == 1 &&
2000 LHSR.first->High == CR.GE &&
2001 cast<ConstantInt>(C)->getValue() ==
2002 (cast<ConstantInt>(CR.GE)->getValue() + 1LL)) {
2003 TrueBB = LHSR.first->BB;
2005 TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2006 CurMF->insert(BBI, TrueBB);
2007 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR));
2009 // Put SV in a virtual register to make it available from the new blocks.
2010 ExportFromCurrentBlock(SV);
2013 // Similar to the optimization above, if the Value being switched on is
2014 // known to be less than the Constant CR.LT, and the current Case Value
2015 // is CR.LT - 1, then we can branch directly to the target block for
2016 // the current Case Value, rather than emitting a RHS leaf node for it.
2017 if ((RHSR.second - RHSR.first) == 1 && CR.LT &&
2018 cast<ConstantInt>(RHSR.first->Low)->getValue() ==
2019 (cast<ConstantInt>(CR.LT)->getValue() - 1LL)) {
2020 FalseBB = RHSR.first->BB;
2022 FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2023 CurMF->insert(BBI, FalseBB);
2024 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR));
2026 // Put SV in a virtual register to make it available from the new blocks.
2027 ExportFromCurrentBlock(SV);
2030 // Create a CaseBlock record representing a conditional branch to
2031 // the LHS node if the value being switched on SV is less than C.
2032 // Otherwise, branch to LHS.
2033 CaseBlock CB(ISD::SETLT, SV, C, NULL, TrueBB, FalseBB, CR.CaseBB);
2035 if (CR.CaseBB == SwitchBB)
2036 visitSwitchCase(CB, SwitchBB);
2038 SwitchCases.push_back(CB);
2043 /// handleBitTestsSwitchCase - if current case range has few destination and
2044 /// range span less, than machine word bitwidth, encode case range into series
2045 /// of masks and emit bit tests with these masks.
2046 bool SelectionDAGBuilder::handleBitTestsSwitchCase(CaseRec& CR,
2047 CaseRecVector& WorkList,
2049 MachineBasicBlock* Default,
2050 MachineBasicBlock *SwitchBB){
2051 EVT PTy = TLI.getPointerTy();
2052 unsigned IntPtrBits = PTy.getSizeInBits();
2054 Case& FrontCase = *CR.Range.first;
2055 Case& BackCase = *(CR.Range.second-1);
2057 // Get the MachineFunction which holds the current MBB. This is used when
2058 // inserting any additional MBBs necessary to represent the switch.
2059 MachineFunction *CurMF = FuncInfo.MF;
2061 // If target does not have legal shift left, do not emit bit tests at all.
2062 if (!TLI.isOperationLegal(ISD::SHL, TLI.getPointerTy()))
2066 for (CaseItr I = CR.Range.first, E = CR.Range.second;
2068 // Single case counts one, case range - two.
2069 numCmps += (I->Low == I->High ? 1 : 2);
2072 // Count unique destinations
2073 SmallSet<MachineBasicBlock*, 4> Dests;
2074 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2075 Dests.insert(I->BB);
2076 if (Dests.size() > 3)
2077 // Don't bother the code below, if there are too much unique destinations
2080 DEBUG(dbgs() << "Total number of unique destinations: "
2081 << Dests.size() << '\n'
2082 << "Total number of comparisons: " << numCmps << '\n');
2084 // Compute span of values.
2085 const APInt& minValue = cast<ConstantInt>(FrontCase.Low)->getValue();
2086 const APInt& maxValue = cast<ConstantInt>(BackCase.High)->getValue();
2087 APInt cmpRange = maxValue - minValue;
2089 DEBUG(dbgs() << "Compare range: " << cmpRange << '\n'
2090 << "Low bound: " << minValue << '\n'
2091 << "High bound: " << maxValue << '\n');
2093 if (cmpRange.uge(IntPtrBits) ||
2094 (!(Dests.size() == 1 && numCmps >= 3) &&
2095 !(Dests.size() == 2 && numCmps >= 5) &&
2096 !(Dests.size() >= 3 && numCmps >= 6)))
2099 DEBUG(dbgs() << "Emitting bit tests\n");
2100 APInt lowBound = APInt::getNullValue(cmpRange.getBitWidth());
2102 // Optimize the case where all the case values fit in a
2103 // word without having to subtract minValue. In this case,
2104 // we can optimize away the subtraction.
2105 if (minValue.isNonNegative() && maxValue.slt(IntPtrBits)) {
2106 cmpRange = maxValue;
2108 lowBound = minValue;
2111 CaseBitsVector CasesBits;
2112 unsigned i, count = 0;
2114 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2115 MachineBasicBlock* Dest = I->BB;
2116 for (i = 0; i < count; ++i)
2117 if (Dest == CasesBits[i].BB)
2121 assert((count < 3) && "Too much destinations to test!");
2122 CasesBits.push_back(CaseBits(0, Dest, 0));
2126 const APInt& lowValue = cast<ConstantInt>(I->Low)->getValue();
2127 const APInt& highValue = cast<ConstantInt>(I->High)->getValue();
2129 uint64_t lo = (lowValue - lowBound).getZExtValue();
2130 uint64_t hi = (highValue - lowBound).getZExtValue();
2132 for (uint64_t j = lo; j <= hi; j++) {
2133 CasesBits[i].Mask |= 1ULL << j;
2134 CasesBits[i].Bits++;
2138 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp());
2142 // Figure out which block is immediately after the current one.
2143 MachineFunction::iterator BBI = CR.CaseBB;
2146 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2148 DEBUG(dbgs() << "Cases:\n");
2149 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) {
2150 DEBUG(dbgs() << "Mask: " << CasesBits[i].Mask
2151 << ", Bits: " << CasesBits[i].Bits
2152 << ", BB: " << CasesBits[i].BB << '\n');
2154 MachineBasicBlock *CaseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2155 CurMF->insert(BBI, CaseBB);
2156 BTC.push_back(BitTestCase(CasesBits[i].Mask,
2160 // Put SV in a virtual register to make it available from the new blocks.
2161 ExportFromCurrentBlock(SV);
2164 BitTestBlock BTB(lowBound, cmpRange, SV,
2165 -1U, (CR.CaseBB == SwitchBB),
2166 CR.CaseBB, Default, BTC);
2168 if (CR.CaseBB == SwitchBB)
2169 visitBitTestHeader(BTB, SwitchBB);
2171 BitTestCases.push_back(BTB);
2176 /// Clusterify - Transform simple list of Cases into list of CaseRange's
2177 size_t SelectionDAGBuilder::Clusterify(CaseVector& Cases,
2178 const SwitchInst& SI) {
2181 // Start with "simple" cases
2182 for (size_t i = 1; i < SI.getNumSuccessors(); ++i) {
2183 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SI.getSuccessor(i)];
2184 Cases.push_back(Case(SI.getSuccessorValue(i),
2185 SI.getSuccessorValue(i),
2188 std::sort(Cases.begin(), Cases.end(), CaseCmp());
2190 // Merge case into clusters
2191 if (Cases.size() >= 2)
2192 // Must recompute end() each iteration because it may be
2193 // invalidated by erase if we hold on to it
2194 for (CaseItr I = Cases.begin(), J = ++(Cases.begin()); J != Cases.end(); ) {
2195 const APInt& nextValue = cast<ConstantInt>(J->Low)->getValue();
2196 const APInt& currentValue = cast<ConstantInt>(I->High)->getValue();
2197 MachineBasicBlock* nextBB = J->BB;
2198 MachineBasicBlock* currentBB = I->BB;
2200 // If the two neighboring cases go to the same destination, merge them
2201 // into a single case.
2202 if ((nextValue - currentValue == 1) && (currentBB == nextBB)) {
2210 for (CaseItr I=Cases.begin(), E=Cases.end(); I!=E; ++I, ++numCmps) {
2211 if (I->Low != I->High)
2212 // A range counts double, since it requires two compares.
2219 void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) {
2220 MachineBasicBlock *SwitchMBB = FuncInfo.MBB;
2222 // Figure out which block is immediately after the current one.
2223 MachineBasicBlock *NextBlock = 0;
2224 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()];
2226 // If there is only the default destination, branch to it if it is not the
2227 // next basic block. Otherwise, just fall through.
2228 if (SI.getNumOperands() == 2) {
2229 // Update machine-CFG edges.
2231 // If this is not a fall-through branch, emit the branch.
2232 SwitchMBB->addSuccessor(Default);
2233 if (Default != NextBlock)
2234 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
2235 MVT::Other, getControlRoot(),
2236 DAG.getBasicBlock(Default)));
2241 // If there are any non-default case statements, create a vector of Cases
2242 // representing each one, and sort the vector so that we can efficiently
2243 // create a binary search tree from them.
2245 size_t numCmps = Clusterify(Cases, SI);
2246 DEBUG(dbgs() << "Clusterify finished. Total clusters: " << Cases.size()
2247 << ". Total compares: " << numCmps << '\n');
2250 // Get the Value to be switched on and default basic blocks, which will be
2251 // inserted into CaseBlock records, representing basic blocks in the binary
2253 const Value *SV = SI.getOperand(0);
2255 // Push the initial CaseRec onto the worklist
2256 CaseRecVector WorkList;
2257 WorkList.push_back(CaseRec(SwitchMBB,0,0,
2258 CaseRange(Cases.begin(),Cases.end())));
2260 while (!WorkList.empty()) {
2261 // Grab a record representing a case range to process off the worklist
2262 CaseRec CR = WorkList.back();
2263 WorkList.pop_back();
2265 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2268 // If the range has few cases (two or less) emit a series of specific
2270 if (handleSmallSwitchRange(CR, WorkList, SV, Default, SwitchMBB))
2273 // If the switch has more than 5 blocks, and at least 40% dense, and the
2274 // target supports indirect branches, then emit a jump table rather than
2275 // lowering the switch to a binary tree of conditional branches.
2276 if (handleJTSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2279 // Emit binary tree. We need to pick a pivot, and push left and right ranges
2280 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call.
2281 handleBTSplitSwitchCase(CR, WorkList, SV, Default, SwitchMBB);
2285 void SelectionDAGBuilder::visitIndirectBr(const IndirectBrInst &I) {
2286 MachineBasicBlock *IndirectBrMBB = FuncInfo.MBB;
2288 // Update machine-CFG edges with unique successors.
2289 SmallVector<BasicBlock*, 32> succs;
2290 succs.reserve(I.getNumSuccessors());
2291 for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i)
2292 succs.push_back(I.getSuccessor(i));
2293 array_pod_sort(succs.begin(), succs.end());
2294 succs.erase(std::unique(succs.begin(), succs.end()), succs.end());
2295 for (unsigned i = 0, e = succs.size(); i != e; ++i)
2296 IndirectBrMBB->addSuccessor(FuncInfo.MBBMap[succs[i]]);
2298 DAG.setRoot(DAG.getNode(ISD::BRIND, getCurDebugLoc(),
2299 MVT::Other, getControlRoot(),
2300 getValue(I.getAddress())));
2303 void SelectionDAGBuilder::visitFSub(const User &I) {
2304 // -0.0 - X --> fneg
2305 const Type *Ty = I.getType();
2306 if (Ty->isVectorTy()) {
2307 if (ConstantVector *CV = dyn_cast<ConstantVector>(I.getOperand(0))) {
2308 const VectorType *DestTy = cast<VectorType>(I.getType());
2309 const Type *ElTy = DestTy->getElementType();
2310 unsigned VL = DestTy->getNumElements();
2311 std::vector<Constant*> NZ(VL, ConstantFP::getNegativeZero(ElTy));
2312 Constant *CNZ = ConstantVector::get(&NZ[0], NZ.size());
2314 SDValue Op2 = getValue(I.getOperand(1));
2315 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(),
2316 Op2.getValueType(), Op2));
2322 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I.getOperand(0)))
2323 if (CFP->isExactlyValue(ConstantFP::getNegativeZero(Ty)->getValueAPF())) {
2324 SDValue Op2 = getValue(I.getOperand(1));
2325 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(),
2326 Op2.getValueType(), Op2));
2330 visitBinary(I, ISD::FSUB);
2333 void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) {
2334 SDValue Op1 = getValue(I.getOperand(0));
2335 SDValue Op2 = getValue(I.getOperand(1));
2336 setValue(&I, DAG.getNode(OpCode, getCurDebugLoc(),
2337 Op1.getValueType(), Op1, Op2));
2340 void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) {
2341 SDValue Op1 = getValue(I.getOperand(0));
2342 SDValue Op2 = getValue(I.getOperand(1));
2343 if (!I.getType()->isVectorTy() &&
2344 Op2.getValueType() != TLI.getShiftAmountTy()) {
2345 // If the operand is smaller than the shift count type, promote it.
2346 EVT PTy = TLI.getPointerTy();
2347 EVT STy = TLI.getShiftAmountTy();
2348 if (STy.bitsGT(Op2.getValueType()))
2349 Op2 = DAG.getNode(ISD::ANY_EXTEND, getCurDebugLoc(),
2350 TLI.getShiftAmountTy(), Op2);
2351 // If the operand is larger than the shift count type but the shift
2352 // count type has enough bits to represent any shift value, truncate
2353 // it now. This is a common case and it exposes the truncate to
2354 // optimization early.
2355 else if (STy.getSizeInBits() >=
2356 Log2_32_Ceil(Op2.getValueType().getSizeInBits()))
2357 Op2 = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2358 TLI.getShiftAmountTy(), Op2);
2359 // Otherwise we'll need to temporarily settle for some other
2360 // convenient type; type legalization will make adjustments as
2362 else if (PTy.bitsLT(Op2.getValueType()))
2363 Op2 = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2364 TLI.getPointerTy(), Op2);
2365 else if (PTy.bitsGT(Op2.getValueType()))
2366 Op2 = DAG.getNode(ISD::ANY_EXTEND, getCurDebugLoc(),
2367 TLI.getPointerTy(), Op2);
2370 setValue(&I, DAG.getNode(Opcode, getCurDebugLoc(),
2371 Op1.getValueType(), Op1, Op2));
2374 void SelectionDAGBuilder::visitICmp(const User &I) {
2375 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2376 if (const ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2377 predicate = IC->getPredicate();
2378 else if (const ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2379 predicate = ICmpInst::Predicate(IC->getPredicate());
2380 SDValue Op1 = getValue(I.getOperand(0));
2381 SDValue Op2 = getValue(I.getOperand(1));
2382 ISD::CondCode Opcode = getICmpCondCode(predicate);
2384 EVT DestVT = TLI.getValueType(I.getType());
2385 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Opcode));
2388 void SelectionDAGBuilder::visitFCmp(const User &I) {
2389 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2390 if (const FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2391 predicate = FC->getPredicate();
2392 else if (const ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2393 predicate = FCmpInst::Predicate(FC->getPredicate());
2394 SDValue Op1 = getValue(I.getOperand(0));
2395 SDValue Op2 = getValue(I.getOperand(1));
2396 ISD::CondCode Condition = getFCmpCondCode(predicate);
2397 EVT DestVT = TLI.getValueType(I.getType());
2398 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Condition));
2401 void SelectionDAGBuilder::visitSelect(const User &I) {
2402 SmallVector<EVT, 4> ValueVTs;
2403 ComputeValueVTs(TLI, I.getType(), ValueVTs);
2404 unsigned NumValues = ValueVTs.size();
2405 if (NumValues == 0) return;
2407 SmallVector<SDValue, 4> Values(NumValues);
2408 SDValue Cond = getValue(I.getOperand(0));
2409 SDValue TrueVal = getValue(I.getOperand(1));
2410 SDValue FalseVal = getValue(I.getOperand(2));
2412 for (unsigned i = 0; i != NumValues; ++i)
2413 Values[i] = DAG.getNode(ISD::SELECT, getCurDebugLoc(),
2414 TrueVal.getNode()->getValueType(TrueVal.getResNo()+i),
2416 SDValue(TrueVal.getNode(),
2417 TrueVal.getResNo() + i),
2418 SDValue(FalseVal.getNode(),
2419 FalseVal.getResNo() + i));
2421 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2422 DAG.getVTList(&ValueVTs[0], NumValues),
2423 &Values[0], NumValues));
2426 void SelectionDAGBuilder::visitTrunc(const User &I) {
2427 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2428 SDValue N = getValue(I.getOperand(0));
2429 EVT DestVT = TLI.getValueType(I.getType());
2430 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), DestVT, N));
2433 void SelectionDAGBuilder::visitZExt(const User &I) {
2434 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2435 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2436 SDValue N = getValue(I.getOperand(0));
2437 EVT DestVT = TLI.getValueType(I.getType());
2438 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(), DestVT, N));
2441 void SelectionDAGBuilder::visitSExt(const User &I) {
2442 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2443 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2444 SDValue N = getValue(I.getOperand(0));
2445 EVT DestVT = TLI.getValueType(I.getType());
2446 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurDebugLoc(), DestVT, N));
2449 void SelectionDAGBuilder::visitFPTrunc(const User &I) {
2450 // FPTrunc is never a no-op cast, no need to check
2451 SDValue N = getValue(I.getOperand(0));
2452 EVT DestVT = TLI.getValueType(I.getType());
2453 setValue(&I, DAG.getNode(ISD::FP_ROUND, getCurDebugLoc(),
2454 DestVT, N, DAG.getIntPtrConstant(0)));
2457 void SelectionDAGBuilder::visitFPExt(const User &I){
2458 // FPTrunc is never a no-op cast, no need to check
2459 SDValue N = getValue(I.getOperand(0));
2460 EVT DestVT = TLI.getValueType(I.getType());
2461 setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurDebugLoc(), DestVT, N));
2464 void SelectionDAGBuilder::visitFPToUI(const User &I) {
2465 // FPToUI is never a no-op cast, no need to check
2466 SDValue N = getValue(I.getOperand(0));
2467 EVT DestVT = TLI.getValueType(I.getType());
2468 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurDebugLoc(), DestVT, N));
2471 void SelectionDAGBuilder::visitFPToSI(const User &I) {
2472 // FPToSI is never a no-op cast, no need to check
2473 SDValue N = getValue(I.getOperand(0));
2474 EVT DestVT = TLI.getValueType(I.getType());
2475 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurDebugLoc(), DestVT, N));
2478 void SelectionDAGBuilder::visitUIToFP(const User &I) {
2479 // UIToFP is never a no-op cast, no need to check
2480 SDValue N = getValue(I.getOperand(0));
2481 EVT DestVT = TLI.getValueType(I.getType());
2482 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurDebugLoc(), DestVT, N));
2485 void SelectionDAGBuilder::visitSIToFP(const User &I){
2486 // SIToFP is never a no-op cast, no need to check
2487 SDValue N = getValue(I.getOperand(0));
2488 EVT DestVT = TLI.getValueType(I.getType());
2489 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurDebugLoc(), DestVT, N));
2492 void SelectionDAGBuilder::visitPtrToInt(const User &I) {
2493 // What to do depends on the size of the integer and the size of the pointer.
2494 // We can either truncate, zero extend, or no-op, accordingly.
2495 SDValue N = getValue(I.getOperand(0));
2496 EVT DestVT = TLI.getValueType(I.getType());
2497 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT));
2500 void SelectionDAGBuilder::visitIntToPtr(const User &I) {
2501 // What to do depends on the size of the integer and the size of the pointer.
2502 // We can either truncate, zero extend, or no-op, accordingly.
2503 SDValue N = getValue(I.getOperand(0));
2504 EVT DestVT = TLI.getValueType(I.getType());
2505 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT));
2508 void SelectionDAGBuilder::visitBitCast(const User &I) {
2509 SDValue N = getValue(I.getOperand(0));
2510 EVT DestVT = TLI.getValueType(I.getType());
2512 // BitCast assures us that source and destination are the same size so this is
2513 // either a BIT_CONVERT or a no-op.
2514 if (DestVT != N.getValueType())
2515 setValue(&I, DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
2516 DestVT, N)); // convert types.
2518 setValue(&I, N); // noop cast.
2521 void SelectionDAGBuilder::visitInsertElement(const User &I) {
2522 SDValue InVec = getValue(I.getOperand(0));
2523 SDValue InVal = getValue(I.getOperand(1));
2524 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2526 getValue(I.getOperand(2)));
2527 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurDebugLoc(),
2528 TLI.getValueType(I.getType()),
2529 InVec, InVal, InIdx));
2532 void SelectionDAGBuilder::visitExtractElement(const User &I) {
2533 SDValue InVec = getValue(I.getOperand(0));
2534 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2536 getValue(I.getOperand(1)));
2537 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2538 TLI.getValueType(I.getType()), InVec, InIdx));
2541 // Utility for visitShuffleVector - Returns true if the mask is mask starting
2542 // from SIndx and increasing to the element length (undefs are allowed).
2543 static bool SequentialMask(SmallVectorImpl<int> &Mask, unsigned SIndx) {
2544 unsigned MaskNumElts = Mask.size();
2545 for (unsigned i = 0; i != MaskNumElts; ++i)
2546 if ((Mask[i] >= 0) && (Mask[i] != (int)(i + SIndx)))
2551 void SelectionDAGBuilder::visitShuffleVector(const User &I) {
2552 SmallVector<int, 8> Mask;
2553 SDValue Src1 = getValue(I.getOperand(0));
2554 SDValue Src2 = getValue(I.getOperand(1));
2556 // Convert the ConstantVector mask operand into an array of ints, with -1
2557 // representing undef values.
2558 SmallVector<Constant*, 8> MaskElts;
2559 cast<Constant>(I.getOperand(2))->getVectorElements(MaskElts);
2560 unsigned MaskNumElts = MaskElts.size();
2561 for (unsigned i = 0; i != MaskNumElts; ++i) {
2562 if (isa<UndefValue>(MaskElts[i]))
2565 Mask.push_back(cast<ConstantInt>(MaskElts[i])->getSExtValue());
2568 EVT VT = TLI.getValueType(I.getType());
2569 EVT SrcVT = Src1.getValueType();
2570 unsigned SrcNumElts = SrcVT.getVectorNumElements();
2572 if (SrcNumElts == MaskNumElts) {
2573 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2578 // Normalize the shuffle vector since mask and vector length don't match.
2579 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) {
2580 // Mask is longer than the source vectors and is a multiple of the source
2581 // vectors. We can use concatenate vector to make the mask and vectors
2583 if (SrcNumElts*2 == MaskNumElts && SequentialMask(Mask, 0)) {
2584 // The shuffle is concatenating two vectors together.
2585 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurDebugLoc(),
2590 // Pad both vectors with undefs to make them the same length as the mask.
2591 unsigned NumConcat = MaskNumElts / SrcNumElts;
2592 bool Src1U = Src1.getOpcode() == ISD::UNDEF;
2593 bool Src2U = Src2.getOpcode() == ISD::UNDEF;
2594 SDValue UndefVal = DAG.getUNDEF(SrcVT);
2596 SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal);
2597 SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal);
2601 Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2602 getCurDebugLoc(), VT,
2603 &MOps1[0], NumConcat);
2604 Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2605 getCurDebugLoc(), VT,
2606 &MOps2[0], NumConcat);
2608 // Readjust mask for new input vector length.
2609 SmallVector<int, 8> MappedOps;
2610 for (unsigned i = 0; i != MaskNumElts; ++i) {
2612 if (Idx < (int)SrcNumElts)
2613 MappedOps.push_back(Idx);
2615 MappedOps.push_back(Idx + MaskNumElts - SrcNumElts);
2618 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2623 if (SrcNumElts > MaskNumElts) {
2624 // Analyze the access pattern of the vector to see if we can extract
2625 // two subvectors and do the shuffle. The analysis is done by calculating
2626 // the range of elements the mask access on both vectors.
2627 int MinRange[2] = { SrcNumElts+1, SrcNumElts+1};
2628 int MaxRange[2] = {-1, -1};
2630 for (unsigned i = 0; i != MaskNumElts; ++i) {
2636 if (Idx >= (int)SrcNumElts) {
2640 if (Idx > MaxRange[Input])
2641 MaxRange[Input] = Idx;
2642 if (Idx < MinRange[Input])
2643 MinRange[Input] = Idx;
2646 // Check if the access is smaller than the vector size and can we find
2647 // a reasonable extract index.
2648 int RangeUse[2] = { 2, 2 }; // 0 = Unused, 1 = Extract, 2 = Can not
2650 int StartIdx[2]; // StartIdx to extract from
2651 for (int Input=0; Input < 2; ++Input) {
2652 if (MinRange[Input] == (int)(SrcNumElts+1) && MaxRange[Input] == -1) {
2653 RangeUse[Input] = 0; // Unused
2654 StartIdx[Input] = 0;
2655 } else if (MaxRange[Input] - MinRange[Input] < (int)MaskNumElts) {
2656 // Fits within range but we should see if we can find a good
2657 // start index that is a multiple of the mask length.
2658 if (MaxRange[Input] < (int)MaskNumElts) {
2659 RangeUse[Input] = 1; // Extract from beginning of the vector
2660 StartIdx[Input] = 0;
2662 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts;
2663 if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts &&
2664 StartIdx[Input] + MaskNumElts < SrcNumElts)
2665 RangeUse[Input] = 1; // Extract from a multiple of the mask length.
2670 if (RangeUse[0] == 0 && RangeUse[1] == 0) {
2671 setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used.
2674 else if (RangeUse[0] < 2 && RangeUse[1] < 2) {
2675 // Extract appropriate subvector and generate a vector shuffle
2676 for (int Input=0; Input < 2; ++Input) {
2677 SDValue &Src = Input == 0 ? Src1 : Src2;
2678 if (RangeUse[Input] == 0)
2679 Src = DAG.getUNDEF(VT);
2681 Src = DAG.getNode(ISD::EXTRACT_SUBVECTOR, getCurDebugLoc(), VT,
2682 Src, DAG.getIntPtrConstant(StartIdx[Input]));
2685 // Calculate new mask.
2686 SmallVector<int, 8> MappedOps;
2687 for (unsigned i = 0; i != MaskNumElts; ++i) {
2690 MappedOps.push_back(Idx);
2691 else if (Idx < (int)SrcNumElts)
2692 MappedOps.push_back(Idx - StartIdx[0]);
2694 MappedOps.push_back(Idx - SrcNumElts - StartIdx[1] + MaskNumElts);
2697 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2703 // We can't use either concat vectors or extract subvectors so fall back to
2704 // replacing the shuffle with extract and build vector.
2705 // to insert and build vector.
2706 EVT EltVT = VT.getVectorElementType();
2707 EVT PtrVT = TLI.getPointerTy();
2708 SmallVector<SDValue,8> Ops;
2709 for (unsigned i = 0; i != MaskNumElts; ++i) {
2711 Ops.push_back(DAG.getUNDEF(EltVT));
2716 if (Idx < (int)SrcNumElts)
2717 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2718 EltVT, Src1, DAG.getConstant(Idx, PtrVT));
2720 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2722 DAG.getConstant(Idx - SrcNumElts, PtrVT));
2728 setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
2729 VT, &Ops[0], Ops.size()));
2732 void SelectionDAGBuilder::visitInsertValue(const InsertValueInst &I) {
2733 const Value *Op0 = I.getOperand(0);
2734 const Value *Op1 = I.getOperand(1);
2735 const Type *AggTy = I.getType();
2736 const Type *ValTy = Op1->getType();
2737 bool IntoUndef = isa<UndefValue>(Op0);
2738 bool FromUndef = isa<UndefValue>(Op1);
2740 unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy,
2741 I.idx_begin(), I.idx_end());
2743 SmallVector<EVT, 4> AggValueVTs;
2744 ComputeValueVTs(TLI, AggTy, AggValueVTs);
2745 SmallVector<EVT, 4> ValValueVTs;
2746 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2748 unsigned NumAggValues = AggValueVTs.size();
2749 unsigned NumValValues = ValValueVTs.size();
2750 SmallVector<SDValue, 4> Values(NumAggValues);
2752 SDValue Agg = getValue(Op0);
2753 SDValue Val = getValue(Op1);
2755 // Copy the beginning value(s) from the original aggregate.
2756 for (; i != LinearIndex; ++i)
2757 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2758 SDValue(Agg.getNode(), Agg.getResNo() + i);
2759 // Copy values from the inserted value(s).
2760 for (; i != LinearIndex + NumValValues; ++i)
2761 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2762 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex);
2763 // Copy remaining value(s) from the original aggregate.
2764 for (; i != NumAggValues; ++i)
2765 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2766 SDValue(Agg.getNode(), Agg.getResNo() + i);
2768 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2769 DAG.getVTList(&AggValueVTs[0], NumAggValues),
2770 &Values[0], NumAggValues));
2773 void SelectionDAGBuilder::visitExtractValue(const ExtractValueInst &I) {
2774 const Value *Op0 = I.getOperand(0);
2775 const Type *AggTy = Op0->getType();
2776 const Type *ValTy = I.getType();
2777 bool OutOfUndef = isa<UndefValue>(Op0);
2779 unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy,
2780 I.idx_begin(), I.idx_end());
2782 SmallVector<EVT, 4> ValValueVTs;
2783 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2785 unsigned NumValValues = ValValueVTs.size();
2786 SmallVector<SDValue, 4> Values(NumValValues);
2788 SDValue Agg = getValue(Op0);
2789 // Copy out the selected value(s).
2790 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i)
2791 Values[i - LinearIndex] =
2793 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) :
2794 SDValue(Agg.getNode(), Agg.getResNo() + i);
2796 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2797 DAG.getVTList(&ValValueVTs[0], NumValValues),
2798 &Values[0], NumValValues));
2801 void SelectionDAGBuilder::visitGetElementPtr(const User &I) {
2802 SDValue N = getValue(I.getOperand(0));
2803 const Type *Ty = I.getOperand(0)->getType();
2805 for (GetElementPtrInst::const_op_iterator OI = I.op_begin()+1, E = I.op_end();
2807 const Value *Idx = *OI;
2808 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
2809 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
2812 uint64_t Offset = TD->getStructLayout(StTy)->getElementOffset(Field);
2813 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
2814 DAG.getIntPtrConstant(Offset));
2817 Ty = StTy->getElementType(Field);
2818 } else if (const UnionType *UnTy = dyn_cast<UnionType>(Ty)) {
2819 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
2821 // Offset canonically 0 for unions, but type changes
2822 Ty = UnTy->getElementType(Field);
2824 Ty = cast<SequentialType>(Ty)->getElementType();
2826 // If this is a constant subscript, handle it quickly.
2827 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
2828 if (CI->isZero()) continue;
2830 TD->getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
2832 EVT PTy = TLI.getPointerTy();
2833 unsigned PtrBits = PTy.getSizeInBits();
2835 OffsVal = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2837 DAG.getConstant(Offs, MVT::i64));
2839 OffsVal = DAG.getIntPtrConstant(Offs);
2841 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
2846 // N = N + Idx * ElementSize;
2847 APInt ElementSize = APInt(TLI.getPointerTy().getSizeInBits(),
2848 TD->getTypeAllocSize(Ty));
2849 SDValue IdxN = getValue(Idx);
2851 // If the index is smaller or larger than intptr_t, truncate or extend
2853 IdxN = DAG.getSExtOrTrunc(IdxN, getCurDebugLoc(), N.getValueType());
2855 // If this is a multiply by a power of two, turn it into a shl
2856 // immediately. This is a very common case.
2857 if (ElementSize != 1) {
2858 if (ElementSize.isPowerOf2()) {
2859 unsigned Amt = ElementSize.logBase2();
2860 IdxN = DAG.getNode(ISD::SHL, getCurDebugLoc(),
2861 N.getValueType(), IdxN,
2862 DAG.getConstant(Amt, TLI.getPointerTy()));
2864 SDValue Scale = DAG.getConstant(ElementSize, TLI.getPointerTy());
2865 IdxN = DAG.getNode(ISD::MUL, getCurDebugLoc(),
2866 N.getValueType(), IdxN, Scale);
2870 N = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2871 N.getValueType(), N, IdxN);
2878 void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) {
2879 // If this is a fixed sized alloca in the entry block of the function,
2880 // allocate it statically on the stack.
2881 if (FuncInfo.StaticAllocaMap.count(&I))
2882 return; // getValue will auto-populate this.
2884 const Type *Ty = I.getAllocatedType();
2885 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
2887 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty),
2890 SDValue AllocSize = getValue(I.getArraySize());
2892 EVT IntPtr = TLI.getPointerTy();
2893 if (AllocSize.getValueType() != IntPtr)
2894 AllocSize = DAG.getZExtOrTrunc(AllocSize, getCurDebugLoc(), IntPtr);
2896 AllocSize = DAG.getNode(ISD::MUL, getCurDebugLoc(), IntPtr,
2898 DAG.getConstant(TySize, IntPtr));
2900 // Handle alignment. If the requested alignment is less than or equal to
2901 // the stack alignment, ignore it. If the size is greater than or equal to
2902 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
2903 unsigned StackAlign = TM.getFrameInfo()->getStackAlignment();
2904 if (Align <= StackAlign)
2907 // Round the size of the allocation up to the stack alignment size
2908 // by add SA-1 to the size.
2909 AllocSize = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2910 AllocSize.getValueType(), AllocSize,
2911 DAG.getIntPtrConstant(StackAlign-1));
2913 // Mask out the low bits for alignment purposes.
2914 AllocSize = DAG.getNode(ISD::AND, getCurDebugLoc(),
2915 AllocSize.getValueType(), AllocSize,
2916 DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1)));
2918 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) };
2919 SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other);
2920 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, getCurDebugLoc(),
2923 DAG.setRoot(DSA.getValue(1));
2925 // Inform the Frame Information that we have just allocated a variable-sized
2927 FuncInfo.MF->getFrameInfo()->CreateVariableSizedObject(Align ? Align : 1);
2930 void SelectionDAGBuilder::visitLoad(const LoadInst &I) {
2931 const Value *SV = I.getOperand(0);
2932 SDValue Ptr = getValue(SV);
2934 const Type *Ty = I.getType();
2936 bool isVolatile = I.isVolatile();
2937 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
2938 unsigned Alignment = I.getAlignment();
2940 SmallVector<EVT, 4> ValueVTs;
2941 SmallVector<uint64_t, 4> Offsets;
2942 ComputeValueVTs(TLI, Ty, ValueVTs, &Offsets);
2943 unsigned NumValues = ValueVTs.size();
2948 bool ConstantMemory = false;
2950 // Serialize volatile loads with other side effects.
2952 else if (AA->pointsToConstantMemory(SV)) {
2953 // Do not serialize (non-volatile) loads of constant memory with anything.
2954 Root = DAG.getEntryNode();
2955 ConstantMemory = true;
2957 // Do not serialize non-volatile loads against each other.
2958 Root = DAG.getRoot();
2961 SmallVector<SDValue, 4> Values(NumValues);
2962 SmallVector<SDValue, 4> Chains(NumValues);
2963 EVT PtrVT = Ptr.getValueType();
2964 for (unsigned i = 0; i != NumValues; ++i) {
2965 SDValue A = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2967 DAG.getConstant(Offsets[i], PtrVT));
2968 SDValue L = DAG.getLoad(ValueVTs[i], getCurDebugLoc(), Root,
2969 A, SV, Offsets[i], isVolatile,
2970 isNonTemporal, Alignment);
2973 Chains[i] = L.getValue(1);
2976 if (!ConstantMemory) {
2977 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
2978 MVT::Other, &Chains[0], NumValues);
2982 PendingLoads.push_back(Chain);
2985 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2986 DAG.getVTList(&ValueVTs[0], NumValues),
2987 &Values[0], NumValues));
2990 void SelectionDAGBuilder::visitStore(const StoreInst &I) {
2991 const Value *SrcV = I.getOperand(0);
2992 const Value *PtrV = I.getOperand(1);
2994 SmallVector<EVT, 4> ValueVTs;
2995 SmallVector<uint64_t, 4> Offsets;
2996 ComputeValueVTs(TLI, SrcV->getType(), ValueVTs, &Offsets);
2997 unsigned NumValues = ValueVTs.size();
3001 // Get the lowered operands. Note that we do this after
3002 // checking if NumResults is zero, because with zero results
3003 // the operands won't have values in the map.
3004 SDValue Src = getValue(SrcV);
3005 SDValue Ptr = getValue(PtrV);
3007 SDValue Root = getRoot();
3008 SmallVector<SDValue, 4> Chains(NumValues);
3009 EVT PtrVT = Ptr.getValueType();
3010 bool isVolatile = I.isVolatile();
3011 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
3012 unsigned Alignment = I.getAlignment();
3014 for (unsigned i = 0; i != NumValues; ++i) {
3015 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT, Ptr,
3016 DAG.getConstant(Offsets[i], PtrVT));
3017 Chains[i] = DAG.getStore(Root, getCurDebugLoc(),
3018 SDValue(Src.getNode(), Src.getResNo() + i),
3019 Add, PtrV, Offsets[i], isVolatile,
3020 isNonTemporal, Alignment);
3023 DAG.setRoot(DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3024 MVT::Other, &Chains[0], NumValues));
3027 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
3029 void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I,
3030 unsigned Intrinsic) {
3031 bool HasChain = !I.doesNotAccessMemory();
3032 bool OnlyLoad = HasChain && I.onlyReadsMemory();
3034 // Build the operand list.
3035 SmallVector<SDValue, 8> Ops;
3036 if (HasChain) { // If this intrinsic has side-effects, chainify it.
3038 // We don't need to serialize loads against other loads.
3039 Ops.push_back(DAG.getRoot());
3041 Ops.push_back(getRoot());
3045 // Info is set by getTgtMemInstrinsic
3046 TargetLowering::IntrinsicInfo Info;
3047 bool IsTgtIntrinsic = TLI.getTgtMemIntrinsic(Info, I, Intrinsic);
3049 // Add the intrinsic ID as an integer operand if it's not a target intrinsic.
3050 if (!IsTgtIntrinsic)
3051 Ops.push_back(DAG.getConstant(Intrinsic, TLI.getPointerTy()));
3053 // Add all operands of the call to the operand list.
3054 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) {
3055 SDValue Op = getValue(I.getArgOperand(i));
3056 assert(TLI.isTypeLegal(Op.getValueType()) &&
3057 "Intrinsic uses a non-legal type?");
3061 SmallVector<EVT, 4> ValueVTs;
3062 ComputeValueVTs(TLI, I.getType(), ValueVTs);
3064 for (unsigned Val = 0, E = ValueVTs.size(); Val != E; ++Val) {
3065 assert(TLI.isTypeLegal(ValueVTs[Val]) &&
3066 "Intrinsic uses a non-legal type?");
3071 ValueVTs.push_back(MVT::Other);
3073 SDVTList VTs = DAG.getVTList(ValueVTs.data(), ValueVTs.size());
3077 if (IsTgtIntrinsic) {
3078 // This is target intrinsic that touches memory
3079 Result = DAG.getMemIntrinsicNode(Info.opc, getCurDebugLoc(),
3080 VTs, &Ops[0], Ops.size(),
3081 Info.memVT, Info.ptrVal, Info.offset,
3082 Info.align, Info.vol,
3083 Info.readMem, Info.writeMem);
3084 } else if (!HasChain) {
3085 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurDebugLoc(),
3086 VTs, &Ops[0], Ops.size());
3087 } else if (!I.getType()->isVoidTy()) {
3088 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurDebugLoc(),
3089 VTs, &Ops[0], Ops.size());
3091 Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurDebugLoc(),
3092 VTs, &Ops[0], Ops.size());
3096 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1);
3098 PendingLoads.push_back(Chain);
3103 if (!I.getType()->isVoidTy()) {
3104 if (const VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
3105 EVT VT = TLI.getValueType(PTy);
3106 Result = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(), VT, Result);
3109 setValue(&I, Result);
3113 /// GetSignificand - Get the significand and build it into a floating-point
3114 /// number with exponent of 1:
3116 /// Op = (Op & 0x007fffff) | 0x3f800000;
3118 /// where Op is the hexidecimal representation of floating point value.
3120 GetSignificand(SelectionDAG &DAG, SDValue Op, DebugLoc dl) {
3121 SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3122 DAG.getConstant(0x007fffff, MVT::i32));
3123 SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1,
3124 DAG.getConstant(0x3f800000, MVT::i32));
3125 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t2);
3128 /// GetExponent - Get the exponent:
3130 /// (float)(int)(((Op & 0x7f800000) >> 23) - 127);
3132 /// where Op is the hexidecimal representation of floating point value.
3134 GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI,
3136 SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3137 DAG.getConstant(0x7f800000, MVT::i32));
3138 SDValue t1 = DAG.getNode(ISD::SRL, dl, MVT::i32, t0,
3139 DAG.getConstant(23, TLI.getPointerTy()));
3140 SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1,
3141 DAG.getConstant(127, MVT::i32));
3142 return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2);
3145 /// getF32Constant - Get 32-bit floating point constant.
3147 getF32Constant(SelectionDAG &DAG, unsigned Flt) {
3148 return DAG.getConstantFP(APFloat(APInt(32, Flt)), MVT::f32);
3151 /// Inlined utility function to implement binary input atomic intrinsics for
3152 /// visitIntrinsicCall: I is a call instruction
3153 /// Op is the associated NodeType for I
3155 SelectionDAGBuilder::implVisitBinaryAtomic(const CallInst& I,
3157 SDValue Root = getRoot();
3159 DAG.getAtomic(Op, getCurDebugLoc(),
3160 getValue(I.getArgOperand(1)).getValueType().getSimpleVT(),
3162 getValue(I.getArgOperand(0)),
3163 getValue(I.getArgOperand(1)),
3164 I.getArgOperand(0));
3166 DAG.setRoot(L.getValue(1));
3170 // implVisitAluOverflow - Lower arithmetic overflow instrinsics.
3172 SelectionDAGBuilder::implVisitAluOverflow(const CallInst &I, ISD::NodeType Op) {
3173 SDValue Op1 = getValue(I.getArgOperand(0));
3174 SDValue Op2 = getValue(I.getArgOperand(1));
3176 SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1);
3177 setValue(&I, DAG.getNode(Op, getCurDebugLoc(), VTs, Op1, Op2));
3181 /// visitExp - Lower an exp intrinsic. Handles the special sequences for
3182 /// limited-precision mode.
3184 SelectionDAGBuilder::visitExp(const CallInst &I) {
3186 DebugLoc dl = getCurDebugLoc();
3188 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3189 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3190 SDValue Op = getValue(I.getArgOperand(0));
3192 // Put the exponent in the right bit position for later addition to the
3195 // #define LOG2OFe 1.4426950f
3196 // IntegerPartOfX = ((int32_t)(X * LOG2OFe));
3197 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3198 getF32Constant(DAG, 0x3fb8aa3b));
3199 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3201 // FractionalPartOfX = (X * LOG2OFe) - (float)IntegerPartOfX;
3202 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3203 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3205 // IntegerPartOfX <<= 23;
3206 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3207 DAG.getConstant(23, TLI.getPointerTy()));
3209 if (LimitFloatPrecision <= 6) {
3210 // For floating-point precision of 6:
3212 // TwoToFractionalPartOfX =
3214 // (0.735607626f + 0.252464424f * x) * x;
3216 // error 0.0144103317, which is 6 bits
3217 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3218 getF32Constant(DAG, 0x3e814304));
3219 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3220 getF32Constant(DAG, 0x3f3c50c8));
3221 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3222 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3223 getF32Constant(DAG, 0x3f7f5e7e));
3224 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,MVT::i32, t5);
3226 // Add the exponent into the result in integer domain.
3227 SDValue t6 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3228 TwoToFracPartOfX, IntegerPartOfX);
3230 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t6);
3231 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3232 // For floating-point precision of 12:
3234 // TwoToFractionalPartOfX =
3237 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3239 // 0.000107046256 error, which is 13 to 14 bits
3240 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3241 getF32Constant(DAG, 0x3da235e3));
3242 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3243 getF32Constant(DAG, 0x3e65b8f3));
3244 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3245 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3246 getF32Constant(DAG, 0x3f324b07));
3247 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3248 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3249 getF32Constant(DAG, 0x3f7ff8fd));
3250 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,MVT::i32, t7);
3252 // Add the exponent into the result in integer domain.
3253 SDValue t8 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3254 TwoToFracPartOfX, IntegerPartOfX);
3256 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t8);
3257 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3258 // For floating-point precision of 18:
3260 // TwoToFractionalPartOfX =
3264 // (0.554906021e-1f +
3265 // (0.961591928e-2f +
3266 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3268 // error 2.47208000*10^(-7), which is better than 18 bits
3269 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3270 getF32Constant(DAG, 0x3924b03e));
3271 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3272 getF32Constant(DAG, 0x3ab24b87));
3273 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3274 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3275 getF32Constant(DAG, 0x3c1d8c17));
3276 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3277 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3278 getF32Constant(DAG, 0x3d634a1d));
3279 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3280 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3281 getF32Constant(DAG, 0x3e75fe14));
3282 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3283 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3284 getF32Constant(DAG, 0x3f317234));
3285 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3286 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3287 getF32Constant(DAG, 0x3f800000));
3288 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,
3291 // Add the exponent into the result in integer domain.
3292 SDValue t14 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3293 TwoToFracPartOfX, IntegerPartOfX);
3295 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t14);
3298 // No special expansion.
3299 result = DAG.getNode(ISD::FEXP, dl,
3300 getValue(I.getArgOperand(0)).getValueType(),
3301 getValue(I.getArgOperand(0)));
3304 setValue(&I, result);
3307 /// visitLog - Lower a log intrinsic. Handles the special sequences for
3308 /// limited-precision mode.
3310 SelectionDAGBuilder::visitLog(const CallInst &I) {
3312 DebugLoc dl = getCurDebugLoc();
3314 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3315 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3316 SDValue Op = getValue(I.getArgOperand(0));
3317 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
3319 // Scale the exponent by log(2) [0.69314718f].
3320 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3321 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3322 getF32Constant(DAG, 0x3f317218));
3324 // Get the significand and build it into a floating-point number with
3326 SDValue X = GetSignificand(DAG, Op1, dl);
3328 if (LimitFloatPrecision <= 6) {
3329 // For floating-point precision of 6:
3333 // (1.4034025f - 0.23903021f * x) * x;
3335 // error 0.0034276066, which is better than 8 bits
3336 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3337 getF32Constant(DAG, 0xbe74c456));
3338 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3339 getF32Constant(DAG, 0x3fb3a2b1));
3340 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3341 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3342 getF32Constant(DAG, 0x3f949a29));
3344 result = DAG.getNode(ISD::FADD, dl,
3345 MVT::f32, LogOfExponent, LogOfMantissa);
3346 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3347 // For floating-point precision of 12:
3353 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x;
3355 // error 0.000061011436, which is 14 bits
3356 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3357 getF32Constant(DAG, 0xbd67b6d6));
3358 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3359 getF32Constant(DAG, 0x3ee4f4b8));
3360 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3361 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3362 getF32Constant(DAG, 0x3fbc278b));
3363 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3364 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3365 getF32Constant(DAG, 0x40348e95));
3366 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3367 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3368 getF32Constant(DAG, 0x3fdef31a));
3370 result = DAG.getNode(ISD::FADD, dl,
3371 MVT::f32, LogOfExponent, LogOfMantissa);
3372 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3373 // For floating-point precision of 18:
3381 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x;
3383 // error 0.0000023660568, which is better than 18 bits
3384 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3385 getF32Constant(DAG, 0xbc91e5ac));
3386 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3387 getF32Constant(DAG, 0x3e4350aa));
3388 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3389 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3390 getF32Constant(DAG, 0x3f60d3e3));
3391 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3392 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3393 getF32Constant(DAG, 0x4011cdf0));
3394 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3395 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3396 getF32Constant(DAG, 0x406cfd1c));
3397 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3398 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3399 getF32Constant(DAG, 0x408797cb));
3400 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3401 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3402 getF32Constant(DAG, 0x4006dcab));
3404 result = DAG.getNode(ISD::FADD, dl,
3405 MVT::f32, LogOfExponent, LogOfMantissa);
3408 // No special expansion.
3409 result = DAG.getNode(ISD::FLOG, dl,
3410 getValue(I.getArgOperand(0)).getValueType(),
3411 getValue(I.getArgOperand(0)));
3414 setValue(&I, result);
3417 /// visitLog2 - Lower a log2 intrinsic. Handles the special sequences for
3418 /// limited-precision mode.
3420 SelectionDAGBuilder::visitLog2(const CallInst &I) {
3422 DebugLoc dl = getCurDebugLoc();
3424 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3425 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3426 SDValue Op = getValue(I.getArgOperand(0));
3427 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
3429 // Get the exponent.
3430 SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl);
3432 // Get the significand and build it into a floating-point number with
3434 SDValue X = GetSignificand(DAG, Op1, dl);
3436 // Different possible minimax approximations of significand in
3437 // floating-point for various degrees of accuracy over [1,2].
3438 if (LimitFloatPrecision <= 6) {
3439 // For floating-point precision of 6:
3441 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x;
3443 // error 0.0049451742, which is more than 7 bits
3444 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3445 getF32Constant(DAG, 0xbeb08fe0));
3446 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3447 getF32Constant(DAG, 0x40019463));
3448 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3449 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3450 getF32Constant(DAG, 0x3fd6633d));
3452 result = DAG.getNode(ISD::FADD, dl,
3453 MVT::f32, LogOfExponent, Log2ofMantissa);
3454 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3455 // For floating-point precision of 12:
3461 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x;
3463 // error 0.0000876136000, which is better than 13 bits
3464 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3465 getF32Constant(DAG, 0xbda7262e));
3466 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3467 getF32Constant(DAG, 0x3f25280b));
3468 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3469 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3470 getF32Constant(DAG, 0x4007b923));
3471 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3472 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3473 getF32Constant(DAG, 0x40823e2f));
3474 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3475 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3476 getF32Constant(DAG, 0x4020d29c));
3478 result = DAG.getNode(ISD::FADD, dl,
3479 MVT::f32, LogOfExponent, Log2ofMantissa);
3480 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3481 // For floating-point precision of 18:
3490 // 0.25691327e-1f * x) * x) * x) * x) * x) * x;
3492 // error 0.0000018516, which is better than 18 bits
3493 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3494 getF32Constant(DAG, 0xbcd2769e));
3495 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3496 getF32Constant(DAG, 0x3e8ce0b9));
3497 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3498 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3499 getF32Constant(DAG, 0x3fa22ae7));
3500 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3501 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3502 getF32Constant(DAG, 0x40525723));
3503 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3504 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3505 getF32Constant(DAG, 0x40aaf200));
3506 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3507 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3508 getF32Constant(DAG, 0x40c39dad));
3509 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3510 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3511 getF32Constant(DAG, 0x4042902c));
3513 result = DAG.getNode(ISD::FADD, dl,
3514 MVT::f32, LogOfExponent, Log2ofMantissa);
3517 // No special expansion.
3518 result = DAG.getNode(ISD::FLOG2, dl,
3519 getValue(I.getArgOperand(0)).getValueType(),
3520 getValue(I.getArgOperand(0)));
3523 setValue(&I, result);
3526 /// visitLog10 - Lower a log10 intrinsic. Handles the special sequences for
3527 /// limited-precision mode.
3529 SelectionDAGBuilder::visitLog10(const CallInst &I) {
3531 DebugLoc dl = getCurDebugLoc();
3533 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3534 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3535 SDValue Op = getValue(I.getArgOperand(0));
3536 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
3538 // Scale the exponent by log10(2) [0.30102999f].
3539 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3540 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3541 getF32Constant(DAG, 0x3e9a209a));
3543 // Get the significand and build it into a floating-point number with
3545 SDValue X = GetSignificand(DAG, Op1, dl);
3547 if (LimitFloatPrecision <= 6) {
3548 // For floating-point precision of 6:
3550 // Log10ofMantissa =
3552 // (0.60948995f - 0.10380950f * x) * x;
3554 // error 0.0014886165, which is 6 bits
3555 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3556 getF32Constant(DAG, 0xbdd49a13));
3557 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3558 getF32Constant(DAG, 0x3f1c0789));
3559 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3560 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3561 getF32Constant(DAG, 0x3f011300));
3563 result = DAG.getNode(ISD::FADD, dl,
3564 MVT::f32, LogOfExponent, Log10ofMantissa);
3565 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3566 // For floating-point precision of 12:
3568 // Log10ofMantissa =
3571 // (-0.31664806f + 0.47637168e-1f * x) * x) * x;
3573 // error 0.00019228036, which is better than 12 bits
3574 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3575 getF32Constant(DAG, 0x3d431f31));
3576 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
3577 getF32Constant(DAG, 0x3ea21fb2));
3578 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3579 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3580 getF32Constant(DAG, 0x3f6ae232));
3581 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3582 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
3583 getF32Constant(DAG, 0x3f25f7c3));
3585 result = DAG.getNode(ISD::FADD, dl,
3586 MVT::f32, LogOfExponent, Log10ofMantissa);
3587 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3588 // For floating-point precision of 18:
3590 // Log10ofMantissa =
3595 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x;
3597 // error 0.0000037995730, which is better than 18 bits
3598 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3599 getF32Constant(DAG, 0x3c5d51ce));
3600 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
3601 getF32Constant(DAG, 0x3e00685a));
3602 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3603 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3604 getF32Constant(DAG, 0x3efb6798));
3605 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3606 SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
3607 getF32Constant(DAG, 0x3f88d192));
3608 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3609 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3610 getF32Constant(DAG, 0x3fc4316c));
3611 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3612 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8,
3613 getF32Constant(DAG, 0x3f57ce70));
3615 result = DAG.getNode(ISD::FADD, dl,
3616 MVT::f32, LogOfExponent, Log10ofMantissa);
3619 // No special expansion.
3620 result = DAG.getNode(ISD::FLOG10, dl,
3621 getValue(I.getArgOperand(0)).getValueType(),
3622 getValue(I.getArgOperand(0)));
3625 setValue(&I, result);
3628 /// visitExp2 - Lower an exp2 intrinsic. Handles the special sequences for
3629 /// limited-precision mode.
3631 SelectionDAGBuilder::visitExp2(const CallInst &I) {
3633 DebugLoc dl = getCurDebugLoc();
3635 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3636 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3637 SDValue Op = getValue(I.getArgOperand(0));
3639 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Op);
3641 // FractionalPartOfX = x - (float)IntegerPartOfX;
3642 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3643 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, Op, t1);
3645 // IntegerPartOfX <<= 23;
3646 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3647 DAG.getConstant(23, TLI.getPointerTy()));
3649 if (LimitFloatPrecision <= 6) {
3650 // For floating-point precision of 6:
3652 // TwoToFractionalPartOfX =
3654 // (0.735607626f + 0.252464424f * x) * x;
3656 // error 0.0144103317, which is 6 bits
3657 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3658 getF32Constant(DAG, 0x3e814304));
3659 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3660 getF32Constant(DAG, 0x3f3c50c8));
3661 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3662 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3663 getF32Constant(DAG, 0x3f7f5e7e));
3664 SDValue t6 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t5);
3665 SDValue TwoToFractionalPartOfX =
3666 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
3668 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3669 MVT::f32, TwoToFractionalPartOfX);
3670 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3671 // For floating-point precision of 12:
3673 // TwoToFractionalPartOfX =
3676 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3678 // error 0.000107046256, which is 13 to 14 bits
3679 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3680 getF32Constant(DAG, 0x3da235e3));
3681 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3682 getF32Constant(DAG, 0x3e65b8f3));
3683 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3684 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3685 getF32Constant(DAG, 0x3f324b07));
3686 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3687 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3688 getF32Constant(DAG, 0x3f7ff8fd));
3689 SDValue t8 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t7);
3690 SDValue TwoToFractionalPartOfX =
3691 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
3693 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3694 MVT::f32, TwoToFractionalPartOfX);
3695 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3696 // For floating-point precision of 18:
3698 // TwoToFractionalPartOfX =
3702 // (0.554906021e-1f +
3703 // (0.961591928e-2f +
3704 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3705 // error 2.47208000*10^(-7), which is better than 18 bits
3706 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3707 getF32Constant(DAG, 0x3924b03e));
3708 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3709 getF32Constant(DAG, 0x3ab24b87));
3710 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3711 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3712 getF32Constant(DAG, 0x3c1d8c17));
3713 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3714 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3715 getF32Constant(DAG, 0x3d634a1d));
3716 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3717 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3718 getF32Constant(DAG, 0x3e75fe14));
3719 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3720 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3721 getF32Constant(DAG, 0x3f317234));
3722 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3723 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3724 getF32Constant(DAG, 0x3f800000));
3725 SDValue t14 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t13);
3726 SDValue TwoToFractionalPartOfX =
3727 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
3729 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3730 MVT::f32, TwoToFractionalPartOfX);
3733 // No special expansion.
3734 result = DAG.getNode(ISD::FEXP2, dl,
3735 getValue(I.getArgOperand(0)).getValueType(),
3736 getValue(I.getArgOperand(0)));
3739 setValue(&I, result);
3742 /// visitPow - Lower a pow intrinsic. Handles the special sequences for
3743 /// limited-precision mode with x == 10.0f.
3745 SelectionDAGBuilder::visitPow(const CallInst &I) {
3747 const Value *Val = I.getArgOperand(0);
3748 DebugLoc dl = getCurDebugLoc();
3749 bool IsExp10 = false;
3751 if (getValue(Val).getValueType() == MVT::f32 &&
3752 getValue(I.getArgOperand(1)).getValueType() == MVT::f32 &&
3753 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3754 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(Val))) {
3755 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
3757 IsExp10 = CFP->getValueAPF().bitwiseIsEqual(Ten);
3762 if (IsExp10 && LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3763 SDValue Op = getValue(I.getArgOperand(1));
3765 // Put the exponent in the right bit position for later addition to the
3768 // #define LOG2OF10 3.3219281f
3769 // IntegerPartOfX = (int32_t)(x * LOG2OF10);
3770 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3771 getF32Constant(DAG, 0x40549a78));
3772 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3774 // FractionalPartOfX = x - (float)IntegerPartOfX;
3775 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3776 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3778 // IntegerPartOfX <<= 23;
3779 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3780 DAG.getConstant(23, TLI.getPointerTy()));
3782 if (LimitFloatPrecision <= 6) {
3783 // For floating-point precision of 6:
3785 // twoToFractionalPartOfX =
3787 // (0.735607626f + 0.252464424f * x) * x;
3789 // error 0.0144103317, which is 6 bits
3790 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3791 getF32Constant(DAG, 0x3e814304));
3792 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3793 getF32Constant(DAG, 0x3f3c50c8));
3794 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3795 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3796 getF32Constant(DAG, 0x3f7f5e7e));
3797 SDValue t6 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t5);
3798 SDValue TwoToFractionalPartOfX =
3799 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
3801 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3802 MVT::f32, TwoToFractionalPartOfX);
3803 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3804 // For floating-point precision of 12:
3806 // TwoToFractionalPartOfX =
3809 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3811 // error 0.000107046256, which is 13 to 14 bits
3812 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3813 getF32Constant(DAG, 0x3da235e3));
3814 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3815 getF32Constant(DAG, 0x3e65b8f3));
3816 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3817 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3818 getF32Constant(DAG, 0x3f324b07));
3819 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3820 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3821 getF32Constant(DAG, 0x3f7ff8fd));
3822 SDValue t8 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t7);
3823 SDValue TwoToFractionalPartOfX =
3824 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
3826 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3827 MVT::f32, TwoToFractionalPartOfX);
3828 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3829 // For floating-point precision of 18:
3831 // TwoToFractionalPartOfX =
3835 // (0.554906021e-1f +
3836 // (0.961591928e-2f +
3837 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3838 // error 2.47208000*10^(-7), which is better than 18 bits
3839 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3840 getF32Constant(DAG, 0x3924b03e));
3841 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3842 getF32Constant(DAG, 0x3ab24b87));
3843 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3844 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3845 getF32Constant(DAG, 0x3c1d8c17));
3846 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3847 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3848 getF32Constant(DAG, 0x3d634a1d));
3849 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3850 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3851 getF32Constant(DAG, 0x3e75fe14));
3852 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3853 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3854 getF32Constant(DAG, 0x3f317234));
3855 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3856 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3857 getF32Constant(DAG, 0x3f800000));
3858 SDValue t14 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t13);
3859 SDValue TwoToFractionalPartOfX =
3860 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
3862 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3863 MVT::f32, TwoToFractionalPartOfX);
3866 // No special expansion.
3867 result = DAG.getNode(ISD::FPOW, dl,
3868 getValue(I.getArgOperand(0)).getValueType(),
3869 getValue(I.getArgOperand(0)),
3870 getValue(I.getArgOperand(1)));
3873 setValue(&I, result);
3877 /// ExpandPowI - Expand a llvm.powi intrinsic.
3878 static SDValue ExpandPowI(DebugLoc DL, SDValue LHS, SDValue RHS,
3879 SelectionDAG &DAG) {
3880 // If RHS is a constant, we can expand this out to a multiplication tree,
3881 // otherwise we end up lowering to a call to __powidf2 (for example). When
3882 // optimizing for size, we only want to do this if the expansion would produce
3883 // a small number of multiplies, otherwise we do the full expansion.
3884 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3885 // Get the exponent as a positive value.
3886 unsigned Val = RHSC->getSExtValue();
3887 if ((int)Val < 0) Val = -Val;
3889 // powi(x, 0) -> 1.0
3891 return DAG.getConstantFP(1.0, LHS.getValueType());
3893 const Function *F = DAG.getMachineFunction().getFunction();
3894 if (!F->hasFnAttr(Attribute::OptimizeForSize) ||
3895 // If optimizing for size, don't insert too many multiplies. This
3896 // inserts up to 5 multiplies.
3897 CountPopulation_32(Val)+Log2_32(Val) < 7) {
3898 // We use the simple binary decomposition method to generate the multiply
3899 // sequence. There are more optimal ways to do this (for example,
3900 // powi(x,15) generates one more multiply than it should), but this has
3901 // the benefit of being both really simple and much better than a libcall.
3902 SDValue Res; // Logically starts equal to 1.0
3903 SDValue CurSquare = LHS;
3907 Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare);
3909 Res = CurSquare; // 1.0*CurSquare.
3912 CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(),
3913 CurSquare, CurSquare);
3917 // If the original was negative, invert the result, producing 1/(x*x*x).
3918 if (RHSC->getSExtValue() < 0)
3919 Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(),
3920 DAG.getConstantFP(1.0, LHS.getValueType()), Res);
3925 // Otherwise, expand to a libcall.
3926 return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS);
3929 /// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a function
3930 /// argument, create the corresponding DBG_VALUE machine instruction for it now.
3931 /// At the end of instruction selection, they will be inserted to the entry BB.
3933 SelectionDAGBuilder::EmitFuncArgumentDbgValue(const Value *V, MDNode *Variable,
3936 if (!isa<Argument>(V))
3939 MachineFunction &MF = DAG.getMachineFunction();
3940 // Ignore inlined function arguments here.
3941 DIVariable DV(Variable);
3942 if (DV.isInlinedFnArgument(MF.getFunction()))
3945 MachineBasicBlock *MBB = FuncInfo.MBB;
3946 if (MBB != &MF.front())
3950 if (N.getNode() && N.getOpcode() == ISD::CopyFromReg) {
3951 Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg();
3952 if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) {
3953 MachineRegisterInfo &RegInfo = MF.getRegInfo();
3954 unsigned PR = RegInfo.getLiveInPhysReg(Reg);
3961 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
3962 if (VMI == FuncInfo.ValueMap.end())
3967 const TargetInstrInfo *TII = DAG.getTarget().getInstrInfo();
3968 MachineInstrBuilder MIB = BuildMI(MF, getCurDebugLoc(),
3969 TII->get(TargetOpcode::DBG_VALUE))
3970 .addReg(Reg, RegState::Debug).addImm(Offset).addMetadata(Variable);
3971 FuncInfo.ArgDbgValues.push_back(&*MIB);
3975 // VisualStudio defines setjmp as _setjmp
3976 #if defined(_MSC_VER) && defined(setjmp)
3977 #define setjmp_undefined_for_visual_studio
3981 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
3982 /// we want to emit this as a call to a named external function, return the name
3983 /// otherwise lower it and return null.
3985 SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) {
3986 DebugLoc dl = getCurDebugLoc();
3989 switch (Intrinsic) {
3991 // By default, turn this into a target intrinsic node.
3992 visitTargetIntrinsic(I, Intrinsic);
3994 case Intrinsic::vastart: visitVAStart(I); return 0;
3995 case Intrinsic::vaend: visitVAEnd(I); return 0;
3996 case Intrinsic::vacopy: visitVACopy(I); return 0;
3997 case Intrinsic::returnaddress:
3998 setValue(&I, DAG.getNode(ISD::RETURNADDR, dl, TLI.getPointerTy(),
3999 getValue(I.getArgOperand(0))));
4001 case Intrinsic::frameaddress:
4002 setValue(&I, DAG.getNode(ISD::FRAMEADDR, dl, TLI.getPointerTy(),
4003 getValue(I.getArgOperand(0))));
4005 case Intrinsic::setjmp:
4006 return "_setjmp"+!TLI.usesUnderscoreSetJmp();
4007 case Intrinsic::longjmp:
4008 return "_longjmp"+!TLI.usesUnderscoreLongJmp();
4009 case Intrinsic::memcpy: {
4010 // Assert for address < 256 since we support only user defined address
4012 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4014 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4016 "Unknown address space");
4017 SDValue Op1 = getValue(I.getArgOperand(0));
4018 SDValue Op2 = getValue(I.getArgOperand(1));
4019 SDValue Op3 = getValue(I.getArgOperand(2));
4020 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4021 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4022 DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, isVol, false,
4023 I.getArgOperand(0), 0, I.getArgOperand(1), 0));
4026 case Intrinsic::memset: {
4027 // Assert for address < 256 since we support only user defined address
4029 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4031 "Unknown address space");
4032 SDValue Op1 = getValue(I.getArgOperand(0));
4033 SDValue Op2 = getValue(I.getArgOperand(1));
4034 SDValue Op3 = getValue(I.getArgOperand(2));
4035 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4036 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4037 DAG.setRoot(DAG.getMemset(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4038 I.getArgOperand(0), 0));
4041 case Intrinsic::memmove: {
4042 // Assert for address < 256 since we support only user defined address
4044 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4046 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4048 "Unknown address space");
4049 SDValue Op1 = getValue(I.getArgOperand(0));
4050 SDValue Op2 = getValue(I.getArgOperand(1));
4051 SDValue Op3 = getValue(I.getArgOperand(2));
4052 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4053 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4055 // If the source and destination are known to not be aliases, we can
4056 // lower memmove as memcpy.
4057 uint64_t Size = -1ULL;
4058 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op3))
4059 Size = C->getZExtValue();
4060 if (AA->alias(I.getArgOperand(0), Size, I.getArgOperand(1), Size) ==
4061 AliasAnalysis::NoAlias) {
4062 DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4063 false, I.getArgOperand(0), 0,
4064 I.getArgOperand(1), 0));
4068 DAG.setRoot(DAG.getMemmove(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4069 I.getArgOperand(0), 0, I.getArgOperand(1), 0));
4072 case Intrinsic::dbg_declare: {
4073 const DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
4074 if (!DIVariable(DI.getVariable()).Verify())
4077 MDNode *Variable = DI.getVariable();
4078 // Parameters are handled specially.
4080 DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable;
4081 const Value *Address = DI.getAddress();
4084 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
4085 Address = BCI->getOperand(0);
4086 const AllocaInst *AI = dyn_cast<AllocaInst>(Address);
4088 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder
4089 // but do not always have a corresponding SDNode built. The SDNodeOrder
4090 // absolute, but not relative, values are different depending on whether
4091 // debug info exists.
4093 SDValue &N = NodeMap[Address];
4096 if (isParameter && !AI) {
4097 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode());
4099 // Byval parameter. We have a frame index at this point.
4100 SDV = DAG.getDbgValue(Variable, FINode->getIndex(),
4101 0, dl, SDNodeOrder);
4103 // Can't do anything with other non-AI cases yet. This might be a
4104 // parameter of a callee function that got inlined, for example.
4107 SDV = DAG.getDbgValue(Variable, N.getNode(), N.getResNo(),
4108 0, dl, SDNodeOrder);
4110 // Can't do anything with other non-AI cases yet.
4112 DAG.AddDbgValue(SDV, N.getNode(), isParameter);
4114 // If Address is an arugment then try to emits its dbg value using
4115 // virtual register info from the FuncInfo.ValueMap. Otherwise add undef
4116 // to help track missing debug info.
4117 if (!EmitFuncArgumentDbgValue(Address, Variable, 0, N)) {
4118 SDV = DAG.getDbgValue(Variable, UndefValue::get(Address->getType()),
4119 0, dl, SDNodeOrder);
4120 DAG.AddDbgValue(SDV, 0, isParameter);
4125 case Intrinsic::dbg_value: {
4126 const DbgValueInst &DI = cast<DbgValueInst>(I);
4127 if (!DIVariable(DI.getVariable()).Verify())
4130 MDNode *Variable = DI.getVariable();
4131 uint64_t Offset = DI.getOffset();
4132 const Value *V = DI.getValue();
4136 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder
4137 // but do not always have a corresponding SDNode built. The SDNodeOrder
4138 // absolute, but not relative, values are different depending on whether
4139 // debug info exists.
4142 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) {
4143 SDV = DAG.getDbgValue(Variable, V, Offset, dl, SDNodeOrder);
4144 DAG.AddDbgValue(SDV, 0, false);
4146 bool createUndef = false;
4147 // Do not use getValue() in here; we don't want to generate code at
4148 // this point if it hasn't been done yet.
4149 SDValue N = NodeMap[V];
4150 if (!N.getNode() && isa<Argument>(V))
4151 // Check unused arguments map.
4152 N = UnusedArgNodeMap[V];
4154 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, N)) {
4155 SDV = DAG.getDbgValue(Variable, N.getNode(),
4156 N.getResNo(), Offset, dl, SDNodeOrder);
4157 DAG.AddDbgValue(SDV, N.getNode(), false);
4159 } else if (isa<PHINode>(V) && !V->use_empty() ) {
4160 // Do not call getValue(V) yet, as we don't want to generate code.
4161 // Remember it for later.
4162 DanglingDebugInfo DDI(&DI, dl, SDNodeOrder);
4163 DanglingDebugInfoMap[V] = DDI;
4167 // We may expand this to cover more cases. One case where we have no
4168 // data available is an unreferenced parameter; we need this fallback.
4169 SDV = DAG.getDbgValue(Variable, UndefValue::get(V->getType()),
4170 Offset, dl, SDNodeOrder);
4171 DAG.AddDbgValue(SDV, 0, false);
4175 // Build a debug info table entry.
4176 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
4177 V = BCI->getOperand(0);
4178 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
4179 // Don't handle byval struct arguments or VLAs, for example.
4182 DenseMap<const AllocaInst*, int>::iterator SI =
4183 FuncInfo.StaticAllocaMap.find(AI);
4184 if (SI == FuncInfo.StaticAllocaMap.end())
4186 int FI = SI->second;
4188 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4189 if (!DI.getDebugLoc().isUnknown() && MMI.hasDebugInfo())
4190 MMI.setVariableDbgInfo(Variable, FI, DI.getDebugLoc());
4193 case Intrinsic::eh_exception: {
4194 // Insert the EXCEPTIONADDR instruction.
4195 assert(FuncInfo.MBB->isLandingPad() &&
4196 "Call to eh.exception not in landing pad!");
4197 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
4199 Ops[0] = DAG.getRoot();
4200 SDValue Op = DAG.getNode(ISD::EXCEPTIONADDR, dl, VTs, Ops, 1);
4202 DAG.setRoot(Op.getValue(1));
4206 case Intrinsic::eh_selector: {
4207 MachineBasicBlock *CallMBB = FuncInfo.MBB;
4208 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4209 if (CallMBB->isLandingPad())
4210 AddCatchInfo(I, &MMI, CallMBB);
4213 FuncInfo.CatchInfoLost.insert(&I);
4215 // FIXME: Mark exception selector register as live in. Hack for PR1508.
4216 unsigned Reg = TLI.getExceptionSelectorRegister();
4217 if (Reg) FuncInfo.MBB->addLiveIn(Reg);
4220 // Insert the EHSELECTION instruction.
4221 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
4223 Ops[0] = getValue(I.getArgOperand(0));
4225 SDValue Op = DAG.getNode(ISD::EHSELECTION, dl, VTs, Ops, 2);
4226 DAG.setRoot(Op.getValue(1));
4227 setValue(&I, DAG.getSExtOrTrunc(Op, dl, MVT::i32));
4231 case Intrinsic::eh_typeid_for: {
4232 // Find the type id for the given typeinfo.
4233 GlobalVariable *GV = ExtractTypeInfo(I.getArgOperand(0));
4234 unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(GV);
4235 Res = DAG.getConstant(TypeID, MVT::i32);
4240 case Intrinsic::eh_return_i32:
4241 case Intrinsic::eh_return_i64:
4242 DAG.getMachineFunction().getMMI().setCallsEHReturn(true);
4243 DAG.setRoot(DAG.getNode(ISD::EH_RETURN, dl,
4246 getValue(I.getArgOperand(0)),
4247 getValue(I.getArgOperand(1))));
4249 case Intrinsic::eh_unwind_init:
4250 DAG.getMachineFunction().getMMI().setCallsUnwindInit(true);
4252 case Intrinsic::eh_dwarf_cfa: {
4253 SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getArgOperand(0)), dl,
4254 TLI.getPointerTy());
4255 SDValue Offset = DAG.getNode(ISD::ADD, dl,
4257 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, dl,
4258 TLI.getPointerTy()),
4260 SDValue FA = DAG.getNode(ISD::FRAMEADDR, dl,
4262 DAG.getConstant(0, TLI.getPointerTy()));
4263 setValue(&I, DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
4267 case Intrinsic::eh_sjlj_callsite: {
4268 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4269 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(0));
4270 assert(CI && "Non-constant call site value in eh.sjlj.callsite!");
4271 assert(MMI.getCurrentCallSite() == 0 && "Overlapping call sites!");
4273 MMI.setCurrentCallSite(CI->getZExtValue());
4276 case Intrinsic::eh_sjlj_setjmp: {
4277 setValue(&I, DAG.getNode(ISD::EH_SJLJ_SETJMP, dl, MVT::i32, getRoot(),
4278 getValue(I.getArgOperand(0))));
4281 case Intrinsic::eh_sjlj_longjmp: {
4282 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, dl, MVT::Other,
4284 getValue(I.getArgOperand(0))));
4288 case Intrinsic::convertff:
4289 case Intrinsic::convertfsi:
4290 case Intrinsic::convertfui:
4291 case Intrinsic::convertsif:
4292 case Intrinsic::convertuif:
4293 case Intrinsic::convertss:
4294 case Intrinsic::convertsu:
4295 case Intrinsic::convertus:
4296 case Intrinsic::convertuu: {
4297 ISD::CvtCode Code = ISD::CVT_INVALID;
4298 switch (Intrinsic) {
4299 case Intrinsic::convertff: Code = ISD::CVT_FF; break;
4300 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break;
4301 case Intrinsic::convertfui: Code = ISD::CVT_FU; break;
4302 case Intrinsic::convertsif: Code = ISD::CVT_SF; break;
4303 case Intrinsic::convertuif: Code = ISD::CVT_UF; break;
4304 case Intrinsic::convertss: Code = ISD::CVT_SS; break;
4305 case Intrinsic::convertsu: Code = ISD::CVT_SU; break;
4306 case Intrinsic::convertus: Code = ISD::CVT_US; break;
4307 case Intrinsic::convertuu: Code = ISD::CVT_UU; break;
4309 EVT DestVT = TLI.getValueType(I.getType());
4310 const Value *Op1 = I.getArgOperand(0);
4311 Res = DAG.getConvertRndSat(DestVT, getCurDebugLoc(), getValue(Op1),
4312 DAG.getValueType(DestVT),
4313 DAG.getValueType(getValue(Op1).getValueType()),
4314 getValue(I.getArgOperand(1)),
4315 getValue(I.getArgOperand(2)),
4320 case Intrinsic::sqrt:
4321 setValue(&I, DAG.getNode(ISD::FSQRT, dl,
4322 getValue(I.getArgOperand(0)).getValueType(),
4323 getValue(I.getArgOperand(0))));
4325 case Intrinsic::powi:
4326 setValue(&I, ExpandPowI(dl, getValue(I.getArgOperand(0)),
4327 getValue(I.getArgOperand(1)), DAG));
4329 case Intrinsic::sin:
4330 setValue(&I, DAG.getNode(ISD::FSIN, dl,
4331 getValue(I.getArgOperand(0)).getValueType(),
4332 getValue(I.getArgOperand(0))));
4334 case Intrinsic::cos:
4335 setValue(&I, DAG.getNode(ISD::FCOS, dl,
4336 getValue(I.getArgOperand(0)).getValueType(),
4337 getValue(I.getArgOperand(0))));
4339 case Intrinsic::log:
4342 case Intrinsic::log2:
4345 case Intrinsic::log10:
4348 case Intrinsic::exp:
4351 case Intrinsic::exp2:
4354 case Intrinsic::pow:
4357 case Intrinsic::convert_to_fp16:
4358 setValue(&I, DAG.getNode(ISD::FP32_TO_FP16, dl,
4359 MVT::i16, getValue(I.getArgOperand(0))));
4361 case Intrinsic::convert_from_fp16:
4362 setValue(&I, DAG.getNode(ISD::FP16_TO_FP32, dl,
4363 MVT::f32, getValue(I.getArgOperand(0))));
4365 case Intrinsic::pcmarker: {
4366 SDValue Tmp = getValue(I.getArgOperand(0));
4367 DAG.setRoot(DAG.getNode(ISD::PCMARKER, dl, MVT::Other, getRoot(), Tmp));
4370 case Intrinsic::readcyclecounter: {
4371 SDValue Op = getRoot();
4372 Res = DAG.getNode(ISD::READCYCLECOUNTER, dl,
4373 DAG.getVTList(MVT::i64, MVT::Other),
4376 DAG.setRoot(Res.getValue(1));
4379 case Intrinsic::bswap:
4380 setValue(&I, DAG.getNode(ISD::BSWAP, dl,
4381 getValue(I.getArgOperand(0)).getValueType(),
4382 getValue(I.getArgOperand(0))));
4384 case Intrinsic::cttz: {
4385 SDValue Arg = getValue(I.getArgOperand(0));
4386 EVT Ty = Arg.getValueType();
4387 setValue(&I, DAG.getNode(ISD::CTTZ, dl, Ty, Arg));
4390 case Intrinsic::ctlz: {
4391 SDValue Arg = getValue(I.getArgOperand(0));
4392 EVT Ty = Arg.getValueType();
4393 setValue(&I, DAG.getNode(ISD::CTLZ, dl, Ty, Arg));
4396 case Intrinsic::ctpop: {
4397 SDValue Arg = getValue(I.getArgOperand(0));
4398 EVT Ty = Arg.getValueType();
4399 setValue(&I, DAG.getNode(ISD::CTPOP, dl, Ty, Arg));
4402 case Intrinsic::stacksave: {
4403 SDValue Op = getRoot();
4404 Res = DAG.getNode(ISD::STACKSAVE, dl,
4405 DAG.getVTList(TLI.getPointerTy(), MVT::Other), &Op, 1);
4407 DAG.setRoot(Res.getValue(1));
4410 case Intrinsic::stackrestore: {
4411 Res = getValue(I.getArgOperand(0));
4412 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, dl, MVT::Other, getRoot(), Res));
4415 case Intrinsic::stackprotector: {
4416 // Emit code into the DAG to store the stack guard onto the stack.
4417 MachineFunction &MF = DAG.getMachineFunction();
4418 MachineFrameInfo *MFI = MF.getFrameInfo();
4419 EVT PtrTy = TLI.getPointerTy();
4421 SDValue Src = getValue(I.getArgOperand(0)); // The guard's value.
4422 AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
4424 int FI = FuncInfo.StaticAllocaMap[Slot];
4425 MFI->setStackProtectorIndex(FI);
4427 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
4429 // Store the stack protector onto the stack.
4430 Res = DAG.getStore(getRoot(), getCurDebugLoc(), Src, FIN,
4431 PseudoSourceValue::getFixedStack(FI),
4437 case Intrinsic::objectsize: {
4438 // If we don't know by now, we're never going to know.
4439 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1));
4441 assert(CI && "Non-constant type in __builtin_object_size?");
4443 SDValue Arg = getValue(I.getCalledValue());
4444 EVT Ty = Arg.getValueType();
4447 Res = DAG.getConstant(-1ULL, Ty);
4449 Res = DAG.getConstant(0, Ty);
4454 case Intrinsic::var_annotation:
4455 // Discard annotate attributes
4458 case Intrinsic::init_trampoline: {
4459 const Function *F = cast<Function>(I.getArgOperand(1)->stripPointerCasts());
4463 Ops[1] = getValue(I.getArgOperand(0));
4464 Ops[2] = getValue(I.getArgOperand(1));
4465 Ops[3] = getValue(I.getArgOperand(2));
4466 Ops[4] = DAG.getSrcValue(I.getArgOperand(0));
4467 Ops[5] = DAG.getSrcValue(F);
4469 Res = DAG.getNode(ISD::TRAMPOLINE, dl,
4470 DAG.getVTList(TLI.getPointerTy(), MVT::Other),
4474 DAG.setRoot(Res.getValue(1));
4477 case Intrinsic::gcroot:
4479 const Value *Alloca = I.getArgOperand(0);
4480 const Constant *TypeMap = cast<Constant>(I.getArgOperand(1));
4482 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode());
4483 GFI->addStackRoot(FI->getIndex(), TypeMap);
4486 case Intrinsic::gcread:
4487 case Intrinsic::gcwrite:
4488 llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!");
4490 case Intrinsic::flt_rounds:
4491 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, dl, MVT::i32));
4493 case Intrinsic::trap:
4494 DAG.setRoot(DAG.getNode(ISD::TRAP, dl,MVT::Other, getRoot()));
4496 case Intrinsic::uadd_with_overflow:
4497 return implVisitAluOverflow(I, ISD::UADDO);
4498 case Intrinsic::sadd_with_overflow:
4499 return implVisitAluOverflow(I, ISD::SADDO);
4500 case Intrinsic::usub_with_overflow:
4501 return implVisitAluOverflow(I, ISD::USUBO);
4502 case Intrinsic::ssub_with_overflow:
4503 return implVisitAluOverflow(I, ISD::SSUBO);
4504 case Intrinsic::umul_with_overflow:
4505 return implVisitAluOverflow(I, ISD::UMULO);
4506 case Intrinsic::smul_with_overflow:
4507 return implVisitAluOverflow(I, ISD::SMULO);
4509 case Intrinsic::prefetch: {
4512 Ops[1] = getValue(I.getArgOperand(0));
4513 Ops[2] = getValue(I.getArgOperand(1));
4514 Ops[3] = getValue(I.getArgOperand(2));
4515 DAG.setRoot(DAG.getNode(ISD::PREFETCH, dl, MVT::Other, &Ops[0], 4));
4519 case Intrinsic::memory_barrier: {
4522 for (int x = 1; x < 6; ++x)
4523 Ops[x] = getValue(I.getArgOperand(x - 1));
4525 DAG.setRoot(DAG.getNode(ISD::MEMBARRIER, dl, MVT::Other, &Ops[0], 6));
4528 case Intrinsic::atomic_cmp_swap: {
4529 SDValue Root = getRoot();
4531 DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, getCurDebugLoc(),
4532 getValue(I.getArgOperand(1)).getValueType().getSimpleVT(),
4534 getValue(I.getArgOperand(0)),
4535 getValue(I.getArgOperand(1)),
4536 getValue(I.getArgOperand(2)),
4537 I.getArgOperand(0));
4539 DAG.setRoot(L.getValue(1));
4542 case Intrinsic::atomic_load_add:
4543 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_ADD);
4544 case Intrinsic::atomic_load_sub:
4545 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_SUB);
4546 case Intrinsic::atomic_load_or:
4547 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_OR);
4548 case Intrinsic::atomic_load_xor:
4549 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_XOR);
4550 case Intrinsic::atomic_load_and:
4551 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_AND);
4552 case Intrinsic::atomic_load_nand:
4553 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_NAND);
4554 case Intrinsic::atomic_load_max:
4555 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MAX);
4556 case Intrinsic::atomic_load_min:
4557 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MIN);
4558 case Intrinsic::atomic_load_umin:
4559 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMIN);
4560 case Intrinsic::atomic_load_umax:
4561 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMAX);
4562 case Intrinsic::atomic_swap:
4563 return implVisitBinaryAtomic(I, ISD::ATOMIC_SWAP);
4565 case Intrinsic::invariant_start:
4566 case Intrinsic::lifetime_start:
4567 // Discard region information.
4568 setValue(&I, DAG.getUNDEF(TLI.getPointerTy()));
4570 case Intrinsic::invariant_end:
4571 case Intrinsic::lifetime_end:
4572 // Discard region information.
4577 void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee,
4579 MachineBasicBlock *LandingPad) {
4580 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
4581 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
4582 const Type *RetTy = FTy->getReturnType();
4583 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4584 MCSymbol *BeginLabel = 0;
4586 TargetLowering::ArgListTy Args;
4587 TargetLowering::ArgListEntry Entry;
4588 Args.reserve(CS.arg_size());
4590 // Check whether the function can return without sret-demotion.
4591 SmallVector<ISD::OutputArg, 4> Outs;
4592 SmallVector<uint64_t, 4> Offsets;
4593 GetReturnInfo(RetTy, CS.getAttributes().getRetAttributes(),
4594 Outs, TLI, &Offsets);
4596 bool CanLowerReturn = TLI.CanLowerReturn(CS.getCallingConv(),
4597 FTy->isVarArg(), Outs, FTy->getContext());
4599 SDValue DemoteStackSlot;
4601 if (!CanLowerReturn) {
4602 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(
4603 FTy->getReturnType());
4604 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(
4605 FTy->getReturnType());
4606 MachineFunction &MF = DAG.getMachineFunction();
4607 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
4608 const Type *StackSlotPtrType = PointerType::getUnqual(FTy->getReturnType());
4610 DemoteStackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
4611 Entry.Node = DemoteStackSlot;
4612 Entry.Ty = StackSlotPtrType;
4613 Entry.isSExt = false;
4614 Entry.isZExt = false;
4615 Entry.isInReg = false;
4616 Entry.isSRet = true;
4617 Entry.isNest = false;
4618 Entry.isByVal = false;
4619 Entry.Alignment = Align;
4620 Args.push_back(Entry);
4621 RetTy = Type::getVoidTy(FTy->getContext());
4624 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
4626 SDValue ArgNode = getValue(*i);
4627 Entry.Node = ArgNode; Entry.Ty = (*i)->getType();
4629 unsigned attrInd = i - CS.arg_begin() + 1;
4630 Entry.isSExt = CS.paramHasAttr(attrInd, Attribute::SExt);
4631 Entry.isZExt = CS.paramHasAttr(attrInd, Attribute::ZExt);
4632 Entry.isInReg = CS.paramHasAttr(attrInd, Attribute::InReg);
4633 Entry.isSRet = CS.paramHasAttr(attrInd, Attribute::StructRet);
4634 Entry.isNest = CS.paramHasAttr(attrInd, Attribute::Nest);
4635 Entry.isByVal = CS.paramHasAttr(attrInd, Attribute::ByVal);
4636 Entry.Alignment = CS.getParamAlignment(attrInd);
4637 Args.push_back(Entry);
4641 // Insert a label before the invoke call to mark the try range. This can be
4642 // used to detect deletion of the invoke via the MachineModuleInfo.
4643 BeginLabel = MMI.getContext().CreateTempSymbol();
4645 // For SjLj, keep track of which landing pads go with which invokes
4646 // so as to maintain the ordering of pads in the LSDA.
4647 unsigned CallSiteIndex = MMI.getCurrentCallSite();
4648 if (CallSiteIndex) {
4649 MMI.setCallSiteBeginLabel(BeginLabel, CallSiteIndex);
4650 // Now that the call site is handled, stop tracking it.
4651 MMI.setCurrentCallSite(0);
4654 // Both PendingLoads and PendingExports must be flushed here;
4655 // this call might not return.
4657 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getControlRoot(), BeginLabel));
4660 // Check if target-independent constraints permit a tail call here.
4661 // Target-dependent constraints are checked within TLI.LowerCallTo.
4663 !isInTailCallPosition(CS, CS.getAttributes().getRetAttributes(), TLI))
4666 std::pair<SDValue,SDValue> Result =
4667 TLI.LowerCallTo(getRoot(), RetTy,
4668 CS.paramHasAttr(0, Attribute::SExt),
4669 CS.paramHasAttr(0, Attribute::ZExt), FTy->isVarArg(),
4670 CS.paramHasAttr(0, Attribute::InReg), FTy->getNumParams(),
4671 CS.getCallingConv(),
4673 !CS.getInstruction()->use_empty(),
4674 Callee, Args, DAG, getCurDebugLoc());
4675 assert((isTailCall || Result.second.getNode()) &&
4676 "Non-null chain expected with non-tail call!");
4677 assert((Result.second.getNode() || !Result.first.getNode()) &&
4678 "Null value expected with tail call!");
4679 if (Result.first.getNode()) {
4680 setValue(CS.getInstruction(), Result.first);
4681 } else if (!CanLowerReturn && Result.second.getNode()) {
4682 // The instruction result is the result of loading from the
4683 // hidden sret parameter.
4684 SmallVector<EVT, 1> PVTs;
4685 const Type *PtrRetTy = PointerType::getUnqual(FTy->getReturnType());
4687 ComputeValueVTs(TLI, PtrRetTy, PVTs);
4688 assert(PVTs.size() == 1 && "Pointers should fit in one register");
4689 EVT PtrVT = PVTs[0];
4690 unsigned NumValues = Outs.size();
4691 SmallVector<SDValue, 4> Values(NumValues);
4692 SmallVector<SDValue, 4> Chains(NumValues);
4694 for (unsigned i = 0; i < NumValues; ++i) {
4695 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT,
4697 DAG.getConstant(Offsets[i], PtrVT));
4698 SDValue L = DAG.getLoad(Outs[i].VT, getCurDebugLoc(), Result.second,
4699 Add, NULL, Offsets[i], false, false, 1);
4701 Chains[i] = L.getValue(1);
4704 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
4705 MVT::Other, &Chains[0], NumValues);
4706 PendingLoads.push_back(Chain);
4708 // Collect the legal value parts into potentially illegal values
4709 // that correspond to the original function's return values.
4710 SmallVector<EVT, 4> RetTys;
4711 RetTy = FTy->getReturnType();
4712 ComputeValueVTs(TLI, RetTy, RetTys);
4713 ISD::NodeType AssertOp = ISD::DELETED_NODE;
4714 SmallVector<SDValue, 4> ReturnValues;
4715 unsigned CurReg = 0;
4716 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
4718 EVT RegisterVT = TLI.getRegisterType(RetTy->getContext(), VT);
4719 unsigned NumRegs = TLI.getNumRegisters(RetTy->getContext(), VT);
4721 SDValue ReturnValue =
4722 getCopyFromParts(DAG, getCurDebugLoc(), &Values[CurReg], NumRegs,
4723 RegisterVT, VT, AssertOp);
4724 ReturnValues.push_back(ReturnValue);
4728 setValue(CS.getInstruction(),
4729 DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
4730 DAG.getVTList(&RetTys[0], RetTys.size()),
4731 &ReturnValues[0], ReturnValues.size()));
4735 // As a special case, a null chain means that a tail call has been emitted and
4736 // the DAG root is already updated.
4737 if (Result.second.getNode())
4738 DAG.setRoot(Result.second);
4743 // Insert a label at the end of the invoke call to mark the try range. This
4744 // can be used to detect deletion of the invoke via the MachineModuleInfo.
4745 MCSymbol *EndLabel = MMI.getContext().CreateTempSymbol();
4746 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getRoot(), EndLabel));
4748 // Inform MachineModuleInfo of range.
4749 MMI.addInvoke(LandingPad, BeginLabel, EndLabel);
4753 /// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the
4754 /// value is equal or not-equal to zero.
4755 static bool IsOnlyUsedInZeroEqualityComparison(const Value *V) {
4756 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end();
4758 if (const ICmpInst *IC = dyn_cast<ICmpInst>(*UI))
4759 if (IC->isEquality())
4760 if (const Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
4761 if (C->isNullValue())
4763 // Unknown instruction.
4769 static SDValue getMemCmpLoad(const Value *PtrVal, MVT LoadVT,
4771 SelectionDAGBuilder &Builder) {
4773 // Check to see if this load can be trivially constant folded, e.g. if the
4774 // input is from a string literal.
4775 if (const Constant *LoadInput = dyn_cast<Constant>(PtrVal)) {
4776 // Cast pointer to the type we really want to load.
4777 LoadInput = ConstantExpr::getBitCast(const_cast<Constant *>(LoadInput),
4778 PointerType::getUnqual(LoadTy));
4780 if (const Constant *LoadCst =
4781 ConstantFoldLoadFromConstPtr(const_cast<Constant *>(LoadInput),
4783 return Builder.getValue(LoadCst);
4786 // Otherwise, we have to emit the load. If the pointer is to unfoldable but
4787 // still constant memory, the input chain can be the entry node.
4789 bool ConstantMemory = false;
4791 // Do not serialize (non-volatile) loads of constant memory with anything.
4792 if (Builder.AA->pointsToConstantMemory(PtrVal)) {
4793 Root = Builder.DAG.getEntryNode();
4794 ConstantMemory = true;
4796 // Do not serialize non-volatile loads against each other.
4797 Root = Builder.DAG.getRoot();
4800 SDValue Ptr = Builder.getValue(PtrVal);
4801 SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurDebugLoc(), Root,
4802 Ptr, PtrVal /*SrcValue*/, 0/*SVOffset*/,
4804 false /*nontemporal*/, 1 /* align=1 */);
4806 if (!ConstantMemory)
4807 Builder.PendingLoads.push_back(LoadVal.getValue(1));
4812 /// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form.
4813 /// If so, return true and lower it, otherwise return false and it will be
4814 /// lowered like a normal call.
4815 bool SelectionDAGBuilder::visitMemCmpCall(const CallInst &I) {
4816 // Verify that the prototype makes sense. int memcmp(void*,void*,size_t)
4817 if (I.getNumArgOperands() != 3)
4820 const Value *LHS = I.getArgOperand(0), *RHS = I.getArgOperand(1);
4821 if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() ||
4822 !I.getArgOperand(2)->getType()->isIntegerTy() ||
4823 !I.getType()->isIntegerTy())
4826 const ConstantInt *Size = dyn_cast<ConstantInt>(I.getArgOperand(2));
4828 // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0
4829 // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0
4830 if (Size && IsOnlyUsedInZeroEqualityComparison(&I)) {
4831 bool ActuallyDoIt = true;
4834 switch (Size->getZExtValue()) {
4836 LoadVT = MVT::Other;
4838 ActuallyDoIt = false;
4842 LoadTy = Type::getInt16Ty(Size->getContext());
4846 LoadTy = Type::getInt32Ty(Size->getContext());
4850 LoadTy = Type::getInt64Ty(Size->getContext());
4854 LoadVT = MVT::v4i32;
4855 LoadTy = Type::getInt32Ty(Size->getContext());
4856 LoadTy = VectorType::get(LoadTy, 4);
4861 // This turns into unaligned loads. We only do this if the target natively
4862 // supports the MVT we'll be loading or if it is small enough (<= 4) that
4863 // we'll only produce a small number of byte loads.
4865 // Require that we can find a legal MVT, and only do this if the target
4866 // supports unaligned loads of that type. Expanding into byte loads would
4868 if (ActuallyDoIt && Size->getZExtValue() > 4) {
4869 // TODO: Handle 5 byte compare as 4-byte + 1 byte.
4870 // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads.
4871 if (!TLI.isTypeLegal(LoadVT) ||!TLI.allowsUnalignedMemoryAccesses(LoadVT))
4872 ActuallyDoIt = false;
4876 SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this);
4877 SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this);
4879 SDValue Res = DAG.getSetCC(getCurDebugLoc(), MVT::i1, LHSVal, RHSVal,
4881 EVT CallVT = TLI.getValueType(I.getType(), true);
4882 setValue(&I, DAG.getZExtOrTrunc(Res, getCurDebugLoc(), CallVT));
4892 void SelectionDAGBuilder::visitCall(const CallInst &I) {
4893 // Handle inline assembly differently.
4894 if (isa<InlineAsm>(I.getCalledValue())) {
4899 const char *RenameFn = 0;
4900 if (Function *F = I.getCalledFunction()) {
4901 if (F->isDeclaration()) {
4902 if (const TargetIntrinsicInfo *II = TM.getIntrinsicInfo()) {
4903 if (unsigned IID = II->getIntrinsicID(F)) {
4904 RenameFn = visitIntrinsicCall(I, IID);
4909 if (unsigned IID = F->getIntrinsicID()) {
4910 RenameFn = visitIntrinsicCall(I, IID);
4916 // Check for well-known libc/libm calls. If the function is internal, it
4917 // can't be a library call.
4918 if (!F->hasLocalLinkage() && F->hasName()) {
4919 StringRef Name = F->getName();
4920 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl") {
4921 if (I.getNumArgOperands() == 2 && // Basic sanity checks.
4922 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
4923 I.getType() == I.getArgOperand(0)->getType() &&
4924 I.getType() == I.getArgOperand(1)->getType()) {
4925 SDValue LHS = getValue(I.getArgOperand(0));
4926 SDValue RHS = getValue(I.getArgOperand(1));
4927 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurDebugLoc(),
4928 LHS.getValueType(), LHS, RHS));
4931 } else if (Name == "fabs" || Name == "fabsf" || Name == "fabsl") {
4932 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
4933 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
4934 I.getType() == I.getArgOperand(0)->getType()) {
4935 SDValue Tmp = getValue(I.getArgOperand(0));
4936 setValue(&I, DAG.getNode(ISD::FABS, getCurDebugLoc(),
4937 Tmp.getValueType(), Tmp));
4940 } else if (Name == "sin" || Name == "sinf" || Name == "sinl") {
4941 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
4942 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
4943 I.getType() == I.getArgOperand(0)->getType() &&
4944 I.onlyReadsMemory()) {
4945 SDValue Tmp = getValue(I.getArgOperand(0));
4946 setValue(&I, DAG.getNode(ISD::FSIN, getCurDebugLoc(),
4947 Tmp.getValueType(), Tmp));
4950 } else if (Name == "cos" || Name == "cosf" || Name == "cosl") {
4951 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
4952 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
4953 I.getType() == I.getArgOperand(0)->getType() &&
4954 I.onlyReadsMemory()) {
4955 SDValue Tmp = getValue(I.getArgOperand(0));
4956 setValue(&I, DAG.getNode(ISD::FCOS, getCurDebugLoc(),
4957 Tmp.getValueType(), Tmp));
4960 } else if (Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl") {
4961 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
4962 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
4963 I.getType() == I.getArgOperand(0)->getType() &&
4964 I.onlyReadsMemory()) {
4965 SDValue Tmp = getValue(I.getArgOperand(0));
4966 setValue(&I, DAG.getNode(ISD::FSQRT, getCurDebugLoc(),
4967 Tmp.getValueType(), Tmp));
4970 } else if (Name == "memcmp") {
4971 if (visitMemCmpCall(I))
4979 Callee = getValue(I.getCalledValue());
4981 Callee = DAG.getExternalSymbol(RenameFn, TLI.getPointerTy());
4983 // Check if we can potentially perform a tail call. More detailed checking is
4984 // be done within LowerCallTo, after more information about the call is known.
4985 LowerCallTo(&I, Callee, I.isTailCall());
4990 /// AsmOperandInfo - This contains information for each constraint that we are
4992 class LLVM_LIBRARY_VISIBILITY SDISelAsmOperandInfo :
4993 public TargetLowering::AsmOperandInfo {
4995 /// CallOperand - If this is the result output operand or a clobber
4996 /// this is null, otherwise it is the incoming operand to the CallInst.
4997 /// This gets modified as the asm is processed.
4998 SDValue CallOperand;
5000 /// AssignedRegs - If this is a register or register class operand, this
5001 /// contains the set of register corresponding to the operand.
5002 RegsForValue AssignedRegs;
5004 explicit SDISelAsmOperandInfo(const InlineAsm::ConstraintInfo &info)
5005 : TargetLowering::AsmOperandInfo(info), CallOperand(0,0) {
5008 /// MarkAllocatedRegs - Once AssignedRegs is set, mark the assigned registers
5009 /// busy in OutputRegs/InputRegs.
5010 void MarkAllocatedRegs(bool isOutReg, bool isInReg,
5011 std::set<unsigned> &OutputRegs,
5012 std::set<unsigned> &InputRegs,
5013 const TargetRegisterInfo &TRI) const {
5015 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
5016 MarkRegAndAliases(AssignedRegs.Regs[i], OutputRegs, TRI);
5019 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
5020 MarkRegAndAliases(AssignedRegs.Regs[i], InputRegs, TRI);
5024 /// getCallOperandValEVT - Return the EVT of the Value* that this operand
5025 /// corresponds to. If there is no Value* for this operand, it returns
5027 EVT getCallOperandValEVT(LLVMContext &Context,
5028 const TargetLowering &TLI,
5029 const TargetData *TD) const {
5030 if (CallOperandVal == 0) return MVT::Other;
5032 if (isa<BasicBlock>(CallOperandVal))
5033 return TLI.getPointerTy();
5035 const llvm::Type *OpTy = CallOperandVal->getType();
5037 // If this is an indirect operand, the operand is a pointer to the
5040 const llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
5042 report_fatal_error("Indirect operand for inline asm not a pointer!");
5043 OpTy = PtrTy->getElementType();
5046 // If OpTy is not a single value, it may be a struct/union that we
5047 // can tile with integers.
5048 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
5049 unsigned BitSize = TD->getTypeSizeInBits(OpTy);
5058 OpTy = IntegerType::get(Context, BitSize);
5063 return TLI.getValueType(OpTy, true);
5067 /// MarkRegAndAliases - Mark the specified register and all aliases in the
5069 static void MarkRegAndAliases(unsigned Reg, std::set<unsigned> &Regs,
5070 const TargetRegisterInfo &TRI) {
5071 assert(TargetRegisterInfo::isPhysicalRegister(Reg) && "Isn't a physreg");
5073 if (const unsigned *Aliases = TRI.getAliasSet(Reg))
5074 for (; *Aliases; ++Aliases)
5075 Regs.insert(*Aliases);
5079 } // end llvm namespace.
5081 /// isAllocatableRegister - If the specified register is safe to allocate,
5082 /// i.e. it isn't a stack pointer or some other special register, return the
5083 /// register class for the register. Otherwise, return null.
5084 static const TargetRegisterClass *
5085 isAllocatableRegister(unsigned Reg, MachineFunction &MF,
5086 const TargetLowering &TLI,
5087 const TargetRegisterInfo *TRI) {
5088 EVT FoundVT = MVT::Other;
5089 const TargetRegisterClass *FoundRC = 0;
5090 for (TargetRegisterInfo::regclass_iterator RCI = TRI->regclass_begin(),
5091 E = TRI->regclass_end(); RCI != E; ++RCI) {
5092 EVT ThisVT = MVT::Other;
5094 const TargetRegisterClass *RC = *RCI;
5095 // If none of the value types for this register class are valid, we
5096 // can't use it. For example, 64-bit reg classes on 32-bit targets.
5097 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
5099 if (TLI.isTypeLegal(*I)) {
5100 // If we have already found this register in a different register class,
5101 // choose the one with the largest VT specified. For example, on
5102 // PowerPC, we favor f64 register classes over f32.
5103 if (FoundVT == MVT::Other || FoundVT.bitsLT(*I)) {
5110 if (ThisVT == MVT::Other) continue;
5112 // NOTE: This isn't ideal. In particular, this might allocate the
5113 // frame pointer in functions that need it (due to them not being taken
5114 // out of allocation, because a variable sized allocation hasn't been seen
5115 // yet). This is a slight code pessimization, but should still work.
5116 for (TargetRegisterClass::iterator I = RC->allocation_order_begin(MF),
5117 E = RC->allocation_order_end(MF); I != E; ++I)
5119 // We found a matching register class. Keep looking at others in case
5120 // we find one with larger registers that this physreg is also in.
5129 /// GetRegistersForValue - Assign registers (virtual or physical) for the
5130 /// specified operand. We prefer to assign virtual registers, to allow the
5131 /// register allocator to handle the assignment process. However, if the asm
5132 /// uses features that we can't model on machineinstrs, we have SDISel do the
5133 /// allocation. This produces generally horrible, but correct, code.
5135 /// OpInfo describes the operand.
5136 /// Input and OutputRegs are the set of already allocated physical registers.
5138 void SelectionDAGBuilder::
5139 GetRegistersForValue(SDISelAsmOperandInfo &OpInfo,
5140 std::set<unsigned> &OutputRegs,
5141 std::set<unsigned> &InputRegs) {
5142 LLVMContext &Context = FuncInfo.Fn->getContext();
5144 // Compute whether this value requires an input register, an output register,
5146 bool isOutReg = false;
5147 bool isInReg = false;
5148 switch (OpInfo.Type) {
5149 case InlineAsm::isOutput:
5152 // If there is an input constraint that matches this, we need to reserve
5153 // the input register so no other inputs allocate to it.
5154 isInReg = OpInfo.hasMatchingInput();
5156 case InlineAsm::isInput:
5160 case InlineAsm::isClobber:
5167 MachineFunction &MF = DAG.getMachineFunction();
5168 SmallVector<unsigned, 4> Regs;
5170 // If this is a constraint for a single physreg, or a constraint for a
5171 // register class, find it.
5172 std::pair<unsigned, const TargetRegisterClass*> PhysReg =
5173 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
5174 OpInfo.ConstraintVT);
5176 unsigned NumRegs = 1;
5177 if (OpInfo.ConstraintVT != MVT::Other) {
5178 // If this is a FP input in an integer register (or visa versa) insert a bit
5179 // cast of the input value. More generally, handle any case where the input
5180 // value disagrees with the register class we plan to stick this in.
5181 if (OpInfo.Type == InlineAsm::isInput &&
5182 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) {
5183 // Try to convert to the first EVT that the reg class contains. If the
5184 // types are identical size, use a bitcast to convert (e.g. two differing
5186 EVT RegVT = *PhysReg.second->vt_begin();
5187 if (RegVT.getSizeInBits() == OpInfo.ConstraintVT.getSizeInBits()) {
5188 OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
5189 RegVT, OpInfo.CallOperand);
5190 OpInfo.ConstraintVT = RegVT;
5191 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) {
5192 // If the input is a FP value and we want it in FP registers, do a
5193 // bitcast to the corresponding integer type. This turns an f64 value
5194 // into i64, which can be passed with two i32 values on a 32-bit
5196 RegVT = EVT::getIntegerVT(Context,
5197 OpInfo.ConstraintVT.getSizeInBits());
5198 OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
5199 RegVT, OpInfo.CallOperand);
5200 OpInfo.ConstraintVT = RegVT;
5204 NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT);
5208 EVT ValueVT = OpInfo.ConstraintVT;
5210 // If this is a constraint for a specific physical register, like {r17},
5212 if (unsigned AssignedReg = PhysReg.first) {
5213 const TargetRegisterClass *RC = PhysReg.second;
5214 if (OpInfo.ConstraintVT == MVT::Other)
5215 ValueVT = *RC->vt_begin();
5217 // Get the actual register value type. This is important, because the user
5218 // may have asked for (e.g.) the AX register in i32 type. We need to
5219 // remember that AX is actually i16 to get the right extension.
5220 RegVT = *RC->vt_begin();
5222 // This is a explicit reference to a physical register.
5223 Regs.push_back(AssignedReg);
5225 // If this is an expanded reference, add the rest of the regs to Regs.
5227 TargetRegisterClass::iterator I = RC->begin();
5228 for (; *I != AssignedReg; ++I)
5229 assert(I != RC->end() && "Didn't find reg!");
5231 // Already added the first reg.
5233 for (; NumRegs; --NumRegs, ++I) {
5234 assert(I != RC->end() && "Ran out of registers to allocate!");
5239 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5240 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
5241 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
5245 // Otherwise, if this was a reference to an LLVM register class, create vregs
5246 // for this reference.
5247 if (const TargetRegisterClass *RC = PhysReg.second) {
5248 RegVT = *RC->vt_begin();
5249 if (OpInfo.ConstraintVT == MVT::Other)
5252 // Create the appropriate number of virtual registers.
5253 MachineRegisterInfo &RegInfo = MF.getRegInfo();
5254 for (; NumRegs; --NumRegs)
5255 Regs.push_back(RegInfo.createVirtualRegister(RC));
5257 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5261 // This is a reference to a register class that doesn't directly correspond
5262 // to an LLVM register class. Allocate NumRegs consecutive, available,
5263 // registers from the class.
5264 std::vector<unsigned> RegClassRegs
5265 = TLI.getRegClassForInlineAsmConstraint(OpInfo.ConstraintCode,
5266 OpInfo.ConstraintVT);
5268 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
5269 unsigned NumAllocated = 0;
5270 for (unsigned i = 0, e = RegClassRegs.size(); i != e; ++i) {
5271 unsigned Reg = RegClassRegs[i];
5272 // See if this register is available.
5273 if ((isOutReg && OutputRegs.count(Reg)) || // Already used.
5274 (isInReg && InputRegs.count(Reg))) { // Already used.
5275 // Make sure we find consecutive registers.
5280 // Check to see if this register is allocatable (i.e. don't give out the
5282 const TargetRegisterClass *RC = isAllocatableRegister(Reg, MF, TLI, TRI);
5283 if (!RC) { // Couldn't allocate this register.
5284 // Reset NumAllocated to make sure we return consecutive registers.
5289 // Okay, this register is good, we can use it.
5292 // If we allocated enough consecutive registers, succeed.
5293 if (NumAllocated == NumRegs) {
5294 unsigned RegStart = (i-NumAllocated)+1;
5295 unsigned RegEnd = i+1;
5296 // Mark all of the allocated registers used.
5297 for (unsigned i = RegStart; i != RegEnd; ++i)
5298 Regs.push_back(RegClassRegs[i]);
5300 OpInfo.AssignedRegs = RegsForValue(Regs, *RC->vt_begin(),
5301 OpInfo.ConstraintVT);
5302 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
5307 // Otherwise, we couldn't allocate enough registers for this.
5310 /// visitInlineAsm - Handle a call to an InlineAsm object.
5312 void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) {
5313 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
5315 /// ConstraintOperands - Information about all of the constraints.
5316 std::vector<SDISelAsmOperandInfo> ConstraintOperands;
5318 std::set<unsigned> OutputRegs, InputRegs;
5320 // Do a prepass over the constraints, canonicalizing them, and building up the
5321 // ConstraintOperands list.
5322 std::vector<InlineAsm::ConstraintInfo>
5323 ConstraintInfos = IA->ParseConstraints();
5325 bool hasMemory = hasInlineAsmMemConstraint(ConstraintInfos, TLI);
5327 SDValue Chain, Flag;
5329 // We won't need to flush pending loads if this asm doesn't touch
5330 // memory and is nonvolatile.
5331 if (hasMemory || IA->hasSideEffects())
5334 Chain = DAG.getRoot();
5336 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
5337 unsigned ResNo = 0; // ResNo - The result number of the next output.
5338 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
5339 ConstraintOperands.push_back(SDISelAsmOperandInfo(ConstraintInfos[i]));
5340 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
5342 EVT OpVT = MVT::Other;
5344 // Compute the value type for each operand.
5345 switch (OpInfo.Type) {
5346 case InlineAsm::isOutput:
5347 // Indirect outputs just consume an argument.
5348 if (OpInfo.isIndirect) {
5349 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5353 // The return value of the call is this value. As such, there is no
5354 // corresponding argument.
5355 assert(!CS.getType()->isVoidTy() &&
5357 if (const StructType *STy = dyn_cast<StructType>(CS.getType())) {
5358 OpVT = TLI.getValueType(STy->getElementType(ResNo));
5360 assert(ResNo == 0 && "Asm only has one result!");
5361 OpVT = TLI.getValueType(CS.getType());
5365 case InlineAsm::isInput:
5366 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5368 case InlineAsm::isClobber:
5373 // If this is an input or an indirect output, process the call argument.
5374 // BasicBlocks are labels, currently appearing only in asm's.
5375 if (OpInfo.CallOperandVal) {
5376 // Strip bitcasts, if any. This mostly comes up for functions.
5377 OpInfo.CallOperandVal = OpInfo.CallOperandVal->stripPointerCasts();
5379 if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) {
5380 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]);
5382 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
5385 OpVT = OpInfo.getCallOperandValEVT(*DAG.getContext(), TLI, TD);
5388 OpInfo.ConstraintVT = OpVT;
5391 // Second pass over the constraints: compute which constraint option to use
5392 // and assign registers to constraints that want a specific physreg.
5393 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
5394 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5396 // If this is an output operand with a matching input operand, look up the
5397 // matching input. If their types mismatch, e.g. one is an integer, the
5398 // other is floating point, or their sizes are different, flag it as an
5400 if (OpInfo.hasMatchingInput()) {
5401 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
5403 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
5404 if ((OpInfo.ConstraintVT.isInteger() !=
5405 Input.ConstraintVT.isInteger()) ||
5406 (OpInfo.ConstraintVT.getSizeInBits() !=
5407 Input.ConstraintVT.getSizeInBits())) {
5408 report_fatal_error("Unsupported asm: input constraint"
5409 " with a matching output constraint of"
5410 " incompatible type!");
5412 Input.ConstraintVT = OpInfo.ConstraintVT;
5416 // Compute the constraint code and ConstraintType to use.
5417 TLI.ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, &DAG);
5419 // If this is a memory input, and if the operand is not indirect, do what we
5420 // need to to provide an address for the memory input.
5421 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5422 !OpInfo.isIndirect) {
5423 assert(OpInfo.Type == InlineAsm::isInput &&
5424 "Can only indirectify direct input operands!");
5426 // Memory operands really want the address of the value. If we don't have
5427 // an indirect input, put it in the constpool if we can, otherwise spill
5428 // it to a stack slot.
5430 // If the operand is a float, integer, or vector constant, spill to a
5431 // constant pool entry to get its address.
5432 const Value *OpVal = OpInfo.CallOperandVal;
5433 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
5434 isa<ConstantVector>(OpVal)) {
5435 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal),
5436 TLI.getPointerTy());
5438 // Otherwise, create a stack slot and emit a store to it before the
5440 const Type *Ty = OpVal->getType();
5441 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
5442 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(Ty);
5443 MachineFunction &MF = DAG.getMachineFunction();
5444 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
5445 SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
5446 Chain = DAG.getStore(Chain, getCurDebugLoc(),
5447 OpInfo.CallOperand, StackSlot, NULL, 0,
5449 OpInfo.CallOperand = StackSlot;
5452 // There is no longer a Value* corresponding to this operand.
5453 OpInfo.CallOperandVal = 0;
5455 // It is now an indirect operand.
5456 OpInfo.isIndirect = true;
5459 // If this constraint is for a specific register, allocate it before
5461 if (OpInfo.ConstraintType == TargetLowering::C_Register)
5462 GetRegistersForValue(OpInfo, OutputRegs, InputRegs);
5465 ConstraintInfos.clear();
5467 // Second pass - Loop over all of the operands, assigning virtual or physregs
5468 // to register class operands.
5469 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5470 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5472 // C_Register operands have already been allocated, Other/Memory don't need
5474 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
5475 GetRegistersForValue(OpInfo, OutputRegs, InputRegs);
5478 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
5479 std::vector<SDValue> AsmNodeOperands;
5480 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain
5481 AsmNodeOperands.push_back(
5482 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(),
5483 TLI.getPointerTy()));
5485 // If we have a !srcloc metadata node associated with it, we want to attach
5486 // this to the ultimately generated inline asm machineinstr. To do this, we
5487 // pass in the third operand as this (potentially null) inline asm MDNode.
5488 const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc");
5489 AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc));
5491 // Remember the AlignStack bit as operand 3.
5492 AsmNodeOperands.push_back(DAG.getTargetConstant(IA->isAlignStack() ? 1 : 0,
5495 // Loop over all of the inputs, copying the operand values into the
5496 // appropriate registers and processing the output regs.
5497 RegsForValue RetValRegs;
5499 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
5500 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
5502 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5503 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5505 switch (OpInfo.Type) {
5506 case InlineAsm::isOutput: {
5507 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
5508 OpInfo.ConstraintType != TargetLowering::C_Register) {
5509 // Memory output, or 'other' output (e.g. 'X' constraint).
5510 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
5512 // Add information to the INLINEASM node to know about this output.
5513 unsigned OpFlags = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
5514 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlags,
5515 TLI.getPointerTy()));
5516 AsmNodeOperands.push_back(OpInfo.CallOperand);
5520 // Otherwise, this is a register or register class output.
5522 // Copy the output from the appropriate register. Find a register that
5524 if (OpInfo.AssignedRegs.Regs.empty())
5525 report_fatal_error("Couldn't allocate output reg for constraint '" +
5526 Twine(OpInfo.ConstraintCode) + "'!");
5528 // If this is an indirect operand, store through the pointer after the
5530 if (OpInfo.isIndirect) {
5531 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
5532 OpInfo.CallOperandVal));
5534 // This is the result value of the call.
5535 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
5536 // Concatenate this output onto the outputs list.
5537 RetValRegs.append(OpInfo.AssignedRegs);
5540 // Add information to the INLINEASM node to know that this register is
5542 OpInfo.AssignedRegs.AddInlineAsmOperands(OpInfo.isEarlyClobber ?
5543 InlineAsm::Kind_RegDefEarlyClobber :
5544 InlineAsm::Kind_RegDef,
5551 case InlineAsm::isInput: {
5552 SDValue InOperandVal = OpInfo.CallOperand;
5554 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint?
5555 // If this is required to match an output register we have already set,
5556 // just use its register.
5557 unsigned OperandNo = OpInfo.getMatchedOperand();
5559 // Scan until we find the definition we already emitted of this operand.
5560 // When we find it, create a RegsForValue operand.
5561 unsigned CurOp = InlineAsm::Op_FirstOperand;
5562 for (; OperandNo; --OperandNo) {
5563 // Advance to the next operand.
5565 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
5566 assert((InlineAsm::isRegDefKind(OpFlag) ||
5567 InlineAsm::isRegDefEarlyClobberKind(OpFlag) ||
5568 InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?");
5569 CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1;
5573 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
5574 if (InlineAsm::isRegDefKind(OpFlag) ||
5575 InlineAsm::isRegDefEarlyClobberKind(OpFlag)) {
5576 // Add (OpFlag&0xffff)>>3 registers to MatchedRegs.
5577 if (OpInfo.isIndirect) {
5578 // This happens on gcc/testsuite/gcc.dg/pr8788-1.c
5579 LLVMContext &Ctx = *DAG.getContext();
5580 Ctx.emitError(CS.getInstruction(), "inline asm not supported yet:"
5581 " don't know how to handle tied "
5582 "indirect register inputs");
5585 RegsForValue MatchedRegs;
5586 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
5587 EVT RegVT = AsmNodeOperands[CurOp+1].getValueType();
5588 MatchedRegs.RegVTs.push_back(RegVT);
5589 MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo();
5590 for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag);
5592 MatchedRegs.Regs.push_back
5593 (RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT)));
5595 // Use the produced MatchedRegs object to
5596 MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
5598 MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse,
5599 true, OpInfo.getMatchedOperand(),
5600 DAG, AsmNodeOperands);
5604 assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!");
5605 assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 &&
5606 "Unexpected number of operands");
5607 // Add information to the INLINEASM node to know about this input.
5608 // See InlineAsm.h isUseOperandTiedToDef.
5609 OpFlag = InlineAsm::getFlagWordForMatchingOp(OpFlag,
5610 OpInfo.getMatchedOperand());
5611 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlag,
5612 TLI.getPointerTy()));
5613 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]);
5617 // Treat indirect 'X' constraint as memory.
5618 if (OpInfo.ConstraintType == TargetLowering::C_Other &&
5620 OpInfo.ConstraintType = TargetLowering::C_Memory;
5622 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
5623 std::vector<SDValue> Ops;
5624 TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode[0],
5627 report_fatal_error("Invalid operand for inline asm constraint '" +
5628 Twine(OpInfo.ConstraintCode) + "'!");
5630 // Add information to the INLINEASM node to know about this input.
5631 unsigned ResOpType =
5632 InlineAsm::getFlagWord(InlineAsm::Kind_Imm, Ops.size());
5633 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
5634 TLI.getPointerTy()));
5635 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
5639 if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
5640 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
5641 assert(InOperandVal.getValueType() == TLI.getPointerTy() &&
5642 "Memory operands expect pointer values");
5644 // Add information to the INLINEASM node to know about this input.
5645 unsigned ResOpType = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
5646 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
5647 TLI.getPointerTy()));
5648 AsmNodeOperands.push_back(InOperandVal);
5652 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
5653 OpInfo.ConstraintType == TargetLowering::C_Register) &&
5654 "Unknown constraint type!");
5655 assert(!OpInfo.isIndirect &&
5656 "Don't know how to handle indirect register inputs yet!");
5658 // Copy the input into the appropriate registers.
5659 if (OpInfo.AssignedRegs.Regs.empty() ||
5660 !OpInfo.AssignedRegs.areValueTypesLegal(TLI))
5661 report_fatal_error("Couldn't allocate input reg for constraint '" +
5662 Twine(OpInfo.ConstraintCode) + "'!");
5664 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
5667 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, false, 0,
5668 DAG, AsmNodeOperands);
5671 case InlineAsm::isClobber: {
5672 // Add the clobbered value to the operand list, so that the register
5673 // allocator is aware that the physreg got clobbered.
5674 if (!OpInfo.AssignedRegs.Regs.empty())
5675 OpInfo.AssignedRegs.AddInlineAsmOperands(
5676 InlineAsm::Kind_RegDefEarlyClobber,
5684 // Finish up input operands. Set the input chain and add the flag last.
5685 AsmNodeOperands[InlineAsm::Op_InputChain] = Chain;
5686 if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
5688 Chain = DAG.getNode(ISD::INLINEASM, getCurDebugLoc(),
5689 DAG.getVTList(MVT::Other, MVT::Flag),
5690 &AsmNodeOperands[0], AsmNodeOperands.size());
5691 Flag = Chain.getValue(1);
5693 // If this asm returns a register value, copy the result from that register
5694 // and set it as the value of the call.
5695 if (!RetValRegs.Regs.empty()) {
5696 SDValue Val = RetValRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(),
5699 // FIXME: Why don't we do this for inline asms with MRVs?
5700 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) {
5701 EVT ResultType = TLI.getValueType(CS.getType());
5703 // If any of the results of the inline asm is a vector, it may have the
5704 // wrong width/num elts. This can happen for register classes that can
5705 // contain multiple different value types. The preg or vreg allocated may
5706 // not have the same VT as was expected. Convert it to the right type
5707 // with bit_convert.
5708 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) {
5709 Val = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
5712 } else if (ResultType != Val.getValueType() &&
5713 ResultType.isInteger() && Val.getValueType().isInteger()) {
5714 // If a result value was tied to an input value, the computed result may
5715 // have a wider width than the expected result. Extract the relevant
5717 Val = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), ResultType, Val);
5720 assert(ResultType == Val.getValueType() && "Asm result value mismatch!");
5723 setValue(CS.getInstruction(), Val);
5724 // Don't need to use this as a chain in this case.
5725 if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty())
5729 std::vector<std::pair<SDValue, const Value *> > StoresToEmit;
5731 // Process indirect outputs, first output all of the flagged copies out of
5733 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
5734 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
5735 const Value *Ptr = IndirectStoresToEmit[i].second;
5736 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(),
5738 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
5741 // Emit the non-flagged stores from the physregs.
5742 SmallVector<SDValue, 8> OutChains;
5743 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) {
5744 SDValue Val = DAG.getStore(Chain, getCurDebugLoc(),
5745 StoresToEmit[i].first,
5746 getValue(StoresToEmit[i].second),
5747 StoresToEmit[i].second, 0,
5749 OutChains.push_back(Val);
5752 if (!OutChains.empty())
5753 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
5754 &OutChains[0], OutChains.size());
5759 void SelectionDAGBuilder::visitVAStart(const CallInst &I) {
5760 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurDebugLoc(),
5761 MVT::Other, getRoot(),
5762 getValue(I.getArgOperand(0)),
5763 DAG.getSrcValue(I.getArgOperand(0))));
5766 void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) {
5767 const TargetData &TD = *TLI.getTargetData();
5768 SDValue V = DAG.getVAArg(TLI.getValueType(I.getType()), getCurDebugLoc(),
5769 getRoot(), getValue(I.getOperand(0)),
5770 DAG.getSrcValue(I.getOperand(0)),
5771 TD.getABITypeAlignment(I.getType()));
5773 DAG.setRoot(V.getValue(1));
5776 void SelectionDAGBuilder::visitVAEnd(const CallInst &I) {
5777 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurDebugLoc(),
5778 MVT::Other, getRoot(),
5779 getValue(I.getArgOperand(0)),
5780 DAG.getSrcValue(I.getArgOperand(0))));
5783 void SelectionDAGBuilder::visitVACopy(const CallInst &I) {
5784 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurDebugLoc(),
5785 MVT::Other, getRoot(),
5786 getValue(I.getArgOperand(0)),
5787 getValue(I.getArgOperand(1)),
5788 DAG.getSrcValue(I.getArgOperand(0)),
5789 DAG.getSrcValue(I.getArgOperand(1))));
5792 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
5793 /// implementation, which just calls LowerCall.
5794 /// FIXME: When all targets are
5795 /// migrated to using LowerCall, this hook should be integrated into SDISel.
5796 std::pair<SDValue, SDValue>
5797 TargetLowering::LowerCallTo(SDValue Chain, const Type *RetTy,
5798 bool RetSExt, bool RetZExt, bool isVarArg,
5799 bool isInreg, unsigned NumFixedArgs,
5800 CallingConv::ID CallConv, bool isTailCall,
5801 bool isReturnValueUsed,
5803 ArgListTy &Args, SelectionDAG &DAG,
5804 DebugLoc dl) const {
5805 // Handle all of the outgoing arguments.
5806 SmallVector<ISD::OutputArg, 32> Outs;
5807 SmallVector<SDValue, 32> OutVals;
5808 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
5809 SmallVector<EVT, 4> ValueVTs;
5810 ComputeValueVTs(*this, Args[i].Ty, ValueVTs);
5811 for (unsigned Value = 0, NumValues = ValueVTs.size();
5812 Value != NumValues; ++Value) {
5813 EVT VT = ValueVTs[Value];
5814 const Type *ArgTy = VT.getTypeForEVT(RetTy->getContext());
5815 SDValue Op = SDValue(Args[i].Node.getNode(),
5816 Args[i].Node.getResNo() + Value);
5817 ISD::ArgFlagsTy Flags;
5818 unsigned OriginalAlignment =
5819 getTargetData()->getABITypeAlignment(ArgTy);
5825 if (Args[i].isInReg)
5829 if (Args[i].isByVal) {
5831 const PointerType *Ty = cast<PointerType>(Args[i].Ty);
5832 const Type *ElementTy = Ty->getElementType();
5833 unsigned FrameAlign = getByValTypeAlignment(ElementTy);
5834 unsigned FrameSize = getTargetData()->getTypeAllocSize(ElementTy);
5835 // For ByVal, alignment should come from FE. BE will guess if this
5836 // info is not there but there are cases it cannot get right.
5837 if (Args[i].Alignment)
5838 FrameAlign = Args[i].Alignment;
5839 Flags.setByValAlign(FrameAlign);
5840 Flags.setByValSize(FrameSize);
5844 Flags.setOrigAlign(OriginalAlignment);
5846 EVT PartVT = getRegisterType(RetTy->getContext(), VT);
5847 unsigned NumParts = getNumRegisters(RetTy->getContext(), VT);
5848 SmallVector<SDValue, 4> Parts(NumParts);
5849 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
5852 ExtendKind = ISD::SIGN_EXTEND;
5853 else if (Args[i].isZExt)
5854 ExtendKind = ISD::ZERO_EXTEND;
5856 getCopyToParts(DAG, dl, Op, &Parts[0], NumParts,
5857 PartVT, ExtendKind);
5859 for (unsigned j = 0; j != NumParts; ++j) {
5860 // if it isn't first piece, alignment must be 1
5861 ISD::OutputArg MyFlags(Flags, Parts[j].getValueType(),
5863 if (NumParts > 1 && j == 0)
5864 MyFlags.Flags.setSplit();
5866 MyFlags.Flags.setOrigAlign(1);
5868 Outs.push_back(MyFlags);
5869 OutVals.push_back(Parts[j]);
5874 // Handle the incoming return values from the call.
5875 SmallVector<ISD::InputArg, 32> Ins;
5876 SmallVector<EVT, 4> RetTys;
5877 ComputeValueVTs(*this, RetTy, RetTys);
5878 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
5880 EVT RegisterVT = getRegisterType(RetTy->getContext(), VT);
5881 unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT);
5882 for (unsigned i = 0; i != NumRegs; ++i) {
5883 ISD::InputArg MyFlags;
5884 MyFlags.VT = RegisterVT;
5885 MyFlags.Used = isReturnValueUsed;
5887 MyFlags.Flags.setSExt();
5889 MyFlags.Flags.setZExt();
5891 MyFlags.Flags.setInReg();
5892 Ins.push_back(MyFlags);
5896 SmallVector<SDValue, 4> InVals;
5897 Chain = LowerCall(Chain, Callee, CallConv, isVarArg, isTailCall,
5898 Outs, OutVals, Ins, dl, DAG, InVals);
5900 // Verify that the target's LowerCall behaved as expected.
5901 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
5902 "LowerCall didn't return a valid chain!");
5903 assert((!isTailCall || InVals.empty()) &&
5904 "LowerCall emitted a return value for a tail call!");
5905 assert((isTailCall || InVals.size() == Ins.size()) &&
5906 "LowerCall didn't emit the correct number of values!");
5908 // For a tail call, the return value is merely live-out and there aren't
5909 // any nodes in the DAG representing it. Return a special value to
5910 // indicate that a tail call has been emitted and no more Instructions
5911 // should be processed in the current block.
5914 return std::make_pair(SDValue(), SDValue());
5917 DEBUG(for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
5918 assert(InVals[i].getNode() &&
5919 "LowerCall emitted a null value!");
5920 assert(Ins[i].VT == InVals[i].getValueType() &&
5921 "LowerCall emitted a value with the wrong type!");
5924 // Collect the legal value parts into potentially illegal values
5925 // that correspond to the original function's return values.
5926 ISD::NodeType AssertOp = ISD::DELETED_NODE;
5928 AssertOp = ISD::AssertSext;
5930 AssertOp = ISD::AssertZext;
5931 SmallVector<SDValue, 4> ReturnValues;
5932 unsigned CurReg = 0;
5933 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
5935 EVT RegisterVT = getRegisterType(RetTy->getContext(), VT);
5936 unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT);
5938 ReturnValues.push_back(getCopyFromParts(DAG, dl, &InVals[CurReg],
5939 NumRegs, RegisterVT, VT,
5944 // For a function returning void, there is no return value. We can't create
5945 // such a node, so we just return a null return value in that case. In
5946 // that case, nothing will actualy look at the value.
5947 if (ReturnValues.empty())
5948 return std::make_pair(SDValue(), Chain);
5950 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, dl,
5951 DAG.getVTList(&RetTys[0], RetTys.size()),
5952 &ReturnValues[0], ReturnValues.size());
5953 return std::make_pair(Res, Chain);
5956 void TargetLowering::LowerOperationWrapper(SDNode *N,
5957 SmallVectorImpl<SDValue> &Results,
5958 SelectionDAG &DAG) const {
5959 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
5961 Results.push_back(Res);
5964 SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5965 llvm_unreachable("LowerOperation not implemented for this target!");
5970 SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) {
5971 SDValue Op = getNonRegisterValue(V);
5972 assert((Op.getOpcode() != ISD::CopyFromReg ||
5973 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
5974 "Copy from a reg to the same reg!");
5975 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg");
5977 RegsForValue RFV(V->getContext(), TLI, Reg, V->getType());
5978 SDValue Chain = DAG.getEntryNode();
5979 RFV.getCopyToRegs(Op, DAG, getCurDebugLoc(), Chain, 0);
5980 PendingExports.push_back(Chain);
5983 #include "llvm/CodeGen/SelectionDAGISel.h"
5985 void SelectionDAGISel::LowerArguments(const BasicBlock *LLVMBB) {
5986 // If this is the entry block, emit arguments.
5987 const Function &F = *LLVMBB->getParent();
5988 SelectionDAG &DAG = SDB->DAG;
5989 DebugLoc dl = SDB->getCurDebugLoc();
5990 const TargetData *TD = TLI.getTargetData();
5991 SmallVector<ISD::InputArg, 16> Ins;
5993 // Check whether the function can return without sret-demotion.
5994 SmallVector<ISD::OutputArg, 4> Outs;
5995 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
5998 if (!FuncInfo->CanLowerReturn) {
5999 // Put in an sret pointer parameter before all the other parameters.
6000 SmallVector<EVT, 1> ValueVTs;
6001 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
6003 // NOTE: Assuming that a pointer will never break down to more than one VT
6005 ISD::ArgFlagsTy Flags;
6007 EVT RegisterVT = TLI.getRegisterType(*DAG.getContext(), ValueVTs[0]);
6008 ISD::InputArg RetArg(Flags, RegisterVT, true);
6009 Ins.push_back(RetArg);
6012 // Set up the incoming argument description vector.
6014 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end();
6015 I != E; ++I, ++Idx) {
6016 SmallVector<EVT, 4> ValueVTs;
6017 ComputeValueVTs(TLI, I->getType(), ValueVTs);
6018 bool isArgValueUsed = !I->use_empty();
6019 for (unsigned Value = 0, NumValues = ValueVTs.size();
6020 Value != NumValues; ++Value) {
6021 EVT VT = ValueVTs[Value];
6022 const Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
6023 ISD::ArgFlagsTy Flags;
6024 unsigned OriginalAlignment =
6025 TD->getABITypeAlignment(ArgTy);
6027 if (F.paramHasAttr(Idx, Attribute::ZExt))
6029 if (F.paramHasAttr(Idx, Attribute::SExt))
6031 if (F.paramHasAttr(Idx, Attribute::InReg))
6033 if (F.paramHasAttr(Idx, Attribute::StructRet))
6035 if (F.paramHasAttr(Idx, Attribute::ByVal)) {
6037 const PointerType *Ty = cast<PointerType>(I->getType());
6038 const Type *ElementTy = Ty->getElementType();
6039 unsigned FrameAlign = TLI.getByValTypeAlignment(ElementTy);
6040 unsigned FrameSize = TD->getTypeAllocSize(ElementTy);
6041 // For ByVal, alignment should be passed from FE. BE will guess if
6042 // this info is not there but there are cases it cannot get right.
6043 if (F.getParamAlignment(Idx))
6044 FrameAlign = F.getParamAlignment(Idx);
6045 Flags.setByValAlign(FrameAlign);
6046 Flags.setByValSize(FrameSize);
6048 if (F.paramHasAttr(Idx, Attribute::Nest))
6050 Flags.setOrigAlign(OriginalAlignment);
6052 EVT RegisterVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6053 unsigned NumRegs = TLI.getNumRegisters(*CurDAG->getContext(), VT);
6054 for (unsigned i = 0; i != NumRegs; ++i) {
6055 ISD::InputArg MyFlags(Flags, RegisterVT, isArgValueUsed);
6056 if (NumRegs > 1 && i == 0)
6057 MyFlags.Flags.setSplit();
6058 // if it isn't first piece, alignment must be 1
6060 MyFlags.Flags.setOrigAlign(1);
6061 Ins.push_back(MyFlags);
6066 // Call the target to set up the argument values.
6067 SmallVector<SDValue, 8> InVals;
6068 SDValue NewRoot = TLI.LowerFormalArguments(DAG.getRoot(), F.getCallingConv(),
6072 // Verify that the target's LowerFormalArguments behaved as expected.
6073 assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other &&
6074 "LowerFormalArguments didn't return a valid chain!");
6075 assert(InVals.size() == Ins.size() &&
6076 "LowerFormalArguments didn't emit the correct number of values!");
6078 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
6079 assert(InVals[i].getNode() &&
6080 "LowerFormalArguments emitted a null value!");
6081 assert(Ins[i].VT == InVals[i].getValueType() &&
6082 "LowerFormalArguments emitted a value with the wrong type!");
6086 // Update the DAG with the new chain value resulting from argument lowering.
6087 DAG.setRoot(NewRoot);
6089 // Set up the argument values.
6092 if (!FuncInfo->CanLowerReturn) {
6093 // Create a virtual register for the sret pointer, and put in a copy
6094 // from the sret argument into it.
6095 SmallVector<EVT, 1> ValueVTs;
6096 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
6097 EVT VT = ValueVTs[0];
6098 EVT RegVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6099 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6100 SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1,
6101 RegVT, VT, AssertOp);
6103 MachineFunction& MF = SDB->DAG.getMachineFunction();
6104 MachineRegisterInfo& RegInfo = MF.getRegInfo();
6105 unsigned SRetReg = RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT));
6106 FuncInfo->DemoteRegister = SRetReg;
6107 NewRoot = SDB->DAG.getCopyToReg(NewRoot, SDB->getCurDebugLoc(),
6109 DAG.setRoot(NewRoot);
6111 // i indexes lowered arguments. Bump it past the hidden sret argument.
6112 // Idx indexes LLVM arguments. Don't touch it.
6116 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
6118 SmallVector<SDValue, 4> ArgValues;
6119 SmallVector<EVT, 4> ValueVTs;
6120 ComputeValueVTs(TLI, I->getType(), ValueVTs);
6121 unsigned NumValues = ValueVTs.size();
6123 // If this argument is unused then remember its value. It is used to generate
6124 // debugging information.
6125 if (I->use_empty() && NumValues)
6126 SDB->setUnusedArgValue(I, InVals[i]);
6128 for (unsigned Value = 0; Value != NumValues; ++Value) {
6129 EVT VT = ValueVTs[Value];
6130 EVT PartVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6131 unsigned NumParts = TLI.getNumRegisters(*CurDAG->getContext(), VT);
6133 if (!I->use_empty()) {
6134 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6135 if (F.paramHasAttr(Idx, Attribute::SExt))
6136 AssertOp = ISD::AssertSext;
6137 else if (F.paramHasAttr(Idx, Attribute::ZExt))
6138 AssertOp = ISD::AssertZext;
6140 ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i],
6141 NumParts, PartVT, VT,
6148 if (!I->use_empty()) {
6150 if (!ArgValues.empty())
6151 Res = DAG.getMergeValues(&ArgValues[0], NumValues,
6152 SDB->getCurDebugLoc());
6153 SDB->setValue(I, Res);
6155 // If this argument is live outside of the entry block, insert a copy from
6156 // whereever we got it to the vreg that other BB's will reference it as.
6157 SDB->CopyToExportRegsIfNeeded(I);
6161 assert(i == InVals.size() && "Argument register count mismatch!");
6163 // Finally, if the target has anything special to do, allow it to do so.
6164 // FIXME: this should insert code into the DAG!
6165 EmitFunctionEntryCode();
6168 /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
6169 /// ensure constants are generated when needed. Remember the virtual registers
6170 /// that need to be added to the Machine PHI nodes as input. We cannot just
6171 /// directly add them, because expansion might result in multiple MBB's for one
6172 /// BB. As such, the start of the BB might correspond to a different MBB than
6176 SelectionDAGBuilder::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
6177 const TerminatorInst *TI = LLVMBB->getTerminator();
6179 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
6181 // Check successor nodes' PHI nodes that expect a constant to be available
6183 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
6184 const BasicBlock *SuccBB = TI->getSuccessor(succ);
6185 if (!isa<PHINode>(SuccBB->begin())) continue;
6186 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
6188 // If this terminator has multiple identical successors (common for
6189 // switches), only handle each succ once.
6190 if (!SuccsHandled.insert(SuccMBB)) continue;
6192 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
6194 // At this point we know that there is a 1-1 correspondence between LLVM PHI
6195 // nodes and Machine PHI nodes, but the incoming operands have not been
6197 for (BasicBlock::const_iterator I = SuccBB->begin();
6198 const PHINode *PN = dyn_cast<PHINode>(I); ++I) {
6199 // Ignore dead phi's.
6200 if (PN->use_empty()) continue;
6203 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
6205 if (const Constant *C = dyn_cast<Constant>(PHIOp)) {
6206 unsigned &RegOut = ConstantsOut[C];
6208 RegOut = FuncInfo.CreateRegs(C->getType());
6209 CopyValueToVirtualRegister(C, RegOut);
6213 DenseMap<const Value *, unsigned>::iterator I =
6214 FuncInfo.ValueMap.find(PHIOp);
6215 if (I != FuncInfo.ValueMap.end())
6218 assert(isa<AllocaInst>(PHIOp) &&
6219 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
6220 "Didn't codegen value into a register!??");
6221 Reg = FuncInfo.CreateRegs(PHIOp->getType());
6222 CopyValueToVirtualRegister(PHIOp, Reg);
6226 // Remember that this register needs to added to the machine PHI node as
6227 // the input for this MBB.
6228 SmallVector<EVT, 4> ValueVTs;
6229 ComputeValueVTs(TLI, PN->getType(), ValueVTs);
6230 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
6231 EVT VT = ValueVTs[vti];
6232 unsigned NumRegisters = TLI.getNumRegisters(*DAG.getContext(), VT);
6233 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
6234 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
6235 Reg += NumRegisters;
6239 ConstantsOut.clear();