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];
880 const DbgValueInst *DI = DDI.getDI();
881 DebugLoc dl = DDI.getdl();
882 unsigned DbgSDNodeOrder = DDI.getSDNodeOrder();
883 MDNode *Variable = DI->getVariable();
884 uint64_t Offset = DI->getOffset();
887 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, Val)) {
888 SDV = DAG.getDbgValue(Variable, Val.getNode(),
889 Val.getResNo(), Offset, dl, DbgSDNodeOrder);
890 DAG.AddDbgValue(SDV, Val.getNode(), false);
893 SDV = DAG.getDbgValue(Variable, UndefValue::get(V->getType()),
894 Offset, dl, SDNodeOrder);
895 DAG.AddDbgValue(SDV, 0, false);
897 DanglingDebugInfoMap[V] = DanglingDebugInfo();
901 // getValue - Return an SDValue for the given Value.
902 SDValue SelectionDAGBuilder::getValue(const Value *V) {
903 // If we already have an SDValue for this value, use it. It's important
904 // to do this first, so that we don't create a CopyFromReg if we already
905 // have a regular SDValue.
906 SDValue &N = NodeMap[V];
907 if (N.getNode()) return N;
909 // If there's a virtual register allocated and initialized for this
911 DenseMap<const Value *, unsigned>::iterator It = FuncInfo.ValueMap.find(V);
912 if (It != FuncInfo.ValueMap.end()) {
913 unsigned InReg = It->second;
914 RegsForValue RFV(*DAG.getContext(), TLI, InReg, V->getType());
915 SDValue Chain = DAG.getEntryNode();
916 return N = RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain,NULL);
919 // Otherwise create a new SDValue and remember it.
920 SDValue Val = getValueImpl(V);
922 resolveDanglingDebugInfo(V, Val);
926 /// getNonRegisterValue - Return an SDValue for the given Value, but
927 /// don't look in FuncInfo.ValueMap for a virtual register.
928 SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) {
929 // If we already have an SDValue for this value, use it.
930 SDValue &N = NodeMap[V];
931 if (N.getNode()) return N;
933 // Otherwise create a new SDValue and remember it.
934 SDValue Val = getValueImpl(V);
936 resolveDanglingDebugInfo(V, Val);
940 /// getValueImpl - Helper function for getValue and getNonRegisterValue.
941 /// Create an SDValue for the given value.
942 SDValue SelectionDAGBuilder::getValueImpl(const Value *V) {
943 if (const Constant *C = dyn_cast<Constant>(V)) {
944 EVT VT = TLI.getValueType(V->getType(), true);
946 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C))
947 return DAG.getConstant(*CI, VT);
949 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
950 return DAG.getGlobalAddress(GV, getCurDebugLoc(), VT);
952 if (isa<ConstantPointerNull>(C))
953 return DAG.getConstant(0, TLI.getPointerTy());
955 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
956 return DAG.getConstantFP(*CFP, VT);
958 if (isa<UndefValue>(C) && !V->getType()->isAggregateType())
959 return DAG.getUNDEF(VT);
961 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
962 visit(CE->getOpcode(), *CE);
963 SDValue N1 = NodeMap[V];
964 assert(N1.getNode() && "visit didn't populate the NodeMap!");
968 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) {
969 SmallVector<SDValue, 4> Constants;
970 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end();
972 SDNode *Val = getValue(*OI).getNode();
973 // If the operand is an empty aggregate, there are no values.
975 // Add each leaf value from the operand to the Constants list
976 // to form a flattened list of all the values.
977 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
978 Constants.push_back(SDValue(Val, i));
981 return DAG.getMergeValues(&Constants[0], Constants.size(),
985 if (C->getType()->isStructTy() || C->getType()->isArrayTy()) {
986 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
987 "Unknown struct or array constant!");
989 SmallVector<EVT, 4> ValueVTs;
990 ComputeValueVTs(TLI, C->getType(), ValueVTs);
991 unsigned NumElts = ValueVTs.size();
993 return SDValue(); // empty struct
994 SmallVector<SDValue, 4> Constants(NumElts);
995 for (unsigned i = 0; i != NumElts; ++i) {
996 EVT EltVT = ValueVTs[i];
997 if (isa<UndefValue>(C))
998 Constants[i] = DAG.getUNDEF(EltVT);
999 else if (EltVT.isFloatingPoint())
1000 Constants[i] = DAG.getConstantFP(0, EltVT);
1002 Constants[i] = DAG.getConstant(0, EltVT);
1005 return DAG.getMergeValues(&Constants[0], NumElts,
1009 if (const BlockAddress *BA = dyn_cast<BlockAddress>(C))
1010 return DAG.getBlockAddress(BA, VT);
1012 const VectorType *VecTy = cast<VectorType>(V->getType());
1013 unsigned NumElements = VecTy->getNumElements();
1015 // Now that we know the number and type of the elements, get that number of
1016 // elements into the Ops array based on what kind of constant it is.
1017 SmallVector<SDValue, 16> Ops;
1018 if (const ConstantVector *CP = dyn_cast<ConstantVector>(C)) {
1019 for (unsigned i = 0; i != NumElements; ++i)
1020 Ops.push_back(getValue(CP->getOperand(i)));
1022 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!");
1023 EVT EltVT = TLI.getValueType(VecTy->getElementType());
1026 if (EltVT.isFloatingPoint())
1027 Op = DAG.getConstantFP(0, EltVT);
1029 Op = DAG.getConstant(0, EltVT);
1030 Ops.assign(NumElements, Op);
1033 // Create a BUILD_VECTOR node.
1034 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
1035 VT, &Ops[0], Ops.size());
1038 // If this is a static alloca, generate it as the frameindex instead of
1040 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
1041 DenseMap<const AllocaInst*, int>::iterator SI =
1042 FuncInfo.StaticAllocaMap.find(AI);
1043 if (SI != FuncInfo.StaticAllocaMap.end())
1044 return DAG.getFrameIndex(SI->second, TLI.getPointerTy());
1047 // If this is an instruction which fast-isel has deferred, select it now.
1048 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
1049 unsigned InReg = FuncInfo.InitializeRegForValue(Inst);
1050 RegsForValue RFV(*DAG.getContext(), TLI, InReg, Inst->getType());
1051 SDValue Chain = DAG.getEntryNode();
1052 return RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain, NULL);
1055 llvm_unreachable("Can't get register for value!");
1059 void SelectionDAGBuilder::visitRet(const ReturnInst &I) {
1060 SDValue Chain = getControlRoot();
1061 SmallVector<ISD::OutputArg, 8> Outs;
1062 SmallVector<SDValue, 8> OutVals;
1064 if (!FuncInfo.CanLowerReturn) {
1065 unsigned DemoteReg = FuncInfo.DemoteRegister;
1066 const Function *F = I.getParent()->getParent();
1068 // Emit a store of the return value through the virtual register.
1069 // Leave Outs empty so that LowerReturn won't try to load return
1070 // registers the usual way.
1071 SmallVector<EVT, 1> PtrValueVTs;
1072 ComputeValueVTs(TLI, PointerType::getUnqual(F->getReturnType()),
1075 SDValue RetPtr = DAG.getRegister(DemoteReg, PtrValueVTs[0]);
1076 SDValue RetOp = getValue(I.getOperand(0));
1078 SmallVector<EVT, 4> ValueVTs;
1079 SmallVector<uint64_t, 4> Offsets;
1080 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs, &Offsets);
1081 unsigned NumValues = ValueVTs.size();
1083 SmallVector<SDValue, 4> Chains(NumValues);
1084 for (unsigned i = 0; i != NumValues; ++i) {
1085 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(),
1086 RetPtr.getValueType(), RetPtr,
1087 DAG.getIntPtrConstant(Offsets[i]));
1089 DAG.getStore(Chain, getCurDebugLoc(),
1090 SDValue(RetOp.getNode(), RetOp.getResNo() + i),
1091 Add, NULL, Offsets[i], false, false, 0);
1094 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
1095 MVT::Other, &Chains[0], NumValues);
1096 } else if (I.getNumOperands() != 0) {
1097 SmallVector<EVT, 4> ValueVTs;
1098 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs);
1099 unsigned NumValues = ValueVTs.size();
1101 SDValue RetOp = getValue(I.getOperand(0));
1102 for (unsigned j = 0, f = NumValues; j != f; ++j) {
1103 EVT VT = ValueVTs[j];
1105 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1107 const Function *F = I.getParent()->getParent();
1108 if (F->paramHasAttr(0, Attribute::SExt))
1109 ExtendKind = ISD::SIGN_EXTEND;
1110 else if (F->paramHasAttr(0, Attribute::ZExt))
1111 ExtendKind = ISD::ZERO_EXTEND;
1113 // FIXME: C calling convention requires the return type to be promoted
1114 // to at least 32-bit. But this is not necessary for non-C calling
1115 // conventions. The frontend should mark functions whose return values
1116 // require promoting with signext or zeroext attributes.
1117 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
1118 EVT MinVT = TLI.getRegisterType(*DAG.getContext(), MVT::i32);
1119 if (VT.bitsLT(MinVT))
1123 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), VT);
1124 EVT PartVT = TLI.getRegisterType(*DAG.getContext(), VT);
1125 SmallVector<SDValue, 4> Parts(NumParts);
1126 getCopyToParts(DAG, getCurDebugLoc(),
1127 SDValue(RetOp.getNode(), RetOp.getResNo() + j),
1128 &Parts[0], NumParts, PartVT, ExtendKind);
1130 // 'inreg' on function refers to return value
1131 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1132 if (F->paramHasAttr(0, Attribute::InReg))
1135 // Propagate extension type if any
1136 if (F->paramHasAttr(0, Attribute::SExt))
1138 else if (F->paramHasAttr(0, Attribute::ZExt))
1141 for (unsigned i = 0; i < NumParts; ++i) {
1142 Outs.push_back(ISD::OutputArg(Flags, Parts[i].getValueType(),
1144 OutVals.push_back(Parts[i]);
1150 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1151 CallingConv::ID CallConv =
1152 DAG.getMachineFunction().getFunction()->getCallingConv();
1153 Chain = TLI.LowerReturn(Chain, CallConv, isVarArg,
1154 Outs, OutVals, getCurDebugLoc(), DAG);
1156 // Verify that the target's LowerReturn behaved as expected.
1157 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
1158 "LowerReturn didn't return a valid chain!");
1160 // Update the DAG with the new chain value resulting from return lowering.
1164 /// CopyToExportRegsIfNeeded - If the given value has virtual registers
1165 /// created for it, emit nodes to copy the value into the virtual
1167 void SelectionDAGBuilder::CopyToExportRegsIfNeeded(const Value *V) {
1168 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
1169 if (VMI != FuncInfo.ValueMap.end()) {
1170 assert(!V->use_empty() && "Unused value assigned virtual registers!");
1171 CopyValueToVirtualRegister(V, VMI->second);
1175 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
1176 /// the current basic block, add it to ValueMap now so that we'll get a
1178 void SelectionDAGBuilder::ExportFromCurrentBlock(const Value *V) {
1179 // No need to export constants.
1180 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
1182 // Already exported?
1183 if (FuncInfo.isExportedInst(V)) return;
1185 unsigned Reg = FuncInfo.InitializeRegForValue(V);
1186 CopyValueToVirtualRegister(V, Reg);
1189 bool SelectionDAGBuilder::isExportableFromCurrentBlock(const Value *V,
1190 const BasicBlock *FromBB) {
1191 // The operands of the setcc have to be in this block. We don't know
1192 // how to export them from some other block.
1193 if (const Instruction *VI = dyn_cast<Instruction>(V)) {
1194 // Can export from current BB.
1195 if (VI->getParent() == FromBB)
1198 // Is already exported, noop.
1199 return FuncInfo.isExportedInst(V);
1202 // If this is an argument, we can export it if the BB is the entry block or
1203 // if it is already exported.
1204 if (isa<Argument>(V)) {
1205 if (FromBB == &FromBB->getParent()->getEntryBlock())
1208 // Otherwise, can only export this if it is already exported.
1209 return FuncInfo.isExportedInst(V);
1212 // Otherwise, constants can always be exported.
1216 static bool InBlock(const Value *V, const BasicBlock *BB) {
1217 if (const Instruction *I = dyn_cast<Instruction>(V))
1218 return I->getParent() == BB;
1222 /// EmitBranchForMergedCondition - Helper method for FindMergedConditions.
1223 /// This function emits a branch and is used at the leaves of an OR or an
1224 /// AND operator tree.
1227 SelectionDAGBuilder::EmitBranchForMergedCondition(const Value *Cond,
1228 MachineBasicBlock *TBB,
1229 MachineBasicBlock *FBB,
1230 MachineBasicBlock *CurBB,
1231 MachineBasicBlock *SwitchBB) {
1232 const BasicBlock *BB = CurBB->getBasicBlock();
1234 // If the leaf of the tree is a comparison, merge the condition into
1236 if (const CmpInst *BOp = dyn_cast<CmpInst>(Cond)) {
1237 // The operands of the cmp have to be in this block. We don't know
1238 // how to export them from some other block. If this is the first block
1239 // of the sequence, no exporting is needed.
1240 if (CurBB == SwitchBB ||
1241 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
1242 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) {
1243 ISD::CondCode Condition;
1244 if (const ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
1245 Condition = getICmpCondCode(IC->getPredicate());
1246 } else if (const FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) {
1247 Condition = getFCmpCondCode(FC->getPredicate());
1249 Condition = ISD::SETEQ; // silence warning.
1250 llvm_unreachable("Unknown compare instruction");
1253 CaseBlock CB(Condition, BOp->getOperand(0),
1254 BOp->getOperand(1), NULL, TBB, FBB, CurBB);
1255 SwitchCases.push_back(CB);
1260 // Create a CaseBlock record representing this branch.
1261 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()),
1262 NULL, TBB, FBB, CurBB);
1263 SwitchCases.push_back(CB);
1266 /// FindMergedConditions - If Cond is an expression like
1267 void SelectionDAGBuilder::FindMergedConditions(const Value *Cond,
1268 MachineBasicBlock *TBB,
1269 MachineBasicBlock *FBB,
1270 MachineBasicBlock *CurBB,
1271 MachineBasicBlock *SwitchBB,
1273 // If this node is not part of the or/and tree, emit it as a branch.
1274 const Instruction *BOp = dyn_cast<Instruction>(Cond);
1275 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
1276 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
1277 BOp->getParent() != CurBB->getBasicBlock() ||
1278 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
1279 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
1280 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB);
1284 // Create TmpBB after CurBB.
1285 MachineFunction::iterator BBI = CurBB;
1286 MachineFunction &MF = DAG.getMachineFunction();
1287 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock());
1288 CurBB->getParent()->insert(++BBI, TmpBB);
1290 if (Opc == Instruction::Or) {
1291 // Codegen X | Y as:
1299 // Emit the LHS condition.
1300 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, SwitchBB, Opc);
1302 // Emit the RHS condition into TmpBB.
1303 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
1305 assert(Opc == Instruction::And && "Unknown merge op!");
1306 // Codegen X & Y as:
1313 // This requires creation of TmpBB after CurBB.
1315 // Emit the LHS condition.
1316 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, SwitchBB, Opc);
1318 // Emit the RHS condition into TmpBB.
1319 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
1323 /// If the set of cases should be emitted as a series of branches, return true.
1324 /// If we should emit this as a bunch of and/or'd together conditions, return
1327 SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases){
1328 if (Cases.size() != 2) return true;
1330 // If this is two comparisons of the same values or'd or and'd together, they
1331 // will get folded into a single comparison, so don't emit two blocks.
1332 if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
1333 Cases[0].CmpRHS == Cases[1].CmpRHS) ||
1334 (Cases[0].CmpRHS == Cases[1].CmpLHS &&
1335 Cases[0].CmpLHS == Cases[1].CmpRHS)) {
1339 // Handle: (X != null) | (Y != null) --> (X|Y) != 0
1340 // Handle: (X == null) & (Y == null) --> (X|Y) == 0
1341 if (Cases[0].CmpRHS == Cases[1].CmpRHS &&
1342 Cases[0].CC == Cases[1].CC &&
1343 isa<Constant>(Cases[0].CmpRHS) &&
1344 cast<Constant>(Cases[0].CmpRHS)->isNullValue()) {
1345 if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB)
1347 if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB)
1354 void SelectionDAGBuilder::visitBr(const BranchInst &I) {
1355 MachineBasicBlock *BrMBB = FuncInfo.MBB;
1357 // Update machine-CFG edges.
1358 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
1360 // Figure out which block is immediately after the current one.
1361 MachineBasicBlock *NextBlock = 0;
1362 MachineFunction::iterator BBI = BrMBB;
1363 if (++BBI != FuncInfo.MF->end())
1366 if (I.isUnconditional()) {
1367 // Update machine-CFG edges.
1368 BrMBB->addSuccessor(Succ0MBB);
1370 // If this is not a fall-through branch, emit the branch.
1371 if (Succ0MBB != NextBlock)
1372 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
1373 MVT::Other, getControlRoot(),
1374 DAG.getBasicBlock(Succ0MBB)));
1379 // If this condition is one of the special cases we handle, do special stuff
1381 const Value *CondVal = I.getCondition();
1382 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
1384 // If this is a series of conditions that are or'd or and'd together, emit
1385 // this as a sequence of branches instead of setcc's with and/or operations.
1386 // For example, instead of something like:
1399 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
1400 if (BOp->hasOneUse() &&
1401 (BOp->getOpcode() == Instruction::And ||
1402 BOp->getOpcode() == Instruction::Or)) {
1403 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB,
1405 // If the compares in later blocks need to use values not currently
1406 // exported from this block, export them now. This block should always
1407 // be the first entry.
1408 assert(SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!");
1410 // Allow some cases to be rejected.
1411 if (ShouldEmitAsBranches(SwitchCases)) {
1412 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1413 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1414 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1417 // Emit the branch for this block.
1418 visitSwitchCase(SwitchCases[0], BrMBB);
1419 SwitchCases.erase(SwitchCases.begin());
1423 // Okay, we decided not to do this, remove any inserted MBB's and clear
1425 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
1426 FuncInfo.MF->erase(SwitchCases[i].ThisBB);
1428 SwitchCases.clear();
1432 // Create a CaseBlock record representing this branch.
1433 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()),
1434 NULL, Succ0MBB, Succ1MBB, BrMBB);
1436 // Use visitSwitchCase to actually insert the fast branch sequence for this
1438 visitSwitchCase(CB, BrMBB);
1441 /// visitSwitchCase - Emits the necessary code to represent a single node in
1442 /// the binary search tree resulting from lowering a switch instruction.
1443 void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB,
1444 MachineBasicBlock *SwitchBB) {
1446 SDValue CondLHS = getValue(CB.CmpLHS);
1447 DebugLoc dl = getCurDebugLoc();
1449 // Build the setcc now.
1450 if (CB.CmpMHS == NULL) {
1451 // Fold "(X == true)" to X and "(X == false)" to !X to
1452 // handle common cases produced by branch lowering.
1453 if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) &&
1454 CB.CC == ISD::SETEQ)
1456 else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) &&
1457 CB.CC == ISD::SETEQ) {
1458 SDValue True = DAG.getConstant(1, CondLHS.getValueType());
1459 Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True);
1461 Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1463 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now");
1465 const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue();
1466 const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue();
1468 SDValue CmpOp = getValue(CB.CmpMHS);
1469 EVT VT = CmpOp.getValueType();
1471 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) {
1472 Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, VT),
1475 SDValue SUB = DAG.getNode(ISD::SUB, dl,
1476 VT, CmpOp, DAG.getConstant(Low, VT));
1477 Cond = DAG.getSetCC(dl, MVT::i1, SUB,
1478 DAG.getConstant(High-Low, VT), ISD::SETULE);
1482 // Update successor info
1483 SwitchBB->addSuccessor(CB.TrueBB);
1484 SwitchBB->addSuccessor(CB.FalseBB);
1486 // Set NextBlock to be the MBB immediately after the current one, if any.
1487 // This is used to avoid emitting unnecessary branches to the next block.
1488 MachineBasicBlock *NextBlock = 0;
1489 MachineFunction::iterator BBI = SwitchBB;
1490 if (++BBI != FuncInfo.MF->end())
1493 // If the lhs block is the next block, invert the condition so that we can
1494 // fall through to the lhs instead of the rhs block.
1495 if (CB.TrueBB == NextBlock) {
1496 std::swap(CB.TrueBB, CB.FalseBB);
1497 SDValue True = DAG.getConstant(1, Cond.getValueType());
1498 Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True);
1501 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1502 MVT::Other, getControlRoot(), Cond,
1503 DAG.getBasicBlock(CB.TrueBB));
1505 // Insert the false branch.
1506 if (CB.FalseBB != NextBlock)
1507 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1508 DAG.getBasicBlock(CB.FalseBB));
1510 DAG.setRoot(BrCond);
1513 /// visitJumpTable - Emit JumpTable node in the current MBB
1514 void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) {
1515 // Emit the code for the jump table
1516 assert(JT.Reg != -1U && "Should lower JT Header first!");
1517 EVT PTy = TLI.getPointerTy();
1518 SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(),
1520 SDValue Table = DAG.getJumpTable(JT.JTI, PTy);
1521 SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurDebugLoc(),
1522 MVT::Other, Index.getValue(1),
1524 DAG.setRoot(BrJumpTable);
1527 /// visitJumpTableHeader - This function emits necessary code to produce index
1528 /// in the JumpTable from switch case.
1529 void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT,
1530 JumpTableHeader &JTH,
1531 MachineBasicBlock *SwitchBB) {
1532 // Subtract the lowest switch case value from the value being switched on and
1533 // conditional branch to default mbb if the result is greater than the
1534 // difference between smallest and largest cases.
1535 SDValue SwitchOp = getValue(JTH.SValue);
1536 EVT VT = SwitchOp.getValueType();
1537 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1538 DAG.getConstant(JTH.First, VT));
1540 // The SDNode we just created, which holds the value being switched on minus
1541 // the smallest case value, needs to be copied to a virtual register so it
1542 // can be used as an index into the jump table in a subsequent basic block.
1543 // This value may be smaller or larger than the target's pointer type, and
1544 // therefore require extension or truncating.
1545 SwitchOp = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(), TLI.getPointerTy());
1547 unsigned JumpTableReg = FuncInfo.CreateReg(TLI.getPointerTy());
1548 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1549 JumpTableReg, SwitchOp);
1550 JT.Reg = JumpTableReg;
1552 // Emit the range check for the jump table, and branch to the default block
1553 // for the switch statement if the value being switched on exceeds the largest
1554 // case in the switch.
1555 SDValue CMP = DAG.getSetCC(getCurDebugLoc(),
1556 TLI.getSetCCResultType(Sub.getValueType()), Sub,
1557 DAG.getConstant(JTH.Last-JTH.First,VT),
1560 // Set NextBlock to be the MBB immediately after the current one, if any.
1561 // This is used to avoid emitting unnecessary branches to the next block.
1562 MachineBasicBlock *NextBlock = 0;
1563 MachineFunction::iterator BBI = SwitchBB;
1565 if (++BBI != FuncInfo.MF->end())
1568 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1569 MVT::Other, CopyTo, CMP,
1570 DAG.getBasicBlock(JT.Default));
1572 if (JT.MBB != NextBlock)
1573 BrCond = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrCond,
1574 DAG.getBasicBlock(JT.MBB));
1576 DAG.setRoot(BrCond);
1579 /// visitBitTestHeader - This function emits necessary code to produce value
1580 /// suitable for "bit tests"
1581 void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B,
1582 MachineBasicBlock *SwitchBB) {
1583 // Subtract the minimum value
1584 SDValue SwitchOp = getValue(B.SValue);
1585 EVT VT = SwitchOp.getValueType();
1586 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1587 DAG.getConstant(B.First, VT));
1590 SDValue RangeCmp = DAG.getSetCC(getCurDebugLoc(),
1591 TLI.getSetCCResultType(Sub.getValueType()),
1592 Sub, DAG.getConstant(B.Range, VT),
1595 SDValue ShiftOp = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(),
1596 TLI.getPointerTy());
1598 B.Reg = FuncInfo.CreateReg(TLI.getPointerTy());
1599 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1602 // Set NextBlock to be the MBB immediately after the current one, if any.
1603 // This is used to avoid emitting unnecessary branches to the next block.
1604 MachineBasicBlock *NextBlock = 0;
1605 MachineFunction::iterator BBI = SwitchBB;
1606 if (++BBI != FuncInfo.MF->end())
1609 MachineBasicBlock* MBB = B.Cases[0].ThisBB;
1611 SwitchBB->addSuccessor(B.Default);
1612 SwitchBB->addSuccessor(MBB);
1614 SDValue BrRange = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1615 MVT::Other, CopyTo, RangeCmp,
1616 DAG.getBasicBlock(B.Default));
1618 if (MBB != NextBlock)
1619 BrRange = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, CopyTo,
1620 DAG.getBasicBlock(MBB));
1622 DAG.setRoot(BrRange);
1625 /// visitBitTestCase - this function produces one "bit test"
1626 void SelectionDAGBuilder::visitBitTestCase(MachineBasicBlock* NextMBB,
1629 MachineBasicBlock *SwitchBB) {
1630 SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(), Reg,
1631 TLI.getPointerTy());
1633 if (CountPopulation_64(B.Mask) == 1) {
1634 // Testing for a single bit; just compare the shift count with what it
1635 // would need to be to shift a 1 bit in that position.
1636 Cmp = DAG.getSetCC(getCurDebugLoc(),
1637 TLI.getSetCCResultType(ShiftOp.getValueType()),
1639 DAG.getConstant(CountTrailingZeros_64(B.Mask),
1640 TLI.getPointerTy()),
1643 // Make desired shift
1644 SDValue SwitchVal = DAG.getNode(ISD::SHL, getCurDebugLoc(),
1646 DAG.getConstant(1, TLI.getPointerTy()),
1649 // Emit bit tests and jumps
1650 SDValue AndOp = DAG.getNode(ISD::AND, getCurDebugLoc(),
1651 TLI.getPointerTy(), SwitchVal,
1652 DAG.getConstant(B.Mask, TLI.getPointerTy()));
1653 Cmp = DAG.getSetCC(getCurDebugLoc(),
1654 TLI.getSetCCResultType(AndOp.getValueType()),
1655 AndOp, DAG.getConstant(0, TLI.getPointerTy()),
1659 SwitchBB->addSuccessor(B.TargetBB);
1660 SwitchBB->addSuccessor(NextMBB);
1662 SDValue BrAnd = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1663 MVT::Other, getControlRoot(),
1664 Cmp, DAG.getBasicBlock(B.TargetBB));
1666 // Set NextBlock to be the MBB immediately after the current one, if any.
1667 // This is used to avoid emitting unnecessary branches to the next block.
1668 MachineBasicBlock *NextBlock = 0;
1669 MachineFunction::iterator BBI = SwitchBB;
1670 if (++BBI != FuncInfo.MF->end())
1673 if (NextMBB != NextBlock)
1674 BrAnd = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrAnd,
1675 DAG.getBasicBlock(NextMBB));
1680 void SelectionDAGBuilder::visitInvoke(const InvokeInst &I) {
1681 MachineBasicBlock *InvokeMBB = FuncInfo.MBB;
1683 // Retrieve successors.
1684 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
1685 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)];
1687 const Value *Callee(I.getCalledValue());
1688 if (isa<InlineAsm>(Callee))
1691 LowerCallTo(&I, getValue(Callee), false, LandingPad);
1693 // If the value of the invoke is used outside of its defining block, make it
1694 // available as a virtual register.
1695 CopyToExportRegsIfNeeded(&I);
1697 // Update successor info
1698 InvokeMBB->addSuccessor(Return);
1699 InvokeMBB->addSuccessor(LandingPad);
1701 // Drop into normal successor.
1702 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
1703 MVT::Other, getControlRoot(),
1704 DAG.getBasicBlock(Return)));
1707 void SelectionDAGBuilder::visitUnwind(const UnwindInst &I) {
1710 /// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for
1711 /// small case ranges).
1712 bool SelectionDAGBuilder::handleSmallSwitchRange(CaseRec& CR,
1713 CaseRecVector& WorkList,
1715 MachineBasicBlock *Default,
1716 MachineBasicBlock *SwitchBB) {
1717 Case& BackCase = *(CR.Range.second-1);
1719 // Size is the number of Cases represented by this range.
1720 size_t Size = CR.Range.second - CR.Range.first;
1724 // Get the MachineFunction which holds the current MBB. This is used when
1725 // inserting any additional MBBs necessary to represent the switch.
1726 MachineFunction *CurMF = FuncInfo.MF;
1728 // Figure out which block is immediately after the current one.
1729 MachineBasicBlock *NextBlock = 0;
1730 MachineFunction::iterator BBI = CR.CaseBB;
1732 if (++BBI != FuncInfo.MF->end())
1735 // TODO: If any two of the cases has the same destination, and if one value
1736 // is the same as the other, but has one bit unset that the other has set,
1737 // use bit manipulation to do two compares at once. For example:
1738 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
1740 // Rearrange the case blocks so that the last one falls through if possible.
1741 if (NextBlock && Default != NextBlock && BackCase.BB != NextBlock) {
1742 // The last case block won't fall through into 'NextBlock' if we emit the
1743 // branches in this order. See if rearranging a case value would help.
1744 for (CaseItr I = CR.Range.first, E = CR.Range.second-1; I != E; ++I) {
1745 if (I->BB == NextBlock) {
1746 std::swap(*I, BackCase);
1752 // Create a CaseBlock record representing a conditional branch to
1753 // the Case's target mbb if the value being switched on SV is equal
1755 MachineBasicBlock *CurBlock = CR.CaseBB;
1756 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
1757 MachineBasicBlock *FallThrough;
1759 FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock());
1760 CurMF->insert(BBI, FallThrough);
1762 // Put SV in a virtual register to make it available from the new blocks.
1763 ExportFromCurrentBlock(SV);
1765 // If the last case doesn't match, go to the default block.
1766 FallThrough = Default;
1769 const Value *RHS, *LHS, *MHS;
1771 if (I->High == I->Low) {
1772 // This is just small small case range :) containing exactly 1 case
1774 LHS = SV; RHS = I->High; MHS = NULL;
1777 LHS = I->Low; MHS = SV; RHS = I->High;
1779 CaseBlock CB(CC, LHS, RHS, MHS, I->BB, FallThrough, CurBlock);
1781 // If emitting the first comparison, just call visitSwitchCase to emit the
1782 // code into the current block. Otherwise, push the CaseBlock onto the
1783 // vector to be later processed by SDISel, and insert the node's MBB
1784 // before the next MBB.
1785 if (CurBlock == SwitchBB)
1786 visitSwitchCase(CB, SwitchBB);
1788 SwitchCases.push_back(CB);
1790 CurBlock = FallThrough;
1796 static inline bool areJTsAllowed(const TargetLowering &TLI) {
1797 return !DisableJumpTables &&
1798 (TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
1799 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other));
1802 static APInt ComputeRange(const APInt &First, const APInt &Last) {
1803 APInt LastExt(Last), FirstExt(First);
1804 uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1;
1805 LastExt.sext(BitWidth); FirstExt.sext(BitWidth);
1806 return (LastExt - FirstExt + 1ULL);
1809 /// handleJTSwitchCase - Emit jumptable for current switch case range
1810 bool SelectionDAGBuilder::handleJTSwitchCase(CaseRec& CR,
1811 CaseRecVector& WorkList,
1813 MachineBasicBlock* Default,
1814 MachineBasicBlock *SwitchBB) {
1815 Case& FrontCase = *CR.Range.first;
1816 Case& BackCase = *(CR.Range.second-1);
1818 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
1819 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
1821 APInt TSize(First.getBitWidth(), 0);
1822 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1826 if (!areJTsAllowed(TLI) || TSize.ult(4))
1829 APInt Range = ComputeRange(First, Last);
1830 double Density = TSize.roundToDouble() / Range.roundToDouble();
1834 DEBUG(dbgs() << "Lowering jump table\n"
1835 << "First entry: " << First << ". Last entry: " << Last << '\n'
1836 << "Range: " << Range
1837 << "Size: " << TSize << ". Density: " << Density << "\n\n");
1839 // Get the MachineFunction which holds the current MBB. This is used when
1840 // inserting any additional MBBs necessary to represent the switch.
1841 MachineFunction *CurMF = FuncInfo.MF;
1843 // Figure out which block is immediately after the current one.
1844 MachineFunction::iterator BBI = CR.CaseBB;
1847 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1849 // Create a new basic block to hold the code for loading the address
1850 // of the jump table, and jumping to it. Update successor information;
1851 // we will either branch to the default case for the switch, or the jump
1853 MachineBasicBlock *JumpTableBB = CurMF->CreateMachineBasicBlock(LLVMBB);
1854 CurMF->insert(BBI, JumpTableBB);
1855 CR.CaseBB->addSuccessor(Default);
1856 CR.CaseBB->addSuccessor(JumpTableBB);
1858 // Build a vector of destination BBs, corresponding to each target
1859 // of the jump table. If the value of the jump table slot corresponds to
1860 // a case statement, push the case's BB onto the vector, otherwise, push
1862 std::vector<MachineBasicBlock*> DestBBs;
1864 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) {
1865 const APInt &Low = cast<ConstantInt>(I->Low)->getValue();
1866 const APInt &High = cast<ConstantInt>(I->High)->getValue();
1868 if (Low.sle(TEI) && TEI.sle(High)) {
1869 DestBBs.push_back(I->BB);
1873 DestBBs.push_back(Default);
1877 // Update successor info. Add one edge to each unique successor.
1878 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs());
1879 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(),
1880 E = DestBBs.end(); I != E; ++I) {
1881 if (!SuccsHandled[(*I)->getNumber()]) {
1882 SuccsHandled[(*I)->getNumber()] = true;
1883 JumpTableBB->addSuccessor(*I);
1887 // Create a jump table index for this jump table.
1888 unsigned JTEncoding = TLI.getJumpTableEncoding();
1889 unsigned JTI = CurMF->getOrCreateJumpTableInfo(JTEncoding)
1890 ->createJumpTableIndex(DestBBs);
1892 // Set the jump table information so that we can codegen it as a second
1893 // MachineBasicBlock
1894 JumpTable JT(-1U, JTI, JumpTableBB, Default);
1895 JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == SwitchBB));
1896 if (CR.CaseBB == SwitchBB)
1897 visitJumpTableHeader(JT, JTH, SwitchBB);
1899 JTCases.push_back(JumpTableBlock(JTH, JT));
1904 /// handleBTSplitSwitchCase - emit comparison and split binary search tree into
1906 bool SelectionDAGBuilder::handleBTSplitSwitchCase(CaseRec& CR,
1907 CaseRecVector& WorkList,
1909 MachineBasicBlock *Default,
1910 MachineBasicBlock *SwitchBB) {
1911 // Get the MachineFunction which holds the current MBB. This is used when
1912 // inserting any additional MBBs necessary to represent the switch.
1913 MachineFunction *CurMF = FuncInfo.MF;
1915 // Figure out which block is immediately after the current one.
1916 MachineFunction::iterator BBI = CR.CaseBB;
1919 Case& FrontCase = *CR.Range.first;
1920 Case& BackCase = *(CR.Range.second-1);
1921 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1923 // Size is the number of Cases represented by this range.
1924 unsigned Size = CR.Range.second - CR.Range.first;
1926 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
1927 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
1929 CaseItr Pivot = CR.Range.first + Size/2;
1931 // Select optimal pivot, maximizing sum density of LHS and RHS. This will
1932 // (heuristically) allow us to emit JumpTable's later.
1933 APInt TSize(First.getBitWidth(), 0);
1934 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1938 APInt LSize = FrontCase.size();
1939 APInt RSize = TSize-LSize;
1940 DEBUG(dbgs() << "Selecting best pivot: \n"
1941 << "First: " << First << ", Last: " << Last <<'\n'
1942 << "LSize: " << LSize << ", RSize: " << RSize << '\n');
1943 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second;
1945 const APInt &LEnd = cast<ConstantInt>(I->High)->getValue();
1946 const APInt &RBegin = cast<ConstantInt>(J->Low)->getValue();
1947 APInt Range = ComputeRange(LEnd, RBegin);
1948 assert((Range - 2ULL).isNonNegative() &&
1949 "Invalid case distance");
1950 double LDensity = (double)LSize.roundToDouble() /
1951 (LEnd - First + 1ULL).roundToDouble();
1952 double RDensity = (double)RSize.roundToDouble() /
1953 (Last - RBegin + 1ULL).roundToDouble();
1954 double Metric = Range.logBase2()*(LDensity+RDensity);
1955 // Should always split in some non-trivial place
1956 DEBUG(dbgs() <<"=>Step\n"
1957 << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n'
1958 << "LDensity: " << LDensity
1959 << ", RDensity: " << RDensity << '\n'
1960 << "Metric: " << Metric << '\n');
1961 if (FMetric < Metric) {
1964 DEBUG(dbgs() << "Current metric set to: " << FMetric << '\n');
1970 if (areJTsAllowed(TLI)) {
1971 // If our case is dense we *really* should handle it earlier!
1972 assert((FMetric > 0) && "Should handle dense range earlier!");
1974 Pivot = CR.Range.first + Size/2;
1977 CaseRange LHSR(CR.Range.first, Pivot);
1978 CaseRange RHSR(Pivot, CR.Range.second);
1979 Constant *C = Pivot->Low;
1980 MachineBasicBlock *FalseBB = 0, *TrueBB = 0;
1982 // We know that we branch to the LHS if the Value being switched on is
1983 // less than the Pivot value, C. We use this to optimize our binary
1984 // tree a bit, by recognizing that if SV is greater than or equal to the
1985 // LHS's Case Value, and that Case Value is exactly one less than the
1986 // Pivot's Value, then we can branch directly to the LHS's Target,
1987 // rather than creating a leaf node for it.
1988 if ((LHSR.second - LHSR.first) == 1 &&
1989 LHSR.first->High == CR.GE &&
1990 cast<ConstantInt>(C)->getValue() ==
1991 (cast<ConstantInt>(CR.GE)->getValue() + 1LL)) {
1992 TrueBB = LHSR.first->BB;
1994 TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB);
1995 CurMF->insert(BBI, TrueBB);
1996 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR));
1998 // Put SV in a virtual register to make it available from the new blocks.
1999 ExportFromCurrentBlock(SV);
2002 // Similar to the optimization above, if the Value being switched on is
2003 // known to be less than the Constant CR.LT, and the current Case Value
2004 // is CR.LT - 1, then we can branch directly to the target block for
2005 // the current Case Value, rather than emitting a RHS leaf node for it.
2006 if ((RHSR.second - RHSR.first) == 1 && CR.LT &&
2007 cast<ConstantInt>(RHSR.first->Low)->getValue() ==
2008 (cast<ConstantInt>(CR.LT)->getValue() - 1LL)) {
2009 FalseBB = RHSR.first->BB;
2011 FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2012 CurMF->insert(BBI, FalseBB);
2013 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR));
2015 // Put SV in a virtual register to make it available from the new blocks.
2016 ExportFromCurrentBlock(SV);
2019 // Create a CaseBlock record representing a conditional branch to
2020 // the LHS node if the value being switched on SV is less than C.
2021 // Otherwise, branch to LHS.
2022 CaseBlock CB(ISD::SETLT, SV, C, NULL, TrueBB, FalseBB, CR.CaseBB);
2024 if (CR.CaseBB == SwitchBB)
2025 visitSwitchCase(CB, SwitchBB);
2027 SwitchCases.push_back(CB);
2032 /// handleBitTestsSwitchCase - if current case range has few destination and
2033 /// range span less, than machine word bitwidth, encode case range into series
2034 /// of masks and emit bit tests with these masks.
2035 bool SelectionDAGBuilder::handleBitTestsSwitchCase(CaseRec& CR,
2036 CaseRecVector& WorkList,
2038 MachineBasicBlock* Default,
2039 MachineBasicBlock *SwitchBB){
2040 EVT PTy = TLI.getPointerTy();
2041 unsigned IntPtrBits = PTy.getSizeInBits();
2043 Case& FrontCase = *CR.Range.first;
2044 Case& BackCase = *(CR.Range.second-1);
2046 // Get the MachineFunction which holds the current MBB. This is used when
2047 // inserting any additional MBBs necessary to represent the switch.
2048 MachineFunction *CurMF = FuncInfo.MF;
2050 // If target does not have legal shift left, do not emit bit tests at all.
2051 if (!TLI.isOperationLegal(ISD::SHL, TLI.getPointerTy()))
2055 for (CaseItr I = CR.Range.first, E = CR.Range.second;
2057 // Single case counts one, case range - two.
2058 numCmps += (I->Low == I->High ? 1 : 2);
2061 // Count unique destinations
2062 SmallSet<MachineBasicBlock*, 4> Dests;
2063 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2064 Dests.insert(I->BB);
2065 if (Dests.size() > 3)
2066 // Don't bother the code below, if there are too much unique destinations
2069 DEBUG(dbgs() << "Total number of unique destinations: "
2070 << Dests.size() << '\n'
2071 << "Total number of comparisons: " << numCmps << '\n');
2073 // Compute span of values.
2074 const APInt& minValue = cast<ConstantInt>(FrontCase.Low)->getValue();
2075 const APInt& maxValue = cast<ConstantInt>(BackCase.High)->getValue();
2076 APInt cmpRange = maxValue - minValue;
2078 DEBUG(dbgs() << "Compare range: " << cmpRange << '\n'
2079 << "Low bound: " << minValue << '\n'
2080 << "High bound: " << maxValue << '\n');
2082 if (cmpRange.uge(IntPtrBits) ||
2083 (!(Dests.size() == 1 && numCmps >= 3) &&
2084 !(Dests.size() == 2 && numCmps >= 5) &&
2085 !(Dests.size() >= 3 && numCmps >= 6)))
2088 DEBUG(dbgs() << "Emitting bit tests\n");
2089 APInt lowBound = APInt::getNullValue(cmpRange.getBitWidth());
2091 // Optimize the case where all the case values fit in a
2092 // word without having to subtract minValue. In this case,
2093 // we can optimize away the subtraction.
2094 if (minValue.isNonNegative() && maxValue.slt(IntPtrBits)) {
2095 cmpRange = maxValue;
2097 lowBound = minValue;
2100 CaseBitsVector CasesBits;
2101 unsigned i, count = 0;
2103 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2104 MachineBasicBlock* Dest = I->BB;
2105 for (i = 0; i < count; ++i)
2106 if (Dest == CasesBits[i].BB)
2110 assert((count < 3) && "Too much destinations to test!");
2111 CasesBits.push_back(CaseBits(0, Dest, 0));
2115 const APInt& lowValue = cast<ConstantInt>(I->Low)->getValue();
2116 const APInt& highValue = cast<ConstantInt>(I->High)->getValue();
2118 uint64_t lo = (lowValue - lowBound).getZExtValue();
2119 uint64_t hi = (highValue - lowBound).getZExtValue();
2121 for (uint64_t j = lo; j <= hi; j++) {
2122 CasesBits[i].Mask |= 1ULL << j;
2123 CasesBits[i].Bits++;
2127 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp());
2131 // Figure out which block is immediately after the current one.
2132 MachineFunction::iterator BBI = CR.CaseBB;
2135 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2137 DEBUG(dbgs() << "Cases:\n");
2138 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) {
2139 DEBUG(dbgs() << "Mask: " << CasesBits[i].Mask
2140 << ", Bits: " << CasesBits[i].Bits
2141 << ", BB: " << CasesBits[i].BB << '\n');
2143 MachineBasicBlock *CaseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2144 CurMF->insert(BBI, CaseBB);
2145 BTC.push_back(BitTestCase(CasesBits[i].Mask,
2149 // Put SV in a virtual register to make it available from the new blocks.
2150 ExportFromCurrentBlock(SV);
2153 BitTestBlock BTB(lowBound, cmpRange, SV,
2154 -1U, (CR.CaseBB == SwitchBB),
2155 CR.CaseBB, Default, BTC);
2157 if (CR.CaseBB == SwitchBB)
2158 visitBitTestHeader(BTB, SwitchBB);
2160 BitTestCases.push_back(BTB);
2165 /// Clusterify - Transform simple list of Cases into list of CaseRange's
2166 size_t SelectionDAGBuilder::Clusterify(CaseVector& Cases,
2167 const SwitchInst& SI) {
2170 // Start with "simple" cases
2171 for (size_t i = 1; i < SI.getNumSuccessors(); ++i) {
2172 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SI.getSuccessor(i)];
2173 Cases.push_back(Case(SI.getSuccessorValue(i),
2174 SI.getSuccessorValue(i),
2177 std::sort(Cases.begin(), Cases.end(), CaseCmp());
2179 // Merge case into clusters
2180 if (Cases.size() >= 2)
2181 // Must recompute end() each iteration because it may be
2182 // invalidated by erase if we hold on to it
2183 for (CaseItr I = Cases.begin(), J = ++(Cases.begin()); J != Cases.end(); ) {
2184 const APInt& nextValue = cast<ConstantInt>(J->Low)->getValue();
2185 const APInt& currentValue = cast<ConstantInt>(I->High)->getValue();
2186 MachineBasicBlock* nextBB = J->BB;
2187 MachineBasicBlock* currentBB = I->BB;
2189 // If the two neighboring cases go to the same destination, merge them
2190 // into a single case.
2191 if ((nextValue - currentValue == 1) && (currentBB == nextBB)) {
2199 for (CaseItr I=Cases.begin(), E=Cases.end(); I!=E; ++I, ++numCmps) {
2200 if (I->Low != I->High)
2201 // A range counts double, since it requires two compares.
2208 void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) {
2209 MachineBasicBlock *SwitchMBB = FuncInfo.MBB;
2211 // Figure out which block is immediately after the current one.
2212 MachineBasicBlock *NextBlock = 0;
2213 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()];
2215 // If there is only the default destination, branch to it if it is not the
2216 // next basic block. Otherwise, just fall through.
2217 if (SI.getNumOperands() == 2) {
2218 // Update machine-CFG edges.
2220 // If this is not a fall-through branch, emit the branch.
2221 SwitchMBB->addSuccessor(Default);
2222 if (Default != NextBlock)
2223 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
2224 MVT::Other, getControlRoot(),
2225 DAG.getBasicBlock(Default)));
2230 // If there are any non-default case statements, create a vector of Cases
2231 // representing each one, and sort the vector so that we can efficiently
2232 // create a binary search tree from them.
2234 size_t numCmps = Clusterify(Cases, SI);
2235 DEBUG(dbgs() << "Clusterify finished. Total clusters: " << Cases.size()
2236 << ". Total compares: " << numCmps << '\n');
2239 // Get the Value to be switched on and default basic blocks, which will be
2240 // inserted into CaseBlock records, representing basic blocks in the binary
2242 const Value *SV = SI.getOperand(0);
2244 // Push the initial CaseRec onto the worklist
2245 CaseRecVector WorkList;
2246 WorkList.push_back(CaseRec(SwitchMBB,0,0,
2247 CaseRange(Cases.begin(),Cases.end())));
2249 while (!WorkList.empty()) {
2250 // Grab a record representing a case range to process off the worklist
2251 CaseRec CR = WorkList.back();
2252 WorkList.pop_back();
2254 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2257 // If the range has few cases (two or less) emit a series of specific
2259 if (handleSmallSwitchRange(CR, WorkList, SV, Default, SwitchMBB))
2262 // If the switch has more than 5 blocks, and at least 40% dense, and the
2263 // target supports indirect branches, then emit a jump table rather than
2264 // lowering the switch to a binary tree of conditional branches.
2265 if (handleJTSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2268 // Emit binary tree. We need to pick a pivot, and push left and right ranges
2269 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call.
2270 handleBTSplitSwitchCase(CR, WorkList, SV, Default, SwitchMBB);
2274 void SelectionDAGBuilder::visitIndirectBr(const IndirectBrInst &I) {
2275 MachineBasicBlock *IndirectBrMBB = FuncInfo.MBB;
2277 // Update machine-CFG edges with unique successors.
2278 SmallVector<BasicBlock*, 32> succs;
2279 succs.reserve(I.getNumSuccessors());
2280 for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i)
2281 succs.push_back(I.getSuccessor(i));
2282 array_pod_sort(succs.begin(), succs.end());
2283 succs.erase(std::unique(succs.begin(), succs.end()), succs.end());
2284 for (unsigned i = 0, e = succs.size(); i != e; ++i)
2285 IndirectBrMBB->addSuccessor(FuncInfo.MBBMap[succs[i]]);
2287 DAG.setRoot(DAG.getNode(ISD::BRIND, getCurDebugLoc(),
2288 MVT::Other, getControlRoot(),
2289 getValue(I.getAddress())));
2292 void SelectionDAGBuilder::visitFSub(const User &I) {
2293 // -0.0 - X --> fneg
2294 const Type *Ty = I.getType();
2295 if (Ty->isVectorTy()) {
2296 if (ConstantVector *CV = dyn_cast<ConstantVector>(I.getOperand(0))) {
2297 const VectorType *DestTy = cast<VectorType>(I.getType());
2298 const Type *ElTy = DestTy->getElementType();
2299 unsigned VL = DestTy->getNumElements();
2300 std::vector<Constant*> NZ(VL, ConstantFP::getNegativeZero(ElTy));
2301 Constant *CNZ = ConstantVector::get(&NZ[0], NZ.size());
2303 SDValue Op2 = getValue(I.getOperand(1));
2304 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(),
2305 Op2.getValueType(), Op2));
2311 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I.getOperand(0)))
2312 if (CFP->isExactlyValue(ConstantFP::getNegativeZero(Ty)->getValueAPF())) {
2313 SDValue Op2 = getValue(I.getOperand(1));
2314 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(),
2315 Op2.getValueType(), Op2));
2319 visitBinary(I, ISD::FSUB);
2322 void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) {
2323 SDValue Op1 = getValue(I.getOperand(0));
2324 SDValue Op2 = getValue(I.getOperand(1));
2325 setValue(&I, DAG.getNode(OpCode, getCurDebugLoc(),
2326 Op1.getValueType(), Op1, Op2));
2329 void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) {
2330 SDValue Op1 = getValue(I.getOperand(0));
2331 SDValue Op2 = getValue(I.getOperand(1));
2332 if (!I.getType()->isVectorTy() &&
2333 Op2.getValueType() != TLI.getShiftAmountTy()) {
2334 // If the operand is smaller than the shift count type, promote it.
2335 EVT PTy = TLI.getPointerTy();
2336 EVT STy = TLI.getShiftAmountTy();
2337 if (STy.bitsGT(Op2.getValueType()))
2338 Op2 = DAG.getNode(ISD::ANY_EXTEND, getCurDebugLoc(),
2339 TLI.getShiftAmountTy(), Op2);
2340 // If the operand is larger than the shift count type but the shift
2341 // count type has enough bits to represent any shift value, truncate
2342 // it now. This is a common case and it exposes the truncate to
2343 // optimization early.
2344 else if (STy.getSizeInBits() >=
2345 Log2_32_Ceil(Op2.getValueType().getSizeInBits()))
2346 Op2 = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2347 TLI.getShiftAmountTy(), Op2);
2348 // Otherwise we'll need to temporarily settle for some other
2349 // convenient type; type legalization will make adjustments as
2351 else if (PTy.bitsLT(Op2.getValueType()))
2352 Op2 = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2353 TLI.getPointerTy(), Op2);
2354 else if (PTy.bitsGT(Op2.getValueType()))
2355 Op2 = DAG.getNode(ISD::ANY_EXTEND, getCurDebugLoc(),
2356 TLI.getPointerTy(), Op2);
2359 setValue(&I, DAG.getNode(Opcode, getCurDebugLoc(),
2360 Op1.getValueType(), Op1, Op2));
2363 void SelectionDAGBuilder::visitICmp(const User &I) {
2364 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2365 if (const ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2366 predicate = IC->getPredicate();
2367 else if (const ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2368 predicate = ICmpInst::Predicate(IC->getPredicate());
2369 SDValue Op1 = getValue(I.getOperand(0));
2370 SDValue Op2 = getValue(I.getOperand(1));
2371 ISD::CondCode Opcode = getICmpCondCode(predicate);
2373 EVT DestVT = TLI.getValueType(I.getType());
2374 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Opcode));
2377 void SelectionDAGBuilder::visitFCmp(const User &I) {
2378 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2379 if (const FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2380 predicate = FC->getPredicate();
2381 else if (const ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2382 predicate = FCmpInst::Predicate(FC->getPredicate());
2383 SDValue Op1 = getValue(I.getOperand(0));
2384 SDValue Op2 = getValue(I.getOperand(1));
2385 ISD::CondCode Condition = getFCmpCondCode(predicate);
2386 EVT DestVT = TLI.getValueType(I.getType());
2387 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Condition));
2390 void SelectionDAGBuilder::visitSelect(const User &I) {
2391 SmallVector<EVT, 4> ValueVTs;
2392 ComputeValueVTs(TLI, I.getType(), ValueVTs);
2393 unsigned NumValues = ValueVTs.size();
2394 if (NumValues == 0) return;
2396 SmallVector<SDValue, 4> Values(NumValues);
2397 SDValue Cond = getValue(I.getOperand(0));
2398 SDValue TrueVal = getValue(I.getOperand(1));
2399 SDValue FalseVal = getValue(I.getOperand(2));
2401 for (unsigned i = 0; i != NumValues; ++i)
2402 Values[i] = DAG.getNode(ISD::SELECT, getCurDebugLoc(),
2403 TrueVal.getNode()->getValueType(TrueVal.getResNo()+i),
2405 SDValue(TrueVal.getNode(),
2406 TrueVal.getResNo() + i),
2407 SDValue(FalseVal.getNode(),
2408 FalseVal.getResNo() + i));
2410 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2411 DAG.getVTList(&ValueVTs[0], NumValues),
2412 &Values[0], NumValues));
2415 void SelectionDAGBuilder::visitTrunc(const User &I) {
2416 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2417 SDValue N = getValue(I.getOperand(0));
2418 EVT DestVT = TLI.getValueType(I.getType());
2419 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), DestVT, N));
2422 void SelectionDAGBuilder::visitZExt(const User &I) {
2423 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2424 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2425 SDValue N = getValue(I.getOperand(0));
2426 EVT DestVT = TLI.getValueType(I.getType());
2427 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(), DestVT, N));
2430 void SelectionDAGBuilder::visitSExt(const User &I) {
2431 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2432 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2433 SDValue N = getValue(I.getOperand(0));
2434 EVT DestVT = TLI.getValueType(I.getType());
2435 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurDebugLoc(), DestVT, N));
2438 void SelectionDAGBuilder::visitFPTrunc(const User &I) {
2439 // FPTrunc is never a no-op cast, no need to check
2440 SDValue N = getValue(I.getOperand(0));
2441 EVT DestVT = TLI.getValueType(I.getType());
2442 setValue(&I, DAG.getNode(ISD::FP_ROUND, getCurDebugLoc(),
2443 DestVT, N, DAG.getIntPtrConstant(0)));
2446 void SelectionDAGBuilder::visitFPExt(const User &I){
2447 // FPTrunc is never a no-op cast, no need to check
2448 SDValue N = getValue(I.getOperand(0));
2449 EVT DestVT = TLI.getValueType(I.getType());
2450 setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurDebugLoc(), DestVT, N));
2453 void SelectionDAGBuilder::visitFPToUI(const User &I) {
2454 // FPToUI is never a no-op cast, no need to check
2455 SDValue N = getValue(I.getOperand(0));
2456 EVT DestVT = TLI.getValueType(I.getType());
2457 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurDebugLoc(), DestVT, N));
2460 void SelectionDAGBuilder::visitFPToSI(const User &I) {
2461 // FPToSI is never a no-op cast, no need to check
2462 SDValue N = getValue(I.getOperand(0));
2463 EVT DestVT = TLI.getValueType(I.getType());
2464 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurDebugLoc(), DestVT, N));
2467 void SelectionDAGBuilder::visitUIToFP(const User &I) {
2468 // UIToFP is never a no-op cast, no need to check
2469 SDValue N = getValue(I.getOperand(0));
2470 EVT DestVT = TLI.getValueType(I.getType());
2471 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurDebugLoc(), DestVT, N));
2474 void SelectionDAGBuilder::visitSIToFP(const User &I){
2475 // SIToFP is never a no-op cast, no need to check
2476 SDValue N = getValue(I.getOperand(0));
2477 EVT DestVT = TLI.getValueType(I.getType());
2478 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurDebugLoc(), DestVT, N));
2481 void SelectionDAGBuilder::visitPtrToInt(const User &I) {
2482 // What to do depends on the size of the integer and the size of the pointer.
2483 // We can either truncate, zero extend, or no-op, accordingly.
2484 SDValue N = getValue(I.getOperand(0));
2485 EVT DestVT = TLI.getValueType(I.getType());
2486 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT));
2489 void SelectionDAGBuilder::visitIntToPtr(const User &I) {
2490 // What to do depends on the size of the integer and the size of the pointer.
2491 // We can either truncate, zero extend, or no-op, accordingly.
2492 SDValue N = getValue(I.getOperand(0));
2493 EVT DestVT = TLI.getValueType(I.getType());
2494 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT));
2497 void SelectionDAGBuilder::visitBitCast(const User &I) {
2498 SDValue N = getValue(I.getOperand(0));
2499 EVT DestVT = TLI.getValueType(I.getType());
2501 // BitCast assures us that source and destination are the same size so this is
2502 // either a BIT_CONVERT or a no-op.
2503 if (DestVT != N.getValueType())
2504 setValue(&I, DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
2505 DestVT, N)); // convert types.
2507 setValue(&I, N); // noop cast.
2510 void SelectionDAGBuilder::visitInsertElement(const User &I) {
2511 SDValue InVec = getValue(I.getOperand(0));
2512 SDValue InVal = getValue(I.getOperand(1));
2513 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2515 getValue(I.getOperand(2)));
2516 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurDebugLoc(),
2517 TLI.getValueType(I.getType()),
2518 InVec, InVal, InIdx));
2521 void SelectionDAGBuilder::visitExtractElement(const User &I) {
2522 SDValue InVec = getValue(I.getOperand(0));
2523 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2525 getValue(I.getOperand(1)));
2526 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2527 TLI.getValueType(I.getType()), InVec, InIdx));
2530 // Utility for visitShuffleVector - Returns true if the mask is mask starting
2531 // from SIndx and increasing to the element length (undefs are allowed).
2532 static bool SequentialMask(SmallVectorImpl<int> &Mask, unsigned SIndx) {
2533 unsigned MaskNumElts = Mask.size();
2534 for (unsigned i = 0; i != MaskNumElts; ++i)
2535 if ((Mask[i] >= 0) && (Mask[i] != (int)(i + SIndx)))
2540 void SelectionDAGBuilder::visitShuffleVector(const User &I) {
2541 SmallVector<int, 8> Mask;
2542 SDValue Src1 = getValue(I.getOperand(0));
2543 SDValue Src2 = getValue(I.getOperand(1));
2545 // Convert the ConstantVector mask operand into an array of ints, with -1
2546 // representing undef values.
2547 SmallVector<Constant*, 8> MaskElts;
2548 cast<Constant>(I.getOperand(2))->getVectorElements(MaskElts);
2549 unsigned MaskNumElts = MaskElts.size();
2550 for (unsigned i = 0; i != MaskNumElts; ++i) {
2551 if (isa<UndefValue>(MaskElts[i]))
2554 Mask.push_back(cast<ConstantInt>(MaskElts[i])->getSExtValue());
2557 EVT VT = TLI.getValueType(I.getType());
2558 EVT SrcVT = Src1.getValueType();
2559 unsigned SrcNumElts = SrcVT.getVectorNumElements();
2561 if (SrcNumElts == MaskNumElts) {
2562 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2567 // Normalize the shuffle vector since mask and vector length don't match.
2568 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) {
2569 // Mask is longer than the source vectors and is a multiple of the source
2570 // vectors. We can use concatenate vector to make the mask and vectors
2572 if (SrcNumElts*2 == MaskNumElts && SequentialMask(Mask, 0)) {
2573 // The shuffle is concatenating two vectors together.
2574 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurDebugLoc(),
2579 // Pad both vectors with undefs to make them the same length as the mask.
2580 unsigned NumConcat = MaskNumElts / SrcNumElts;
2581 bool Src1U = Src1.getOpcode() == ISD::UNDEF;
2582 bool Src2U = Src2.getOpcode() == ISD::UNDEF;
2583 SDValue UndefVal = DAG.getUNDEF(SrcVT);
2585 SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal);
2586 SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal);
2590 Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2591 getCurDebugLoc(), VT,
2592 &MOps1[0], NumConcat);
2593 Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2594 getCurDebugLoc(), VT,
2595 &MOps2[0], NumConcat);
2597 // Readjust mask for new input vector length.
2598 SmallVector<int, 8> MappedOps;
2599 for (unsigned i = 0; i != MaskNumElts; ++i) {
2601 if (Idx < (int)SrcNumElts)
2602 MappedOps.push_back(Idx);
2604 MappedOps.push_back(Idx + MaskNumElts - SrcNumElts);
2607 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2612 if (SrcNumElts > MaskNumElts) {
2613 // Analyze the access pattern of the vector to see if we can extract
2614 // two subvectors and do the shuffle. The analysis is done by calculating
2615 // the range of elements the mask access on both vectors.
2616 int MinRange[2] = { SrcNumElts+1, SrcNumElts+1};
2617 int MaxRange[2] = {-1, -1};
2619 for (unsigned i = 0; i != MaskNumElts; ++i) {
2625 if (Idx >= (int)SrcNumElts) {
2629 if (Idx > MaxRange[Input])
2630 MaxRange[Input] = Idx;
2631 if (Idx < MinRange[Input])
2632 MinRange[Input] = Idx;
2635 // Check if the access is smaller than the vector size and can we find
2636 // a reasonable extract index.
2637 int RangeUse[2] = { 2, 2 }; // 0 = Unused, 1 = Extract, 2 = Can not
2639 int StartIdx[2]; // StartIdx to extract from
2640 for (int Input=0; Input < 2; ++Input) {
2641 if (MinRange[Input] == (int)(SrcNumElts+1) && MaxRange[Input] == -1) {
2642 RangeUse[Input] = 0; // Unused
2643 StartIdx[Input] = 0;
2644 } else if (MaxRange[Input] - MinRange[Input] < (int)MaskNumElts) {
2645 // Fits within range but we should see if we can find a good
2646 // start index that is a multiple of the mask length.
2647 if (MaxRange[Input] < (int)MaskNumElts) {
2648 RangeUse[Input] = 1; // Extract from beginning of the vector
2649 StartIdx[Input] = 0;
2651 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts;
2652 if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts &&
2653 StartIdx[Input] + MaskNumElts < SrcNumElts)
2654 RangeUse[Input] = 1; // Extract from a multiple of the mask length.
2659 if (RangeUse[0] == 0 && RangeUse[1] == 0) {
2660 setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used.
2663 else if (RangeUse[0] < 2 && RangeUse[1] < 2) {
2664 // Extract appropriate subvector and generate a vector shuffle
2665 for (int Input=0; Input < 2; ++Input) {
2666 SDValue &Src = Input == 0 ? Src1 : Src2;
2667 if (RangeUse[Input] == 0)
2668 Src = DAG.getUNDEF(VT);
2670 Src = DAG.getNode(ISD::EXTRACT_SUBVECTOR, getCurDebugLoc(), VT,
2671 Src, DAG.getIntPtrConstant(StartIdx[Input]));
2674 // Calculate new mask.
2675 SmallVector<int, 8> MappedOps;
2676 for (unsigned i = 0; i != MaskNumElts; ++i) {
2679 MappedOps.push_back(Idx);
2680 else if (Idx < (int)SrcNumElts)
2681 MappedOps.push_back(Idx - StartIdx[0]);
2683 MappedOps.push_back(Idx - SrcNumElts - StartIdx[1] + MaskNumElts);
2686 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2692 // We can't use either concat vectors or extract subvectors so fall back to
2693 // replacing the shuffle with extract and build vector.
2694 // to insert and build vector.
2695 EVT EltVT = VT.getVectorElementType();
2696 EVT PtrVT = TLI.getPointerTy();
2697 SmallVector<SDValue,8> Ops;
2698 for (unsigned i = 0; i != MaskNumElts; ++i) {
2700 Ops.push_back(DAG.getUNDEF(EltVT));
2705 if (Idx < (int)SrcNumElts)
2706 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2707 EltVT, Src1, DAG.getConstant(Idx, PtrVT));
2709 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2711 DAG.getConstant(Idx - SrcNumElts, PtrVT));
2717 setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
2718 VT, &Ops[0], Ops.size()));
2721 void SelectionDAGBuilder::visitInsertValue(const InsertValueInst &I) {
2722 const Value *Op0 = I.getOperand(0);
2723 const Value *Op1 = I.getOperand(1);
2724 const Type *AggTy = I.getType();
2725 const Type *ValTy = Op1->getType();
2726 bool IntoUndef = isa<UndefValue>(Op0);
2727 bool FromUndef = isa<UndefValue>(Op1);
2729 unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy,
2730 I.idx_begin(), I.idx_end());
2732 SmallVector<EVT, 4> AggValueVTs;
2733 ComputeValueVTs(TLI, AggTy, AggValueVTs);
2734 SmallVector<EVT, 4> ValValueVTs;
2735 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2737 unsigned NumAggValues = AggValueVTs.size();
2738 unsigned NumValValues = ValValueVTs.size();
2739 SmallVector<SDValue, 4> Values(NumAggValues);
2741 SDValue Agg = getValue(Op0);
2742 SDValue Val = getValue(Op1);
2744 // Copy the beginning value(s) from the original aggregate.
2745 for (; i != LinearIndex; ++i)
2746 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2747 SDValue(Agg.getNode(), Agg.getResNo() + i);
2748 // Copy values from the inserted value(s).
2749 for (; i != LinearIndex + NumValValues; ++i)
2750 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2751 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex);
2752 // Copy remaining value(s) from the original aggregate.
2753 for (; i != NumAggValues; ++i)
2754 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2755 SDValue(Agg.getNode(), Agg.getResNo() + i);
2757 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2758 DAG.getVTList(&AggValueVTs[0], NumAggValues),
2759 &Values[0], NumAggValues));
2762 void SelectionDAGBuilder::visitExtractValue(const ExtractValueInst &I) {
2763 const Value *Op0 = I.getOperand(0);
2764 const Type *AggTy = Op0->getType();
2765 const Type *ValTy = I.getType();
2766 bool OutOfUndef = isa<UndefValue>(Op0);
2768 unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy,
2769 I.idx_begin(), I.idx_end());
2771 SmallVector<EVT, 4> ValValueVTs;
2772 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2774 unsigned NumValValues = ValValueVTs.size();
2775 SmallVector<SDValue, 4> Values(NumValValues);
2777 SDValue Agg = getValue(Op0);
2778 // Copy out the selected value(s).
2779 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i)
2780 Values[i - LinearIndex] =
2782 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) :
2783 SDValue(Agg.getNode(), Agg.getResNo() + i);
2785 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2786 DAG.getVTList(&ValValueVTs[0], NumValValues),
2787 &Values[0], NumValValues));
2790 void SelectionDAGBuilder::visitGetElementPtr(const User &I) {
2791 SDValue N = getValue(I.getOperand(0));
2792 const Type *Ty = I.getOperand(0)->getType();
2794 for (GetElementPtrInst::const_op_iterator OI = I.op_begin()+1, E = I.op_end();
2796 const Value *Idx = *OI;
2797 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
2798 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
2801 uint64_t Offset = TD->getStructLayout(StTy)->getElementOffset(Field);
2802 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
2803 DAG.getIntPtrConstant(Offset));
2806 Ty = StTy->getElementType(Field);
2808 Ty = cast<SequentialType>(Ty)->getElementType();
2810 // If this is a constant subscript, handle it quickly.
2811 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
2812 if (CI->isZero()) continue;
2814 TD->getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
2816 EVT PTy = TLI.getPointerTy();
2817 unsigned PtrBits = PTy.getSizeInBits();
2819 OffsVal = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2821 DAG.getConstant(Offs, MVT::i64));
2823 OffsVal = DAG.getIntPtrConstant(Offs);
2825 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
2830 // N = N + Idx * ElementSize;
2831 APInt ElementSize = APInt(TLI.getPointerTy().getSizeInBits(),
2832 TD->getTypeAllocSize(Ty));
2833 SDValue IdxN = getValue(Idx);
2835 // If the index is smaller or larger than intptr_t, truncate or extend
2837 IdxN = DAG.getSExtOrTrunc(IdxN, getCurDebugLoc(), N.getValueType());
2839 // If this is a multiply by a power of two, turn it into a shl
2840 // immediately. This is a very common case.
2841 if (ElementSize != 1) {
2842 if (ElementSize.isPowerOf2()) {
2843 unsigned Amt = ElementSize.logBase2();
2844 IdxN = DAG.getNode(ISD::SHL, getCurDebugLoc(),
2845 N.getValueType(), IdxN,
2846 DAG.getConstant(Amt, TLI.getPointerTy()));
2848 SDValue Scale = DAG.getConstant(ElementSize, TLI.getPointerTy());
2849 IdxN = DAG.getNode(ISD::MUL, getCurDebugLoc(),
2850 N.getValueType(), IdxN, Scale);
2854 N = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2855 N.getValueType(), N, IdxN);
2862 void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) {
2863 // If this is a fixed sized alloca in the entry block of the function,
2864 // allocate it statically on the stack.
2865 if (FuncInfo.StaticAllocaMap.count(&I))
2866 return; // getValue will auto-populate this.
2868 const Type *Ty = I.getAllocatedType();
2869 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
2871 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty),
2874 SDValue AllocSize = getValue(I.getArraySize());
2876 EVT IntPtr = TLI.getPointerTy();
2877 if (AllocSize.getValueType() != IntPtr)
2878 AllocSize = DAG.getZExtOrTrunc(AllocSize, getCurDebugLoc(), IntPtr);
2880 AllocSize = DAG.getNode(ISD::MUL, getCurDebugLoc(), IntPtr,
2882 DAG.getConstant(TySize, IntPtr));
2884 // Handle alignment. If the requested alignment is less than or equal to
2885 // the stack alignment, ignore it. If the size is greater than or equal to
2886 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
2887 unsigned StackAlign = TM.getFrameInfo()->getStackAlignment();
2888 if (Align <= StackAlign)
2891 // Round the size of the allocation up to the stack alignment size
2892 // by add SA-1 to the size.
2893 AllocSize = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2894 AllocSize.getValueType(), AllocSize,
2895 DAG.getIntPtrConstant(StackAlign-1));
2897 // Mask out the low bits for alignment purposes.
2898 AllocSize = DAG.getNode(ISD::AND, getCurDebugLoc(),
2899 AllocSize.getValueType(), AllocSize,
2900 DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1)));
2902 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) };
2903 SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other);
2904 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, getCurDebugLoc(),
2907 DAG.setRoot(DSA.getValue(1));
2909 // Inform the Frame Information that we have just allocated a variable-sized
2911 FuncInfo.MF->getFrameInfo()->CreateVariableSizedObject(Align ? Align : 1);
2914 void SelectionDAGBuilder::visitLoad(const LoadInst &I) {
2915 const Value *SV = I.getOperand(0);
2916 SDValue Ptr = getValue(SV);
2918 const Type *Ty = I.getType();
2920 bool isVolatile = I.isVolatile();
2921 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
2922 unsigned Alignment = I.getAlignment();
2924 SmallVector<EVT, 4> ValueVTs;
2925 SmallVector<uint64_t, 4> Offsets;
2926 ComputeValueVTs(TLI, Ty, ValueVTs, &Offsets);
2927 unsigned NumValues = ValueVTs.size();
2932 bool ConstantMemory = false;
2934 // Serialize volatile loads with other side effects.
2936 else if (AA->pointsToConstantMemory(SV)) {
2937 // Do not serialize (non-volatile) loads of constant memory with anything.
2938 Root = DAG.getEntryNode();
2939 ConstantMemory = true;
2941 // Do not serialize non-volatile loads against each other.
2942 Root = DAG.getRoot();
2945 SmallVector<SDValue, 4> Values(NumValues);
2946 SmallVector<SDValue, 4> Chains(NumValues);
2947 EVT PtrVT = Ptr.getValueType();
2948 for (unsigned i = 0; i != NumValues; ++i) {
2949 SDValue A = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2951 DAG.getConstant(Offsets[i], PtrVT));
2952 SDValue L = DAG.getLoad(ValueVTs[i], getCurDebugLoc(), Root,
2953 A, SV, Offsets[i], isVolatile,
2954 isNonTemporal, Alignment);
2957 Chains[i] = L.getValue(1);
2960 if (!ConstantMemory) {
2961 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
2962 MVT::Other, &Chains[0], NumValues);
2966 PendingLoads.push_back(Chain);
2969 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2970 DAG.getVTList(&ValueVTs[0], NumValues),
2971 &Values[0], NumValues));
2974 void SelectionDAGBuilder::visitStore(const StoreInst &I) {
2975 const Value *SrcV = I.getOperand(0);
2976 const Value *PtrV = I.getOperand(1);
2978 SmallVector<EVT, 4> ValueVTs;
2979 SmallVector<uint64_t, 4> Offsets;
2980 ComputeValueVTs(TLI, SrcV->getType(), ValueVTs, &Offsets);
2981 unsigned NumValues = ValueVTs.size();
2985 // Get the lowered operands. Note that we do this after
2986 // checking if NumResults is zero, because with zero results
2987 // the operands won't have values in the map.
2988 SDValue Src = getValue(SrcV);
2989 SDValue Ptr = getValue(PtrV);
2991 SDValue Root = getRoot();
2992 SmallVector<SDValue, 4> Chains(NumValues);
2993 EVT PtrVT = Ptr.getValueType();
2994 bool isVolatile = I.isVolatile();
2995 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
2996 unsigned Alignment = I.getAlignment();
2998 for (unsigned i = 0; i != NumValues; ++i) {
2999 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT, Ptr,
3000 DAG.getConstant(Offsets[i], PtrVT));
3001 Chains[i] = DAG.getStore(Root, getCurDebugLoc(),
3002 SDValue(Src.getNode(), Src.getResNo() + i),
3003 Add, PtrV, Offsets[i], isVolatile,
3004 isNonTemporal, Alignment);
3007 DAG.setRoot(DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3008 MVT::Other, &Chains[0], NumValues));
3011 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
3013 void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I,
3014 unsigned Intrinsic) {
3015 bool HasChain = !I.doesNotAccessMemory();
3016 bool OnlyLoad = HasChain && I.onlyReadsMemory();
3018 // Build the operand list.
3019 SmallVector<SDValue, 8> Ops;
3020 if (HasChain) { // If this intrinsic has side-effects, chainify it.
3022 // We don't need to serialize loads against other loads.
3023 Ops.push_back(DAG.getRoot());
3025 Ops.push_back(getRoot());
3029 // Info is set by getTgtMemInstrinsic
3030 TargetLowering::IntrinsicInfo Info;
3031 bool IsTgtIntrinsic = TLI.getTgtMemIntrinsic(Info, I, Intrinsic);
3033 // Add the intrinsic ID as an integer operand if it's not a target intrinsic.
3034 if (!IsTgtIntrinsic)
3035 Ops.push_back(DAG.getConstant(Intrinsic, TLI.getPointerTy()));
3037 // Add all operands of the call to the operand list.
3038 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) {
3039 SDValue Op = getValue(I.getArgOperand(i));
3040 assert(TLI.isTypeLegal(Op.getValueType()) &&
3041 "Intrinsic uses a non-legal type?");
3045 SmallVector<EVT, 4> ValueVTs;
3046 ComputeValueVTs(TLI, I.getType(), ValueVTs);
3048 for (unsigned Val = 0, E = ValueVTs.size(); Val != E; ++Val) {
3049 assert(TLI.isTypeLegal(ValueVTs[Val]) &&
3050 "Intrinsic uses a non-legal type?");
3055 ValueVTs.push_back(MVT::Other);
3057 SDVTList VTs = DAG.getVTList(ValueVTs.data(), ValueVTs.size());
3061 if (IsTgtIntrinsic) {
3062 // This is target intrinsic that touches memory
3063 Result = DAG.getMemIntrinsicNode(Info.opc, getCurDebugLoc(),
3064 VTs, &Ops[0], Ops.size(),
3065 Info.memVT, Info.ptrVal, Info.offset,
3066 Info.align, Info.vol,
3067 Info.readMem, Info.writeMem);
3068 } else if (!HasChain) {
3069 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurDebugLoc(),
3070 VTs, &Ops[0], Ops.size());
3071 } else if (!I.getType()->isVoidTy()) {
3072 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurDebugLoc(),
3073 VTs, &Ops[0], Ops.size());
3075 Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurDebugLoc(),
3076 VTs, &Ops[0], Ops.size());
3080 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1);
3082 PendingLoads.push_back(Chain);
3087 if (!I.getType()->isVoidTy()) {
3088 if (const VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
3089 EVT VT = TLI.getValueType(PTy);
3090 Result = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(), VT, Result);
3093 setValue(&I, Result);
3097 /// GetSignificand - Get the significand and build it into a floating-point
3098 /// number with exponent of 1:
3100 /// Op = (Op & 0x007fffff) | 0x3f800000;
3102 /// where Op is the hexidecimal representation of floating point value.
3104 GetSignificand(SelectionDAG &DAG, SDValue Op, DebugLoc dl) {
3105 SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3106 DAG.getConstant(0x007fffff, MVT::i32));
3107 SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1,
3108 DAG.getConstant(0x3f800000, MVT::i32));
3109 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t2);
3112 /// GetExponent - Get the exponent:
3114 /// (float)(int)(((Op & 0x7f800000) >> 23) - 127);
3116 /// where Op is the hexidecimal representation of floating point value.
3118 GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI,
3120 SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3121 DAG.getConstant(0x7f800000, MVT::i32));
3122 SDValue t1 = DAG.getNode(ISD::SRL, dl, MVT::i32, t0,
3123 DAG.getConstant(23, TLI.getPointerTy()));
3124 SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1,
3125 DAG.getConstant(127, MVT::i32));
3126 return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2);
3129 /// getF32Constant - Get 32-bit floating point constant.
3131 getF32Constant(SelectionDAG &DAG, unsigned Flt) {
3132 return DAG.getConstantFP(APFloat(APInt(32, Flt)), MVT::f32);
3135 /// Inlined utility function to implement binary input atomic intrinsics for
3136 /// visitIntrinsicCall: I is a call instruction
3137 /// Op is the associated NodeType for I
3139 SelectionDAGBuilder::implVisitBinaryAtomic(const CallInst& I,
3141 SDValue Root = getRoot();
3143 DAG.getAtomic(Op, getCurDebugLoc(),
3144 getValue(I.getArgOperand(1)).getValueType().getSimpleVT(),
3146 getValue(I.getArgOperand(0)),
3147 getValue(I.getArgOperand(1)),
3148 I.getArgOperand(0));
3150 DAG.setRoot(L.getValue(1));
3154 // implVisitAluOverflow - Lower arithmetic overflow instrinsics.
3156 SelectionDAGBuilder::implVisitAluOverflow(const CallInst &I, ISD::NodeType Op) {
3157 SDValue Op1 = getValue(I.getArgOperand(0));
3158 SDValue Op2 = getValue(I.getArgOperand(1));
3160 SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1);
3161 setValue(&I, DAG.getNode(Op, getCurDebugLoc(), VTs, Op1, Op2));
3165 /// visitExp - Lower an exp intrinsic. Handles the special sequences for
3166 /// limited-precision mode.
3168 SelectionDAGBuilder::visitExp(const CallInst &I) {
3170 DebugLoc dl = getCurDebugLoc();
3172 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3173 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3174 SDValue Op = getValue(I.getArgOperand(0));
3176 // Put the exponent in the right bit position for later addition to the
3179 // #define LOG2OFe 1.4426950f
3180 // IntegerPartOfX = ((int32_t)(X * LOG2OFe));
3181 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3182 getF32Constant(DAG, 0x3fb8aa3b));
3183 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3185 // FractionalPartOfX = (X * LOG2OFe) - (float)IntegerPartOfX;
3186 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3187 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3189 // IntegerPartOfX <<= 23;
3190 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3191 DAG.getConstant(23, TLI.getPointerTy()));
3193 if (LimitFloatPrecision <= 6) {
3194 // For floating-point precision of 6:
3196 // TwoToFractionalPartOfX =
3198 // (0.735607626f + 0.252464424f * x) * x;
3200 // error 0.0144103317, which is 6 bits
3201 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3202 getF32Constant(DAG, 0x3e814304));
3203 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3204 getF32Constant(DAG, 0x3f3c50c8));
3205 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3206 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3207 getF32Constant(DAG, 0x3f7f5e7e));
3208 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,MVT::i32, t5);
3210 // Add the exponent into the result in integer domain.
3211 SDValue t6 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3212 TwoToFracPartOfX, IntegerPartOfX);
3214 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t6);
3215 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3216 // For floating-point precision of 12:
3218 // TwoToFractionalPartOfX =
3221 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3223 // 0.000107046256 error, which is 13 to 14 bits
3224 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3225 getF32Constant(DAG, 0x3da235e3));
3226 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3227 getF32Constant(DAG, 0x3e65b8f3));
3228 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3229 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3230 getF32Constant(DAG, 0x3f324b07));
3231 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3232 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3233 getF32Constant(DAG, 0x3f7ff8fd));
3234 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,MVT::i32, t7);
3236 // Add the exponent into the result in integer domain.
3237 SDValue t8 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3238 TwoToFracPartOfX, IntegerPartOfX);
3240 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t8);
3241 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3242 // For floating-point precision of 18:
3244 // TwoToFractionalPartOfX =
3248 // (0.554906021e-1f +
3249 // (0.961591928e-2f +
3250 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3252 // error 2.47208000*10^(-7), which is better than 18 bits
3253 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3254 getF32Constant(DAG, 0x3924b03e));
3255 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3256 getF32Constant(DAG, 0x3ab24b87));
3257 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3258 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3259 getF32Constant(DAG, 0x3c1d8c17));
3260 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3261 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3262 getF32Constant(DAG, 0x3d634a1d));
3263 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3264 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3265 getF32Constant(DAG, 0x3e75fe14));
3266 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3267 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3268 getF32Constant(DAG, 0x3f317234));
3269 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3270 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3271 getF32Constant(DAG, 0x3f800000));
3272 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,
3275 // Add the exponent into the result in integer domain.
3276 SDValue t14 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3277 TwoToFracPartOfX, IntegerPartOfX);
3279 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t14);
3282 // No special expansion.
3283 result = DAG.getNode(ISD::FEXP, dl,
3284 getValue(I.getArgOperand(0)).getValueType(),
3285 getValue(I.getArgOperand(0)));
3288 setValue(&I, result);
3291 /// visitLog - Lower a log intrinsic. Handles the special sequences for
3292 /// limited-precision mode.
3294 SelectionDAGBuilder::visitLog(const CallInst &I) {
3296 DebugLoc dl = getCurDebugLoc();
3298 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3299 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3300 SDValue Op = getValue(I.getArgOperand(0));
3301 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
3303 // Scale the exponent by log(2) [0.69314718f].
3304 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3305 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3306 getF32Constant(DAG, 0x3f317218));
3308 // Get the significand and build it into a floating-point number with
3310 SDValue X = GetSignificand(DAG, Op1, dl);
3312 if (LimitFloatPrecision <= 6) {
3313 // For floating-point precision of 6:
3317 // (1.4034025f - 0.23903021f * x) * x;
3319 // error 0.0034276066, which is better than 8 bits
3320 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3321 getF32Constant(DAG, 0xbe74c456));
3322 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3323 getF32Constant(DAG, 0x3fb3a2b1));
3324 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3325 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3326 getF32Constant(DAG, 0x3f949a29));
3328 result = DAG.getNode(ISD::FADD, dl,
3329 MVT::f32, LogOfExponent, LogOfMantissa);
3330 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3331 // For floating-point precision of 12:
3337 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x;
3339 // error 0.000061011436, which is 14 bits
3340 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3341 getF32Constant(DAG, 0xbd67b6d6));
3342 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3343 getF32Constant(DAG, 0x3ee4f4b8));
3344 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3345 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3346 getF32Constant(DAG, 0x3fbc278b));
3347 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3348 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3349 getF32Constant(DAG, 0x40348e95));
3350 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3351 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3352 getF32Constant(DAG, 0x3fdef31a));
3354 result = DAG.getNode(ISD::FADD, dl,
3355 MVT::f32, LogOfExponent, LogOfMantissa);
3356 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3357 // For floating-point precision of 18:
3365 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x;
3367 // error 0.0000023660568, which is better than 18 bits
3368 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3369 getF32Constant(DAG, 0xbc91e5ac));
3370 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3371 getF32Constant(DAG, 0x3e4350aa));
3372 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3373 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3374 getF32Constant(DAG, 0x3f60d3e3));
3375 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3376 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3377 getF32Constant(DAG, 0x4011cdf0));
3378 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3379 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3380 getF32Constant(DAG, 0x406cfd1c));
3381 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3382 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3383 getF32Constant(DAG, 0x408797cb));
3384 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3385 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3386 getF32Constant(DAG, 0x4006dcab));
3388 result = DAG.getNode(ISD::FADD, dl,
3389 MVT::f32, LogOfExponent, LogOfMantissa);
3392 // No special expansion.
3393 result = DAG.getNode(ISD::FLOG, dl,
3394 getValue(I.getArgOperand(0)).getValueType(),
3395 getValue(I.getArgOperand(0)));
3398 setValue(&I, result);
3401 /// visitLog2 - Lower a log2 intrinsic. Handles the special sequences for
3402 /// limited-precision mode.
3404 SelectionDAGBuilder::visitLog2(const CallInst &I) {
3406 DebugLoc dl = getCurDebugLoc();
3408 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3409 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3410 SDValue Op = getValue(I.getArgOperand(0));
3411 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
3413 // Get the exponent.
3414 SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl);
3416 // Get the significand and build it into a floating-point number with
3418 SDValue X = GetSignificand(DAG, Op1, dl);
3420 // Different possible minimax approximations of significand in
3421 // floating-point for various degrees of accuracy over [1,2].
3422 if (LimitFloatPrecision <= 6) {
3423 // For floating-point precision of 6:
3425 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x;
3427 // error 0.0049451742, which is more than 7 bits
3428 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3429 getF32Constant(DAG, 0xbeb08fe0));
3430 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3431 getF32Constant(DAG, 0x40019463));
3432 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3433 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3434 getF32Constant(DAG, 0x3fd6633d));
3436 result = DAG.getNode(ISD::FADD, dl,
3437 MVT::f32, LogOfExponent, Log2ofMantissa);
3438 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3439 // For floating-point precision of 12:
3445 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x;
3447 // error 0.0000876136000, which is better than 13 bits
3448 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3449 getF32Constant(DAG, 0xbda7262e));
3450 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3451 getF32Constant(DAG, 0x3f25280b));
3452 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3453 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3454 getF32Constant(DAG, 0x4007b923));
3455 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3456 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3457 getF32Constant(DAG, 0x40823e2f));
3458 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3459 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3460 getF32Constant(DAG, 0x4020d29c));
3462 result = DAG.getNode(ISD::FADD, dl,
3463 MVT::f32, LogOfExponent, Log2ofMantissa);
3464 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3465 // For floating-point precision of 18:
3474 // 0.25691327e-1f * x) * x) * x) * x) * x) * x;
3476 // error 0.0000018516, which is better than 18 bits
3477 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3478 getF32Constant(DAG, 0xbcd2769e));
3479 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3480 getF32Constant(DAG, 0x3e8ce0b9));
3481 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3482 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3483 getF32Constant(DAG, 0x3fa22ae7));
3484 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3485 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3486 getF32Constant(DAG, 0x40525723));
3487 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3488 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3489 getF32Constant(DAG, 0x40aaf200));
3490 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3491 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3492 getF32Constant(DAG, 0x40c39dad));
3493 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3494 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3495 getF32Constant(DAG, 0x4042902c));
3497 result = DAG.getNode(ISD::FADD, dl,
3498 MVT::f32, LogOfExponent, Log2ofMantissa);
3501 // No special expansion.
3502 result = DAG.getNode(ISD::FLOG2, dl,
3503 getValue(I.getArgOperand(0)).getValueType(),
3504 getValue(I.getArgOperand(0)));
3507 setValue(&I, result);
3510 /// visitLog10 - Lower a log10 intrinsic. Handles the special sequences for
3511 /// limited-precision mode.
3513 SelectionDAGBuilder::visitLog10(const CallInst &I) {
3515 DebugLoc dl = getCurDebugLoc();
3517 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3518 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3519 SDValue Op = getValue(I.getArgOperand(0));
3520 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
3522 // Scale the exponent by log10(2) [0.30102999f].
3523 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3524 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3525 getF32Constant(DAG, 0x3e9a209a));
3527 // Get the significand and build it into a floating-point number with
3529 SDValue X = GetSignificand(DAG, Op1, dl);
3531 if (LimitFloatPrecision <= 6) {
3532 // For floating-point precision of 6:
3534 // Log10ofMantissa =
3536 // (0.60948995f - 0.10380950f * x) * x;
3538 // error 0.0014886165, which is 6 bits
3539 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3540 getF32Constant(DAG, 0xbdd49a13));
3541 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3542 getF32Constant(DAG, 0x3f1c0789));
3543 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3544 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3545 getF32Constant(DAG, 0x3f011300));
3547 result = DAG.getNode(ISD::FADD, dl,
3548 MVT::f32, LogOfExponent, Log10ofMantissa);
3549 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3550 // For floating-point precision of 12:
3552 // Log10ofMantissa =
3555 // (-0.31664806f + 0.47637168e-1f * x) * x) * x;
3557 // error 0.00019228036, which is better than 12 bits
3558 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3559 getF32Constant(DAG, 0x3d431f31));
3560 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
3561 getF32Constant(DAG, 0x3ea21fb2));
3562 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3563 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3564 getF32Constant(DAG, 0x3f6ae232));
3565 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3566 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
3567 getF32Constant(DAG, 0x3f25f7c3));
3569 result = DAG.getNode(ISD::FADD, dl,
3570 MVT::f32, LogOfExponent, Log10ofMantissa);
3571 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3572 // For floating-point precision of 18:
3574 // Log10ofMantissa =
3579 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x;
3581 // error 0.0000037995730, which is better than 18 bits
3582 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3583 getF32Constant(DAG, 0x3c5d51ce));
3584 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
3585 getF32Constant(DAG, 0x3e00685a));
3586 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3587 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3588 getF32Constant(DAG, 0x3efb6798));
3589 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3590 SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
3591 getF32Constant(DAG, 0x3f88d192));
3592 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3593 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3594 getF32Constant(DAG, 0x3fc4316c));
3595 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3596 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8,
3597 getF32Constant(DAG, 0x3f57ce70));
3599 result = DAG.getNode(ISD::FADD, dl,
3600 MVT::f32, LogOfExponent, Log10ofMantissa);
3603 // No special expansion.
3604 result = DAG.getNode(ISD::FLOG10, dl,
3605 getValue(I.getArgOperand(0)).getValueType(),
3606 getValue(I.getArgOperand(0)));
3609 setValue(&I, result);
3612 /// visitExp2 - Lower an exp2 intrinsic. Handles the special sequences for
3613 /// limited-precision mode.
3615 SelectionDAGBuilder::visitExp2(const CallInst &I) {
3617 DebugLoc dl = getCurDebugLoc();
3619 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3620 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3621 SDValue Op = getValue(I.getArgOperand(0));
3623 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Op);
3625 // FractionalPartOfX = x - (float)IntegerPartOfX;
3626 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3627 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, Op, t1);
3629 // IntegerPartOfX <<= 23;
3630 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3631 DAG.getConstant(23, TLI.getPointerTy()));
3633 if (LimitFloatPrecision <= 6) {
3634 // For floating-point precision of 6:
3636 // TwoToFractionalPartOfX =
3638 // (0.735607626f + 0.252464424f * x) * x;
3640 // error 0.0144103317, which is 6 bits
3641 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3642 getF32Constant(DAG, 0x3e814304));
3643 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3644 getF32Constant(DAG, 0x3f3c50c8));
3645 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3646 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3647 getF32Constant(DAG, 0x3f7f5e7e));
3648 SDValue t6 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t5);
3649 SDValue TwoToFractionalPartOfX =
3650 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
3652 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3653 MVT::f32, TwoToFractionalPartOfX);
3654 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3655 // For floating-point precision of 12:
3657 // TwoToFractionalPartOfX =
3660 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3662 // error 0.000107046256, which is 13 to 14 bits
3663 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3664 getF32Constant(DAG, 0x3da235e3));
3665 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3666 getF32Constant(DAG, 0x3e65b8f3));
3667 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3668 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3669 getF32Constant(DAG, 0x3f324b07));
3670 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3671 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3672 getF32Constant(DAG, 0x3f7ff8fd));
3673 SDValue t8 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t7);
3674 SDValue TwoToFractionalPartOfX =
3675 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
3677 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3678 MVT::f32, TwoToFractionalPartOfX);
3679 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3680 // For floating-point precision of 18:
3682 // TwoToFractionalPartOfX =
3686 // (0.554906021e-1f +
3687 // (0.961591928e-2f +
3688 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3689 // error 2.47208000*10^(-7), which is better than 18 bits
3690 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3691 getF32Constant(DAG, 0x3924b03e));
3692 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3693 getF32Constant(DAG, 0x3ab24b87));
3694 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3695 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3696 getF32Constant(DAG, 0x3c1d8c17));
3697 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3698 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3699 getF32Constant(DAG, 0x3d634a1d));
3700 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3701 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3702 getF32Constant(DAG, 0x3e75fe14));
3703 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3704 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3705 getF32Constant(DAG, 0x3f317234));
3706 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3707 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3708 getF32Constant(DAG, 0x3f800000));
3709 SDValue t14 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t13);
3710 SDValue TwoToFractionalPartOfX =
3711 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
3713 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3714 MVT::f32, TwoToFractionalPartOfX);
3717 // No special expansion.
3718 result = DAG.getNode(ISD::FEXP2, dl,
3719 getValue(I.getArgOperand(0)).getValueType(),
3720 getValue(I.getArgOperand(0)));
3723 setValue(&I, result);
3726 /// visitPow - Lower a pow intrinsic. Handles the special sequences for
3727 /// limited-precision mode with x == 10.0f.
3729 SelectionDAGBuilder::visitPow(const CallInst &I) {
3731 const Value *Val = I.getArgOperand(0);
3732 DebugLoc dl = getCurDebugLoc();
3733 bool IsExp10 = false;
3735 if (getValue(Val).getValueType() == MVT::f32 &&
3736 getValue(I.getArgOperand(1)).getValueType() == MVT::f32 &&
3737 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3738 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(Val))) {
3739 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
3741 IsExp10 = CFP->getValueAPF().bitwiseIsEqual(Ten);
3746 if (IsExp10 && LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3747 SDValue Op = getValue(I.getArgOperand(1));
3749 // Put the exponent in the right bit position for later addition to the
3752 // #define LOG2OF10 3.3219281f
3753 // IntegerPartOfX = (int32_t)(x * LOG2OF10);
3754 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3755 getF32Constant(DAG, 0x40549a78));
3756 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3758 // FractionalPartOfX = x - (float)IntegerPartOfX;
3759 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3760 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3762 // IntegerPartOfX <<= 23;
3763 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3764 DAG.getConstant(23, TLI.getPointerTy()));
3766 if (LimitFloatPrecision <= 6) {
3767 // For floating-point precision of 6:
3769 // twoToFractionalPartOfX =
3771 // (0.735607626f + 0.252464424f * x) * x;
3773 // error 0.0144103317, which is 6 bits
3774 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3775 getF32Constant(DAG, 0x3e814304));
3776 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3777 getF32Constant(DAG, 0x3f3c50c8));
3778 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3779 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3780 getF32Constant(DAG, 0x3f7f5e7e));
3781 SDValue t6 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t5);
3782 SDValue TwoToFractionalPartOfX =
3783 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
3785 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3786 MVT::f32, TwoToFractionalPartOfX);
3787 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3788 // For floating-point precision of 12:
3790 // TwoToFractionalPartOfX =
3793 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3795 // error 0.000107046256, which is 13 to 14 bits
3796 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3797 getF32Constant(DAG, 0x3da235e3));
3798 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3799 getF32Constant(DAG, 0x3e65b8f3));
3800 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3801 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3802 getF32Constant(DAG, 0x3f324b07));
3803 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3804 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3805 getF32Constant(DAG, 0x3f7ff8fd));
3806 SDValue t8 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t7);
3807 SDValue TwoToFractionalPartOfX =
3808 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
3810 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3811 MVT::f32, TwoToFractionalPartOfX);
3812 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3813 // For floating-point precision of 18:
3815 // TwoToFractionalPartOfX =
3819 // (0.554906021e-1f +
3820 // (0.961591928e-2f +
3821 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3822 // error 2.47208000*10^(-7), which is better than 18 bits
3823 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3824 getF32Constant(DAG, 0x3924b03e));
3825 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3826 getF32Constant(DAG, 0x3ab24b87));
3827 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3828 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3829 getF32Constant(DAG, 0x3c1d8c17));
3830 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3831 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3832 getF32Constant(DAG, 0x3d634a1d));
3833 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3834 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3835 getF32Constant(DAG, 0x3e75fe14));
3836 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3837 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3838 getF32Constant(DAG, 0x3f317234));
3839 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3840 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3841 getF32Constant(DAG, 0x3f800000));
3842 SDValue t14 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t13);
3843 SDValue TwoToFractionalPartOfX =
3844 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
3846 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3847 MVT::f32, TwoToFractionalPartOfX);
3850 // No special expansion.
3851 result = DAG.getNode(ISD::FPOW, dl,
3852 getValue(I.getArgOperand(0)).getValueType(),
3853 getValue(I.getArgOperand(0)),
3854 getValue(I.getArgOperand(1)));
3857 setValue(&I, result);
3861 /// ExpandPowI - Expand a llvm.powi intrinsic.
3862 static SDValue ExpandPowI(DebugLoc DL, SDValue LHS, SDValue RHS,
3863 SelectionDAG &DAG) {
3864 // If RHS is a constant, we can expand this out to a multiplication tree,
3865 // otherwise we end up lowering to a call to __powidf2 (for example). When
3866 // optimizing for size, we only want to do this if the expansion would produce
3867 // a small number of multiplies, otherwise we do the full expansion.
3868 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3869 // Get the exponent as a positive value.
3870 unsigned Val = RHSC->getSExtValue();
3871 if ((int)Val < 0) Val = -Val;
3873 // powi(x, 0) -> 1.0
3875 return DAG.getConstantFP(1.0, LHS.getValueType());
3877 const Function *F = DAG.getMachineFunction().getFunction();
3878 if (!F->hasFnAttr(Attribute::OptimizeForSize) ||
3879 // If optimizing for size, don't insert too many multiplies. This
3880 // inserts up to 5 multiplies.
3881 CountPopulation_32(Val)+Log2_32(Val) < 7) {
3882 // We use the simple binary decomposition method to generate the multiply
3883 // sequence. There are more optimal ways to do this (for example,
3884 // powi(x,15) generates one more multiply than it should), but this has
3885 // the benefit of being both really simple and much better than a libcall.
3886 SDValue Res; // Logically starts equal to 1.0
3887 SDValue CurSquare = LHS;
3891 Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare);
3893 Res = CurSquare; // 1.0*CurSquare.
3896 CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(),
3897 CurSquare, CurSquare);
3901 // If the original was negative, invert the result, producing 1/(x*x*x).
3902 if (RHSC->getSExtValue() < 0)
3903 Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(),
3904 DAG.getConstantFP(1.0, LHS.getValueType()), Res);
3909 // Otherwise, expand to a libcall.
3910 return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS);
3913 /// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a function
3914 /// argument, create the corresponding DBG_VALUE machine instruction for it now.
3915 /// At the end of instruction selection, they will be inserted to the entry BB.
3917 SelectionDAGBuilder::EmitFuncArgumentDbgValue(const Value *V, MDNode *Variable,
3920 if (!isa<Argument>(V))
3923 MachineFunction &MF = DAG.getMachineFunction();
3924 // Ignore inlined function arguments here.
3925 DIVariable DV(Variable);
3926 if (DV.isInlinedFnArgument(MF.getFunction()))
3929 MachineBasicBlock *MBB = FuncInfo.MBB;
3930 if (MBB != &MF.front())
3934 if (N.getNode() && N.getOpcode() == ISD::CopyFromReg) {
3935 Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg();
3936 if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) {
3937 MachineRegisterInfo &RegInfo = MF.getRegInfo();
3938 unsigned PR = RegInfo.getLiveInPhysReg(Reg);
3945 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
3946 if (VMI == FuncInfo.ValueMap.end())
3951 const TargetInstrInfo *TII = DAG.getTarget().getInstrInfo();
3952 MachineInstrBuilder MIB = BuildMI(MF, getCurDebugLoc(),
3953 TII->get(TargetOpcode::DBG_VALUE))
3954 .addReg(Reg, RegState::Debug).addImm(Offset).addMetadata(Variable);
3955 FuncInfo.ArgDbgValues.push_back(&*MIB);
3959 // VisualStudio defines setjmp as _setjmp
3960 #if defined(_MSC_VER) && defined(setjmp)
3961 #define setjmp_undefined_for_visual_studio
3965 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
3966 /// we want to emit this as a call to a named external function, return the name
3967 /// otherwise lower it and return null.
3969 SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) {
3970 DebugLoc dl = getCurDebugLoc();
3973 switch (Intrinsic) {
3975 // By default, turn this into a target intrinsic node.
3976 visitTargetIntrinsic(I, Intrinsic);
3978 case Intrinsic::vastart: visitVAStart(I); return 0;
3979 case Intrinsic::vaend: visitVAEnd(I); return 0;
3980 case Intrinsic::vacopy: visitVACopy(I); return 0;
3981 case Intrinsic::returnaddress:
3982 setValue(&I, DAG.getNode(ISD::RETURNADDR, dl, TLI.getPointerTy(),
3983 getValue(I.getArgOperand(0))));
3985 case Intrinsic::frameaddress:
3986 setValue(&I, DAG.getNode(ISD::FRAMEADDR, dl, TLI.getPointerTy(),
3987 getValue(I.getArgOperand(0))));
3989 case Intrinsic::setjmp:
3990 return "_setjmp"+!TLI.usesUnderscoreSetJmp();
3991 case Intrinsic::longjmp:
3992 return "_longjmp"+!TLI.usesUnderscoreLongJmp();
3993 case Intrinsic::memcpy: {
3994 // Assert for address < 256 since we support only user defined address
3996 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
3998 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4000 "Unknown address space");
4001 SDValue Op1 = getValue(I.getArgOperand(0));
4002 SDValue Op2 = getValue(I.getArgOperand(1));
4003 SDValue Op3 = getValue(I.getArgOperand(2));
4004 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4005 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4006 DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, isVol, false,
4007 I.getArgOperand(0), 0, I.getArgOperand(1), 0));
4010 case Intrinsic::memset: {
4011 // Assert for address < 256 since we support only user defined address
4013 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4015 "Unknown address space");
4016 SDValue Op1 = getValue(I.getArgOperand(0));
4017 SDValue Op2 = getValue(I.getArgOperand(1));
4018 SDValue Op3 = getValue(I.getArgOperand(2));
4019 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4020 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4021 DAG.setRoot(DAG.getMemset(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4022 I.getArgOperand(0), 0));
4025 case Intrinsic::memmove: {
4026 // Assert for address < 256 since we support only user defined address
4028 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4030 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4032 "Unknown address space");
4033 SDValue Op1 = getValue(I.getArgOperand(0));
4034 SDValue Op2 = getValue(I.getArgOperand(1));
4035 SDValue Op3 = getValue(I.getArgOperand(2));
4036 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4037 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4039 // If the source and destination are known to not be aliases, we can
4040 // lower memmove as memcpy.
4041 uint64_t Size = -1ULL;
4042 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op3))
4043 Size = C->getZExtValue();
4044 if (AA->alias(I.getArgOperand(0), Size, I.getArgOperand(1), Size) ==
4045 AliasAnalysis::NoAlias) {
4046 DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4047 false, I.getArgOperand(0), 0,
4048 I.getArgOperand(1), 0));
4052 DAG.setRoot(DAG.getMemmove(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4053 I.getArgOperand(0), 0, I.getArgOperand(1), 0));
4056 case Intrinsic::dbg_declare: {
4057 const DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
4058 if (!DIVariable(DI.getVariable()).Verify())
4061 MDNode *Variable = DI.getVariable();
4062 // Parameters are handled specially.
4064 DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable;
4065 const Value *Address = DI.getAddress();
4068 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
4069 Address = BCI->getOperand(0);
4070 const AllocaInst *AI = dyn_cast<AllocaInst>(Address);
4072 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder
4073 // but do not always have a corresponding SDNode built. The SDNodeOrder
4074 // absolute, but not relative, values are different depending on whether
4075 // debug info exists.
4077 SDValue &N = NodeMap[Address];
4080 if (isParameter && !AI) {
4081 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode());
4083 // Byval parameter. We have a frame index at this point.
4084 SDV = DAG.getDbgValue(Variable, FINode->getIndex(),
4085 0, dl, SDNodeOrder);
4087 // Can't do anything with other non-AI cases yet. This might be a
4088 // parameter of a callee function that got inlined, for example.
4091 SDV = DAG.getDbgValue(Variable, N.getNode(), N.getResNo(),
4092 0, dl, SDNodeOrder);
4094 // Can't do anything with other non-AI cases yet.
4096 DAG.AddDbgValue(SDV, N.getNode(), isParameter);
4098 // If Address is an arugment then try to emits its dbg value using
4099 // virtual register info from the FuncInfo.ValueMap. Otherwise add undef
4100 // to help track missing debug info.
4101 if (!EmitFuncArgumentDbgValue(Address, Variable, 0, N)) {
4102 SDV = DAG.getDbgValue(Variable, UndefValue::get(Address->getType()),
4103 0, dl, SDNodeOrder);
4104 DAG.AddDbgValue(SDV, 0, isParameter);
4109 case Intrinsic::dbg_value: {
4110 const DbgValueInst &DI = cast<DbgValueInst>(I);
4111 if (!DIVariable(DI.getVariable()).Verify())
4114 MDNode *Variable = DI.getVariable();
4115 uint64_t Offset = DI.getOffset();
4116 const Value *V = DI.getValue();
4120 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder
4121 // but do not always have a corresponding SDNode built. The SDNodeOrder
4122 // absolute, but not relative, values are different depending on whether
4123 // debug info exists.
4126 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) {
4127 SDV = DAG.getDbgValue(Variable, V, Offset, dl, SDNodeOrder);
4128 DAG.AddDbgValue(SDV, 0, false);
4130 // Do not use getValue() in here; we don't want to generate code at
4131 // this point if it hasn't been done yet.
4132 SDValue N = NodeMap[V];
4133 if (!N.getNode() && isa<Argument>(V))
4134 // Check unused arguments map.
4135 N = UnusedArgNodeMap[V];
4137 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, N)) {
4138 SDV = DAG.getDbgValue(Variable, N.getNode(),
4139 N.getResNo(), Offset, dl, SDNodeOrder);
4140 DAG.AddDbgValue(SDV, N.getNode(), false);
4142 } else if (isa<PHINode>(V) && !V->use_empty() ) {
4143 // Do not call getValue(V) yet, as we don't want to generate code.
4144 // Remember it for later.
4145 DanglingDebugInfo DDI(&DI, dl, SDNodeOrder);
4146 DanglingDebugInfoMap[V] = DDI;
4148 // We may expand this to cover more cases. One case where we have no
4149 // data available is an unreferenced parameter; we need this fallback.
4150 SDV = DAG.getDbgValue(Variable, UndefValue::get(V->getType()),
4151 Offset, dl, SDNodeOrder);
4152 DAG.AddDbgValue(SDV, 0, false);
4156 // Build a debug info table entry.
4157 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
4158 V = BCI->getOperand(0);
4159 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
4160 // Don't handle byval struct arguments or VLAs, for example.
4163 DenseMap<const AllocaInst*, int>::iterator SI =
4164 FuncInfo.StaticAllocaMap.find(AI);
4165 if (SI == FuncInfo.StaticAllocaMap.end())
4167 int FI = SI->second;
4169 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4170 if (!DI.getDebugLoc().isUnknown() && MMI.hasDebugInfo())
4171 MMI.setVariableDbgInfo(Variable, FI, DI.getDebugLoc());
4174 case Intrinsic::eh_exception: {
4175 // Insert the EXCEPTIONADDR instruction.
4176 assert(FuncInfo.MBB->isLandingPad() &&
4177 "Call to eh.exception not in landing pad!");
4178 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
4180 Ops[0] = DAG.getRoot();
4181 SDValue Op = DAG.getNode(ISD::EXCEPTIONADDR, dl, VTs, Ops, 1);
4183 DAG.setRoot(Op.getValue(1));
4187 case Intrinsic::eh_selector: {
4188 MachineBasicBlock *CallMBB = FuncInfo.MBB;
4189 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4190 if (CallMBB->isLandingPad())
4191 AddCatchInfo(I, &MMI, CallMBB);
4194 FuncInfo.CatchInfoLost.insert(&I);
4196 // FIXME: Mark exception selector register as live in. Hack for PR1508.
4197 unsigned Reg = TLI.getExceptionSelectorRegister();
4198 if (Reg) FuncInfo.MBB->addLiveIn(Reg);
4201 // Insert the EHSELECTION instruction.
4202 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
4204 Ops[0] = getValue(I.getArgOperand(0));
4206 SDValue Op = DAG.getNode(ISD::EHSELECTION, dl, VTs, Ops, 2);
4207 DAG.setRoot(Op.getValue(1));
4208 setValue(&I, DAG.getSExtOrTrunc(Op, dl, MVT::i32));
4212 case Intrinsic::eh_typeid_for: {
4213 // Find the type id for the given typeinfo.
4214 GlobalVariable *GV = ExtractTypeInfo(I.getArgOperand(0));
4215 unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(GV);
4216 Res = DAG.getConstant(TypeID, MVT::i32);
4221 case Intrinsic::eh_return_i32:
4222 case Intrinsic::eh_return_i64:
4223 DAG.getMachineFunction().getMMI().setCallsEHReturn(true);
4224 DAG.setRoot(DAG.getNode(ISD::EH_RETURN, dl,
4227 getValue(I.getArgOperand(0)),
4228 getValue(I.getArgOperand(1))));
4230 case Intrinsic::eh_unwind_init:
4231 DAG.getMachineFunction().getMMI().setCallsUnwindInit(true);
4233 case Intrinsic::eh_dwarf_cfa: {
4234 SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getArgOperand(0)), dl,
4235 TLI.getPointerTy());
4236 SDValue Offset = DAG.getNode(ISD::ADD, dl,
4238 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, dl,
4239 TLI.getPointerTy()),
4241 SDValue FA = DAG.getNode(ISD::FRAMEADDR, dl,
4243 DAG.getConstant(0, TLI.getPointerTy()));
4244 setValue(&I, DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
4248 case Intrinsic::eh_sjlj_callsite: {
4249 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4250 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(0));
4251 assert(CI && "Non-constant call site value in eh.sjlj.callsite!");
4252 assert(MMI.getCurrentCallSite() == 0 && "Overlapping call sites!");
4254 MMI.setCurrentCallSite(CI->getZExtValue());
4257 case Intrinsic::eh_sjlj_setjmp: {
4258 setValue(&I, DAG.getNode(ISD::EH_SJLJ_SETJMP, dl, MVT::i32, getRoot(),
4259 getValue(I.getArgOperand(0))));
4262 case Intrinsic::eh_sjlj_longjmp: {
4263 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, dl, MVT::Other,
4265 getValue(I.getArgOperand(0))));
4269 case Intrinsic::convertff:
4270 case Intrinsic::convertfsi:
4271 case Intrinsic::convertfui:
4272 case Intrinsic::convertsif:
4273 case Intrinsic::convertuif:
4274 case Intrinsic::convertss:
4275 case Intrinsic::convertsu:
4276 case Intrinsic::convertus:
4277 case Intrinsic::convertuu: {
4278 ISD::CvtCode Code = ISD::CVT_INVALID;
4279 switch (Intrinsic) {
4280 case Intrinsic::convertff: Code = ISD::CVT_FF; break;
4281 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break;
4282 case Intrinsic::convertfui: Code = ISD::CVT_FU; break;
4283 case Intrinsic::convertsif: Code = ISD::CVT_SF; break;
4284 case Intrinsic::convertuif: Code = ISD::CVT_UF; break;
4285 case Intrinsic::convertss: Code = ISD::CVT_SS; break;
4286 case Intrinsic::convertsu: Code = ISD::CVT_SU; break;
4287 case Intrinsic::convertus: Code = ISD::CVT_US; break;
4288 case Intrinsic::convertuu: Code = ISD::CVT_UU; break;
4290 EVT DestVT = TLI.getValueType(I.getType());
4291 const Value *Op1 = I.getArgOperand(0);
4292 Res = DAG.getConvertRndSat(DestVT, getCurDebugLoc(), getValue(Op1),
4293 DAG.getValueType(DestVT),
4294 DAG.getValueType(getValue(Op1).getValueType()),
4295 getValue(I.getArgOperand(1)),
4296 getValue(I.getArgOperand(2)),
4301 case Intrinsic::sqrt:
4302 setValue(&I, DAG.getNode(ISD::FSQRT, dl,
4303 getValue(I.getArgOperand(0)).getValueType(),
4304 getValue(I.getArgOperand(0))));
4306 case Intrinsic::powi:
4307 setValue(&I, ExpandPowI(dl, getValue(I.getArgOperand(0)),
4308 getValue(I.getArgOperand(1)), DAG));
4310 case Intrinsic::sin:
4311 setValue(&I, DAG.getNode(ISD::FSIN, dl,
4312 getValue(I.getArgOperand(0)).getValueType(),
4313 getValue(I.getArgOperand(0))));
4315 case Intrinsic::cos:
4316 setValue(&I, DAG.getNode(ISD::FCOS, dl,
4317 getValue(I.getArgOperand(0)).getValueType(),
4318 getValue(I.getArgOperand(0))));
4320 case Intrinsic::log:
4323 case Intrinsic::log2:
4326 case Intrinsic::log10:
4329 case Intrinsic::exp:
4332 case Intrinsic::exp2:
4335 case Intrinsic::pow:
4338 case Intrinsic::convert_to_fp16:
4339 setValue(&I, DAG.getNode(ISD::FP32_TO_FP16, dl,
4340 MVT::i16, getValue(I.getArgOperand(0))));
4342 case Intrinsic::convert_from_fp16:
4343 setValue(&I, DAG.getNode(ISD::FP16_TO_FP32, dl,
4344 MVT::f32, getValue(I.getArgOperand(0))));
4346 case Intrinsic::pcmarker: {
4347 SDValue Tmp = getValue(I.getArgOperand(0));
4348 DAG.setRoot(DAG.getNode(ISD::PCMARKER, dl, MVT::Other, getRoot(), Tmp));
4351 case Intrinsic::readcyclecounter: {
4352 SDValue Op = getRoot();
4353 Res = DAG.getNode(ISD::READCYCLECOUNTER, dl,
4354 DAG.getVTList(MVT::i64, MVT::Other),
4357 DAG.setRoot(Res.getValue(1));
4360 case Intrinsic::bswap:
4361 setValue(&I, DAG.getNode(ISD::BSWAP, dl,
4362 getValue(I.getArgOperand(0)).getValueType(),
4363 getValue(I.getArgOperand(0))));
4365 case Intrinsic::cttz: {
4366 SDValue Arg = getValue(I.getArgOperand(0));
4367 EVT Ty = Arg.getValueType();
4368 setValue(&I, DAG.getNode(ISD::CTTZ, dl, Ty, Arg));
4371 case Intrinsic::ctlz: {
4372 SDValue Arg = getValue(I.getArgOperand(0));
4373 EVT Ty = Arg.getValueType();
4374 setValue(&I, DAG.getNode(ISD::CTLZ, dl, Ty, Arg));
4377 case Intrinsic::ctpop: {
4378 SDValue Arg = getValue(I.getArgOperand(0));
4379 EVT Ty = Arg.getValueType();
4380 setValue(&I, DAG.getNode(ISD::CTPOP, dl, Ty, Arg));
4383 case Intrinsic::stacksave: {
4384 SDValue Op = getRoot();
4385 Res = DAG.getNode(ISD::STACKSAVE, dl,
4386 DAG.getVTList(TLI.getPointerTy(), MVT::Other), &Op, 1);
4388 DAG.setRoot(Res.getValue(1));
4391 case Intrinsic::stackrestore: {
4392 Res = getValue(I.getArgOperand(0));
4393 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, dl, MVT::Other, getRoot(), Res));
4396 case Intrinsic::stackprotector: {
4397 // Emit code into the DAG to store the stack guard onto the stack.
4398 MachineFunction &MF = DAG.getMachineFunction();
4399 MachineFrameInfo *MFI = MF.getFrameInfo();
4400 EVT PtrTy = TLI.getPointerTy();
4402 SDValue Src = getValue(I.getArgOperand(0)); // The guard's value.
4403 AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
4405 int FI = FuncInfo.StaticAllocaMap[Slot];
4406 MFI->setStackProtectorIndex(FI);
4408 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
4410 // Store the stack protector onto the stack.
4411 Res = DAG.getStore(getRoot(), getCurDebugLoc(), Src, FIN,
4412 PseudoSourceValue::getFixedStack(FI),
4418 case Intrinsic::objectsize: {
4419 // If we don't know by now, we're never going to know.
4420 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1));
4422 assert(CI && "Non-constant type in __builtin_object_size?");
4424 SDValue Arg = getValue(I.getCalledValue());
4425 EVT Ty = Arg.getValueType();
4428 Res = DAG.getConstant(-1ULL, Ty);
4430 Res = DAG.getConstant(0, Ty);
4435 case Intrinsic::var_annotation:
4436 // Discard annotate attributes
4439 case Intrinsic::init_trampoline: {
4440 const Function *F = cast<Function>(I.getArgOperand(1)->stripPointerCasts());
4444 Ops[1] = getValue(I.getArgOperand(0));
4445 Ops[2] = getValue(I.getArgOperand(1));
4446 Ops[3] = getValue(I.getArgOperand(2));
4447 Ops[4] = DAG.getSrcValue(I.getArgOperand(0));
4448 Ops[5] = DAG.getSrcValue(F);
4450 Res = DAG.getNode(ISD::TRAMPOLINE, dl,
4451 DAG.getVTList(TLI.getPointerTy(), MVT::Other),
4455 DAG.setRoot(Res.getValue(1));
4458 case Intrinsic::gcroot:
4460 const Value *Alloca = I.getArgOperand(0);
4461 const Constant *TypeMap = cast<Constant>(I.getArgOperand(1));
4463 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode());
4464 GFI->addStackRoot(FI->getIndex(), TypeMap);
4467 case Intrinsic::gcread:
4468 case Intrinsic::gcwrite:
4469 llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!");
4471 case Intrinsic::flt_rounds:
4472 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, dl, MVT::i32));
4474 case Intrinsic::trap:
4475 DAG.setRoot(DAG.getNode(ISD::TRAP, dl,MVT::Other, getRoot()));
4477 case Intrinsic::uadd_with_overflow:
4478 return implVisitAluOverflow(I, ISD::UADDO);
4479 case Intrinsic::sadd_with_overflow:
4480 return implVisitAluOverflow(I, ISD::SADDO);
4481 case Intrinsic::usub_with_overflow:
4482 return implVisitAluOverflow(I, ISD::USUBO);
4483 case Intrinsic::ssub_with_overflow:
4484 return implVisitAluOverflow(I, ISD::SSUBO);
4485 case Intrinsic::umul_with_overflow:
4486 return implVisitAluOverflow(I, ISD::UMULO);
4487 case Intrinsic::smul_with_overflow:
4488 return implVisitAluOverflow(I, ISD::SMULO);
4490 case Intrinsic::prefetch: {
4493 Ops[1] = getValue(I.getArgOperand(0));
4494 Ops[2] = getValue(I.getArgOperand(1));
4495 Ops[3] = getValue(I.getArgOperand(2));
4496 DAG.setRoot(DAG.getNode(ISD::PREFETCH, dl, MVT::Other, &Ops[0], 4));
4500 case Intrinsic::memory_barrier: {
4503 for (int x = 1; x < 6; ++x)
4504 Ops[x] = getValue(I.getArgOperand(x - 1));
4506 DAG.setRoot(DAG.getNode(ISD::MEMBARRIER, dl, MVT::Other, &Ops[0], 6));
4509 case Intrinsic::atomic_cmp_swap: {
4510 SDValue Root = getRoot();
4512 DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, getCurDebugLoc(),
4513 getValue(I.getArgOperand(1)).getValueType().getSimpleVT(),
4515 getValue(I.getArgOperand(0)),
4516 getValue(I.getArgOperand(1)),
4517 getValue(I.getArgOperand(2)),
4518 I.getArgOperand(0));
4520 DAG.setRoot(L.getValue(1));
4523 case Intrinsic::atomic_load_add:
4524 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_ADD);
4525 case Intrinsic::atomic_load_sub:
4526 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_SUB);
4527 case Intrinsic::atomic_load_or:
4528 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_OR);
4529 case Intrinsic::atomic_load_xor:
4530 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_XOR);
4531 case Intrinsic::atomic_load_and:
4532 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_AND);
4533 case Intrinsic::atomic_load_nand:
4534 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_NAND);
4535 case Intrinsic::atomic_load_max:
4536 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MAX);
4537 case Intrinsic::atomic_load_min:
4538 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MIN);
4539 case Intrinsic::atomic_load_umin:
4540 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMIN);
4541 case Intrinsic::atomic_load_umax:
4542 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMAX);
4543 case Intrinsic::atomic_swap:
4544 return implVisitBinaryAtomic(I, ISD::ATOMIC_SWAP);
4546 case Intrinsic::invariant_start:
4547 case Intrinsic::lifetime_start:
4548 // Discard region information.
4549 setValue(&I, DAG.getUNDEF(TLI.getPointerTy()));
4551 case Intrinsic::invariant_end:
4552 case Intrinsic::lifetime_end:
4553 // Discard region information.
4558 void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee,
4560 MachineBasicBlock *LandingPad) {
4561 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
4562 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
4563 const Type *RetTy = FTy->getReturnType();
4564 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4565 MCSymbol *BeginLabel = 0;
4567 TargetLowering::ArgListTy Args;
4568 TargetLowering::ArgListEntry Entry;
4569 Args.reserve(CS.arg_size());
4571 // Check whether the function can return without sret-demotion.
4572 SmallVector<ISD::OutputArg, 4> Outs;
4573 SmallVector<uint64_t, 4> Offsets;
4574 GetReturnInfo(RetTy, CS.getAttributes().getRetAttributes(),
4575 Outs, TLI, &Offsets);
4577 bool CanLowerReturn = TLI.CanLowerReturn(CS.getCallingConv(),
4578 FTy->isVarArg(), Outs, FTy->getContext());
4580 SDValue DemoteStackSlot;
4582 if (!CanLowerReturn) {
4583 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(
4584 FTy->getReturnType());
4585 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(
4586 FTy->getReturnType());
4587 MachineFunction &MF = DAG.getMachineFunction();
4588 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
4589 const Type *StackSlotPtrType = PointerType::getUnqual(FTy->getReturnType());
4591 DemoteStackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
4592 Entry.Node = DemoteStackSlot;
4593 Entry.Ty = StackSlotPtrType;
4594 Entry.isSExt = false;
4595 Entry.isZExt = false;
4596 Entry.isInReg = false;
4597 Entry.isSRet = true;
4598 Entry.isNest = false;
4599 Entry.isByVal = false;
4600 Entry.Alignment = Align;
4601 Args.push_back(Entry);
4602 RetTy = Type::getVoidTy(FTy->getContext());
4605 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
4607 SDValue ArgNode = getValue(*i);
4608 Entry.Node = ArgNode; Entry.Ty = (*i)->getType();
4610 unsigned attrInd = i - CS.arg_begin() + 1;
4611 Entry.isSExt = CS.paramHasAttr(attrInd, Attribute::SExt);
4612 Entry.isZExt = CS.paramHasAttr(attrInd, Attribute::ZExt);
4613 Entry.isInReg = CS.paramHasAttr(attrInd, Attribute::InReg);
4614 Entry.isSRet = CS.paramHasAttr(attrInd, Attribute::StructRet);
4615 Entry.isNest = CS.paramHasAttr(attrInd, Attribute::Nest);
4616 Entry.isByVal = CS.paramHasAttr(attrInd, Attribute::ByVal);
4617 Entry.Alignment = CS.getParamAlignment(attrInd);
4618 Args.push_back(Entry);
4622 // Insert a label before the invoke call to mark the try range. This can be
4623 // used to detect deletion of the invoke via the MachineModuleInfo.
4624 BeginLabel = MMI.getContext().CreateTempSymbol();
4626 // For SjLj, keep track of which landing pads go with which invokes
4627 // so as to maintain the ordering of pads in the LSDA.
4628 unsigned CallSiteIndex = MMI.getCurrentCallSite();
4629 if (CallSiteIndex) {
4630 MMI.setCallSiteBeginLabel(BeginLabel, CallSiteIndex);
4631 // Now that the call site is handled, stop tracking it.
4632 MMI.setCurrentCallSite(0);
4635 // Both PendingLoads and PendingExports must be flushed here;
4636 // this call might not return.
4638 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getControlRoot(), BeginLabel));
4641 // Check if target-independent constraints permit a tail call here.
4642 // Target-dependent constraints are checked within TLI.LowerCallTo.
4644 !isInTailCallPosition(CS, CS.getAttributes().getRetAttributes(), TLI))
4647 // If there's a possibility that fast-isel has already selected some amount
4648 // of the current basic block, don't emit a tail call.
4649 if (isTailCall && EnableFastISel)
4652 std::pair<SDValue,SDValue> Result =
4653 TLI.LowerCallTo(getRoot(), RetTy,
4654 CS.paramHasAttr(0, Attribute::SExt),
4655 CS.paramHasAttr(0, Attribute::ZExt), FTy->isVarArg(),
4656 CS.paramHasAttr(0, Attribute::InReg), FTy->getNumParams(),
4657 CS.getCallingConv(),
4659 !CS.getInstruction()->use_empty(),
4660 Callee, Args, DAG, getCurDebugLoc());
4661 assert((isTailCall || Result.second.getNode()) &&
4662 "Non-null chain expected with non-tail call!");
4663 assert((Result.second.getNode() || !Result.first.getNode()) &&
4664 "Null value expected with tail call!");
4665 if (Result.first.getNode()) {
4666 setValue(CS.getInstruction(), Result.first);
4667 } else if (!CanLowerReturn && Result.second.getNode()) {
4668 // The instruction result is the result of loading from the
4669 // hidden sret parameter.
4670 SmallVector<EVT, 1> PVTs;
4671 const Type *PtrRetTy = PointerType::getUnqual(FTy->getReturnType());
4673 ComputeValueVTs(TLI, PtrRetTy, PVTs);
4674 assert(PVTs.size() == 1 && "Pointers should fit in one register");
4675 EVT PtrVT = PVTs[0];
4676 unsigned NumValues = Outs.size();
4677 SmallVector<SDValue, 4> Values(NumValues);
4678 SmallVector<SDValue, 4> Chains(NumValues);
4680 for (unsigned i = 0; i < NumValues; ++i) {
4681 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT,
4683 DAG.getConstant(Offsets[i], PtrVT));
4684 SDValue L = DAG.getLoad(Outs[i].VT, getCurDebugLoc(), Result.second,
4685 Add, NULL, Offsets[i], false, false, 1);
4687 Chains[i] = L.getValue(1);
4690 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
4691 MVT::Other, &Chains[0], NumValues);
4692 PendingLoads.push_back(Chain);
4694 // Collect the legal value parts into potentially illegal values
4695 // that correspond to the original function's return values.
4696 SmallVector<EVT, 4> RetTys;
4697 RetTy = FTy->getReturnType();
4698 ComputeValueVTs(TLI, RetTy, RetTys);
4699 ISD::NodeType AssertOp = ISD::DELETED_NODE;
4700 SmallVector<SDValue, 4> ReturnValues;
4701 unsigned CurReg = 0;
4702 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
4704 EVT RegisterVT = TLI.getRegisterType(RetTy->getContext(), VT);
4705 unsigned NumRegs = TLI.getNumRegisters(RetTy->getContext(), VT);
4707 SDValue ReturnValue =
4708 getCopyFromParts(DAG, getCurDebugLoc(), &Values[CurReg], NumRegs,
4709 RegisterVT, VT, AssertOp);
4710 ReturnValues.push_back(ReturnValue);
4714 setValue(CS.getInstruction(),
4715 DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
4716 DAG.getVTList(&RetTys[0], RetTys.size()),
4717 &ReturnValues[0], ReturnValues.size()));
4721 // As a special case, a null chain means that a tail call has been emitted and
4722 // the DAG root is already updated.
4723 if (Result.second.getNode())
4724 DAG.setRoot(Result.second);
4729 // Insert a label at the end of the invoke call to mark the try range. This
4730 // can be used to detect deletion of the invoke via the MachineModuleInfo.
4731 MCSymbol *EndLabel = MMI.getContext().CreateTempSymbol();
4732 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getRoot(), EndLabel));
4734 // Inform MachineModuleInfo of range.
4735 MMI.addInvoke(LandingPad, BeginLabel, EndLabel);
4739 /// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the
4740 /// value is equal or not-equal to zero.
4741 static bool IsOnlyUsedInZeroEqualityComparison(const Value *V) {
4742 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end();
4744 if (const ICmpInst *IC = dyn_cast<ICmpInst>(*UI))
4745 if (IC->isEquality())
4746 if (const Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
4747 if (C->isNullValue())
4749 // Unknown instruction.
4755 static SDValue getMemCmpLoad(const Value *PtrVal, MVT LoadVT,
4757 SelectionDAGBuilder &Builder) {
4759 // Check to see if this load can be trivially constant folded, e.g. if the
4760 // input is from a string literal.
4761 if (const Constant *LoadInput = dyn_cast<Constant>(PtrVal)) {
4762 // Cast pointer to the type we really want to load.
4763 LoadInput = ConstantExpr::getBitCast(const_cast<Constant *>(LoadInput),
4764 PointerType::getUnqual(LoadTy));
4766 if (const Constant *LoadCst =
4767 ConstantFoldLoadFromConstPtr(const_cast<Constant *>(LoadInput),
4769 return Builder.getValue(LoadCst);
4772 // Otherwise, we have to emit the load. If the pointer is to unfoldable but
4773 // still constant memory, the input chain can be the entry node.
4775 bool ConstantMemory = false;
4777 // Do not serialize (non-volatile) loads of constant memory with anything.
4778 if (Builder.AA->pointsToConstantMemory(PtrVal)) {
4779 Root = Builder.DAG.getEntryNode();
4780 ConstantMemory = true;
4782 // Do not serialize non-volatile loads against each other.
4783 Root = Builder.DAG.getRoot();
4786 SDValue Ptr = Builder.getValue(PtrVal);
4787 SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurDebugLoc(), Root,
4788 Ptr, PtrVal /*SrcValue*/, 0/*SVOffset*/,
4790 false /*nontemporal*/, 1 /* align=1 */);
4792 if (!ConstantMemory)
4793 Builder.PendingLoads.push_back(LoadVal.getValue(1));
4798 /// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form.
4799 /// If so, return true and lower it, otherwise return false and it will be
4800 /// lowered like a normal call.
4801 bool SelectionDAGBuilder::visitMemCmpCall(const CallInst &I) {
4802 // Verify that the prototype makes sense. int memcmp(void*,void*,size_t)
4803 if (I.getNumArgOperands() != 3)
4806 const Value *LHS = I.getArgOperand(0), *RHS = I.getArgOperand(1);
4807 if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() ||
4808 !I.getArgOperand(2)->getType()->isIntegerTy() ||
4809 !I.getType()->isIntegerTy())
4812 const ConstantInt *Size = dyn_cast<ConstantInt>(I.getArgOperand(2));
4814 // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0
4815 // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0
4816 if (Size && IsOnlyUsedInZeroEqualityComparison(&I)) {
4817 bool ActuallyDoIt = true;
4820 switch (Size->getZExtValue()) {
4822 LoadVT = MVT::Other;
4824 ActuallyDoIt = false;
4828 LoadTy = Type::getInt16Ty(Size->getContext());
4832 LoadTy = Type::getInt32Ty(Size->getContext());
4836 LoadTy = Type::getInt64Ty(Size->getContext());
4840 LoadVT = MVT::v4i32;
4841 LoadTy = Type::getInt32Ty(Size->getContext());
4842 LoadTy = VectorType::get(LoadTy, 4);
4847 // This turns into unaligned loads. We only do this if the target natively
4848 // supports the MVT we'll be loading or if it is small enough (<= 4) that
4849 // we'll only produce a small number of byte loads.
4851 // Require that we can find a legal MVT, and only do this if the target
4852 // supports unaligned loads of that type. Expanding into byte loads would
4854 if (ActuallyDoIt && Size->getZExtValue() > 4) {
4855 // TODO: Handle 5 byte compare as 4-byte + 1 byte.
4856 // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads.
4857 if (!TLI.isTypeLegal(LoadVT) ||!TLI.allowsUnalignedMemoryAccesses(LoadVT))
4858 ActuallyDoIt = false;
4862 SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this);
4863 SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this);
4865 SDValue Res = DAG.getSetCC(getCurDebugLoc(), MVT::i1, LHSVal, RHSVal,
4867 EVT CallVT = TLI.getValueType(I.getType(), true);
4868 setValue(&I, DAG.getZExtOrTrunc(Res, getCurDebugLoc(), CallVT));
4878 void SelectionDAGBuilder::visitCall(const CallInst &I) {
4879 // Handle inline assembly differently.
4880 if (isa<InlineAsm>(I.getCalledValue())) {
4885 const char *RenameFn = 0;
4886 if (Function *F = I.getCalledFunction()) {
4887 if (F->isDeclaration()) {
4888 if (const TargetIntrinsicInfo *II = TM.getIntrinsicInfo()) {
4889 if (unsigned IID = II->getIntrinsicID(F)) {
4890 RenameFn = visitIntrinsicCall(I, IID);
4895 if (unsigned IID = F->getIntrinsicID()) {
4896 RenameFn = visitIntrinsicCall(I, IID);
4902 // Check for well-known libc/libm calls. If the function is internal, it
4903 // can't be a library call.
4904 if (!F->hasLocalLinkage() && F->hasName()) {
4905 StringRef Name = F->getName();
4906 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl") {
4907 if (I.getNumArgOperands() == 2 && // Basic sanity checks.
4908 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
4909 I.getType() == I.getArgOperand(0)->getType() &&
4910 I.getType() == I.getArgOperand(1)->getType()) {
4911 SDValue LHS = getValue(I.getArgOperand(0));
4912 SDValue RHS = getValue(I.getArgOperand(1));
4913 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurDebugLoc(),
4914 LHS.getValueType(), LHS, RHS));
4917 } else if (Name == "fabs" || Name == "fabsf" || Name == "fabsl") {
4918 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
4919 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
4920 I.getType() == I.getArgOperand(0)->getType()) {
4921 SDValue Tmp = getValue(I.getArgOperand(0));
4922 setValue(&I, DAG.getNode(ISD::FABS, getCurDebugLoc(),
4923 Tmp.getValueType(), Tmp));
4926 } else if (Name == "sin" || Name == "sinf" || Name == "sinl") {
4927 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
4928 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
4929 I.getType() == I.getArgOperand(0)->getType() &&
4930 I.onlyReadsMemory()) {
4931 SDValue Tmp = getValue(I.getArgOperand(0));
4932 setValue(&I, DAG.getNode(ISD::FSIN, getCurDebugLoc(),
4933 Tmp.getValueType(), Tmp));
4936 } else if (Name == "cos" || Name == "cosf" || Name == "cosl") {
4937 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
4938 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
4939 I.getType() == I.getArgOperand(0)->getType() &&
4940 I.onlyReadsMemory()) {
4941 SDValue Tmp = getValue(I.getArgOperand(0));
4942 setValue(&I, DAG.getNode(ISD::FCOS, getCurDebugLoc(),
4943 Tmp.getValueType(), Tmp));
4946 } else if (Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl") {
4947 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
4948 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
4949 I.getType() == I.getArgOperand(0)->getType() &&
4950 I.onlyReadsMemory()) {
4951 SDValue Tmp = getValue(I.getArgOperand(0));
4952 setValue(&I, DAG.getNode(ISD::FSQRT, getCurDebugLoc(),
4953 Tmp.getValueType(), Tmp));
4956 } else if (Name == "memcmp") {
4957 if (visitMemCmpCall(I))
4965 Callee = getValue(I.getCalledValue());
4967 Callee = DAG.getExternalSymbol(RenameFn, TLI.getPointerTy());
4969 // Check if we can potentially perform a tail call. More detailed checking is
4970 // be done within LowerCallTo, after more information about the call is known.
4971 LowerCallTo(&I, Callee, I.isTailCall());
4976 /// AsmOperandInfo - This contains information for each constraint that we are
4978 class LLVM_LIBRARY_VISIBILITY SDISelAsmOperandInfo :
4979 public TargetLowering::AsmOperandInfo {
4981 /// CallOperand - If this is the result output operand or a clobber
4982 /// this is null, otherwise it is the incoming operand to the CallInst.
4983 /// This gets modified as the asm is processed.
4984 SDValue CallOperand;
4986 /// AssignedRegs - If this is a register or register class operand, this
4987 /// contains the set of register corresponding to the operand.
4988 RegsForValue AssignedRegs;
4990 explicit SDISelAsmOperandInfo(const InlineAsm::ConstraintInfo &info)
4991 : TargetLowering::AsmOperandInfo(info), CallOperand(0,0) {
4994 /// MarkAllocatedRegs - Once AssignedRegs is set, mark the assigned registers
4995 /// busy in OutputRegs/InputRegs.
4996 void MarkAllocatedRegs(bool isOutReg, bool isInReg,
4997 std::set<unsigned> &OutputRegs,
4998 std::set<unsigned> &InputRegs,
4999 const TargetRegisterInfo &TRI) const {
5001 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
5002 MarkRegAndAliases(AssignedRegs.Regs[i], OutputRegs, TRI);
5005 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
5006 MarkRegAndAliases(AssignedRegs.Regs[i], InputRegs, TRI);
5010 /// getCallOperandValEVT - Return the EVT of the Value* that this operand
5011 /// corresponds to. If there is no Value* for this operand, it returns
5013 EVT getCallOperandValEVT(LLVMContext &Context,
5014 const TargetLowering &TLI,
5015 const TargetData *TD) const {
5016 if (CallOperandVal == 0) return MVT::Other;
5018 if (isa<BasicBlock>(CallOperandVal))
5019 return TLI.getPointerTy();
5021 const llvm::Type *OpTy = CallOperandVal->getType();
5023 // If this is an indirect operand, the operand is a pointer to the
5026 const llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
5028 report_fatal_error("Indirect operand for inline asm not a pointer!");
5029 OpTy = PtrTy->getElementType();
5032 // If OpTy is not a single value, it may be a struct/union that we
5033 // can tile with integers.
5034 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
5035 unsigned BitSize = TD->getTypeSizeInBits(OpTy);
5044 OpTy = IntegerType::get(Context, BitSize);
5049 return TLI.getValueType(OpTy, true);
5053 /// MarkRegAndAliases - Mark the specified register and all aliases in the
5055 static void MarkRegAndAliases(unsigned Reg, std::set<unsigned> &Regs,
5056 const TargetRegisterInfo &TRI) {
5057 assert(TargetRegisterInfo::isPhysicalRegister(Reg) && "Isn't a physreg");
5059 if (const unsigned *Aliases = TRI.getAliasSet(Reg))
5060 for (; *Aliases; ++Aliases)
5061 Regs.insert(*Aliases);
5065 } // end llvm namespace.
5067 /// isAllocatableRegister - If the specified register is safe to allocate,
5068 /// i.e. it isn't a stack pointer or some other special register, return the
5069 /// register class for the register. Otherwise, return null.
5070 static const TargetRegisterClass *
5071 isAllocatableRegister(unsigned Reg, MachineFunction &MF,
5072 const TargetLowering &TLI,
5073 const TargetRegisterInfo *TRI) {
5074 EVT FoundVT = MVT::Other;
5075 const TargetRegisterClass *FoundRC = 0;
5076 for (TargetRegisterInfo::regclass_iterator RCI = TRI->regclass_begin(),
5077 E = TRI->regclass_end(); RCI != E; ++RCI) {
5078 EVT ThisVT = MVT::Other;
5080 const TargetRegisterClass *RC = *RCI;
5081 // If none of the value types for this register class are valid, we
5082 // can't use it. For example, 64-bit reg classes on 32-bit targets.
5083 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
5085 if (TLI.isTypeLegal(*I)) {
5086 // If we have already found this register in a different register class,
5087 // choose the one with the largest VT specified. For example, on
5088 // PowerPC, we favor f64 register classes over f32.
5089 if (FoundVT == MVT::Other || FoundVT.bitsLT(*I)) {
5096 if (ThisVT == MVT::Other) continue;
5098 // NOTE: This isn't ideal. In particular, this might allocate the
5099 // frame pointer in functions that need it (due to them not being taken
5100 // out of allocation, because a variable sized allocation hasn't been seen
5101 // yet). This is a slight code pessimization, but should still work.
5102 for (TargetRegisterClass::iterator I = RC->allocation_order_begin(MF),
5103 E = RC->allocation_order_end(MF); I != E; ++I)
5105 // We found a matching register class. Keep looking at others in case
5106 // we find one with larger registers that this physreg is also in.
5115 /// GetRegistersForValue - Assign registers (virtual or physical) for the
5116 /// specified operand. We prefer to assign virtual registers, to allow the
5117 /// register allocator to handle the assignment process. However, if the asm
5118 /// uses features that we can't model on machineinstrs, we have SDISel do the
5119 /// allocation. This produces generally horrible, but correct, code.
5121 /// OpInfo describes the operand.
5122 /// Input and OutputRegs are the set of already allocated physical registers.
5124 void SelectionDAGBuilder::
5125 GetRegistersForValue(SDISelAsmOperandInfo &OpInfo,
5126 std::set<unsigned> &OutputRegs,
5127 std::set<unsigned> &InputRegs) {
5128 LLVMContext &Context = FuncInfo.Fn->getContext();
5130 // Compute whether this value requires an input register, an output register,
5132 bool isOutReg = false;
5133 bool isInReg = false;
5134 switch (OpInfo.Type) {
5135 case InlineAsm::isOutput:
5138 // If there is an input constraint that matches this, we need to reserve
5139 // the input register so no other inputs allocate to it.
5140 isInReg = OpInfo.hasMatchingInput();
5142 case InlineAsm::isInput:
5146 case InlineAsm::isClobber:
5153 MachineFunction &MF = DAG.getMachineFunction();
5154 SmallVector<unsigned, 4> Regs;
5156 // If this is a constraint for a single physreg, or a constraint for a
5157 // register class, find it.
5158 std::pair<unsigned, const TargetRegisterClass*> PhysReg =
5159 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
5160 OpInfo.ConstraintVT);
5162 unsigned NumRegs = 1;
5163 if (OpInfo.ConstraintVT != MVT::Other) {
5164 // If this is a FP input in an integer register (or visa versa) insert a bit
5165 // cast of the input value. More generally, handle any case where the input
5166 // value disagrees with the register class we plan to stick this in.
5167 if (OpInfo.Type == InlineAsm::isInput &&
5168 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) {
5169 // Try to convert to the first EVT that the reg class contains. If the
5170 // types are identical size, use a bitcast to convert (e.g. two differing
5172 EVT RegVT = *PhysReg.second->vt_begin();
5173 if (RegVT.getSizeInBits() == OpInfo.ConstraintVT.getSizeInBits()) {
5174 OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
5175 RegVT, OpInfo.CallOperand);
5176 OpInfo.ConstraintVT = RegVT;
5177 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) {
5178 // If the input is a FP value and we want it in FP registers, do a
5179 // bitcast to the corresponding integer type. This turns an f64 value
5180 // into i64, which can be passed with two i32 values on a 32-bit
5182 RegVT = EVT::getIntegerVT(Context,
5183 OpInfo.ConstraintVT.getSizeInBits());
5184 OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
5185 RegVT, OpInfo.CallOperand);
5186 OpInfo.ConstraintVT = RegVT;
5190 NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT);
5194 EVT ValueVT = OpInfo.ConstraintVT;
5196 // If this is a constraint for a specific physical register, like {r17},
5198 if (unsigned AssignedReg = PhysReg.first) {
5199 const TargetRegisterClass *RC = PhysReg.second;
5200 if (OpInfo.ConstraintVT == MVT::Other)
5201 ValueVT = *RC->vt_begin();
5203 // Get the actual register value type. This is important, because the user
5204 // may have asked for (e.g.) the AX register in i32 type. We need to
5205 // remember that AX is actually i16 to get the right extension.
5206 RegVT = *RC->vt_begin();
5208 // This is a explicit reference to a physical register.
5209 Regs.push_back(AssignedReg);
5211 // If this is an expanded reference, add the rest of the regs to Regs.
5213 TargetRegisterClass::iterator I = RC->begin();
5214 for (; *I != AssignedReg; ++I)
5215 assert(I != RC->end() && "Didn't find reg!");
5217 // Already added the first reg.
5219 for (; NumRegs; --NumRegs, ++I) {
5220 assert(I != RC->end() && "Ran out of registers to allocate!");
5225 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5226 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
5227 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
5231 // Otherwise, if this was a reference to an LLVM register class, create vregs
5232 // for this reference.
5233 if (const TargetRegisterClass *RC = PhysReg.second) {
5234 RegVT = *RC->vt_begin();
5235 if (OpInfo.ConstraintVT == MVT::Other)
5238 // Create the appropriate number of virtual registers.
5239 MachineRegisterInfo &RegInfo = MF.getRegInfo();
5240 for (; NumRegs; --NumRegs)
5241 Regs.push_back(RegInfo.createVirtualRegister(RC));
5243 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5247 // This is a reference to a register class that doesn't directly correspond
5248 // to an LLVM register class. Allocate NumRegs consecutive, available,
5249 // registers from the class.
5250 std::vector<unsigned> RegClassRegs
5251 = TLI.getRegClassForInlineAsmConstraint(OpInfo.ConstraintCode,
5252 OpInfo.ConstraintVT);
5254 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
5255 unsigned NumAllocated = 0;
5256 for (unsigned i = 0, e = RegClassRegs.size(); i != e; ++i) {
5257 unsigned Reg = RegClassRegs[i];
5258 // See if this register is available.
5259 if ((isOutReg && OutputRegs.count(Reg)) || // Already used.
5260 (isInReg && InputRegs.count(Reg))) { // Already used.
5261 // Make sure we find consecutive registers.
5266 // Check to see if this register is allocatable (i.e. don't give out the
5268 const TargetRegisterClass *RC = isAllocatableRegister(Reg, MF, TLI, TRI);
5269 if (!RC) { // Couldn't allocate this register.
5270 // Reset NumAllocated to make sure we return consecutive registers.
5275 // Okay, this register is good, we can use it.
5278 // If we allocated enough consecutive registers, succeed.
5279 if (NumAllocated == NumRegs) {
5280 unsigned RegStart = (i-NumAllocated)+1;
5281 unsigned RegEnd = i+1;
5282 // Mark all of the allocated registers used.
5283 for (unsigned i = RegStart; i != RegEnd; ++i)
5284 Regs.push_back(RegClassRegs[i]);
5286 OpInfo.AssignedRegs = RegsForValue(Regs, *RC->vt_begin(),
5287 OpInfo.ConstraintVT);
5288 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
5293 // Otherwise, we couldn't allocate enough registers for this.
5296 /// visitInlineAsm - Handle a call to an InlineAsm object.
5298 void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) {
5299 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
5301 /// ConstraintOperands - Information about all of the constraints.
5302 std::vector<SDISelAsmOperandInfo> ConstraintOperands;
5304 std::set<unsigned> OutputRegs, InputRegs;
5306 // Do a prepass over the constraints, canonicalizing them, and building up the
5307 // ConstraintOperands list.
5308 std::vector<InlineAsm::ConstraintInfo>
5309 ConstraintInfos = IA->ParseConstraints();
5311 bool hasMemory = hasInlineAsmMemConstraint(ConstraintInfos, TLI);
5313 SDValue Chain, Flag;
5315 // We won't need to flush pending loads if this asm doesn't touch
5316 // memory and is nonvolatile.
5317 if (hasMemory || IA->hasSideEffects())
5320 Chain = DAG.getRoot();
5322 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
5323 unsigned ResNo = 0; // ResNo - The result number of the next output.
5324 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
5325 ConstraintOperands.push_back(SDISelAsmOperandInfo(ConstraintInfos[i]));
5326 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
5328 EVT OpVT = MVT::Other;
5330 // Compute the value type for each operand.
5331 switch (OpInfo.Type) {
5332 case InlineAsm::isOutput:
5333 // Indirect outputs just consume an argument.
5334 if (OpInfo.isIndirect) {
5335 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5339 // The return value of the call is this value. As such, there is no
5340 // corresponding argument.
5341 assert(!CS.getType()->isVoidTy() &&
5343 if (const StructType *STy = dyn_cast<StructType>(CS.getType())) {
5344 OpVT = TLI.getValueType(STy->getElementType(ResNo));
5346 assert(ResNo == 0 && "Asm only has one result!");
5347 OpVT = TLI.getValueType(CS.getType());
5351 case InlineAsm::isInput:
5352 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5354 case InlineAsm::isClobber:
5359 // If this is an input or an indirect output, process the call argument.
5360 // BasicBlocks are labels, currently appearing only in asm's.
5361 if (OpInfo.CallOperandVal) {
5362 // Strip bitcasts, if any. This mostly comes up for functions.
5363 OpInfo.CallOperandVal = OpInfo.CallOperandVal->stripPointerCasts();
5365 if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) {
5366 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]);
5368 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
5371 OpVT = OpInfo.getCallOperandValEVT(*DAG.getContext(), TLI, TD);
5374 OpInfo.ConstraintVT = OpVT;
5377 // Second pass over the constraints: compute which constraint option to use
5378 // and assign registers to constraints that want a specific physreg.
5379 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
5380 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5382 // If this is an output operand with a matching input operand, look up the
5383 // matching input. If their types mismatch, e.g. one is an integer, the
5384 // other is floating point, or their sizes are different, flag it as an
5386 if (OpInfo.hasMatchingInput()) {
5387 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
5389 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
5390 if ((OpInfo.ConstraintVT.isInteger() !=
5391 Input.ConstraintVT.isInteger()) ||
5392 (OpInfo.ConstraintVT.getSizeInBits() !=
5393 Input.ConstraintVT.getSizeInBits())) {
5394 report_fatal_error("Unsupported asm: input constraint"
5395 " with a matching output constraint of"
5396 " incompatible type!");
5398 Input.ConstraintVT = OpInfo.ConstraintVT;
5402 // Compute the constraint code and ConstraintType to use.
5403 TLI.ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, &DAG);
5405 // If this is a memory input, and if the operand is not indirect, do what we
5406 // need to to provide an address for the memory input.
5407 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5408 !OpInfo.isIndirect) {
5409 assert(OpInfo.Type == InlineAsm::isInput &&
5410 "Can only indirectify direct input operands!");
5412 // Memory operands really want the address of the value. If we don't have
5413 // an indirect input, put it in the constpool if we can, otherwise spill
5414 // it to a stack slot.
5416 // If the operand is a float, integer, or vector constant, spill to a
5417 // constant pool entry to get its address.
5418 const Value *OpVal = OpInfo.CallOperandVal;
5419 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
5420 isa<ConstantVector>(OpVal)) {
5421 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal),
5422 TLI.getPointerTy());
5424 // Otherwise, create a stack slot and emit a store to it before the
5426 const Type *Ty = OpVal->getType();
5427 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
5428 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(Ty);
5429 MachineFunction &MF = DAG.getMachineFunction();
5430 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
5431 SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
5432 Chain = DAG.getStore(Chain, getCurDebugLoc(),
5433 OpInfo.CallOperand, StackSlot, NULL, 0,
5435 OpInfo.CallOperand = StackSlot;
5438 // There is no longer a Value* corresponding to this operand.
5439 OpInfo.CallOperandVal = 0;
5441 // It is now an indirect operand.
5442 OpInfo.isIndirect = true;
5445 // If this constraint is for a specific register, allocate it before
5447 if (OpInfo.ConstraintType == TargetLowering::C_Register)
5448 GetRegistersForValue(OpInfo, OutputRegs, InputRegs);
5451 ConstraintInfos.clear();
5453 // Second pass - Loop over all of the operands, assigning virtual or physregs
5454 // to register class operands.
5455 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5456 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5458 // C_Register operands have already been allocated, Other/Memory don't need
5460 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
5461 GetRegistersForValue(OpInfo, OutputRegs, InputRegs);
5464 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
5465 std::vector<SDValue> AsmNodeOperands;
5466 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain
5467 AsmNodeOperands.push_back(
5468 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(),
5469 TLI.getPointerTy()));
5471 // If we have a !srcloc metadata node associated with it, we want to attach
5472 // this to the ultimately generated inline asm machineinstr. To do this, we
5473 // pass in the third operand as this (potentially null) inline asm MDNode.
5474 const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc");
5475 AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc));
5477 // Remember the AlignStack bit as operand 3.
5478 AsmNodeOperands.push_back(DAG.getTargetConstant(IA->isAlignStack() ? 1 : 0,
5481 // Loop over all of the inputs, copying the operand values into the
5482 // appropriate registers and processing the output regs.
5483 RegsForValue RetValRegs;
5485 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
5486 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
5488 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5489 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5491 switch (OpInfo.Type) {
5492 case InlineAsm::isOutput: {
5493 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
5494 OpInfo.ConstraintType != TargetLowering::C_Register) {
5495 // Memory output, or 'other' output (e.g. 'X' constraint).
5496 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
5498 // Add information to the INLINEASM node to know about this output.
5499 unsigned OpFlags = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
5500 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlags,
5501 TLI.getPointerTy()));
5502 AsmNodeOperands.push_back(OpInfo.CallOperand);
5506 // Otherwise, this is a register or register class output.
5508 // Copy the output from the appropriate register. Find a register that
5510 if (OpInfo.AssignedRegs.Regs.empty())
5511 report_fatal_error("Couldn't allocate output reg for constraint '" +
5512 Twine(OpInfo.ConstraintCode) + "'!");
5514 // If this is an indirect operand, store through the pointer after the
5516 if (OpInfo.isIndirect) {
5517 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
5518 OpInfo.CallOperandVal));
5520 // This is the result value of the call.
5521 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
5522 // Concatenate this output onto the outputs list.
5523 RetValRegs.append(OpInfo.AssignedRegs);
5526 // Add information to the INLINEASM node to know that this register is
5528 OpInfo.AssignedRegs.AddInlineAsmOperands(OpInfo.isEarlyClobber ?
5529 InlineAsm::Kind_RegDefEarlyClobber :
5530 InlineAsm::Kind_RegDef,
5537 case InlineAsm::isInput: {
5538 SDValue InOperandVal = OpInfo.CallOperand;
5540 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint?
5541 // If this is required to match an output register we have already set,
5542 // just use its register.
5543 unsigned OperandNo = OpInfo.getMatchedOperand();
5545 // Scan until we find the definition we already emitted of this operand.
5546 // When we find it, create a RegsForValue operand.
5547 unsigned CurOp = InlineAsm::Op_FirstOperand;
5548 for (; OperandNo; --OperandNo) {
5549 // Advance to the next operand.
5551 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
5552 assert((InlineAsm::isRegDefKind(OpFlag) ||
5553 InlineAsm::isRegDefEarlyClobberKind(OpFlag) ||
5554 InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?");
5555 CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1;
5559 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
5560 if (InlineAsm::isRegDefKind(OpFlag) ||
5561 InlineAsm::isRegDefEarlyClobberKind(OpFlag)) {
5562 // Add (OpFlag&0xffff)>>3 registers to MatchedRegs.
5563 if (OpInfo.isIndirect) {
5564 // This happens on gcc/testsuite/gcc.dg/pr8788-1.c
5565 LLVMContext &Ctx = *DAG.getContext();
5566 Ctx.emitError(CS.getInstruction(), "inline asm not supported yet:"
5567 " don't know how to handle tied "
5568 "indirect register inputs");
5571 RegsForValue MatchedRegs;
5572 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
5573 EVT RegVT = AsmNodeOperands[CurOp+1].getValueType();
5574 MatchedRegs.RegVTs.push_back(RegVT);
5575 MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo();
5576 for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag);
5578 MatchedRegs.Regs.push_back
5579 (RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT)));
5581 // Use the produced MatchedRegs object to
5582 MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
5584 MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse,
5585 true, OpInfo.getMatchedOperand(),
5586 DAG, AsmNodeOperands);
5590 assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!");
5591 assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 &&
5592 "Unexpected number of operands");
5593 // Add information to the INLINEASM node to know about this input.
5594 // See InlineAsm.h isUseOperandTiedToDef.
5595 OpFlag = InlineAsm::getFlagWordForMatchingOp(OpFlag,
5596 OpInfo.getMatchedOperand());
5597 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlag,
5598 TLI.getPointerTy()));
5599 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]);
5603 // Treat indirect 'X' constraint as memory.
5604 if (OpInfo.ConstraintType == TargetLowering::C_Other &&
5606 OpInfo.ConstraintType = TargetLowering::C_Memory;
5608 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
5609 std::vector<SDValue> Ops;
5610 TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode[0],
5613 report_fatal_error("Invalid operand for inline asm constraint '" +
5614 Twine(OpInfo.ConstraintCode) + "'!");
5616 // Add information to the INLINEASM node to know about this input.
5617 unsigned ResOpType =
5618 InlineAsm::getFlagWord(InlineAsm::Kind_Imm, Ops.size());
5619 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
5620 TLI.getPointerTy()));
5621 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
5625 if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
5626 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
5627 assert(InOperandVal.getValueType() == TLI.getPointerTy() &&
5628 "Memory operands expect pointer values");
5630 // Add information to the INLINEASM node to know about this input.
5631 unsigned ResOpType = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
5632 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
5633 TLI.getPointerTy()));
5634 AsmNodeOperands.push_back(InOperandVal);
5638 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
5639 OpInfo.ConstraintType == TargetLowering::C_Register) &&
5640 "Unknown constraint type!");
5641 assert(!OpInfo.isIndirect &&
5642 "Don't know how to handle indirect register inputs yet!");
5644 // Copy the input into the appropriate registers.
5645 if (OpInfo.AssignedRegs.Regs.empty() ||
5646 !OpInfo.AssignedRegs.areValueTypesLegal(TLI))
5647 report_fatal_error("Couldn't allocate input reg for constraint '" +
5648 Twine(OpInfo.ConstraintCode) + "'!");
5650 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
5653 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, false, 0,
5654 DAG, AsmNodeOperands);
5657 case InlineAsm::isClobber: {
5658 // Add the clobbered value to the operand list, so that the register
5659 // allocator is aware that the physreg got clobbered.
5660 if (!OpInfo.AssignedRegs.Regs.empty())
5661 OpInfo.AssignedRegs.AddInlineAsmOperands(
5662 InlineAsm::Kind_RegDefEarlyClobber,
5670 // Finish up input operands. Set the input chain and add the flag last.
5671 AsmNodeOperands[InlineAsm::Op_InputChain] = Chain;
5672 if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
5674 Chain = DAG.getNode(ISD::INLINEASM, getCurDebugLoc(),
5675 DAG.getVTList(MVT::Other, MVT::Flag),
5676 &AsmNodeOperands[0], AsmNodeOperands.size());
5677 Flag = Chain.getValue(1);
5679 // If this asm returns a register value, copy the result from that register
5680 // and set it as the value of the call.
5681 if (!RetValRegs.Regs.empty()) {
5682 SDValue Val = RetValRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(),
5685 // FIXME: Why don't we do this for inline asms with MRVs?
5686 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) {
5687 EVT ResultType = TLI.getValueType(CS.getType());
5689 // If any of the results of the inline asm is a vector, it may have the
5690 // wrong width/num elts. This can happen for register classes that can
5691 // contain multiple different value types. The preg or vreg allocated may
5692 // not have the same VT as was expected. Convert it to the right type
5693 // with bit_convert.
5694 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) {
5695 Val = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
5698 } else if (ResultType != Val.getValueType() &&
5699 ResultType.isInteger() && Val.getValueType().isInteger()) {
5700 // If a result value was tied to an input value, the computed result may
5701 // have a wider width than the expected result. Extract the relevant
5703 Val = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), ResultType, Val);
5706 assert(ResultType == Val.getValueType() && "Asm result value mismatch!");
5709 setValue(CS.getInstruction(), Val);
5710 // Don't need to use this as a chain in this case.
5711 if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty())
5715 std::vector<std::pair<SDValue, const Value *> > StoresToEmit;
5717 // Process indirect outputs, first output all of the flagged copies out of
5719 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
5720 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
5721 const Value *Ptr = IndirectStoresToEmit[i].second;
5722 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(),
5724 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
5727 // Emit the non-flagged stores from the physregs.
5728 SmallVector<SDValue, 8> OutChains;
5729 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) {
5730 SDValue Val = DAG.getStore(Chain, getCurDebugLoc(),
5731 StoresToEmit[i].first,
5732 getValue(StoresToEmit[i].second),
5733 StoresToEmit[i].second, 0,
5735 OutChains.push_back(Val);
5738 if (!OutChains.empty())
5739 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
5740 &OutChains[0], OutChains.size());
5745 void SelectionDAGBuilder::visitVAStart(const CallInst &I) {
5746 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurDebugLoc(),
5747 MVT::Other, getRoot(),
5748 getValue(I.getArgOperand(0)),
5749 DAG.getSrcValue(I.getArgOperand(0))));
5752 void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) {
5753 const TargetData &TD = *TLI.getTargetData();
5754 SDValue V = DAG.getVAArg(TLI.getValueType(I.getType()), getCurDebugLoc(),
5755 getRoot(), getValue(I.getOperand(0)),
5756 DAG.getSrcValue(I.getOperand(0)),
5757 TD.getABITypeAlignment(I.getType()));
5759 DAG.setRoot(V.getValue(1));
5762 void SelectionDAGBuilder::visitVAEnd(const CallInst &I) {
5763 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurDebugLoc(),
5764 MVT::Other, getRoot(),
5765 getValue(I.getArgOperand(0)),
5766 DAG.getSrcValue(I.getArgOperand(0))));
5769 void SelectionDAGBuilder::visitVACopy(const CallInst &I) {
5770 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurDebugLoc(),
5771 MVT::Other, getRoot(),
5772 getValue(I.getArgOperand(0)),
5773 getValue(I.getArgOperand(1)),
5774 DAG.getSrcValue(I.getArgOperand(0)),
5775 DAG.getSrcValue(I.getArgOperand(1))));
5778 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
5779 /// implementation, which just calls LowerCall.
5780 /// FIXME: When all targets are
5781 /// migrated to using LowerCall, this hook should be integrated into SDISel.
5782 std::pair<SDValue, SDValue>
5783 TargetLowering::LowerCallTo(SDValue Chain, const Type *RetTy,
5784 bool RetSExt, bool RetZExt, bool isVarArg,
5785 bool isInreg, unsigned NumFixedArgs,
5786 CallingConv::ID CallConv, bool isTailCall,
5787 bool isReturnValueUsed,
5789 ArgListTy &Args, SelectionDAG &DAG,
5790 DebugLoc dl) const {
5791 // Handle all of the outgoing arguments.
5792 SmallVector<ISD::OutputArg, 32> Outs;
5793 SmallVector<SDValue, 32> OutVals;
5794 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
5795 SmallVector<EVT, 4> ValueVTs;
5796 ComputeValueVTs(*this, Args[i].Ty, ValueVTs);
5797 for (unsigned Value = 0, NumValues = ValueVTs.size();
5798 Value != NumValues; ++Value) {
5799 EVT VT = ValueVTs[Value];
5800 const Type *ArgTy = VT.getTypeForEVT(RetTy->getContext());
5801 SDValue Op = SDValue(Args[i].Node.getNode(),
5802 Args[i].Node.getResNo() + Value);
5803 ISD::ArgFlagsTy Flags;
5804 unsigned OriginalAlignment =
5805 getTargetData()->getABITypeAlignment(ArgTy);
5811 if (Args[i].isInReg)
5815 if (Args[i].isByVal) {
5817 const PointerType *Ty = cast<PointerType>(Args[i].Ty);
5818 const Type *ElementTy = Ty->getElementType();
5819 unsigned FrameAlign = getByValTypeAlignment(ElementTy);
5820 unsigned FrameSize = getTargetData()->getTypeAllocSize(ElementTy);
5821 // For ByVal, alignment should come from FE. BE will guess if this
5822 // info is not there but there are cases it cannot get right.
5823 if (Args[i].Alignment)
5824 FrameAlign = Args[i].Alignment;
5825 Flags.setByValAlign(FrameAlign);
5826 Flags.setByValSize(FrameSize);
5830 Flags.setOrigAlign(OriginalAlignment);
5832 EVT PartVT = getRegisterType(RetTy->getContext(), VT);
5833 unsigned NumParts = getNumRegisters(RetTy->getContext(), VT);
5834 SmallVector<SDValue, 4> Parts(NumParts);
5835 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
5838 ExtendKind = ISD::SIGN_EXTEND;
5839 else if (Args[i].isZExt)
5840 ExtendKind = ISD::ZERO_EXTEND;
5842 getCopyToParts(DAG, dl, Op, &Parts[0], NumParts,
5843 PartVT, ExtendKind);
5845 for (unsigned j = 0; j != NumParts; ++j) {
5846 // if it isn't first piece, alignment must be 1
5847 ISD::OutputArg MyFlags(Flags, Parts[j].getValueType(),
5849 if (NumParts > 1 && j == 0)
5850 MyFlags.Flags.setSplit();
5852 MyFlags.Flags.setOrigAlign(1);
5854 Outs.push_back(MyFlags);
5855 OutVals.push_back(Parts[j]);
5860 // Handle the incoming return values from the call.
5861 SmallVector<ISD::InputArg, 32> Ins;
5862 SmallVector<EVT, 4> RetTys;
5863 ComputeValueVTs(*this, RetTy, RetTys);
5864 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
5866 EVT RegisterVT = getRegisterType(RetTy->getContext(), VT);
5867 unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT);
5868 for (unsigned i = 0; i != NumRegs; ++i) {
5869 ISD::InputArg MyFlags;
5870 MyFlags.VT = RegisterVT;
5871 MyFlags.Used = isReturnValueUsed;
5873 MyFlags.Flags.setSExt();
5875 MyFlags.Flags.setZExt();
5877 MyFlags.Flags.setInReg();
5878 Ins.push_back(MyFlags);
5882 SmallVector<SDValue, 4> InVals;
5883 Chain = LowerCall(Chain, Callee, CallConv, isVarArg, isTailCall,
5884 Outs, OutVals, Ins, dl, DAG, InVals);
5886 // Verify that the target's LowerCall behaved as expected.
5887 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
5888 "LowerCall didn't return a valid chain!");
5889 assert((!isTailCall || InVals.empty()) &&
5890 "LowerCall emitted a return value for a tail call!");
5891 assert((isTailCall || InVals.size() == Ins.size()) &&
5892 "LowerCall didn't emit the correct number of values!");
5894 // For a tail call, the return value is merely live-out and there aren't
5895 // any nodes in the DAG representing it. Return a special value to
5896 // indicate that a tail call has been emitted and no more Instructions
5897 // should be processed in the current block.
5900 return std::make_pair(SDValue(), SDValue());
5903 DEBUG(for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
5904 assert(InVals[i].getNode() &&
5905 "LowerCall emitted a null value!");
5906 assert(Ins[i].VT == InVals[i].getValueType() &&
5907 "LowerCall emitted a value with the wrong type!");
5910 // Collect the legal value parts into potentially illegal values
5911 // that correspond to the original function's return values.
5912 ISD::NodeType AssertOp = ISD::DELETED_NODE;
5914 AssertOp = ISD::AssertSext;
5916 AssertOp = ISD::AssertZext;
5917 SmallVector<SDValue, 4> ReturnValues;
5918 unsigned CurReg = 0;
5919 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
5921 EVT RegisterVT = getRegisterType(RetTy->getContext(), VT);
5922 unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT);
5924 ReturnValues.push_back(getCopyFromParts(DAG, dl, &InVals[CurReg],
5925 NumRegs, RegisterVT, VT,
5930 // For a function returning void, there is no return value. We can't create
5931 // such a node, so we just return a null return value in that case. In
5932 // that case, nothing will actualy look at the value.
5933 if (ReturnValues.empty())
5934 return std::make_pair(SDValue(), Chain);
5936 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, dl,
5937 DAG.getVTList(&RetTys[0], RetTys.size()),
5938 &ReturnValues[0], ReturnValues.size());
5939 return std::make_pair(Res, Chain);
5942 void TargetLowering::LowerOperationWrapper(SDNode *N,
5943 SmallVectorImpl<SDValue> &Results,
5944 SelectionDAG &DAG) const {
5945 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
5947 Results.push_back(Res);
5950 SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5951 llvm_unreachable("LowerOperation not implemented for this target!");
5956 SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) {
5957 SDValue Op = getNonRegisterValue(V);
5958 assert((Op.getOpcode() != ISD::CopyFromReg ||
5959 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
5960 "Copy from a reg to the same reg!");
5961 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg");
5963 RegsForValue RFV(V->getContext(), TLI, Reg, V->getType());
5964 SDValue Chain = DAG.getEntryNode();
5965 RFV.getCopyToRegs(Op, DAG, getCurDebugLoc(), Chain, 0);
5966 PendingExports.push_back(Chain);
5969 #include "llvm/CodeGen/SelectionDAGISel.h"
5971 void SelectionDAGISel::LowerArguments(const BasicBlock *LLVMBB) {
5972 // If this is the entry block, emit arguments.
5973 const Function &F = *LLVMBB->getParent();
5974 SelectionDAG &DAG = SDB->DAG;
5975 DebugLoc dl = SDB->getCurDebugLoc();
5976 const TargetData *TD = TLI.getTargetData();
5977 SmallVector<ISD::InputArg, 16> Ins;
5979 // Check whether the function can return without sret-demotion.
5980 SmallVector<ISD::OutputArg, 4> Outs;
5981 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
5984 if (!FuncInfo->CanLowerReturn) {
5985 // Put in an sret pointer parameter before all the other parameters.
5986 SmallVector<EVT, 1> ValueVTs;
5987 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
5989 // NOTE: Assuming that a pointer will never break down to more than one VT
5991 ISD::ArgFlagsTy Flags;
5993 EVT RegisterVT = TLI.getRegisterType(*DAG.getContext(), ValueVTs[0]);
5994 ISD::InputArg RetArg(Flags, RegisterVT, true);
5995 Ins.push_back(RetArg);
5998 // Set up the incoming argument description vector.
6000 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end();
6001 I != E; ++I, ++Idx) {
6002 SmallVector<EVT, 4> ValueVTs;
6003 ComputeValueVTs(TLI, I->getType(), ValueVTs);
6004 bool isArgValueUsed = !I->use_empty();
6005 for (unsigned Value = 0, NumValues = ValueVTs.size();
6006 Value != NumValues; ++Value) {
6007 EVT VT = ValueVTs[Value];
6008 const Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
6009 ISD::ArgFlagsTy Flags;
6010 unsigned OriginalAlignment =
6011 TD->getABITypeAlignment(ArgTy);
6013 if (F.paramHasAttr(Idx, Attribute::ZExt))
6015 if (F.paramHasAttr(Idx, Attribute::SExt))
6017 if (F.paramHasAttr(Idx, Attribute::InReg))
6019 if (F.paramHasAttr(Idx, Attribute::StructRet))
6021 if (F.paramHasAttr(Idx, Attribute::ByVal)) {
6023 const PointerType *Ty = cast<PointerType>(I->getType());
6024 const Type *ElementTy = Ty->getElementType();
6025 unsigned FrameAlign = TLI.getByValTypeAlignment(ElementTy);
6026 unsigned FrameSize = TD->getTypeAllocSize(ElementTy);
6027 // For ByVal, alignment should be passed from FE. BE will guess if
6028 // this info is not there but there are cases it cannot get right.
6029 if (F.getParamAlignment(Idx))
6030 FrameAlign = F.getParamAlignment(Idx);
6031 Flags.setByValAlign(FrameAlign);
6032 Flags.setByValSize(FrameSize);
6034 if (F.paramHasAttr(Idx, Attribute::Nest))
6036 Flags.setOrigAlign(OriginalAlignment);
6038 EVT RegisterVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6039 unsigned NumRegs = TLI.getNumRegisters(*CurDAG->getContext(), VT);
6040 for (unsigned i = 0; i != NumRegs; ++i) {
6041 ISD::InputArg MyFlags(Flags, RegisterVT, isArgValueUsed);
6042 if (NumRegs > 1 && i == 0)
6043 MyFlags.Flags.setSplit();
6044 // if it isn't first piece, alignment must be 1
6046 MyFlags.Flags.setOrigAlign(1);
6047 Ins.push_back(MyFlags);
6052 // Call the target to set up the argument values.
6053 SmallVector<SDValue, 8> InVals;
6054 SDValue NewRoot = TLI.LowerFormalArguments(DAG.getRoot(), F.getCallingConv(),
6058 // Verify that the target's LowerFormalArguments behaved as expected.
6059 assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other &&
6060 "LowerFormalArguments didn't return a valid chain!");
6061 assert(InVals.size() == Ins.size() &&
6062 "LowerFormalArguments didn't emit the correct number of values!");
6064 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
6065 assert(InVals[i].getNode() &&
6066 "LowerFormalArguments emitted a null value!");
6067 assert(Ins[i].VT == InVals[i].getValueType() &&
6068 "LowerFormalArguments emitted a value with the wrong type!");
6072 // Update the DAG with the new chain value resulting from argument lowering.
6073 DAG.setRoot(NewRoot);
6075 // Set up the argument values.
6078 if (!FuncInfo->CanLowerReturn) {
6079 // Create a virtual register for the sret pointer, and put in a copy
6080 // from the sret argument into it.
6081 SmallVector<EVT, 1> ValueVTs;
6082 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
6083 EVT VT = ValueVTs[0];
6084 EVT RegVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6085 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6086 SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1,
6087 RegVT, VT, AssertOp);
6089 MachineFunction& MF = SDB->DAG.getMachineFunction();
6090 MachineRegisterInfo& RegInfo = MF.getRegInfo();
6091 unsigned SRetReg = RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT));
6092 FuncInfo->DemoteRegister = SRetReg;
6093 NewRoot = SDB->DAG.getCopyToReg(NewRoot, SDB->getCurDebugLoc(),
6095 DAG.setRoot(NewRoot);
6097 // i indexes lowered arguments. Bump it past the hidden sret argument.
6098 // Idx indexes LLVM arguments. Don't touch it.
6102 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
6104 SmallVector<SDValue, 4> ArgValues;
6105 SmallVector<EVT, 4> ValueVTs;
6106 ComputeValueVTs(TLI, I->getType(), ValueVTs);
6107 unsigned NumValues = ValueVTs.size();
6109 // If this argument is unused then remember its value. It is used to generate
6110 // debugging information.
6111 if (I->use_empty() && NumValues)
6112 SDB->setUnusedArgValue(I, InVals[i]);
6114 for (unsigned Value = 0; Value != NumValues; ++Value) {
6115 EVT VT = ValueVTs[Value];
6116 EVT PartVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6117 unsigned NumParts = TLI.getNumRegisters(*CurDAG->getContext(), VT);
6119 if (!I->use_empty()) {
6120 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6121 if (F.paramHasAttr(Idx, Attribute::SExt))
6122 AssertOp = ISD::AssertSext;
6123 else if (F.paramHasAttr(Idx, Attribute::ZExt))
6124 AssertOp = ISD::AssertZext;
6126 ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i],
6127 NumParts, PartVT, VT,
6134 if (!I->use_empty()) {
6136 if (!ArgValues.empty())
6137 Res = DAG.getMergeValues(&ArgValues[0], NumValues,
6138 SDB->getCurDebugLoc());
6139 SDB->setValue(I, Res);
6141 // If this argument is live outside of the entry block, insert a copy from
6142 // whereever we got it to the vreg that other BB's will reference it as.
6143 SDB->CopyToExportRegsIfNeeded(I);
6147 assert(i == InVals.size() && "Argument register count mismatch!");
6149 // Finally, if the target has anything special to do, allow it to do so.
6150 // FIXME: this should insert code into the DAG!
6151 EmitFunctionEntryCode();
6154 /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
6155 /// ensure constants are generated when needed. Remember the virtual registers
6156 /// that need to be added to the Machine PHI nodes as input. We cannot just
6157 /// directly add them, because expansion might result in multiple MBB's for one
6158 /// BB. As such, the start of the BB might correspond to a different MBB than
6162 SelectionDAGBuilder::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
6163 const TerminatorInst *TI = LLVMBB->getTerminator();
6165 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
6167 // Check successor nodes' PHI nodes that expect a constant to be available
6169 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
6170 const BasicBlock *SuccBB = TI->getSuccessor(succ);
6171 if (!isa<PHINode>(SuccBB->begin())) continue;
6172 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
6174 // If this terminator has multiple identical successors (common for
6175 // switches), only handle each succ once.
6176 if (!SuccsHandled.insert(SuccMBB)) continue;
6178 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
6180 // At this point we know that there is a 1-1 correspondence between LLVM PHI
6181 // nodes and Machine PHI nodes, but the incoming operands have not been
6183 for (BasicBlock::const_iterator I = SuccBB->begin();
6184 const PHINode *PN = dyn_cast<PHINode>(I); ++I) {
6185 // Ignore dead phi's.
6186 if (PN->use_empty()) continue;
6189 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
6191 if (const Constant *C = dyn_cast<Constant>(PHIOp)) {
6192 unsigned &RegOut = ConstantsOut[C];
6194 RegOut = FuncInfo.CreateRegs(C->getType());
6195 CopyValueToVirtualRegister(C, RegOut);
6199 DenseMap<const Value *, unsigned>::iterator I =
6200 FuncInfo.ValueMap.find(PHIOp);
6201 if (I != FuncInfo.ValueMap.end())
6204 assert(isa<AllocaInst>(PHIOp) &&
6205 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
6206 "Didn't codegen value into a register!??");
6207 Reg = FuncInfo.CreateRegs(PHIOp->getType());
6208 CopyValueToVirtualRegister(PHIOp, Reg);
6212 // Remember that this register needs to added to the machine PHI node as
6213 // the input for this MBB.
6214 SmallVector<EVT, 4> ValueVTs;
6215 ComputeValueVTs(TLI, PN->getType(), ValueVTs);
6216 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
6217 EVT VT = ValueVTs[vti];
6218 unsigned NumRegisters = TLI.getNumRegisters(*DAG.getContext(), VT);
6219 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
6220 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
6221 Reg += NumRegisters;
6225 ConstantsOut.clear();