1 //===-- SelectionDAGBuilder.cpp - Selection-DAG building ------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements routines for translating from LLVM IR into SelectionDAG IR.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "isel"
15 #include "SDNodeDbgValue.h"
16 #include "SelectionDAGBuilder.h"
17 #include "llvm/ADT/BitVector.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/Analysis/AliasAnalysis.h"
20 #include "llvm/Analysis/ConstantFolding.h"
21 #include "llvm/Constants.h"
22 #include "llvm/CallingConv.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/Function.h"
25 #include "llvm/GlobalVariable.h"
26 #include "llvm/InlineAsm.h"
27 #include "llvm/Instructions.h"
28 #include "llvm/Intrinsics.h"
29 #include "llvm/IntrinsicInst.h"
30 #include "llvm/LLVMContext.h"
31 #include "llvm/Module.h"
32 #include "llvm/CodeGen/Analysis.h"
33 #include "llvm/CodeGen/FastISel.h"
34 #include "llvm/CodeGen/FunctionLoweringInfo.h"
35 #include "llvm/CodeGen/GCStrategy.h"
36 #include "llvm/CodeGen/GCMetadata.h"
37 #include "llvm/CodeGen/MachineFunction.h"
38 #include "llvm/CodeGen/MachineFrameInfo.h"
39 #include "llvm/CodeGen/MachineInstrBuilder.h"
40 #include "llvm/CodeGen/MachineJumpTableInfo.h"
41 #include "llvm/CodeGen/MachineModuleInfo.h"
42 #include "llvm/CodeGen/MachineRegisterInfo.h"
43 #include "llvm/CodeGen/PseudoSourceValue.h"
44 #include "llvm/CodeGen/SelectionDAG.h"
45 #include "llvm/Analysis/DebugInfo.h"
46 #include "llvm/Target/TargetRegisterInfo.h"
47 #include "llvm/Target/TargetData.h"
48 #include "llvm/Target/TargetFrameInfo.h"
49 #include "llvm/Target/TargetInstrInfo.h"
50 #include "llvm/Target/TargetIntrinsicInfo.h"
51 #include "llvm/Target/TargetLowering.h"
52 #include "llvm/Target/TargetOptions.h"
53 #include "llvm/Support/Compiler.h"
54 #include "llvm/Support/CommandLine.h"
55 #include "llvm/Support/Debug.h"
56 #include "llvm/Support/ErrorHandling.h"
57 #include "llvm/Support/MathExtras.h"
58 #include "llvm/Support/raw_ostream.h"
62 /// LimitFloatPrecision - Generate low-precision inline sequences for
63 /// some float libcalls (6, 8 or 12 bits).
64 static unsigned LimitFloatPrecision;
66 static cl::opt<unsigned, true>
67 LimitFPPrecision("limit-float-precision",
68 cl::desc("Generate low-precision inline sequences "
69 "for some float libcalls"),
70 cl::location(LimitFloatPrecision),
73 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, DebugLoc DL,
74 const SDValue *Parts, unsigned NumParts,
75 EVT PartVT, EVT ValueVT);
77 /// getCopyFromParts - Create a value that contains the specified legal parts
78 /// combined into the value they represent. If the parts combine to a type
79 /// larger then ValueVT then AssertOp can be used to specify whether the extra
80 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
81 /// (ISD::AssertSext).
82 static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc DL,
84 unsigned NumParts, EVT PartVT, EVT ValueVT,
85 ISD::NodeType AssertOp = ISD::DELETED_NODE) {
86 if (ValueVT.isVector())
87 return getCopyFromPartsVector(DAG, DL, Parts, NumParts, PartVT, ValueVT);
89 assert(NumParts > 0 && "No parts to assemble!");
90 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
91 SDValue Val = Parts[0];
94 // Assemble the value from multiple parts.
95 if (ValueVT.isInteger()) {
96 unsigned PartBits = PartVT.getSizeInBits();
97 unsigned ValueBits = ValueVT.getSizeInBits();
99 // Assemble the power of 2 part.
100 unsigned RoundParts = NumParts & (NumParts - 1) ?
101 1 << Log2_32(NumParts) : NumParts;
102 unsigned RoundBits = PartBits * RoundParts;
103 EVT RoundVT = RoundBits == ValueBits ?
104 ValueVT : EVT::getIntegerVT(*DAG.getContext(), RoundBits);
107 EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2);
109 if (RoundParts > 2) {
110 Lo = getCopyFromParts(DAG, DL, Parts, RoundParts / 2,
112 Hi = getCopyFromParts(DAG, DL, Parts + RoundParts / 2,
113 RoundParts / 2, PartVT, HalfVT);
115 Lo = DAG.getNode(ISD::BIT_CONVERT, DL, HalfVT, Parts[0]);
116 Hi = DAG.getNode(ISD::BIT_CONVERT, DL, HalfVT, Parts[1]);
119 if (TLI.isBigEndian())
122 Val = DAG.getNode(ISD::BUILD_PAIR, DL, RoundVT, Lo, Hi);
124 if (RoundParts < NumParts) {
125 // Assemble the trailing non-power-of-2 part.
126 unsigned OddParts = NumParts - RoundParts;
127 EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits);
128 Hi = getCopyFromParts(DAG, DL,
129 Parts + RoundParts, OddParts, PartVT, OddVT);
131 // Combine the round and odd parts.
133 if (TLI.isBigEndian())
135 EVT TotalVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
136 Hi = DAG.getNode(ISD::ANY_EXTEND, DL, TotalVT, Hi);
137 Hi = DAG.getNode(ISD::SHL, DL, TotalVT, Hi,
138 DAG.getConstant(Lo.getValueType().getSizeInBits(),
139 TLI.getPointerTy()));
140 Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, TotalVT, Lo);
141 Val = DAG.getNode(ISD::OR, DL, TotalVT, Lo, Hi);
143 } else if (PartVT.isFloatingPoint()) {
144 // FP split into multiple FP parts (for ppcf128)
145 assert(ValueVT == EVT(MVT::ppcf128) && PartVT == EVT(MVT::f64) &&
148 Lo = DAG.getNode(ISD::BIT_CONVERT, DL, EVT(MVT::f64), Parts[0]);
149 Hi = DAG.getNode(ISD::BIT_CONVERT, DL, EVT(MVT::f64), Parts[1]);
150 if (TLI.isBigEndian())
152 Val = DAG.getNode(ISD::BUILD_PAIR, DL, ValueVT, Lo, Hi);
154 // FP split into integer parts (soft fp)
155 assert(ValueVT.isFloatingPoint() && PartVT.isInteger() &&
156 !PartVT.isVector() && "Unexpected split");
157 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits());
158 Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT);
162 // There is now one part, held in Val. Correct it to match ValueVT.
163 PartVT = Val.getValueType();
165 if (PartVT == ValueVT)
168 if (PartVT.isInteger() && ValueVT.isInteger()) {
169 if (ValueVT.bitsLT(PartVT)) {
170 // For a truncate, see if we have any information to
171 // indicate whether the truncated bits will always be
172 // zero or sign-extension.
173 if (AssertOp != ISD::DELETED_NODE)
174 Val = DAG.getNode(AssertOp, DL, PartVT, Val,
175 DAG.getValueType(ValueVT));
176 return DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
178 return DAG.getNode(ISD::ANY_EXTEND, DL, ValueVT, Val);
181 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
182 // FP_ROUND's are always exact here.
183 if (ValueVT.bitsLT(Val.getValueType()))
184 return DAG.getNode(ISD::FP_ROUND, DL, ValueVT, Val,
185 DAG.getIntPtrConstant(1));
187 return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val);
190 if (PartVT.getSizeInBits() == ValueVT.getSizeInBits())
191 return DAG.getNode(ISD::BIT_CONVERT, DL, ValueVT, Val);
193 llvm_unreachable("Unknown mismatch!");
197 /// getCopyFromParts - Create a value that contains the specified legal parts
198 /// combined into the value they represent. If the parts combine to a type
199 /// larger then ValueVT then AssertOp can be used to specify whether the extra
200 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
201 /// (ISD::AssertSext).
202 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, DebugLoc DL,
203 const SDValue *Parts, unsigned NumParts,
204 EVT PartVT, EVT ValueVT) {
205 assert(ValueVT.isVector() && "Not a vector value");
206 assert(NumParts > 0 && "No parts to assemble!");
207 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
208 SDValue Val = Parts[0];
210 // Handle a multi-element vector.
212 EVT IntermediateVT, RegisterVT;
213 unsigned NumIntermediates;
215 TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT,
216 NumIntermediates, RegisterVT);
217 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
218 NumParts = NumRegs; // Silence a compiler warning.
219 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
220 assert(RegisterVT == Parts[0].getValueType() &&
221 "Part type doesn't match part!");
223 // Assemble the parts into intermediate operands.
224 SmallVector<SDValue, 8> Ops(NumIntermediates);
225 if (NumIntermediates == NumParts) {
226 // If the register was not expanded, truncate or copy the value,
228 for (unsigned i = 0; i != NumParts; ++i)
229 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i], 1,
230 PartVT, IntermediateVT);
231 } else if (NumParts > 0) {
232 // If the intermediate type was expanded, build the intermediate
233 // operands from the parts.
234 assert(NumParts % NumIntermediates == 0 &&
235 "Must expand into a divisible number of parts!");
236 unsigned Factor = NumParts / NumIntermediates;
237 for (unsigned i = 0; i != NumIntermediates; ++i)
238 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor,
239 PartVT, IntermediateVT);
242 // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the
243 // intermediate operands.
244 Val = DAG.getNode(IntermediateVT.isVector() ?
245 ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR, DL,
246 ValueVT, &Ops[0], NumIntermediates);
249 // There is now one part, held in Val. Correct it to match ValueVT.
250 PartVT = Val.getValueType();
252 if (PartVT == ValueVT)
255 if (PartVT.isVector()) {
256 // If the element type of the source/dest vectors are the same, but the
257 // parts vector has more elements than the value vector, then we have a
258 // vector widening case (e.g. <2 x float> -> <4 x float>). Extract the
260 if (PartVT.getVectorElementType() == ValueVT.getVectorElementType()) {
261 assert(PartVT.getVectorNumElements() > ValueVT.getVectorNumElements() &&
262 "Cannot narrow, it would be a lossy transformation");
263 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val,
264 DAG.getIntPtrConstant(0));
267 // Vector/Vector bitcast.
268 return DAG.getNode(ISD::BIT_CONVERT, DL, ValueVT, Val);
271 assert(ValueVT.getVectorElementType() == PartVT &&
272 ValueVT.getVectorNumElements() == 1 &&
273 "Only trivial scalar-to-vector conversions should get here!");
274 return DAG.getNode(ISD::BUILD_VECTOR, DL, ValueVT, Val);
280 static void getCopyToPartsVector(SelectionDAG &DAG, DebugLoc dl,
281 SDValue Val, SDValue *Parts, unsigned NumParts,
284 /// getCopyToParts - Create a series of nodes that contain the specified value
285 /// split into legal parts. If the parts contain more bits than Val, then, for
286 /// integers, ExtendKind can be used to specify how to generate the extra bits.
287 static void getCopyToParts(SelectionDAG &DAG, DebugLoc DL,
288 SDValue Val, SDValue *Parts, unsigned NumParts,
290 ISD::NodeType ExtendKind = ISD::ANY_EXTEND) {
291 EVT ValueVT = Val.getValueType();
293 // Handle the vector case separately.
294 if (ValueVT.isVector())
295 return getCopyToPartsVector(DAG, DL, Val, Parts, NumParts, PartVT);
297 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
298 unsigned PartBits = PartVT.getSizeInBits();
299 unsigned OrigNumParts = NumParts;
300 assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!");
305 assert(!ValueVT.isVector() && "Vector case handled elsewhere");
306 if (PartVT == ValueVT) {
307 assert(NumParts == 1 && "No-op copy with multiple parts!");
312 if (NumParts * PartBits > ValueVT.getSizeInBits()) {
313 // If the parts cover more bits than the value has, promote the value.
314 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
315 assert(NumParts == 1 && "Do not know what to promote to!");
316 Val = DAG.getNode(ISD::FP_EXTEND, DL, PartVT, Val);
318 assert(PartVT.isInteger() && ValueVT.isInteger() &&
319 "Unknown mismatch!");
320 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
321 Val = DAG.getNode(ExtendKind, DL, ValueVT, Val);
323 } else if (PartBits == ValueVT.getSizeInBits()) {
324 // Different types of the same size.
325 assert(NumParts == 1 && PartVT != ValueVT);
326 Val = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Val);
327 } else if (NumParts * PartBits < ValueVT.getSizeInBits()) {
328 // If the parts cover less bits than value has, truncate the value.
329 assert(PartVT.isInteger() && ValueVT.isInteger() &&
330 "Unknown mismatch!");
331 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
332 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
335 // The value may have changed - recompute ValueVT.
336 ValueVT = Val.getValueType();
337 assert(NumParts * PartBits == ValueVT.getSizeInBits() &&
338 "Failed to tile the value with PartVT!");
341 assert(PartVT == ValueVT && "Type conversion failed!");
346 // Expand the value into multiple parts.
347 if (NumParts & (NumParts - 1)) {
348 // The number of parts is not a power of 2. Split off and copy the tail.
349 assert(PartVT.isInteger() && ValueVT.isInteger() &&
350 "Do not know what to expand to!");
351 unsigned RoundParts = 1 << Log2_32(NumParts);
352 unsigned RoundBits = RoundParts * PartBits;
353 unsigned OddParts = NumParts - RoundParts;
354 SDValue OddVal = DAG.getNode(ISD::SRL, DL, ValueVT, Val,
355 DAG.getIntPtrConstant(RoundBits));
356 getCopyToParts(DAG, DL, OddVal, Parts + RoundParts, OddParts, PartVT);
358 if (TLI.isBigEndian())
359 // The odd parts were reversed by getCopyToParts - unreverse them.
360 std::reverse(Parts + RoundParts, Parts + NumParts);
362 NumParts = RoundParts;
363 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
364 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
367 // The number of parts is a power of 2. Repeatedly bisect the value using
369 Parts[0] = DAG.getNode(ISD::BIT_CONVERT, DL,
370 EVT::getIntegerVT(*DAG.getContext(),
371 ValueVT.getSizeInBits()),
374 for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) {
375 for (unsigned i = 0; i < NumParts; i += StepSize) {
376 unsigned ThisBits = StepSize * PartBits / 2;
377 EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits);
378 SDValue &Part0 = Parts[i];
379 SDValue &Part1 = Parts[i+StepSize/2];
381 Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
382 ThisVT, Part0, DAG.getIntPtrConstant(1));
383 Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
384 ThisVT, Part0, DAG.getIntPtrConstant(0));
386 if (ThisBits == PartBits && ThisVT != PartVT) {
387 Part0 = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Part0);
388 Part1 = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Part1);
393 if (TLI.isBigEndian())
394 std::reverse(Parts, Parts + OrigNumParts);
398 /// getCopyToPartsVector - Create a series of nodes that contain the specified
399 /// value split into legal parts.
400 static void getCopyToPartsVector(SelectionDAG &DAG, DebugLoc DL,
401 SDValue Val, SDValue *Parts, unsigned NumParts,
403 EVT ValueVT = Val.getValueType();
404 assert(ValueVT.isVector() && "Not a vector");
405 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
408 if (PartVT == ValueVT) {
410 } else if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) {
411 // Bitconvert vector->vector case.
412 Val = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Val);
413 } else if (PartVT.isVector() &&
414 PartVT.getVectorElementType() == ValueVT.getVectorElementType()&&
415 PartVT.getVectorNumElements() > ValueVT.getVectorNumElements()) {
416 EVT ElementVT = PartVT.getVectorElementType();
417 // Vector widening case, e.g. <2 x float> -> <4 x float>. Shuffle in
419 SmallVector<SDValue, 16> Ops;
420 for (unsigned i = 0, e = ValueVT.getVectorNumElements(); i != e; ++i)
421 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
422 ElementVT, Val, DAG.getIntPtrConstant(i)));
424 for (unsigned i = ValueVT.getVectorNumElements(),
425 e = PartVT.getVectorNumElements(); i != e; ++i)
426 Ops.push_back(DAG.getUNDEF(ElementVT));
428 Val = DAG.getNode(ISD::BUILD_VECTOR, DL, PartVT, &Ops[0], Ops.size());
430 // FIXME: Use CONCAT for 2x -> 4x.
432 //SDValue UndefElts = DAG.getUNDEF(VectorTy);
433 //Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts);
435 // Vector -> scalar conversion.
436 assert(ValueVT.getVectorElementType() == PartVT &&
437 ValueVT.getVectorNumElements() == 1 &&
438 "Only trivial vector-to-scalar conversions should get here!");
439 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
440 PartVT, Val, DAG.getIntPtrConstant(0));
447 // Handle a multi-element vector.
448 EVT IntermediateVT, RegisterVT;
449 unsigned NumIntermediates;
450 unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT,
452 NumIntermediates, RegisterVT);
453 unsigned NumElements = ValueVT.getVectorNumElements();
455 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
456 NumParts = NumRegs; // Silence a compiler warning.
457 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
459 // Split the vector into intermediate operands.
460 SmallVector<SDValue, 8> Ops(NumIntermediates);
461 for (unsigned i = 0; i != NumIntermediates; ++i) {
462 if (IntermediateVT.isVector())
463 Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL,
465 DAG.getIntPtrConstant(i * (NumElements / NumIntermediates)));
467 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
468 IntermediateVT, Val, DAG.getIntPtrConstant(i));
471 // Split the intermediate operands into legal parts.
472 if (NumParts == NumIntermediates) {
473 // If the register was not expanded, promote or copy the value,
475 for (unsigned i = 0; i != NumParts; ++i)
476 getCopyToParts(DAG, DL, Ops[i], &Parts[i], 1, PartVT);
477 } else if (NumParts > 0) {
478 // If the intermediate type was expanded, split each the value into
480 assert(NumParts % NumIntermediates == 0 &&
481 "Must expand into a divisible number of parts!");
482 unsigned Factor = NumParts / NumIntermediates;
483 for (unsigned i = 0; i != NumIntermediates; ++i)
484 getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT);
492 /// RegsForValue - This struct represents the registers (physical or virtual)
493 /// that a particular set of values is assigned, and the type information
494 /// about the value. The most common situation is to represent one value at a
495 /// time, but struct or array values are handled element-wise as multiple
496 /// values. The splitting of aggregates is performed recursively, so that we
497 /// never have aggregate-typed registers. The values at this point do not
498 /// necessarily have legal types, so each value may require one or more
499 /// registers of some legal type.
501 struct RegsForValue {
502 /// ValueVTs - The value types of the values, which may not be legal, and
503 /// may need be promoted or synthesized from one or more registers.
505 SmallVector<EVT, 4> ValueVTs;
507 /// RegVTs - The value types of the registers. This is the same size as
508 /// ValueVTs and it records, for each value, what the type of the assigned
509 /// register or registers are. (Individual values are never synthesized
510 /// from more than one type of register.)
512 /// With virtual registers, the contents of RegVTs is redundant with TLI's
513 /// getRegisterType member function, however when with physical registers
514 /// it is necessary to have a separate record of the types.
516 SmallVector<EVT, 4> RegVTs;
518 /// Regs - This list holds the registers assigned to the values.
519 /// Each legal or promoted value requires one register, and each
520 /// expanded value requires multiple registers.
522 SmallVector<unsigned, 4> Regs;
526 RegsForValue(const SmallVector<unsigned, 4> ®s,
527 EVT regvt, EVT valuevt)
528 : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {}
530 RegsForValue(LLVMContext &Context, const TargetLowering &tli,
531 unsigned Reg, const Type *Ty) {
532 ComputeValueVTs(tli, Ty, ValueVTs);
534 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
535 EVT ValueVT = ValueVTs[Value];
536 unsigned NumRegs = tli.getNumRegisters(Context, ValueVT);
537 EVT RegisterVT = tli.getRegisterType(Context, ValueVT);
538 for (unsigned i = 0; i != NumRegs; ++i)
539 Regs.push_back(Reg + i);
540 RegVTs.push_back(RegisterVT);
545 /// areValueTypesLegal - Return true if types of all the values are legal.
546 bool areValueTypesLegal(const TargetLowering &TLI) {
547 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
548 EVT RegisterVT = RegVTs[Value];
549 if (!TLI.isTypeLegal(RegisterVT))
555 /// append - Add the specified values to this one.
556 void append(const RegsForValue &RHS) {
557 ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end());
558 RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end());
559 Regs.append(RHS.Regs.begin(), RHS.Regs.end());
562 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
563 /// this value and returns the result as a ValueVTs value. This uses
564 /// Chain/Flag as the input and updates them for the output Chain/Flag.
565 /// If the Flag pointer is NULL, no flag is used.
566 SDValue getCopyFromRegs(SelectionDAG &DAG, FunctionLoweringInfo &FuncInfo,
568 SDValue &Chain, SDValue *Flag) const;
570 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
571 /// specified value into the registers specified by this object. This uses
572 /// Chain/Flag as the input and updates them for the output Chain/Flag.
573 /// If the Flag pointer is NULL, no flag is used.
574 void getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
575 SDValue &Chain, SDValue *Flag) const;
577 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
578 /// operand list. This adds the code marker, matching input operand index
579 /// (if applicable), and includes the number of values added into it.
580 void AddInlineAsmOperands(unsigned Kind,
581 bool HasMatching, unsigned MatchingIdx,
583 std::vector<SDValue> &Ops) const;
587 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
588 /// this value and returns the result as a ValueVT value. This uses
589 /// Chain/Flag as the input and updates them for the output Chain/Flag.
590 /// If the Flag pointer is NULL, no flag is used.
591 SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
592 FunctionLoweringInfo &FuncInfo,
594 SDValue &Chain, SDValue *Flag) const {
595 // A Value with type {} or [0 x %t] needs no registers.
596 if (ValueVTs.empty())
599 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
601 // Assemble the legal parts into the final values.
602 SmallVector<SDValue, 4> Values(ValueVTs.size());
603 SmallVector<SDValue, 8> Parts;
604 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
605 // Copy the legal parts from the registers.
606 EVT ValueVT = ValueVTs[Value];
607 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
608 EVT RegisterVT = RegVTs[Value];
610 Parts.resize(NumRegs);
611 for (unsigned i = 0; i != NumRegs; ++i) {
614 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT);
616 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag);
617 *Flag = P.getValue(2);
620 Chain = P.getValue(1);
622 // If the source register was virtual and if we know something about it,
623 // add an assert node.
624 if (TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) &&
625 RegisterVT.isInteger() && !RegisterVT.isVector()) {
626 unsigned SlotNo = Regs[Part+i]-TargetRegisterInfo::FirstVirtualRegister;
627 if (FuncInfo.LiveOutRegInfo.size() > SlotNo) {
628 const FunctionLoweringInfo::LiveOutInfo &LOI =
629 FuncInfo.LiveOutRegInfo[SlotNo];
631 unsigned RegSize = RegisterVT.getSizeInBits();
632 unsigned NumSignBits = LOI.NumSignBits;
633 unsigned NumZeroBits = LOI.KnownZero.countLeadingOnes();
635 // FIXME: We capture more information than the dag can represent. For
636 // now, just use the tightest assertzext/assertsext possible.
638 EVT FromVT(MVT::Other);
639 if (NumSignBits == RegSize)
640 isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1
641 else if (NumZeroBits >= RegSize-1)
642 isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1
643 else if (NumSignBits > RegSize-8)
644 isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8
645 else if (NumZeroBits >= RegSize-8)
646 isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8
647 else if (NumSignBits > RegSize-16)
648 isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16
649 else if (NumZeroBits >= RegSize-16)
650 isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16
651 else if (NumSignBits > RegSize-32)
652 isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32
653 else if (NumZeroBits >= RegSize-32)
654 isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32
656 if (FromVT != MVT::Other)
657 P = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl,
658 RegisterVT, P, DAG.getValueType(FromVT));
665 Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(),
666 NumRegs, RegisterVT, ValueVT);
671 return DAG.getNode(ISD::MERGE_VALUES, dl,
672 DAG.getVTList(&ValueVTs[0], ValueVTs.size()),
673 &Values[0], ValueVTs.size());
676 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
677 /// specified value into the registers specified by this object. This uses
678 /// Chain/Flag as the input and updates them for the output Chain/Flag.
679 /// If the Flag pointer is NULL, no flag is used.
680 void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
681 SDValue &Chain, SDValue *Flag) const {
682 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
684 // Get the list of the values's legal parts.
685 unsigned NumRegs = Regs.size();
686 SmallVector<SDValue, 8> Parts(NumRegs);
687 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
688 EVT ValueVT = ValueVTs[Value];
689 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
690 EVT RegisterVT = RegVTs[Value];
692 getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value),
693 &Parts[Part], NumParts, RegisterVT);
697 // Copy the parts into the registers.
698 SmallVector<SDValue, 8> Chains(NumRegs);
699 for (unsigned i = 0; i != NumRegs; ++i) {
702 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]);
704 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag);
705 *Flag = Part.getValue(1);
708 Chains[i] = Part.getValue(0);
711 if (NumRegs == 1 || Flag)
712 // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is
713 // flagged to it. That is the CopyToReg nodes and the user are considered
714 // a single scheduling unit. If we create a TokenFactor and return it as
715 // chain, then the TokenFactor is both a predecessor (operand) of the
716 // user as well as a successor (the TF operands are flagged to the user).
717 // c1, f1 = CopyToReg
718 // c2, f2 = CopyToReg
719 // c3 = TokenFactor c1, c2
722 Chain = Chains[NumRegs-1];
724 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0], NumRegs);
727 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
728 /// operand list. This adds the code marker and includes the number of
729 /// values added into it.
730 void RegsForValue::AddInlineAsmOperands(unsigned Code, bool HasMatching,
731 unsigned MatchingIdx,
733 std::vector<SDValue> &Ops) const {
734 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
736 unsigned Flag = InlineAsm::getFlagWord(Code, Regs.size());
738 Flag = InlineAsm::getFlagWordForMatchingOp(Flag, MatchingIdx);
739 SDValue Res = DAG.getTargetConstant(Flag, MVT::i32);
742 for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) {
743 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVTs[Value]);
744 EVT RegisterVT = RegVTs[Value];
745 for (unsigned i = 0; i != NumRegs; ++i) {
746 assert(Reg < Regs.size() && "Mismatch in # registers expected");
747 Ops.push_back(DAG.getRegister(Regs[Reg++], RegisterVT));
752 void SelectionDAGBuilder::init(GCFunctionInfo *gfi, AliasAnalysis &aa) {
755 TD = DAG.getTarget().getTargetData();
758 /// clear - Clear out the current SelectionDAG and the associated
759 /// state and prepare this SelectionDAGBuilder object to be used
760 /// for a new block. This doesn't clear out information about
761 /// additional blocks that are needed to complete switch lowering
762 /// or PHI node updating; that information is cleared out as it is
764 void SelectionDAGBuilder::clear() {
766 UnusedArgNodeMap.clear();
767 PendingLoads.clear();
768 PendingExports.clear();
769 DanglingDebugInfoMap.clear();
770 CurDebugLoc = DebugLoc();
774 /// getRoot - Return the current virtual root of the Selection DAG,
775 /// flushing any PendingLoad items. This must be done before emitting
776 /// a store or any other node that may need to be ordered after any
777 /// prior load instructions.
779 SDValue SelectionDAGBuilder::getRoot() {
780 if (PendingLoads.empty())
781 return DAG.getRoot();
783 if (PendingLoads.size() == 1) {
784 SDValue Root = PendingLoads[0];
786 PendingLoads.clear();
790 // Otherwise, we have to make a token factor node.
791 SDValue Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
792 &PendingLoads[0], PendingLoads.size());
793 PendingLoads.clear();
798 /// getControlRoot - Similar to getRoot, but instead of flushing all the
799 /// PendingLoad items, flush all the PendingExports items. It is necessary
800 /// to do this before emitting a terminator instruction.
802 SDValue SelectionDAGBuilder::getControlRoot() {
803 SDValue Root = DAG.getRoot();
805 if (PendingExports.empty())
808 // Turn all of the CopyToReg chains into one factored node.
809 if (Root.getOpcode() != ISD::EntryToken) {
810 unsigned i = 0, e = PendingExports.size();
811 for (; i != e; ++i) {
812 assert(PendingExports[i].getNode()->getNumOperands() > 1);
813 if (PendingExports[i].getNode()->getOperand(0) == Root)
814 break; // Don't add the root if we already indirectly depend on it.
818 PendingExports.push_back(Root);
821 Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
823 PendingExports.size());
824 PendingExports.clear();
829 void SelectionDAGBuilder::AssignOrderingToNode(const SDNode *Node) {
830 if (DAG.GetOrdering(Node) != 0) return; // Already has ordering.
831 DAG.AssignOrdering(Node, SDNodeOrder);
833 for (unsigned I = 0, E = Node->getNumOperands(); I != E; ++I)
834 AssignOrderingToNode(Node->getOperand(I).getNode());
837 void SelectionDAGBuilder::visit(const Instruction &I) {
838 // Set up outgoing PHI node register values before emitting the terminator.
839 if (isa<TerminatorInst>(&I))
840 HandlePHINodesInSuccessorBlocks(I.getParent());
842 CurDebugLoc = I.getDebugLoc();
844 visit(I.getOpcode(), I);
846 if (!isa<TerminatorInst>(&I) && !HasTailCall)
847 CopyToExportRegsIfNeeded(&I);
849 CurDebugLoc = DebugLoc();
852 void SelectionDAGBuilder::visitPHI(const PHINode &) {
853 llvm_unreachable("SelectionDAGBuilder shouldn't visit PHI nodes!");
856 void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) {
857 // Note: this doesn't use InstVisitor, because it has to work with
858 // ConstantExpr's in addition to instructions.
860 default: llvm_unreachable("Unknown instruction type encountered!");
861 // Build the switch statement using the Instruction.def file.
862 #define HANDLE_INST(NUM, OPCODE, CLASS) \
863 case Instruction::OPCODE: visit##OPCODE((CLASS&)I); break;
864 #include "llvm/Instruction.def"
867 // Assign the ordering to the freshly created DAG nodes.
868 if (NodeMap.count(&I)) {
870 AssignOrderingToNode(getValue(&I).getNode());
874 // resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V,
875 // generate the debug data structures now that we've seen its definition.
876 void SelectionDAGBuilder::resolveDanglingDebugInfo(const Value *V,
878 DanglingDebugInfo &DDI = DanglingDebugInfoMap[V];
879 MDNode *Variable = NULL;
882 if (const DbgValueInst *DI = dyn_cast_or_null<DbgValueInst>(DDI.getDI())) {
883 Variable = DI->getVariable();
884 Offset = DI->getOffset();
885 } else if (const DbgDeclareInst *DI =
886 dyn_cast_or_null<DbgDeclareInst>(DDI.getDI()))
887 Variable = DI->getVariable();
889 assert (DDI.getDI() == NULL && "Invalid debug info intrinsic!");
894 DebugLoc dl = DDI.getdl();
895 unsigned DbgSDNodeOrder = DDI.getSDNodeOrder();
898 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, Val)) {
899 SDV = DAG.getDbgValue(Variable, Val.getNode(),
900 Val.getResNo(), Offset, dl, DbgSDNodeOrder);
901 DAG.AddDbgValue(SDV, Val.getNode(), false);
904 SDV = DAG.getDbgValue(Variable, UndefValue::get(V->getType()),
905 Offset, dl, SDNodeOrder);
906 DAG.AddDbgValue(SDV, 0, false);
908 DanglingDebugInfoMap[V] = DanglingDebugInfo();
912 // getValue - Return an SDValue for the given Value.
913 SDValue SelectionDAGBuilder::getValue(const Value *V) {
914 // If we already have an SDValue for this value, use it. It's important
915 // to do this first, so that we don't create a CopyFromReg if we already
916 // have a regular SDValue.
917 SDValue &N = NodeMap[V];
918 if (N.getNode()) return N;
920 // If there's a virtual register allocated and initialized for this
922 DenseMap<const Value *, unsigned>::iterator It = FuncInfo.ValueMap.find(V);
923 if (It != FuncInfo.ValueMap.end()) {
924 unsigned InReg = It->second;
925 RegsForValue RFV(*DAG.getContext(), TLI, InReg, V->getType());
926 SDValue Chain = DAG.getEntryNode();
927 N = RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain,NULL);
928 resolveDanglingDebugInfo(V, N);
932 // Otherwise create a new SDValue and remember it.
933 SDValue Val = getValueImpl(V);
935 resolveDanglingDebugInfo(V, Val);
939 /// getNonRegisterValue - Return an SDValue for the given Value, but
940 /// don't look in FuncInfo.ValueMap for a virtual register.
941 SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) {
942 // If we already have an SDValue for this value, use it.
943 SDValue &N = NodeMap[V];
944 if (N.getNode()) return N;
946 // Otherwise create a new SDValue and remember it.
947 SDValue Val = getValueImpl(V);
949 resolveDanglingDebugInfo(V, Val);
953 /// getValueImpl - Helper function for getValue and getNonRegisterValue.
954 /// Create an SDValue for the given value.
955 SDValue SelectionDAGBuilder::getValueImpl(const Value *V) {
956 if (const Constant *C = dyn_cast<Constant>(V)) {
957 EVT VT = TLI.getValueType(V->getType(), true);
959 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C))
960 return DAG.getConstant(*CI, VT);
962 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
963 return DAG.getGlobalAddress(GV, getCurDebugLoc(), VT);
965 if (isa<ConstantPointerNull>(C))
966 return DAG.getConstant(0, TLI.getPointerTy());
968 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
969 return DAG.getConstantFP(*CFP, VT);
971 if (isa<UndefValue>(C) && !V->getType()->isAggregateType())
972 return DAG.getUNDEF(VT);
974 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
975 visit(CE->getOpcode(), *CE);
976 SDValue N1 = NodeMap[V];
977 assert(N1.getNode() && "visit didn't populate the NodeMap!");
981 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) {
982 SmallVector<SDValue, 4> Constants;
983 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end();
985 SDNode *Val = getValue(*OI).getNode();
986 // If the operand is an empty aggregate, there are no values.
988 // Add each leaf value from the operand to the Constants list
989 // to form a flattened list of all the values.
990 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
991 Constants.push_back(SDValue(Val, i));
994 return DAG.getMergeValues(&Constants[0], Constants.size(),
998 if (C->getType()->isStructTy() || C->getType()->isArrayTy()) {
999 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
1000 "Unknown struct or array constant!");
1002 SmallVector<EVT, 4> ValueVTs;
1003 ComputeValueVTs(TLI, C->getType(), ValueVTs);
1004 unsigned NumElts = ValueVTs.size();
1006 return SDValue(); // empty struct
1007 SmallVector<SDValue, 4> Constants(NumElts);
1008 for (unsigned i = 0; i != NumElts; ++i) {
1009 EVT EltVT = ValueVTs[i];
1010 if (isa<UndefValue>(C))
1011 Constants[i] = DAG.getUNDEF(EltVT);
1012 else if (EltVT.isFloatingPoint())
1013 Constants[i] = DAG.getConstantFP(0, EltVT);
1015 Constants[i] = DAG.getConstant(0, EltVT);
1018 return DAG.getMergeValues(&Constants[0], NumElts,
1022 if (const BlockAddress *BA = dyn_cast<BlockAddress>(C))
1023 return DAG.getBlockAddress(BA, VT);
1025 const VectorType *VecTy = cast<VectorType>(V->getType());
1026 unsigned NumElements = VecTy->getNumElements();
1028 // Now that we know the number and type of the elements, get that number of
1029 // elements into the Ops array based on what kind of constant it is.
1030 SmallVector<SDValue, 16> Ops;
1031 if (const ConstantVector *CP = dyn_cast<ConstantVector>(C)) {
1032 for (unsigned i = 0; i != NumElements; ++i)
1033 Ops.push_back(getValue(CP->getOperand(i)));
1035 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!");
1036 EVT EltVT = TLI.getValueType(VecTy->getElementType());
1039 if (EltVT.isFloatingPoint())
1040 Op = DAG.getConstantFP(0, EltVT);
1042 Op = DAG.getConstant(0, EltVT);
1043 Ops.assign(NumElements, Op);
1046 // Create a BUILD_VECTOR node.
1047 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
1048 VT, &Ops[0], Ops.size());
1051 // If this is a static alloca, generate it as the frameindex instead of
1053 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
1054 DenseMap<const AllocaInst*, int>::iterator SI =
1055 FuncInfo.StaticAllocaMap.find(AI);
1056 if (SI != FuncInfo.StaticAllocaMap.end())
1057 return DAG.getFrameIndex(SI->second, TLI.getPointerTy());
1060 // If this is an instruction which fast-isel has deferred, select it now.
1061 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
1062 unsigned InReg = FuncInfo.InitializeRegForValue(Inst);
1063 RegsForValue RFV(*DAG.getContext(), TLI, InReg, Inst->getType());
1064 SDValue Chain = DAG.getEntryNode();
1065 return RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain, NULL);
1068 llvm_unreachable("Can't get register for value!");
1072 void SelectionDAGBuilder::visitRet(const ReturnInst &I) {
1073 SDValue Chain = getControlRoot();
1074 SmallVector<ISD::OutputArg, 8> Outs;
1075 SmallVector<SDValue, 8> OutVals;
1077 if (!FuncInfo.CanLowerReturn) {
1078 unsigned DemoteReg = FuncInfo.DemoteRegister;
1079 const Function *F = I.getParent()->getParent();
1081 // Emit a store of the return value through the virtual register.
1082 // Leave Outs empty so that LowerReturn won't try to load return
1083 // registers the usual way.
1084 SmallVector<EVT, 1> PtrValueVTs;
1085 ComputeValueVTs(TLI, PointerType::getUnqual(F->getReturnType()),
1088 SDValue RetPtr = DAG.getRegister(DemoteReg, PtrValueVTs[0]);
1089 SDValue RetOp = getValue(I.getOperand(0));
1091 SmallVector<EVT, 4> ValueVTs;
1092 SmallVector<uint64_t, 4> Offsets;
1093 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs, &Offsets);
1094 unsigned NumValues = ValueVTs.size();
1096 SmallVector<SDValue, 4> Chains(NumValues);
1097 for (unsigned i = 0; i != NumValues; ++i) {
1098 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(),
1099 RetPtr.getValueType(), RetPtr,
1100 DAG.getIntPtrConstant(Offsets[i]));
1102 DAG.getStore(Chain, getCurDebugLoc(),
1103 SDValue(RetOp.getNode(), RetOp.getResNo() + i),
1104 Add, NULL, Offsets[i], false, false, 0);
1107 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
1108 MVT::Other, &Chains[0], NumValues);
1109 } else if (I.getNumOperands() != 0) {
1110 SmallVector<EVT, 4> ValueVTs;
1111 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs);
1112 unsigned NumValues = ValueVTs.size();
1114 SDValue RetOp = getValue(I.getOperand(0));
1115 for (unsigned j = 0, f = NumValues; j != f; ++j) {
1116 EVT VT = ValueVTs[j];
1118 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1120 const Function *F = I.getParent()->getParent();
1121 if (F->paramHasAttr(0, Attribute::SExt))
1122 ExtendKind = ISD::SIGN_EXTEND;
1123 else if (F->paramHasAttr(0, Attribute::ZExt))
1124 ExtendKind = ISD::ZERO_EXTEND;
1126 // FIXME: C calling convention requires the return type to be promoted
1127 // to at least 32-bit. But this is not necessary for non-C calling
1128 // conventions. The frontend should mark functions whose return values
1129 // require promoting with signext or zeroext attributes.
1130 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
1131 EVT MinVT = TLI.getRegisterType(*DAG.getContext(), MVT::i32);
1132 if (VT.bitsLT(MinVT))
1136 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), VT);
1137 EVT PartVT = TLI.getRegisterType(*DAG.getContext(), VT);
1138 SmallVector<SDValue, 4> Parts(NumParts);
1139 getCopyToParts(DAG, getCurDebugLoc(),
1140 SDValue(RetOp.getNode(), RetOp.getResNo() + j),
1141 &Parts[0], NumParts, PartVT, ExtendKind);
1143 // 'inreg' on function refers to return value
1144 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1145 if (F->paramHasAttr(0, Attribute::InReg))
1148 // Propagate extension type if any
1149 if (F->paramHasAttr(0, Attribute::SExt))
1151 else if (F->paramHasAttr(0, Attribute::ZExt))
1154 for (unsigned i = 0; i < NumParts; ++i) {
1155 Outs.push_back(ISD::OutputArg(Flags, Parts[i].getValueType(),
1157 OutVals.push_back(Parts[i]);
1163 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1164 CallingConv::ID CallConv =
1165 DAG.getMachineFunction().getFunction()->getCallingConv();
1166 Chain = TLI.LowerReturn(Chain, CallConv, isVarArg,
1167 Outs, OutVals, getCurDebugLoc(), DAG);
1169 // Verify that the target's LowerReturn behaved as expected.
1170 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
1171 "LowerReturn didn't return a valid chain!");
1173 // Update the DAG with the new chain value resulting from return lowering.
1177 /// CopyToExportRegsIfNeeded - If the given value has virtual registers
1178 /// created for it, emit nodes to copy the value into the virtual
1180 void SelectionDAGBuilder::CopyToExportRegsIfNeeded(const Value *V) {
1181 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
1182 if (VMI != FuncInfo.ValueMap.end()) {
1183 assert(!V->use_empty() && "Unused value assigned virtual registers!");
1184 CopyValueToVirtualRegister(V, VMI->second);
1188 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
1189 /// the current basic block, add it to ValueMap now so that we'll get a
1191 void SelectionDAGBuilder::ExportFromCurrentBlock(const Value *V) {
1192 // No need to export constants.
1193 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
1195 // Already exported?
1196 if (FuncInfo.isExportedInst(V)) return;
1198 unsigned Reg = FuncInfo.InitializeRegForValue(V);
1199 CopyValueToVirtualRegister(V, Reg);
1202 bool SelectionDAGBuilder::isExportableFromCurrentBlock(const Value *V,
1203 const BasicBlock *FromBB) {
1204 // The operands of the setcc have to be in this block. We don't know
1205 // how to export them from some other block.
1206 if (const Instruction *VI = dyn_cast<Instruction>(V)) {
1207 // Can export from current BB.
1208 if (VI->getParent() == FromBB)
1211 // Is already exported, noop.
1212 return FuncInfo.isExportedInst(V);
1215 // If this is an argument, we can export it if the BB is the entry block or
1216 // if it is already exported.
1217 if (isa<Argument>(V)) {
1218 if (FromBB == &FromBB->getParent()->getEntryBlock())
1221 // Otherwise, can only export this if it is already exported.
1222 return FuncInfo.isExportedInst(V);
1225 // Otherwise, constants can always be exported.
1229 static bool InBlock(const Value *V, const BasicBlock *BB) {
1230 if (const Instruction *I = dyn_cast<Instruction>(V))
1231 return I->getParent() == BB;
1235 /// EmitBranchForMergedCondition - Helper method for FindMergedConditions.
1236 /// This function emits a branch and is used at the leaves of an OR or an
1237 /// AND operator tree.
1240 SelectionDAGBuilder::EmitBranchForMergedCondition(const Value *Cond,
1241 MachineBasicBlock *TBB,
1242 MachineBasicBlock *FBB,
1243 MachineBasicBlock *CurBB,
1244 MachineBasicBlock *SwitchBB) {
1245 const BasicBlock *BB = CurBB->getBasicBlock();
1247 // If the leaf of the tree is a comparison, merge the condition into
1249 if (const CmpInst *BOp = dyn_cast<CmpInst>(Cond)) {
1250 // The operands of the cmp have to be in this block. We don't know
1251 // how to export them from some other block. If this is the first block
1252 // of the sequence, no exporting is needed.
1253 if (CurBB == SwitchBB ||
1254 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
1255 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) {
1256 ISD::CondCode Condition;
1257 if (const ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
1258 Condition = getICmpCondCode(IC->getPredicate());
1259 } else if (const FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) {
1260 Condition = getFCmpCondCode(FC->getPredicate());
1262 Condition = ISD::SETEQ; // silence warning.
1263 llvm_unreachable("Unknown compare instruction");
1266 CaseBlock CB(Condition, BOp->getOperand(0),
1267 BOp->getOperand(1), NULL, TBB, FBB, CurBB);
1268 SwitchCases.push_back(CB);
1273 // Create a CaseBlock record representing this branch.
1274 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()),
1275 NULL, TBB, FBB, CurBB);
1276 SwitchCases.push_back(CB);
1279 /// FindMergedConditions - If Cond is an expression like
1280 void SelectionDAGBuilder::FindMergedConditions(const Value *Cond,
1281 MachineBasicBlock *TBB,
1282 MachineBasicBlock *FBB,
1283 MachineBasicBlock *CurBB,
1284 MachineBasicBlock *SwitchBB,
1286 // If this node is not part of the or/and tree, emit it as a branch.
1287 const Instruction *BOp = dyn_cast<Instruction>(Cond);
1288 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
1289 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
1290 BOp->getParent() != CurBB->getBasicBlock() ||
1291 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
1292 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
1293 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB);
1297 // Create TmpBB after CurBB.
1298 MachineFunction::iterator BBI = CurBB;
1299 MachineFunction &MF = DAG.getMachineFunction();
1300 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock());
1301 CurBB->getParent()->insert(++BBI, TmpBB);
1303 if (Opc == Instruction::Or) {
1304 // Codegen X | Y as:
1312 // Emit the LHS condition.
1313 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, SwitchBB, Opc);
1315 // Emit the RHS condition into TmpBB.
1316 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
1318 assert(Opc == Instruction::And && "Unknown merge op!");
1319 // Codegen X & Y as:
1326 // This requires creation of TmpBB after CurBB.
1328 // Emit the LHS condition.
1329 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, SwitchBB, Opc);
1331 // Emit the RHS condition into TmpBB.
1332 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
1336 /// If the set of cases should be emitted as a series of branches, return true.
1337 /// If we should emit this as a bunch of and/or'd together conditions, return
1340 SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases){
1341 if (Cases.size() != 2) return true;
1343 // If this is two comparisons of the same values or'd or and'd together, they
1344 // will get folded into a single comparison, so don't emit two blocks.
1345 if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
1346 Cases[0].CmpRHS == Cases[1].CmpRHS) ||
1347 (Cases[0].CmpRHS == Cases[1].CmpLHS &&
1348 Cases[0].CmpLHS == Cases[1].CmpRHS)) {
1352 // Handle: (X != null) | (Y != null) --> (X|Y) != 0
1353 // Handle: (X == null) & (Y == null) --> (X|Y) == 0
1354 if (Cases[0].CmpRHS == Cases[1].CmpRHS &&
1355 Cases[0].CC == Cases[1].CC &&
1356 isa<Constant>(Cases[0].CmpRHS) &&
1357 cast<Constant>(Cases[0].CmpRHS)->isNullValue()) {
1358 if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB)
1360 if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB)
1367 void SelectionDAGBuilder::visitBr(const BranchInst &I) {
1368 MachineBasicBlock *BrMBB = FuncInfo.MBB;
1370 // Update machine-CFG edges.
1371 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
1373 // Figure out which block is immediately after the current one.
1374 MachineBasicBlock *NextBlock = 0;
1375 MachineFunction::iterator BBI = BrMBB;
1376 if (++BBI != FuncInfo.MF->end())
1379 if (I.isUnconditional()) {
1380 // Update machine-CFG edges.
1381 BrMBB->addSuccessor(Succ0MBB);
1383 // If this is not a fall-through branch, emit the branch.
1384 if (Succ0MBB != NextBlock)
1385 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
1386 MVT::Other, getControlRoot(),
1387 DAG.getBasicBlock(Succ0MBB)));
1392 // If this condition is one of the special cases we handle, do special stuff
1394 const Value *CondVal = I.getCondition();
1395 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
1397 // If this is a series of conditions that are or'd or and'd together, emit
1398 // this as a sequence of branches instead of setcc's with and/or operations.
1399 // For example, instead of something like:
1412 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
1413 if (BOp->hasOneUse() &&
1414 (BOp->getOpcode() == Instruction::And ||
1415 BOp->getOpcode() == Instruction::Or)) {
1416 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB,
1418 // If the compares in later blocks need to use values not currently
1419 // exported from this block, export them now. This block should always
1420 // be the first entry.
1421 assert(SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!");
1423 // Allow some cases to be rejected.
1424 if (ShouldEmitAsBranches(SwitchCases)) {
1425 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1426 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1427 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1430 // Emit the branch for this block.
1431 visitSwitchCase(SwitchCases[0], BrMBB);
1432 SwitchCases.erase(SwitchCases.begin());
1436 // Okay, we decided not to do this, remove any inserted MBB's and clear
1438 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
1439 FuncInfo.MF->erase(SwitchCases[i].ThisBB);
1441 SwitchCases.clear();
1445 // Create a CaseBlock record representing this branch.
1446 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()),
1447 NULL, Succ0MBB, Succ1MBB, BrMBB);
1449 // Use visitSwitchCase to actually insert the fast branch sequence for this
1451 visitSwitchCase(CB, BrMBB);
1454 /// visitSwitchCase - Emits the necessary code to represent a single node in
1455 /// the binary search tree resulting from lowering a switch instruction.
1456 void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB,
1457 MachineBasicBlock *SwitchBB) {
1459 SDValue CondLHS = getValue(CB.CmpLHS);
1460 DebugLoc dl = getCurDebugLoc();
1462 // Build the setcc now.
1463 if (CB.CmpMHS == NULL) {
1464 // Fold "(X == true)" to X and "(X == false)" to !X to
1465 // handle common cases produced by branch lowering.
1466 if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) &&
1467 CB.CC == ISD::SETEQ)
1469 else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) &&
1470 CB.CC == ISD::SETEQ) {
1471 SDValue True = DAG.getConstant(1, CondLHS.getValueType());
1472 Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True);
1474 Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1476 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now");
1478 const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue();
1479 const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue();
1481 SDValue CmpOp = getValue(CB.CmpMHS);
1482 EVT VT = CmpOp.getValueType();
1484 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) {
1485 Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, VT),
1488 SDValue SUB = DAG.getNode(ISD::SUB, dl,
1489 VT, CmpOp, DAG.getConstant(Low, VT));
1490 Cond = DAG.getSetCC(dl, MVT::i1, SUB,
1491 DAG.getConstant(High-Low, VT), ISD::SETULE);
1495 // Update successor info
1496 SwitchBB->addSuccessor(CB.TrueBB);
1497 SwitchBB->addSuccessor(CB.FalseBB);
1499 // Set NextBlock to be the MBB immediately after the current one, if any.
1500 // This is used to avoid emitting unnecessary branches to the next block.
1501 MachineBasicBlock *NextBlock = 0;
1502 MachineFunction::iterator BBI = SwitchBB;
1503 if (++BBI != FuncInfo.MF->end())
1506 // If the lhs block is the next block, invert the condition so that we can
1507 // fall through to the lhs instead of the rhs block.
1508 if (CB.TrueBB == NextBlock) {
1509 std::swap(CB.TrueBB, CB.FalseBB);
1510 SDValue True = DAG.getConstant(1, Cond.getValueType());
1511 Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True);
1514 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1515 MVT::Other, getControlRoot(), Cond,
1516 DAG.getBasicBlock(CB.TrueBB));
1518 // Insert the false branch.
1519 if (CB.FalseBB != NextBlock)
1520 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1521 DAG.getBasicBlock(CB.FalseBB));
1523 DAG.setRoot(BrCond);
1526 /// visitJumpTable - Emit JumpTable node in the current MBB
1527 void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) {
1528 // Emit the code for the jump table
1529 assert(JT.Reg != -1U && "Should lower JT Header first!");
1530 EVT PTy = TLI.getPointerTy();
1531 SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(),
1533 SDValue Table = DAG.getJumpTable(JT.JTI, PTy);
1534 SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurDebugLoc(),
1535 MVT::Other, Index.getValue(1),
1537 DAG.setRoot(BrJumpTable);
1540 /// visitJumpTableHeader - This function emits necessary code to produce index
1541 /// in the JumpTable from switch case.
1542 void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT,
1543 JumpTableHeader &JTH,
1544 MachineBasicBlock *SwitchBB) {
1545 // Subtract the lowest switch case value from the value being switched on and
1546 // conditional branch to default mbb if the result is greater than the
1547 // difference between smallest and largest cases.
1548 SDValue SwitchOp = getValue(JTH.SValue);
1549 EVT VT = SwitchOp.getValueType();
1550 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1551 DAG.getConstant(JTH.First, VT));
1553 // The SDNode we just created, which holds the value being switched on minus
1554 // the smallest case value, needs to be copied to a virtual register so it
1555 // can be used as an index into the jump table in a subsequent basic block.
1556 // This value may be smaller or larger than the target's pointer type, and
1557 // therefore require extension or truncating.
1558 SwitchOp = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(), TLI.getPointerTy());
1560 unsigned JumpTableReg = FuncInfo.CreateReg(TLI.getPointerTy());
1561 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1562 JumpTableReg, SwitchOp);
1563 JT.Reg = JumpTableReg;
1565 // Emit the range check for the jump table, and branch to the default block
1566 // for the switch statement if the value being switched on exceeds the largest
1567 // case in the switch.
1568 SDValue CMP = DAG.getSetCC(getCurDebugLoc(),
1569 TLI.getSetCCResultType(Sub.getValueType()), Sub,
1570 DAG.getConstant(JTH.Last-JTH.First,VT),
1573 // Set NextBlock to be the MBB immediately after the current one, if any.
1574 // This is used to avoid emitting unnecessary branches to the next block.
1575 MachineBasicBlock *NextBlock = 0;
1576 MachineFunction::iterator BBI = SwitchBB;
1578 if (++BBI != FuncInfo.MF->end())
1581 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1582 MVT::Other, CopyTo, CMP,
1583 DAG.getBasicBlock(JT.Default));
1585 if (JT.MBB != NextBlock)
1586 BrCond = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrCond,
1587 DAG.getBasicBlock(JT.MBB));
1589 DAG.setRoot(BrCond);
1592 /// visitBitTestHeader - This function emits necessary code to produce value
1593 /// suitable for "bit tests"
1594 void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B,
1595 MachineBasicBlock *SwitchBB) {
1596 // Subtract the minimum value
1597 SDValue SwitchOp = getValue(B.SValue);
1598 EVT VT = SwitchOp.getValueType();
1599 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1600 DAG.getConstant(B.First, VT));
1603 SDValue RangeCmp = DAG.getSetCC(getCurDebugLoc(),
1604 TLI.getSetCCResultType(Sub.getValueType()),
1605 Sub, DAG.getConstant(B.Range, VT),
1608 SDValue ShiftOp = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(),
1609 TLI.getPointerTy());
1611 B.Reg = FuncInfo.CreateReg(TLI.getPointerTy());
1612 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1615 // Set NextBlock to be the MBB immediately after the current one, if any.
1616 // This is used to avoid emitting unnecessary branches to the next block.
1617 MachineBasicBlock *NextBlock = 0;
1618 MachineFunction::iterator BBI = SwitchBB;
1619 if (++BBI != FuncInfo.MF->end())
1622 MachineBasicBlock* MBB = B.Cases[0].ThisBB;
1624 SwitchBB->addSuccessor(B.Default);
1625 SwitchBB->addSuccessor(MBB);
1627 SDValue BrRange = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1628 MVT::Other, CopyTo, RangeCmp,
1629 DAG.getBasicBlock(B.Default));
1631 if (MBB != NextBlock)
1632 BrRange = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, CopyTo,
1633 DAG.getBasicBlock(MBB));
1635 DAG.setRoot(BrRange);
1638 /// visitBitTestCase - this function produces one "bit test"
1639 void SelectionDAGBuilder::visitBitTestCase(MachineBasicBlock* NextMBB,
1642 MachineBasicBlock *SwitchBB) {
1643 SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(), Reg,
1644 TLI.getPointerTy());
1646 if (CountPopulation_64(B.Mask) == 1) {
1647 // Testing for a single bit; just compare the shift count with what it
1648 // would need to be to shift a 1 bit in that position.
1649 Cmp = DAG.getSetCC(getCurDebugLoc(),
1650 TLI.getSetCCResultType(ShiftOp.getValueType()),
1652 DAG.getConstant(CountTrailingZeros_64(B.Mask),
1653 TLI.getPointerTy()),
1656 // Make desired shift
1657 SDValue SwitchVal = DAG.getNode(ISD::SHL, getCurDebugLoc(),
1659 DAG.getConstant(1, TLI.getPointerTy()),
1662 // Emit bit tests and jumps
1663 SDValue AndOp = DAG.getNode(ISD::AND, getCurDebugLoc(),
1664 TLI.getPointerTy(), SwitchVal,
1665 DAG.getConstant(B.Mask, TLI.getPointerTy()));
1666 Cmp = DAG.getSetCC(getCurDebugLoc(),
1667 TLI.getSetCCResultType(AndOp.getValueType()),
1668 AndOp, DAG.getConstant(0, TLI.getPointerTy()),
1672 SwitchBB->addSuccessor(B.TargetBB);
1673 SwitchBB->addSuccessor(NextMBB);
1675 SDValue BrAnd = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1676 MVT::Other, getControlRoot(),
1677 Cmp, DAG.getBasicBlock(B.TargetBB));
1679 // Set NextBlock to be the MBB immediately after the current one, if any.
1680 // This is used to avoid emitting unnecessary branches to the next block.
1681 MachineBasicBlock *NextBlock = 0;
1682 MachineFunction::iterator BBI = SwitchBB;
1683 if (++BBI != FuncInfo.MF->end())
1686 if (NextMBB != NextBlock)
1687 BrAnd = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrAnd,
1688 DAG.getBasicBlock(NextMBB));
1693 void SelectionDAGBuilder::visitInvoke(const InvokeInst &I) {
1694 MachineBasicBlock *InvokeMBB = FuncInfo.MBB;
1696 // Retrieve successors.
1697 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
1698 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)];
1700 const Value *Callee(I.getCalledValue());
1701 if (isa<InlineAsm>(Callee))
1704 LowerCallTo(&I, getValue(Callee), false, LandingPad);
1706 // If the value of the invoke is used outside of its defining block, make it
1707 // available as a virtual register.
1708 CopyToExportRegsIfNeeded(&I);
1710 // Update successor info
1711 InvokeMBB->addSuccessor(Return);
1712 InvokeMBB->addSuccessor(LandingPad);
1714 // Drop into normal successor.
1715 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
1716 MVT::Other, getControlRoot(),
1717 DAG.getBasicBlock(Return)));
1720 void SelectionDAGBuilder::visitUnwind(const UnwindInst &I) {
1723 /// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for
1724 /// small case ranges).
1725 bool SelectionDAGBuilder::handleSmallSwitchRange(CaseRec& CR,
1726 CaseRecVector& WorkList,
1728 MachineBasicBlock *Default,
1729 MachineBasicBlock *SwitchBB) {
1730 Case& BackCase = *(CR.Range.second-1);
1732 // Size is the number of Cases represented by this range.
1733 size_t Size = CR.Range.second - CR.Range.first;
1737 // Get the MachineFunction which holds the current MBB. This is used when
1738 // inserting any additional MBBs necessary to represent the switch.
1739 MachineFunction *CurMF = FuncInfo.MF;
1741 // Figure out which block is immediately after the current one.
1742 MachineBasicBlock *NextBlock = 0;
1743 MachineFunction::iterator BBI = CR.CaseBB;
1745 if (++BBI != FuncInfo.MF->end())
1748 // TODO: If any two of the cases has the same destination, and if one value
1749 // is the same as the other, but has one bit unset that the other has set,
1750 // use bit manipulation to do two compares at once. For example:
1751 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
1753 // Rearrange the case blocks so that the last one falls through if possible.
1754 if (NextBlock && Default != NextBlock && BackCase.BB != NextBlock) {
1755 // The last case block won't fall through into 'NextBlock' if we emit the
1756 // branches in this order. See if rearranging a case value would help.
1757 for (CaseItr I = CR.Range.first, E = CR.Range.second-1; I != E; ++I) {
1758 if (I->BB == NextBlock) {
1759 std::swap(*I, BackCase);
1765 // Create a CaseBlock record representing a conditional branch to
1766 // the Case's target mbb if the value being switched on SV is equal
1768 MachineBasicBlock *CurBlock = CR.CaseBB;
1769 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
1770 MachineBasicBlock *FallThrough;
1772 FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock());
1773 CurMF->insert(BBI, FallThrough);
1775 // Put SV in a virtual register to make it available from the new blocks.
1776 ExportFromCurrentBlock(SV);
1778 // If the last case doesn't match, go to the default block.
1779 FallThrough = Default;
1782 const Value *RHS, *LHS, *MHS;
1784 if (I->High == I->Low) {
1785 // This is just small small case range :) containing exactly 1 case
1787 LHS = SV; RHS = I->High; MHS = NULL;
1790 LHS = I->Low; MHS = SV; RHS = I->High;
1792 CaseBlock CB(CC, LHS, RHS, MHS, I->BB, FallThrough, CurBlock);
1794 // If emitting the first comparison, just call visitSwitchCase to emit the
1795 // code into the current block. Otherwise, push the CaseBlock onto the
1796 // vector to be later processed by SDISel, and insert the node's MBB
1797 // before the next MBB.
1798 if (CurBlock == SwitchBB)
1799 visitSwitchCase(CB, SwitchBB);
1801 SwitchCases.push_back(CB);
1803 CurBlock = FallThrough;
1809 static inline bool areJTsAllowed(const TargetLowering &TLI) {
1810 return !DisableJumpTables &&
1811 (TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
1812 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other));
1815 static APInt ComputeRange(const APInt &First, const APInt &Last) {
1816 APInt LastExt(Last), FirstExt(First);
1817 uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1;
1818 LastExt.sext(BitWidth); FirstExt.sext(BitWidth);
1819 return (LastExt - FirstExt + 1ULL);
1822 /// handleJTSwitchCase - Emit jumptable for current switch case range
1823 bool SelectionDAGBuilder::handleJTSwitchCase(CaseRec& CR,
1824 CaseRecVector& WorkList,
1826 MachineBasicBlock* Default,
1827 MachineBasicBlock *SwitchBB) {
1828 Case& FrontCase = *CR.Range.first;
1829 Case& BackCase = *(CR.Range.second-1);
1831 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
1832 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
1834 APInt TSize(First.getBitWidth(), 0);
1835 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1839 if (!areJTsAllowed(TLI) || TSize.ult(4))
1842 APInt Range = ComputeRange(First, Last);
1843 double Density = TSize.roundToDouble() / Range.roundToDouble();
1847 DEBUG(dbgs() << "Lowering jump table\n"
1848 << "First entry: " << First << ". Last entry: " << Last << '\n'
1849 << "Range: " << Range
1850 << "Size: " << TSize << ". Density: " << Density << "\n\n");
1852 // Get the MachineFunction which holds the current MBB. This is used when
1853 // inserting any additional MBBs necessary to represent the switch.
1854 MachineFunction *CurMF = FuncInfo.MF;
1856 // Figure out which block is immediately after the current one.
1857 MachineFunction::iterator BBI = CR.CaseBB;
1860 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1862 // Create a new basic block to hold the code for loading the address
1863 // of the jump table, and jumping to it. Update successor information;
1864 // we will either branch to the default case for the switch, or the jump
1866 MachineBasicBlock *JumpTableBB = CurMF->CreateMachineBasicBlock(LLVMBB);
1867 CurMF->insert(BBI, JumpTableBB);
1868 CR.CaseBB->addSuccessor(Default);
1869 CR.CaseBB->addSuccessor(JumpTableBB);
1871 // Build a vector of destination BBs, corresponding to each target
1872 // of the jump table. If the value of the jump table slot corresponds to
1873 // a case statement, push the case's BB onto the vector, otherwise, push
1875 std::vector<MachineBasicBlock*> DestBBs;
1877 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) {
1878 const APInt &Low = cast<ConstantInt>(I->Low)->getValue();
1879 const APInt &High = cast<ConstantInt>(I->High)->getValue();
1881 if (Low.sle(TEI) && TEI.sle(High)) {
1882 DestBBs.push_back(I->BB);
1886 DestBBs.push_back(Default);
1890 // Update successor info. Add one edge to each unique successor.
1891 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs());
1892 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(),
1893 E = DestBBs.end(); I != E; ++I) {
1894 if (!SuccsHandled[(*I)->getNumber()]) {
1895 SuccsHandled[(*I)->getNumber()] = true;
1896 JumpTableBB->addSuccessor(*I);
1900 // Create a jump table index for this jump table.
1901 unsigned JTEncoding = TLI.getJumpTableEncoding();
1902 unsigned JTI = CurMF->getOrCreateJumpTableInfo(JTEncoding)
1903 ->createJumpTableIndex(DestBBs);
1905 // Set the jump table information so that we can codegen it as a second
1906 // MachineBasicBlock
1907 JumpTable JT(-1U, JTI, JumpTableBB, Default);
1908 JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == SwitchBB));
1909 if (CR.CaseBB == SwitchBB)
1910 visitJumpTableHeader(JT, JTH, SwitchBB);
1912 JTCases.push_back(JumpTableBlock(JTH, JT));
1917 /// handleBTSplitSwitchCase - emit comparison and split binary search tree into
1919 bool SelectionDAGBuilder::handleBTSplitSwitchCase(CaseRec& CR,
1920 CaseRecVector& WorkList,
1922 MachineBasicBlock *Default,
1923 MachineBasicBlock *SwitchBB) {
1924 // Get the MachineFunction which holds the current MBB. This is used when
1925 // inserting any additional MBBs necessary to represent the switch.
1926 MachineFunction *CurMF = FuncInfo.MF;
1928 // Figure out which block is immediately after the current one.
1929 MachineFunction::iterator BBI = CR.CaseBB;
1932 Case& FrontCase = *CR.Range.first;
1933 Case& BackCase = *(CR.Range.second-1);
1934 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1936 // Size is the number of Cases represented by this range.
1937 unsigned Size = CR.Range.second - CR.Range.first;
1939 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
1940 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
1942 CaseItr Pivot = CR.Range.first + Size/2;
1944 // Select optimal pivot, maximizing sum density of LHS and RHS. This will
1945 // (heuristically) allow us to emit JumpTable's later.
1946 APInt TSize(First.getBitWidth(), 0);
1947 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1951 APInt LSize = FrontCase.size();
1952 APInt RSize = TSize-LSize;
1953 DEBUG(dbgs() << "Selecting best pivot: \n"
1954 << "First: " << First << ", Last: " << Last <<'\n'
1955 << "LSize: " << LSize << ", RSize: " << RSize << '\n');
1956 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second;
1958 const APInt &LEnd = cast<ConstantInt>(I->High)->getValue();
1959 const APInt &RBegin = cast<ConstantInt>(J->Low)->getValue();
1960 APInt Range = ComputeRange(LEnd, RBegin);
1961 assert((Range - 2ULL).isNonNegative() &&
1962 "Invalid case distance");
1963 double LDensity = (double)LSize.roundToDouble() /
1964 (LEnd - First + 1ULL).roundToDouble();
1965 double RDensity = (double)RSize.roundToDouble() /
1966 (Last - RBegin + 1ULL).roundToDouble();
1967 double Metric = Range.logBase2()*(LDensity+RDensity);
1968 // Should always split in some non-trivial place
1969 DEBUG(dbgs() <<"=>Step\n"
1970 << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n'
1971 << "LDensity: " << LDensity
1972 << ", RDensity: " << RDensity << '\n'
1973 << "Metric: " << Metric << '\n');
1974 if (FMetric < Metric) {
1977 DEBUG(dbgs() << "Current metric set to: " << FMetric << '\n');
1983 if (areJTsAllowed(TLI)) {
1984 // If our case is dense we *really* should handle it earlier!
1985 assert((FMetric > 0) && "Should handle dense range earlier!");
1987 Pivot = CR.Range.first + Size/2;
1990 CaseRange LHSR(CR.Range.first, Pivot);
1991 CaseRange RHSR(Pivot, CR.Range.second);
1992 Constant *C = Pivot->Low;
1993 MachineBasicBlock *FalseBB = 0, *TrueBB = 0;
1995 // We know that we branch to the LHS if the Value being switched on is
1996 // less than the Pivot value, C. We use this to optimize our binary
1997 // tree a bit, by recognizing that if SV is greater than or equal to the
1998 // LHS's Case Value, and that Case Value is exactly one less than the
1999 // Pivot's Value, then we can branch directly to the LHS's Target,
2000 // rather than creating a leaf node for it.
2001 if ((LHSR.second - LHSR.first) == 1 &&
2002 LHSR.first->High == CR.GE &&
2003 cast<ConstantInt>(C)->getValue() ==
2004 (cast<ConstantInt>(CR.GE)->getValue() + 1LL)) {
2005 TrueBB = LHSR.first->BB;
2007 TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2008 CurMF->insert(BBI, TrueBB);
2009 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR));
2011 // Put SV in a virtual register to make it available from the new blocks.
2012 ExportFromCurrentBlock(SV);
2015 // Similar to the optimization above, if the Value being switched on is
2016 // known to be less than the Constant CR.LT, and the current Case Value
2017 // is CR.LT - 1, then we can branch directly to the target block for
2018 // the current Case Value, rather than emitting a RHS leaf node for it.
2019 if ((RHSR.second - RHSR.first) == 1 && CR.LT &&
2020 cast<ConstantInt>(RHSR.first->Low)->getValue() ==
2021 (cast<ConstantInt>(CR.LT)->getValue() - 1LL)) {
2022 FalseBB = RHSR.first->BB;
2024 FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2025 CurMF->insert(BBI, FalseBB);
2026 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR));
2028 // Put SV in a virtual register to make it available from the new blocks.
2029 ExportFromCurrentBlock(SV);
2032 // Create a CaseBlock record representing a conditional branch to
2033 // the LHS node if the value being switched on SV is less than C.
2034 // Otherwise, branch to LHS.
2035 CaseBlock CB(ISD::SETLT, SV, C, NULL, TrueBB, FalseBB, CR.CaseBB);
2037 if (CR.CaseBB == SwitchBB)
2038 visitSwitchCase(CB, SwitchBB);
2040 SwitchCases.push_back(CB);
2045 /// handleBitTestsSwitchCase - if current case range has few destination and
2046 /// range span less, than machine word bitwidth, encode case range into series
2047 /// of masks and emit bit tests with these masks.
2048 bool SelectionDAGBuilder::handleBitTestsSwitchCase(CaseRec& CR,
2049 CaseRecVector& WorkList,
2051 MachineBasicBlock* Default,
2052 MachineBasicBlock *SwitchBB){
2053 EVT PTy = TLI.getPointerTy();
2054 unsigned IntPtrBits = PTy.getSizeInBits();
2056 Case& FrontCase = *CR.Range.first;
2057 Case& BackCase = *(CR.Range.second-1);
2059 // Get the MachineFunction which holds the current MBB. This is used when
2060 // inserting any additional MBBs necessary to represent the switch.
2061 MachineFunction *CurMF = FuncInfo.MF;
2063 // If target does not have legal shift left, do not emit bit tests at all.
2064 if (!TLI.isOperationLegal(ISD::SHL, TLI.getPointerTy()))
2068 for (CaseItr I = CR.Range.first, E = CR.Range.second;
2070 // Single case counts one, case range - two.
2071 numCmps += (I->Low == I->High ? 1 : 2);
2074 // Count unique destinations
2075 SmallSet<MachineBasicBlock*, 4> Dests;
2076 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2077 Dests.insert(I->BB);
2078 if (Dests.size() > 3)
2079 // Don't bother the code below, if there are too much unique destinations
2082 DEBUG(dbgs() << "Total number of unique destinations: "
2083 << Dests.size() << '\n'
2084 << "Total number of comparisons: " << numCmps << '\n');
2086 // Compute span of values.
2087 const APInt& minValue = cast<ConstantInt>(FrontCase.Low)->getValue();
2088 const APInt& maxValue = cast<ConstantInt>(BackCase.High)->getValue();
2089 APInt cmpRange = maxValue - minValue;
2091 DEBUG(dbgs() << "Compare range: " << cmpRange << '\n'
2092 << "Low bound: " << minValue << '\n'
2093 << "High bound: " << maxValue << '\n');
2095 if (cmpRange.uge(IntPtrBits) ||
2096 (!(Dests.size() == 1 && numCmps >= 3) &&
2097 !(Dests.size() == 2 && numCmps >= 5) &&
2098 !(Dests.size() >= 3 && numCmps >= 6)))
2101 DEBUG(dbgs() << "Emitting bit tests\n");
2102 APInt lowBound = APInt::getNullValue(cmpRange.getBitWidth());
2104 // Optimize the case where all the case values fit in a
2105 // word without having to subtract minValue. In this case,
2106 // we can optimize away the subtraction.
2107 if (minValue.isNonNegative() && maxValue.slt(IntPtrBits)) {
2108 cmpRange = maxValue;
2110 lowBound = minValue;
2113 CaseBitsVector CasesBits;
2114 unsigned i, count = 0;
2116 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2117 MachineBasicBlock* Dest = I->BB;
2118 for (i = 0; i < count; ++i)
2119 if (Dest == CasesBits[i].BB)
2123 assert((count < 3) && "Too much destinations to test!");
2124 CasesBits.push_back(CaseBits(0, Dest, 0));
2128 const APInt& lowValue = cast<ConstantInt>(I->Low)->getValue();
2129 const APInt& highValue = cast<ConstantInt>(I->High)->getValue();
2131 uint64_t lo = (lowValue - lowBound).getZExtValue();
2132 uint64_t hi = (highValue - lowBound).getZExtValue();
2134 for (uint64_t j = lo; j <= hi; j++) {
2135 CasesBits[i].Mask |= 1ULL << j;
2136 CasesBits[i].Bits++;
2140 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp());
2144 // Figure out which block is immediately after the current one.
2145 MachineFunction::iterator BBI = CR.CaseBB;
2148 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2150 DEBUG(dbgs() << "Cases:\n");
2151 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) {
2152 DEBUG(dbgs() << "Mask: " << CasesBits[i].Mask
2153 << ", Bits: " << CasesBits[i].Bits
2154 << ", BB: " << CasesBits[i].BB << '\n');
2156 MachineBasicBlock *CaseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2157 CurMF->insert(BBI, CaseBB);
2158 BTC.push_back(BitTestCase(CasesBits[i].Mask,
2162 // Put SV in a virtual register to make it available from the new blocks.
2163 ExportFromCurrentBlock(SV);
2166 BitTestBlock BTB(lowBound, cmpRange, SV,
2167 -1U, (CR.CaseBB == SwitchBB),
2168 CR.CaseBB, Default, BTC);
2170 if (CR.CaseBB == SwitchBB)
2171 visitBitTestHeader(BTB, SwitchBB);
2173 BitTestCases.push_back(BTB);
2178 /// Clusterify - Transform simple list of Cases into list of CaseRange's
2179 size_t SelectionDAGBuilder::Clusterify(CaseVector& Cases,
2180 const SwitchInst& SI) {
2183 // Start with "simple" cases
2184 for (size_t i = 1; i < SI.getNumSuccessors(); ++i) {
2185 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SI.getSuccessor(i)];
2186 Cases.push_back(Case(SI.getSuccessorValue(i),
2187 SI.getSuccessorValue(i),
2190 std::sort(Cases.begin(), Cases.end(), CaseCmp());
2192 // Merge case into clusters
2193 if (Cases.size() >= 2)
2194 // Must recompute end() each iteration because it may be
2195 // invalidated by erase if we hold on to it
2196 for (CaseItr I = Cases.begin(), J = ++(Cases.begin()); J != Cases.end(); ) {
2197 const APInt& nextValue = cast<ConstantInt>(J->Low)->getValue();
2198 const APInt& currentValue = cast<ConstantInt>(I->High)->getValue();
2199 MachineBasicBlock* nextBB = J->BB;
2200 MachineBasicBlock* currentBB = I->BB;
2202 // If the two neighboring cases go to the same destination, merge them
2203 // into a single case.
2204 if ((nextValue - currentValue == 1) && (currentBB == nextBB)) {
2212 for (CaseItr I=Cases.begin(), E=Cases.end(); I!=E; ++I, ++numCmps) {
2213 if (I->Low != I->High)
2214 // A range counts double, since it requires two compares.
2221 void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) {
2222 MachineBasicBlock *SwitchMBB = FuncInfo.MBB;
2224 // Figure out which block is immediately after the current one.
2225 MachineBasicBlock *NextBlock = 0;
2226 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()];
2228 // If there is only the default destination, branch to it if it is not the
2229 // next basic block. Otherwise, just fall through.
2230 if (SI.getNumOperands() == 2) {
2231 // Update machine-CFG edges.
2233 // If this is not a fall-through branch, emit the branch.
2234 SwitchMBB->addSuccessor(Default);
2235 if (Default != NextBlock)
2236 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
2237 MVT::Other, getControlRoot(),
2238 DAG.getBasicBlock(Default)));
2243 // If there are any non-default case statements, create a vector of Cases
2244 // representing each one, and sort the vector so that we can efficiently
2245 // create a binary search tree from them.
2247 size_t numCmps = Clusterify(Cases, SI);
2248 DEBUG(dbgs() << "Clusterify finished. Total clusters: " << Cases.size()
2249 << ". Total compares: " << numCmps << '\n');
2252 // Get the Value to be switched on and default basic blocks, which will be
2253 // inserted into CaseBlock records, representing basic blocks in the binary
2255 const Value *SV = SI.getOperand(0);
2257 // Push the initial CaseRec onto the worklist
2258 CaseRecVector WorkList;
2259 WorkList.push_back(CaseRec(SwitchMBB,0,0,
2260 CaseRange(Cases.begin(),Cases.end())));
2262 while (!WorkList.empty()) {
2263 // Grab a record representing a case range to process off the worklist
2264 CaseRec CR = WorkList.back();
2265 WorkList.pop_back();
2267 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2270 // If the range has few cases (two or less) emit a series of specific
2272 if (handleSmallSwitchRange(CR, WorkList, SV, Default, SwitchMBB))
2275 // If the switch has more than 5 blocks, and at least 40% dense, and the
2276 // target supports indirect branches, then emit a jump table rather than
2277 // lowering the switch to a binary tree of conditional branches.
2278 if (handleJTSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2281 // Emit binary tree. We need to pick a pivot, and push left and right ranges
2282 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call.
2283 handleBTSplitSwitchCase(CR, WorkList, SV, Default, SwitchMBB);
2287 void SelectionDAGBuilder::visitIndirectBr(const IndirectBrInst &I) {
2288 MachineBasicBlock *IndirectBrMBB = FuncInfo.MBB;
2290 // Update machine-CFG edges with unique successors.
2291 SmallVector<BasicBlock*, 32> succs;
2292 succs.reserve(I.getNumSuccessors());
2293 for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i)
2294 succs.push_back(I.getSuccessor(i));
2295 array_pod_sort(succs.begin(), succs.end());
2296 succs.erase(std::unique(succs.begin(), succs.end()), succs.end());
2297 for (unsigned i = 0, e = succs.size(); i != e; ++i)
2298 IndirectBrMBB->addSuccessor(FuncInfo.MBBMap[succs[i]]);
2300 DAG.setRoot(DAG.getNode(ISD::BRIND, getCurDebugLoc(),
2301 MVT::Other, getControlRoot(),
2302 getValue(I.getAddress())));
2305 void SelectionDAGBuilder::visitFSub(const User &I) {
2306 // -0.0 - X --> fneg
2307 const Type *Ty = I.getType();
2308 if (Ty->isVectorTy()) {
2309 if (ConstantVector *CV = dyn_cast<ConstantVector>(I.getOperand(0))) {
2310 const VectorType *DestTy = cast<VectorType>(I.getType());
2311 const Type *ElTy = DestTy->getElementType();
2312 unsigned VL = DestTy->getNumElements();
2313 std::vector<Constant*> NZ(VL, ConstantFP::getNegativeZero(ElTy));
2314 Constant *CNZ = ConstantVector::get(&NZ[0], NZ.size());
2316 SDValue Op2 = getValue(I.getOperand(1));
2317 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(),
2318 Op2.getValueType(), Op2));
2324 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I.getOperand(0)))
2325 if (CFP->isExactlyValue(ConstantFP::getNegativeZero(Ty)->getValueAPF())) {
2326 SDValue Op2 = getValue(I.getOperand(1));
2327 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(),
2328 Op2.getValueType(), Op2));
2332 visitBinary(I, ISD::FSUB);
2335 void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) {
2336 SDValue Op1 = getValue(I.getOperand(0));
2337 SDValue Op2 = getValue(I.getOperand(1));
2338 setValue(&I, DAG.getNode(OpCode, getCurDebugLoc(),
2339 Op1.getValueType(), Op1, Op2));
2342 void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) {
2343 SDValue Op1 = getValue(I.getOperand(0));
2344 SDValue Op2 = getValue(I.getOperand(1));
2345 if (!I.getType()->isVectorTy() &&
2346 Op2.getValueType() != TLI.getShiftAmountTy()) {
2347 // If the operand is smaller than the shift count type, promote it.
2348 EVT PTy = TLI.getPointerTy();
2349 EVT STy = TLI.getShiftAmountTy();
2350 if (STy.bitsGT(Op2.getValueType()))
2351 Op2 = DAG.getNode(ISD::ANY_EXTEND, getCurDebugLoc(),
2352 TLI.getShiftAmountTy(), Op2);
2353 // If the operand is larger than the shift count type but the shift
2354 // count type has enough bits to represent any shift value, truncate
2355 // it now. This is a common case and it exposes the truncate to
2356 // optimization early.
2357 else if (STy.getSizeInBits() >=
2358 Log2_32_Ceil(Op2.getValueType().getSizeInBits()))
2359 Op2 = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2360 TLI.getShiftAmountTy(), Op2);
2361 // Otherwise we'll need to temporarily settle for some other
2362 // convenient type; type legalization will make adjustments as
2364 else if (PTy.bitsLT(Op2.getValueType()))
2365 Op2 = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2366 TLI.getPointerTy(), Op2);
2367 else if (PTy.bitsGT(Op2.getValueType()))
2368 Op2 = DAG.getNode(ISD::ANY_EXTEND, getCurDebugLoc(),
2369 TLI.getPointerTy(), Op2);
2372 setValue(&I, DAG.getNode(Opcode, getCurDebugLoc(),
2373 Op1.getValueType(), Op1, Op2));
2376 void SelectionDAGBuilder::visitICmp(const User &I) {
2377 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2378 if (const ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2379 predicate = IC->getPredicate();
2380 else if (const ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2381 predicate = ICmpInst::Predicate(IC->getPredicate());
2382 SDValue Op1 = getValue(I.getOperand(0));
2383 SDValue Op2 = getValue(I.getOperand(1));
2384 ISD::CondCode Opcode = getICmpCondCode(predicate);
2386 EVT DestVT = TLI.getValueType(I.getType());
2387 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Opcode));
2390 void SelectionDAGBuilder::visitFCmp(const User &I) {
2391 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2392 if (const FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2393 predicate = FC->getPredicate();
2394 else if (const ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2395 predicate = FCmpInst::Predicate(FC->getPredicate());
2396 SDValue Op1 = getValue(I.getOperand(0));
2397 SDValue Op2 = getValue(I.getOperand(1));
2398 ISD::CondCode Condition = getFCmpCondCode(predicate);
2399 EVT DestVT = TLI.getValueType(I.getType());
2400 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Condition));
2403 void SelectionDAGBuilder::visitSelect(const User &I) {
2404 SmallVector<EVT, 4> ValueVTs;
2405 ComputeValueVTs(TLI, I.getType(), ValueVTs);
2406 unsigned NumValues = ValueVTs.size();
2407 if (NumValues == 0) return;
2409 SmallVector<SDValue, 4> Values(NumValues);
2410 SDValue Cond = getValue(I.getOperand(0));
2411 SDValue TrueVal = getValue(I.getOperand(1));
2412 SDValue FalseVal = getValue(I.getOperand(2));
2414 for (unsigned i = 0; i != NumValues; ++i)
2415 Values[i] = DAG.getNode(ISD::SELECT, getCurDebugLoc(),
2416 TrueVal.getNode()->getValueType(TrueVal.getResNo()+i),
2418 SDValue(TrueVal.getNode(),
2419 TrueVal.getResNo() + i),
2420 SDValue(FalseVal.getNode(),
2421 FalseVal.getResNo() + i));
2423 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2424 DAG.getVTList(&ValueVTs[0], NumValues),
2425 &Values[0], NumValues));
2428 void SelectionDAGBuilder::visitTrunc(const User &I) {
2429 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2430 SDValue N = getValue(I.getOperand(0));
2431 EVT DestVT = TLI.getValueType(I.getType());
2432 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), DestVT, N));
2435 void SelectionDAGBuilder::visitZExt(const User &I) {
2436 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2437 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2438 SDValue N = getValue(I.getOperand(0));
2439 EVT DestVT = TLI.getValueType(I.getType());
2440 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(), DestVT, N));
2443 void SelectionDAGBuilder::visitSExt(const User &I) {
2444 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2445 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2446 SDValue N = getValue(I.getOperand(0));
2447 EVT DestVT = TLI.getValueType(I.getType());
2448 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurDebugLoc(), DestVT, N));
2451 void SelectionDAGBuilder::visitFPTrunc(const User &I) {
2452 // FPTrunc is never a no-op cast, no need to check
2453 SDValue N = getValue(I.getOperand(0));
2454 EVT DestVT = TLI.getValueType(I.getType());
2455 setValue(&I, DAG.getNode(ISD::FP_ROUND, getCurDebugLoc(),
2456 DestVT, N, DAG.getIntPtrConstant(0)));
2459 void SelectionDAGBuilder::visitFPExt(const User &I){
2460 // FPTrunc is never a no-op cast, no need to check
2461 SDValue N = getValue(I.getOperand(0));
2462 EVT DestVT = TLI.getValueType(I.getType());
2463 setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurDebugLoc(), DestVT, N));
2466 void SelectionDAGBuilder::visitFPToUI(const User &I) {
2467 // FPToUI is never a no-op cast, no need to check
2468 SDValue N = getValue(I.getOperand(0));
2469 EVT DestVT = TLI.getValueType(I.getType());
2470 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurDebugLoc(), DestVT, N));
2473 void SelectionDAGBuilder::visitFPToSI(const User &I) {
2474 // FPToSI is never a no-op cast, no need to check
2475 SDValue N = getValue(I.getOperand(0));
2476 EVT DestVT = TLI.getValueType(I.getType());
2477 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurDebugLoc(), DestVT, N));
2480 void SelectionDAGBuilder::visitUIToFP(const User &I) {
2481 // UIToFP is never a no-op cast, no need to check
2482 SDValue N = getValue(I.getOperand(0));
2483 EVT DestVT = TLI.getValueType(I.getType());
2484 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurDebugLoc(), DestVT, N));
2487 void SelectionDAGBuilder::visitSIToFP(const User &I){
2488 // SIToFP is never a no-op cast, no need to check
2489 SDValue N = getValue(I.getOperand(0));
2490 EVT DestVT = TLI.getValueType(I.getType());
2491 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurDebugLoc(), DestVT, N));
2494 void SelectionDAGBuilder::visitPtrToInt(const User &I) {
2495 // What to do depends on the size of the integer and the size of the pointer.
2496 // We can either truncate, zero extend, or no-op, accordingly.
2497 SDValue N = getValue(I.getOperand(0));
2498 EVT DestVT = TLI.getValueType(I.getType());
2499 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT));
2502 void SelectionDAGBuilder::visitIntToPtr(const User &I) {
2503 // What to do depends on the size of the integer and the size of the pointer.
2504 // We can either truncate, zero extend, or no-op, accordingly.
2505 SDValue N = getValue(I.getOperand(0));
2506 EVT DestVT = TLI.getValueType(I.getType());
2507 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT));
2510 void SelectionDAGBuilder::visitBitCast(const User &I) {
2511 SDValue N = getValue(I.getOperand(0));
2512 EVT DestVT = TLI.getValueType(I.getType());
2514 // BitCast assures us that source and destination are the same size so this is
2515 // either a BIT_CONVERT or a no-op.
2516 if (DestVT != N.getValueType())
2517 setValue(&I, DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
2518 DestVT, N)); // convert types.
2520 setValue(&I, N); // noop cast.
2523 void SelectionDAGBuilder::visitInsertElement(const User &I) {
2524 SDValue InVec = getValue(I.getOperand(0));
2525 SDValue InVal = getValue(I.getOperand(1));
2526 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2528 getValue(I.getOperand(2)));
2529 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurDebugLoc(),
2530 TLI.getValueType(I.getType()),
2531 InVec, InVal, InIdx));
2534 void SelectionDAGBuilder::visitExtractElement(const User &I) {
2535 SDValue InVec = getValue(I.getOperand(0));
2536 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2538 getValue(I.getOperand(1)));
2539 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2540 TLI.getValueType(I.getType()), InVec, InIdx));
2543 // Utility for visitShuffleVector - Returns true if the mask is mask starting
2544 // from SIndx and increasing to the element length (undefs are allowed).
2545 static bool SequentialMask(SmallVectorImpl<int> &Mask, unsigned SIndx) {
2546 unsigned MaskNumElts = Mask.size();
2547 for (unsigned i = 0; i != MaskNumElts; ++i)
2548 if ((Mask[i] >= 0) && (Mask[i] != (int)(i + SIndx)))
2553 void SelectionDAGBuilder::visitShuffleVector(const User &I) {
2554 SmallVector<int, 8> Mask;
2555 SDValue Src1 = getValue(I.getOperand(0));
2556 SDValue Src2 = getValue(I.getOperand(1));
2558 // Convert the ConstantVector mask operand into an array of ints, with -1
2559 // representing undef values.
2560 SmallVector<Constant*, 8> MaskElts;
2561 cast<Constant>(I.getOperand(2))->getVectorElements(MaskElts);
2562 unsigned MaskNumElts = MaskElts.size();
2563 for (unsigned i = 0; i != MaskNumElts; ++i) {
2564 if (isa<UndefValue>(MaskElts[i]))
2567 Mask.push_back(cast<ConstantInt>(MaskElts[i])->getSExtValue());
2570 EVT VT = TLI.getValueType(I.getType());
2571 EVT SrcVT = Src1.getValueType();
2572 unsigned SrcNumElts = SrcVT.getVectorNumElements();
2574 if (SrcNumElts == MaskNumElts) {
2575 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2580 // Normalize the shuffle vector since mask and vector length don't match.
2581 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) {
2582 // Mask is longer than the source vectors and is a multiple of the source
2583 // vectors. We can use concatenate vector to make the mask and vectors
2585 if (SrcNumElts*2 == MaskNumElts && SequentialMask(Mask, 0)) {
2586 // The shuffle is concatenating two vectors together.
2587 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurDebugLoc(),
2592 // Pad both vectors with undefs to make them the same length as the mask.
2593 unsigned NumConcat = MaskNumElts / SrcNumElts;
2594 bool Src1U = Src1.getOpcode() == ISD::UNDEF;
2595 bool Src2U = Src2.getOpcode() == ISD::UNDEF;
2596 SDValue UndefVal = DAG.getUNDEF(SrcVT);
2598 SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal);
2599 SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal);
2603 Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2604 getCurDebugLoc(), VT,
2605 &MOps1[0], NumConcat);
2606 Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2607 getCurDebugLoc(), VT,
2608 &MOps2[0], NumConcat);
2610 // Readjust mask for new input vector length.
2611 SmallVector<int, 8> MappedOps;
2612 for (unsigned i = 0; i != MaskNumElts; ++i) {
2614 if (Idx < (int)SrcNumElts)
2615 MappedOps.push_back(Idx);
2617 MappedOps.push_back(Idx + MaskNumElts - SrcNumElts);
2620 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2625 if (SrcNumElts > MaskNumElts) {
2626 // Analyze the access pattern of the vector to see if we can extract
2627 // two subvectors and do the shuffle. The analysis is done by calculating
2628 // the range of elements the mask access on both vectors.
2629 int MinRange[2] = { SrcNumElts+1, SrcNumElts+1};
2630 int MaxRange[2] = {-1, -1};
2632 for (unsigned i = 0; i != MaskNumElts; ++i) {
2638 if (Idx >= (int)SrcNumElts) {
2642 if (Idx > MaxRange[Input])
2643 MaxRange[Input] = Idx;
2644 if (Idx < MinRange[Input])
2645 MinRange[Input] = Idx;
2648 // Check if the access is smaller than the vector size and can we find
2649 // a reasonable extract index.
2650 int RangeUse[2] = { 2, 2 }; // 0 = Unused, 1 = Extract, 2 = Can not
2652 int StartIdx[2]; // StartIdx to extract from
2653 for (int Input=0; Input < 2; ++Input) {
2654 if (MinRange[Input] == (int)(SrcNumElts+1) && MaxRange[Input] == -1) {
2655 RangeUse[Input] = 0; // Unused
2656 StartIdx[Input] = 0;
2657 } else if (MaxRange[Input] - MinRange[Input] < (int)MaskNumElts) {
2658 // Fits within range but we should see if we can find a good
2659 // start index that is a multiple of the mask length.
2660 if (MaxRange[Input] < (int)MaskNumElts) {
2661 RangeUse[Input] = 1; // Extract from beginning of the vector
2662 StartIdx[Input] = 0;
2664 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts;
2665 if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts &&
2666 StartIdx[Input] + MaskNumElts < SrcNumElts)
2667 RangeUse[Input] = 1; // Extract from a multiple of the mask length.
2672 if (RangeUse[0] == 0 && RangeUse[1] == 0) {
2673 setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used.
2676 else if (RangeUse[0] < 2 && RangeUse[1] < 2) {
2677 // Extract appropriate subvector and generate a vector shuffle
2678 for (int Input=0; Input < 2; ++Input) {
2679 SDValue &Src = Input == 0 ? Src1 : Src2;
2680 if (RangeUse[Input] == 0)
2681 Src = DAG.getUNDEF(VT);
2683 Src = DAG.getNode(ISD::EXTRACT_SUBVECTOR, getCurDebugLoc(), VT,
2684 Src, DAG.getIntPtrConstant(StartIdx[Input]));
2687 // Calculate new mask.
2688 SmallVector<int, 8> MappedOps;
2689 for (unsigned i = 0; i != MaskNumElts; ++i) {
2692 MappedOps.push_back(Idx);
2693 else if (Idx < (int)SrcNumElts)
2694 MappedOps.push_back(Idx - StartIdx[0]);
2696 MappedOps.push_back(Idx - SrcNumElts - StartIdx[1] + MaskNumElts);
2699 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2705 // We can't use either concat vectors or extract subvectors so fall back to
2706 // replacing the shuffle with extract and build vector.
2707 // to insert and build vector.
2708 EVT EltVT = VT.getVectorElementType();
2709 EVT PtrVT = TLI.getPointerTy();
2710 SmallVector<SDValue,8> Ops;
2711 for (unsigned i = 0; i != MaskNumElts; ++i) {
2713 Ops.push_back(DAG.getUNDEF(EltVT));
2718 if (Idx < (int)SrcNumElts)
2719 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2720 EltVT, Src1, DAG.getConstant(Idx, PtrVT));
2722 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2724 DAG.getConstant(Idx - SrcNumElts, PtrVT));
2730 setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
2731 VT, &Ops[0], Ops.size()));
2734 void SelectionDAGBuilder::visitInsertValue(const InsertValueInst &I) {
2735 const Value *Op0 = I.getOperand(0);
2736 const Value *Op1 = I.getOperand(1);
2737 const Type *AggTy = I.getType();
2738 const Type *ValTy = Op1->getType();
2739 bool IntoUndef = isa<UndefValue>(Op0);
2740 bool FromUndef = isa<UndefValue>(Op1);
2742 unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy,
2743 I.idx_begin(), I.idx_end());
2745 SmallVector<EVT, 4> AggValueVTs;
2746 ComputeValueVTs(TLI, AggTy, AggValueVTs);
2747 SmallVector<EVT, 4> ValValueVTs;
2748 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2750 unsigned NumAggValues = AggValueVTs.size();
2751 unsigned NumValValues = ValValueVTs.size();
2752 SmallVector<SDValue, 4> Values(NumAggValues);
2754 SDValue Agg = getValue(Op0);
2755 SDValue Val = getValue(Op1);
2757 // Copy the beginning value(s) from the original aggregate.
2758 for (; i != LinearIndex; ++i)
2759 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2760 SDValue(Agg.getNode(), Agg.getResNo() + i);
2761 // Copy values from the inserted value(s).
2762 for (; i != LinearIndex + NumValValues; ++i)
2763 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2764 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex);
2765 // Copy remaining value(s) from the original aggregate.
2766 for (; i != NumAggValues; ++i)
2767 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2768 SDValue(Agg.getNode(), Agg.getResNo() + i);
2770 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2771 DAG.getVTList(&AggValueVTs[0], NumAggValues),
2772 &Values[0], NumAggValues));
2775 void SelectionDAGBuilder::visitExtractValue(const ExtractValueInst &I) {
2776 const Value *Op0 = I.getOperand(0);
2777 const Type *AggTy = Op0->getType();
2778 const Type *ValTy = I.getType();
2779 bool OutOfUndef = isa<UndefValue>(Op0);
2781 unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy,
2782 I.idx_begin(), I.idx_end());
2784 SmallVector<EVT, 4> ValValueVTs;
2785 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2787 unsigned NumValValues = ValValueVTs.size();
2788 SmallVector<SDValue, 4> Values(NumValValues);
2790 SDValue Agg = getValue(Op0);
2791 // Copy out the selected value(s).
2792 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i)
2793 Values[i - LinearIndex] =
2795 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) :
2796 SDValue(Agg.getNode(), Agg.getResNo() + i);
2798 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2799 DAG.getVTList(&ValValueVTs[0], NumValValues),
2800 &Values[0], NumValValues));
2803 void SelectionDAGBuilder::visitGetElementPtr(const User &I) {
2804 SDValue N = getValue(I.getOperand(0));
2805 const Type *Ty = I.getOperand(0)->getType();
2807 for (GetElementPtrInst::const_op_iterator OI = I.op_begin()+1, E = I.op_end();
2809 const Value *Idx = *OI;
2810 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
2811 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
2814 uint64_t Offset = TD->getStructLayout(StTy)->getElementOffset(Field);
2815 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
2816 DAG.getIntPtrConstant(Offset));
2819 Ty = StTy->getElementType(Field);
2820 } else if (const UnionType *UnTy = dyn_cast<UnionType>(Ty)) {
2821 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
2823 // Offset canonically 0 for unions, but type changes
2824 Ty = UnTy->getElementType(Field);
2826 Ty = cast<SequentialType>(Ty)->getElementType();
2828 // If this is a constant subscript, handle it quickly.
2829 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
2830 if (CI->isZero()) continue;
2832 TD->getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
2834 EVT PTy = TLI.getPointerTy();
2835 unsigned PtrBits = PTy.getSizeInBits();
2837 OffsVal = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2839 DAG.getConstant(Offs, MVT::i64));
2841 OffsVal = DAG.getIntPtrConstant(Offs);
2843 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
2848 // N = N + Idx * ElementSize;
2849 APInt ElementSize = APInt(TLI.getPointerTy().getSizeInBits(),
2850 TD->getTypeAllocSize(Ty));
2851 SDValue IdxN = getValue(Idx);
2853 // If the index is smaller or larger than intptr_t, truncate or extend
2855 IdxN = DAG.getSExtOrTrunc(IdxN, getCurDebugLoc(), N.getValueType());
2857 // If this is a multiply by a power of two, turn it into a shl
2858 // immediately. This is a very common case.
2859 if (ElementSize != 1) {
2860 if (ElementSize.isPowerOf2()) {
2861 unsigned Amt = ElementSize.logBase2();
2862 IdxN = DAG.getNode(ISD::SHL, getCurDebugLoc(),
2863 N.getValueType(), IdxN,
2864 DAG.getConstant(Amt, TLI.getPointerTy()));
2866 SDValue Scale = DAG.getConstant(ElementSize, TLI.getPointerTy());
2867 IdxN = DAG.getNode(ISD::MUL, getCurDebugLoc(),
2868 N.getValueType(), IdxN, Scale);
2872 N = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2873 N.getValueType(), N, IdxN);
2880 void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) {
2881 // If this is a fixed sized alloca in the entry block of the function,
2882 // allocate it statically on the stack.
2883 if (FuncInfo.StaticAllocaMap.count(&I))
2884 return; // getValue will auto-populate this.
2886 const Type *Ty = I.getAllocatedType();
2887 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
2889 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty),
2892 SDValue AllocSize = getValue(I.getArraySize());
2894 EVT IntPtr = TLI.getPointerTy();
2895 if (AllocSize.getValueType() != IntPtr)
2896 AllocSize = DAG.getZExtOrTrunc(AllocSize, getCurDebugLoc(), IntPtr);
2898 AllocSize = DAG.getNode(ISD::MUL, getCurDebugLoc(), IntPtr,
2900 DAG.getConstant(TySize, IntPtr));
2902 // Handle alignment. If the requested alignment is less than or equal to
2903 // the stack alignment, ignore it. If the size is greater than or equal to
2904 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
2905 unsigned StackAlign = TM.getFrameInfo()->getStackAlignment();
2906 if (Align <= StackAlign)
2909 // Round the size of the allocation up to the stack alignment size
2910 // by add SA-1 to the size.
2911 AllocSize = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2912 AllocSize.getValueType(), AllocSize,
2913 DAG.getIntPtrConstant(StackAlign-1));
2915 // Mask out the low bits for alignment purposes.
2916 AllocSize = DAG.getNode(ISD::AND, getCurDebugLoc(),
2917 AllocSize.getValueType(), AllocSize,
2918 DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1)));
2920 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) };
2921 SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other);
2922 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, getCurDebugLoc(),
2925 DAG.setRoot(DSA.getValue(1));
2927 // Inform the Frame Information that we have just allocated a variable-sized
2929 FuncInfo.MF->getFrameInfo()->CreateVariableSizedObject(Align ? Align : 1);
2932 void SelectionDAGBuilder::visitLoad(const LoadInst &I) {
2933 const Value *SV = I.getOperand(0);
2934 SDValue Ptr = getValue(SV);
2936 const Type *Ty = I.getType();
2938 bool isVolatile = I.isVolatile();
2939 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
2940 unsigned Alignment = I.getAlignment();
2942 SmallVector<EVT, 4> ValueVTs;
2943 SmallVector<uint64_t, 4> Offsets;
2944 ComputeValueVTs(TLI, Ty, ValueVTs, &Offsets);
2945 unsigned NumValues = ValueVTs.size();
2950 bool ConstantMemory = false;
2952 // Serialize volatile loads with other side effects.
2954 else if (AA->pointsToConstantMemory(SV)) {
2955 // Do not serialize (non-volatile) loads of constant memory with anything.
2956 Root = DAG.getEntryNode();
2957 ConstantMemory = true;
2959 // Do not serialize non-volatile loads against each other.
2960 Root = DAG.getRoot();
2963 SmallVector<SDValue, 4> Values(NumValues);
2964 SmallVector<SDValue, 4> Chains(NumValues);
2965 EVT PtrVT = Ptr.getValueType();
2966 for (unsigned i = 0; i != NumValues; ++i) {
2967 SDValue A = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2969 DAG.getConstant(Offsets[i], PtrVT));
2970 SDValue L = DAG.getLoad(ValueVTs[i], getCurDebugLoc(), Root,
2971 A, SV, Offsets[i], isVolatile,
2972 isNonTemporal, Alignment);
2975 Chains[i] = L.getValue(1);
2978 if (!ConstantMemory) {
2979 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
2980 MVT::Other, &Chains[0], NumValues);
2984 PendingLoads.push_back(Chain);
2987 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2988 DAG.getVTList(&ValueVTs[0], NumValues),
2989 &Values[0], NumValues));
2992 void SelectionDAGBuilder::visitStore(const StoreInst &I) {
2993 const Value *SrcV = I.getOperand(0);
2994 const Value *PtrV = I.getOperand(1);
2996 SmallVector<EVT, 4> ValueVTs;
2997 SmallVector<uint64_t, 4> Offsets;
2998 ComputeValueVTs(TLI, SrcV->getType(), ValueVTs, &Offsets);
2999 unsigned NumValues = ValueVTs.size();
3003 // Get the lowered operands. Note that we do this after
3004 // checking if NumResults is zero, because with zero results
3005 // the operands won't have values in the map.
3006 SDValue Src = getValue(SrcV);
3007 SDValue Ptr = getValue(PtrV);
3009 SDValue Root = getRoot();
3010 SmallVector<SDValue, 4> Chains(NumValues);
3011 EVT PtrVT = Ptr.getValueType();
3012 bool isVolatile = I.isVolatile();
3013 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
3014 unsigned Alignment = I.getAlignment();
3016 for (unsigned i = 0; i != NumValues; ++i) {
3017 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT, Ptr,
3018 DAG.getConstant(Offsets[i], PtrVT));
3019 Chains[i] = DAG.getStore(Root, getCurDebugLoc(),
3020 SDValue(Src.getNode(), Src.getResNo() + i),
3021 Add, PtrV, Offsets[i], isVolatile,
3022 isNonTemporal, Alignment);
3025 DAG.setRoot(DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3026 MVT::Other, &Chains[0], NumValues));
3029 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
3031 void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I,
3032 unsigned Intrinsic) {
3033 bool HasChain = !I.doesNotAccessMemory();
3034 bool OnlyLoad = HasChain && I.onlyReadsMemory();
3036 // Build the operand list.
3037 SmallVector<SDValue, 8> Ops;
3038 if (HasChain) { // If this intrinsic has side-effects, chainify it.
3040 // We don't need to serialize loads against other loads.
3041 Ops.push_back(DAG.getRoot());
3043 Ops.push_back(getRoot());
3047 // Info is set by getTgtMemInstrinsic
3048 TargetLowering::IntrinsicInfo Info;
3049 bool IsTgtIntrinsic = TLI.getTgtMemIntrinsic(Info, I, Intrinsic);
3051 // Add the intrinsic ID as an integer operand if it's not a target intrinsic.
3052 if (!IsTgtIntrinsic)
3053 Ops.push_back(DAG.getConstant(Intrinsic, TLI.getPointerTy()));
3055 // Add all operands of the call to the operand list.
3056 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) {
3057 SDValue Op = getValue(I.getArgOperand(i));
3058 assert(TLI.isTypeLegal(Op.getValueType()) &&
3059 "Intrinsic uses a non-legal type?");
3063 SmallVector<EVT, 4> ValueVTs;
3064 ComputeValueVTs(TLI, I.getType(), ValueVTs);
3066 for (unsigned Val = 0, E = ValueVTs.size(); Val != E; ++Val) {
3067 assert(TLI.isTypeLegal(ValueVTs[Val]) &&
3068 "Intrinsic uses a non-legal type?");
3073 ValueVTs.push_back(MVT::Other);
3075 SDVTList VTs = DAG.getVTList(ValueVTs.data(), ValueVTs.size());
3079 if (IsTgtIntrinsic) {
3080 // This is target intrinsic that touches memory
3081 Result = DAG.getMemIntrinsicNode(Info.opc, getCurDebugLoc(),
3082 VTs, &Ops[0], Ops.size(),
3083 Info.memVT, Info.ptrVal, Info.offset,
3084 Info.align, Info.vol,
3085 Info.readMem, Info.writeMem);
3086 } else if (!HasChain) {
3087 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurDebugLoc(),
3088 VTs, &Ops[0], Ops.size());
3089 } else if (!I.getType()->isVoidTy()) {
3090 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurDebugLoc(),
3091 VTs, &Ops[0], Ops.size());
3093 Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurDebugLoc(),
3094 VTs, &Ops[0], Ops.size());
3098 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1);
3100 PendingLoads.push_back(Chain);
3105 if (!I.getType()->isVoidTy()) {
3106 if (const VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
3107 EVT VT = TLI.getValueType(PTy);
3108 Result = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(), VT, Result);
3111 setValue(&I, Result);
3115 /// GetSignificand - Get the significand and build it into a floating-point
3116 /// number with exponent of 1:
3118 /// Op = (Op & 0x007fffff) | 0x3f800000;
3120 /// where Op is the hexidecimal representation of floating point value.
3122 GetSignificand(SelectionDAG &DAG, SDValue Op, DebugLoc dl) {
3123 SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3124 DAG.getConstant(0x007fffff, MVT::i32));
3125 SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1,
3126 DAG.getConstant(0x3f800000, MVT::i32));
3127 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t2);
3130 /// GetExponent - Get the exponent:
3132 /// (float)(int)(((Op & 0x7f800000) >> 23) - 127);
3134 /// where Op is the hexidecimal representation of floating point value.
3136 GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI,
3138 SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3139 DAG.getConstant(0x7f800000, MVT::i32));
3140 SDValue t1 = DAG.getNode(ISD::SRL, dl, MVT::i32, t0,
3141 DAG.getConstant(23, TLI.getPointerTy()));
3142 SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1,
3143 DAG.getConstant(127, MVT::i32));
3144 return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2);
3147 /// getF32Constant - Get 32-bit floating point constant.
3149 getF32Constant(SelectionDAG &DAG, unsigned Flt) {
3150 return DAG.getConstantFP(APFloat(APInt(32, Flt)), MVT::f32);
3153 /// Inlined utility function to implement binary input atomic intrinsics for
3154 /// visitIntrinsicCall: I is a call instruction
3155 /// Op is the associated NodeType for I
3157 SelectionDAGBuilder::implVisitBinaryAtomic(const CallInst& I,
3159 SDValue Root = getRoot();
3161 DAG.getAtomic(Op, getCurDebugLoc(),
3162 getValue(I.getArgOperand(1)).getValueType().getSimpleVT(),
3164 getValue(I.getArgOperand(0)),
3165 getValue(I.getArgOperand(1)),
3166 I.getArgOperand(0));
3168 DAG.setRoot(L.getValue(1));
3172 // implVisitAluOverflow - Lower arithmetic overflow instrinsics.
3174 SelectionDAGBuilder::implVisitAluOverflow(const CallInst &I, ISD::NodeType Op) {
3175 SDValue Op1 = getValue(I.getArgOperand(0));
3176 SDValue Op2 = getValue(I.getArgOperand(1));
3178 SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1);
3179 setValue(&I, DAG.getNode(Op, getCurDebugLoc(), VTs, Op1, Op2));
3183 /// visitExp - Lower an exp intrinsic. Handles the special sequences for
3184 /// limited-precision mode.
3186 SelectionDAGBuilder::visitExp(const CallInst &I) {
3188 DebugLoc dl = getCurDebugLoc();
3190 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3191 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3192 SDValue Op = getValue(I.getArgOperand(0));
3194 // Put the exponent in the right bit position for later addition to the
3197 // #define LOG2OFe 1.4426950f
3198 // IntegerPartOfX = ((int32_t)(X * LOG2OFe));
3199 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3200 getF32Constant(DAG, 0x3fb8aa3b));
3201 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3203 // FractionalPartOfX = (X * LOG2OFe) - (float)IntegerPartOfX;
3204 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3205 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3207 // IntegerPartOfX <<= 23;
3208 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3209 DAG.getConstant(23, TLI.getPointerTy()));
3211 if (LimitFloatPrecision <= 6) {
3212 // For floating-point precision of 6:
3214 // TwoToFractionalPartOfX =
3216 // (0.735607626f + 0.252464424f * x) * x;
3218 // error 0.0144103317, which is 6 bits
3219 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3220 getF32Constant(DAG, 0x3e814304));
3221 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3222 getF32Constant(DAG, 0x3f3c50c8));
3223 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3224 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3225 getF32Constant(DAG, 0x3f7f5e7e));
3226 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,MVT::i32, t5);
3228 // Add the exponent into the result in integer domain.
3229 SDValue t6 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3230 TwoToFracPartOfX, IntegerPartOfX);
3232 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t6);
3233 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3234 // For floating-point precision of 12:
3236 // TwoToFractionalPartOfX =
3239 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3241 // 0.000107046256 error, which is 13 to 14 bits
3242 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3243 getF32Constant(DAG, 0x3da235e3));
3244 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3245 getF32Constant(DAG, 0x3e65b8f3));
3246 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3247 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3248 getF32Constant(DAG, 0x3f324b07));
3249 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3250 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3251 getF32Constant(DAG, 0x3f7ff8fd));
3252 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,MVT::i32, t7);
3254 // Add the exponent into the result in integer domain.
3255 SDValue t8 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3256 TwoToFracPartOfX, IntegerPartOfX);
3258 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t8);
3259 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3260 // For floating-point precision of 18:
3262 // TwoToFractionalPartOfX =
3266 // (0.554906021e-1f +
3267 // (0.961591928e-2f +
3268 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3270 // error 2.47208000*10^(-7), which is better than 18 bits
3271 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3272 getF32Constant(DAG, 0x3924b03e));
3273 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3274 getF32Constant(DAG, 0x3ab24b87));
3275 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3276 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3277 getF32Constant(DAG, 0x3c1d8c17));
3278 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3279 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3280 getF32Constant(DAG, 0x3d634a1d));
3281 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3282 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3283 getF32Constant(DAG, 0x3e75fe14));
3284 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3285 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3286 getF32Constant(DAG, 0x3f317234));
3287 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3288 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3289 getF32Constant(DAG, 0x3f800000));
3290 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,
3293 // Add the exponent into the result in integer domain.
3294 SDValue t14 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3295 TwoToFracPartOfX, IntegerPartOfX);
3297 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t14);
3300 // No special expansion.
3301 result = DAG.getNode(ISD::FEXP, dl,
3302 getValue(I.getArgOperand(0)).getValueType(),
3303 getValue(I.getArgOperand(0)));
3306 setValue(&I, result);
3309 /// visitLog - Lower a log intrinsic. Handles the special sequences for
3310 /// limited-precision mode.
3312 SelectionDAGBuilder::visitLog(const CallInst &I) {
3314 DebugLoc dl = getCurDebugLoc();
3316 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3317 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3318 SDValue Op = getValue(I.getArgOperand(0));
3319 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
3321 // Scale the exponent by log(2) [0.69314718f].
3322 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3323 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3324 getF32Constant(DAG, 0x3f317218));
3326 // Get the significand and build it into a floating-point number with
3328 SDValue X = GetSignificand(DAG, Op1, dl);
3330 if (LimitFloatPrecision <= 6) {
3331 // For floating-point precision of 6:
3335 // (1.4034025f - 0.23903021f * x) * x;
3337 // error 0.0034276066, which is better than 8 bits
3338 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3339 getF32Constant(DAG, 0xbe74c456));
3340 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3341 getF32Constant(DAG, 0x3fb3a2b1));
3342 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3343 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3344 getF32Constant(DAG, 0x3f949a29));
3346 result = DAG.getNode(ISD::FADD, dl,
3347 MVT::f32, LogOfExponent, LogOfMantissa);
3348 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3349 // For floating-point precision of 12:
3355 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x;
3357 // error 0.000061011436, which is 14 bits
3358 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3359 getF32Constant(DAG, 0xbd67b6d6));
3360 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3361 getF32Constant(DAG, 0x3ee4f4b8));
3362 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3363 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3364 getF32Constant(DAG, 0x3fbc278b));
3365 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3366 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3367 getF32Constant(DAG, 0x40348e95));
3368 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3369 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3370 getF32Constant(DAG, 0x3fdef31a));
3372 result = DAG.getNode(ISD::FADD, dl,
3373 MVT::f32, LogOfExponent, LogOfMantissa);
3374 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3375 // For floating-point precision of 18:
3383 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x;
3385 // error 0.0000023660568, which is better than 18 bits
3386 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3387 getF32Constant(DAG, 0xbc91e5ac));
3388 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3389 getF32Constant(DAG, 0x3e4350aa));
3390 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3391 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3392 getF32Constant(DAG, 0x3f60d3e3));
3393 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3394 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3395 getF32Constant(DAG, 0x4011cdf0));
3396 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3397 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3398 getF32Constant(DAG, 0x406cfd1c));
3399 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3400 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3401 getF32Constant(DAG, 0x408797cb));
3402 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3403 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3404 getF32Constant(DAG, 0x4006dcab));
3406 result = DAG.getNode(ISD::FADD, dl,
3407 MVT::f32, LogOfExponent, LogOfMantissa);
3410 // No special expansion.
3411 result = DAG.getNode(ISD::FLOG, dl,
3412 getValue(I.getArgOperand(0)).getValueType(),
3413 getValue(I.getArgOperand(0)));
3416 setValue(&I, result);
3419 /// visitLog2 - Lower a log2 intrinsic. Handles the special sequences for
3420 /// limited-precision mode.
3422 SelectionDAGBuilder::visitLog2(const CallInst &I) {
3424 DebugLoc dl = getCurDebugLoc();
3426 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3427 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3428 SDValue Op = getValue(I.getArgOperand(0));
3429 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
3431 // Get the exponent.
3432 SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl);
3434 // Get the significand and build it into a floating-point number with
3436 SDValue X = GetSignificand(DAG, Op1, dl);
3438 // Different possible minimax approximations of significand in
3439 // floating-point for various degrees of accuracy over [1,2].
3440 if (LimitFloatPrecision <= 6) {
3441 // For floating-point precision of 6:
3443 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x;
3445 // error 0.0049451742, which is more than 7 bits
3446 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3447 getF32Constant(DAG, 0xbeb08fe0));
3448 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3449 getF32Constant(DAG, 0x40019463));
3450 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3451 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3452 getF32Constant(DAG, 0x3fd6633d));
3454 result = DAG.getNode(ISD::FADD, dl,
3455 MVT::f32, LogOfExponent, Log2ofMantissa);
3456 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3457 // For floating-point precision of 12:
3463 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x;
3465 // error 0.0000876136000, which is better than 13 bits
3466 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3467 getF32Constant(DAG, 0xbda7262e));
3468 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3469 getF32Constant(DAG, 0x3f25280b));
3470 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3471 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3472 getF32Constant(DAG, 0x4007b923));
3473 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3474 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3475 getF32Constant(DAG, 0x40823e2f));
3476 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3477 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3478 getF32Constant(DAG, 0x4020d29c));
3480 result = DAG.getNode(ISD::FADD, dl,
3481 MVT::f32, LogOfExponent, Log2ofMantissa);
3482 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3483 // For floating-point precision of 18:
3492 // 0.25691327e-1f * x) * x) * x) * x) * x) * x;
3494 // error 0.0000018516, which is better than 18 bits
3495 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3496 getF32Constant(DAG, 0xbcd2769e));
3497 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3498 getF32Constant(DAG, 0x3e8ce0b9));
3499 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3500 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3501 getF32Constant(DAG, 0x3fa22ae7));
3502 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3503 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3504 getF32Constant(DAG, 0x40525723));
3505 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3506 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3507 getF32Constant(DAG, 0x40aaf200));
3508 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3509 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3510 getF32Constant(DAG, 0x40c39dad));
3511 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3512 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3513 getF32Constant(DAG, 0x4042902c));
3515 result = DAG.getNode(ISD::FADD, dl,
3516 MVT::f32, LogOfExponent, Log2ofMantissa);
3519 // No special expansion.
3520 result = DAG.getNode(ISD::FLOG2, dl,
3521 getValue(I.getArgOperand(0)).getValueType(),
3522 getValue(I.getArgOperand(0)));
3525 setValue(&I, result);
3528 /// visitLog10 - Lower a log10 intrinsic. Handles the special sequences for
3529 /// limited-precision mode.
3531 SelectionDAGBuilder::visitLog10(const CallInst &I) {
3533 DebugLoc dl = getCurDebugLoc();
3535 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3536 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3537 SDValue Op = getValue(I.getArgOperand(0));
3538 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
3540 // Scale the exponent by log10(2) [0.30102999f].
3541 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3542 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3543 getF32Constant(DAG, 0x3e9a209a));
3545 // Get the significand and build it into a floating-point number with
3547 SDValue X = GetSignificand(DAG, Op1, dl);
3549 if (LimitFloatPrecision <= 6) {
3550 // For floating-point precision of 6:
3552 // Log10ofMantissa =
3554 // (0.60948995f - 0.10380950f * x) * x;
3556 // error 0.0014886165, which is 6 bits
3557 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3558 getF32Constant(DAG, 0xbdd49a13));
3559 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3560 getF32Constant(DAG, 0x3f1c0789));
3561 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3562 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3563 getF32Constant(DAG, 0x3f011300));
3565 result = DAG.getNode(ISD::FADD, dl,
3566 MVT::f32, LogOfExponent, Log10ofMantissa);
3567 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3568 // For floating-point precision of 12:
3570 // Log10ofMantissa =
3573 // (-0.31664806f + 0.47637168e-1f * x) * x) * x;
3575 // error 0.00019228036, which is better than 12 bits
3576 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3577 getF32Constant(DAG, 0x3d431f31));
3578 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
3579 getF32Constant(DAG, 0x3ea21fb2));
3580 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3581 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3582 getF32Constant(DAG, 0x3f6ae232));
3583 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3584 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
3585 getF32Constant(DAG, 0x3f25f7c3));
3587 result = DAG.getNode(ISD::FADD, dl,
3588 MVT::f32, LogOfExponent, Log10ofMantissa);
3589 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3590 // For floating-point precision of 18:
3592 // Log10ofMantissa =
3597 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x;
3599 // error 0.0000037995730, which is better than 18 bits
3600 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3601 getF32Constant(DAG, 0x3c5d51ce));
3602 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
3603 getF32Constant(DAG, 0x3e00685a));
3604 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3605 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3606 getF32Constant(DAG, 0x3efb6798));
3607 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3608 SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
3609 getF32Constant(DAG, 0x3f88d192));
3610 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3611 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3612 getF32Constant(DAG, 0x3fc4316c));
3613 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3614 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8,
3615 getF32Constant(DAG, 0x3f57ce70));
3617 result = DAG.getNode(ISD::FADD, dl,
3618 MVT::f32, LogOfExponent, Log10ofMantissa);
3621 // No special expansion.
3622 result = DAG.getNode(ISD::FLOG10, dl,
3623 getValue(I.getArgOperand(0)).getValueType(),
3624 getValue(I.getArgOperand(0)));
3627 setValue(&I, result);
3630 /// visitExp2 - Lower an exp2 intrinsic. Handles the special sequences for
3631 /// limited-precision mode.
3633 SelectionDAGBuilder::visitExp2(const CallInst &I) {
3635 DebugLoc dl = getCurDebugLoc();
3637 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3638 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3639 SDValue Op = getValue(I.getArgOperand(0));
3641 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Op);
3643 // FractionalPartOfX = x - (float)IntegerPartOfX;
3644 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3645 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, Op, t1);
3647 // IntegerPartOfX <<= 23;
3648 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3649 DAG.getConstant(23, TLI.getPointerTy()));
3651 if (LimitFloatPrecision <= 6) {
3652 // For floating-point precision of 6:
3654 // TwoToFractionalPartOfX =
3656 // (0.735607626f + 0.252464424f * x) * x;
3658 // error 0.0144103317, which is 6 bits
3659 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3660 getF32Constant(DAG, 0x3e814304));
3661 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3662 getF32Constant(DAG, 0x3f3c50c8));
3663 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3664 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3665 getF32Constant(DAG, 0x3f7f5e7e));
3666 SDValue t6 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t5);
3667 SDValue TwoToFractionalPartOfX =
3668 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
3670 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3671 MVT::f32, TwoToFractionalPartOfX);
3672 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3673 // For floating-point precision of 12:
3675 // TwoToFractionalPartOfX =
3678 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3680 // error 0.000107046256, which is 13 to 14 bits
3681 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3682 getF32Constant(DAG, 0x3da235e3));
3683 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3684 getF32Constant(DAG, 0x3e65b8f3));
3685 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3686 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3687 getF32Constant(DAG, 0x3f324b07));
3688 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3689 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3690 getF32Constant(DAG, 0x3f7ff8fd));
3691 SDValue t8 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t7);
3692 SDValue TwoToFractionalPartOfX =
3693 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
3695 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3696 MVT::f32, TwoToFractionalPartOfX);
3697 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3698 // For floating-point precision of 18:
3700 // TwoToFractionalPartOfX =
3704 // (0.554906021e-1f +
3705 // (0.961591928e-2f +
3706 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3707 // error 2.47208000*10^(-7), which is better than 18 bits
3708 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3709 getF32Constant(DAG, 0x3924b03e));
3710 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3711 getF32Constant(DAG, 0x3ab24b87));
3712 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3713 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3714 getF32Constant(DAG, 0x3c1d8c17));
3715 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3716 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3717 getF32Constant(DAG, 0x3d634a1d));
3718 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3719 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3720 getF32Constant(DAG, 0x3e75fe14));
3721 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3722 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3723 getF32Constant(DAG, 0x3f317234));
3724 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3725 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3726 getF32Constant(DAG, 0x3f800000));
3727 SDValue t14 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t13);
3728 SDValue TwoToFractionalPartOfX =
3729 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
3731 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3732 MVT::f32, TwoToFractionalPartOfX);
3735 // No special expansion.
3736 result = DAG.getNode(ISD::FEXP2, dl,
3737 getValue(I.getArgOperand(0)).getValueType(),
3738 getValue(I.getArgOperand(0)));
3741 setValue(&I, result);
3744 /// visitPow - Lower a pow intrinsic. Handles the special sequences for
3745 /// limited-precision mode with x == 10.0f.
3747 SelectionDAGBuilder::visitPow(const CallInst &I) {
3749 const Value *Val = I.getArgOperand(0);
3750 DebugLoc dl = getCurDebugLoc();
3751 bool IsExp10 = false;
3753 if (getValue(Val).getValueType() == MVT::f32 &&
3754 getValue(I.getArgOperand(1)).getValueType() == MVT::f32 &&
3755 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3756 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(Val))) {
3757 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
3759 IsExp10 = CFP->getValueAPF().bitwiseIsEqual(Ten);
3764 if (IsExp10 && LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3765 SDValue Op = getValue(I.getArgOperand(1));
3767 // Put the exponent in the right bit position for later addition to the
3770 // #define LOG2OF10 3.3219281f
3771 // IntegerPartOfX = (int32_t)(x * LOG2OF10);
3772 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3773 getF32Constant(DAG, 0x40549a78));
3774 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3776 // FractionalPartOfX = x - (float)IntegerPartOfX;
3777 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3778 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3780 // IntegerPartOfX <<= 23;
3781 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3782 DAG.getConstant(23, TLI.getPointerTy()));
3784 if (LimitFloatPrecision <= 6) {
3785 // For floating-point precision of 6:
3787 // twoToFractionalPartOfX =
3789 // (0.735607626f + 0.252464424f * x) * x;
3791 // error 0.0144103317, which is 6 bits
3792 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3793 getF32Constant(DAG, 0x3e814304));
3794 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3795 getF32Constant(DAG, 0x3f3c50c8));
3796 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3797 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3798 getF32Constant(DAG, 0x3f7f5e7e));
3799 SDValue t6 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t5);
3800 SDValue TwoToFractionalPartOfX =
3801 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
3803 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3804 MVT::f32, TwoToFractionalPartOfX);
3805 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3806 // For floating-point precision of 12:
3808 // TwoToFractionalPartOfX =
3811 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3813 // error 0.000107046256, which is 13 to 14 bits
3814 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3815 getF32Constant(DAG, 0x3da235e3));
3816 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3817 getF32Constant(DAG, 0x3e65b8f3));
3818 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3819 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3820 getF32Constant(DAG, 0x3f324b07));
3821 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3822 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3823 getF32Constant(DAG, 0x3f7ff8fd));
3824 SDValue t8 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t7);
3825 SDValue TwoToFractionalPartOfX =
3826 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
3828 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3829 MVT::f32, TwoToFractionalPartOfX);
3830 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3831 // For floating-point precision of 18:
3833 // TwoToFractionalPartOfX =
3837 // (0.554906021e-1f +
3838 // (0.961591928e-2f +
3839 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3840 // error 2.47208000*10^(-7), which is better than 18 bits
3841 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3842 getF32Constant(DAG, 0x3924b03e));
3843 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3844 getF32Constant(DAG, 0x3ab24b87));
3845 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3846 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3847 getF32Constant(DAG, 0x3c1d8c17));
3848 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3849 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3850 getF32Constant(DAG, 0x3d634a1d));
3851 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3852 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3853 getF32Constant(DAG, 0x3e75fe14));
3854 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3855 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3856 getF32Constant(DAG, 0x3f317234));
3857 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3858 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3859 getF32Constant(DAG, 0x3f800000));
3860 SDValue t14 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t13);
3861 SDValue TwoToFractionalPartOfX =
3862 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
3864 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3865 MVT::f32, TwoToFractionalPartOfX);
3868 // No special expansion.
3869 result = DAG.getNode(ISD::FPOW, dl,
3870 getValue(I.getArgOperand(0)).getValueType(),
3871 getValue(I.getArgOperand(0)),
3872 getValue(I.getArgOperand(1)));
3875 setValue(&I, result);
3879 /// ExpandPowI - Expand a llvm.powi intrinsic.
3880 static SDValue ExpandPowI(DebugLoc DL, SDValue LHS, SDValue RHS,
3881 SelectionDAG &DAG) {
3882 // If RHS is a constant, we can expand this out to a multiplication tree,
3883 // otherwise we end up lowering to a call to __powidf2 (for example). When
3884 // optimizing for size, we only want to do this if the expansion would produce
3885 // a small number of multiplies, otherwise we do the full expansion.
3886 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3887 // Get the exponent as a positive value.
3888 unsigned Val = RHSC->getSExtValue();
3889 if ((int)Val < 0) Val = -Val;
3891 // powi(x, 0) -> 1.0
3893 return DAG.getConstantFP(1.0, LHS.getValueType());
3895 const Function *F = DAG.getMachineFunction().getFunction();
3896 if (!F->hasFnAttr(Attribute::OptimizeForSize) ||
3897 // If optimizing for size, don't insert too many multiplies. This
3898 // inserts up to 5 multiplies.
3899 CountPopulation_32(Val)+Log2_32(Val) < 7) {
3900 // We use the simple binary decomposition method to generate the multiply
3901 // sequence. There are more optimal ways to do this (for example,
3902 // powi(x,15) generates one more multiply than it should), but this has
3903 // the benefit of being both really simple and much better than a libcall.
3904 SDValue Res; // Logically starts equal to 1.0
3905 SDValue CurSquare = LHS;
3909 Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare);
3911 Res = CurSquare; // 1.0*CurSquare.
3914 CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(),
3915 CurSquare, CurSquare);
3919 // If the original was negative, invert the result, producing 1/(x*x*x).
3920 if (RHSC->getSExtValue() < 0)
3921 Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(),
3922 DAG.getConstantFP(1.0, LHS.getValueType()), Res);
3927 // Otherwise, expand to a libcall.
3928 return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS);
3931 /// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a function
3932 /// argument, create the corresponding DBG_VALUE machine instruction for it now.
3933 /// At the end of instruction selection, they will be inserted to the entry BB.
3935 SelectionDAGBuilder::EmitFuncArgumentDbgValue(const Value *V, MDNode *Variable,
3938 if (!isa<Argument>(V))
3941 MachineFunction &MF = DAG.getMachineFunction();
3942 // Ignore inlined function arguments here.
3943 DIVariable DV(Variable);
3944 if (DV.isInlinedFnArgument(MF.getFunction()))
3947 MachineBasicBlock *MBB = FuncInfo.MBB;
3948 if (MBB != &MF.front())
3952 if (N.getOpcode() == ISD::CopyFromReg) {
3953 Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg();
3954 if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) {
3955 MachineRegisterInfo &RegInfo = MF.getRegInfo();
3956 unsigned PR = RegInfo.getLiveInPhysReg(Reg);
3963 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
3964 if (VMI == FuncInfo.ValueMap.end())
3969 const TargetInstrInfo *TII = DAG.getTarget().getInstrInfo();
3970 MachineInstrBuilder MIB = BuildMI(MF, getCurDebugLoc(),
3971 TII->get(TargetOpcode::DBG_VALUE))
3972 .addReg(Reg, RegState::Debug).addImm(Offset).addMetadata(Variable);
3973 FuncInfo.ArgDbgValues.push_back(&*MIB);
3977 // VisualStudio defines setjmp as _setjmp
3978 #if defined(_MSC_VER) && defined(setjmp)
3979 #define setjmp_undefined_for_visual_studio
3983 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
3984 /// we want to emit this as a call to a named external function, return the name
3985 /// otherwise lower it and return null.
3987 SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) {
3988 DebugLoc dl = getCurDebugLoc();
3991 switch (Intrinsic) {
3993 // By default, turn this into a target intrinsic node.
3994 visitTargetIntrinsic(I, Intrinsic);
3996 case Intrinsic::vastart: visitVAStart(I); return 0;
3997 case Intrinsic::vaend: visitVAEnd(I); return 0;
3998 case Intrinsic::vacopy: visitVACopy(I); return 0;
3999 case Intrinsic::returnaddress:
4000 setValue(&I, DAG.getNode(ISD::RETURNADDR, dl, TLI.getPointerTy(),
4001 getValue(I.getArgOperand(0))));
4003 case Intrinsic::frameaddress:
4004 setValue(&I, DAG.getNode(ISD::FRAMEADDR, dl, TLI.getPointerTy(),
4005 getValue(I.getArgOperand(0))));
4007 case Intrinsic::setjmp:
4008 return "_setjmp"+!TLI.usesUnderscoreSetJmp();
4009 case Intrinsic::longjmp:
4010 return "_longjmp"+!TLI.usesUnderscoreLongJmp();
4011 case Intrinsic::memcpy: {
4012 // Assert for address < 256 since we support only user defined address
4014 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4016 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4018 "Unknown address space");
4019 SDValue Op1 = getValue(I.getArgOperand(0));
4020 SDValue Op2 = getValue(I.getArgOperand(1));
4021 SDValue Op3 = getValue(I.getArgOperand(2));
4022 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4023 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4024 DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, isVol, false,
4025 I.getArgOperand(0), 0, I.getArgOperand(1), 0));
4028 case Intrinsic::memset: {
4029 // Assert for address < 256 since we support only user defined address
4031 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4033 "Unknown address space");
4034 SDValue Op1 = getValue(I.getArgOperand(0));
4035 SDValue Op2 = getValue(I.getArgOperand(1));
4036 SDValue Op3 = getValue(I.getArgOperand(2));
4037 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4038 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4039 DAG.setRoot(DAG.getMemset(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4040 I.getArgOperand(0), 0));
4043 case Intrinsic::memmove: {
4044 // Assert for address < 256 since we support only user defined address
4046 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4048 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4050 "Unknown address space");
4051 SDValue Op1 = getValue(I.getArgOperand(0));
4052 SDValue Op2 = getValue(I.getArgOperand(1));
4053 SDValue Op3 = getValue(I.getArgOperand(2));
4054 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4055 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4057 // If the source and destination are known to not be aliases, we can
4058 // lower memmove as memcpy.
4059 uint64_t Size = -1ULL;
4060 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op3))
4061 Size = C->getZExtValue();
4062 if (AA->alias(I.getArgOperand(0), Size, I.getArgOperand(1), Size) ==
4063 AliasAnalysis::NoAlias) {
4064 DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4065 false, I.getArgOperand(0), 0,
4066 I.getArgOperand(1), 0));
4070 DAG.setRoot(DAG.getMemmove(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4071 I.getArgOperand(0), 0, I.getArgOperand(1), 0));
4074 case Intrinsic::dbg_declare: {
4075 const DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
4076 if (!DIVariable(DI.getVariable()).Verify())
4079 MDNode *Variable = DI.getVariable();
4080 // Parameters are handled specially.
4082 DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable;
4083 const Value *Address = DI.getAddress();
4086 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
4087 Address = BCI->getOperand(0);
4088 const AllocaInst *AI = dyn_cast<AllocaInst>(Address);
4090 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder
4091 // but do not always have a corresponding SDNode built. The SDNodeOrder
4092 // absolute, but not relative, values are different depending on whether
4093 // debug info exists.
4095 SDValue &N = NodeMap[Address];
4098 if (isParameter && !AI) {
4099 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode());
4101 // Byval parameter. We have a frame index at this point.
4102 SDV = DAG.getDbgValue(Variable, FINode->getIndex(),
4103 0, dl, SDNodeOrder);
4105 // Can't do anything with other non-AI cases yet. This might be a
4106 // parameter of a callee function that got inlined, for example.
4109 SDV = DAG.getDbgValue(Variable, N.getNode(), N.getResNo(),
4110 0, dl, SDNodeOrder);
4112 // Can't do anything with other non-AI cases yet.
4114 DAG.AddDbgValue(SDV, N.getNode(), isParameter);
4116 // This isn't useful, but it shows what we're missing.
4117 SDV = DAG.getDbgValue(Variable, UndefValue::get(Address->getType()),
4118 0, dl, SDNodeOrder);
4119 DAG.AddDbgValue(SDV, 0, isParameter);
4123 case Intrinsic::dbg_value: {
4124 const DbgValueInst &DI = cast<DbgValueInst>(I);
4125 if (!DIVariable(DI.getVariable()).Verify())
4128 MDNode *Variable = DI.getVariable();
4129 uint64_t Offset = DI.getOffset();
4130 const Value *V = DI.getValue();
4134 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder
4135 // but do not always have a corresponding SDNode built. The SDNodeOrder
4136 // absolute, but not relative, values are different depending on whether
4137 // debug info exists.
4140 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) {
4141 SDV = DAG.getDbgValue(Variable, V, Offset, dl, SDNodeOrder);
4142 DAG.AddDbgValue(SDV, 0, false);
4144 bool createUndef = false;
4145 // Do not use getValue() in here; we don't want to generate code at
4146 // this point if it hasn't been done yet.
4147 SDValue N = NodeMap[V];
4148 if (!N.getNode() && isa<Argument>(V))
4149 // Check unused arguments map.
4150 N = UnusedArgNodeMap[V];
4152 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, N)) {
4153 SDV = DAG.getDbgValue(Variable, N.getNode(),
4154 N.getResNo(), Offset, dl, SDNodeOrder);
4155 DAG.AddDbgValue(SDV, N.getNode(), false);
4157 } else if (isa<PHINode>(V) && !V->use_empty() ) {
4158 // Do not call getValue(V) yet, as we don't want to generate code.
4159 // Remember it for later.
4160 DanglingDebugInfo DDI(&DI, dl, SDNodeOrder);
4161 DanglingDebugInfoMap[V] = DDI;
4165 // We may expand this to cover more cases. One case where we have no
4166 // data available is an unreferenced parameter; we need this fallback.
4167 SDV = DAG.getDbgValue(Variable, UndefValue::get(V->getType()),
4168 Offset, dl, SDNodeOrder);
4169 DAG.AddDbgValue(SDV, 0, false);
4173 // Build a debug info table entry.
4174 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
4175 V = BCI->getOperand(0);
4176 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
4177 // Don't handle byval struct arguments or VLAs, for example.
4180 DenseMap<const AllocaInst*, int>::iterator SI =
4181 FuncInfo.StaticAllocaMap.find(AI);
4182 if (SI == FuncInfo.StaticAllocaMap.end())
4184 int FI = SI->second;
4186 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4187 if (!DI.getDebugLoc().isUnknown() && MMI.hasDebugInfo())
4188 MMI.setVariableDbgInfo(Variable, FI, DI.getDebugLoc());
4191 case Intrinsic::eh_exception: {
4192 // Insert the EXCEPTIONADDR instruction.
4193 assert(FuncInfo.MBB->isLandingPad() &&
4194 "Call to eh.exception not in landing pad!");
4195 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
4197 Ops[0] = DAG.getRoot();
4198 SDValue Op = DAG.getNode(ISD::EXCEPTIONADDR, dl, VTs, Ops, 1);
4200 DAG.setRoot(Op.getValue(1));
4204 case Intrinsic::eh_selector: {
4205 MachineBasicBlock *CallMBB = FuncInfo.MBB;
4206 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4207 if (CallMBB->isLandingPad())
4208 AddCatchInfo(I, &MMI, CallMBB);
4211 FuncInfo.CatchInfoLost.insert(&I);
4213 // FIXME: Mark exception selector register as live in. Hack for PR1508.
4214 unsigned Reg = TLI.getExceptionSelectorRegister();
4215 if (Reg) FuncInfo.MBB->addLiveIn(Reg);
4218 // Insert the EHSELECTION instruction.
4219 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
4221 Ops[0] = getValue(I.getArgOperand(0));
4223 SDValue Op = DAG.getNode(ISD::EHSELECTION, dl, VTs, Ops, 2);
4224 DAG.setRoot(Op.getValue(1));
4225 setValue(&I, DAG.getSExtOrTrunc(Op, dl, MVT::i32));
4229 case Intrinsic::eh_typeid_for: {
4230 // Find the type id for the given typeinfo.
4231 GlobalVariable *GV = ExtractTypeInfo(I.getArgOperand(0));
4232 unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(GV);
4233 Res = DAG.getConstant(TypeID, MVT::i32);
4238 case Intrinsic::eh_return_i32:
4239 case Intrinsic::eh_return_i64:
4240 DAG.getMachineFunction().getMMI().setCallsEHReturn(true);
4241 DAG.setRoot(DAG.getNode(ISD::EH_RETURN, dl,
4244 getValue(I.getArgOperand(0)),
4245 getValue(I.getArgOperand(1))));
4247 case Intrinsic::eh_unwind_init:
4248 DAG.getMachineFunction().getMMI().setCallsUnwindInit(true);
4250 case Intrinsic::eh_dwarf_cfa: {
4251 SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getArgOperand(0)), dl,
4252 TLI.getPointerTy());
4253 SDValue Offset = DAG.getNode(ISD::ADD, dl,
4255 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, dl,
4256 TLI.getPointerTy()),
4258 SDValue FA = DAG.getNode(ISD::FRAMEADDR, dl,
4260 DAG.getConstant(0, TLI.getPointerTy()));
4261 setValue(&I, DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
4265 case Intrinsic::eh_sjlj_callsite: {
4266 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4267 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(0));
4268 assert(CI && "Non-constant call site value in eh.sjlj.callsite!");
4269 assert(MMI.getCurrentCallSite() == 0 && "Overlapping call sites!");
4271 MMI.setCurrentCallSite(CI->getZExtValue());
4274 case Intrinsic::eh_sjlj_setjmp: {
4275 setValue(&I, DAG.getNode(ISD::EH_SJLJ_SETJMP, dl, MVT::i32, getRoot(),
4276 getValue(I.getArgOperand(0))));
4279 case Intrinsic::eh_sjlj_longjmp: {
4280 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, dl, MVT::Other,
4282 getValue(I.getArgOperand(0))));
4286 case Intrinsic::convertff:
4287 case Intrinsic::convertfsi:
4288 case Intrinsic::convertfui:
4289 case Intrinsic::convertsif:
4290 case Intrinsic::convertuif:
4291 case Intrinsic::convertss:
4292 case Intrinsic::convertsu:
4293 case Intrinsic::convertus:
4294 case Intrinsic::convertuu: {
4295 ISD::CvtCode Code = ISD::CVT_INVALID;
4296 switch (Intrinsic) {
4297 case Intrinsic::convertff: Code = ISD::CVT_FF; break;
4298 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break;
4299 case Intrinsic::convertfui: Code = ISD::CVT_FU; break;
4300 case Intrinsic::convertsif: Code = ISD::CVT_SF; break;
4301 case Intrinsic::convertuif: Code = ISD::CVT_UF; break;
4302 case Intrinsic::convertss: Code = ISD::CVT_SS; break;
4303 case Intrinsic::convertsu: Code = ISD::CVT_SU; break;
4304 case Intrinsic::convertus: Code = ISD::CVT_US; break;
4305 case Intrinsic::convertuu: Code = ISD::CVT_UU; break;
4307 EVT DestVT = TLI.getValueType(I.getType());
4308 const Value *Op1 = I.getArgOperand(0);
4309 Res = DAG.getConvertRndSat(DestVT, getCurDebugLoc(), getValue(Op1),
4310 DAG.getValueType(DestVT),
4311 DAG.getValueType(getValue(Op1).getValueType()),
4312 getValue(I.getArgOperand(1)),
4313 getValue(I.getArgOperand(2)),
4318 case Intrinsic::sqrt:
4319 setValue(&I, DAG.getNode(ISD::FSQRT, dl,
4320 getValue(I.getArgOperand(0)).getValueType(),
4321 getValue(I.getArgOperand(0))));
4323 case Intrinsic::powi:
4324 setValue(&I, ExpandPowI(dl, getValue(I.getArgOperand(0)),
4325 getValue(I.getArgOperand(1)), DAG));
4327 case Intrinsic::sin:
4328 setValue(&I, DAG.getNode(ISD::FSIN, dl,
4329 getValue(I.getArgOperand(0)).getValueType(),
4330 getValue(I.getArgOperand(0))));
4332 case Intrinsic::cos:
4333 setValue(&I, DAG.getNode(ISD::FCOS, dl,
4334 getValue(I.getArgOperand(0)).getValueType(),
4335 getValue(I.getArgOperand(0))));
4337 case Intrinsic::log:
4340 case Intrinsic::log2:
4343 case Intrinsic::log10:
4346 case Intrinsic::exp:
4349 case Intrinsic::exp2:
4352 case Intrinsic::pow:
4355 case Intrinsic::convert_to_fp16:
4356 setValue(&I, DAG.getNode(ISD::FP32_TO_FP16, dl,
4357 MVT::i16, getValue(I.getArgOperand(0))));
4359 case Intrinsic::convert_from_fp16:
4360 setValue(&I, DAG.getNode(ISD::FP16_TO_FP32, dl,
4361 MVT::f32, getValue(I.getArgOperand(0))));
4363 case Intrinsic::pcmarker: {
4364 SDValue Tmp = getValue(I.getArgOperand(0));
4365 DAG.setRoot(DAG.getNode(ISD::PCMARKER, dl, MVT::Other, getRoot(), Tmp));
4368 case Intrinsic::readcyclecounter: {
4369 SDValue Op = getRoot();
4370 Res = DAG.getNode(ISD::READCYCLECOUNTER, dl,
4371 DAG.getVTList(MVT::i64, MVT::Other),
4374 DAG.setRoot(Res.getValue(1));
4377 case Intrinsic::bswap:
4378 setValue(&I, DAG.getNode(ISD::BSWAP, dl,
4379 getValue(I.getArgOperand(0)).getValueType(),
4380 getValue(I.getArgOperand(0))));
4382 case Intrinsic::cttz: {
4383 SDValue Arg = getValue(I.getArgOperand(0));
4384 EVT Ty = Arg.getValueType();
4385 setValue(&I, DAG.getNode(ISD::CTTZ, dl, Ty, Arg));
4388 case Intrinsic::ctlz: {
4389 SDValue Arg = getValue(I.getArgOperand(0));
4390 EVT Ty = Arg.getValueType();
4391 setValue(&I, DAG.getNode(ISD::CTLZ, dl, Ty, Arg));
4394 case Intrinsic::ctpop: {
4395 SDValue Arg = getValue(I.getArgOperand(0));
4396 EVT Ty = Arg.getValueType();
4397 setValue(&I, DAG.getNode(ISD::CTPOP, dl, Ty, Arg));
4400 case Intrinsic::stacksave: {
4401 SDValue Op = getRoot();
4402 Res = DAG.getNode(ISD::STACKSAVE, dl,
4403 DAG.getVTList(TLI.getPointerTy(), MVT::Other), &Op, 1);
4405 DAG.setRoot(Res.getValue(1));
4408 case Intrinsic::stackrestore: {
4409 Res = getValue(I.getArgOperand(0));
4410 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, dl, MVT::Other, getRoot(), Res));
4413 case Intrinsic::stackprotector: {
4414 // Emit code into the DAG to store the stack guard onto the stack.
4415 MachineFunction &MF = DAG.getMachineFunction();
4416 MachineFrameInfo *MFI = MF.getFrameInfo();
4417 EVT PtrTy = TLI.getPointerTy();
4419 SDValue Src = getValue(I.getArgOperand(0)); // The guard's value.
4420 AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
4422 int FI = FuncInfo.StaticAllocaMap[Slot];
4423 MFI->setStackProtectorIndex(FI);
4425 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
4427 // Store the stack protector onto the stack.
4428 Res = DAG.getStore(getRoot(), getCurDebugLoc(), Src, FIN,
4429 PseudoSourceValue::getFixedStack(FI),
4435 case Intrinsic::objectsize: {
4436 // If we don't know by now, we're never going to know.
4437 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1));
4439 assert(CI && "Non-constant type in __builtin_object_size?");
4441 SDValue Arg = getValue(I.getCalledValue());
4442 EVT Ty = Arg.getValueType();
4445 Res = DAG.getConstant(-1ULL, Ty);
4447 Res = DAG.getConstant(0, Ty);
4452 case Intrinsic::var_annotation:
4453 // Discard annotate attributes
4456 case Intrinsic::init_trampoline: {
4457 const Function *F = cast<Function>(I.getArgOperand(1)->stripPointerCasts());
4461 Ops[1] = getValue(I.getArgOperand(0));
4462 Ops[2] = getValue(I.getArgOperand(1));
4463 Ops[3] = getValue(I.getArgOperand(2));
4464 Ops[4] = DAG.getSrcValue(I.getArgOperand(0));
4465 Ops[5] = DAG.getSrcValue(F);
4467 Res = DAG.getNode(ISD::TRAMPOLINE, dl,
4468 DAG.getVTList(TLI.getPointerTy(), MVT::Other),
4472 DAG.setRoot(Res.getValue(1));
4475 case Intrinsic::gcroot:
4477 const Value *Alloca = I.getArgOperand(0);
4478 const Constant *TypeMap = cast<Constant>(I.getArgOperand(1));
4480 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode());
4481 GFI->addStackRoot(FI->getIndex(), TypeMap);
4484 case Intrinsic::gcread:
4485 case Intrinsic::gcwrite:
4486 llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!");
4488 case Intrinsic::flt_rounds:
4489 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, dl, MVT::i32));
4491 case Intrinsic::trap:
4492 DAG.setRoot(DAG.getNode(ISD::TRAP, dl,MVT::Other, getRoot()));
4494 case Intrinsic::uadd_with_overflow:
4495 return implVisitAluOverflow(I, ISD::UADDO);
4496 case Intrinsic::sadd_with_overflow:
4497 return implVisitAluOverflow(I, ISD::SADDO);
4498 case Intrinsic::usub_with_overflow:
4499 return implVisitAluOverflow(I, ISD::USUBO);
4500 case Intrinsic::ssub_with_overflow:
4501 return implVisitAluOverflow(I, ISD::SSUBO);
4502 case Intrinsic::umul_with_overflow:
4503 return implVisitAluOverflow(I, ISD::UMULO);
4504 case Intrinsic::smul_with_overflow:
4505 return implVisitAluOverflow(I, ISD::SMULO);
4507 case Intrinsic::prefetch: {
4510 Ops[1] = getValue(I.getArgOperand(0));
4511 Ops[2] = getValue(I.getArgOperand(1));
4512 Ops[3] = getValue(I.getArgOperand(2));
4513 DAG.setRoot(DAG.getNode(ISD::PREFETCH, dl, MVT::Other, &Ops[0], 4));
4517 case Intrinsic::memory_barrier: {
4520 for (int x = 1; x < 6; ++x)
4521 Ops[x] = getValue(I.getArgOperand(x - 1));
4523 DAG.setRoot(DAG.getNode(ISD::MEMBARRIER, dl, MVT::Other, &Ops[0], 6));
4526 case Intrinsic::atomic_cmp_swap: {
4527 SDValue Root = getRoot();
4529 DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, getCurDebugLoc(),
4530 getValue(I.getArgOperand(1)).getValueType().getSimpleVT(),
4532 getValue(I.getArgOperand(0)),
4533 getValue(I.getArgOperand(1)),
4534 getValue(I.getArgOperand(2)),
4535 I.getArgOperand(0));
4537 DAG.setRoot(L.getValue(1));
4540 case Intrinsic::atomic_load_add:
4541 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_ADD);
4542 case Intrinsic::atomic_load_sub:
4543 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_SUB);
4544 case Intrinsic::atomic_load_or:
4545 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_OR);
4546 case Intrinsic::atomic_load_xor:
4547 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_XOR);
4548 case Intrinsic::atomic_load_and:
4549 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_AND);
4550 case Intrinsic::atomic_load_nand:
4551 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_NAND);
4552 case Intrinsic::atomic_load_max:
4553 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MAX);
4554 case Intrinsic::atomic_load_min:
4555 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MIN);
4556 case Intrinsic::atomic_load_umin:
4557 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMIN);
4558 case Intrinsic::atomic_load_umax:
4559 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMAX);
4560 case Intrinsic::atomic_swap:
4561 return implVisitBinaryAtomic(I, ISD::ATOMIC_SWAP);
4563 case Intrinsic::invariant_start:
4564 case Intrinsic::lifetime_start:
4565 // Discard region information.
4566 setValue(&I, DAG.getUNDEF(TLI.getPointerTy()));
4568 case Intrinsic::invariant_end:
4569 case Intrinsic::lifetime_end:
4570 // Discard region information.
4575 void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee,
4577 MachineBasicBlock *LandingPad) {
4578 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
4579 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
4580 const Type *RetTy = FTy->getReturnType();
4581 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4582 MCSymbol *BeginLabel = 0;
4584 TargetLowering::ArgListTy Args;
4585 TargetLowering::ArgListEntry Entry;
4586 Args.reserve(CS.arg_size());
4588 // Check whether the function can return without sret-demotion.
4589 SmallVector<ISD::OutputArg, 4> Outs;
4590 SmallVector<uint64_t, 4> Offsets;
4591 GetReturnInfo(RetTy, CS.getAttributes().getRetAttributes(),
4592 Outs, TLI, &Offsets);
4594 bool CanLowerReturn = TLI.CanLowerReturn(CS.getCallingConv(),
4595 FTy->isVarArg(), Outs, FTy->getContext());
4597 SDValue DemoteStackSlot;
4599 if (!CanLowerReturn) {
4600 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(
4601 FTy->getReturnType());
4602 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(
4603 FTy->getReturnType());
4604 MachineFunction &MF = DAG.getMachineFunction();
4605 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
4606 const Type *StackSlotPtrType = PointerType::getUnqual(FTy->getReturnType());
4608 DemoteStackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
4609 Entry.Node = DemoteStackSlot;
4610 Entry.Ty = StackSlotPtrType;
4611 Entry.isSExt = false;
4612 Entry.isZExt = false;
4613 Entry.isInReg = false;
4614 Entry.isSRet = true;
4615 Entry.isNest = false;
4616 Entry.isByVal = false;
4617 Entry.Alignment = Align;
4618 Args.push_back(Entry);
4619 RetTy = Type::getVoidTy(FTy->getContext());
4622 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
4624 SDValue ArgNode = getValue(*i);
4625 Entry.Node = ArgNode; Entry.Ty = (*i)->getType();
4627 unsigned attrInd = i - CS.arg_begin() + 1;
4628 Entry.isSExt = CS.paramHasAttr(attrInd, Attribute::SExt);
4629 Entry.isZExt = CS.paramHasAttr(attrInd, Attribute::ZExt);
4630 Entry.isInReg = CS.paramHasAttr(attrInd, Attribute::InReg);
4631 Entry.isSRet = CS.paramHasAttr(attrInd, Attribute::StructRet);
4632 Entry.isNest = CS.paramHasAttr(attrInd, Attribute::Nest);
4633 Entry.isByVal = CS.paramHasAttr(attrInd, Attribute::ByVal);
4634 Entry.Alignment = CS.getParamAlignment(attrInd);
4635 Args.push_back(Entry);
4639 // Insert a label before the invoke call to mark the try range. This can be
4640 // used to detect deletion of the invoke via the MachineModuleInfo.
4641 BeginLabel = MMI.getContext().CreateTempSymbol();
4643 // For SjLj, keep track of which landing pads go with which invokes
4644 // so as to maintain the ordering of pads in the LSDA.
4645 unsigned CallSiteIndex = MMI.getCurrentCallSite();
4646 if (CallSiteIndex) {
4647 MMI.setCallSiteBeginLabel(BeginLabel, CallSiteIndex);
4648 // Now that the call site is handled, stop tracking it.
4649 MMI.setCurrentCallSite(0);
4652 // Both PendingLoads and PendingExports must be flushed here;
4653 // this call might not return.
4655 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getControlRoot(), BeginLabel));
4658 // Check if target-independent constraints permit a tail call here.
4659 // Target-dependent constraints are checked within TLI.LowerCallTo.
4661 !isInTailCallPosition(CS, CS.getAttributes().getRetAttributes(), TLI))
4664 std::pair<SDValue,SDValue> Result =
4665 TLI.LowerCallTo(getRoot(), RetTy,
4666 CS.paramHasAttr(0, Attribute::SExt),
4667 CS.paramHasAttr(0, Attribute::ZExt), FTy->isVarArg(),
4668 CS.paramHasAttr(0, Attribute::InReg), FTy->getNumParams(),
4669 CS.getCallingConv(),
4671 !CS.getInstruction()->use_empty(),
4672 Callee, Args, DAG, getCurDebugLoc());
4673 assert((isTailCall || Result.second.getNode()) &&
4674 "Non-null chain expected with non-tail call!");
4675 assert((Result.second.getNode() || !Result.first.getNode()) &&
4676 "Null value expected with tail call!");
4677 if (Result.first.getNode()) {
4678 setValue(CS.getInstruction(), Result.first);
4679 } else if (!CanLowerReturn && Result.second.getNode()) {
4680 // The instruction result is the result of loading from the
4681 // hidden sret parameter.
4682 SmallVector<EVT, 1> PVTs;
4683 const Type *PtrRetTy = PointerType::getUnqual(FTy->getReturnType());
4685 ComputeValueVTs(TLI, PtrRetTy, PVTs);
4686 assert(PVTs.size() == 1 && "Pointers should fit in one register");
4687 EVT PtrVT = PVTs[0];
4688 unsigned NumValues = Outs.size();
4689 SmallVector<SDValue, 4> Values(NumValues);
4690 SmallVector<SDValue, 4> Chains(NumValues);
4692 for (unsigned i = 0; i < NumValues; ++i) {
4693 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT,
4695 DAG.getConstant(Offsets[i], PtrVT));
4696 SDValue L = DAG.getLoad(Outs[i].VT, getCurDebugLoc(), Result.second,
4697 Add, NULL, Offsets[i], false, false, 1);
4699 Chains[i] = L.getValue(1);
4702 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
4703 MVT::Other, &Chains[0], NumValues);
4704 PendingLoads.push_back(Chain);
4706 // Collect the legal value parts into potentially illegal values
4707 // that correspond to the original function's return values.
4708 SmallVector<EVT, 4> RetTys;
4709 RetTy = FTy->getReturnType();
4710 ComputeValueVTs(TLI, RetTy, RetTys);
4711 ISD::NodeType AssertOp = ISD::DELETED_NODE;
4712 SmallVector<SDValue, 4> ReturnValues;
4713 unsigned CurReg = 0;
4714 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
4716 EVT RegisterVT = TLI.getRegisterType(RetTy->getContext(), VT);
4717 unsigned NumRegs = TLI.getNumRegisters(RetTy->getContext(), VT);
4719 SDValue ReturnValue =
4720 getCopyFromParts(DAG, getCurDebugLoc(), &Values[CurReg], NumRegs,
4721 RegisterVT, VT, AssertOp);
4722 ReturnValues.push_back(ReturnValue);
4726 setValue(CS.getInstruction(),
4727 DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
4728 DAG.getVTList(&RetTys[0], RetTys.size()),
4729 &ReturnValues[0], ReturnValues.size()));
4733 // As a special case, a null chain means that a tail call has been emitted and
4734 // the DAG root is already updated.
4735 if (Result.second.getNode())
4736 DAG.setRoot(Result.second);
4741 // Insert a label at the end of the invoke call to mark the try range. This
4742 // can be used to detect deletion of the invoke via the MachineModuleInfo.
4743 MCSymbol *EndLabel = MMI.getContext().CreateTempSymbol();
4744 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getRoot(), EndLabel));
4746 // Inform MachineModuleInfo of range.
4747 MMI.addInvoke(LandingPad, BeginLabel, EndLabel);
4751 /// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the
4752 /// value is equal or not-equal to zero.
4753 static bool IsOnlyUsedInZeroEqualityComparison(const Value *V) {
4754 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end();
4756 if (const ICmpInst *IC = dyn_cast<ICmpInst>(*UI))
4757 if (IC->isEquality())
4758 if (const Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
4759 if (C->isNullValue())
4761 // Unknown instruction.
4767 static SDValue getMemCmpLoad(const Value *PtrVal, MVT LoadVT,
4769 SelectionDAGBuilder &Builder) {
4771 // Check to see if this load can be trivially constant folded, e.g. if the
4772 // input is from a string literal.
4773 if (const Constant *LoadInput = dyn_cast<Constant>(PtrVal)) {
4774 // Cast pointer to the type we really want to load.
4775 LoadInput = ConstantExpr::getBitCast(const_cast<Constant *>(LoadInput),
4776 PointerType::getUnqual(LoadTy));
4778 if (const Constant *LoadCst =
4779 ConstantFoldLoadFromConstPtr(const_cast<Constant *>(LoadInput),
4781 return Builder.getValue(LoadCst);
4784 // Otherwise, we have to emit the load. If the pointer is to unfoldable but
4785 // still constant memory, the input chain can be the entry node.
4787 bool ConstantMemory = false;
4789 // Do not serialize (non-volatile) loads of constant memory with anything.
4790 if (Builder.AA->pointsToConstantMemory(PtrVal)) {
4791 Root = Builder.DAG.getEntryNode();
4792 ConstantMemory = true;
4794 // Do not serialize non-volatile loads against each other.
4795 Root = Builder.DAG.getRoot();
4798 SDValue Ptr = Builder.getValue(PtrVal);
4799 SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurDebugLoc(), Root,
4800 Ptr, PtrVal /*SrcValue*/, 0/*SVOffset*/,
4802 false /*nontemporal*/, 1 /* align=1 */);
4804 if (!ConstantMemory)
4805 Builder.PendingLoads.push_back(LoadVal.getValue(1));
4810 /// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form.
4811 /// If so, return true and lower it, otherwise return false and it will be
4812 /// lowered like a normal call.
4813 bool SelectionDAGBuilder::visitMemCmpCall(const CallInst &I) {
4814 // Verify that the prototype makes sense. int memcmp(void*,void*,size_t)
4815 if (I.getNumArgOperands() != 3)
4818 const Value *LHS = I.getArgOperand(0), *RHS = I.getArgOperand(1);
4819 if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() ||
4820 !I.getArgOperand(2)->getType()->isIntegerTy() ||
4821 !I.getType()->isIntegerTy())
4824 const ConstantInt *Size = dyn_cast<ConstantInt>(I.getArgOperand(2));
4826 // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0
4827 // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0
4828 if (Size && IsOnlyUsedInZeroEqualityComparison(&I)) {
4829 bool ActuallyDoIt = true;
4832 switch (Size->getZExtValue()) {
4834 LoadVT = MVT::Other;
4836 ActuallyDoIt = false;
4840 LoadTy = Type::getInt16Ty(Size->getContext());
4844 LoadTy = Type::getInt32Ty(Size->getContext());
4848 LoadTy = Type::getInt64Ty(Size->getContext());
4852 LoadVT = MVT::v4i32;
4853 LoadTy = Type::getInt32Ty(Size->getContext());
4854 LoadTy = VectorType::get(LoadTy, 4);
4859 // This turns into unaligned loads. We only do this if the target natively
4860 // supports the MVT we'll be loading or if it is small enough (<= 4) that
4861 // we'll only produce a small number of byte loads.
4863 // Require that we can find a legal MVT, and only do this if the target
4864 // supports unaligned loads of that type. Expanding into byte loads would
4866 if (ActuallyDoIt && Size->getZExtValue() > 4) {
4867 // TODO: Handle 5 byte compare as 4-byte + 1 byte.
4868 // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads.
4869 if (!TLI.isTypeLegal(LoadVT) ||!TLI.allowsUnalignedMemoryAccesses(LoadVT))
4870 ActuallyDoIt = false;
4874 SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this);
4875 SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this);
4877 SDValue Res = DAG.getSetCC(getCurDebugLoc(), MVT::i1, LHSVal, RHSVal,
4879 EVT CallVT = TLI.getValueType(I.getType(), true);
4880 setValue(&I, DAG.getZExtOrTrunc(Res, getCurDebugLoc(), CallVT));
4890 void SelectionDAGBuilder::visitCall(const CallInst &I) {
4891 // Handle inline assembly differently.
4892 if (isa<InlineAsm>(I.getCalledValue())) {
4897 const char *RenameFn = 0;
4898 if (Function *F = I.getCalledFunction()) {
4899 if (F->isDeclaration()) {
4900 if (const TargetIntrinsicInfo *II = TM.getIntrinsicInfo()) {
4901 if (unsigned IID = II->getIntrinsicID(F)) {
4902 RenameFn = visitIntrinsicCall(I, IID);
4907 if (unsigned IID = F->getIntrinsicID()) {
4908 RenameFn = visitIntrinsicCall(I, IID);
4914 // Check for well-known libc/libm calls. If the function is internal, it
4915 // can't be a library call.
4916 if (!F->hasLocalLinkage() && F->hasName()) {
4917 StringRef Name = F->getName();
4918 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl") {
4919 if (I.getNumArgOperands() == 2 && // Basic sanity checks.
4920 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
4921 I.getType() == I.getArgOperand(0)->getType() &&
4922 I.getType() == I.getArgOperand(1)->getType()) {
4923 SDValue LHS = getValue(I.getArgOperand(0));
4924 SDValue RHS = getValue(I.getArgOperand(1));
4925 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurDebugLoc(),
4926 LHS.getValueType(), LHS, RHS));
4929 } else if (Name == "fabs" || Name == "fabsf" || Name == "fabsl") {
4930 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
4931 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
4932 I.getType() == I.getArgOperand(0)->getType()) {
4933 SDValue Tmp = getValue(I.getArgOperand(0));
4934 setValue(&I, DAG.getNode(ISD::FABS, getCurDebugLoc(),
4935 Tmp.getValueType(), Tmp));
4938 } else if (Name == "sin" || Name == "sinf" || Name == "sinl") {
4939 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
4940 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
4941 I.getType() == I.getArgOperand(0)->getType() &&
4942 I.onlyReadsMemory()) {
4943 SDValue Tmp = getValue(I.getArgOperand(0));
4944 setValue(&I, DAG.getNode(ISD::FSIN, getCurDebugLoc(),
4945 Tmp.getValueType(), Tmp));
4948 } else if (Name == "cos" || Name == "cosf" || Name == "cosl") {
4949 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
4950 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
4951 I.getType() == I.getArgOperand(0)->getType() &&
4952 I.onlyReadsMemory()) {
4953 SDValue Tmp = getValue(I.getArgOperand(0));
4954 setValue(&I, DAG.getNode(ISD::FCOS, getCurDebugLoc(),
4955 Tmp.getValueType(), Tmp));
4958 } else if (Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl") {
4959 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
4960 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
4961 I.getType() == I.getArgOperand(0)->getType() &&
4962 I.onlyReadsMemory()) {
4963 SDValue Tmp = getValue(I.getArgOperand(0));
4964 setValue(&I, DAG.getNode(ISD::FSQRT, getCurDebugLoc(),
4965 Tmp.getValueType(), Tmp));
4968 } else if (Name == "memcmp") {
4969 if (visitMemCmpCall(I))
4977 Callee = getValue(I.getCalledValue());
4979 Callee = DAG.getExternalSymbol(RenameFn, TLI.getPointerTy());
4981 // Check if we can potentially perform a tail call. More detailed checking is
4982 // be done within LowerCallTo, after more information about the call is known.
4983 LowerCallTo(&I, Callee, I.isTailCall());
4988 /// AsmOperandInfo - This contains information for each constraint that we are
4990 class LLVM_LIBRARY_VISIBILITY SDISelAsmOperandInfo :
4991 public TargetLowering::AsmOperandInfo {
4993 /// CallOperand - If this is the result output operand or a clobber
4994 /// this is null, otherwise it is the incoming operand to the CallInst.
4995 /// This gets modified as the asm is processed.
4996 SDValue CallOperand;
4998 /// AssignedRegs - If this is a register or register class operand, this
4999 /// contains the set of register corresponding to the operand.
5000 RegsForValue AssignedRegs;
5002 explicit SDISelAsmOperandInfo(const InlineAsm::ConstraintInfo &info)
5003 : TargetLowering::AsmOperandInfo(info), CallOperand(0,0) {
5006 /// MarkAllocatedRegs - Once AssignedRegs is set, mark the assigned registers
5007 /// busy in OutputRegs/InputRegs.
5008 void MarkAllocatedRegs(bool isOutReg, bool isInReg,
5009 std::set<unsigned> &OutputRegs,
5010 std::set<unsigned> &InputRegs,
5011 const TargetRegisterInfo &TRI) const {
5013 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
5014 MarkRegAndAliases(AssignedRegs.Regs[i], OutputRegs, TRI);
5017 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
5018 MarkRegAndAliases(AssignedRegs.Regs[i], InputRegs, TRI);
5022 /// getCallOperandValEVT - Return the EVT of the Value* that this operand
5023 /// corresponds to. If there is no Value* for this operand, it returns
5025 EVT getCallOperandValEVT(LLVMContext &Context,
5026 const TargetLowering &TLI,
5027 const TargetData *TD) const {
5028 if (CallOperandVal == 0) return MVT::Other;
5030 if (isa<BasicBlock>(CallOperandVal))
5031 return TLI.getPointerTy();
5033 const llvm::Type *OpTy = CallOperandVal->getType();
5035 // If this is an indirect operand, the operand is a pointer to the
5038 const llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
5040 report_fatal_error("Indirect operand for inline asm not a pointer!");
5041 OpTy = PtrTy->getElementType();
5044 // If OpTy is not a single value, it may be a struct/union that we
5045 // can tile with integers.
5046 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
5047 unsigned BitSize = TD->getTypeSizeInBits(OpTy);
5056 OpTy = IntegerType::get(Context, BitSize);
5061 return TLI.getValueType(OpTy, true);
5065 /// MarkRegAndAliases - Mark the specified register and all aliases in the
5067 static void MarkRegAndAliases(unsigned Reg, std::set<unsigned> &Regs,
5068 const TargetRegisterInfo &TRI) {
5069 assert(TargetRegisterInfo::isPhysicalRegister(Reg) && "Isn't a physreg");
5071 if (const unsigned *Aliases = TRI.getAliasSet(Reg))
5072 for (; *Aliases; ++Aliases)
5073 Regs.insert(*Aliases);
5077 } // end llvm namespace.
5079 /// isAllocatableRegister - If the specified register is safe to allocate,
5080 /// i.e. it isn't a stack pointer or some other special register, return the
5081 /// register class for the register. Otherwise, return null.
5082 static const TargetRegisterClass *
5083 isAllocatableRegister(unsigned Reg, MachineFunction &MF,
5084 const TargetLowering &TLI,
5085 const TargetRegisterInfo *TRI) {
5086 EVT FoundVT = MVT::Other;
5087 const TargetRegisterClass *FoundRC = 0;
5088 for (TargetRegisterInfo::regclass_iterator RCI = TRI->regclass_begin(),
5089 E = TRI->regclass_end(); RCI != E; ++RCI) {
5090 EVT ThisVT = MVT::Other;
5092 const TargetRegisterClass *RC = *RCI;
5093 // If none of the value types for this register class are valid, we
5094 // can't use it. For example, 64-bit reg classes on 32-bit targets.
5095 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
5097 if (TLI.isTypeLegal(*I)) {
5098 // If we have already found this register in a different register class,
5099 // choose the one with the largest VT specified. For example, on
5100 // PowerPC, we favor f64 register classes over f32.
5101 if (FoundVT == MVT::Other || FoundVT.bitsLT(*I)) {
5108 if (ThisVT == MVT::Other) continue;
5110 // NOTE: This isn't ideal. In particular, this might allocate the
5111 // frame pointer in functions that need it (due to them not being taken
5112 // out of allocation, because a variable sized allocation hasn't been seen
5113 // yet). This is a slight code pessimization, but should still work.
5114 for (TargetRegisterClass::iterator I = RC->allocation_order_begin(MF),
5115 E = RC->allocation_order_end(MF); I != E; ++I)
5117 // We found a matching register class. Keep looking at others in case
5118 // we find one with larger registers that this physreg is also in.
5127 /// GetRegistersForValue - Assign registers (virtual or physical) for the
5128 /// specified operand. We prefer to assign virtual registers, to allow the
5129 /// register allocator to handle the assignment process. However, if the asm
5130 /// uses features that we can't model on machineinstrs, we have SDISel do the
5131 /// allocation. This produces generally horrible, but correct, code.
5133 /// OpInfo describes the operand.
5134 /// Input and OutputRegs are the set of already allocated physical registers.
5136 void SelectionDAGBuilder::
5137 GetRegistersForValue(SDISelAsmOperandInfo &OpInfo,
5138 std::set<unsigned> &OutputRegs,
5139 std::set<unsigned> &InputRegs) {
5140 LLVMContext &Context = FuncInfo.Fn->getContext();
5142 // Compute whether this value requires an input register, an output register,
5144 bool isOutReg = false;
5145 bool isInReg = false;
5146 switch (OpInfo.Type) {
5147 case InlineAsm::isOutput:
5150 // If there is an input constraint that matches this, we need to reserve
5151 // the input register so no other inputs allocate to it.
5152 isInReg = OpInfo.hasMatchingInput();
5154 case InlineAsm::isInput:
5158 case InlineAsm::isClobber:
5165 MachineFunction &MF = DAG.getMachineFunction();
5166 SmallVector<unsigned, 4> Regs;
5168 // If this is a constraint for a single physreg, or a constraint for a
5169 // register class, find it.
5170 std::pair<unsigned, const TargetRegisterClass*> PhysReg =
5171 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
5172 OpInfo.ConstraintVT);
5174 unsigned NumRegs = 1;
5175 if (OpInfo.ConstraintVT != MVT::Other) {
5176 // If this is a FP input in an integer register (or visa versa) insert a bit
5177 // cast of the input value. More generally, handle any case where the input
5178 // value disagrees with the register class we plan to stick this in.
5179 if (OpInfo.Type == InlineAsm::isInput &&
5180 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) {
5181 // Try to convert to the first EVT that the reg class contains. If the
5182 // types are identical size, use a bitcast to convert (e.g. two differing
5184 EVT RegVT = *PhysReg.second->vt_begin();
5185 if (RegVT.getSizeInBits() == OpInfo.ConstraintVT.getSizeInBits()) {
5186 OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
5187 RegVT, OpInfo.CallOperand);
5188 OpInfo.ConstraintVT = RegVT;
5189 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) {
5190 // If the input is a FP value and we want it in FP registers, do a
5191 // bitcast to the corresponding integer type. This turns an f64 value
5192 // into i64, which can be passed with two i32 values on a 32-bit
5194 RegVT = EVT::getIntegerVT(Context,
5195 OpInfo.ConstraintVT.getSizeInBits());
5196 OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
5197 RegVT, OpInfo.CallOperand);
5198 OpInfo.ConstraintVT = RegVT;
5202 NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT);
5206 EVT ValueVT = OpInfo.ConstraintVT;
5208 // If this is a constraint for a specific physical register, like {r17},
5210 if (unsigned AssignedReg = PhysReg.first) {
5211 const TargetRegisterClass *RC = PhysReg.second;
5212 if (OpInfo.ConstraintVT == MVT::Other)
5213 ValueVT = *RC->vt_begin();
5215 // Get the actual register value type. This is important, because the user
5216 // may have asked for (e.g.) the AX register in i32 type. We need to
5217 // remember that AX is actually i16 to get the right extension.
5218 RegVT = *RC->vt_begin();
5220 // This is a explicit reference to a physical register.
5221 Regs.push_back(AssignedReg);
5223 // If this is an expanded reference, add the rest of the regs to Regs.
5225 TargetRegisterClass::iterator I = RC->begin();
5226 for (; *I != AssignedReg; ++I)
5227 assert(I != RC->end() && "Didn't find reg!");
5229 // Already added the first reg.
5231 for (; NumRegs; --NumRegs, ++I) {
5232 assert(I != RC->end() && "Ran out of registers to allocate!");
5237 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5238 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
5239 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
5243 // Otherwise, if this was a reference to an LLVM register class, create vregs
5244 // for this reference.
5245 if (const TargetRegisterClass *RC = PhysReg.second) {
5246 RegVT = *RC->vt_begin();
5247 if (OpInfo.ConstraintVT == MVT::Other)
5250 // Create the appropriate number of virtual registers.
5251 MachineRegisterInfo &RegInfo = MF.getRegInfo();
5252 for (; NumRegs; --NumRegs)
5253 Regs.push_back(RegInfo.createVirtualRegister(RC));
5255 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5259 // This is a reference to a register class that doesn't directly correspond
5260 // to an LLVM register class. Allocate NumRegs consecutive, available,
5261 // registers from the class.
5262 std::vector<unsigned> RegClassRegs
5263 = TLI.getRegClassForInlineAsmConstraint(OpInfo.ConstraintCode,
5264 OpInfo.ConstraintVT);
5266 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
5267 unsigned NumAllocated = 0;
5268 for (unsigned i = 0, e = RegClassRegs.size(); i != e; ++i) {
5269 unsigned Reg = RegClassRegs[i];
5270 // See if this register is available.
5271 if ((isOutReg && OutputRegs.count(Reg)) || // Already used.
5272 (isInReg && InputRegs.count(Reg))) { // Already used.
5273 // Make sure we find consecutive registers.
5278 // Check to see if this register is allocatable (i.e. don't give out the
5280 const TargetRegisterClass *RC = isAllocatableRegister(Reg, MF, TLI, TRI);
5281 if (!RC) { // Couldn't allocate this register.
5282 // Reset NumAllocated to make sure we return consecutive registers.
5287 // Okay, this register is good, we can use it.
5290 // If we allocated enough consecutive registers, succeed.
5291 if (NumAllocated == NumRegs) {
5292 unsigned RegStart = (i-NumAllocated)+1;
5293 unsigned RegEnd = i+1;
5294 // Mark all of the allocated registers used.
5295 for (unsigned i = RegStart; i != RegEnd; ++i)
5296 Regs.push_back(RegClassRegs[i]);
5298 OpInfo.AssignedRegs = RegsForValue(Regs, *RC->vt_begin(),
5299 OpInfo.ConstraintVT);
5300 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
5305 // Otherwise, we couldn't allocate enough registers for this.
5308 /// visitInlineAsm - Handle a call to an InlineAsm object.
5310 void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) {
5311 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
5313 /// ConstraintOperands - Information about all of the constraints.
5314 std::vector<SDISelAsmOperandInfo> ConstraintOperands;
5316 std::set<unsigned> OutputRegs, InputRegs;
5318 // Do a prepass over the constraints, canonicalizing them, and building up the
5319 // ConstraintOperands list.
5320 std::vector<InlineAsm::ConstraintInfo>
5321 ConstraintInfos = IA->ParseConstraints();
5323 bool hasMemory = hasInlineAsmMemConstraint(ConstraintInfos, TLI);
5325 SDValue Chain, Flag;
5327 // We won't need to flush pending loads if this asm doesn't touch
5328 // memory and is nonvolatile.
5329 if (hasMemory || IA->hasSideEffects())
5332 Chain = DAG.getRoot();
5334 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
5335 unsigned ResNo = 0; // ResNo - The result number of the next output.
5336 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
5337 ConstraintOperands.push_back(SDISelAsmOperandInfo(ConstraintInfos[i]));
5338 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
5340 EVT OpVT = MVT::Other;
5342 // Compute the value type for each operand.
5343 switch (OpInfo.Type) {
5344 case InlineAsm::isOutput:
5345 // Indirect outputs just consume an argument.
5346 if (OpInfo.isIndirect) {
5347 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5351 // The return value of the call is this value. As such, there is no
5352 // corresponding argument.
5353 assert(!CS.getType()->isVoidTy() &&
5355 if (const StructType *STy = dyn_cast<StructType>(CS.getType())) {
5356 OpVT = TLI.getValueType(STy->getElementType(ResNo));
5358 assert(ResNo == 0 && "Asm only has one result!");
5359 OpVT = TLI.getValueType(CS.getType());
5363 case InlineAsm::isInput:
5364 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5366 case InlineAsm::isClobber:
5371 // If this is an input or an indirect output, process the call argument.
5372 // BasicBlocks are labels, currently appearing only in asm's.
5373 if (OpInfo.CallOperandVal) {
5374 // Strip bitcasts, if any. This mostly comes up for functions.
5375 OpInfo.CallOperandVal = OpInfo.CallOperandVal->stripPointerCasts();
5377 if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) {
5378 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]);
5380 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
5383 OpVT = OpInfo.getCallOperandValEVT(*DAG.getContext(), TLI, TD);
5386 OpInfo.ConstraintVT = OpVT;
5389 // Second pass over the constraints: compute which constraint option to use
5390 // and assign registers to constraints that want a specific physreg.
5391 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
5392 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5394 // If this is an output operand with a matching input operand, look up the
5395 // matching input. If their types mismatch, e.g. one is an integer, the
5396 // other is floating point, or their sizes are different, flag it as an
5398 if (OpInfo.hasMatchingInput()) {
5399 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
5401 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
5402 if ((OpInfo.ConstraintVT.isInteger() !=
5403 Input.ConstraintVT.isInteger()) ||
5404 (OpInfo.ConstraintVT.getSizeInBits() !=
5405 Input.ConstraintVT.getSizeInBits())) {
5406 report_fatal_error("Unsupported asm: input constraint"
5407 " with a matching output constraint of"
5408 " incompatible type!");
5410 Input.ConstraintVT = OpInfo.ConstraintVT;
5414 // Compute the constraint code and ConstraintType to use.
5415 TLI.ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, &DAG);
5417 // If this is a memory input, and if the operand is not indirect, do what we
5418 // need to to provide an address for the memory input.
5419 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5420 !OpInfo.isIndirect) {
5421 assert(OpInfo.Type == InlineAsm::isInput &&
5422 "Can only indirectify direct input operands!");
5424 // Memory operands really want the address of the value. If we don't have
5425 // an indirect input, put it in the constpool if we can, otherwise spill
5426 // it to a stack slot.
5428 // If the operand is a float, integer, or vector constant, spill to a
5429 // constant pool entry to get its address.
5430 const Value *OpVal = OpInfo.CallOperandVal;
5431 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
5432 isa<ConstantVector>(OpVal)) {
5433 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal),
5434 TLI.getPointerTy());
5436 // Otherwise, create a stack slot and emit a store to it before the
5438 const Type *Ty = OpVal->getType();
5439 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
5440 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(Ty);
5441 MachineFunction &MF = DAG.getMachineFunction();
5442 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
5443 SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
5444 Chain = DAG.getStore(Chain, getCurDebugLoc(),
5445 OpInfo.CallOperand, StackSlot, NULL, 0,
5447 OpInfo.CallOperand = StackSlot;
5450 // There is no longer a Value* corresponding to this operand.
5451 OpInfo.CallOperandVal = 0;
5453 // It is now an indirect operand.
5454 OpInfo.isIndirect = true;
5457 // If this constraint is for a specific register, allocate it before
5459 if (OpInfo.ConstraintType == TargetLowering::C_Register)
5460 GetRegistersForValue(OpInfo, OutputRegs, InputRegs);
5463 ConstraintInfos.clear();
5465 // Second pass - Loop over all of the operands, assigning virtual or physregs
5466 // to register class operands.
5467 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5468 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5470 // C_Register operands have already been allocated, Other/Memory don't need
5472 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
5473 GetRegistersForValue(OpInfo, OutputRegs, InputRegs);
5476 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
5477 std::vector<SDValue> AsmNodeOperands;
5478 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain
5479 AsmNodeOperands.push_back(
5480 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(),
5481 TLI.getPointerTy()));
5483 // If we have a !srcloc metadata node associated with it, we want to attach
5484 // this to the ultimately generated inline asm machineinstr. To do this, we
5485 // pass in the third operand as this (potentially null) inline asm MDNode.
5486 const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc");
5487 AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc));
5489 // Remember the AlignStack bit as operand 3.
5490 AsmNodeOperands.push_back(DAG.getTargetConstant(IA->isAlignStack() ? 1 : 0,
5493 // Loop over all of the inputs, copying the operand values into the
5494 // appropriate registers and processing the output regs.
5495 RegsForValue RetValRegs;
5497 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
5498 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
5500 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5501 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5503 switch (OpInfo.Type) {
5504 case InlineAsm::isOutput: {
5505 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
5506 OpInfo.ConstraintType != TargetLowering::C_Register) {
5507 // Memory output, or 'other' output (e.g. 'X' constraint).
5508 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
5510 // Add information to the INLINEASM node to know about this output.
5511 unsigned OpFlags = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
5512 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlags,
5513 TLI.getPointerTy()));
5514 AsmNodeOperands.push_back(OpInfo.CallOperand);
5518 // Otherwise, this is a register or register class output.
5520 // Copy the output from the appropriate register. Find a register that
5522 if (OpInfo.AssignedRegs.Regs.empty())
5523 report_fatal_error("Couldn't allocate output reg for constraint '" +
5524 Twine(OpInfo.ConstraintCode) + "'!");
5526 // If this is an indirect operand, store through the pointer after the
5528 if (OpInfo.isIndirect) {
5529 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
5530 OpInfo.CallOperandVal));
5532 // This is the result value of the call.
5533 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
5534 // Concatenate this output onto the outputs list.
5535 RetValRegs.append(OpInfo.AssignedRegs);
5538 // Add information to the INLINEASM node to know that this register is
5540 OpInfo.AssignedRegs.AddInlineAsmOperands(OpInfo.isEarlyClobber ?
5541 InlineAsm::Kind_RegDefEarlyClobber :
5542 InlineAsm::Kind_RegDef,
5549 case InlineAsm::isInput: {
5550 SDValue InOperandVal = OpInfo.CallOperand;
5552 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint?
5553 // If this is required to match an output register we have already set,
5554 // just use its register.
5555 unsigned OperandNo = OpInfo.getMatchedOperand();
5557 // Scan until we find the definition we already emitted of this operand.
5558 // When we find it, create a RegsForValue operand.
5559 unsigned CurOp = InlineAsm::Op_FirstOperand;
5560 for (; OperandNo; --OperandNo) {
5561 // Advance to the next operand.
5563 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
5564 assert((InlineAsm::isRegDefKind(OpFlag) ||
5565 InlineAsm::isRegDefEarlyClobberKind(OpFlag) ||
5566 InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?");
5567 CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1;
5571 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
5572 if (InlineAsm::isRegDefKind(OpFlag) ||
5573 InlineAsm::isRegDefEarlyClobberKind(OpFlag)) {
5574 // Add (OpFlag&0xffff)>>3 registers to MatchedRegs.
5575 if (OpInfo.isIndirect) {
5576 // This happens on gcc/testsuite/gcc.dg/pr8788-1.c
5577 LLVMContext &Ctx = *DAG.getContext();
5578 Ctx.emitError(CS.getInstruction(), "inline asm not supported yet:"
5579 " don't know how to handle tied "
5580 "indirect register inputs");
5583 RegsForValue MatchedRegs;
5584 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
5585 EVT RegVT = AsmNodeOperands[CurOp+1].getValueType();
5586 MatchedRegs.RegVTs.push_back(RegVT);
5587 MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo();
5588 for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag);
5590 MatchedRegs.Regs.push_back
5591 (RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT)));
5593 // Use the produced MatchedRegs object to
5594 MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
5596 MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse,
5597 true, OpInfo.getMatchedOperand(),
5598 DAG, AsmNodeOperands);
5602 assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!");
5603 assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 &&
5604 "Unexpected number of operands");
5605 // Add information to the INLINEASM node to know about this input.
5606 // See InlineAsm.h isUseOperandTiedToDef.
5607 OpFlag = InlineAsm::getFlagWordForMatchingOp(OpFlag,
5608 OpInfo.getMatchedOperand());
5609 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlag,
5610 TLI.getPointerTy()));
5611 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]);
5615 // Treat indirect 'X' constraint as memory.
5616 if (OpInfo.ConstraintType == TargetLowering::C_Other &&
5618 OpInfo.ConstraintType = TargetLowering::C_Memory;
5620 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
5621 std::vector<SDValue> Ops;
5622 TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode[0],
5625 report_fatal_error("Invalid operand for inline asm constraint '" +
5626 Twine(OpInfo.ConstraintCode) + "'!");
5628 // Add information to the INLINEASM node to know about this input.
5629 unsigned ResOpType =
5630 InlineAsm::getFlagWord(InlineAsm::Kind_Imm, Ops.size());
5631 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
5632 TLI.getPointerTy()));
5633 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
5637 if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
5638 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
5639 assert(InOperandVal.getValueType() == TLI.getPointerTy() &&
5640 "Memory operands expect pointer values");
5642 // Add information to the INLINEASM node to know about this input.
5643 unsigned ResOpType = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
5644 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
5645 TLI.getPointerTy()));
5646 AsmNodeOperands.push_back(InOperandVal);
5650 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
5651 OpInfo.ConstraintType == TargetLowering::C_Register) &&
5652 "Unknown constraint type!");
5653 assert(!OpInfo.isIndirect &&
5654 "Don't know how to handle indirect register inputs yet!");
5656 // Copy the input into the appropriate registers.
5657 if (OpInfo.AssignedRegs.Regs.empty() ||
5658 !OpInfo.AssignedRegs.areValueTypesLegal(TLI))
5659 report_fatal_error("Couldn't allocate input reg for constraint '" +
5660 Twine(OpInfo.ConstraintCode) + "'!");
5662 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
5665 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, false, 0,
5666 DAG, AsmNodeOperands);
5669 case InlineAsm::isClobber: {
5670 // Add the clobbered value to the operand list, so that the register
5671 // allocator is aware that the physreg got clobbered.
5672 if (!OpInfo.AssignedRegs.Regs.empty())
5673 OpInfo.AssignedRegs.AddInlineAsmOperands(
5674 InlineAsm::Kind_RegDefEarlyClobber,
5682 // Finish up input operands. Set the input chain and add the flag last.
5683 AsmNodeOperands[InlineAsm::Op_InputChain] = Chain;
5684 if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
5686 Chain = DAG.getNode(ISD::INLINEASM, getCurDebugLoc(),
5687 DAG.getVTList(MVT::Other, MVT::Flag),
5688 &AsmNodeOperands[0], AsmNodeOperands.size());
5689 Flag = Chain.getValue(1);
5691 // If this asm returns a register value, copy the result from that register
5692 // and set it as the value of the call.
5693 if (!RetValRegs.Regs.empty()) {
5694 SDValue Val = RetValRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(),
5697 // FIXME: Why don't we do this for inline asms with MRVs?
5698 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) {
5699 EVT ResultType = TLI.getValueType(CS.getType());
5701 // If any of the results of the inline asm is a vector, it may have the
5702 // wrong width/num elts. This can happen for register classes that can
5703 // contain multiple different value types. The preg or vreg allocated may
5704 // not have the same VT as was expected. Convert it to the right type
5705 // with bit_convert.
5706 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) {
5707 Val = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
5710 } else if (ResultType != Val.getValueType() &&
5711 ResultType.isInteger() && Val.getValueType().isInteger()) {
5712 // If a result value was tied to an input value, the computed result may
5713 // have a wider width than the expected result. Extract the relevant
5715 Val = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), ResultType, Val);
5718 assert(ResultType == Val.getValueType() && "Asm result value mismatch!");
5721 setValue(CS.getInstruction(), Val);
5722 // Don't need to use this as a chain in this case.
5723 if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty())
5727 std::vector<std::pair<SDValue, const Value *> > StoresToEmit;
5729 // Process indirect outputs, first output all of the flagged copies out of
5731 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
5732 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
5733 const Value *Ptr = IndirectStoresToEmit[i].second;
5734 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(),
5736 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
5739 // Emit the non-flagged stores from the physregs.
5740 SmallVector<SDValue, 8> OutChains;
5741 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) {
5742 SDValue Val = DAG.getStore(Chain, getCurDebugLoc(),
5743 StoresToEmit[i].first,
5744 getValue(StoresToEmit[i].second),
5745 StoresToEmit[i].second, 0,
5747 OutChains.push_back(Val);
5750 if (!OutChains.empty())
5751 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
5752 &OutChains[0], OutChains.size());
5757 void SelectionDAGBuilder::visitVAStart(const CallInst &I) {
5758 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurDebugLoc(),
5759 MVT::Other, getRoot(),
5760 getValue(I.getArgOperand(0)),
5761 DAG.getSrcValue(I.getArgOperand(0))));
5764 void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) {
5765 const TargetData &TD = *TLI.getTargetData();
5766 SDValue V = DAG.getVAArg(TLI.getValueType(I.getType()), getCurDebugLoc(),
5767 getRoot(), getValue(I.getOperand(0)),
5768 DAG.getSrcValue(I.getOperand(0)),
5769 TD.getABITypeAlignment(I.getType()));
5771 DAG.setRoot(V.getValue(1));
5774 void SelectionDAGBuilder::visitVAEnd(const CallInst &I) {
5775 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurDebugLoc(),
5776 MVT::Other, getRoot(),
5777 getValue(I.getArgOperand(0)),
5778 DAG.getSrcValue(I.getArgOperand(0))));
5781 void SelectionDAGBuilder::visitVACopy(const CallInst &I) {
5782 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurDebugLoc(),
5783 MVT::Other, getRoot(),
5784 getValue(I.getArgOperand(0)),
5785 getValue(I.getArgOperand(1)),
5786 DAG.getSrcValue(I.getArgOperand(0)),
5787 DAG.getSrcValue(I.getArgOperand(1))));
5790 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
5791 /// implementation, which just calls LowerCall.
5792 /// FIXME: When all targets are
5793 /// migrated to using LowerCall, this hook should be integrated into SDISel.
5794 std::pair<SDValue, SDValue>
5795 TargetLowering::LowerCallTo(SDValue Chain, const Type *RetTy,
5796 bool RetSExt, bool RetZExt, bool isVarArg,
5797 bool isInreg, unsigned NumFixedArgs,
5798 CallingConv::ID CallConv, bool isTailCall,
5799 bool isReturnValueUsed,
5801 ArgListTy &Args, SelectionDAG &DAG,
5802 DebugLoc dl) const {
5803 // Handle all of the outgoing arguments.
5804 SmallVector<ISD::OutputArg, 32> Outs;
5805 SmallVector<SDValue, 32> OutVals;
5806 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
5807 SmallVector<EVT, 4> ValueVTs;
5808 ComputeValueVTs(*this, Args[i].Ty, ValueVTs);
5809 for (unsigned Value = 0, NumValues = ValueVTs.size();
5810 Value != NumValues; ++Value) {
5811 EVT VT = ValueVTs[Value];
5812 const Type *ArgTy = VT.getTypeForEVT(RetTy->getContext());
5813 SDValue Op = SDValue(Args[i].Node.getNode(),
5814 Args[i].Node.getResNo() + Value);
5815 ISD::ArgFlagsTy Flags;
5816 unsigned OriginalAlignment =
5817 getTargetData()->getABITypeAlignment(ArgTy);
5823 if (Args[i].isInReg)
5827 if (Args[i].isByVal) {
5829 const PointerType *Ty = cast<PointerType>(Args[i].Ty);
5830 const Type *ElementTy = Ty->getElementType();
5831 unsigned FrameAlign = getByValTypeAlignment(ElementTy);
5832 unsigned FrameSize = getTargetData()->getTypeAllocSize(ElementTy);
5833 // For ByVal, alignment should come from FE. BE will guess if this
5834 // info is not there but there are cases it cannot get right.
5835 if (Args[i].Alignment)
5836 FrameAlign = Args[i].Alignment;
5837 Flags.setByValAlign(FrameAlign);
5838 Flags.setByValSize(FrameSize);
5842 Flags.setOrigAlign(OriginalAlignment);
5844 EVT PartVT = getRegisterType(RetTy->getContext(), VT);
5845 unsigned NumParts = getNumRegisters(RetTy->getContext(), VT);
5846 SmallVector<SDValue, 4> Parts(NumParts);
5847 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
5850 ExtendKind = ISD::SIGN_EXTEND;
5851 else if (Args[i].isZExt)
5852 ExtendKind = ISD::ZERO_EXTEND;
5854 getCopyToParts(DAG, dl, Op, &Parts[0], NumParts,
5855 PartVT, ExtendKind);
5857 for (unsigned j = 0; j != NumParts; ++j) {
5858 // if it isn't first piece, alignment must be 1
5859 ISD::OutputArg MyFlags(Flags, Parts[j].getValueType(),
5861 if (NumParts > 1 && j == 0)
5862 MyFlags.Flags.setSplit();
5864 MyFlags.Flags.setOrigAlign(1);
5866 Outs.push_back(MyFlags);
5867 OutVals.push_back(Parts[j]);
5872 // Handle the incoming return values from the call.
5873 SmallVector<ISD::InputArg, 32> Ins;
5874 SmallVector<EVT, 4> RetTys;
5875 ComputeValueVTs(*this, RetTy, RetTys);
5876 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
5878 EVT RegisterVT = getRegisterType(RetTy->getContext(), VT);
5879 unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT);
5880 for (unsigned i = 0; i != NumRegs; ++i) {
5881 ISD::InputArg MyFlags;
5882 MyFlags.VT = RegisterVT;
5883 MyFlags.Used = isReturnValueUsed;
5885 MyFlags.Flags.setSExt();
5887 MyFlags.Flags.setZExt();
5889 MyFlags.Flags.setInReg();
5890 Ins.push_back(MyFlags);
5894 SmallVector<SDValue, 4> InVals;
5895 Chain = LowerCall(Chain, Callee, CallConv, isVarArg, isTailCall,
5896 Outs, OutVals, Ins, dl, DAG, InVals);
5898 // Verify that the target's LowerCall behaved as expected.
5899 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
5900 "LowerCall didn't return a valid chain!");
5901 assert((!isTailCall || InVals.empty()) &&
5902 "LowerCall emitted a return value for a tail call!");
5903 assert((isTailCall || InVals.size() == Ins.size()) &&
5904 "LowerCall didn't emit the correct number of values!");
5906 // For a tail call, the return value is merely live-out and there aren't
5907 // any nodes in the DAG representing it. Return a special value to
5908 // indicate that a tail call has been emitted and no more Instructions
5909 // should be processed in the current block.
5912 return std::make_pair(SDValue(), SDValue());
5915 DEBUG(for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
5916 assert(InVals[i].getNode() &&
5917 "LowerCall emitted a null value!");
5918 assert(Ins[i].VT == InVals[i].getValueType() &&
5919 "LowerCall emitted a value with the wrong type!");
5922 // Collect the legal value parts into potentially illegal values
5923 // that correspond to the original function's return values.
5924 ISD::NodeType AssertOp = ISD::DELETED_NODE;
5926 AssertOp = ISD::AssertSext;
5928 AssertOp = ISD::AssertZext;
5929 SmallVector<SDValue, 4> ReturnValues;
5930 unsigned CurReg = 0;
5931 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
5933 EVT RegisterVT = getRegisterType(RetTy->getContext(), VT);
5934 unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT);
5936 ReturnValues.push_back(getCopyFromParts(DAG, dl, &InVals[CurReg],
5937 NumRegs, RegisterVT, VT,
5942 // For a function returning void, there is no return value. We can't create
5943 // such a node, so we just return a null return value in that case. In
5944 // that case, nothing will actualy look at the value.
5945 if (ReturnValues.empty())
5946 return std::make_pair(SDValue(), Chain);
5948 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, dl,
5949 DAG.getVTList(&RetTys[0], RetTys.size()),
5950 &ReturnValues[0], ReturnValues.size());
5951 return std::make_pair(Res, Chain);
5954 void TargetLowering::LowerOperationWrapper(SDNode *N,
5955 SmallVectorImpl<SDValue> &Results,
5956 SelectionDAG &DAG) const {
5957 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
5959 Results.push_back(Res);
5962 SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5963 llvm_unreachable("LowerOperation not implemented for this target!");
5968 SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) {
5969 SDValue Op = getNonRegisterValue(V);
5970 assert((Op.getOpcode() != ISD::CopyFromReg ||
5971 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
5972 "Copy from a reg to the same reg!");
5973 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg");
5975 RegsForValue RFV(V->getContext(), TLI, Reg, V->getType());
5976 SDValue Chain = DAG.getEntryNode();
5977 RFV.getCopyToRegs(Op, DAG, getCurDebugLoc(), Chain, 0);
5978 PendingExports.push_back(Chain);
5981 #include "llvm/CodeGen/SelectionDAGISel.h"
5983 void SelectionDAGISel::LowerArguments(const BasicBlock *LLVMBB) {
5984 // If this is the entry block, emit arguments.
5985 const Function &F = *LLVMBB->getParent();
5986 SelectionDAG &DAG = SDB->DAG;
5987 DebugLoc dl = SDB->getCurDebugLoc();
5988 const TargetData *TD = TLI.getTargetData();
5989 SmallVector<ISD::InputArg, 16> Ins;
5991 // Check whether the function can return without sret-demotion.
5992 SmallVector<ISD::OutputArg, 4> Outs;
5993 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
5996 if (!FuncInfo->CanLowerReturn) {
5997 // Put in an sret pointer parameter before all the other parameters.
5998 SmallVector<EVT, 1> ValueVTs;
5999 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
6001 // NOTE: Assuming that a pointer will never break down to more than one VT
6003 ISD::ArgFlagsTy Flags;
6005 EVT RegisterVT = TLI.getRegisterType(*DAG.getContext(), ValueVTs[0]);
6006 ISD::InputArg RetArg(Flags, RegisterVT, true);
6007 Ins.push_back(RetArg);
6010 // Set up the incoming argument description vector.
6012 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end();
6013 I != E; ++I, ++Idx) {
6014 SmallVector<EVT, 4> ValueVTs;
6015 ComputeValueVTs(TLI, I->getType(), ValueVTs);
6016 bool isArgValueUsed = !I->use_empty();
6017 for (unsigned Value = 0, NumValues = ValueVTs.size();
6018 Value != NumValues; ++Value) {
6019 EVT VT = ValueVTs[Value];
6020 const Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
6021 ISD::ArgFlagsTy Flags;
6022 unsigned OriginalAlignment =
6023 TD->getABITypeAlignment(ArgTy);
6025 if (F.paramHasAttr(Idx, Attribute::ZExt))
6027 if (F.paramHasAttr(Idx, Attribute::SExt))
6029 if (F.paramHasAttr(Idx, Attribute::InReg))
6031 if (F.paramHasAttr(Idx, Attribute::StructRet))
6033 if (F.paramHasAttr(Idx, Attribute::ByVal)) {
6035 const PointerType *Ty = cast<PointerType>(I->getType());
6036 const Type *ElementTy = Ty->getElementType();
6037 unsigned FrameAlign = TLI.getByValTypeAlignment(ElementTy);
6038 unsigned FrameSize = TD->getTypeAllocSize(ElementTy);
6039 // For ByVal, alignment should be passed from FE. BE will guess if
6040 // this info is not there but there are cases it cannot get right.
6041 if (F.getParamAlignment(Idx))
6042 FrameAlign = F.getParamAlignment(Idx);
6043 Flags.setByValAlign(FrameAlign);
6044 Flags.setByValSize(FrameSize);
6046 if (F.paramHasAttr(Idx, Attribute::Nest))
6048 Flags.setOrigAlign(OriginalAlignment);
6050 EVT RegisterVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6051 unsigned NumRegs = TLI.getNumRegisters(*CurDAG->getContext(), VT);
6052 for (unsigned i = 0; i != NumRegs; ++i) {
6053 ISD::InputArg MyFlags(Flags, RegisterVT, isArgValueUsed);
6054 if (NumRegs > 1 && i == 0)
6055 MyFlags.Flags.setSplit();
6056 // if it isn't first piece, alignment must be 1
6058 MyFlags.Flags.setOrigAlign(1);
6059 Ins.push_back(MyFlags);
6064 // Call the target to set up the argument values.
6065 SmallVector<SDValue, 8> InVals;
6066 SDValue NewRoot = TLI.LowerFormalArguments(DAG.getRoot(), F.getCallingConv(),
6070 // Verify that the target's LowerFormalArguments behaved as expected.
6071 assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other &&
6072 "LowerFormalArguments didn't return a valid chain!");
6073 assert(InVals.size() == Ins.size() &&
6074 "LowerFormalArguments didn't emit the correct number of values!");
6076 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
6077 assert(InVals[i].getNode() &&
6078 "LowerFormalArguments emitted a null value!");
6079 assert(Ins[i].VT == InVals[i].getValueType() &&
6080 "LowerFormalArguments emitted a value with the wrong type!");
6084 // Update the DAG with the new chain value resulting from argument lowering.
6085 DAG.setRoot(NewRoot);
6087 // Set up the argument values.
6090 if (!FuncInfo->CanLowerReturn) {
6091 // Create a virtual register for the sret pointer, and put in a copy
6092 // from the sret argument into it.
6093 SmallVector<EVT, 1> ValueVTs;
6094 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
6095 EVT VT = ValueVTs[0];
6096 EVT RegVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6097 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6098 SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1,
6099 RegVT, VT, AssertOp);
6101 MachineFunction& MF = SDB->DAG.getMachineFunction();
6102 MachineRegisterInfo& RegInfo = MF.getRegInfo();
6103 unsigned SRetReg = RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT));
6104 FuncInfo->DemoteRegister = SRetReg;
6105 NewRoot = SDB->DAG.getCopyToReg(NewRoot, SDB->getCurDebugLoc(),
6107 DAG.setRoot(NewRoot);
6109 // i indexes lowered arguments. Bump it past the hidden sret argument.
6110 // Idx indexes LLVM arguments. Don't touch it.
6114 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
6116 SmallVector<SDValue, 4> ArgValues;
6117 SmallVector<EVT, 4> ValueVTs;
6118 ComputeValueVTs(TLI, I->getType(), ValueVTs);
6119 unsigned NumValues = ValueVTs.size();
6121 // If this argument is unused then remember its value. It is used to generate
6122 // debugging information.
6123 if (I->use_empty() && NumValues)
6124 SDB->setUnusedArgValue(I, InVals[i]);
6126 for (unsigned Value = 0; Value != NumValues; ++Value) {
6127 EVT VT = ValueVTs[Value];
6128 EVT PartVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6129 unsigned NumParts = TLI.getNumRegisters(*CurDAG->getContext(), VT);
6131 if (!I->use_empty()) {
6132 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6133 if (F.paramHasAttr(Idx, Attribute::SExt))
6134 AssertOp = ISD::AssertSext;
6135 else if (F.paramHasAttr(Idx, Attribute::ZExt))
6136 AssertOp = ISD::AssertZext;
6138 ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i],
6139 NumParts, PartVT, VT,
6146 if (!I->use_empty()) {
6148 if (!ArgValues.empty())
6149 Res = DAG.getMergeValues(&ArgValues[0], NumValues,
6150 SDB->getCurDebugLoc());
6151 SDB->setValue(I, Res);
6153 // If this argument is live outside of the entry block, insert a copy from
6154 // whereever we got it to the vreg that other BB's will reference it as.
6155 SDB->CopyToExportRegsIfNeeded(I);
6159 assert(i == InVals.size() && "Argument register count mismatch!");
6161 // Finally, if the target has anything special to do, allow it to do so.
6162 // FIXME: this should insert code into the DAG!
6163 EmitFunctionEntryCode();
6166 /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
6167 /// ensure constants are generated when needed. Remember the virtual registers
6168 /// that need to be added to the Machine PHI nodes as input. We cannot just
6169 /// directly add them, because expansion might result in multiple MBB's for one
6170 /// BB. As such, the start of the BB might correspond to a different MBB than
6174 SelectionDAGBuilder::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
6175 const TerminatorInst *TI = LLVMBB->getTerminator();
6177 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
6179 // Check successor nodes' PHI nodes that expect a constant to be available
6181 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
6182 const BasicBlock *SuccBB = TI->getSuccessor(succ);
6183 if (!isa<PHINode>(SuccBB->begin())) continue;
6184 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
6186 // If this terminator has multiple identical successors (common for
6187 // switches), only handle each succ once.
6188 if (!SuccsHandled.insert(SuccMBB)) continue;
6190 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
6192 // At this point we know that there is a 1-1 correspondence between LLVM PHI
6193 // nodes and Machine PHI nodes, but the incoming operands have not been
6195 for (BasicBlock::const_iterator I = SuccBB->begin();
6196 const PHINode *PN = dyn_cast<PHINode>(I); ++I) {
6197 // Ignore dead phi's.
6198 if (PN->use_empty()) continue;
6201 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
6203 if (const Constant *C = dyn_cast<Constant>(PHIOp)) {
6204 unsigned &RegOut = ConstantsOut[C];
6206 RegOut = FuncInfo.CreateRegs(C->getType());
6207 CopyValueToVirtualRegister(C, RegOut);
6211 DenseMap<const Value *, unsigned>::iterator I =
6212 FuncInfo.ValueMap.find(PHIOp);
6213 if (I != FuncInfo.ValueMap.end())
6216 assert(isa<AllocaInst>(PHIOp) &&
6217 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
6218 "Didn't codegen value into a register!??");
6219 Reg = FuncInfo.CreateRegs(PHIOp->getType());
6220 CopyValueToVirtualRegister(PHIOp, Reg);
6224 // Remember that this register needs to added to the machine PHI node as
6225 // the input for this MBB.
6226 SmallVector<EVT, 4> ValueVTs;
6227 ComputeValueVTs(TLI, PN->getType(), ValueVTs);
6228 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
6229 EVT VT = ValueVTs[vti];
6230 unsigned NumRegisters = TLI.getNumRegisters(*DAG.getContext(), VT);
6231 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
6232 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
6233 Reg += NumRegisters;
6237 ConstantsOut.clear();