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,
451 IntermediateVT, NumIntermediates, RegisterVT);
452 unsigned NumElements = ValueVT.getVectorNumElements();
454 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
455 NumParts = NumRegs; // Silence a compiler warning.
456 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
458 // Split the vector into intermediate operands.
459 SmallVector<SDValue, 8> Ops(NumIntermediates);
460 for (unsigned i = 0; i != NumIntermediates; ++i) {
461 if (IntermediateVT.isVector())
462 Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL,
464 DAG.getIntPtrConstant(i * (NumElements / NumIntermediates)));
466 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
467 IntermediateVT, Val, DAG.getIntPtrConstant(i));
470 // Split the intermediate operands into legal parts.
471 if (NumParts == NumIntermediates) {
472 // If the register was not expanded, promote or copy the value,
474 for (unsigned i = 0; i != NumParts; ++i)
475 getCopyToParts(DAG, DL, Ops[i], &Parts[i], 1, PartVT);
476 } else if (NumParts > 0) {
477 // If the intermediate type was expanded, split each the value into
479 assert(NumParts % NumIntermediates == 0 &&
480 "Must expand into a divisible number of parts!");
481 unsigned Factor = NumParts / NumIntermediates;
482 for (unsigned i = 0; i != NumIntermediates; ++i)
483 getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT);
491 /// RegsForValue - This struct represents the registers (physical or virtual)
492 /// that a particular set of values is assigned, and the type information
493 /// about the value. The most common situation is to represent one value at a
494 /// time, but struct or array values are handled element-wise as multiple
495 /// values. The splitting of aggregates is performed recursively, so that we
496 /// never have aggregate-typed registers. The values at this point do not
497 /// necessarily have legal types, so each value may require one or more
498 /// registers of some legal type.
500 struct RegsForValue {
501 /// ValueVTs - The value types of the values, which may not be legal, and
502 /// may need be promoted or synthesized from one or more registers.
504 SmallVector<EVT, 4> ValueVTs;
506 /// RegVTs - The value types of the registers. This is the same size as
507 /// ValueVTs and it records, for each value, what the type of the assigned
508 /// register or registers are. (Individual values are never synthesized
509 /// from more than one type of register.)
511 /// With virtual registers, the contents of RegVTs is redundant with TLI's
512 /// getRegisterType member function, however when with physical registers
513 /// it is necessary to have a separate record of the types.
515 SmallVector<EVT, 4> RegVTs;
517 /// Regs - This list holds the registers assigned to the values.
518 /// Each legal or promoted value requires one register, and each
519 /// expanded value requires multiple registers.
521 SmallVector<unsigned, 4> Regs;
525 RegsForValue(const SmallVector<unsigned, 4> ®s,
526 EVT regvt, EVT valuevt)
527 : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {}
529 RegsForValue(LLVMContext &Context, const TargetLowering &tli,
530 unsigned Reg, const Type *Ty) {
531 ComputeValueVTs(tli, Ty, ValueVTs);
533 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
534 EVT ValueVT = ValueVTs[Value];
535 unsigned NumRegs = tli.getNumRegisters(Context, ValueVT);
536 EVT RegisterVT = tli.getRegisterType(Context, ValueVT);
537 for (unsigned i = 0; i != NumRegs; ++i)
538 Regs.push_back(Reg + i);
539 RegVTs.push_back(RegisterVT);
544 /// areValueTypesLegal - Return true if types of all the values are legal.
545 bool areValueTypesLegal(const TargetLowering &TLI) {
546 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
547 EVT RegisterVT = RegVTs[Value];
548 if (!TLI.isTypeLegal(RegisterVT))
554 /// append - Add the specified values to this one.
555 void append(const RegsForValue &RHS) {
556 ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end());
557 RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end());
558 Regs.append(RHS.Regs.begin(), RHS.Regs.end());
561 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
562 /// this value and returns the result as a ValueVTs value. This uses
563 /// Chain/Flag as the input and updates them for the output Chain/Flag.
564 /// If the Flag pointer is NULL, no flag is used.
565 SDValue getCopyFromRegs(SelectionDAG &DAG, FunctionLoweringInfo &FuncInfo,
567 SDValue &Chain, SDValue *Flag) const;
569 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
570 /// specified value into the registers specified by this object. This uses
571 /// Chain/Flag as the input and updates them for the output Chain/Flag.
572 /// If the Flag pointer is NULL, no flag is used.
573 void getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
574 SDValue &Chain, SDValue *Flag) const;
576 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
577 /// operand list. This adds the code marker, matching input operand index
578 /// (if applicable), and includes the number of values added into it.
579 void AddInlineAsmOperands(unsigned Kind,
580 bool HasMatching, unsigned MatchingIdx,
582 std::vector<SDValue> &Ops) const;
586 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
587 /// this value and returns the result as a ValueVT value. This uses
588 /// Chain/Flag as the input and updates them for the output Chain/Flag.
589 /// If the Flag pointer is NULL, no flag is used.
590 SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
591 FunctionLoweringInfo &FuncInfo,
593 SDValue &Chain, SDValue *Flag) const {
594 // A Value with type {} or [0 x %t] needs no registers.
595 if (ValueVTs.empty())
598 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
600 // Assemble the legal parts into the final values.
601 SmallVector<SDValue, 4> Values(ValueVTs.size());
602 SmallVector<SDValue, 8> Parts;
603 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
604 // Copy the legal parts from the registers.
605 EVT ValueVT = ValueVTs[Value];
606 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
607 EVT RegisterVT = RegVTs[Value];
609 Parts.resize(NumRegs);
610 for (unsigned i = 0; i != NumRegs; ++i) {
613 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT);
615 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag);
616 *Flag = P.getValue(2);
619 Chain = P.getValue(1);
621 // If the source register was virtual and if we know something about it,
622 // add an assert node.
623 if (TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) &&
624 RegisterVT.isInteger() && !RegisterVT.isVector()) {
625 unsigned SlotNo = Regs[Part+i]-TargetRegisterInfo::FirstVirtualRegister;
626 if (FuncInfo.LiveOutRegInfo.size() > SlotNo) {
627 const FunctionLoweringInfo::LiveOutInfo &LOI =
628 FuncInfo.LiveOutRegInfo[SlotNo];
630 unsigned RegSize = RegisterVT.getSizeInBits();
631 unsigned NumSignBits = LOI.NumSignBits;
632 unsigned NumZeroBits = LOI.KnownZero.countLeadingOnes();
634 // FIXME: We capture more information than the dag can represent. For
635 // now, just use the tightest assertzext/assertsext possible.
637 EVT FromVT(MVT::Other);
638 if (NumSignBits == RegSize)
639 isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1
640 else if (NumZeroBits >= RegSize-1)
641 isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1
642 else if (NumSignBits > RegSize-8)
643 isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8
644 else if (NumZeroBits >= RegSize-8)
645 isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8
646 else if (NumSignBits > RegSize-16)
647 isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16
648 else if (NumZeroBits >= RegSize-16)
649 isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16
650 else if (NumSignBits > RegSize-32)
651 isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32
652 else if (NumZeroBits >= RegSize-32)
653 isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32
655 if (FromVT != MVT::Other)
656 P = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl,
657 RegisterVT, P, DAG.getValueType(FromVT));
664 Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(),
665 NumRegs, RegisterVT, ValueVT);
670 return DAG.getNode(ISD::MERGE_VALUES, dl,
671 DAG.getVTList(&ValueVTs[0], ValueVTs.size()),
672 &Values[0], ValueVTs.size());
675 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
676 /// specified value into the registers specified by this object. This uses
677 /// Chain/Flag as the input and updates them for the output Chain/Flag.
678 /// If the Flag pointer is NULL, no flag is used.
679 void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
680 SDValue &Chain, SDValue *Flag) const {
681 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
683 // Get the list of the values's legal parts.
684 unsigned NumRegs = Regs.size();
685 SmallVector<SDValue, 8> Parts(NumRegs);
686 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
687 EVT ValueVT = ValueVTs[Value];
688 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
689 EVT RegisterVT = RegVTs[Value];
691 getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value),
692 &Parts[Part], NumParts, RegisterVT);
696 // Copy the parts into the registers.
697 SmallVector<SDValue, 8> Chains(NumRegs);
698 for (unsigned i = 0; i != NumRegs; ++i) {
701 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]);
703 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag);
704 *Flag = Part.getValue(1);
707 Chains[i] = Part.getValue(0);
710 if (NumRegs == 1 || Flag)
711 // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is
712 // flagged to it. That is the CopyToReg nodes and the user are considered
713 // a single scheduling unit. If we create a TokenFactor and return it as
714 // chain, then the TokenFactor is both a predecessor (operand) of the
715 // user as well as a successor (the TF operands are flagged to the user).
716 // c1, f1 = CopyToReg
717 // c2, f2 = CopyToReg
718 // c3 = TokenFactor c1, c2
721 Chain = Chains[NumRegs-1];
723 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0], NumRegs);
726 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
727 /// operand list. This adds the code marker and includes the number of
728 /// values added into it.
729 void RegsForValue::AddInlineAsmOperands(unsigned Code, bool HasMatching,
730 unsigned MatchingIdx,
732 std::vector<SDValue> &Ops) const {
733 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
735 unsigned Flag = InlineAsm::getFlagWord(Code, Regs.size());
737 Flag = InlineAsm::getFlagWordForMatchingOp(Flag, MatchingIdx);
738 SDValue Res = DAG.getTargetConstant(Flag, MVT::i32);
741 for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) {
742 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVTs[Value]);
743 EVT RegisterVT = RegVTs[Value];
744 for (unsigned i = 0; i != NumRegs; ++i) {
745 assert(Reg < Regs.size() && "Mismatch in # registers expected");
746 Ops.push_back(DAG.getRegister(Regs[Reg++], RegisterVT));
751 void SelectionDAGBuilder::init(GCFunctionInfo *gfi, AliasAnalysis &aa) {
754 TD = DAG.getTarget().getTargetData();
757 /// clear - Clear out the current SelectionDAG and the associated
758 /// state and prepare this SelectionDAGBuilder object to be used
759 /// for a new block. This doesn't clear out information about
760 /// additional blocks that are needed to complete switch lowering
761 /// or PHI node updating; that information is cleared out as it is
763 void SelectionDAGBuilder::clear() {
765 UnusedArgNodeMap.clear();
766 PendingLoads.clear();
767 PendingExports.clear();
768 DanglingDebugInfoMap.clear();
769 CurDebugLoc = DebugLoc();
773 /// getRoot - Return the current virtual root of the Selection DAG,
774 /// flushing any PendingLoad items. This must be done before emitting
775 /// a store or any other node that may need to be ordered after any
776 /// prior load instructions.
778 SDValue SelectionDAGBuilder::getRoot() {
779 if (PendingLoads.empty())
780 return DAG.getRoot();
782 if (PendingLoads.size() == 1) {
783 SDValue Root = PendingLoads[0];
785 PendingLoads.clear();
789 // Otherwise, we have to make a token factor node.
790 SDValue Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
791 &PendingLoads[0], PendingLoads.size());
792 PendingLoads.clear();
797 /// getControlRoot - Similar to getRoot, but instead of flushing all the
798 /// PendingLoad items, flush all the PendingExports items. It is necessary
799 /// to do this before emitting a terminator instruction.
801 SDValue SelectionDAGBuilder::getControlRoot() {
802 SDValue Root = DAG.getRoot();
804 if (PendingExports.empty())
807 // Turn all of the CopyToReg chains into one factored node.
808 if (Root.getOpcode() != ISD::EntryToken) {
809 unsigned i = 0, e = PendingExports.size();
810 for (; i != e; ++i) {
811 assert(PendingExports[i].getNode()->getNumOperands() > 1);
812 if (PendingExports[i].getNode()->getOperand(0) == Root)
813 break; // Don't add the root if we already indirectly depend on it.
817 PendingExports.push_back(Root);
820 Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
822 PendingExports.size());
823 PendingExports.clear();
828 void SelectionDAGBuilder::AssignOrderingToNode(const SDNode *Node) {
829 if (DAG.GetOrdering(Node) != 0) return; // Already has ordering.
830 DAG.AssignOrdering(Node, SDNodeOrder);
832 for (unsigned I = 0, E = Node->getNumOperands(); I != E; ++I)
833 AssignOrderingToNode(Node->getOperand(I).getNode());
836 void SelectionDAGBuilder::visit(const Instruction &I) {
837 // Set up outgoing PHI node register values before emitting the terminator.
838 if (isa<TerminatorInst>(&I))
839 HandlePHINodesInSuccessorBlocks(I.getParent());
841 CurDebugLoc = I.getDebugLoc();
843 visit(I.getOpcode(), I);
845 if (!isa<TerminatorInst>(&I) && !HasTailCall)
846 CopyToExportRegsIfNeeded(&I);
848 CurDebugLoc = DebugLoc();
851 void SelectionDAGBuilder::visitPHI(const PHINode &) {
852 llvm_unreachable("SelectionDAGBuilder shouldn't visit PHI nodes!");
855 void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) {
856 // Note: this doesn't use InstVisitor, because it has to work with
857 // ConstantExpr's in addition to instructions.
859 default: llvm_unreachable("Unknown instruction type encountered!");
860 // Build the switch statement using the Instruction.def file.
861 #define HANDLE_INST(NUM, OPCODE, CLASS) \
862 case Instruction::OPCODE: visit##OPCODE((CLASS&)I); break;
863 #include "llvm/Instruction.def"
866 // Assign the ordering to the freshly created DAG nodes.
867 if (NodeMap.count(&I)) {
869 AssignOrderingToNode(getValue(&I).getNode());
873 // resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V,
874 // generate the debug data structures now that we've seen its definition.
875 void SelectionDAGBuilder::resolveDanglingDebugInfo(const Value *V,
877 DanglingDebugInfo &DDI = DanglingDebugInfoMap[V];
878 MDNode *Variable = NULL;
881 if (const DbgValueInst *DI = dyn_cast_or_null<DbgValueInst>(DDI.getDI())) {
882 Variable = DI->getVariable();
883 Offset = DI->getOffset();
884 } else if (const DbgDeclareInst *DI =
885 dyn_cast_or_null<DbgDeclareInst>(DDI.getDI()))
886 Variable = DI->getVariable();
888 assert (DDI.getDI() == NULL && "Invalid debug info intrinsic!");
893 DebugLoc dl = DDI.getdl();
894 unsigned DbgSDNodeOrder = DDI.getSDNodeOrder();
897 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, Val)) {
898 SDV = DAG.getDbgValue(Variable, Val.getNode(),
899 Val.getResNo(), Offset, dl, DbgSDNodeOrder);
900 DAG.AddDbgValue(SDV, Val.getNode(), false);
903 SDV = DAG.getDbgValue(Variable, UndefValue::get(V->getType()),
904 Offset, dl, SDNodeOrder);
905 DAG.AddDbgValue(SDV, 0, false);
907 DanglingDebugInfoMap[V] = DanglingDebugInfo();
911 // getValue - Return an SDValue for the given Value.
912 SDValue SelectionDAGBuilder::getValue(const Value *V) {
913 // If we already have an SDValue for this value, use it. It's important
914 // to do this first, so that we don't create a CopyFromReg if we already
915 // have a regular SDValue.
916 SDValue &N = NodeMap[V];
917 if (N.getNode()) return N;
919 // If there's a virtual register allocated and initialized for this
921 DenseMap<const Value *, unsigned>::iterator It = FuncInfo.ValueMap.find(V);
922 if (It != FuncInfo.ValueMap.end()) {
923 unsigned InReg = It->second;
924 RegsForValue RFV(*DAG.getContext(), TLI, InReg, V->getType());
925 SDValue Chain = DAG.getEntryNode();
926 N = RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain,NULL);
927 resolveDanglingDebugInfo(V, N);
931 // Otherwise create a new SDValue and remember it.
932 SDValue Val = getValueImpl(V);
934 resolveDanglingDebugInfo(V, Val);
938 /// getNonRegisterValue - Return an SDValue for the given Value, but
939 /// don't look in FuncInfo.ValueMap for a virtual register.
940 SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) {
941 // If we already have an SDValue for this value, use it.
942 SDValue &N = NodeMap[V];
943 if (N.getNode()) return N;
945 // Otherwise create a new SDValue and remember it.
946 SDValue Val = getValueImpl(V);
948 resolveDanglingDebugInfo(V, Val);
952 /// getValueImpl - Helper function for getValue and getNonRegisterValue.
953 /// Create an SDValue for the given value.
954 SDValue SelectionDAGBuilder::getValueImpl(const Value *V) {
955 if (const Constant *C = dyn_cast<Constant>(V)) {
956 EVT VT = TLI.getValueType(V->getType(), true);
958 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C))
959 return DAG.getConstant(*CI, VT);
961 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
962 return DAG.getGlobalAddress(GV, getCurDebugLoc(), VT);
964 if (isa<ConstantPointerNull>(C))
965 return DAG.getConstant(0, TLI.getPointerTy());
967 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
968 return DAG.getConstantFP(*CFP, VT);
970 if (isa<UndefValue>(C) && !V->getType()->isAggregateType())
971 return DAG.getUNDEF(VT);
973 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
974 visit(CE->getOpcode(), *CE);
975 SDValue N1 = NodeMap[V];
976 assert(N1.getNode() && "visit didn't populate the NodeMap!");
980 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) {
981 SmallVector<SDValue, 4> Constants;
982 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end();
984 SDNode *Val = getValue(*OI).getNode();
985 // If the operand is an empty aggregate, there are no values.
987 // Add each leaf value from the operand to the Constants list
988 // to form a flattened list of all the values.
989 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
990 Constants.push_back(SDValue(Val, i));
993 return DAG.getMergeValues(&Constants[0], Constants.size(),
997 if (C->getType()->isStructTy() || C->getType()->isArrayTy()) {
998 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
999 "Unknown struct or array constant!");
1001 SmallVector<EVT, 4> ValueVTs;
1002 ComputeValueVTs(TLI, C->getType(), ValueVTs);
1003 unsigned NumElts = ValueVTs.size();
1005 return SDValue(); // empty struct
1006 SmallVector<SDValue, 4> Constants(NumElts);
1007 for (unsigned i = 0; i != NumElts; ++i) {
1008 EVT EltVT = ValueVTs[i];
1009 if (isa<UndefValue>(C))
1010 Constants[i] = DAG.getUNDEF(EltVT);
1011 else if (EltVT.isFloatingPoint())
1012 Constants[i] = DAG.getConstantFP(0, EltVT);
1014 Constants[i] = DAG.getConstant(0, EltVT);
1017 return DAG.getMergeValues(&Constants[0], NumElts,
1021 if (const BlockAddress *BA = dyn_cast<BlockAddress>(C))
1022 return DAG.getBlockAddress(BA, VT);
1024 const VectorType *VecTy = cast<VectorType>(V->getType());
1025 unsigned NumElements = VecTy->getNumElements();
1027 // Now that we know the number and type of the elements, get that number of
1028 // elements into the Ops array based on what kind of constant it is.
1029 SmallVector<SDValue, 16> Ops;
1030 if (const ConstantVector *CP = dyn_cast<ConstantVector>(C)) {
1031 for (unsigned i = 0; i != NumElements; ++i)
1032 Ops.push_back(getValue(CP->getOperand(i)));
1034 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!");
1035 EVT EltVT = TLI.getValueType(VecTy->getElementType());
1038 if (EltVT.isFloatingPoint())
1039 Op = DAG.getConstantFP(0, EltVT);
1041 Op = DAG.getConstant(0, EltVT);
1042 Ops.assign(NumElements, Op);
1045 // Create a BUILD_VECTOR node.
1046 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
1047 VT, &Ops[0], Ops.size());
1050 // If this is a static alloca, generate it as the frameindex instead of
1052 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
1053 DenseMap<const AllocaInst*, int>::iterator SI =
1054 FuncInfo.StaticAllocaMap.find(AI);
1055 if (SI != FuncInfo.StaticAllocaMap.end())
1056 return DAG.getFrameIndex(SI->second, TLI.getPointerTy());
1059 // If this is an instruction which fast-isel has deferred, select it now.
1060 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
1061 unsigned InReg = FuncInfo.InitializeRegForValue(Inst);
1062 RegsForValue RFV(*DAG.getContext(), TLI, InReg, Inst->getType());
1063 SDValue Chain = DAG.getEntryNode();
1064 return RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain, NULL);
1067 llvm_unreachable("Can't get register for value!");
1071 void SelectionDAGBuilder::visitRet(const ReturnInst &I) {
1072 SDValue Chain = getControlRoot();
1073 SmallVector<ISD::OutputArg, 8> Outs;
1074 SmallVector<SDValue, 8> OutVals;
1076 if (!FuncInfo.CanLowerReturn) {
1077 unsigned DemoteReg = FuncInfo.DemoteRegister;
1078 const Function *F = I.getParent()->getParent();
1080 // Emit a store of the return value through the virtual register.
1081 // Leave Outs empty so that LowerReturn won't try to load return
1082 // registers the usual way.
1083 SmallVector<EVT, 1> PtrValueVTs;
1084 ComputeValueVTs(TLI, PointerType::getUnqual(F->getReturnType()),
1087 SDValue RetPtr = DAG.getRegister(DemoteReg, PtrValueVTs[0]);
1088 SDValue RetOp = getValue(I.getOperand(0));
1090 SmallVector<EVT, 4> ValueVTs;
1091 SmallVector<uint64_t, 4> Offsets;
1092 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs, &Offsets);
1093 unsigned NumValues = ValueVTs.size();
1095 SmallVector<SDValue, 4> Chains(NumValues);
1096 for (unsigned i = 0; i != NumValues; ++i) {
1097 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(),
1098 RetPtr.getValueType(), RetPtr,
1099 DAG.getIntPtrConstant(Offsets[i]));
1101 DAG.getStore(Chain, getCurDebugLoc(),
1102 SDValue(RetOp.getNode(), RetOp.getResNo() + i),
1103 Add, NULL, Offsets[i], false, false, 0);
1106 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
1107 MVT::Other, &Chains[0], NumValues);
1108 } else if (I.getNumOperands() != 0) {
1109 SmallVector<EVT, 4> ValueVTs;
1110 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs);
1111 unsigned NumValues = ValueVTs.size();
1113 SDValue RetOp = getValue(I.getOperand(0));
1114 for (unsigned j = 0, f = NumValues; j != f; ++j) {
1115 EVT VT = ValueVTs[j];
1117 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1119 const Function *F = I.getParent()->getParent();
1120 if (F->paramHasAttr(0, Attribute::SExt))
1121 ExtendKind = ISD::SIGN_EXTEND;
1122 else if (F->paramHasAttr(0, Attribute::ZExt))
1123 ExtendKind = ISD::ZERO_EXTEND;
1125 // FIXME: C calling convention requires the return type to be promoted
1126 // to at least 32-bit. But this is not necessary for non-C calling
1127 // conventions. The frontend should mark functions whose return values
1128 // require promoting with signext or zeroext attributes.
1129 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
1130 EVT MinVT = TLI.getRegisterType(*DAG.getContext(), MVT::i32);
1131 if (VT.bitsLT(MinVT))
1135 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), VT);
1136 EVT PartVT = TLI.getRegisterType(*DAG.getContext(), VT);
1137 SmallVector<SDValue, 4> Parts(NumParts);
1138 getCopyToParts(DAG, getCurDebugLoc(),
1139 SDValue(RetOp.getNode(), RetOp.getResNo() + j),
1140 &Parts[0], NumParts, PartVT, ExtendKind);
1142 // 'inreg' on function refers to return value
1143 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1144 if (F->paramHasAttr(0, Attribute::InReg))
1147 // Propagate extension type if any
1148 if (F->paramHasAttr(0, Attribute::SExt))
1150 else if (F->paramHasAttr(0, Attribute::ZExt))
1153 for (unsigned i = 0; i < NumParts; ++i) {
1154 Outs.push_back(ISD::OutputArg(Flags, Parts[i].getValueType(),
1156 OutVals.push_back(Parts[i]);
1162 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1163 CallingConv::ID CallConv =
1164 DAG.getMachineFunction().getFunction()->getCallingConv();
1165 Chain = TLI.LowerReturn(Chain, CallConv, isVarArg,
1166 Outs, OutVals, getCurDebugLoc(), DAG);
1168 // Verify that the target's LowerReturn behaved as expected.
1169 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
1170 "LowerReturn didn't return a valid chain!");
1172 // Update the DAG with the new chain value resulting from return lowering.
1176 /// CopyToExportRegsIfNeeded - If the given value has virtual registers
1177 /// created for it, emit nodes to copy the value into the virtual
1179 void SelectionDAGBuilder::CopyToExportRegsIfNeeded(const Value *V) {
1180 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
1181 if (VMI != FuncInfo.ValueMap.end()) {
1182 assert(!V->use_empty() && "Unused value assigned virtual registers!");
1183 CopyValueToVirtualRegister(V, VMI->second);
1187 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
1188 /// the current basic block, add it to ValueMap now so that we'll get a
1190 void SelectionDAGBuilder::ExportFromCurrentBlock(const Value *V) {
1191 // No need to export constants.
1192 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
1194 // Already exported?
1195 if (FuncInfo.isExportedInst(V)) return;
1197 unsigned Reg = FuncInfo.InitializeRegForValue(V);
1198 CopyValueToVirtualRegister(V, Reg);
1201 bool SelectionDAGBuilder::isExportableFromCurrentBlock(const Value *V,
1202 const BasicBlock *FromBB) {
1203 // The operands of the setcc have to be in this block. We don't know
1204 // how to export them from some other block.
1205 if (const Instruction *VI = dyn_cast<Instruction>(V)) {
1206 // Can export from current BB.
1207 if (VI->getParent() == FromBB)
1210 // Is already exported, noop.
1211 return FuncInfo.isExportedInst(V);
1214 // If this is an argument, we can export it if the BB is the entry block or
1215 // if it is already exported.
1216 if (isa<Argument>(V)) {
1217 if (FromBB == &FromBB->getParent()->getEntryBlock())
1220 // Otherwise, can only export this if it is already exported.
1221 return FuncInfo.isExportedInst(V);
1224 // Otherwise, constants can always be exported.
1228 static bool InBlock(const Value *V, const BasicBlock *BB) {
1229 if (const Instruction *I = dyn_cast<Instruction>(V))
1230 return I->getParent() == BB;
1234 /// EmitBranchForMergedCondition - Helper method for FindMergedConditions.
1235 /// This function emits a branch and is used at the leaves of an OR or an
1236 /// AND operator tree.
1239 SelectionDAGBuilder::EmitBranchForMergedCondition(const Value *Cond,
1240 MachineBasicBlock *TBB,
1241 MachineBasicBlock *FBB,
1242 MachineBasicBlock *CurBB,
1243 MachineBasicBlock *SwitchBB) {
1244 const BasicBlock *BB = CurBB->getBasicBlock();
1246 // If the leaf of the tree is a comparison, merge the condition into
1248 if (const CmpInst *BOp = dyn_cast<CmpInst>(Cond)) {
1249 // The operands of the cmp have to be in this block. We don't know
1250 // how to export them from some other block. If this is the first block
1251 // of the sequence, no exporting is needed.
1252 if (CurBB == SwitchBB ||
1253 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
1254 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) {
1255 ISD::CondCode Condition;
1256 if (const ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
1257 Condition = getICmpCondCode(IC->getPredicate());
1258 } else if (const FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) {
1259 Condition = getFCmpCondCode(FC->getPredicate());
1261 Condition = ISD::SETEQ; // silence warning.
1262 llvm_unreachable("Unknown compare instruction");
1265 CaseBlock CB(Condition, BOp->getOperand(0),
1266 BOp->getOperand(1), NULL, TBB, FBB, CurBB);
1267 SwitchCases.push_back(CB);
1272 // Create a CaseBlock record representing this branch.
1273 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()),
1274 NULL, TBB, FBB, CurBB);
1275 SwitchCases.push_back(CB);
1278 /// FindMergedConditions - If Cond is an expression like
1279 void SelectionDAGBuilder::FindMergedConditions(const Value *Cond,
1280 MachineBasicBlock *TBB,
1281 MachineBasicBlock *FBB,
1282 MachineBasicBlock *CurBB,
1283 MachineBasicBlock *SwitchBB,
1285 // If this node is not part of the or/and tree, emit it as a branch.
1286 const Instruction *BOp = dyn_cast<Instruction>(Cond);
1287 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
1288 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
1289 BOp->getParent() != CurBB->getBasicBlock() ||
1290 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
1291 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
1292 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB);
1296 // Create TmpBB after CurBB.
1297 MachineFunction::iterator BBI = CurBB;
1298 MachineFunction &MF = DAG.getMachineFunction();
1299 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock());
1300 CurBB->getParent()->insert(++BBI, TmpBB);
1302 if (Opc == Instruction::Or) {
1303 // Codegen X | Y as:
1311 // Emit the LHS condition.
1312 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, SwitchBB, Opc);
1314 // Emit the RHS condition into TmpBB.
1315 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
1317 assert(Opc == Instruction::And && "Unknown merge op!");
1318 // Codegen X & Y as:
1325 // This requires creation of TmpBB after CurBB.
1327 // Emit the LHS condition.
1328 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, SwitchBB, Opc);
1330 // Emit the RHS condition into TmpBB.
1331 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
1335 /// If the set of cases should be emitted as a series of branches, return true.
1336 /// If we should emit this as a bunch of and/or'd together conditions, return
1339 SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases){
1340 if (Cases.size() != 2) return true;
1342 // If this is two comparisons of the same values or'd or and'd together, they
1343 // will get folded into a single comparison, so don't emit two blocks.
1344 if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
1345 Cases[0].CmpRHS == Cases[1].CmpRHS) ||
1346 (Cases[0].CmpRHS == Cases[1].CmpLHS &&
1347 Cases[0].CmpLHS == Cases[1].CmpRHS)) {
1351 // Handle: (X != null) | (Y != null) --> (X|Y) != 0
1352 // Handle: (X == null) & (Y == null) --> (X|Y) == 0
1353 if (Cases[0].CmpRHS == Cases[1].CmpRHS &&
1354 Cases[0].CC == Cases[1].CC &&
1355 isa<Constant>(Cases[0].CmpRHS) &&
1356 cast<Constant>(Cases[0].CmpRHS)->isNullValue()) {
1357 if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB)
1359 if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB)
1366 void SelectionDAGBuilder::visitBr(const BranchInst &I) {
1367 MachineBasicBlock *BrMBB = FuncInfo.MBB;
1369 // Update machine-CFG edges.
1370 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
1372 // Figure out which block is immediately after the current one.
1373 MachineBasicBlock *NextBlock = 0;
1374 MachineFunction::iterator BBI = BrMBB;
1375 if (++BBI != FuncInfo.MF->end())
1378 if (I.isUnconditional()) {
1379 // Update machine-CFG edges.
1380 BrMBB->addSuccessor(Succ0MBB);
1382 // If this is not a fall-through branch, emit the branch.
1383 if (Succ0MBB != NextBlock)
1384 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
1385 MVT::Other, getControlRoot(),
1386 DAG.getBasicBlock(Succ0MBB)));
1391 // If this condition is one of the special cases we handle, do special stuff
1393 const Value *CondVal = I.getCondition();
1394 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
1396 // If this is a series of conditions that are or'd or and'd together, emit
1397 // this as a sequence of branches instead of setcc's with and/or operations.
1398 // For example, instead of something like:
1411 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
1412 if (BOp->hasOneUse() &&
1413 (BOp->getOpcode() == Instruction::And ||
1414 BOp->getOpcode() == Instruction::Or)) {
1415 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB,
1417 // If the compares in later blocks need to use values not currently
1418 // exported from this block, export them now. This block should always
1419 // be the first entry.
1420 assert(SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!");
1422 // Allow some cases to be rejected.
1423 if (ShouldEmitAsBranches(SwitchCases)) {
1424 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1425 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1426 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1429 // Emit the branch for this block.
1430 visitSwitchCase(SwitchCases[0], BrMBB);
1431 SwitchCases.erase(SwitchCases.begin());
1435 // Okay, we decided not to do this, remove any inserted MBB's and clear
1437 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
1438 FuncInfo.MF->erase(SwitchCases[i].ThisBB);
1440 SwitchCases.clear();
1444 // Create a CaseBlock record representing this branch.
1445 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()),
1446 NULL, Succ0MBB, Succ1MBB, BrMBB);
1448 // Use visitSwitchCase to actually insert the fast branch sequence for this
1450 visitSwitchCase(CB, BrMBB);
1453 /// visitSwitchCase - Emits the necessary code to represent a single node in
1454 /// the binary search tree resulting from lowering a switch instruction.
1455 void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB,
1456 MachineBasicBlock *SwitchBB) {
1458 SDValue CondLHS = getValue(CB.CmpLHS);
1459 DebugLoc dl = getCurDebugLoc();
1461 // Build the setcc now.
1462 if (CB.CmpMHS == NULL) {
1463 // Fold "(X == true)" to X and "(X == false)" to !X to
1464 // handle common cases produced by branch lowering.
1465 if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) &&
1466 CB.CC == ISD::SETEQ)
1468 else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) &&
1469 CB.CC == ISD::SETEQ) {
1470 SDValue True = DAG.getConstant(1, CondLHS.getValueType());
1471 Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True);
1473 Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1475 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now");
1477 const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue();
1478 const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue();
1480 SDValue CmpOp = getValue(CB.CmpMHS);
1481 EVT VT = CmpOp.getValueType();
1483 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) {
1484 Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, VT),
1487 SDValue SUB = DAG.getNode(ISD::SUB, dl,
1488 VT, CmpOp, DAG.getConstant(Low, VT));
1489 Cond = DAG.getSetCC(dl, MVT::i1, SUB,
1490 DAG.getConstant(High-Low, VT), ISD::SETULE);
1494 // Update successor info
1495 SwitchBB->addSuccessor(CB.TrueBB);
1496 SwitchBB->addSuccessor(CB.FalseBB);
1498 // Set NextBlock to be the MBB immediately after the current one, if any.
1499 // This is used to avoid emitting unnecessary branches to the next block.
1500 MachineBasicBlock *NextBlock = 0;
1501 MachineFunction::iterator BBI = SwitchBB;
1502 if (++BBI != FuncInfo.MF->end())
1505 // If the lhs block is the next block, invert the condition so that we can
1506 // fall through to the lhs instead of the rhs block.
1507 if (CB.TrueBB == NextBlock) {
1508 std::swap(CB.TrueBB, CB.FalseBB);
1509 SDValue True = DAG.getConstant(1, Cond.getValueType());
1510 Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True);
1513 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1514 MVT::Other, getControlRoot(), Cond,
1515 DAG.getBasicBlock(CB.TrueBB));
1517 // Insert the false branch.
1518 if (CB.FalseBB != NextBlock)
1519 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1520 DAG.getBasicBlock(CB.FalseBB));
1522 DAG.setRoot(BrCond);
1525 /// visitJumpTable - Emit JumpTable node in the current MBB
1526 void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) {
1527 // Emit the code for the jump table
1528 assert(JT.Reg != -1U && "Should lower JT Header first!");
1529 EVT PTy = TLI.getPointerTy();
1530 SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(),
1532 SDValue Table = DAG.getJumpTable(JT.JTI, PTy);
1533 SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurDebugLoc(),
1534 MVT::Other, Index.getValue(1),
1536 DAG.setRoot(BrJumpTable);
1539 /// visitJumpTableHeader - This function emits necessary code to produce index
1540 /// in the JumpTable from switch case.
1541 void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT,
1542 JumpTableHeader &JTH,
1543 MachineBasicBlock *SwitchBB) {
1544 // Subtract the lowest switch case value from the value being switched on and
1545 // conditional branch to default mbb if the result is greater than the
1546 // difference between smallest and largest cases.
1547 SDValue SwitchOp = getValue(JTH.SValue);
1548 EVT VT = SwitchOp.getValueType();
1549 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1550 DAG.getConstant(JTH.First, VT));
1552 // The SDNode we just created, which holds the value being switched on minus
1553 // the smallest case value, needs to be copied to a virtual register so it
1554 // can be used as an index into the jump table in a subsequent basic block.
1555 // This value may be smaller or larger than the target's pointer type, and
1556 // therefore require extension or truncating.
1557 SwitchOp = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(), TLI.getPointerTy());
1559 unsigned JumpTableReg = FuncInfo.CreateReg(TLI.getPointerTy());
1560 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1561 JumpTableReg, SwitchOp);
1562 JT.Reg = JumpTableReg;
1564 // Emit the range check for the jump table, and branch to the default block
1565 // for the switch statement if the value being switched on exceeds the largest
1566 // case in the switch.
1567 SDValue CMP = DAG.getSetCC(getCurDebugLoc(),
1568 TLI.getSetCCResultType(Sub.getValueType()), Sub,
1569 DAG.getConstant(JTH.Last-JTH.First,VT),
1572 // Set NextBlock to be the MBB immediately after the current one, if any.
1573 // This is used to avoid emitting unnecessary branches to the next block.
1574 MachineBasicBlock *NextBlock = 0;
1575 MachineFunction::iterator BBI = SwitchBB;
1577 if (++BBI != FuncInfo.MF->end())
1580 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1581 MVT::Other, CopyTo, CMP,
1582 DAG.getBasicBlock(JT.Default));
1584 if (JT.MBB != NextBlock)
1585 BrCond = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrCond,
1586 DAG.getBasicBlock(JT.MBB));
1588 DAG.setRoot(BrCond);
1591 /// visitBitTestHeader - This function emits necessary code to produce value
1592 /// suitable for "bit tests"
1593 void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B,
1594 MachineBasicBlock *SwitchBB) {
1595 // Subtract the minimum value
1596 SDValue SwitchOp = getValue(B.SValue);
1597 EVT VT = SwitchOp.getValueType();
1598 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1599 DAG.getConstant(B.First, VT));
1602 SDValue RangeCmp = DAG.getSetCC(getCurDebugLoc(),
1603 TLI.getSetCCResultType(Sub.getValueType()),
1604 Sub, DAG.getConstant(B.Range, VT),
1607 SDValue ShiftOp = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(),
1608 TLI.getPointerTy());
1610 B.Reg = FuncInfo.CreateReg(TLI.getPointerTy());
1611 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1614 // Set NextBlock to be the MBB immediately after the current one, if any.
1615 // This is used to avoid emitting unnecessary branches to the next block.
1616 MachineBasicBlock *NextBlock = 0;
1617 MachineFunction::iterator BBI = SwitchBB;
1618 if (++BBI != FuncInfo.MF->end())
1621 MachineBasicBlock* MBB = B.Cases[0].ThisBB;
1623 SwitchBB->addSuccessor(B.Default);
1624 SwitchBB->addSuccessor(MBB);
1626 SDValue BrRange = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1627 MVT::Other, CopyTo, RangeCmp,
1628 DAG.getBasicBlock(B.Default));
1630 if (MBB != NextBlock)
1631 BrRange = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, CopyTo,
1632 DAG.getBasicBlock(MBB));
1634 DAG.setRoot(BrRange);
1637 /// visitBitTestCase - this function produces one "bit test"
1638 void SelectionDAGBuilder::visitBitTestCase(MachineBasicBlock* NextMBB,
1641 MachineBasicBlock *SwitchBB) {
1642 SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(), Reg,
1643 TLI.getPointerTy());
1645 if (CountPopulation_64(B.Mask) == 1) {
1646 // Testing for a single bit; just compare the shift count with what it
1647 // would need to be to shift a 1 bit in that position.
1648 Cmp = DAG.getSetCC(getCurDebugLoc(),
1649 TLI.getSetCCResultType(ShiftOp.getValueType()),
1651 DAG.getConstant(CountTrailingZeros_64(B.Mask),
1652 TLI.getPointerTy()),
1655 // Make desired shift
1656 SDValue SwitchVal = DAG.getNode(ISD::SHL, getCurDebugLoc(),
1658 DAG.getConstant(1, TLI.getPointerTy()),
1661 // Emit bit tests and jumps
1662 SDValue AndOp = DAG.getNode(ISD::AND, getCurDebugLoc(),
1663 TLI.getPointerTy(), SwitchVal,
1664 DAG.getConstant(B.Mask, TLI.getPointerTy()));
1665 Cmp = DAG.getSetCC(getCurDebugLoc(),
1666 TLI.getSetCCResultType(AndOp.getValueType()),
1667 AndOp, DAG.getConstant(0, TLI.getPointerTy()),
1671 SwitchBB->addSuccessor(B.TargetBB);
1672 SwitchBB->addSuccessor(NextMBB);
1674 SDValue BrAnd = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1675 MVT::Other, getControlRoot(),
1676 Cmp, DAG.getBasicBlock(B.TargetBB));
1678 // Set NextBlock to be the MBB immediately after the current one, if any.
1679 // This is used to avoid emitting unnecessary branches to the next block.
1680 MachineBasicBlock *NextBlock = 0;
1681 MachineFunction::iterator BBI = SwitchBB;
1682 if (++BBI != FuncInfo.MF->end())
1685 if (NextMBB != NextBlock)
1686 BrAnd = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrAnd,
1687 DAG.getBasicBlock(NextMBB));
1692 void SelectionDAGBuilder::visitInvoke(const InvokeInst &I) {
1693 MachineBasicBlock *InvokeMBB = FuncInfo.MBB;
1695 // Retrieve successors.
1696 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
1697 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)];
1699 const Value *Callee(I.getCalledValue());
1700 if (isa<InlineAsm>(Callee))
1703 LowerCallTo(&I, getValue(Callee), false, LandingPad);
1705 // If the value of the invoke is used outside of its defining block, make it
1706 // available as a virtual register.
1707 CopyToExportRegsIfNeeded(&I);
1709 // Update successor info
1710 InvokeMBB->addSuccessor(Return);
1711 InvokeMBB->addSuccessor(LandingPad);
1713 // Drop into normal successor.
1714 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
1715 MVT::Other, getControlRoot(),
1716 DAG.getBasicBlock(Return)));
1719 void SelectionDAGBuilder::visitUnwind(const UnwindInst &I) {
1722 /// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for
1723 /// small case ranges).
1724 bool SelectionDAGBuilder::handleSmallSwitchRange(CaseRec& CR,
1725 CaseRecVector& WorkList,
1727 MachineBasicBlock *Default,
1728 MachineBasicBlock *SwitchBB) {
1729 Case& BackCase = *(CR.Range.second-1);
1731 // Size is the number of Cases represented by this range.
1732 size_t Size = CR.Range.second - CR.Range.first;
1736 // Get the MachineFunction which holds the current MBB. This is used when
1737 // inserting any additional MBBs necessary to represent the switch.
1738 MachineFunction *CurMF = FuncInfo.MF;
1740 // Figure out which block is immediately after the current one.
1741 MachineBasicBlock *NextBlock = 0;
1742 MachineFunction::iterator BBI = CR.CaseBB;
1744 if (++BBI != FuncInfo.MF->end())
1747 // TODO: If any two of the cases has the same destination, and if one value
1748 // is the same as the other, but has one bit unset that the other has set,
1749 // use bit manipulation to do two compares at once. For example:
1750 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
1752 // Rearrange the case blocks so that the last one falls through if possible.
1753 if (NextBlock && Default != NextBlock && BackCase.BB != NextBlock) {
1754 // The last case block won't fall through into 'NextBlock' if we emit the
1755 // branches in this order. See if rearranging a case value would help.
1756 for (CaseItr I = CR.Range.first, E = CR.Range.second-1; I != E; ++I) {
1757 if (I->BB == NextBlock) {
1758 std::swap(*I, BackCase);
1764 // Create a CaseBlock record representing a conditional branch to
1765 // the Case's target mbb if the value being switched on SV is equal
1767 MachineBasicBlock *CurBlock = CR.CaseBB;
1768 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
1769 MachineBasicBlock *FallThrough;
1771 FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock());
1772 CurMF->insert(BBI, FallThrough);
1774 // Put SV in a virtual register to make it available from the new blocks.
1775 ExportFromCurrentBlock(SV);
1777 // If the last case doesn't match, go to the default block.
1778 FallThrough = Default;
1781 const Value *RHS, *LHS, *MHS;
1783 if (I->High == I->Low) {
1784 // This is just small small case range :) containing exactly 1 case
1786 LHS = SV; RHS = I->High; MHS = NULL;
1789 LHS = I->Low; MHS = SV; RHS = I->High;
1791 CaseBlock CB(CC, LHS, RHS, MHS, I->BB, FallThrough, CurBlock);
1793 // If emitting the first comparison, just call visitSwitchCase to emit the
1794 // code into the current block. Otherwise, push the CaseBlock onto the
1795 // vector to be later processed by SDISel, and insert the node's MBB
1796 // before the next MBB.
1797 if (CurBlock == SwitchBB)
1798 visitSwitchCase(CB, SwitchBB);
1800 SwitchCases.push_back(CB);
1802 CurBlock = FallThrough;
1808 static inline bool areJTsAllowed(const TargetLowering &TLI) {
1809 return !DisableJumpTables &&
1810 (TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
1811 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other));
1814 static APInt ComputeRange(const APInt &First, const APInt &Last) {
1815 APInt LastExt(Last), FirstExt(First);
1816 uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1;
1817 LastExt.sext(BitWidth); FirstExt.sext(BitWidth);
1818 return (LastExt - FirstExt + 1ULL);
1821 /// handleJTSwitchCase - Emit jumptable for current switch case range
1822 bool SelectionDAGBuilder::handleJTSwitchCase(CaseRec& CR,
1823 CaseRecVector& WorkList,
1825 MachineBasicBlock* Default,
1826 MachineBasicBlock *SwitchBB) {
1827 Case& FrontCase = *CR.Range.first;
1828 Case& BackCase = *(CR.Range.second-1);
1830 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
1831 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
1833 APInt TSize(First.getBitWidth(), 0);
1834 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1838 if (!areJTsAllowed(TLI) || TSize.ult(4))
1841 APInt Range = ComputeRange(First, Last);
1842 double Density = TSize.roundToDouble() / Range.roundToDouble();
1846 DEBUG(dbgs() << "Lowering jump table\n"
1847 << "First entry: " << First << ". Last entry: " << Last << '\n'
1848 << "Range: " << Range
1849 << "Size: " << TSize << ". Density: " << Density << "\n\n");
1851 // Get the MachineFunction which holds the current MBB. This is used when
1852 // inserting any additional MBBs necessary to represent the switch.
1853 MachineFunction *CurMF = FuncInfo.MF;
1855 // Figure out which block is immediately after the current one.
1856 MachineFunction::iterator BBI = CR.CaseBB;
1859 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1861 // Create a new basic block to hold the code for loading the address
1862 // of the jump table, and jumping to it. Update successor information;
1863 // we will either branch to the default case for the switch, or the jump
1865 MachineBasicBlock *JumpTableBB = CurMF->CreateMachineBasicBlock(LLVMBB);
1866 CurMF->insert(BBI, JumpTableBB);
1867 CR.CaseBB->addSuccessor(Default);
1868 CR.CaseBB->addSuccessor(JumpTableBB);
1870 // Build a vector of destination BBs, corresponding to each target
1871 // of the jump table. If the value of the jump table slot corresponds to
1872 // a case statement, push the case's BB onto the vector, otherwise, push
1874 std::vector<MachineBasicBlock*> DestBBs;
1876 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) {
1877 const APInt &Low = cast<ConstantInt>(I->Low)->getValue();
1878 const APInt &High = cast<ConstantInt>(I->High)->getValue();
1880 if (Low.sle(TEI) && TEI.sle(High)) {
1881 DestBBs.push_back(I->BB);
1885 DestBBs.push_back(Default);
1889 // Update successor info. Add one edge to each unique successor.
1890 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs());
1891 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(),
1892 E = DestBBs.end(); I != E; ++I) {
1893 if (!SuccsHandled[(*I)->getNumber()]) {
1894 SuccsHandled[(*I)->getNumber()] = true;
1895 JumpTableBB->addSuccessor(*I);
1899 // Create a jump table index for this jump table.
1900 unsigned JTEncoding = TLI.getJumpTableEncoding();
1901 unsigned JTI = CurMF->getOrCreateJumpTableInfo(JTEncoding)
1902 ->createJumpTableIndex(DestBBs);
1904 // Set the jump table information so that we can codegen it as a second
1905 // MachineBasicBlock
1906 JumpTable JT(-1U, JTI, JumpTableBB, Default);
1907 JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == SwitchBB));
1908 if (CR.CaseBB == SwitchBB)
1909 visitJumpTableHeader(JT, JTH, SwitchBB);
1911 JTCases.push_back(JumpTableBlock(JTH, JT));
1916 /// handleBTSplitSwitchCase - emit comparison and split binary search tree into
1918 bool SelectionDAGBuilder::handleBTSplitSwitchCase(CaseRec& CR,
1919 CaseRecVector& WorkList,
1921 MachineBasicBlock *Default,
1922 MachineBasicBlock *SwitchBB) {
1923 // Get the MachineFunction which holds the current MBB. This is used when
1924 // inserting any additional MBBs necessary to represent the switch.
1925 MachineFunction *CurMF = FuncInfo.MF;
1927 // Figure out which block is immediately after the current one.
1928 MachineFunction::iterator BBI = CR.CaseBB;
1931 Case& FrontCase = *CR.Range.first;
1932 Case& BackCase = *(CR.Range.second-1);
1933 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1935 // Size is the number of Cases represented by this range.
1936 unsigned Size = CR.Range.second - CR.Range.first;
1938 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
1939 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
1941 CaseItr Pivot = CR.Range.first + Size/2;
1943 // Select optimal pivot, maximizing sum density of LHS and RHS. This will
1944 // (heuristically) allow us to emit JumpTable's later.
1945 APInt TSize(First.getBitWidth(), 0);
1946 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1950 APInt LSize = FrontCase.size();
1951 APInt RSize = TSize-LSize;
1952 DEBUG(dbgs() << "Selecting best pivot: \n"
1953 << "First: " << First << ", Last: " << Last <<'\n'
1954 << "LSize: " << LSize << ", RSize: " << RSize << '\n');
1955 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second;
1957 const APInt &LEnd = cast<ConstantInt>(I->High)->getValue();
1958 const APInt &RBegin = cast<ConstantInt>(J->Low)->getValue();
1959 APInt Range = ComputeRange(LEnd, RBegin);
1960 assert((Range - 2ULL).isNonNegative() &&
1961 "Invalid case distance");
1962 double LDensity = (double)LSize.roundToDouble() /
1963 (LEnd - First + 1ULL).roundToDouble();
1964 double RDensity = (double)RSize.roundToDouble() /
1965 (Last - RBegin + 1ULL).roundToDouble();
1966 double Metric = Range.logBase2()*(LDensity+RDensity);
1967 // Should always split in some non-trivial place
1968 DEBUG(dbgs() <<"=>Step\n"
1969 << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n'
1970 << "LDensity: " << LDensity
1971 << ", RDensity: " << RDensity << '\n'
1972 << "Metric: " << Metric << '\n');
1973 if (FMetric < Metric) {
1976 DEBUG(dbgs() << "Current metric set to: " << FMetric << '\n');
1982 if (areJTsAllowed(TLI)) {
1983 // If our case is dense we *really* should handle it earlier!
1984 assert((FMetric > 0) && "Should handle dense range earlier!");
1986 Pivot = CR.Range.first + Size/2;
1989 CaseRange LHSR(CR.Range.first, Pivot);
1990 CaseRange RHSR(Pivot, CR.Range.second);
1991 Constant *C = Pivot->Low;
1992 MachineBasicBlock *FalseBB = 0, *TrueBB = 0;
1994 // We know that we branch to the LHS if the Value being switched on is
1995 // less than the Pivot value, C. We use this to optimize our binary
1996 // tree a bit, by recognizing that if SV is greater than or equal to the
1997 // LHS's Case Value, and that Case Value is exactly one less than the
1998 // Pivot's Value, then we can branch directly to the LHS's Target,
1999 // rather than creating a leaf node for it.
2000 if ((LHSR.second - LHSR.first) == 1 &&
2001 LHSR.first->High == CR.GE &&
2002 cast<ConstantInt>(C)->getValue() ==
2003 (cast<ConstantInt>(CR.GE)->getValue() + 1LL)) {
2004 TrueBB = LHSR.first->BB;
2006 TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2007 CurMF->insert(BBI, TrueBB);
2008 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR));
2010 // Put SV in a virtual register to make it available from the new blocks.
2011 ExportFromCurrentBlock(SV);
2014 // Similar to the optimization above, if the Value being switched on is
2015 // known to be less than the Constant CR.LT, and the current Case Value
2016 // is CR.LT - 1, then we can branch directly to the target block for
2017 // the current Case Value, rather than emitting a RHS leaf node for it.
2018 if ((RHSR.second - RHSR.first) == 1 && CR.LT &&
2019 cast<ConstantInt>(RHSR.first->Low)->getValue() ==
2020 (cast<ConstantInt>(CR.LT)->getValue() - 1LL)) {
2021 FalseBB = RHSR.first->BB;
2023 FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2024 CurMF->insert(BBI, FalseBB);
2025 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR));
2027 // Put SV in a virtual register to make it available from the new blocks.
2028 ExportFromCurrentBlock(SV);
2031 // Create a CaseBlock record representing a conditional branch to
2032 // the LHS node if the value being switched on SV is less than C.
2033 // Otherwise, branch to LHS.
2034 CaseBlock CB(ISD::SETLT, SV, C, NULL, TrueBB, FalseBB, CR.CaseBB);
2036 if (CR.CaseBB == SwitchBB)
2037 visitSwitchCase(CB, SwitchBB);
2039 SwitchCases.push_back(CB);
2044 /// handleBitTestsSwitchCase - if current case range has few destination and
2045 /// range span less, than machine word bitwidth, encode case range into series
2046 /// of masks and emit bit tests with these masks.
2047 bool SelectionDAGBuilder::handleBitTestsSwitchCase(CaseRec& CR,
2048 CaseRecVector& WorkList,
2050 MachineBasicBlock* Default,
2051 MachineBasicBlock *SwitchBB){
2052 EVT PTy = TLI.getPointerTy();
2053 unsigned IntPtrBits = PTy.getSizeInBits();
2055 Case& FrontCase = *CR.Range.first;
2056 Case& BackCase = *(CR.Range.second-1);
2058 // Get the MachineFunction which holds the current MBB. This is used when
2059 // inserting any additional MBBs necessary to represent the switch.
2060 MachineFunction *CurMF = FuncInfo.MF;
2062 // If target does not have legal shift left, do not emit bit tests at all.
2063 if (!TLI.isOperationLegal(ISD::SHL, TLI.getPointerTy()))
2067 for (CaseItr I = CR.Range.first, E = CR.Range.second;
2069 // Single case counts one, case range - two.
2070 numCmps += (I->Low == I->High ? 1 : 2);
2073 // Count unique destinations
2074 SmallSet<MachineBasicBlock*, 4> Dests;
2075 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2076 Dests.insert(I->BB);
2077 if (Dests.size() > 3)
2078 // Don't bother the code below, if there are too much unique destinations
2081 DEBUG(dbgs() << "Total number of unique destinations: "
2082 << Dests.size() << '\n'
2083 << "Total number of comparisons: " << numCmps << '\n');
2085 // Compute span of values.
2086 const APInt& minValue = cast<ConstantInt>(FrontCase.Low)->getValue();
2087 const APInt& maxValue = cast<ConstantInt>(BackCase.High)->getValue();
2088 APInt cmpRange = maxValue - minValue;
2090 DEBUG(dbgs() << "Compare range: " << cmpRange << '\n'
2091 << "Low bound: " << minValue << '\n'
2092 << "High bound: " << maxValue << '\n');
2094 if (cmpRange.uge(IntPtrBits) ||
2095 (!(Dests.size() == 1 && numCmps >= 3) &&
2096 !(Dests.size() == 2 && numCmps >= 5) &&
2097 !(Dests.size() >= 3 && numCmps >= 6)))
2100 DEBUG(dbgs() << "Emitting bit tests\n");
2101 APInt lowBound = APInt::getNullValue(cmpRange.getBitWidth());
2103 // Optimize the case where all the case values fit in a
2104 // word without having to subtract minValue. In this case,
2105 // we can optimize away the subtraction.
2106 if (minValue.isNonNegative() && maxValue.slt(IntPtrBits)) {
2107 cmpRange = maxValue;
2109 lowBound = minValue;
2112 CaseBitsVector CasesBits;
2113 unsigned i, count = 0;
2115 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2116 MachineBasicBlock* Dest = I->BB;
2117 for (i = 0; i < count; ++i)
2118 if (Dest == CasesBits[i].BB)
2122 assert((count < 3) && "Too much destinations to test!");
2123 CasesBits.push_back(CaseBits(0, Dest, 0));
2127 const APInt& lowValue = cast<ConstantInt>(I->Low)->getValue();
2128 const APInt& highValue = cast<ConstantInt>(I->High)->getValue();
2130 uint64_t lo = (lowValue - lowBound).getZExtValue();
2131 uint64_t hi = (highValue - lowBound).getZExtValue();
2133 for (uint64_t j = lo; j <= hi; j++) {
2134 CasesBits[i].Mask |= 1ULL << j;
2135 CasesBits[i].Bits++;
2139 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp());
2143 // Figure out which block is immediately after the current one.
2144 MachineFunction::iterator BBI = CR.CaseBB;
2147 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2149 DEBUG(dbgs() << "Cases:\n");
2150 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) {
2151 DEBUG(dbgs() << "Mask: " << CasesBits[i].Mask
2152 << ", Bits: " << CasesBits[i].Bits
2153 << ", BB: " << CasesBits[i].BB << '\n');
2155 MachineBasicBlock *CaseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2156 CurMF->insert(BBI, CaseBB);
2157 BTC.push_back(BitTestCase(CasesBits[i].Mask,
2161 // Put SV in a virtual register to make it available from the new blocks.
2162 ExportFromCurrentBlock(SV);
2165 BitTestBlock BTB(lowBound, cmpRange, SV,
2166 -1U, (CR.CaseBB == SwitchBB),
2167 CR.CaseBB, Default, BTC);
2169 if (CR.CaseBB == SwitchBB)
2170 visitBitTestHeader(BTB, SwitchBB);
2172 BitTestCases.push_back(BTB);
2177 /// Clusterify - Transform simple list of Cases into list of CaseRange's
2178 size_t SelectionDAGBuilder::Clusterify(CaseVector& Cases,
2179 const SwitchInst& SI) {
2182 // Start with "simple" cases
2183 for (size_t i = 1; i < SI.getNumSuccessors(); ++i) {
2184 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SI.getSuccessor(i)];
2185 Cases.push_back(Case(SI.getSuccessorValue(i),
2186 SI.getSuccessorValue(i),
2189 std::sort(Cases.begin(), Cases.end(), CaseCmp());
2191 // Merge case into clusters
2192 if (Cases.size() >= 2)
2193 // Must recompute end() each iteration because it may be
2194 // invalidated by erase if we hold on to it
2195 for (CaseItr I = Cases.begin(), J = ++(Cases.begin()); J != Cases.end(); ) {
2196 const APInt& nextValue = cast<ConstantInt>(J->Low)->getValue();
2197 const APInt& currentValue = cast<ConstantInt>(I->High)->getValue();
2198 MachineBasicBlock* nextBB = J->BB;
2199 MachineBasicBlock* currentBB = I->BB;
2201 // If the two neighboring cases go to the same destination, merge them
2202 // into a single case.
2203 if ((nextValue - currentValue == 1) && (currentBB == nextBB)) {
2211 for (CaseItr I=Cases.begin(), E=Cases.end(); I!=E; ++I, ++numCmps) {
2212 if (I->Low != I->High)
2213 // A range counts double, since it requires two compares.
2220 void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) {
2221 MachineBasicBlock *SwitchMBB = FuncInfo.MBB;
2223 // Figure out which block is immediately after the current one.
2224 MachineBasicBlock *NextBlock = 0;
2225 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()];
2227 // If there is only the default destination, branch to it if it is not the
2228 // next basic block. Otherwise, just fall through.
2229 if (SI.getNumOperands() == 2) {
2230 // Update machine-CFG edges.
2232 // If this is not a fall-through branch, emit the branch.
2233 SwitchMBB->addSuccessor(Default);
2234 if (Default != NextBlock)
2235 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
2236 MVT::Other, getControlRoot(),
2237 DAG.getBasicBlock(Default)));
2242 // If there are any non-default case statements, create a vector of Cases
2243 // representing each one, and sort the vector so that we can efficiently
2244 // create a binary search tree from them.
2246 size_t numCmps = Clusterify(Cases, SI);
2247 DEBUG(dbgs() << "Clusterify finished. Total clusters: " << Cases.size()
2248 << ". Total compares: " << numCmps << '\n');
2251 // Get the Value to be switched on and default basic blocks, which will be
2252 // inserted into CaseBlock records, representing basic blocks in the binary
2254 const Value *SV = SI.getOperand(0);
2256 // Push the initial CaseRec onto the worklist
2257 CaseRecVector WorkList;
2258 WorkList.push_back(CaseRec(SwitchMBB,0,0,
2259 CaseRange(Cases.begin(),Cases.end())));
2261 while (!WorkList.empty()) {
2262 // Grab a record representing a case range to process off the worklist
2263 CaseRec CR = WorkList.back();
2264 WorkList.pop_back();
2266 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2269 // If the range has few cases (two or less) emit a series of specific
2271 if (handleSmallSwitchRange(CR, WorkList, SV, Default, SwitchMBB))
2274 // If the switch has more than 5 blocks, and at least 40% dense, and the
2275 // target supports indirect branches, then emit a jump table rather than
2276 // lowering the switch to a binary tree of conditional branches.
2277 if (handleJTSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2280 // Emit binary tree. We need to pick a pivot, and push left and right ranges
2281 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call.
2282 handleBTSplitSwitchCase(CR, WorkList, SV, Default, SwitchMBB);
2286 void SelectionDAGBuilder::visitIndirectBr(const IndirectBrInst &I) {
2287 MachineBasicBlock *IndirectBrMBB = FuncInfo.MBB;
2289 // Update machine-CFG edges with unique successors.
2290 SmallVector<BasicBlock*, 32> succs;
2291 succs.reserve(I.getNumSuccessors());
2292 for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i)
2293 succs.push_back(I.getSuccessor(i));
2294 array_pod_sort(succs.begin(), succs.end());
2295 succs.erase(std::unique(succs.begin(), succs.end()), succs.end());
2296 for (unsigned i = 0, e = succs.size(); i != e; ++i)
2297 IndirectBrMBB->addSuccessor(FuncInfo.MBBMap[succs[i]]);
2299 DAG.setRoot(DAG.getNode(ISD::BRIND, getCurDebugLoc(),
2300 MVT::Other, getControlRoot(),
2301 getValue(I.getAddress())));
2304 void SelectionDAGBuilder::visitFSub(const User &I) {
2305 // -0.0 - X --> fneg
2306 const Type *Ty = I.getType();
2307 if (Ty->isVectorTy()) {
2308 if (ConstantVector *CV = dyn_cast<ConstantVector>(I.getOperand(0))) {
2309 const VectorType *DestTy = cast<VectorType>(I.getType());
2310 const Type *ElTy = DestTy->getElementType();
2311 unsigned VL = DestTy->getNumElements();
2312 std::vector<Constant*> NZ(VL, ConstantFP::getNegativeZero(ElTy));
2313 Constant *CNZ = ConstantVector::get(&NZ[0], NZ.size());
2315 SDValue Op2 = getValue(I.getOperand(1));
2316 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(),
2317 Op2.getValueType(), Op2));
2323 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I.getOperand(0)))
2324 if (CFP->isExactlyValue(ConstantFP::getNegativeZero(Ty)->getValueAPF())) {
2325 SDValue Op2 = getValue(I.getOperand(1));
2326 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(),
2327 Op2.getValueType(), Op2));
2331 visitBinary(I, ISD::FSUB);
2334 void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) {
2335 SDValue Op1 = getValue(I.getOperand(0));
2336 SDValue Op2 = getValue(I.getOperand(1));
2337 setValue(&I, DAG.getNode(OpCode, getCurDebugLoc(),
2338 Op1.getValueType(), Op1, Op2));
2341 void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) {
2342 SDValue Op1 = getValue(I.getOperand(0));
2343 SDValue Op2 = getValue(I.getOperand(1));
2344 if (!I.getType()->isVectorTy() &&
2345 Op2.getValueType() != TLI.getShiftAmountTy()) {
2346 // If the operand is smaller than the shift count type, promote it.
2347 EVT PTy = TLI.getPointerTy();
2348 EVT STy = TLI.getShiftAmountTy();
2349 if (STy.bitsGT(Op2.getValueType()))
2350 Op2 = DAG.getNode(ISD::ANY_EXTEND, getCurDebugLoc(),
2351 TLI.getShiftAmountTy(), Op2);
2352 // If the operand is larger than the shift count type but the shift
2353 // count type has enough bits to represent any shift value, truncate
2354 // it now. This is a common case and it exposes the truncate to
2355 // optimization early.
2356 else if (STy.getSizeInBits() >=
2357 Log2_32_Ceil(Op2.getValueType().getSizeInBits()))
2358 Op2 = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2359 TLI.getShiftAmountTy(), Op2);
2360 // Otherwise we'll need to temporarily settle for some other
2361 // convenient type; type legalization will make adjustments as
2363 else if (PTy.bitsLT(Op2.getValueType()))
2364 Op2 = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2365 TLI.getPointerTy(), Op2);
2366 else if (PTy.bitsGT(Op2.getValueType()))
2367 Op2 = DAG.getNode(ISD::ANY_EXTEND, getCurDebugLoc(),
2368 TLI.getPointerTy(), Op2);
2371 setValue(&I, DAG.getNode(Opcode, getCurDebugLoc(),
2372 Op1.getValueType(), Op1, Op2));
2375 void SelectionDAGBuilder::visitICmp(const User &I) {
2376 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2377 if (const ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2378 predicate = IC->getPredicate();
2379 else if (const ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2380 predicate = ICmpInst::Predicate(IC->getPredicate());
2381 SDValue Op1 = getValue(I.getOperand(0));
2382 SDValue Op2 = getValue(I.getOperand(1));
2383 ISD::CondCode Opcode = getICmpCondCode(predicate);
2385 EVT DestVT = TLI.getValueType(I.getType());
2386 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Opcode));
2389 void SelectionDAGBuilder::visitFCmp(const User &I) {
2390 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2391 if (const FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2392 predicate = FC->getPredicate();
2393 else if (const ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2394 predicate = FCmpInst::Predicate(FC->getPredicate());
2395 SDValue Op1 = getValue(I.getOperand(0));
2396 SDValue Op2 = getValue(I.getOperand(1));
2397 ISD::CondCode Condition = getFCmpCondCode(predicate);
2398 EVT DestVT = TLI.getValueType(I.getType());
2399 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Condition));
2402 void SelectionDAGBuilder::visitSelect(const User &I) {
2403 SmallVector<EVT, 4> ValueVTs;
2404 ComputeValueVTs(TLI, I.getType(), ValueVTs);
2405 unsigned NumValues = ValueVTs.size();
2406 if (NumValues == 0) return;
2408 SmallVector<SDValue, 4> Values(NumValues);
2409 SDValue Cond = getValue(I.getOperand(0));
2410 SDValue TrueVal = getValue(I.getOperand(1));
2411 SDValue FalseVal = getValue(I.getOperand(2));
2413 for (unsigned i = 0; i != NumValues; ++i)
2414 Values[i] = DAG.getNode(ISD::SELECT, getCurDebugLoc(),
2415 TrueVal.getNode()->getValueType(TrueVal.getResNo()+i),
2417 SDValue(TrueVal.getNode(),
2418 TrueVal.getResNo() + i),
2419 SDValue(FalseVal.getNode(),
2420 FalseVal.getResNo() + i));
2422 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2423 DAG.getVTList(&ValueVTs[0], NumValues),
2424 &Values[0], NumValues));
2427 void SelectionDAGBuilder::visitTrunc(const User &I) {
2428 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2429 SDValue N = getValue(I.getOperand(0));
2430 EVT DestVT = TLI.getValueType(I.getType());
2431 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), DestVT, N));
2434 void SelectionDAGBuilder::visitZExt(const User &I) {
2435 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2436 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2437 SDValue N = getValue(I.getOperand(0));
2438 EVT DestVT = TLI.getValueType(I.getType());
2439 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(), DestVT, N));
2442 void SelectionDAGBuilder::visitSExt(const User &I) {
2443 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2444 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2445 SDValue N = getValue(I.getOperand(0));
2446 EVT DestVT = TLI.getValueType(I.getType());
2447 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurDebugLoc(), DestVT, N));
2450 void SelectionDAGBuilder::visitFPTrunc(const User &I) {
2451 // FPTrunc is never a no-op cast, no need to check
2452 SDValue N = getValue(I.getOperand(0));
2453 EVT DestVT = TLI.getValueType(I.getType());
2454 setValue(&I, DAG.getNode(ISD::FP_ROUND, getCurDebugLoc(),
2455 DestVT, N, DAG.getIntPtrConstant(0)));
2458 void SelectionDAGBuilder::visitFPExt(const User &I){
2459 // FPTrunc is never a no-op cast, no need to check
2460 SDValue N = getValue(I.getOperand(0));
2461 EVT DestVT = TLI.getValueType(I.getType());
2462 setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurDebugLoc(), DestVT, N));
2465 void SelectionDAGBuilder::visitFPToUI(const User &I) {
2466 // FPToUI is never a no-op cast, no need to check
2467 SDValue N = getValue(I.getOperand(0));
2468 EVT DestVT = TLI.getValueType(I.getType());
2469 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurDebugLoc(), DestVT, N));
2472 void SelectionDAGBuilder::visitFPToSI(const User &I) {
2473 // FPToSI is never a no-op cast, no need to check
2474 SDValue N = getValue(I.getOperand(0));
2475 EVT DestVT = TLI.getValueType(I.getType());
2476 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurDebugLoc(), DestVT, N));
2479 void SelectionDAGBuilder::visitUIToFP(const User &I) {
2480 // UIToFP is never a no-op cast, no need to check
2481 SDValue N = getValue(I.getOperand(0));
2482 EVT DestVT = TLI.getValueType(I.getType());
2483 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurDebugLoc(), DestVT, N));
2486 void SelectionDAGBuilder::visitSIToFP(const User &I){
2487 // SIToFP is never a no-op cast, no need to check
2488 SDValue N = getValue(I.getOperand(0));
2489 EVT DestVT = TLI.getValueType(I.getType());
2490 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurDebugLoc(), DestVT, N));
2493 void SelectionDAGBuilder::visitPtrToInt(const User &I) {
2494 // What to do depends on the size of the integer and the size of the pointer.
2495 // We can either truncate, zero extend, or no-op, accordingly.
2496 SDValue N = getValue(I.getOperand(0));
2497 EVT DestVT = TLI.getValueType(I.getType());
2498 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT));
2501 void SelectionDAGBuilder::visitIntToPtr(const User &I) {
2502 // What to do depends on the size of the integer and the size of the pointer.
2503 // We can either truncate, zero extend, or no-op, accordingly.
2504 SDValue N = getValue(I.getOperand(0));
2505 EVT DestVT = TLI.getValueType(I.getType());
2506 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT));
2509 void SelectionDAGBuilder::visitBitCast(const User &I) {
2510 SDValue N = getValue(I.getOperand(0));
2511 EVT DestVT = TLI.getValueType(I.getType());
2513 // BitCast assures us that source and destination are the same size so this is
2514 // either a BIT_CONVERT or a no-op.
2515 if (DestVT != N.getValueType())
2516 setValue(&I, DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
2517 DestVT, N)); // convert types.
2519 setValue(&I, N); // noop cast.
2522 void SelectionDAGBuilder::visitInsertElement(const User &I) {
2523 SDValue InVec = getValue(I.getOperand(0));
2524 SDValue InVal = getValue(I.getOperand(1));
2525 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2527 getValue(I.getOperand(2)));
2528 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurDebugLoc(),
2529 TLI.getValueType(I.getType()),
2530 InVec, InVal, InIdx));
2533 void SelectionDAGBuilder::visitExtractElement(const User &I) {
2534 SDValue InVec = getValue(I.getOperand(0));
2535 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2537 getValue(I.getOperand(1)));
2538 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2539 TLI.getValueType(I.getType()), InVec, InIdx));
2542 // Utility for visitShuffleVector - Returns true if the mask is mask starting
2543 // from SIndx and increasing to the element length (undefs are allowed).
2544 static bool SequentialMask(SmallVectorImpl<int> &Mask, unsigned SIndx) {
2545 unsigned MaskNumElts = Mask.size();
2546 for (unsigned i = 0; i != MaskNumElts; ++i)
2547 if ((Mask[i] >= 0) && (Mask[i] != (int)(i + SIndx)))
2552 void SelectionDAGBuilder::visitShuffleVector(const User &I) {
2553 SmallVector<int, 8> Mask;
2554 SDValue Src1 = getValue(I.getOperand(0));
2555 SDValue Src2 = getValue(I.getOperand(1));
2557 // Convert the ConstantVector mask operand into an array of ints, with -1
2558 // representing undef values.
2559 SmallVector<Constant*, 8> MaskElts;
2560 cast<Constant>(I.getOperand(2))->getVectorElements(MaskElts);
2561 unsigned MaskNumElts = MaskElts.size();
2562 for (unsigned i = 0; i != MaskNumElts; ++i) {
2563 if (isa<UndefValue>(MaskElts[i]))
2566 Mask.push_back(cast<ConstantInt>(MaskElts[i])->getSExtValue());
2569 EVT VT = TLI.getValueType(I.getType());
2570 EVT SrcVT = Src1.getValueType();
2571 unsigned SrcNumElts = SrcVT.getVectorNumElements();
2573 if (SrcNumElts == MaskNumElts) {
2574 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2579 // Normalize the shuffle vector since mask and vector length don't match.
2580 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) {
2581 // Mask is longer than the source vectors and is a multiple of the source
2582 // vectors. We can use concatenate vector to make the mask and vectors
2584 if (SrcNumElts*2 == MaskNumElts && SequentialMask(Mask, 0)) {
2585 // The shuffle is concatenating two vectors together.
2586 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurDebugLoc(),
2591 // Pad both vectors with undefs to make them the same length as the mask.
2592 unsigned NumConcat = MaskNumElts / SrcNumElts;
2593 bool Src1U = Src1.getOpcode() == ISD::UNDEF;
2594 bool Src2U = Src2.getOpcode() == ISD::UNDEF;
2595 SDValue UndefVal = DAG.getUNDEF(SrcVT);
2597 SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal);
2598 SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal);
2602 Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2603 getCurDebugLoc(), VT,
2604 &MOps1[0], NumConcat);
2605 Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2606 getCurDebugLoc(), VT,
2607 &MOps2[0], NumConcat);
2609 // Readjust mask for new input vector length.
2610 SmallVector<int, 8> MappedOps;
2611 for (unsigned i = 0; i != MaskNumElts; ++i) {
2613 if (Idx < (int)SrcNumElts)
2614 MappedOps.push_back(Idx);
2616 MappedOps.push_back(Idx + MaskNumElts - SrcNumElts);
2619 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2624 if (SrcNumElts > MaskNumElts) {
2625 // Analyze the access pattern of the vector to see if we can extract
2626 // two subvectors and do the shuffle. The analysis is done by calculating
2627 // the range of elements the mask access on both vectors.
2628 int MinRange[2] = { SrcNumElts+1, SrcNumElts+1};
2629 int MaxRange[2] = {-1, -1};
2631 for (unsigned i = 0; i != MaskNumElts; ++i) {
2637 if (Idx >= (int)SrcNumElts) {
2641 if (Idx > MaxRange[Input])
2642 MaxRange[Input] = Idx;
2643 if (Idx < MinRange[Input])
2644 MinRange[Input] = Idx;
2647 // Check if the access is smaller than the vector size and can we find
2648 // a reasonable extract index.
2649 int RangeUse[2] = { 2, 2 }; // 0 = Unused, 1 = Extract, 2 = Can not
2651 int StartIdx[2]; // StartIdx to extract from
2652 for (int Input=0; Input < 2; ++Input) {
2653 if (MinRange[Input] == (int)(SrcNumElts+1) && MaxRange[Input] == -1) {
2654 RangeUse[Input] = 0; // Unused
2655 StartIdx[Input] = 0;
2656 } else if (MaxRange[Input] - MinRange[Input] < (int)MaskNumElts) {
2657 // Fits within range but we should see if we can find a good
2658 // start index that is a multiple of the mask length.
2659 if (MaxRange[Input] < (int)MaskNumElts) {
2660 RangeUse[Input] = 1; // Extract from beginning of the vector
2661 StartIdx[Input] = 0;
2663 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts;
2664 if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts &&
2665 StartIdx[Input] + MaskNumElts < SrcNumElts)
2666 RangeUse[Input] = 1; // Extract from a multiple of the mask length.
2671 if (RangeUse[0] == 0 && RangeUse[1] == 0) {
2672 setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used.
2675 else if (RangeUse[0] < 2 && RangeUse[1] < 2) {
2676 // Extract appropriate subvector and generate a vector shuffle
2677 for (int Input=0; Input < 2; ++Input) {
2678 SDValue &Src = Input == 0 ? Src1 : Src2;
2679 if (RangeUse[Input] == 0)
2680 Src = DAG.getUNDEF(VT);
2682 Src = DAG.getNode(ISD::EXTRACT_SUBVECTOR, getCurDebugLoc(), VT,
2683 Src, DAG.getIntPtrConstant(StartIdx[Input]));
2686 // Calculate new mask.
2687 SmallVector<int, 8> MappedOps;
2688 for (unsigned i = 0; i != MaskNumElts; ++i) {
2691 MappedOps.push_back(Idx);
2692 else if (Idx < (int)SrcNumElts)
2693 MappedOps.push_back(Idx - StartIdx[0]);
2695 MappedOps.push_back(Idx - SrcNumElts - StartIdx[1] + MaskNumElts);
2698 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2704 // We can't use either concat vectors or extract subvectors so fall back to
2705 // replacing the shuffle with extract and build vector.
2706 // to insert and build vector.
2707 EVT EltVT = VT.getVectorElementType();
2708 EVT PtrVT = TLI.getPointerTy();
2709 SmallVector<SDValue,8> Ops;
2710 for (unsigned i = 0; i != MaskNumElts; ++i) {
2712 Ops.push_back(DAG.getUNDEF(EltVT));
2717 if (Idx < (int)SrcNumElts)
2718 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2719 EltVT, Src1, DAG.getConstant(Idx, PtrVT));
2721 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2723 DAG.getConstant(Idx - SrcNumElts, PtrVT));
2729 setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
2730 VT, &Ops[0], Ops.size()));
2733 void SelectionDAGBuilder::visitInsertValue(const InsertValueInst &I) {
2734 const Value *Op0 = I.getOperand(0);
2735 const Value *Op1 = I.getOperand(1);
2736 const Type *AggTy = I.getType();
2737 const Type *ValTy = Op1->getType();
2738 bool IntoUndef = isa<UndefValue>(Op0);
2739 bool FromUndef = isa<UndefValue>(Op1);
2741 unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy,
2742 I.idx_begin(), I.idx_end());
2744 SmallVector<EVT, 4> AggValueVTs;
2745 ComputeValueVTs(TLI, AggTy, AggValueVTs);
2746 SmallVector<EVT, 4> ValValueVTs;
2747 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2749 unsigned NumAggValues = AggValueVTs.size();
2750 unsigned NumValValues = ValValueVTs.size();
2751 SmallVector<SDValue, 4> Values(NumAggValues);
2753 SDValue Agg = getValue(Op0);
2754 SDValue Val = getValue(Op1);
2756 // Copy the beginning value(s) from the original aggregate.
2757 for (; i != LinearIndex; ++i)
2758 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2759 SDValue(Agg.getNode(), Agg.getResNo() + i);
2760 // Copy values from the inserted value(s).
2761 for (; i != LinearIndex + NumValValues; ++i)
2762 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2763 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex);
2764 // Copy remaining value(s) from the original aggregate.
2765 for (; i != NumAggValues; ++i)
2766 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2767 SDValue(Agg.getNode(), Agg.getResNo() + i);
2769 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2770 DAG.getVTList(&AggValueVTs[0], NumAggValues),
2771 &Values[0], NumAggValues));
2774 void SelectionDAGBuilder::visitExtractValue(const ExtractValueInst &I) {
2775 const Value *Op0 = I.getOperand(0);
2776 const Type *AggTy = Op0->getType();
2777 const Type *ValTy = I.getType();
2778 bool OutOfUndef = isa<UndefValue>(Op0);
2780 unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy,
2781 I.idx_begin(), I.idx_end());
2783 SmallVector<EVT, 4> ValValueVTs;
2784 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2786 unsigned NumValValues = ValValueVTs.size();
2787 SmallVector<SDValue, 4> Values(NumValValues);
2789 SDValue Agg = getValue(Op0);
2790 // Copy out the selected value(s).
2791 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i)
2792 Values[i - LinearIndex] =
2794 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) :
2795 SDValue(Agg.getNode(), Agg.getResNo() + i);
2797 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2798 DAG.getVTList(&ValValueVTs[0], NumValValues),
2799 &Values[0], NumValValues));
2802 void SelectionDAGBuilder::visitGetElementPtr(const User &I) {
2803 SDValue N = getValue(I.getOperand(0));
2804 const Type *Ty = I.getOperand(0)->getType();
2806 for (GetElementPtrInst::const_op_iterator OI = I.op_begin()+1, E = I.op_end();
2808 const Value *Idx = *OI;
2809 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
2810 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
2813 uint64_t Offset = TD->getStructLayout(StTy)->getElementOffset(Field);
2814 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
2815 DAG.getIntPtrConstant(Offset));
2818 Ty = StTy->getElementType(Field);
2819 } else if (const UnionType *UnTy = dyn_cast<UnionType>(Ty)) {
2820 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
2822 // Offset canonically 0 for unions, but type changes
2823 Ty = UnTy->getElementType(Field);
2825 Ty = cast<SequentialType>(Ty)->getElementType();
2827 // If this is a constant subscript, handle it quickly.
2828 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
2829 if (CI->isZero()) continue;
2831 TD->getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
2833 EVT PTy = TLI.getPointerTy();
2834 unsigned PtrBits = PTy.getSizeInBits();
2836 OffsVal = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2838 DAG.getConstant(Offs, MVT::i64));
2840 OffsVal = DAG.getIntPtrConstant(Offs);
2842 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
2847 // N = N + Idx * ElementSize;
2848 APInt ElementSize = APInt(TLI.getPointerTy().getSizeInBits(),
2849 TD->getTypeAllocSize(Ty));
2850 SDValue IdxN = getValue(Idx);
2852 // If the index is smaller or larger than intptr_t, truncate or extend
2854 IdxN = DAG.getSExtOrTrunc(IdxN, getCurDebugLoc(), N.getValueType());
2856 // If this is a multiply by a power of two, turn it into a shl
2857 // immediately. This is a very common case.
2858 if (ElementSize != 1) {
2859 if (ElementSize.isPowerOf2()) {
2860 unsigned Amt = ElementSize.logBase2();
2861 IdxN = DAG.getNode(ISD::SHL, getCurDebugLoc(),
2862 N.getValueType(), IdxN,
2863 DAG.getConstant(Amt, TLI.getPointerTy()));
2865 SDValue Scale = DAG.getConstant(ElementSize, TLI.getPointerTy());
2866 IdxN = DAG.getNode(ISD::MUL, getCurDebugLoc(),
2867 N.getValueType(), IdxN, Scale);
2871 N = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2872 N.getValueType(), N, IdxN);
2879 void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) {
2880 // If this is a fixed sized alloca in the entry block of the function,
2881 // allocate it statically on the stack.
2882 if (FuncInfo.StaticAllocaMap.count(&I))
2883 return; // getValue will auto-populate this.
2885 const Type *Ty = I.getAllocatedType();
2886 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
2888 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty),
2891 SDValue AllocSize = getValue(I.getArraySize());
2893 EVT IntPtr = TLI.getPointerTy();
2894 if (AllocSize.getValueType() != IntPtr)
2895 AllocSize = DAG.getZExtOrTrunc(AllocSize, getCurDebugLoc(), IntPtr);
2897 AllocSize = DAG.getNode(ISD::MUL, getCurDebugLoc(), IntPtr,
2899 DAG.getConstant(TySize, IntPtr));
2901 // Handle alignment. If the requested alignment is less than or equal to
2902 // the stack alignment, ignore it. If the size is greater than or equal to
2903 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
2904 unsigned StackAlign = TM.getFrameInfo()->getStackAlignment();
2905 if (Align <= StackAlign)
2908 // Round the size of the allocation up to the stack alignment size
2909 // by add SA-1 to the size.
2910 AllocSize = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2911 AllocSize.getValueType(), AllocSize,
2912 DAG.getIntPtrConstant(StackAlign-1));
2914 // Mask out the low bits for alignment purposes.
2915 AllocSize = DAG.getNode(ISD::AND, getCurDebugLoc(),
2916 AllocSize.getValueType(), AllocSize,
2917 DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1)));
2919 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) };
2920 SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other);
2921 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, getCurDebugLoc(),
2924 DAG.setRoot(DSA.getValue(1));
2926 // Inform the Frame Information that we have just allocated a variable-sized
2928 FuncInfo.MF->getFrameInfo()->CreateVariableSizedObject(Align ? Align : 1);
2931 void SelectionDAGBuilder::visitLoad(const LoadInst &I) {
2932 const Value *SV = I.getOperand(0);
2933 SDValue Ptr = getValue(SV);
2935 const Type *Ty = I.getType();
2937 bool isVolatile = I.isVolatile();
2938 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
2939 unsigned Alignment = I.getAlignment();
2941 SmallVector<EVT, 4> ValueVTs;
2942 SmallVector<uint64_t, 4> Offsets;
2943 ComputeValueVTs(TLI, Ty, ValueVTs, &Offsets);
2944 unsigned NumValues = ValueVTs.size();
2949 bool ConstantMemory = false;
2951 // Serialize volatile loads with other side effects.
2953 else if (AA->pointsToConstantMemory(SV)) {
2954 // Do not serialize (non-volatile) loads of constant memory with anything.
2955 Root = DAG.getEntryNode();
2956 ConstantMemory = true;
2958 // Do not serialize non-volatile loads against each other.
2959 Root = DAG.getRoot();
2962 SmallVector<SDValue, 4> Values(NumValues);
2963 SmallVector<SDValue, 4> Chains(NumValues);
2964 EVT PtrVT = Ptr.getValueType();
2965 for (unsigned i = 0; i != NumValues; ++i) {
2966 SDValue A = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2968 DAG.getConstant(Offsets[i], PtrVT));
2969 SDValue L = DAG.getLoad(ValueVTs[i], getCurDebugLoc(), Root,
2970 A, SV, Offsets[i], isVolatile,
2971 isNonTemporal, Alignment);
2974 Chains[i] = L.getValue(1);
2977 if (!ConstantMemory) {
2978 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
2979 MVT::Other, &Chains[0], NumValues);
2983 PendingLoads.push_back(Chain);
2986 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2987 DAG.getVTList(&ValueVTs[0], NumValues),
2988 &Values[0], NumValues));
2991 void SelectionDAGBuilder::visitStore(const StoreInst &I) {
2992 const Value *SrcV = I.getOperand(0);
2993 const Value *PtrV = I.getOperand(1);
2995 SmallVector<EVT, 4> ValueVTs;
2996 SmallVector<uint64_t, 4> Offsets;
2997 ComputeValueVTs(TLI, SrcV->getType(), ValueVTs, &Offsets);
2998 unsigned NumValues = ValueVTs.size();
3002 // Get the lowered operands. Note that we do this after
3003 // checking if NumResults is zero, because with zero results
3004 // the operands won't have values in the map.
3005 SDValue Src = getValue(SrcV);
3006 SDValue Ptr = getValue(PtrV);
3008 SDValue Root = getRoot();
3009 SmallVector<SDValue, 4> Chains(NumValues);
3010 EVT PtrVT = Ptr.getValueType();
3011 bool isVolatile = I.isVolatile();
3012 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
3013 unsigned Alignment = I.getAlignment();
3015 for (unsigned i = 0; i != NumValues; ++i) {
3016 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT, Ptr,
3017 DAG.getConstant(Offsets[i], PtrVT));
3018 Chains[i] = DAG.getStore(Root, getCurDebugLoc(),
3019 SDValue(Src.getNode(), Src.getResNo() + i),
3020 Add, PtrV, Offsets[i], isVolatile,
3021 isNonTemporal, Alignment);
3024 DAG.setRoot(DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3025 MVT::Other, &Chains[0], NumValues));
3028 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
3030 void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I,
3031 unsigned Intrinsic) {
3032 bool HasChain = !I.doesNotAccessMemory();
3033 bool OnlyLoad = HasChain && I.onlyReadsMemory();
3035 // Build the operand list.
3036 SmallVector<SDValue, 8> Ops;
3037 if (HasChain) { // If this intrinsic has side-effects, chainify it.
3039 // We don't need to serialize loads against other loads.
3040 Ops.push_back(DAG.getRoot());
3042 Ops.push_back(getRoot());
3046 // Info is set by getTgtMemInstrinsic
3047 TargetLowering::IntrinsicInfo Info;
3048 bool IsTgtIntrinsic = TLI.getTgtMemIntrinsic(Info, I, Intrinsic);
3050 // Add the intrinsic ID as an integer operand if it's not a target intrinsic.
3051 if (!IsTgtIntrinsic)
3052 Ops.push_back(DAG.getConstant(Intrinsic, TLI.getPointerTy()));
3054 // Add all operands of the call to the operand list.
3055 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) {
3056 SDValue Op = getValue(I.getArgOperand(i));
3057 assert(TLI.isTypeLegal(Op.getValueType()) &&
3058 "Intrinsic uses a non-legal type?");
3062 SmallVector<EVT, 4> ValueVTs;
3063 ComputeValueVTs(TLI, I.getType(), ValueVTs);
3065 for (unsigned Val = 0, E = ValueVTs.size(); Val != E; ++Val) {
3066 assert(TLI.isTypeLegal(ValueVTs[Val]) &&
3067 "Intrinsic uses a non-legal type?");
3072 ValueVTs.push_back(MVT::Other);
3074 SDVTList VTs = DAG.getVTList(ValueVTs.data(), ValueVTs.size());
3078 if (IsTgtIntrinsic) {
3079 // This is target intrinsic that touches memory
3080 Result = DAG.getMemIntrinsicNode(Info.opc, getCurDebugLoc(),
3081 VTs, &Ops[0], Ops.size(),
3082 Info.memVT, Info.ptrVal, Info.offset,
3083 Info.align, Info.vol,
3084 Info.readMem, Info.writeMem);
3085 } else if (!HasChain) {
3086 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurDebugLoc(),
3087 VTs, &Ops[0], Ops.size());
3088 } else if (!I.getType()->isVoidTy()) {
3089 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurDebugLoc(),
3090 VTs, &Ops[0], Ops.size());
3092 Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurDebugLoc(),
3093 VTs, &Ops[0], Ops.size());
3097 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1);
3099 PendingLoads.push_back(Chain);
3104 if (!I.getType()->isVoidTy()) {
3105 if (const VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
3106 EVT VT = TLI.getValueType(PTy);
3107 Result = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(), VT, Result);
3110 setValue(&I, Result);
3114 /// GetSignificand - Get the significand and build it into a floating-point
3115 /// number with exponent of 1:
3117 /// Op = (Op & 0x007fffff) | 0x3f800000;
3119 /// where Op is the hexidecimal representation of floating point value.
3121 GetSignificand(SelectionDAG &DAG, SDValue Op, DebugLoc dl) {
3122 SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3123 DAG.getConstant(0x007fffff, MVT::i32));
3124 SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1,
3125 DAG.getConstant(0x3f800000, MVT::i32));
3126 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t2);
3129 /// GetExponent - Get the exponent:
3131 /// (float)(int)(((Op & 0x7f800000) >> 23) - 127);
3133 /// where Op is the hexidecimal representation of floating point value.
3135 GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI,
3137 SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3138 DAG.getConstant(0x7f800000, MVT::i32));
3139 SDValue t1 = DAG.getNode(ISD::SRL, dl, MVT::i32, t0,
3140 DAG.getConstant(23, TLI.getPointerTy()));
3141 SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1,
3142 DAG.getConstant(127, MVT::i32));
3143 return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2);
3146 /// getF32Constant - Get 32-bit floating point constant.
3148 getF32Constant(SelectionDAG &DAG, unsigned Flt) {
3149 return DAG.getConstantFP(APFloat(APInt(32, Flt)), MVT::f32);
3152 /// Inlined utility function to implement binary input atomic intrinsics for
3153 /// visitIntrinsicCall: I is a call instruction
3154 /// Op is the associated NodeType for I
3156 SelectionDAGBuilder::implVisitBinaryAtomic(const CallInst& I,
3158 SDValue Root = getRoot();
3160 DAG.getAtomic(Op, getCurDebugLoc(),
3161 getValue(I.getArgOperand(1)).getValueType().getSimpleVT(),
3163 getValue(I.getArgOperand(0)),
3164 getValue(I.getArgOperand(1)),
3165 I.getArgOperand(0));
3167 DAG.setRoot(L.getValue(1));
3171 // implVisitAluOverflow - Lower arithmetic overflow instrinsics.
3173 SelectionDAGBuilder::implVisitAluOverflow(const CallInst &I, ISD::NodeType Op) {
3174 SDValue Op1 = getValue(I.getArgOperand(0));
3175 SDValue Op2 = getValue(I.getArgOperand(1));
3177 SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1);
3178 setValue(&I, DAG.getNode(Op, getCurDebugLoc(), VTs, Op1, Op2));
3182 /// visitExp - Lower an exp intrinsic. Handles the special sequences for
3183 /// limited-precision mode.
3185 SelectionDAGBuilder::visitExp(const CallInst &I) {
3187 DebugLoc dl = getCurDebugLoc();
3189 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3190 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3191 SDValue Op = getValue(I.getArgOperand(0));
3193 // Put the exponent in the right bit position for later addition to the
3196 // #define LOG2OFe 1.4426950f
3197 // IntegerPartOfX = ((int32_t)(X * LOG2OFe));
3198 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3199 getF32Constant(DAG, 0x3fb8aa3b));
3200 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3202 // FractionalPartOfX = (X * LOG2OFe) - (float)IntegerPartOfX;
3203 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3204 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3206 // IntegerPartOfX <<= 23;
3207 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3208 DAG.getConstant(23, TLI.getPointerTy()));
3210 if (LimitFloatPrecision <= 6) {
3211 // For floating-point precision of 6:
3213 // TwoToFractionalPartOfX =
3215 // (0.735607626f + 0.252464424f * x) * x;
3217 // error 0.0144103317, which is 6 bits
3218 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3219 getF32Constant(DAG, 0x3e814304));
3220 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3221 getF32Constant(DAG, 0x3f3c50c8));
3222 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3223 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3224 getF32Constant(DAG, 0x3f7f5e7e));
3225 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,MVT::i32, t5);
3227 // Add the exponent into the result in integer domain.
3228 SDValue t6 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3229 TwoToFracPartOfX, IntegerPartOfX);
3231 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t6);
3232 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3233 // For floating-point precision of 12:
3235 // TwoToFractionalPartOfX =
3238 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3240 // 0.000107046256 error, which is 13 to 14 bits
3241 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3242 getF32Constant(DAG, 0x3da235e3));
3243 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3244 getF32Constant(DAG, 0x3e65b8f3));
3245 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3246 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3247 getF32Constant(DAG, 0x3f324b07));
3248 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3249 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3250 getF32Constant(DAG, 0x3f7ff8fd));
3251 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,MVT::i32, t7);
3253 // Add the exponent into the result in integer domain.
3254 SDValue t8 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3255 TwoToFracPartOfX, IntegerPartOfX);
3257 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t8);
3258 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3259 // For floating-point precision of 18:
3261 // TwoToFractionalPartOfX =
3265 // (0.554906021e-1f +
3266 // (0.961591928e-2f +
3267 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3269 // error 2.47208000*10^(-7), which is better than 18 bits
3270 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3271 getF32Constant(DAG, 0x3924b03e));
3272 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3273 getF32Constant(DAG, 0x3ab24b87));
3274 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3275 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3276 getF32Constant(DAG, 0x3c1d8c17));
3277 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3278 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3279 getF32Constant(DAG, 0x3d634a1d));
3280 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3281 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3282 getF32Constant(DAG, 0x3e75fe14));
3283 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3284 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3285 getF32Constant(DAG, 0x3f317234));
3286 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3287 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3288 getF32Constant(DAG, 0x3f800000));
3289 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,
3292 // Add the exponent into the result in integer domain.
3293 SDValue t14 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3294 TwoToFracPartOfX, IntegerPartOfX);
3296 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t14);
3299 // No special expansion.
3300 result = DAG.getNode(ISD::FEXP, dl,
3301 getValue(I.getArgOperand(0)).getValueType(),
3302 getValue(I.getArgOperand(0)));
3305 setValue(&I, result);
3308 /// visitLog - Lower a log intrinsic. Handles the special sequences for
3309 /// limited-precision mode.
3311 SelectionDAGBuilder::visitLog(const CallInst &I) {
3313 DebugLoc dl = getCurDebugLoc();
3315 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3316 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3317 SDValue Op = getValue(I.getArgOperand(0));
3318 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
3320 // Scale the exponent by log(2) [0.69314718f].
3321 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3322 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3323 getF32Constant(DAG, 0x3f317218));
3325 // Get the significand and build it into a floating-point number with
3327 SDValue X = GetSignificand(DAG, Op1, dl);
3329 if (LimitFloatPrecision <= 6) {
3330 // For floating-point precision of 6:
3334 // (1.4034025f - 0.23903021f * x) * x;
3336 // error 0.0034276066, which is better than 8 bits
3337 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3338 getF32Constant(DAG, 0xbe74c456));
3339 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3340 getF32Constant(DAG, 0x3fb3a2b1));
3341 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3342 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3343 getF32Constant(DAG, 0x3f949a29));
3345 result = DAG.getNode(ISD::FADD, dl,
3346 MVT::f32, LogOfExponent, LogOfMantissa);
3347 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3348 // For floating-point precision of 12:
3354 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x;
3356 // error 0.000061011436, which is 14 bits
3357 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3358 getF32Constant(DAG, 0xbd67b6d6));
3359 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3360 getF32Constant(DAG, 0x3ee4f4b8));
3361 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3362 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3363 getF32Constant(DAG, 0x3fbc278b));
3364 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3365 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3366 getF32Constant(DAG, 0x40348e95));
3367 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3368 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3369 getF32Constant(DAG, 0x3fdef31a));
3371 result = DAG.getNode(ISD::FADD, dl,
3372 MVT::f32, LogOfExponent, LogOfMantissa);
3373 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3374 // For floating-point precision of 18:
3382 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x;
3384 // error 0.0000023660568, which is better than 18 bits
3385 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3386 getF32Constant(DAG, 0xbc91e5ac));
3387 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3388 getF32Constant(DAG, 0x3e4350aa));
3389 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3390 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3391 getF32Constant(DAG, 0x3f60d3e3));
3392 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3393 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3394 getF32Constant(DAG, 0x4011cdf0));
3395 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3396 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3397 getF32Constant(DAG, 0x406cfd1c));
3398 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3399 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3400 getF32Constant(DAG, 0x408797cb));
3401 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3402 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3403 getF32Constant(DAG, 0x4006dcab));
3405 result = DAG.getNode(ISD::FADD, dl,
3406 MVT::f32, LogOfExponent, LogOfMantissa);
3409 // No special expansion.
3410 result = DAG.getNode(ISD::FLOG, dl,
3411 getValue(I.getArgOperand(0)).getValueType(),
3412 getValue(I.getArgOperand(0)));
3415 setValue(&I, result);
3418 /// visitLog2 - Lower a log2 intrinsic. Handles the special sequences for
3419 /// limited-precision mode.
3421 SelectionDAGBuilder::visitLog2(const CallInst &I) {
3423 DebugLoc dl = getCurDebugLoc();
3425 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3426 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3427 SDValue Op = getValue(I.getArgOperand(0));
3428 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
3430 // Get the exponent.
3431 SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl);
3433 // Get the significand and build it into a floating-point number with
3435 SDValue X = GetSignificand(DAG, Op1, dl);
3437 // Different possible minimax approximations of significand in
3438 // floating-point for various degrees of accuracy over [1,2].
3439 if (LimitFloatPrecision <= 6) {
3440 // For floating-point precision of 6:
3442 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x;
3444 // error 0.0049451742, which is more than 7 bits
3445 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3446 getF32Constant(DAG, 0xbeb08fe0));
3447 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3448 getF32Constant(DAG, 0x40019463));
3449 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3450 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3451 getF32Constant(DAG, 0x3fd6633d));
3453 result = DAG.getNode(ISD::FADD, dl,
3454 MVT::f32, LogOfExponent, Log2ofMantissa);
3455 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3456 // For floating-point precision of 12:
3462 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x;
3464 // error 0.0000876136000, which is better than 13 bits
3465 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3466 getF32Constant(DAG, 0xbda7262e));
3467 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3468 getF32Constant(DAG, 0x3f25280b));
3469 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3470 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3471 getF32Constant(DAG, 0x4007b923));
3472 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3473 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3474 getF32Constant(DAG, 0x40823e2f));
3475 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3476 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3477 getF32Constant(DAG, 0x4020d29c));
3479 result = DAG.getNode(ISD::FADD, dl,
3480 MVT::f32, LogOfExponent, Log2ofMantissa);
3481 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3482 // For floating-point precision of 18:
3491 // 0.25691327e-1f * x) * x) * x) * x) * x) * x;
3493 // error 0.0000018516, which is better than 18 bits
3494 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3495 getF32Constant(DAG, 0xbcd2769e));
3496 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3497 getF32Constant(DAG, 0x3e8ce0b9));
3498 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3499 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3500 getF32Constant(DAG, 0x3fa22ae7));
3501 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3502 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3503 getF32Constant(DAG, 0x40525723));
3504 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3505 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3506 getF32Constant(DAG, 0x40aaf200));
3507 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3508 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3509 getF32Constant(DAG, 0x40c39dad));
3510 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3511 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3512 getF32Constant(DAG, 0x4042902c));
3514 result = DAG.getNode(ISD::FADD, dl,
3515 MVT::f32, LogOfExponent, Log2ofMantissa);
3518 // No special expansion.
3519 result = DAG.getNode(ISD::FLOG2, dl,
3520 getValue(I.getArgOperand(0)).getValueType(),
3521 getValue(I.getArgOperand(0)));
3524 setValue(&I, result);
3527 /// visitLog10 - Lower a log10 intrinsic. Handles the special sequences for
3528 /// limited-precision mode.
3530 SelectionDAGBuilder::visitLog10(const CallInst &I) {
3532 DebugLoc dl = getCurDebugLoc();
3534 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3535 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3536 SDValue Op = getValue(I.getArgOperand(0));
3537 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
3539 // Scale the exponent by log10(2) [0.30102999f].
3540 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3541 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3542 getF32Constant(DAG, 0x3e9a209a));
3544 // Get the significand and build it into a floating-point number with
3546 SDValue X = GetSignificand(DAG, Op1, dl);
3548 if (LimitFloatPrecision <= 6) {
3549 // For floating-point precision of 6:
3551 // Log10ofMantissa =
3553 // (0.60948995f - 0.10380950f * x) * x;
3555 // error 0.0014886165, which is 6 bits
3556 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3557 getF32Constant(DAG, 0xbdd49a13));
3558 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3559 getF32Constant(DAG, 0x3f1c0789));
3560 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3561 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3562 getF32Constant(DAG, 0x3f011300));
3564 result = DAG.getNode(ISD::FADD, dl,
3565 MVT::f32, LogOfExponent, Log10ofMantissa);
3566 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3567 // For floating-point precision of 12:
3569 // Log10ofMantissa =
3572 // (-0.31664806f + 0.47637168e-1f * x) * x) * x;
3574 // error 0.00019228036, which is better than 12 bits
3575 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3576 getF32Constant(DAG, 0x3d431f31));
3577 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
3578 getF32Constant(DAG, 0x3ea21fb2));
3579 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3580 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3581 getF32Constant(DAG, 0x3f6ae232));
3582 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3583 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
3584 getF32Constant(DAG, 0x3f25f7c3));
3586 result = DAG.getNode(ISD::FADD, dl,
3587 MVT::f32, LogOfExponent, Log10ofMantissa);
3588 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3589 // For floating-point precision of 18:
3591 // Log10ofMantissa =
3596 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x;
3598 // error 0.0000037995730, which is better than 18 bits
3599 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3600 getF32Constant(DAG, 0x3c5d51ce));
3601 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
3602 getF32Constant(DAG, 0x3e00685a));
3603 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3604 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3605 getF32Constant(DAG, 0x3efb6798));
3606 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3607 SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
3608 getF32Constant(DAG, 0x3f88d192));
3609 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3610 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3611 getF32Constant(DAG, 0x3fc4316c));
3612 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3613 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8,
3614 getF32Constant(DAG, 0x3f57ce70));
3616 result = DAG.getNode(ISD::FADD, dl,
3617 MVT::f32, LogOfExponent, Log10ofMantissa);
3620 // No special expansion.
3621 result = DAG.getNode(ISD::FLOG10, dl,
3622 getValue(I.getArgOperand(0)).getValueType(),
3623 getValue(I.getArgOperand(0)));
3626 setValue(&I, result);
3629 /// visitExp2 - Lower an exp2 intrinsic. Handles the special sequences for
3630 /// limited-precision mode.
3632 SelectionDAGBuilder::visitExp2(const CallInst &I) {
3634 DebugLoc dl = getCurDebugLoc();
3636 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3637 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3638 SDValue Op = getValue(I.getArgOperand(0));
3640 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Op);
3642 // FractionalPartOfX = x - (float)IntegerPartOfX;
3643 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3644 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, Op, t1);
3646 // IntegerPartOfX <<= 23;
3647 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3648 DAG.getConstant(23, TLI.getPointerTy()));
3650 if (LimitFloatPrecision <= 6) {
3651 // For floating-point precision of 6:
3653 // TwoToFractionalPartOfX =
3655 // (0.735607626f + 0.252464424f * x) * x;
3657 // error 0.0144103317, which is 6 bits
3658 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3659 getF32Constant(DAG, 0x3e814304));
3660 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3661 getF32Constant(DAG, 0x3f3c50c8));
3662 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3663 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3664 getF32Constant(DAG, 0x3f7f5e7e));
3665 SDValue t6 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t5);
3666 SDValue TwoToFractionalPartOfX =
3667 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
3669 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3670 MVT::f32, TwoToFractionalPartOfX);
3671 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3672 // For floating-point precision of 12:
3674 // TwoToFractionalPartOfX =
3677 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3679 // error 0.000107046256, which is 13 to 14 bits
3680 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3681 getF32Constant(DAG, 0x3da235e3));
3682 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3683 getF32Constant(DAG, 0x3e65b8f3));
3684 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3685 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3686 getF32Constant(DAG, 0x3f324b07));
3687 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3688 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3689 getF32Constant(DAG, 0x3f7ff8fd));
3690 SDValue t8 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t7);
3691 SDValue TwoToFractionalPartOfX =
3692 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
3694 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3695 MVT::f32, TwoToFractionalPartOfX);
3696 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3697 // For floating-point precision of 18:
3699 // TwoToFractionalPartOfX =
3703 // (0.554906021e-1f +
3704 // (0.961591928e-2f +
3705 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3706 // error 2.47208000*10^(-7), which is better than 18 bits
3707 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3708 getF32Constant(DAG, 0x3924b03e));
3709 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3710 getF32Constant(DAG, 0x3ab24b87));
3711 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3712 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3713 getF32Constant(DAG, 0x3c1d8c17));
3714 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3715 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3716 getF32Constant(DAG, 0x3d634a1d));
3717 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3718 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3719 getF32Constant(DAG, 0x3e75fe14));
3720 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3721 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3722 getF32Constant(DAG, 0x3f317234));
3723 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3724 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3725 getF32Constant(DAG, 0x3f800000));
3726 SDValue t14 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t13);
3727 SDValue TwoToFractionalPartOfX =
3728 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
3730 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3731 MVT::f32, TwoToFractionalPartOfX);
3734 // No special expansion.
3735 result = DAG.getNode(ISD::FEXP2, dl,
3736 getValue(I.getArgOperand(0)).getValueType(),
3737 getValue(I.getArgOperand(0)));
3740 setValue(&I, result);
3743 /// visitPow - Lower a pow intrinsic. Handles the special sequences for
3744 /// limited-precision mode with x == 10.0f.
3746 SelectionDAGBuilder::visitPow(const CallInst &I) {
3748 const Value *Val = I.getArgOperand(0);
3749 DebugLoc dl = getCurDebugLoc();
3750 bool IsExp10 = false;
3752 if (getValue(Val).getValueType() == MVT::f32 &&
3753 getValue(I.getArgOperand(1)).getValueType() == MVT::f32 &&
3754 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3755 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(Val))) {
3756 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
3758 IsExp10 = CFP->getValueAPF().bitwiseIsEqual(Ten);
3763 if (IsExp10 && LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3764 SDValue Op = getValue(I.getArgOperand(1));
3766 // Put the exponent in the right bit position for later addition to the
3769 // #define LOG2OF10 3.3219281f
3770 // IntegerPartOfX = (int32_t)(x * LOG2OF10);
3771 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3772 getF32Constant(DAG, 0x40549a78));
3773 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3775 // FractionalPartOfX = x - (float)IntegerPartOfX;
3776 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3777 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3779 // IntegerPartOfX <<= 23;
3780 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3781 DAG.getConstant(23, TLI.getPointerTy()));
3783 if (LimitFloatPrecision <= 6) {
3784 // For floating-point precision of 6:
3786 // twoToFractionalPartOfX =
3788 // (0.735607626f + 0.252464424f * x) * x;
3790 // error 0.0144103317, which is 6 bits
3791 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3792 getF32Constant(DAG, 0x3e814304));
3793 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3794 getF32Constant(DAG, 0x3f3c50c8));
3795 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3796 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3797 getF32Constant(DAG, 0x3f7f5e7e));
3798 SDValue t6 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t5);
3799 SDValue TwoToFractionalPartOfX =
3800 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
3802 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3803 MVT::f32, TwoToFractionalPartOfX);
3804 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3805 // For floating-point precision of 12:
3807 // TwoToFractionalPartOfX =
3810 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3812 // error 0.000107046256, which is 13 to 14 bits
3813 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3814 getF32Constant(DAG, 0x3da235e3));
3815 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3816 getF32Constant(DAG, 0x3e65b8f3));
3817 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3818 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3819 getF32Constant(DAG, 0x3f324b07));
3820 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3821 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3822 getF32Constant(DAG, 0x3f7ff8fd));
3823 SDValue t8 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t7);
3824 SDValue TwoToFractionalPartOfX =
3825 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
3827 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3828 MVT::f32, TwoToFractionalPartOfX);
3829 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3830 // For floating-point precision of 18:
3832 // TwoToFractionalPartOfX =
3836 // (0.554906021e-1f +
3837 // (0.961591928e-2f +
3838 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3839 // error 2.47208000*10^(-7), which is better than 18 bits
3840 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3841 getF32Constant(DAG, 0x3924b03e));
3842 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3843 getF32Constant(DAG, 0x3ab24b87));
3844 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3845 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3846 getF32Constant(DAG, 0x3c1d8c17));
3847 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3848 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3849 getF32Constant(DAG, 0x3d634a1d));
3850 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3851 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3852 getF32Constant(DAG, 0x3e75fe14));
3853 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3854 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3855 getF32Constant(DAG, 0x3f317234));
3856 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3857 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3858 getF32Constant(DAG, 0x3f800000));
3859 SDValue t14 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t13);
3860 SDValue TwoToFractionalPartOfX =
3861 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
3863 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3864 MVT::f32, TwoToFractionalPartOfX);
3867 // No special expansion.
3868 result = DAG.getNode(ISD::FPOW, dl,
3869 getValue(I.getArgOperand(0)).getValueType(),
3870 getValue(I.getArgOperand(0)),
3871 getValue(I.getArgOperand(1)));
3874 setValue(&I, result);
3878 /// ExpandPowI - Expand a llvm.powi intrinsic.
3879 static SDValue ExpandPowI(DebugLoc DL, SDValue LHS, SDValue RHS,
3880 SelectionDAG &DAG) {
3881 // If RHS is a constant, we can expand this out to a multiplication tree,
3882 // otherwise we end up lowering to a call to __powidf2 (for example). When
3883 // optimizing for size, we only want to do this if the expansion would produce
3884 // a small number of multiplies, otherwise we do the full expansion.
3885 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3886 // Get the exponent as a positive value.
3887 unsigned Val = RHSC->getSExtValue();
3888 if ((int)Val < 0) Val = -Val;
3890 // powi(x, 0) -> 1.0
3892 return DAG.getConstantFP(1.0, LHS.getValueType());
3894 const Function *F = DAG.getMachineFunction().getFunction();
3895 if (!F->hasFnAttr(Attribute::OptimizeForSize) ||
3896 // If optimizing for size, don't insert too many multiplies. This
3897 // inserts up to 5 multiplies.
3898 CountPopulation_32(Val)+Log2_32(Val) < 7) {
3899 // We use the simple binary decomposition method to generate the multiply
3900 // sequence. There are more optimal ways to do this (for example,
3901 // powi(x,15) generates one more multiply than it should), but this has
3902 // the benefit of being both really simple and much better than a libcall.
3903 SDValue Res; // Logically starts equal to 1.0
3904 SDValue CurSquare = LHS;
3908 Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare);
3910 Res = CurSquare; // 1.0*CurSquare.
3913 CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(),
3914 CurSquare, CurSquare);
3918 // If the original was negative, invert the result, producing 1/(x*x*x).
3919 if (RHSC->getSExtValue() < 0)
3920 Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(),
3921 DAG.getConstantFP(1.0, LHS.getValueType()), Res);
3926 // Otherwise, expand to a libcall.
3927 return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS);
3930 /// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a function
3931 /// argument, create the corresponding DBG_VALUE machine instruction for it now.
3932 /// At the end of instruction selection, they will be inserted to the entry BB.
3934 SelectionDAGBuilder::EmitFuncArgumentDbgValue(const Value *V, MDNode *Variable,
3937 if (!isa<Argument>(V))
3940 MachineFunction &MF = DAG.getMachineFunction();
3941 // Ignore inlined function arguments here.
3942 DIVariable DV(Variable);
3943 if (DV.isInlinedFnArgument(MF.getFunction()))
3946 MachineBasicBlock *MBB = FuncInfo.MBB;
3947 if (MBB != &MF.front())
3951 if (N.getOpcode() == ISD::CopyFromReg) {
3952 Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg();
3953 if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) {
3954 MachineRegisterInfo &RegInfo = MF.getRegInfo();
3955 unsigned PR = RegInfo.getLiveInPhysReg(Reg);
3962 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
3963 if (VMI == FuncInfo.ValueMap.end())
3968 const TargetInstrInfo *TII = DAG.getTarget().getInstrInfo();
3969 MachineInstrBuilder MIB = BuildMI(MF, getCurDebugLoc(),
3970 TII->get(TargetOpcode::DBG_VALUE))
3971 .addReg(Reg, RegState::Debug).addImm(Offset).addMetadata(Variable);
3972 FuncInfo.ArgDbgValues.push_back(&*MIB);
3976 // VisualStudio defines setjmp as _setjmp
3977 #if defined(_MSC_VER) && defined(setjmp)
3978 #define setjmp_undefined_for_visual_studio
3982 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
3983 /// we want to emit this as a call to a named external function, return the name
3984 /// otherwise lower it and return null.
3986 SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) {
3987 DebugLoc dl = getCurDebugLoc();
3990 switch (Intrinsic) {
3992 // By default, turn this into a target intrinsic node.
3993 visitTargetIntrinsic(I, Intrinsic);
3995 case Intrinsic::vastart: visitVAStart(I); return 0;
3996 case Intrinsic::vaend: visitVAEnd(I); return 0;
3997 case Intrinsic::vacopy: visitVACopy(I); return 0;
3998 case Intrinsic::returnaddress:
3999 setValue(&I, DAG.getNode(ISD::RETURNADDR, dl, TLI.getPointerTy(),
4000 getValue(I.getArgOperand(0))));
4002 case Intrinsic::frameaddress:
4003 setValue(&I, DAG.getNode(ISD::FRAMEADDR, dl, TLI.getPointerTy(),
4004 getValue(I.getArgOperand(0))));
4006 case Intrinsic::setjmp:
4007 return "_setjmp"+!TLI.usesUnderscoreSetJmp();
4008 case Intrinsic::longjmp:
4009 return "_longjmp"+!TLI.usesUnderscoreLongJmp();
4010 case Intrinsic::memcpy: {
4011 // Assert for address < 256 since we support only user defined address
4013 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4015 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4017 "Unknown address space");
4018 SDValue Op1 = getValue(I.getArgOperand(0));
4019 SDValue Op2 = getValue(I.getArgOperand(1));
4020 SDValue Op3 = getValue(I.getArgOperand(2));
4021 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4022 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4023 DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, isVol, false,
4024 I.getArgOperand(0), 0, I.getArgOperand(1), 0));
4027 case Intrinsic::memset: {
4028 // Assert for address < 256 since we support only user defined address
4030 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4032 "Unknown address space");
4033 SDValue Op1 = getValue(I.getArgOperand(0));
4034 SDValue Op2 = getValue(I.getArgOperand(1));
4035 SDValue Op3 = getValue(I.getArgOperand(2));
4036 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4037 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4038 DAG.setRoot(DAG.getMemset(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4039 I.getArgOperand(0), 0));
4042 case Intrinsic::memmove: {
4043 // Assert for address < 256 since we support only user defined address
4045 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4047 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4049 "Unknown address space");
4050 SDValue Op1 = getValue(I.getArgOperand(0));
4051 SDValue Op2 = getValue(I.getArgOperand(1));
4052 SDValue Op3 = getValue(I.getArgOperand(2));
4053 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4054 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4056 // If the source and destination are known to not be aliases, we can
4057 // lower memmove as memcpy.
4058 uint64_t Size = -1ULL;
4059 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op3))
4060 Size = C->getZExtValue();
4061 if (AA->alias(I.getArgOperand(0), Size, I.getArgOperand(1), Size) ==
4062 AliasAnalysis::NoAlias) {
4063 DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4064 false, I.getArgOperand(0), 0,
4065 I.getArgOperand(1), 0));
4069 DAG.setRoot(DAG.getMemmove(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4070 I.getArgOperand(0), 0, I.getArgOperand(1), 0));
4073 case Intrinsic::dbg_declare: {
4074 const DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
4075 if (!DIVariable(DI.getVariable()).Verify())
4078 MDNode *Variable = DI.getVariable();
4079 // Parameters are handled specially.
4081 DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable;
4082 const Value *Address = DI.getAddress();
4085 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
4086 Address = BCI->getOperand(0);
4087 const AllocaInst *AI = dyn_cast<AllocaInst>(Address);
4089 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder
4090 // but do not always have a corresponding SDNode built. The SDNodeOrder
4091 // absolute, but not relative, values are different depending on whether
4092 // debug info exists.
4094 SDValue &N = NodeMap[Address];
4097 if (isParameter && !AI) {
4098 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode());
4100 // Byval parameter. We have a frame index at this point.
4101 SDV = DAG.getDbgValue(Variable, FINode->getIndex(),
4102 0, dl, SDNodeOrder);
4104 // Can't do anything with other non-AI cases yet. This might be a
4105 // parameter of a callee function that got inlined, for example.
4108 SDV = DAG.getDbgValue(Variable, N.getNode(), N.getResNo(),
4109 0, dl, SDNodeOrder);
4111 // Can't do anything with other non-AI cases yet.
4113 DAG.AddDbgValue(SDV, N.getNode(), isParameter);
4115 // This isn't useful, but it shows what we're missing.
4116 SDV = DAG.getDbgValue(Variable, UndefValue::get(Address->getType()),
4117 0, dl, SDNodeOrder);
4118 DAG.AddDbgValue(SDV, 0, isParameter);
4122 case Intrinsic::dbg_value: {
4123 const DbgValueInst &DI = cast<DbgValueInst>(I);
4124 if (!DIVariable(DI.getVariable()).Verify())
4127 MDNode *Variable = DI.getVariable();
4128 uint64_t Offset = DI.getOffset();
4129 const Value *V = DI.getValue();
4133 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder
4134 // but do not always have a corresponding SDNode built. The SDNodeOrder
4135 // absolute, but not relative, values are different depending on whether
4136 // debug info exists.
4139 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) {
4140 SDV = DAG.getDbgValue(Variable, V, Offset, dl, SDNodeOrder);
4141 DAG.AddDbgValue(SDV, 0, false);
4143 bool createUndef = false;
4144 // Do not use getValue() in here; we don't want to generate code at
4145 // this point if it hasn't been done yet.
4146 SDValue N = NodeMap[V];
4147 if (!N.getNode() && isa<Argument>(V))
4148 // Check unused arguments map.
4149 N = UnusedArgNodeMap[V];
4151 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, N)) {
4152 SDV = DAG.getDbgValue(Variable, N.getNode(),
4153 N.getResNo(), Offset, dl, SDNodeOrder);
4154 DAG.AddDbgValue(SDV, N.getNode(), false);
4156 } else if (isa<PHINode>(V) && !V->use_empty() ) {
4157 // Do not call getValue(V) yet, as we don't want to generate code.
4158 // Remember it for later.
4159 DanglingDebugInfo DDI(&DI, dl, SDNodeOrder);
4160 DanglingDebugInfoMap[V] = DDI;
4164 // We may expand this to cover more cases. One case where we have no
4165 // data available is an unreferenced parameter; we need this fallback.
4166 SDV = DAG.getDbgValue(Variable, UndefValue::get(V->getType()),
4167 Offset, dl, SDNodeOrder);
4168 DAG.AddDbgValue(SDV, 0, false);
4172 // Build a debug info table entry.
4173 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
4174 V = BCI->getOperand(0);
4175 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
4176 // Don't handle byval struct arguments or VLAs, for example.
4179 DenseMap<const AllocaInst*, int>::iterator SI =
4180 FuncInfo.StaticAllocaMap.find(AI);
4181 if (SI == FuncInfo.StaticAllocaMap.end())
4183 int FI = SI->second;
4185 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4186 if (!DI.getDebugLoc().isUnknown() && MMI.hasDebugInfo())
4187 MMI.setVariableDbgInfo(Variable, FI, DI.getDebugLoc());
4190 case Intrinsic::eh_exception: {
4191 // Insert the EXCEPTIONADDR instruction.
4192 assert(FuncInfo.MBB->isLandingPad() &&
4193 "Call to eh.exception not in landing pad!");
4194 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
4196 Ops[0] = DAG.getRoot();
4197 SDValue Op = DAG.getNode(ISD::EXCEPTIONADDR, dl, VTs, Ops, 1);
4199 DAG.setRoot(Op.getValue(1));
4203 case Intrinsic::eh_selector: {
4204 MachineBasicBlock *CallMBB = FuncInfo.MBB;
4205 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4206 if (CallMBB->isLandingPad())
4207 AddCatchInfo(I, &MMI, CallMBB);
4210 FuncInfo.CatchInfoLost.insert(&I);
4212 // FIXME: Mark exception selector register as live in. Hack for PR1508.
4213 unsigned Reg = TLI.getExceptionSelectorRegister();
4214 if (Reg) FuncInfo.MBB->addLiveIn(Reg);
4217 // Insert the EHSELECTION instruction.
4218 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
4220 Ops[0] = getValue(I.getArgOperand(0));
4222 SDValue Op = DAG.getNode(ISD::EHSELECTION, dl, VTs, Ops, 2);
4223 DAG.setRoot(Op.getValue(1));
4224 setValue(&I, DAG.getSExtOrTrunc(Op, dl, MVT::i32));
4228 case Intrinsic::eh_typeid_for: {
4229 // Find the type id for the given typeinfo.
4230 GlobalVariable *GV = ExtractTypeInfo(I.getArgOperand(0));
4231 unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(GV);
4232 Res = DAG.getConstant(TypeID, MVT::i32);
4237 case Intrinsic::eh_return_i32:
4238 case Intrinsic::eh_return_i64:
4239 DAG.getMachineFunction().getMMI().setCallsEHReturn(true);
4240 DAG.setRoot(DAG.getNode(ISD::EH_RETURN, dl,
4243 getValue(I.getArgOperand(0)),
4244 getValue(I.getArgOperand(1))));
4246 case Intrinsic::eh_unwind_init:
4247 DAG.getMachineFunction().getMMI().setCallsUnwindInit(true);
4249 case Intrinsic::eh_dwarf_cfa: {
4250 SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getArgOperand(0)), dl,
4251 TLI.getPointerTy());
4252 SDValue Offset = DAG.getNode(ISD::ADD, dl,
4254 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, dl,
4255 TLI.getPointerTy()),
4257 SDValue FA = DAG.getNode(ISD::FRAMEADDR, dl,
4259 DAG.getConstant(0, TLI.getPointerTy()));
4260 setValue(&I, DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
4264 case Intrinsic::eh_sjlj_callsite: {
4265 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4266 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(0));
4267 assert(CI && "Non-constant call site value in eh.sjlj.callsite!");
4268 assert(MMI.getCurrentCallSite() == 0 && "Overlapping call sites!");
4270 MMI.setCurrentCallSite(CI->getZExtValue());
4273 case Intrinsic::eh_sjlj_setjmp: {
4274 setValue(&I, DAG.getNode(ISD::EH_SJLJ_SETJMP, dl, MVT::i32, getRoot(),
4275 getValue(I.getArgOperand(0))));
4278 case Intrinsic::eh_sjlj_longjmp: {
4279 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, dl, MVT::Other,
4281 getValue(I.getArgOperand(0))));
4285 case Intrinsic::convertff:
4286 case Intrinsic::convertfsi:
4287 case Intrinsic::convertfui:
4288 case Intrinsic::convertsif:
4289 case Intrinsic::convertuif:
4290 case Intrinsic::convertss:
4291 case Intrinsic::convertsu:
4292 case Intrinsic::convertus:
4293 case Intrinsic::convertuu: {
4294 ISD::CvtCode Code = ISD::CVT_INVALID;
4295 switch (Intrinsic) {
4296 case Intrinsic::convertff: Code = ISD::CVT_FF; break;
4297 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break;
4298 case Intrinsic::convertfui: Code = ISD::CVT_FU; break;
4299 case Intrinsic::convertsif: Code = ISD::CVT_SF; break;
4300 case Intrinsic::convertuif: Code = ISD::CVT_UF; break;
4301 case Intrinsic::convertss: Code = ISD::CVT_SS; break;
4302 case Intrinsic::convertsu: Code = ISD::CVT_SU; break;
4303 case Intrinsic::convertus: Code = ISD::CVT_US; break;
4304 case Intrinsic::convertuu: Code = ISD::CVT_UU; break;
4306 EVT DestVT = TLI.getValueType(I.getType());
4307 const Value *Op1 = I.getArgOperand(0);
4308 Res = DAG.getConvertRndSat(DestVT, getCurDebugLoc(), getValue(Op1),
4309 DAG.getValueType(DestVT),
4310 DAG.getValueType(getValue(Op1).getValueType()),
4311 getValue(I.getArgOperand(1)),
4312 getValue(I.getArgOperand(2)),
4317 case Intrinsic::sqrt:
4318 setValue(&I, DAG.getNode(ISD::FSQRT, dl,
4319 getValue(I.getArgOperand(0)).getValueType(),
4320 getValue(I.getArgOperand(0))));
4322 case Intrinsic::powi:
4323 setValue(&I, ExpandPowI(dl, getValue(I.getArgOperand(0)),
4324 getValue(I.getArgOperand(1)), DAG));
4326 case Intrinsic::sin:
4327 setValue(&I, DAG.getNode(ISD::FSIN, dl,
4328 getValue(I.getArgOperand(0)).getValueType(),
4329 getValue(I.getArgOperand(0))));
4331 case Intrinsic::cos:
4332 setValue(&I, DAG.getNode(ISD::FCOS, dl,
4333 getValue(I.getArgOperand(0)).getValueType(),
4334 getValue(I.getArgOperand(0))));
4336 case Intrinsic::log:
4339 case Intrinsic::log2:
4342 case Intrinsic::log10:
4345 case Intrinsic::exp:
4348 case Intrinsic::exp2:
4351 case Intrinsic::pow:
4354 case Intrinsic::convert_to_fp16:
4355 setValue(&I, DAG.getNode(ISD::FP32_TO_FP16, dl,
4356 MVT::i16, getValue(I.getArgOperand(0))));
4358 case Intrinsic::convert_from_fp16:
4359 setValue(&I, DAG.getNode(ISD::FP16_TO_FP32, dl,
4360 MVT::f32, getValue(I.getArgOperand(0))));
4362 case Intrinsic::pcmarker: {
4363 SDValue Tmp = getValue(I.getArgOperand(0));
4364 DAG.setRoot(DAG.getNode(ISD::PCMARKER, dl, MVT::Other, getRoot(), Tmp));
4367 case Intrinsic::readcyclecounter: {
4368 SDValue Op = getRoot();
4369 Res = DAG.getNode(ISD::READCYCLECOUNTER, dl,
4370 DAG.getVTList(MVT::i64, MVT::Other),
4373 DAG.setRoot(Res.getValue(1));
4376 case Intrinsic::bswap:
4377 setValue(&I, DAG.getNode(ISD::BSWAP, dl,
4378 getValue(I.getArgOperand(0)).getValueType(),
4379 getValue(I.getArgOperand(0))));
4381 case Intrinsic::cttz: {
4382 SDValue Arg = getValue(I.getArgOperand(0));
4383 EVT Ty = Arg.getValueType();
4384 setValue(&I, DAG.getNode(ISD::CTTZ, dl, Ty, Arg));
4387 case Intrinsic::ctlz: {
4388 SDValue Arg = getValue(I.getArgOperand(0));
4389 EVT Ty = Arg.getValueType();
4390 setValue(&I, DAG.getNode(ISD::CTLZ, dl, Ty, Arg));
4393 case Intrinsic::ctpop: {
4394 SDValue Arg = getValue(I.getArgOperand(0));
4395 EVT Ty = Arg.getValueType();
4396 setValue(&I, DAG.getNode(ISD::CTPOP, dl, Ty, Arg));
4399 case Intrinsic::stacksave: {
4400 SDValue Op = getRoot();
4401 Res = DAG.getNode(ISD::STACKSAVE, dl,
4402 DAG.getVTList(TLI.getPointerTy(), MVT::Other), &Op, 1);
4404 DAG.setRoot(Res.getValue(1));
4407 case Intrinsic::stackrestore: {
4408 Res = getValue(I.getArgOperand(0));
4409 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, dl, MVT::Other, getRoot(), Res));
4412 case Intrinsic::stackprotector: {
4413 // Emit code into the DAG to store the stack guard onto the stack.
4414 MachineFunction &MF = DAG.getMachineFunction();
4415 MachineFrameInfo *MFI = MF.getFrameInfo();
4416 EVT PtrTy = TLI.getPointerTy();
4418 SDValue Src = getValue(I.getArgOperand(0)); // The guard's value.
4419 AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
4421 int FI = FuncInfo.StaticAllocaMap[Slot];
4422 MFI->setStackProtectorIndex(FI);
4424 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
4426 // Store the stack protector onto the stack.
4427 Res = DAG.getStore(getRoot(), getCurDebugLoc(), Src, FIN,
4428 PseudoSourceValue::getFixedStack(FI),
4434 case Intrinsic::objectsize: {
4435 // If we don't know by now, we're never going to know.
4436 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1));
4438 assert(CI && "Non-constant type in __builtin_object_size?");
4440 SDValue Arg = getValue(I.getCalledValue());
4441 EVT Ty = Arg.getValueType();
4444 Res = DAG.getConstant(-1ULL, Ty);
4446 Res = DAG.getConstant(0, Ty);
4451 case Intrinsic::var_annotation:
4452 // Discard annotate attributes
4455 case Intrinsic::init_trampoline: {
4456 const Function *F = cast<Function>(I.getArgOperand(1)->stripPointerCasts());
4460 Ops[1] = getValue(I.getArgOperand(0));
4461 Ops[2] = getValue(I.getArgOperand(1));
4462 Ops[3] = getValue(I.getArgOperand(2));
4463 Ops[4] = DAG.getSrcValue(I.getArgOperand(0));
4464 Ops[5] = DAG.getSrcValue(F);
4466 Res = DAG.getNode(ISD::TRAMPOLINE, dl,
4467 DAG.getVTList(TLI.getPointerTy(), MVT::Other),
4471 DAG.setRoot(Res.getValue(1));
4474 case Intrinsic::gcroot:
4476 const Value *Alloca = I.getArgOperand(0);
4477 const Constant *TypeMap = cast<Constant>(I.getArgOperand(1));
4479 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode());
4480 GFI->addStackRoot(FI->getIndex(), TypeMap);
4483 case Intrinsic::gcread:
4484 case Intrinsic::gcwrite:
4485 llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!");
4487 case Intrinsic::flt_rounds:
4488 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, dl, MVT::i32));
4490 case Intrinsic::trap:
4491 DAG.setRoot(DAG.getNode(ISD::TRAP, dl,MVT::Other, getRoot()));
4493 case Intrinsic::uadd_with_overflow:
4494 return implVisitAluOverflow(I, ISD::UADDO);
4495 case Intrinsic::sadd_with_overflow:
4496 return implVisitAluOverflow(I, ISD::SADDO);
4497 case Intrinsic::usub_with_overflow:
4498 return implVisitAluOverflow(I, ISD::USUBO);
4499 case Intrinsic::ssub_with_overflow:
4500 return implVisitAluOverflow(I, ISD::SSUBO);
4501 case Intrinsic::umul_with_overflow:
4502 return implVisitAluOverflow(I, ISD::UMULO);
4503 case Intrinsic::smul_with_overflow:
4504 return implVisitAluOverflow(I, ISD::SMULO);
4506 case Intrinsic::prefetch: {
4509 Ops[1] = getValue(I.getArgOperand(0));
4510 Ops[2] = getValue(I.getArgOperand(1));
4511 Ops[3] = getValue(I.getArgOperand(2));
4512 DAG.setRoot(DAG.getNode(ISD::PREFETCH, dl, MVT::Other, &Ops[0], 4));
4516 case Intrinsic::memory_barrier: {
4519 for (int x = 1; x < 6; ++x)
4520 Ops[x] = getValue(I.getArgOperand(x - 1));
4522 DAG.setRoot(DAG.getNode(ISD::MEMBARRIER, dl, MVT::Other, &Ops[0], 6));
4525 case Intrinsic::atomic_cmp_swap: {
4526 SDValue Root = getRoot();
4528 DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, getCurDebugLoc(),
4529 getValue(I.getArgOperand(1)).getValueType().getSimpleVT(),
4531 getValue(I.getArgOperand(0)),
4532 getValue(I.getArgOperand(1)),
4533 getValue(I.getArgOperand(2)),
4534 I.getArgOperand(0));
4536 DAG.setRoot(L.getValue(1));
4539 case Intrinsic::atomic_load_add:
4540 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_ADD);
4541 case Intrinsic::atomic_load_sub:
4542 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_SUB);
4543 case Intrinsic::atomic_load_or:
4544 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_OR);
4545 case Intrinsic::atomic_load_xor:
4546 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_XOR);
4547 case Intrinsic::atomic_load_and:
4548 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_AND);
4549 case Intrinsic::atomic_load_nand:
4550 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_NAND);
4551 case Intrinsic::atomic_load_max:
4552 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MAX);
4553 case Intrinsic::atomic_load_min:
4554 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MIN);
4555 case Intrinsic::atomic_load_umin:
4556 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMIN);
4557 case Intrinsic::atomic_load_umax:
4558 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMAX);
4559 case Intrinsic::atomic_swap:
4560 return implVisitBinaryAtomic(I, ISD::ATOMIC_SWAP);
4562 case Intrinsic::invariant_start:
4563 case Intrinsic::lifetime_start:
4564 // Discard region information.
4565 setValue(&I, DAG.getUNDEF(TLI.getPointerTy()));
4567 case Intrinsic::invariant_end:
4568 case Intrinsic::lifetime_end:
4569 // Discard region information.
4574 void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee,
4576 MachineBasicBlock *LandingPad) {
4577 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
4578 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
4579 const Type *RetTy = FTy->getReturnType();
4580 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4581 MCSymbol *BeginLabel = 0;
4583 TargetLowering::ArgListTy Args;
4584 TargetLowering::ArgListEntry Entry;
4585 Args.reserve(CS.arg_size());
4587 // Check whether the function can return without sret-demotion.
4588 SmallVector<ISD::OutputArg, 4> Outs;
4589 SmallVector<uint64_t, 4> Offsets;
4590 GetReturnInfo(RetTy, CS.getAttributes().getRetAttributes(),
4591 Outs, TLI, &Offsets);
4593 bool CanLowerReturn = TLI.CanLowerReturn(CS.getCallingConv(),
4594 FTy->isVarArg(), Outs, FTy->getContext());
4596 SDValue DemoteStackSlot;
4598 if (!CanLowerReturn) {
4599 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(
4600 FTy->getReturnType());
4601 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(
4602 FTy->getReturnType());
4603 MachineFunction &MF = DAG.getMachineFunction();
4604 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
4605 const Type *StackSlotPtrType = PointerType::getUnqual(FTy->getReturnType());
4607 DemoteStackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
4608 Entry.Node = DemoteStackSlot;
4609 Entry.Ty = StackSlotPtrType;
4610 Entry.isSExt = false;
4611 Entry.isZExt = false;
4612 Entry.isInReg = false;
4613 Entry.isSRet = true;
4614 Entry.isNest = false;
4615 Entry.isByVal = false;
4616 Entry.Alignment = Align;
4617 Args.push_back(Entry);
4618 RetTy = Type::getVoidTy(FTy->getContext());
4621 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
4623 SDValue ArgNode = getValue(*i);
4624 Entry.Node = ArgNode; Entry.Ty = (*i)->getType();
4626 unsigned attrInd = i - CS.arg_begin() + 1;
4627 Entry.isSExt = CS.paramHasAttr(attrInd, Attribute::SExt);
4628 Entry.isZExt = CS.paramHasAttr(attrInd, Attribute::ZExt);
4629 Entry.isInReg = CS.paramHasAttr(attrInd, Attribute::InReg);
4630 Entry.isSRet = CS.paramHasAttr(attrInd, Attribute::StructRet);
4631 Entry.isNest = CS.paramHasAttr(attrInd, Attribute::Nest);
4632 Entry.isByVal = CS.paramHasAttr(attrInd, Attribute::ByVal);
4633 Entry.Alignment = CS.getParamAlignment(attrInd);
4634 Args.push_back(Entry);
4638 // Insert a label before the invoke call to mark the try range. This can be
4639 // used to detect deletion of the invoke via the MachineModuleInfo.
4640 BeginLabel = MMI.getContext().CreateTempSymbol();
4642 // For SjLj, keep track of which landing pads go with which invokes
4643 // so as to maintain the ordering of pads in the LSDA.
4644 unsigned CallSiteIndex = MMI.getCurrentCallSite();
4645 if (CallSiteIndex) {
4646 MMI.setCallSiteBeginLabel(BeginLabel, CallSiteIndex);
4647 // Now that the call site is handled, stop tracking it.
4648 MMI.setCurrentCallSite(0);
4651 // Both PendingLoads and PendingExports must be flushed here;
4652 // this call might not return.
4654 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getControlRoot(), BeginLabel));
4657 // Check if target-independent constraints permit a tail call here.
4658 // Target-dependent constraints are checked within TLI.LowerCallTo.
4660 !isInTailCallPosition(CS, CS.getAttributes().getRetAttributes(), TLI))
4663 std::pair<SDValue,SDValue> Result =
4664 TLI.LowerCallTo(getRoot(), RetTy,
4665 CS.paramHasAttr(0, Attribute::SExt),
4666 CS.paramHasAttr(0, Attribute::ZExt), FTy->isVarArg(),
4667 CS.paramHasAttr(0, Attribute::InReg), FTy->getNumParams(),
4668 CS.getCallingConv(),
4670 !CS.getInstruction()->use_empty(),
4671 Callee, Args, DAG, getCurDebugLoc());
4672 assert((isTailCall || Result.second.getNode()) &&
4673 "Non-null chain expected with non-tail call!");
4674 assert((Result.second.getNode() || !Result.first.getNode()) &&
4675 "Null value expected with tail call!");
4676 if (Result.first.getNode()) {
4677 setValue(CS.getInstruction(), Result.first);
4678 } else if (!CanLowerReturn && Result.second.getNode()) {
4679 // The instruction result is the result of loading from the
4680 // hidden sret parameter.
4681 SmallVector<EVT, 1> PVTs;
4682 const Type *PtrRetTy = PointerType::getUnqual(FTy->getReturnType());
4684 ComputeValueVTs(TLI, PtrRetTy, PVTs);
4685 assert(PVTs.size() == 1 && "Pointers should fit in one register");
4686 EVT PtrVT = PVTs[0];
4687 unsigned NumValues = Outs.size();
4688 SmallVector<SDValue, 4> Values(NumValues);
4689 SmallVector<SDValue, 4> Chains(NumValues);
4691 for (unsigned i = 0; i < NumValues; ++i) {
4692 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT,
4694 DAG.getConstant(Offsets[i], PtrVT));
4695 SDValue L = DAG.getLoad(Outs[i].VT, getCurDebugLoc(), Result.second,
4696 Add, NULL, Offsets[i], false, false, 1);
4698 Chains[i] = L.getValue(1);
4701 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
4702 MVT::Other, &Chains[0], NumValues);
4703 PendingLoads.push_back(Chain);
4705 // Collect the legal value parts into potentially illegal values
4706 // that correspond to the original function's return values.
4707 SmallVector<EVT, 4> RetTys;
4708 RetTy = FTy->getReturnType();
4709 ComputeValueVTs(TLI, RetTy, RetTys);
4710 ISD::NodeType AssertOp = ISD::DELETED_NODE;
4711 SmallVector<SDValue, 4> ReturnValues;
4712 unsigned CurReg = 0;
4713 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
4715 EVT RegisterVT = TLI.getRegisterType(RetTy->getContext(), VT);
4716 unsigned NumRegs = TLI.getNumRegisters(RetTy->getContext(), VT);
4718 SDValue ReturnValue =
4719 getCopyFromParts(DAG, getCurDebugLoc(), &Values[CurReg], NumRegs,
4720 RegisterVT, VT, AssertOp);
4721 ReturnValues.push_back(ReturnValue);
4725 setValue(CS.getInstruction(),
4726 DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
4727 DAG.getVTList(&RetTys[0], RetTys.size()),
4728 &ReturnValues[0], ReturnValues.size()));
4732 // As a special case, a null chain means that a tail call has been emitted and
4733 // the DAG root is already updated.
4734 if (Result.second.getNode())
4735 DAG.setRoot(Result.second);
4740 // Insert a label at the end of the invoke call to mark the try range. This
4741 // can be used to detect deletion of the invoke via the MachineModuleInfo.
4742 MCSymbol *EndLabel = MMI.getContext().CreateTempSymbol();
4743 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getRoot(), EndLabel));
4745 // Inform MachineModuleInfo of range.
4746 MMI.addInvoke(LandingPad, BeginLabel, EndLabel);
4750 /// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the
4751 /// value is equal or not-equal to zero.
4752 static bool IsOnlyUsedInZeroEqualityComparison(const Value *V) {
4753 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end();
4755 if (const ICmpInst *IC = dyn_cast<ICmpInst>(*UI))
4756 if (IC->isEquality())
4757 if (const Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
4758 if (C->isNullValue())
4760 // Unknown instruction.
4766 static SDValue getMemCmpLoad(const Value *PtrVal, MVT LoadVT,
4768 SelectionDAGBuilder &Builder) {
4770 // Check to see if this load can be trivially constant folded, e.g. if the
4771 // input is from a string literal.
4772 if (const Constant *LoadInput = dyn_cast<Constant>(PtrVal)) {
4773 // Cast pointer to the type we really want to load.
4774 LoadInput = ConstantExpr::getBitCast(const_cast<Constant *>(LoadInput),
4775 PointerType::getUnqual(LoadTy));
4777 if (const Constant *LoadCst =
4778 ConstantFoldLoadFromConstPtr(const_cast<Constant *>(LoadInput),
4780 return Builder.getValue(LoadCst);
4783 // Otherwise, we have to emit the load. If the pointer is to unfoldable but
4784 // still constant memory, the input chain can be the entry node.
4786 bool ConstantMemory = false;
4788 // Do not serialize (non-volatile) loads of constant memory with anything.
4789 if (Builder.AA->pointsToConstantMemory(PtrVal)) {
4790 Root = Builder.DAG.getEntryNode();
4791 ConstantMemory = true;
4793 // Do not serialize non-volatile loads against each other.
4794 Root = Builder.DAG.getRoot();
4797 SDValue Ptr = Builder.getValue(PtrVal);
4798 SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurDebugLoc(), Root,
4799 Ptr, PtrVal /*SrcValue*/, 0/*SVOffset*/,
4801 false /*nontemporal*/, 1 /* align=1 */);
4803 if (!ConstantMemory)
4804 Builder.PendingLoads.push_back(LoadVal.getValue(1));
4809 /// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form.
4810 /// If so, return true and lower it, otherwise return false and it will be
4811 /// lowered like a normal call.
4812 bool SelectionDAGBuilder::visitMemCmpCall(const CallInst &I) {
4813 // Verify that the prototype makes sense. int memcmp(void*,void*,size_t)
4814 if (I.getNumArgOperands() != 3)
4817 const Value *LHS = I.getArgOperand(0), *RHS = I.getArgOperand(1);
4818 if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() ||
4819 !I.getArgOperand(2)->getType()->isIntegerTy() ||
4820 !I.getType()->isIntegerTy())
4823 const ConstantInt *Size = dyn_cast<ConstantInt>(I.getArgOperand(2));
4825 // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0
4826 // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0
4827 if (Size && IsOnlyUsedInZeroEqualityComparison(&I)) {
4828 bool ActuallyDoIt = true;
4831 switch (Size->getZExtValue()) {
4833 LoadVT = MVT::Other;
4835 ActuallyDoIt = false;
4839 LoadTy = Type::getInt16Ty(Size->getContext());
4843 LoadTy = Type::getInt32Ty(Size->getContext());
4847 LoadTy = Type::getInt64Ty(Size->getContext());
4851 LoadVT = MVT::v4i32;
4852 LoadTy = Type::getInt32Ty(Size->getContext());
4853 LoadTy = VectorType::get(LoadTy, 4);
4858 // This turns into unaligned loads. We only do this if the target natively
4859 // supports the MVT we'll be loading or if it is small enough (<= 4) that
4860 // we'll only produce a small number of byte loads.
4862 // Require that we can find a legal MVT, and only do this if the target
4863 // supports unaligned loads of that type. Expanding into byte loads would
4865 if (ActuallyDoIt && Size->getZExtValue() > 4) {
4866 // TODO: Handle 5 byte compare as 4-byte + 1 byte.
4867 // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads.
4868 if (!TLI.isTypeLegal(LoadVT) ||!TLI.allowsUnalignedMemoryAccesses(LoadVT))
4869 ActuallyDoIt = false;
4873 SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this);
4874 SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this);
4876 SDValue Res = DAG.getSetCC(getCurDebugLoc(), MVT::i1, LHSVal, RHSVal,
4878 EVT CallVT = TLI.getValueType(I.getType(), true);
4879 setValue(&I, DAG.getZExtOrTrunc(Res, getCurDebugLoc(), CallVT));
4889 void SelectionDAGBuilder::visitCall(const CallInst &I) {
4890 // Handle inline assembly differently.
4891 if (isa<InlineAsm>(I.getCalledValue())) {
4896 const char *RenameFn = 0;
4897 if (Function *F = I.getCalledFunction()) {
4898 if (F->isDeclaration()) {
4899 if (const TargetIntrinsicInfo *II = TM.getIntrinsicInfo()) {
4900 if (unsigned IID = II->getIntrinsicID(F)) {
4901 RenameFn = visitIntrinsicCall(I, IID);
4906 if (unsigned IID = F->getIntrinsicID()) {
4907 RenameFn = visitIntrinsicCall(I, IID);
4913 // Check for well-known libc/libm calls. If the function is internal, it
4914 // can't be a library call.
4915 if (!F->hasLocalLinkage() && F->hasName()) {
4916 StringRef Name = F->getName();
4917 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl") {
4918 if (I.getNumArgOperands() == 2 && // Basic sanity checks.
4919 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
4920 I.getType() == I.getArgOperand(0)->getType() &&
4921 I.getType() == I.getArgOperand(1)->getType()) {
4922 SDValue LHS = getValue(I.getArgOperand(0));
4923 SDValue RHS = getValue(I.getArgOperand(1));
4924 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurDebugLoc(),
4925 LHS.getValueType(), LHS, RHS));
4928 } else if (Name == "fabs" || Name == "fabsf" || Name == "fabsl") {
4929 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
4930 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
4931 I.getType() == I.getArgOperand(0)->getType()) {
4932 SDValue Tmp = getValue(I.getArgOperand(0));
4933 setValue(&I, DAG.getNode(ISD::FABS, getCurDebugLoc(),
4934 Tmp.getValueType(), Tmp));
4937 } else if (Name == "sin" || Name == "sinf" || Name == "sinl") {
4938 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
4939 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
4940 I.getType() == I.getArgOperand(0)->getType() &&
4941 I.onlyReadsMemory()) {
4942 SDValue Tmp = getValue(I.getArgOperand(0));
4943 setValue(&I, DAG.getNode(ISD::FSIN, getCurDebugLoc(),
4944 Tmp.getValueType(), Tmp));
4947 } else if (Name == "cos" || Name == "cosf" || Name == "cosl") {
4948 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
4949 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
4950 I.getType() == I.getArgOperand(0)->getType() &&
4951 I.onlyReadsMemory()) {
4952 SDValue Tmp = getValue(I.getArgOperand(0));
4953 setValue(&I, DAG.getNode(ISD::FCOS, getCurDebugLoc(),
4954 Tmp.getValueType(), Tmp));
4957 } else if (Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl") {
4958 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
4959 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
4960 I.getType() == I.getArgOperand(0)->getType() &&
4961 I.onlyReadsMemory()) {
4962 SDValue Tmp = getValue(I.getArgOperand(0));
4963 setValue(&I, DAG.getNode(ISD::FSQRT, getCurDebugLoc(),
4964 Tmp.getValueType(), Tmp));
4967 } else if (Name == "memcmp") {
4968 if (visitMemCmpCall(I))
4976 Callee = getValue(I.getCalledValue());
4978 Callee = DAG.getExternalSymbol(RenameFn, TLI.getPointerTy());
4980 // Check if we can potentially perform a tail call. More detailed checking is
4981 // be done within LowerCallTo, after more information about the call is known.
4982 LowerCallTo(&I, Callee, I.isTailCall());
4987 /// AsmOperandInfo - This contains information for each constraint that we are
4989 class LLVM_LIBRARY_VISIBILITY SDISelAsmOperandInfo :
4990 public TargetLowering::AsmOperandInfo {
4992 /// CallOperand - If this is the result output operand or a clobber
4993 /// this is null, otherwise it is the incoming operand to the CallInst.
4994 /// This gets modified as the asm is processed.
4995 SDValue CallOperand;
4997 /// AssignedRegs - If this is a register or register class operand, this
4998 /// contains the set of register corresponding to the operand.
4999 RegsForValue AssignedRegs;
5001 explicit SDISelAsmOperandInfo(const InlineAsm::ConstraintInfo &info)
5002 : TargetLowering::AsmOperandInfo(info), CallOperand(0,0) {
5005 /// MarkAllocatedRegs - Once AssignedRegs is set, mark the assigned registers
5006 /// busy in OutputRegs/InputRegs.
5007 void MarkAllocatedRegs(bool isOutReg, bool isInReg,
5008 std::set<unsigned> &OutputRegs,
5009 std::set<unsigned> &InputRegs,
5010 const TargetRegisterInfo &TRI) const {
5012 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
5013 MarkRegAndAliases(AssignedRegs.Regs[i], OutputRegs, TRI);
5016 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
5017 MarkRegAndAliases(AssignedRegs.Regs[i], InputRegs, TRI);
5021 /// getCallOperandValEVT - Return the EVT of the Value* that this operand
5022 /// corresponds to. If there is no Value* for this operand, it returns
5024 EVT getCallOperandValEVT(LLVMContext &Context,
5025 const TargetLowering &TLI,
5026 const TargetData *TD) const {
5027 if (CallOperandVal == 0) return MVT::Other;
5029 if (isa<BasicBlock>(CallOperandVal))
5030 return TLI.getPointerTy();
5032 const llvm::Type *OpTy = CallOperandVal->getType();
5034 // If this is an indirect operand, the operand is a pointer to the
5037 const llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
5039 report_fatal_error("Indirect operand for inline asm not a pointer!");
5040 OpTy = PtrTy->getElementType();
5043 // If OpTy is not a single value, it may be a struct/union that we
5044 // can tile with integers.
5045 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
5046 unsigned BitSize = TD->getTypeSizeInBits(OpTy);
5055 OpTy = IntegerType::get(Context, BitSize);
5060 return TLI.getValueType(OpTy, true);
5064 /// MarkRegAndAliases - Mark the specified register and all aliases in the
5066 static void MarkRegAndAliases(unsigned Reg, std::set<unsigned> &Regs,
5067 const TargetRegisterInfo &TRI) {
5068 assert(TargetRegisterInfo::isPhysicalRegister(Reg) && "Isn't a physreg");
5070 if (const unsigned *Aliases = TRI.getAliasSet(Reg))
5071 for (; *Aliases; ++Aliases)
5072 Regs.insert(*Aliases);
5076 } // end llvm namespace.
5078 /// isAllocatableRegister - If the specified register is safe to allocate,
5079 /// i.e. it isn't a stack pointer or some other special register, return the
5080 /// register class for the register. Otherwise, return null.
5081 static const TargetRegisterClass *
5082 isAllocatableRegister(unsigned Reg, MachineFunction &MF,
5083 const TargetLowering &TLI,
5084 const TargetRegisterInfo *TRI) {
5085 EVT FoundVT = MVT::Other;
5086 const TargetRegisterClass *FoundRC = 0;
5087 for (TargetRegisterInfo::regclass_iterator RCI = TRI->regclass_begin(),
5088 E = TRI->regclass_end(); RCI != E; ++RCI) {
5089 EVT ThisVT = MVT::Other;
5091 const TargetRegisterClass *RC = *RCI;
5092 // If none of the value types for this register class are valid, we
5093 // can't use it. For example, 64-bit reg classes on 32-bit targets.
5094 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
5096 if (TLI.isTypeLegal(*I)) {
5097 // If we have already found this register in a different register class,
5098 // choose the one with the largest VT specified. For example, on
5099 // PowerPC, we favor f64 register classes over f32.
5100 if (FoundVT == MVT::Other || FoundVT.bitsLT(*I)) {
5107 if (ThisVT == MVT::Other) continue;
5109 // NOTE: This isn't ideal. In particular, this might allocate the
5110 // frame pointer in functions that need it (due to them not being taken
5111 // out of allocation, because a variable sized allocation hasn't been seen
5112 // yet). This is a slight code pessimization, but should still work.
5113 for (TargetRegisterClass::iterator I = RC->allocation_order_begin(MF),
5114 E = RC->allocation_order_end(MF); I != E; ++I)
5116 // We found a matching register class. Keep looking at others in case
5117 // we find one with larger registers that this physreg is also in.
5126 /// GetRegistersForValue - Assign registers (virtual or physical) for the
5127 /// specified operand. We prefer to assign virtual registers, to allow the
5128 /// register allocator to handle the assignment process. However, if the asm
5129 /// uses features that we can't model on machineinstrs, we have SDISel do the
5130 /// allocation. This produces generally horrible, but correct, code.
5132 /// OpInfo describes the operand.
5133 /// Input and OutputRegs are the set of already allocated physical registers.
5135 void SelectionDAGBuilder::
5136 GetRegistersForValue(SDISelAsmOperandInfo &OpInfo,
5137 std::set<unsigned> &OutputRegs,
5138 std::set<unsigned> &InputRegs) {
5139 LLVMContext &Context = FuncInfo.Fn->getContext();
5141 // Compute whether this value requires an input register, an output register,
5143 bool isOutReg = false;
5144 bool isInReg = false;
5145 switch (OpInfo.Type) {
5146 case InlineAsm::isOutput:
5149 // If there is an input constraint that matches this, we need to reserve
5150 // the input register so no other inputs allocate to it.
5151 isInReg = OpInfo.hasMatchingInput();
5153 case InlineAsm::isInput:
5157 case InlineAsm::isClobber:
5164 MachineFunction &MF = DAG.getMachineFunction();
5165 SmallVector<unsigned, 4> Regs;
5167 // If this is a constraint for a single physreg, or a constraint for a
5168 // register class, find it.
5169 std::pair<unsigned, const TargetRegisterClass*> PhysReg =
5170 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
5171 OpInfo.ConstraintVT);
5173 unsigned NumRegs = 1;
5174 if (OpInfo.ConstraintVT != MVT::Other) {
5175 // If this is a FP input in an integer register (or visa versa) insert a bit
5176 // cast of the input value. More generally, handle any case where the input
5177 // value disagrees with the register class we plan to stick this in.
5178 if (OpInfo.Type == InlineAsm::isInput &&
5179 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) {
5180 // Try to convert to the first EVT that the reg class contains. If the
5181 // types are identical size, use a bitcast to convert (e.g. two differing
5183 EVT RegVT = *PhysReg.second->vt_begin();
5184 if (RegVT.getSizeInBits() == OpInfo.ConstraintVT.getSizeInBits()) {
5185 OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
5186 RegVT, OpInfo.CallOperand);
5187 OpInfo.ConstraintVT = RegVT;
5188 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) {
5189 // If the input is a FP value and we want it in FP registers, do a
5190 // bitcast to the corresponding integer type. This turns an f64 value
5191 // into i64, which can be passed with two i32 values on a 32-bit
5193 RegVT = EVT::getIntegerVT(Context,
5194 OpInfo.ConstraintVT.getSizeInBits());
5195 OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
5196 RegVT, OpInfo.CallOperand);
5197 OpInfo.ConstraintVT = RegVT;
5201 NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT);
5205 EVT ValueVT = OpInfo.ConstraintVT;
5207 // If this is a constraint for a specific physical register, like {r17},
5209 if (unsigned AssignedReg = PhysReg.first) {
5210 const TargetRegisterClass *RC = PhysReg.second;
5211 if (OpInfo.ConstraintVT == MVT::Other)
5212 ValueVT = *RC->vt_begin();
5214 // Get the actual register value type. This is important, because the user
5215 // may have asked for (e.g.) the AX register in i32 type. We need to
5216 // remember that AX is actually i16 to get the right extension.
5217 RegVT = *RC->vt_begin();
5219 // This is a explicit reference to a physical register.
5220 Regs.push_back(AssignedReg);
5222 // If this is an expanded reference, add the rest of the regs to Regs.
5224 TargetRegisterClass::iterator I = RC->begin();
5225 for (; *I != AssignedReg; ++I)
5226 assert(I != RC->end() && "Didn't find reg!");
5228 // Already added the first reg.
5230 for (; NumRegs; --NumRegs, ++I) {
5231 assert(I != RC->end() && "Ran out of registers to allocate!");
5236 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5237 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
5238 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
5242 // Otherwise, if this was a reference to an LLVM register class, create vregs
5243 // for this reference.
5244 if (const TargetRegisterClass *RC = PhysReg.second) {
5245 RegVT = *RC->vt_begin();
5246 if (OpInfo.ConstraintVT == MVT::Other)
5249 // Create the appropriate number of virtual registers.
5250 MachineRegisterInfo &RegInfo = MF.getRegInfo();
5251 for (; NumRegs; --NumRegs)
5252 Regs.push_back(RegInfo.createVirtualRegister(RC));
5254 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5258 // This is a reference to a register class that doesn't directly correspond
5259 // to an LLVM register class. Allocate NumRegs consecutive, available,
5260 // registers from the class.
5261 std::vector<unsigned> RegClassRegs
5262 = TLI.getRegClassForInlineAsmConstraint(OpInfo.ConstraintCode,
5263 OpInfo.ConstraintVT);
5265 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
5266 unsigned NumAllocated = 0;
5267 for (unsigned i = 0, e = RegClassRegs.size(); i != e; ++i) {
5268 unsigned Reg = RegClassRegs[i];
5269 // See if this register is available.
5270 if ((isOutReg && OutputRegs.count(Reg)) || // Already used.
5271 (isInReg && InputRegs.count(Reg))) { // Already used.
5272 // Make sure we find consecutive registers.
5277 // Check to see if this register is allocatable (i.e. don't give out the
5279 const TargetRegisterClass *RC = isAllocatableRegister(Reg, MF, TLI, TRI);
5280 if (!RC) { // Couldn't allocate this register.
5281 // Reset NumAllocated to make sure we return consecutive registers.
5286 // Okay, this register is good, we can use it.
5289 // If we allocated enough consecutive registers, succeed.
5290 if (NumAllocated == NumRegs) {
5291 unsigned RegStart = (i-NumAllocated)+1;
5292 unsigned RegEnd = i+1;
5293 // Mark all of the allocated registers used.
5294 for (unsigned i = RegStart; i != RegEnd; ++i)
5295 Regs.push_back(RegClassRegs[i]);
5297 OpInfo.AssignedRegs = RegsForValue(Regs, *RC->vt_begin(),
5298 OpInfo.ConstraintVT);
5299 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
5304 // Otherwise, we couldn't allocate enough registers for this.
5307 /// visitInlineAsm - Handle a call to an InlineAsm object.
5309 void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) {
5310 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
5312 /// ConstraintOperands - Information about all of the constraints.
5313 std::vector<SDISelAsmOperandInfo> ConstraintOperands;
5315 std::set<unsigned> OutputRegs, InputRegs;
5317 // Do a prepass over the constraints, canonicalizing them, and building up the
5318 // ConstraintOperands list.
5319 std::vector<InlineAsm::ConstraintInfo>
5320 ConstraintInfos = IA->ParseConstraints();
5322 bool hasMemory = hasInlineAsmMemConstraint(ConstraintInfos, TLI);
5324 SDValue Chain, Flag;
5326 // We won't need to flush pending loads if this asm doesn't touch
5327 // memory and is nonvolatile.
5328 if (hasMemory || IA->hasSideEffects())
5331 Chain = DAG.getRoot();
5333 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
5334 unsigned ResNo = 0; // ResNo - The result number of the next output.
5335 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
5336 ConstraintOperands.push_back(SDISelAsmOperandInfo(ConstraintInfos[i]));
5337 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
5339 EVT OpVT = MVT::Other;
5341 // Compute the value type for each operand.
5342 switch (OpInfo.Type) {
5343 case InlineAsm::isOutput:
5344 // Indirect outputs just consume an argument.
5345 if (OpInfo.isIndirect) {
5346 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5350 // The return value of the call is this value. As such, there is no
5351 // corresponding argument.
5352 assert(!CS.getType()->isVoidTy() &&
5354 if (const StructType *STy = dyn_cast<StructType>(CS.getType())) {
5355 OpVT = TLI.getValueType(STy->getElementType(ResNo));
5357 assert(ResNo == 0 && "Asm only has one result!");
5358 OpVT = TLI.getValueType(CS.getType());
5362 case InlineAsm::isInput:
5363 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5365 case InlineAsm::isClobber:
5370 // If this is an input or an indirect output, process the call argument.
5371 // BasicBlocks are labels, currently appearing only in asm's.
5372 if (OpInfo.CallOperandVal) {
5373 // Strip bitcasts, if any. This mostly comes up for functions.
5374 OpInfo.CallOperandVal = OpInfo.CallOperandVal->stripPointerCasts();
5376 if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) {
5377 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]);
5379 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
5382 OpVT = OpInfo.getCallOperandValEVT(*DAG.getContext(), TLI, TD);
5385 OpInfo.ConstraintVT = OpVT;
5388 // Second pass over the constraints: compute which constraint option to use
5389 // and assign registers to constraints that want a specific physreg.
5390 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
5391 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5393 // If this is an output operand with a matching input operand, look up the
5394 // matching input. If their types mismatch, e.g. one is an integer, the
5395 // other is floating point, or their sizes are different, flag it as an
5397 if (OpInfo.hasMatchingInput()) {
5398 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
5400 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
5401 if ((OpInfo.ConstraintVT.isInteger() !=
5402 Input.ConstraintVT.isInteger()) ||
5403 (OpInfo.ConstraintVT.getSizeInBits() !=
5404 Input.ConstraintVT.getSizeInBits())) {
5405 report_fatal_error("Unsupported asm: input constraint"
5406 " with a matching output constraint of"
5407 " incompatible type!");
5409 Input.ConstraintVT = OpInfo.ConstraintVT;
5413 // Compute the constraint code and ConstraintType to use.
5414 TLI.ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, &DAG);
5416 // If this is a memory input, and if the operand is not indirect, do what we
5417 // need to to provide an address for the memory input.
5418 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5419 !OpInfo.isIndirect) {
5420 assert(OpInfo.Type == InlineAsm::isInput &&
5421 "Can only indirectify direct input operands!");
5423 // Memory operands really want the address of the value. If we don't have
5424 // an indirect input, put it in the constpool if we can, otherwise spill
5425 // it to a stack slot.
5427 // If the operand is a float, integer, or vector constant, spill to a
5428 // constant pool entry to get its address.
5429 const Value *OpVal = OpInfo.CallOperandVal;
5430 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
5431 isa<ConstantVector>(OpVal)) {
5432 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal),
5433 TLI.getPointerTy());
5435 // Otherwise, create a stack slot and emit a store to it before the
5437 const Type *Ty = OpVal->getType();
5438 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
5439 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(Ty);
5440 MachineFunction &MF = DAG.getMachineFunction();
5441 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
5442 SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
5443 Chain = DAG.getStore(Chain, getCurDebugLoc(),
5444 OpInfo.CallOperand, StackSlot, NULL, 0,
5446 OpInfo.CallOperand = StackSlot;
5449 // There is no longer a Value* corresponding to this operand.
5450 OpInfo.CallOperandVal = 0;
5452 // It is now an indirect operand.
5453 OpInfo.isIndirect = true;
5456 // If this constraint is for a specific register, allocate it before
5458 if (OpInfo.ConstraintType == TargetLowering::C_Register)
5459 GetRegistersForValue(OpInfo, OutputRegs, InputRegs);
5462 ConstraintInfos.clear();
5464 // Second pass - Loop over all of the operands, assigning virtual or physregs
5465 // to register class operands.
5466 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5467 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5469 // C_Register operands have already been allocated, Other/Memory don't need
5471 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
5472 GetRegistersForValue(OpInfo, OutputRegs, InputRegs);
5475 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
5476 std::vector<SDValue> AsmNodeOperands;
5477 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain
5478 AsmNodeOperands.push_back(
5479 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(),
5480 TLI.getPointerTy()));
5482 // If we have a !srcloc metadata node associated with it, we want to attach
5483 // this to the ultimately generated inline asm machineinstr. To do this, we
5484 // pass in the third operand as this (potentially null) inline asm MDNode.
5485 const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc");
5486 AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc));
5488 // Remember the AlignStack bit as operand 3.
5489 AsmNodeOperands.push_back(DAG.getTargetConstant(IA->isAlignStack() ? 1 : 0,
5492 // Loop over all of the inputs, copying the operand values into the
5493 // appropriate registers and processing the output regs.
5494 RegsForValue RetValRegs;
5496 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
5497 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
5499 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5500 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5502 switch (OpInfo.Type) {
5503 case InlineAsm::isOutput: {
5504 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
5505 OpInfo.ConstraintType != TargetLowering::C_Register) {
5506 // Memory output, or 'other' output (e.g. 'X' constraint).
5507 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
5509 // Add information to the INLINEASM node to know about this output.
5510 unsigned OpFlags = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
5511 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlags,
5512 TLI.getPointerTy()));
5513 AsmNodeOperands.push_back(OpInfo.CallOperand);
5517 // Otherwise, this is a register or register class output.
5519 // Copy the output from the appropriate register. Find a register that
5521 if (OpInfo.AssignedRegs.Regs.empty())
5522 report_fatal_error("Couldn't allocate output reg for constraint '" +
5523 Twine(OpInfo.ConstraintCode) + "'!");
5525 // If this is an indirect operand, store through the pointer after the
5527 if (OpInfo.isIndirect) {
5528 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
5529 OpInfo.CallOperandVal));
5531 // This is the result value of the call.
5532 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
5533 // Concatenate this output onto the outputs list.
5534 RetValRegs.append(OpInfo.AssignedRegs);
5537 // Add information to the INLINEASM node to know that this register is
5539 OpInfo.AssignedRegs.AddInlineAsmOperands(OpInfo.isEarlyClobber ?
5540 InlineAsm::Kind_RegDefEarlyClobber :
5541 InlineAsm::Kind_RegDef,
5548 case InlineAsm::isInput: {
5549 SDValue InOperandVal = OpInfo.CallOperand;
5551 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint?
5552 // If this is required to match an output register we have already set,
5553 // just use its register.
5554 unsigned OperandNo = OpInfo.getMatchedOperand();
5556 // Scan until we find the definition we already emitted of this operand.
5557 // When we find it, create a RegsForValue operand.
5558 unsigned CurOp = InlineAsm::Op_FirstOperand;
5559 for (; OperandNo; --OperandNo) {
5560 // Advance to the next operand.
5562 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
5563 assert((InlineAsm::isRegDefKind(OpFlag) ||
5564 InlineAsm::isRegDefEarlyClobberKind(OpFlag) ||
5565 InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?");
5566 CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1;
5570 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
5571 if (InlineAsm::isRegDefKind(OpFlag) ||
5572 InlineAsm::isRegDefEarlyClobberKind(OpFlag)) {
5573 // Add (OpFlag&0xffff)>>3 registers to MatchedRegs.
5574 if (OpInfo.isIndirect) {
5575 // This happens on gcc/testsuite/gcc.dg/pr8788-1.c
5576 LLVMContext &Ctx = *DAG.getContext();
5577 Ctx.emitError(CS.getInstruction(), "inline asm not supported yet:"
5578 " don't know how to handle tied "
5579 "indirect register inputs");
5582 RegsForValue MatchedRegs;
5583 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
5584 EVT RegVT = AsmNodeOperands[CurOp+1].getValueType();
5585 MatchedRegs.RegVTs.push_back(RegVT);
5586 MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo();
5587 for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag);
5589 MatchedRegs.Regs.push_back
5590 (RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT)));
5592 // Use the produced MatchedRegs object to
5593 MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
5595 MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse,
5596 true, OpInfo.getMatchedOperand(),
5597 DAG, AsmNodeOperands);
5601 assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!");
5602 assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 &&
5603 "Unexpected number of operands");
5604 // Add information to the INLINEASM node to know about this input.
5605 // See InlineAsm.h isUseOperandTiedToDef.
5606 OpFlag = InlineAsm::getFlagWordForMatchingOp(OpFlag,
5607 OpInfo.getMatchedOperand());
5608 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlag,
5609 TLI.getPointerTy()));
5610 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]);
5614 // Treat indirect 'X' constraint as memory.
5615 if (OpInfo.ConstraintType == TargetLowering::C_Other &&
5617 OpInfo.ConstraintType = TargetLowering::C_Memory;
5619 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
5620 std::vector<SDValue> Ops;
5621 TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode[0],
5624 report_fatal_error("Invalid operand for inline asm constraint '" +
5625 Twine(OpInfo.ConstraintCode) + "'!");
5627 // Add information to the INLINEASM node to know about this input.
5628 unsigned ResOpType =
5629 InlineAsm::getFlagWord(InlineAsm::Kind_Imm, Ops.size());
5630 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
5631 TLI.getPointerTy()));
5632 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
5636 if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
5637 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
5638 assert(InOperandVal.getValueType() == TLI.getPointerTy() &&
5639 "Memory operands expect pointer values");
5641 // Add information to the INLINEASM node to know about this input.
5642 unsigned ResOpType = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
5643 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
5644 TLI.getPointerTy()));
5645 AsmNodeOperands.push_back(InOperandVal);
5649 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
5650 OpInfo.ConstraintType == TargetLowering::C_Register) &&
5651 "Unknown constraint type!");
5652 assert(!OpInfo.isIndirect &&
5653 "Don't know how to handle indirect register inputs yet!");
5655 // Copy the input into the appropriate registers.
5656 if (OpInfo.AssignedRegs.Regs.empty() ||
5657 !OpInfo.AssignedRegs.areValueTypesLegal(TLI))
5658 report_fatal_error("Couldn't allocate input reg for constraint '" +
5659 Twine(OpInfo.ConstraintCode) + "'!");
5661 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
5664 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, false, 0,
5665 DAG, AsmNodeOperands);
5668 case InlineAsm::isClobber: {
5669 // Add the clobbered value to the operand list, so that the register
5670 // allocator is aware that the physreg got clobbered.
5671 if (!OpInfo.AssignedRegs.Regs.empty())
5672 OpInfo.AssignedRegs.AddInlineAsmOperands(
5673 InlineAsm::Kind_RegDefEarlyClobber,
5681 // Finish up input operands. Set the input chain and add the flag last.
5682 AsmNodeOperands[InlineAsm::Op_InputChain] = Chain;
5683 if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
5685 Chain = DAG.getNode(ISD::INLINEASM, getCurDebugLoc(),
5686 DAG.getVTList(MVT::Other, MVT::Flag),
5687 &AsmNodeOperands[0], AsmNodeOperands.size());
5688 Flag = Chain.getValue(1);
5690 // If this asm returns a register value, copy the result from that register
5691 // and set it as the value of the call.
5692 if (!RetValRegs.Regs.empty()) {
5693 SDValue Val = RetValRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(),
5696 // FIXME: Why don't we do this for inline asms with MRVs?
5697 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) {
5698 EVT ResultType = TLI.getValueType(CS.getType());
5700 // If any of the results of the inline asm is a vector, it may have the
5701 // wrong width/num elts. This can happen for register classes that can
5702 // contain multiple different value types. The preg or vreg allocated may
5703 // not have the same VT as was expected. Convert it to the right type
5704 // with bit_convert.
5705 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) {
5706 Val = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
5709 } else if (ResultType != Val.getValueType() &&
5710 ResultType.isInteger() && Val.getValueType().isInteger()) {
5711 // If a result value was tied to an input value, the computed result may
5712 // have a wider width than the expected result. Extract the relevant
5714 Val = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), ResultType, Val);
5717 assert(ResultType == Val.getValueType() && "Asm result value mismatch!");
5720 setValue(CS.getInstruction(), Val);
5721 // Don't need to use this as a chain in this case.
5722 if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty())
5726 std::vector<std::pair<SDValue, const Value *> > StoresToEmit;
5728 // Process indirect outputs, first output all of the flagged copies out of
5730 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
5731 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
5732 const Value *Ptr = IndirectStoresToEmit[i].second;
5733 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(),
5735 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
5738 // Emit the non-flagged stores from the physregs.
5739 SmallVector<SDValue, 8> OutChains;
5740 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) {
5741 SDValue Val = DAG.getStore(Chain, getCurDebugLoc(),
5742 StoresToEmit[i].first,
5743 getValue(StoresToEmit[i].second),
5744 StoresToEmit[i].second, 0,
5746 OutChains.push_back(Val);
5749 if (!OutChains.empty())
5750 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
5751 &OutChains[0], OutChains.size());
5756 void SelectionDAGBuilder::visitVAStart(const CallInst &I) {
5757 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurDebugLoc(),
5758 MVT::Other, getRoot(),
5759 getValue(I.getArgOperand(0)),
5760 DAG.getSrcValue(I.getArgOperand(0))));
5763 void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) {
5764 const TargetData &TD = *TLI.getTargetData();
5765 SDValue V = DAG.getVAArg(TLI.getValueType(I.getType()), getCurDebugLoc(),
5766 getRoot(), getValue(I.getOperand(0)),
5767 DAG.getSrcValue(I.getOperand(0)),
5768 TD.getABITypeAlignment(I.getType()));
5770 DAG.setRoot(V.getValue(1));
5773 void SelectionDAGBuilder::visitVAEnd(const CallInst &I) {
5774 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurDebugLoc(),
5775 MVT::Other, getRoot(),
5776 getValue(I.getArgOperand(0)),
5777 DAG.getSrcValue(I.getArgOperand(0))));
5780 void SelectionDAGBuilder::visitVACopy(const CallInst &I) {
5781 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurDebugLoc(),
5782 MVT::Other, getRoot(),
5783 getValue(I.getArgOperand(0)),
5784 getValue(I.getArgOperand(1)),
5785 DAG.getSrcValue(I.getArgOperand(0)),
5786 DAG.getSrcValue(I.getArgOperand(1))));
5789 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
5790 /// implementation, which just calls LowerCall.
5791 /// FIXME: When all targets are
5792 /// migrated to using LowerCall, this hook should be integrated into SDISel.
5793 std::pair<SDValue, SDValue>
5794 TargetLowering::LowerCallTo(SDValue Chain, const Type *RetTy,
5795 bool RetSExt, bool RetZExt, bool isVarArg,
5796 bool isInreg, unsigned NumFixedArgs,
5797 CallingConv::ID CallConv, bool isTailCall,
5798 bool isReturnValueUsed,
5800 ArgListTy &Args, SelectionDAG &DAG,
5801 DebugLoc dl) const {
5802 // Handle all of the outgoing arguments.
5803 SmallVector<ISD::OutputArg, 32> Outs;
5804 SmallVector<SDValue, 32> OutVals;
5805 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
5806 SmallVector<EVT, 4> ValueVTs;
5807 ComputeValueVTs(*this, Args[i].Ty, ValueVTs);
5808 for (unsigned Value = 0, NumValues = ValueVTs.size();
5809 Value != NumValues; ++Value) {
5810 EVT VT = ValueVTs[Value];
5811 const Type *ArgTy = VT.getTypeForEVT(RetTy->getContext());
5812 SDValue Op = SDValue(Args[i].Node.getNode(),
5813 Args[i].Node.getResNo() + Value);
5814 ISD::ArgFlagsTy Flags;
5815 unsigned OriginalAlignment =
5816 getTargetData()->getABITypeAlignment(ArgTy);
5822 if (Args[i].isInReg)
5826 if (Args[i].isByVal) {
5828 const PointerType *Ty = cast<PointerType>(Args[i].Ty);
5829 const Type *ElementTy = Ty->getElementType();
5830 unsigned FrameAlign = getByValTypeAlignment(ElementTy);
5831 unsigned FrameSize = getTargetData()->getTypeAllocSize(ElementTy);
5832 // For ByVal, alignment should come from FE. BE will guess if this
5833 // info is not there but there are cases it cannot get right.
5834 if (Args[i].Alignment)
5835 FrameAlign = Args[i].Alignment;
5836 Flags.setByValAlign(FrameAlign);
5837 Flags.setByValSize(FrameSize);
5841 Flags.setOrigAlign(OriginalAlignment);
5843 EVT PartVT = getRegisterType(RetTy->getContext(), VT);
5844 unsigned NumParts = getNumRegisters(RetTy->getContext(), VT);
5845 SmallVector<SDValue, 4> Parts(NumParts);
5846 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
5849 ExtendKind = ISD::SIGN_EXTEND;
5850 else if (Args[i].isZExt)
5851 ExtendKind = ISD::ZERO_EXTEND;
5853 getCopyToParts(DAG, dl, Op, &Parts[0], NumParts,
5854 PartVT, ExtendKind);
5856 for (unsigned j = 0; j != NumParts; ++j) {
5857 // if it isn't first piece, alignment must be 1
5858 ISD::OutputArg MyFlags(Flags, Parts[j].getValueType(),
5860 if (NumParts > 1 && j == 0)
5861 MyFlags.Flags.setSplit();
5863 MyFlags.Flags.setOrigAlign(1);
5865 Outs.push_back(MyFlags);
5866 OutVals.push_back(Parts[j]);
5871 // Handle the incoming return values from the call.
5872 SmallVector<ISD::InputArg, 32> Ins;
5873 SmallVector<EVT, 4> RetTys;
5874 ComputeValueVTs(*this, RetTy, RetTys);
5875 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
5877 EVT RegisterVT = getRegisterType(RetTy->getContext(), VT);
5878 unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT);
5879 for (unsigned i = 0; i != NumRegs; ++i) {
5880 ISD::InputArg MyFlags;
5881 MyFlags.VT = RegisterVT;
5882 MyFlags.Used = isReturnValueUsed;
5884 MyFlags.Flags.setSExt();
5886 MyFlags.Flags.setZExt();
5888 MyFlags.Flags.setInReg();
5889 Ins.push_back(MyFlags);
5893 SmallVector<SDValue, 4> InVals;
5894 Chain = LowerCall(Chain, Callee, CallConv, isVarArg, isTailCall,
5895 Outs, OutVals, Ins, dl, DAG, InVals);
5897 // Verify that the target's LowerCall behaved as expected.
5898 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
5899 "LowerCall didn't return a valid chain!");
5900 assert((!isTailCall || InVals.empty()) &&
5901 "LowerCall emitted a return value for a tail call!");
5902 assert((isTailCall || InVals.size() == Ins.size()) &&
5903 "LowerCall didn't emit the correct number of values!");
5905 // For a tail call, the return value is merely live-out and there aren't
5906 // any nodes in the DAG representing it. Return a special value to
5907 // indicate that a tail call has been emitted and no more Instructions
5908 // should be processed in the current block.
5911 return std::make_pair(SDValue(), SDValue());
5914 DEBUG(for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
5915 assert(InVals[i].getNode() &&
5916 "LowerCall emitted a null value!");
5917 assert(Ins[i].VT == InVals[i].getValueType() &&
5918 "LowerCall emitted a value with the wrong type!");
5921 // Collect the legal value parts into potentially illegal values
5922 // that correspond to the original function's return values.
5923 ISD::NodeType AssertOp = ISD::DELETED_NODE;
5925 AssertOp = ISD::AssertSext;
5927 AssertOp = ISD::AssertZext;
5928 SmallVector<SDValue, 4> ReturnValues;
5929 unsigned CurReg = 0;
5930 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
5932 EVT RegisterVT = getRegisterType(RetTy->getContext(), VT);
5933 unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT);
5935 ReturnValues.push_back(getCopyFromParts(DAG, dl, &InVals[CurReg],
5936 NumRegs, RegisterVT, VT,
5941 // For a function returning void, there is no return value. We can't create
5942 // such a node, so we just return a null return value in that case. In
5943 // that case, nothing will actualy look at the value.
5944 if (ReturnValues.empty())
5945 return std::make_pair(SDValue(), Chain);
5947 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, dl,
5948 DAG.getVTList(&RetTys[0], RetTys.size()),
5949 &ReturnValues[0], ReturnValues.size());
5950 return std::make_pair(Res, Chain);
5953 void TargetLowering::LowerOperationWrapper(SDNode *N,
5954 SmallVectorImpl<SDValue> &Results,
5955 SelectionDAG &DAG) const {
5956 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
5958 Results.push_back(Res);
5961 SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5962 llvm_unreachable("LowerOperation not implemented for this target!");
5967 SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) {
5968 SDValue Op = getNonRegisterValue(V);
5969 assert((Op.getOpcode() != ISD::CopyFromReg ||
5970 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
5971 "Copy from a reg to the same reg!");
5972 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg");
5974 RegsForValue RFV(V->getContext(), TLI, Reg, V->getType());
5975 SDValue Chain = DAG.getEntryNode();
5976 RFV.getCopyToRegs(Op, DAG, getCurDebugLoc(), Chain, 0);
5977 PendingExports.push_back(Chain);
5980 #include "llvm/CodeGen/SelectionDAGISel.h"
5982 void SelectionDAGISel::LowerArguments(const BasicBlock *LLVMBB) {
5983 // If this is the entry block, emit arguments.
5984 const Function &F = *LLVMBB->getParent();
5985 SelectionDAG &DAG = SDB->DAG;
5986 DebugLoc dl = SDB->getCurDebugLoc();
5987 const TargetData *TD = TLI.getTargetData();
5988 SmallVector<ISD::InputArg, 16> Ins;
5990 // Check whether the function can return without sret-demotion.
5991 SmallVector<ISD::OutputArg, 4> Outs;
5992 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
5995 if (!FuncInfo->CanLowerReturn) {
5996 // Put in an sret pointer parameter before all the other parameters.
5997 SmallVector<EVT, 1> ValueVTs;
5998 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
6000 // NOTE: Assuming that a pointer will never break down to more than one VT
6002 ISD::ArgFlagsTy Flags;
6004 EVT RegisterVT = TLI.getRegisterType(*DAG.getContext(), ValueVTs[0]);
6005 ISD::InputArg RetArg(Flags, RegisterVT, true);
6006 Ins.push_back(RetArg);
6009 // Set up the incoming argument description vector.
6011 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end();
6012 I != E; ++I, ++Idx) {
6013 SmallVector<EVT, 4> ValueVTs;
6014 ComputeValueVTs(TLI, I->getType(), ValueVTs);
6015 bool isArgValueUsed = !I->use_empty();
6016 for (unsigned Value = 0, NumValues = ValueVTs.size();
6017 Value != NumValues; ++Value) {
6018 EVT VT = ValueVTs[Value];
6019 const Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
6020 ISD::ArgFlagsTy Flags;
6021 unsigned OriginalAlignment =
6022 TD->getABITypeAlignment(ArgTy);
6024 if (F.paramHasAttr(Idx, Attribute::ZExt))
6026 if (F.paramHasAttr(Idx, Attribute::SExt))
6028 if (F.paramHasAttr(Idx, Attribute::InReg))
6030 if (F.paramHasAttr(Idx, Attribute::StructRet))
6032 if (F.paramHasAttr(Idx, Attribute::ByVal)) {
6034 const PointerType *Ty = cast<PointerType>(I->getType());
6035 const Type *ElementTy = Ty->getElementType();
6036 unsigned FrameAlign = TLI.getByValTypeAlignment(ElementTy);
6037 unsigned FrameSize = TD->getTypeAllocSize(ElementTy);
6038 // For ByVal, alignment should be passed from FE. BE will guess if
6039 // this info is not there but there are cases it cannot get right.
6040 if (F.getParamAlignment(Idx))
6041 FrameAlign = F.getParamAlignment(Idx);
6042 Flags.setByValAlign(FrameAlign);
6043 Flags.setByValSize(FrameSize);
6045 if (F.paramHasAttr(Idx, Attribute::Nest))
6047 Flags.setOrigAlign(OriginalAlignment);
6049 EVT RegisterVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6050 unsigned NumRegs = TLI.getNumRegisters(*CurDAG->getContext(), VT);
6051 for (unsigned i = 0; i != NumRegs; ++i) {
6052 ISD::InputArg MyFlags(Flags, RegisterVT, isArgValueUsed);
6053 if (NumRegs > 1 && i == 0)
6054 MyFlags.Flags.setSplit();
6055 // if it isn't first piece, alignment must be 1
6057 MyFlags.Flags.setOrigAlign(1);
6058 Ins.push_back(MyFlags);
6063 // Call the target to set up the argument values.
6064 SmallVector<SDValue, 8> InVals;
6065 SDValue NewRoot = TLI.LowerFormalArguments(DAG.getRoot(), F.getCallingConv(),
6069 // Verify that the target's LowerFormalArguments behaved as expected.
6070 assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other &&
6071 "LowerFormalArguments didn't return a valid chain!");
6072 assert(InVals.size() == Ins.size() &&
6073 "LowerFormalArguments didn't emit the correct number of values!");
6075 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
6076 assert(InVals[i].getNode() &&
6077 "LowerFormalArguments emitted a null value!");
6078 assert(Ins[i].VT == InVals[i].getValueType() &&
6079 "LowerFormalArguments emitted a value with the wrong type!");
6083 // Update the DAG with the new chain value resulting from argument lowering.
6084 DAG.setRoot(NewRoot);
6086 // Set up the argument values.
6089 if (!FuncInfo->CanLowerReturn) {
6090 // Create a virtual register for the sret pointer, and put in a copy
6091 // from the sret argument into it.
6092 SmallVector<EVT, 1> ValueVTs;
6093 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
6094 EVT VT = ValueVTs[0];
6095 EVT RegVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6096 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6097 SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1,
6098 RegVT, VT, AssertOp);
6100 MachineFunction& MF = SDB->DAG.getMachineFunction();
6101 MachineRegisterInfo& RegInfo = MF.getRegInfo();
6102 unsigned SRetReg = RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT));
6103 FuncInfo->DemoteRegister = SRetReg;
6104 NewRoot = SDB->DAG.getCopyToReg(NewRoot, SDB->getCurDebugLoc(),
6106 DAG.setRoot(NewRoot);
6108 // i indexes lowered arguments. Bump it past the hidden sret argument.
6109 // Idx indexes LLVM arguments. Don't touch it.
6113 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
6115 SmallVector<SDValue, 4> ArgValues;
6116 SmallVector<EVT, 4> ValueVTs;
6117 ComputeValueVTs(TLI, I->getType(), ValueVTs);
6118 unsigned NumValues = ValueVTs.size();
6120 // If this argument is unused then remember its value. It is used to generate
6121 // debugging information.
6122 if (I->use_empty() && NumValues)
6123 SDB->setUnusedArgValue(I, InVals[i]);
6125 for (unsigned Value = 0; Value != NumValues; ++Value) {
6126 EVT VT = ValueVTs[Value];
6127 EVT PartVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6128 unsigned NumParts = TLI.getNumRegisters(*CurDAG->getContext(), VT);
6130 if (!I->use_empty()) {
6131 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6132 if (F.paramHasAttr(Idx, Attribute::SExt))
6133 AssertOp = ISD::AssertSext;
6134 else if (F.paramHasAttr(Idx, Attribute::ZExt))
6135 AssertOp = ISD::AssertZext;
6137 ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i],
6138 NumParts, PartVT, VT,
6145 if (!I->use_empty()) {
6147 if (!ArgValues.empty())
6148 Res = DAG.getMergeValues(&ArgValues[0], NumValues,
6149 SDB->getCurDebugLoc());
6150 SDB->setValue(I, Res);
6152 // If this argument is live outside of the entry block, insert a copy from
6153 // whereever we got it to the vreg that other BB's will reference it as.
6154 SDB->CopyToExportRegsIfNeeded(I);
6158 assert(i == InVals.size() && "Argument register count mismatch!");
6160 // Finally, if the target has anything special to do, allow it to do so.
6161 // FIXME: this should insert code into the DAG!
6162 EmitFunctionEntryCode();
6165 /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
6166 /// ensure constants are generated when needed. Remember the virtual registers
6167 /// that need to be added to the Machine PHI nodes as input. We cannot just
6168 /// directly add them, because expansion might result in multiple MBB's for one
6169 /// BB. As such, the start of the BB might correspond to a different MBB than
6173 SelectionDAGBuilder::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
6174 const TerminatorInst *TI = LLVMBB->getTerminator();
6176 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
6178 // Check successor nodes' PHI nodes that expect a constant to be available
6180 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
6181 const BasicBlock *SuccBB = TI->getSuccessor(succ);
6182 if (!isa<PHINode>(SuccBB->begin())) continue;
6183 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
6185 // If this terminator has multiple identical successors (common for
6186 // switches), only handle each succ once.
6187 if (!SuccsHandled.insert(SuccMBB)) continue;
6189 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
6191 // At this point we know that there is a 1-1 correspondence between LLVM PHI
6192 // nodes and Machine PHI nodes, but the incoming operands have not been
6194 for (BasicBlock::const_iterator I = SuccBB->begin();
6195 const PHINode *PN = dyn_cast<PHINode>(I); ++I) {
6196 // Ignore dead phi's.
6197 if (PN->use_empty()) continue;
6200 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
6202 if (const Constant *C = dyn_cast<Constant>(PHIOp)) {
6203 unsigned &RegOut = ConstantsOut[C];
6205 RegOut = FuncInfo.CreateRegs(C->getType());
6206 CopyValueToVirtualRegister(C, RegOut);
6210 DenseMap<const Value *, unsigned>::iterator I =
6211 FuncInfo.ValueMap.find(PHIOp);
6212 if (I != FuncInfo.ValueMap.end())
6215 assert(isa<AllocaInst>(PHIOp) &&
6216 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
6217 "Didn't codegen value into a register!??");
6218 Reg = FuncInfo.CreateRegs(PHIOp->getType());
6219 CopyValueToVirtualRegister(PHIOp, Reg);
6223 // Remember that this register needs to added to the machine PHI node as
6224 // the input for this MBB.
6225 SmallVector<EVT, 4> ValueVTs;
6226 ComputeValueVTs(TLI, PN->getType(), ValueVTs);
6227 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
6228 EVT VT = ValueVTs[vti];
6229 unsigned NumRegisters = TLI.getNumRegisters(*DAG.getContext(), VT);
6230 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
6231 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
6232 Reg += NumRegisters;
6236 ConstantsOut.clear();