1 //===-- SelectionDAGBuilder.cpp - Selection-DAG building ------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements routines for translating from LLVM IR into SelectionDAG IR.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "isel"
15 #include "SDNodeDbgValue.h"
16 #include "SelectionDAGBuilder.h"
17 #include "llvm/ADT/BitVector.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/Analysis/AliasAnalysis.h"
20 #include "llvm/Analysis/ConstantFolding.h"
21 #include "llvm/Constants.h"
22 #include "llvm/CallingConv.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/Function.h"
25 #include "llvm/GlobalVariable.h"
26 #include "llvm/InlineAsm.h"
27 #include "llvm/Instructions.h"
28 #include "llvm/Intrinsics.h"
29 #include "llvm/IntrinsicInst.h"
30 #include "llvm/LLVMContext.h"
31 #include "llvm/Module.h"
32 #include "llvm/CodeGen/Analysis.h"
33 #include "llvm/CodeGen/FastISel.h"
34 #include "llvm/CodeGen/FunctionLoweringInfo.h"
35 #include "llvm/CodeGen/GCStrategy.h"
36 #include "llvm/CodeGen/GCMetadata.h"
37 #include "llvm/CodeGen/MachineFunction.h"
38 #include "llvm/CodeGen/MachineFrameInfo.h"
39 #include "llvm/CodeGen/MachineInstrBuilder.h"
40 #include "llvm/CodeGen/MachineJumpTableInfo.h"
41 #include "llvm/CodeGen/MachineModuleInfo.h"
42 #include "llvm/CodeGen/MachineRegisterInfo.h"
43 #include "llvm/CodeGen/PseudoSourceValue.h"
44 #include "llvm/CodeGen/SelectionDAG.h"
45 #include "llvm/Analysis/DebugInfo.h"
46 #include "llvm/Target/TargetRegisterInfo.h"
47 #include "llvm/Target/TargetData.h"
48 #include "llvm/Target/TargetFrameInfo.h"
49 #include "llvm/Target/TargetInstrInfo.h"
50 #include "llvm/Target/TargetIntrinsicInfo.h"
51 #include "llvm/Target/TargetLowering.h"
52 #include "llvm/Target/TargetOptions.h"
53 #include "llvm/Support/Compiler.h"
54 #include "llvm/Support/CommandLine.h"
55 #include "llvm/Support/Debug.h"
56 #include "llvm/Support/ErrorHandling.h"
57 #include "llvm/Support/MathExtras.h"
58 #include "llvm/Support/raw_ostream.h"
62 /// LimitFloatPrecision - Generate low-precision inline sequences for
63 /// some float libcalls (6, 8 or 12 bits).
64 static unsigned LimitFloatPrecision;
66 static cl::opt<unsigned, true>
67 LimitFPPrecision("limit-float-precision",
68 cl::desc("Generate low-precision inline sequences "
69 "for some float libcalls"),
70 cl::location(LimitFloatPrecision),
73 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, DebugLoc DL,
74 const SDValue *Parts, unsigned NumParts,
75 EVT PartVT, EVT ValueVT);
77 /// getCopyFromParts - Create a value that contains the specified legal parts
78 /// combined into the value they represent. If the parts combine to a type
79 /// larger then ValueVT then AssertOp can be used to specify whether the extra
80 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
81 /// (ISD::AssertSext).
82 static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc DL,
84 unsigned NumParts, EVT PartVT, EVT ValueVT,
85 ISD::NodeType AssertOp = ISD::DELETED_NODE) {
86 if (ValueVT.isVector())
87 return getCopyFromPartsVector(DAG, DL, Parts, NumParts, PartVT, ValueVT);
89 assert(NumParts > 0 && "No parts to assemble!");
90 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
91 SDValue Val = Parts[0];
94 // Assemble the value from multiple parts.
95 if (ValueVT.isInteger()) {
96 unsigned PartBits = PartVT.getSizeInBits();
97 unsigned ValueBits = ValueVT.getSizeInBits();
99 // Assemble the power of 2 part.
100 unsigned RoundParts = NumParts & (NumParts - 1) ?
101 1 << Log2_32(NumParts) : NumParts;
102 unsigned RoundBits = PartBits * RoundParts;
103 EVT RoundVT = RoundBits == ValueBits ?
104 ValueVT : EVT::getIntegerVT(*DAG.getContext(), RoundBits);
107 EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2);
109 if (RoundParts > 2) {
110 Lo = getCopyFromParts(DAG, DL, Parts, RoundParts / 2,
112 Hi = getCopyFromParts(DAG, DL, Parts + RoundParts / 2,
113 RoundParts / 2, PartVT, HalfVT);
115 Lo = DAG.getNode(ISD::BIT_CONVERT, DL, HalfVT, Parts[0]);
116 Hi = DAG.getNode(ISD::BIT_CONVERT, DL, HalfVT, Parts[1]);
119 if (TLI.isBigEndian())
122 Val = DAG.getNode(ISD::BUILD_PAIR, DL, RoundVT, Lo, Hi);
124 if (RoundParts < NumParts) {
125 // Assemble the trailing non-power-of-2 part.
126 unsigned OddParts = NumParts - RoundParts;
127 EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits);
128 Hi = getCopyFromParts(DAG, DL,
129 Parts + RoundParts, OddParts, PartVT, OddVT);
131 // Combine the round and odd parts.
133 if (TLI.isBigEndian())
135 EVT TotalVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
136 Hi = DAG.getNode(ISD::ANY_EXTEND, DL, TotalVT, Hi);
137 Hi = DAG.getNode(ISD::SHL, DL, TotalVT, Hi,
138 DAG.getConstant(Lo.getValueType().getSizeInBits(),
139 TLI.getPointerTy()));
140 Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, TotalVT, Lo);
141 Val = DAG.getNode(ISD::OR, DL, TotalVT, Lo, Hi);
143 } else if (PartVT.isFloatingPoint()) {
144 // FP split into multiple FP parts (for ppcf128)
145 assert(ValueVT == EVT(MVT::ppcf128) && PartVT == EVT(MVT::f64) &&
148 Lo = DAG.getNode(ISD::BIT_CONVERT, DL, EVT(MVT::f64), Parts[0]);
149 Hi = DAG.getNode(ISD::BIT_CONVERT, DL, EVT(MVT::f64), Parts[1]);
150 if (TLI.isBigEndian())
152 Val = DAG.getNode(ISD::BUILD_PAIR, DL, ValueVT, Lo, Hi);
154 // FP split into integer parts (soft fp)
155 assert(ValueVT.isFloatingPoint() && PartVT.isInteger() &&
156 !PartVT.isVector() && "Unexpected split");
157 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits());
158 Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT);
162 // There is now one part, held in Val. Correct it to match ValueVT.
163 PartVT = Val.getValueType();
165 if (PartVT == ValueVT)
168 if (PartVT.isInteger() && ValueVT.isInteger()) {
169 if (ValueVT.bitsLT(PartVT)) {
170 // For a truncate, see if we have any information to
171 // indicate whether the truncated bits will always be
172 // zero or sign-extension.
173 if (AssertOp != ISD::DELETED_NODE)
174 Val = DAG.getNode(AssertOp, DL, PartVT, Val,
175 DAG.getValueType(ValueVT));
176 return DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
178 return DAG.getNode(ISD::ANY_EXTEND, DL, ValueVT, Val);
181 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
182 // FP_ROUND's are always exact here.
183 if (ValueVT.bitsLT(Val.getValueType()))
184 return DAG.getNode(ISD::FP_ROUND, DL, ValueVT, Val,
185 DAG.getIntPtrConstant(1));
187 return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val);
190 if (PartVT.getSizeInBits() == ValueVT.getSizeInBits())
191 return DAG.getNode(ISD::BIT_CONVERT, DL, ValueVT, Val);
193 llvm_unreachable("Unknown mismatch!");
197 /// getCopyFromParts - Create a value that contains the specified legal parts
198 /// combined into the value they represent. If the parts combine to a type
199 /// larger then ValueVT then AssertOp can be used to specify whether the extra
200 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
201 /// (ISD::AssertSext).
202 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, DebugLoc DL,
203 const SDValue *Parts, unsigned NumParts,
204 EVT PartVT, EVT ValueVT) {
205 assert(ValueVT.isVector() && "Not a vector value");
206 assert(NumParts > 0 && "No parts to assemble!");
207 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
208 SDValue Val = Parts[0];
210 // Handle a multi-element vector.
212 EVT IntermediateVT, RegisterVT;
213 unsigned NumIntermediates;
215 TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT,
216 NumIntermediates, RegisterVT);
217 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
218 NumParts = NumRegs; // Silence a compiler warning.
219 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
220 assert(RegisterVT == Parts[0].getValueType() &&
221 "Part type doesn't match part!");
223 // Assemble the parts into intermediate operands.
224 SmallVector<SDValue, 8> Ops(NumIntermediates);
225 if (NumIntermediates == NumParts) {
226 // If the register was not expanded, truncate or copy the value,
228 for (unsigned i = 0; i != NumParts; ++i)
229 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i], 1,
230 PartVT, IntermediateVT);
231 } else if (NumParts > 0) {
232 // If the intermediate type was expanded, build the intermediate
233 // operands from the parts.
234 assert(NumParts % NumIntermediates == 0 &&
235 "Must expand into a divisible number of parts!");
236 unsigned Factor = NumParts / NumIntermediates;
237 for (unsigned i = 0; i != NumIntermediates; ++i)
238 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor,
239 PartVT, IntermediateVT);
242 // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the
243 // intermediate operands.
244 Val = DAG.getNode(IntermediateVT.isVector() ?
245 ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR, DL,
246 ValueVT, &Ops[0], NumIntermediates);
249 // There is now one part, held in Val. Correct it to match ValueVT.
250 PartVT = Val.getValueType();
252 if (PartVT == ValueVT)
255 if (PartVT.isVector()) {
256 // If the element type of the source/dest vectors are the same, but the
257 // parts vector has more elements than the value vector, then we have a
258 // vector widening case (e.g. <2 x float> -> <4 x float>). Extract the
260 if (PartVT.getVectorElementType() == ValueVT.getVectorElementType()) {
261 assert(PartVT.getVectorNumElements() > ValueVT.getVectorNumElements() &&
262 "Cannot narrow, it would be a lossy transformation");
263 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val,
264 DAG.getIntPtrConstant(0));
267 // Vector/Vector bitcast.
268 return DAG.getNode(ISD::BIT_CONVERT, DL, ValueVT, Val);
271 assert(ValueVT.getVectorElementType() == PartVT &&
272 ValueVT.getVectorNumElements() == 1 &&
273 "Only trivial scalar-to-vector conversions should get here!");
274 return DAG.getNode(ISD::BUILD_VECTOR, DL, ValueVT, Val);
280 static void getCopyToPartsVector(SelectionDAG &DAG, DebugLoc dl,
281 SDValue Val, SDValue *Parts, unsigned NumParts,
284 /// getCopyToParts - Create a series of nodes that contain the specified value
285 /// split into legal parts. If the parts contain more bits than Val, then, for
286 /// integers, ExtendKind can be used to specify how to generate the extra bits.
287 static void getCopyToParts(SelectionDAG &DAG, DebugLoc DL,
288 SDValue Val, SDValue *Parts, unsigned NumParts,
290 ISD::NodeType ExtendKind = ISD::ANY_EXTEND) {
291 EVT ValueVT = Val.getValueType();
293 // Handle the vector case separately.
294 if (ValueVT.isVector())
295 return getCopyToPartsVector(DAG, DL, Val, Parts, NumParts, PartVT);
297 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
298 unsigned PartBits = PartVT.getSizeInBits();
299 unsigned OrigNumParts = NumParts;
300 assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!");
305 assert(!ValueVT.isVector() && "Vector case handled elsewhere");
306 if (PartVT == ValueVT) {
307 assert(NumParts == 1 && "No-op copy with multiple parts!");
312 if (NumParts * PartBits > ValueVT.getSizeInBits()) {
313 // If the parts cover more bits than the value has, promote the value.
314 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
315 assert(NumParts == 1 && "Do not know what to promote to!");
316 Val = DAG.getNode(ISD::FP_EXTEND, DL, PartVT, Val);
318 assert(PartVT.isInteger() && ValueVT.isInteger() &&
319 "Unknown mismatch!");
320 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
321 Val = DAG.getNode(ExtendKind, DL, ValueVT, Val);
323 } else if (PartBits == ValueVT.getSizeInBits()) {
324 // Different types of the same size.
325 assert(NumParts == 1 && PartVT != ValueVT);
326 Val = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Val);
327 } else if (NumParts * PartBits < ValueVT.getSizeInBits()) {
328 // If the parts cover less bits than value has, truncate the value.
329 assert(PartVT.isInteger() && ValueVT.isInteger() &&
330 "Unknown mismatch!");
331 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
332 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
335 // The value may have changed - recompute ValueVT.
336 ValueVT = Val.getValueType();
337 assert(NumParts * PartBits == ValueVT.getSizeInBits() &&
338 "Failed to tile the value with PartVT!");
341 assert(PartVT == ValueVT && "Type conversion failed!");
346 // Expand the value into multiple parts.
347 if (NumParts & (NumParts - 1)) {
348 // The number of parts is not a power of 2. Split off and copy the tail.
349 assert(PartVT.isInteger() && ValueVT.isInteger() &&
350 "Do not know what to expand to!");
351 unsigned RoundParts = 1 << Log2_32(NumParts);
352 unsigned RoundBits = RoundParts * PartBits;
353 unsigned OddParts = NumParts - RoundParts;
354 SDValue OddVal = DAG.getNode(ISD::SRL, DL, ValueVT, Val,
355 DAG.getIntPtrConstant(RoundBits));
356 getCopyToParts(DAG, DL, OddVal, Parts + RoundParts, OddParts, PartVT);
358 if (TLI.isBigEndian())
359 // The odd parts were reversed by getCopyToParts - unreverse them.
360 std::reverse(Parts + RoundParts, Parts + NumParts);
362 NumParts = RoundParts;
363 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
364 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
367 // The number of parts is a power of 2. Repeatedly bisect the value using
369 Parts[0] = DAG.getNode(ISD::BIT_CONVERT, DL,
370 EVT::getIntegerVT(*DAG.getContext(),
371 ValueVT.getSizeInBits()),
374 for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) {
375 for (unsigned i = 0; i < NumParts; i += StepSize) {
376 unsigned ThisBits = StepSize * PartBits / 2;
377 EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits);
378 SDValue &Part0 = Parts[i];
379 SDValue &Part1 = Parts[i+StepSize/2];
381 Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
382 ThisVT, Part0, DAG.getIntPtrConstant(1));
383 Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
384 ThisVT, Part0, DAG.getIntPtrConstant(0));
386 if (ThisBits == PartBits && ThisVT != PartVT) {
387 Part0 = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Part0);
388 Part1 = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Part1);
393 if (TLI.isBigEndian())
394 std::reverse(Parts, Parts + OrigNumParts);
398 /// getCopyToPartsVector - Create a series of nodes that contain the specified
399 /// value split into legal parts.
400 static void getCopyToPartsVector(SelectionDAG &DAG, DebugLoc DL,
401 SDValue Val, SDValue *Parts, unsigned NumParts,
403 EVT ValueVT = Val.getValueType();
404 assert(ValueVT.isVector() && "Not a vector");
405 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
408 if (PartVT == ValueVT) {
410 } else if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) {
411 // Bitconvert vector->vector case.
412 Val = DAG.getNode(ISD::BIT_CONVERT, DL, PartVT, Val);
413 } else if (PartVT.isVector() &&
414 PartVT.getVectorElementType() == ValueVT.getVectorElementType()&&
415 PartVT.getVectorNumElements() > ValueVT.getVectorNumElements()) {
416 EVT ElementVT = PartVT.getVectorElementType();
417 // Vector widening case, e.g. <2 x float> -> <4 x float>. Shuffle in
419 SmallVector<SDValue, 16> Ops;
420 for (unsigned i = 0, e = ValueVT.getVectorNumElements(); i != e; ++i)
421 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
422 ElementVT, Val, DAG.getIntPtrConstant(i)));
424 for (unsigned i = ValueVT.getVectorNumElements(),
425 e = PartVT.getVectorNumElements(); i != e; ++i)
426 Ops.push_back(DAG.getUNDEF(ElementVT));
428 Val = DAG.getNode(ISD::BUILD_VECTOR, DL, PartVT, &Ops[0], Ops.size());
430 // FIXME: Use CONCAT for 2x -> 4x.
432 //SDValue UndefElts = DAG.getUNDEF(VectorTy);
433 //Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts);
435 // Vector -> scalar conversion.
436 assert(ValueVT.getVectorElementType() == PartVT &&
437 ValueVT.getVectorNumElements() == 1 &&
438 "Only trivial vector-to-scalar conversions should get here!");
439 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
440 PartVT, Val, DAG.getIntPtrConstant(0));
447 // Handle a multi-element vector.
448 EVT IntermediateVT, RegisterVT;
449 unsigned NumIntermediates;
450 unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT,
452 NumIntermediates, RegisterVT);
453 unsigned NumElements = ValueVT.getVectorNumElements();
455 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
456 NumParts = NumRegs; // Silence a compiler warning.
457 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
459 // Split the vector into intermediate operands.
460 SmallVector<SDValue, 8> Ops(NumIntermediates);
461 for (unsigned i = 0; i != NumIntermediates; ++i) {
462 if (IntermediateVT.isVector())
463 Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL,
465 DAG.getIntPtrConstant(i * (NumElements / NumIntermediates)));
467 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
468 IntermediateVT, Val, DAG.getIntPtrConstant(i));
471 // Split the intermediate operands into legal parts.
472 if (NumParts == NumIntermediates) {
473 // If the register was not expanded, promote or copy the value,
475 for (unsigned i = 0; i != NumParts; ++i)
476 getCopyToParts(DAG, DL, Ops[i], &Parts[i], 1, PartVT);
477 } else if (NumParts > 0) {
478 // If the intermediate type was expanded, split each the value into
480 assert(NumParts % NumIntermediates == 0 &&
481 "Must expand into a divisible number of parts!");
482 unsigned Factor = NumParts / NumIntermediates;
483 for (unsigned i = 0; i != NumIntermediates; ++i)
484 getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT);
492 /// RegsForValue - This struct represents the registers (physical or virtual)
493 /// that a particular set of values is assigned, and the type information
494 /// about the value. The most common situation is to represent one value at a
495 /// time, but struct or array values are handled element-wise as multiple
496 /// values. The splitting of aggregates is performed recursively, so that we
497 /// never have aggregate-typed registers. The values at this point do not
498 /// necessarily have legal types, so each value may require one or more
499 /// registers of some legal type.
501 struct RegsForValue {
502 /// ValueVTs - The value types of the values, which may not be legal, and
503 /// may need be promoted or synthesized from one or more registers.
505 SmallVector<EVT, 4> ValueVTs;
507 /// RegVTs - The value types of the registers. This is the same size as
508 /// ValueVTs and it records, for each value, what the type of the assigned
509 /// register or registers are. (Individual values are never synthesized
510 /// from more than one type of register.)
512 /// With virtual registers, the contents of RegVTs is redundant with TLI's
513 /// getRegisterType member function, however when with physical registers
514 /// it is necessary to have a separate record of the types.
516 SmallVector<EVT, 4> RegVTs;
518 /// Regs - This list holds the registers assigned to the values.
519 /// Each legal or promoted value requires one register, and each
520 /// expanded value requires multiple registers.
522 SmallVector<unsigned, 4> Regs;
526 RegsForValue(const SmallVector<unsigned, 4> ®s,
527 EVT regvt, EVT valuevt)
528 : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {}
530 RegsForValue(LLVMContext &Context, const TargetLowering &tli,
531 unsigned Reg, const Type *Ty) {
532 ComputeValueVTs(tli, Ty, ValueVTs);
534 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
535 EVT ValueVT = ValueVTs[Value];
536 unsigned NumRegs = tli.getNumRegisters(Context, ValueVT);
537 EVT RegisterVT = tli.getRegisterType(Context, ValueVT);
538 for (unsigned i = 0; i != NumRegs; ++i)
539 Regs.push_back(Reg + i);
540 RegVTs.push_back(RegisterVT);
545 /// areValueTypesLegal - Return true if types of all the values are legal.
546 bool areValueTypesLegal(const TargetLowering &TLI) {
547 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
548 EVT RegisterVT = RegVTs[Value];
549 if (!TLI.isTypeLegal(RegisterVT))
555 /// append - Add the specified values to this one.
556 void append(const RegsForValue &RHS) {
557 ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end());
558 RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end());
559 Regs.append(RHS.Regs.begin(), RHS.Regs.end());
562 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
563 /// this value and returns the result as a ValueVTs value. This uses
564 /// Chain/Flag as the input and updates them for the output Chain/Flag.
565 /// If the Flag pointer is NULL, no flag is used.
566 SDValue getCopyFromRegs(SelectionDAG &DAG, FunctionLoweringInfo &FuncInfo,
568 SDValue &Chain, SDValue *Flag) const;
570 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
571 /// specified value into the registers specified by this object. This uses
572 /// Chain/Flag as the input and updates them for the output Chain/Flag.
573 /// If the Flag pointer is NULL, no flag is used.
574 void getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
575 SDValue &Chain, SDValue *Flag) const;
577 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
578 /// operand list. This adds the code marker, matching input operand index
579 /// (if applicable), and includes the number of values added into it.
580 void AddInlineAsmOperands(unsigned Kind,
581 bool HasMatching, unsigned MatchingIdx,
583 std::vector<SDValue> &Ops) const;
587 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
588 /// this value and returns the result as a ValueVT value. This uses
589 /// Chain/Flag as the input and updates them for the output Chain/Flag.
590 /// If the Flag pointer is NULL, no flag is used.
591 SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
592 FunctionLoweringInfo &FuncInfo,
594 SDValue &Chain, SDValue *Flag) const {
595 // A Value with type {} or [0 x %t] needs no registers.
596 if (ValueVTs.empty())
599 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
601 // Assemble the legal parts into the final values.
602 SmallVector<SDValue, 4> Values(ValueVTs.size());
603 SmallVector<SDValue, 8> Parts;
604 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
605 // Copy the legal parts from the registers.
606 EVT ValueVT = ValueVTs[Value];
607 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
608 EVT RegisterVT = RegVTs[Value];
610 Parts.resize(NumRegs);
611 for (unsigned i = 0; i != NumRegs; ++i) {
614 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT);
616 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag);
617 *Flag = P.getValue(2);
620 Chain = P.getValue(1);
622 // If the source register was virtual and if we know something about it,
623 // add an assert node.
624 if (TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) &&
625 RegisterVT.isInteger() && !RegisterVT.isVector()) {
626 unsigned SlotNo = Regs[Part+i]-TargetRegisterInfo::FirstVirtualRegister;
627 if (FuncInfo.LiveOutRegInfo.size() > SlotNo) {
628 const FunctionLoweringInfo::LiveOutInfo &LOI =
629 FuncInfo.LiveOutRegInfo[SlotNo];
631 unsigned RegSize = RegisterVT.getSizeInBits();
632 unsigned NumSignBits = LOI.NumSignBits;
633 unsigned NumZeroBits = LOI.KnownZero.countLeadingOnes();
635 // FIXME: We capture more information than the dag can represent. For
636 // now, just use the tightest assertzext/assertsext possible.
638 EVT FromVT(MVT::Other);
639 if (NumSignBits == RegSize)
640 isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1
641 else if (NumZeroBits >= RegSize-1)
642 isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1
643 else if (NumSignBits > RegSize-8)
644 isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8
645 else if (NumZeroBits >= RegSize-8)
646 isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8
647 else if (NumSignBits > RegSize-16)
648 isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16
649 else if (NumZeroBits >= RegSize-16)
650 isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16
651 else if (NumSignBits > RegSize-32)
652 isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32
653 else if (NumZeroBits >= RegSize-32)
654 isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32
656 if (FromVT != MVT::Other)
657 P = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl,
658 RegisterVT, P, DAG.getValueType(FromVT));
665 Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(),
666 NumRegs, RegisterVT, ValueVT);
671 return DAG.getNode(ISD::MERGE_VALUES, dl,
672 DAG.getVTList(&ValueVTs[0], ValueVTs.size()),
673 &Values[0], ValueVTs.size());
676 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
677 /// specified value into the registers specified by this object. This uses
678 /// Chain/Flag as the input and updates them for the output Chain/Flag.
679 /// If the Flag pointer is NULL, no flag is used.
680 void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
681 SDValue &Chain, SDValue *Flag) const {
682 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
684 // Get the list of the values's legal parts.
685 unsigned NumRegs = Regs.size();
686 SmallVector<SDValue, 8> Parts(NumRegs);
687 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
688 EVT ValueVT = ValueVTs[Value];
689 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
690 EVT RegisterVT = RegVTs[Value];
692 getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value),
693 &Parts[Part], NumParts, RegisterVT);
697 // Copy the parts into the registers.
698 SmallVector<SDValue, 8> Chains(NumRegs);
699 for (unsigned i = 0; i != NumRegs; ++i) {
702 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]);
704 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag);
705 *Flag = Part.getValue(1);
708 Chains[i] = Part.getValue(0);
711 if (NumRegs == 1 || Flag)
712 // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is
713 // flagged to it. That is the CopyToReg nodes and the user are considered
714 // a single scheduling unit. If we create a TokenFactor and return it as
715 // chain, then the TokenFactor is both a predecessor (operand) of the
716 // user as well as a successor (the TF operands are flagged to the user).
717 // c1, f1 = CopyToReg
718 // c2, f2 = CopyToReg
719 // c3 = TokenFactor c1, c2
722 Chain = Chains[NumRegs-1];
724 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0], NumRegs);
727 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
728 /// operand list. This adds the code marker and includes the number of
729 /// values added into it.
730 void RegsForValue::AddInlineAsmOperands(unsigned Code, bool HasMatching,
731 unsigned MatchingIdx,
733 std::vector<SDValue> &Ops) const {
734 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
736 unsigned Flag = InlineAsm::getFlagWord(Code, Regs.size());
738 Flag = InlineAsm::getFlagWordForMatchingOp(Flag, MatchingIdx);
739 SDValue Res = DAG.getTargetConstant(Flag, MVT::i32);
742 for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) {
743 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVTs[Value]);
744 EVT RegisterVT = RegVTs[Value];
745 for (unsigned i = 0; i != NumRegs; ++i) {
746 assert(Reg < Regs.size() && "Mismatch in # registers expected");
747 Ops.push_back(DAG.getRegister(Regs[Reg++], RegisterVT));
752 void SelectionDAGBuilder::init(GCFunctionInfo *gfi, AliasAnalysis &aa) {
755 TD = DAG.getTarget().getTargetData();
758 /// clear - Clear out the current SelectionDAG and the associated
759 /// state and prepare this SelectionDAGBuilder object to be used
760 /// for a new block. This doesn't clear out information about
761 /// additional blocks that are needed to complete switch lowering
762 /// or PHI node updating; that information is cleared out as it is
764 void SelectionDAGBuilder::clear() {
766 UnusedArgNodeMap.clear();
767 PendingLoads.clear();
768 PendingExports.clear();
769 DanglingDebugInfoMap.clear();
770 CurDebugLoc = DebugLoc();
774 /// getRoot - Return the current virtual root of the Selection DAG,
775 /// flushing any PendingLoad items. This must be done before emitting
776 /// a store or any other node that may need to be ordered after any
777 /// prior load instructions.
779 SDValue SelectionDAGBuilder::getRoot() {
780 if (PendingLoads.empty())
781 return DAG.getRoot();
783 if (PendingLoads.size() == 1) {
784 SDValue Root = PendingLoads[0];
786 PendingLoads.clear();
790 // Otherwise, we have to make a token factor node.
791 SDValue Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
792 &PendingLoads[0], PendingLoads.size());
793 PendingLoads.clear();
798 /// getControlRoot - Similar to getRoot, but instead of flushing all the
799 /// PendingLoad items, flush all the PendingExports items. It is necessary
800 /// to do this before emitting a terminator instruction.
802 SDValue SelectionDAGBuilder::getControlRoot() {
803 SDValue Root = DAG.getRoot();
805 if (PendingExports.empty())
808 // Turn all of the CopyToReg chains into one factored node.
809 if (Root.getOpcode() != ISD::EntryToken) {
810 unsigned i = 0, e = PendingExports.size();
811 for (; i != e; ++i) {
812 assert(PendingExports[i].getNode()->getNumOperands() > 1);
813 if (PendingExports[i].getNode()->getOperand(0) == Root)
814 break; // Don't add the root if we already indirectly depend on it.
818 PendingExports.push_back(Root);
821 Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
823 PendingExports.size());
824 PendingExports.clear();
829 void SelectionDAGBuilder::AssignOrderingToNode(const SDNode *Node) {
830 if (DAG.GetOrdering(Node) != 0) return; // Already has ordering.
831 DAG.AssignOrdering(Node, SDNodeOrder);
833 for (unsigned I = 0, E = Node->getNumOperands(); I != E; ++I)
834 AssignOrderingToNode(Node->getOperand(I).getNode());
837 void SelectionDAGBuilder::visit(const Instruction &I) {
838 // Set up outgoing PHI node register values before emitting the terminator.
839 if (isa<TerminatorInst>(&I))
840 HandlePHINodesInSuccessorBlocks(I.getParent());
842 CurDebugLoc = I.getDebugLoc();
844 visit(I.getOpcode(), I);
846 if (!isa<TerminatorInst>(&I) && !HasTailCall)
847 CopyToExportRegsIfNeeded(&I);
849 CurDebugLoc = DebugLoc();
852 void SelectionDAGBuilder::visitPHI(const PHINode &) {
853 llvm_unreachable("SelectionDAGBuilder shouldn't visit PHI nodes!");
856 void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) {
857 // Note: this doesn't use InstVisitor, because it has to work with
858 // ConstantExpr's in addition to instructions.
860 default: llvm_unreachable("Unknown instruction type encountered!");
861 // Build the switch statement using the Instruction.def file.
862 #define HANDLE_INST(NUM, OPCODE, CLASS) \
863 case Instruction::OPCODE: visit##OPCODE((CLASS&)I); break;
864 #include "llvm/Instruction.def"
867 // Assign the ordering to the freshly created DAG nodes.
868 if (NodeMap.count(&I)) {
870 AssignOrderingToNode(getValue(&I).getNode());
874 // resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V,
875 // generate the debug data structures now that we've seen its definition.
876 void SelectionDAGBuilder::resolveDanglingDebugInfo(const Value *V,
878 DanglingDebugInfo &DDI = DanglingDebugInfoMap[V];
880 const DbgValueInst *DI = DDI.getDI();
881 DebugLoc dl = DDI.getdl();
882 unsigned DbgSDNodeOrder = DDI.getSDNodeOrder();
883 MDNode *Variable = DI->getVariable();
884 uint64_t Offset = DI->getOffset();
887 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, Val)) {
888 SDV = DAG.getDbgValue(Variable, Val.getNode(),
889 Val.getResNo(), Offset, dl, DbgSDNodeOrder);
890 DAG.AddDbgValue(SDV, Val.getNode(), false);
893 SDV = DAG.getDbgValue(Variable, UndefValue::get(V->getType()),
894 Offset, dl, SDNodeOrder);
895 DAG.AddDbgValue(SDV, 0, false);
897 DanglingDebugInfoMap[V] = DanglingDebugInfo();
901 // getValue - Return an SDValue for the given Value.
902 SDValue SelectionDAGBuilder::getValue(const Value *V) {
903 // If we already have an SDValue for this value, use it. It's important
904 // to do this first, so that we don't create a CopyFromReg if we already
905 // have a regular SDValue.
906 SDValue &N = NodeMap[V];
907 if (N.getNode()) return N;
909 // If there's a virtual register allocated and initialized for this
911 DenseMap<const Value *, unsigned>::iterator It = FuncInfo.ValueMap.find(V);
912 if (It != FuncInfo.ValueMap.end()) {
913 unsigned InReg = It->second;
914 RegsForValue RFV(*DAG.getContext(), TLI, InReg, V->getType());
915 SDValue Chain = DAG.getEntryNode();
916 return N = RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain,NULL);
919 // Otherwise create a new SDValue and remember it.
920 SDValue Val = getValueImpl(V);
922 resolveDanglingDebugInfo(V, Val);
926 /// getNonRegisterValue - Return an SDValue for the given Value, but
927 /// don't look in FuncInfo.ValueMap for a virtual register.
928 SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) {
929 // If we already have an SDValue for this value, use it.
930 SDValue &N = NodeMap[V];
931 if (N.getNode()) return N;
933 // Otherwise create a new SDValue and remember it.
934 SDValue Val = getValueImpl(V);
936 resolveDanglingDebugInfo(V, Val);
940 /// getValueImpl - Helper function for getValue and getNonRegisterValue.
941 /// Create an SDValue for the given value.
942 SDValue SelectionDAGBuilder::getValueImpl(const Value *V) {
943 if (const Constant *C = dyn_cast<Constant>(V)) {
944 EVT VT = TLI.getValueType(V->getType(), true);
946 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C))
947 return DAG.getConstant(*CI, VT);
949 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
950 return DAG.getGlobalAddress(GV, getCurDebugLoc(), VT);
952 if (isa<ConstantPointerNull>(C))
953 return DAG.getConstant(0, TLI.getPointerTy());
955 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
956 return DAG.getConstantFP(*CFP, VT);
958 if (isa<UndefValue>(C) && !V->getType()->isAggregateType())
959 return DAG.getUNDEF(VT);
961 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
962 visit(CE->getOpcode(), *CE);
963 SDValue N1 = NodeMap[V];
964 assert(N1.getNode() && "visit didn't populate the NodeMap!");
968 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) {
969 SmallVector<SDValue, 4> Constants;
970 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end();
972 SDNode *Val = getValue(*OI).getNode();
973 // If the operand is an empty aggregate, there are no values.
975 // Add each leaf value from the operand to the Constants list
976 // to form a flattened list of all the values.
977 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
978 Constants.push_back(SDValue(Val, i));
981 return DAG.getMergeValues(&Constants[0], Constants.size(),
985 if (C->getType()->isStructTy() || C->getType()->isArrayTy()) {
986 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
987 "Unknown struct or array constant!");
989 SmallVector<EVT, 4> ValueVTs;
990 ComputeValueVTs(TLI, C->getType(), ValueVTs);
991 unsigned NumElts = ValueVTs.size();
993 return SDValue(); // empty struct
994 SmallVector<SDValue, 4> Constants(NumElts);
995 for (unsigned i = 0; i != NumElts; ++i) {
996 EVT EltVT = ValueVTs[i];
997 if (isa<UndefValue>(C))
998 Constants[i] = DAG.getUNDEF(EltVT);
999 else if (EltVT.isFloatingPoint())
1000 Constants[i] = DAG.getConstantFP(0, EltVT);
1002 Constants[i] = DAG.getConstant(0, EltVT);
1005 return DAG.getMergeValues(&Constants[0], NumElts,
1009 if (const BlockAddress *BA = dyn_cast<BlockAddress>(C))
1010 return DAG.getBlockAddress(BA, VT);
1012 const VectorType *VecTy = cast<VectorType>(V->getType());
1013 unsigned NumElements = VecTy->getNumElements();
1015 // Now that we know the number and type of the elements, get that number of
1016 // elements into the Ops array based on what kind of constant it is.
1017 SmallVector<SDValue, 16> Ops;
1018 if (const ConstantVector *CP = dyn_cast<ConstantVector>(C)) {
1019 for (unsigned i = 0; i != NumElements; ++i)
1020 Ops.push_back(getValue(CP->getOperand(i)));
1022 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!");
1023 EVT EltVT = TLI.getValueType(VecTy->getElementType());
1026 if (EltVT.isFloatingPoint())
1027 Op = DAG.getConstantFP(0, EltVT);
1029 Op = DAG.getConstant(0, EltVT);
1030 Ops.assign(NumElements, Op);
1033 // Create a BUILD_VECTOR node.
1034 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
1035 VT, &Ops[0], Ops.size());
1038 // If this is a static alloca, generate it as the frameindex instead of
1040 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
1041 DenseMap<const AllocaInst*, int>::iterator SI =
1042 FuncInfo.StaticAllocaMap.find(AI);
1043 if (SI != FuncInfo.StaticAllocaMap.end())
1044 return DAG.getFrameIndex(SI->second, TLI.getPointerTy());
1047 // If this is an instruction which fast-isel has deferred, select it now.
1048 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
1049 unsigned InReg = FuncInfo.InitializeRegForValue(Inst);
1050 RegsForValue RFV(*DAG.getContext(), TLI, InReg, Inst->getType());
1051 SDValue Chain = DAG.getEntryNode();
1052 return RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain, NULL);
1055 llvm_unreachable("Can't get register for value!");
1059 void SelectionDAGBuilder::visitRet(const ReturnInst &I) {
1060 SDValue Chain = getControlRoot();
1061 SmallVector<ISD::OutputArg, 8> Outs;
1062 SmallVector<SDValue, 8> OutVals;
1064 if (!FuncInfo.CanLowerReturn) {
1065 unsigned DemoteReg = FuncInfo.DemoteRegister;
1066 const Function *F = I.getParent()->getParent();
1068 // Emit a store of the return value through the virtual register.
1069 // Leave Outs empty so that LowerReturn won't try to load return
1070 // registers the usual way.
1071 SmallVector<EVT, 1> PtrValueVTs;
1072 ComputeValueVTs(TLI, PointerType::getUnqual(F->getReturnType()),
1075 SDValue RetPtr = DAG.getRegister(DemoteReg, PtrValueVTs[0]);
1076 SDValue RetOp = getValue(I.getOperand(0));
1078 SmallVector<EVT, 4> ValueVTs;
1079 SmallVector<uint64_t, 4> Offsets;
1080 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs, &Offsets);
1081 unsigned NumValues = ValueVTs.size();
1083 SmallVector<SDValue, 4> Chains(NumValues);
1084 for (unsigned i = 0; i != NumValues; ++i) {
1085 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(),
1086 RetPtr.getValueType(), RetPtr,
1087 DAG.getIntPtrConstant(Offsets[i]));
1089 DAG.getStore(Chain, getCurDebugLoc(),
1090 SDValue(RetOp.getNode(), RetOp.getResNo() + i),
1091 // FIXME: better loc info would be nice.
1092 Add, MachinePointerInfo(), false, false, 0);
1095 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
1096 MVT::Other, &Chains[0], NumValues);
1097 } else if (I.getNumOperands() != 0) {
1098 SmallVector<EVT, 4> ValueVTs;
1099 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs);
1100 unsigned NumValues = ValueVTs.size();
1102 SDValue RetOp = getValue(I.getOperand(0));
1103 for (unsigned j = 0, f = NumValues; j != f; ++j) {
1104 EVT VT = ValueVTs[j];
1106 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1108 const Function *F = I.getParent()->getParent();
1109 if (F->paramHasAttr(0, Attribute::SExt))
1110 ExtendKind = ISD::SIGN_EXTEND;
1111 else if (F->paramHasAttr(0, Attribute::ZExt))
1112 ExtendKind = ISD::ZERO_EXTEND;
1114 // FIXME: C calling convention requires the return type to be promoted
1115 // to at least 32-bit. But this is not necessary for non-C calling
1116 // conventions. The frontend should mark functions whose return values
1117 // require promoting with signext or zeroext attributes.
1118 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
1119 EVT MinVT = TLI.getRegisterType(*DAG.getContext(), MVT::i32);
1120 if (VT.bitsLT(MinVT))
1124 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), VT);
1125 EVT PartVT = TLI.getRegisterType(*DAG.getContext(), VT);
1126 SmallVector<SDValue, 4> Parts(NumParts);
1127 getCopyToParts(DAG, getCurDebugLoc(),
1128 SDValue(RetOp.getNode(), RetOp.getResNo() + j),
1129 &Parts[0], NumParts, PartVT, ExtendKind);
1131 // 'inreg' on function refers to return value
1132 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1133 if (F->paramHasAttr(0, Attribute::InReg))
1136 // Propagate extension type if any
1137 if (F->paramHasAttr(0, Attribute::SExt))
1139 else if (F->paramHasAttr(0, Attribute::ZExt))
1142 for (unsigned i = 0; i < NumParts; ++i) {
1143 Outs.push_back(ISD::OutputArg(Flags, Parts[i].getValueType(),
1145 OutVals.push_back(Parts[i]);
1151 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1152 CallingConv::ID CallConv =
1153 DAG.getMachineFunction().getFunction()->getCallingConv();
1154 Chain = TLI.LowerReturn(Chain, CallConv, isVarArg,
1155 Outs, OutVals, getCurDebugLoc(), DAG);
1157 // Verify that the target's LowerReturn behaved as expected.
1158 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
1159 "LowerReturn didn't return a valid chain!");
1161 // Update the DAG with the new chain value resulting from return lowering.
1165 /// CopyToExportRegsIfNeeded - If the given value has virtual registers
1166 /// created for it, emit nodes to copy the value into the virtual
1168 void SelectionDAGBuilder::CopyToExportRegsIfNeeded(const Value *V) {
1169 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
1170 if (VMI != FuncInfo.ValueMap.end()) {
1171 assert(!V->use_empty() && "Unused value assigned virtual registers!");
1172 CopyValueToVirtualRegister(V, VMI->second);
1176 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
1177 /// the current basic block, add it to ValueMap now so that we'll get a
1179 void SelectionDAGBuilder::ExportFromCurrentBlock(const Value *V) {
1180 // No need to export constants.
1181 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
1183 // Already exported?
1184 if (FuncInfo.isExportedInst(V)) return;
1186 unsigned Reg = FuncInfo.InitializeRegForValue(V);
1187 CopyValueToVirtualRegister(V, Reg);
1190 bool SelectionDAGBuilder::isExportableFromCurrentBlock(const Value *V,
1191 const BasicBlock *FromBB) {
1192 // The operands of the setcc have to be in this block. We don't know
1193 // how to export them from some other block.
1194 if (const Instruction *VI = dyn_cast<Instruction>(V)) {
1195 // Can export from current BB.
1196 if (VI->getParent() == FromBB)
1199 // Is already exported, noop.
1200 return FuncInfo.isExportedInst(V);
1203 // If this is an argument, we can export it if the BB is the entry block or
1204 // if it is already exported.
1205 if (isa<Argument>(V)) {
1206 if (FromBB == &FromBB->getParent()->getEntryBlock())
1209 // Otherwise, can only export this if it is already exported.
1210 return FuncInfo.isExportedInst(V);
1213 // Otherwise, constants can always be exported.
1217 static bool InBlock(const Value *V, const BasicBlock *BB) {
1218 if (const Instruction *I = dyn_cast<Instruction>(V))
1219 return I->getParent() == BB;
1223 /// EmitBranchForMergedCondition - Helper method for FindMergedConditions.
1224 /// This function emits a branch and is used at the leaves of an OR or an
1225 /// AND operator tree.
1228 SelectionDAGBuilder::EmitBranchForMergedCondition(const Value *Cond,
1229 MachineBasicBlock *TBB,
1230 MachineBasicBlock *FBB,
1231 MachineBasicBlock *CurBB,
1232 MachineBasicBlock *SwitchBB) {
1233 const BasicBlock *BB = CurBB->getBasicBlock();
1235 // If the leaf of the tree is a comparison, merge the condition into
1237 if (const CmpInst *BOp = dyn_cast<CmpInst>(Cond)) {
1238 // The operands of the cmp have to be in this block. We don't know
1239 // how to export them from some other block. If this is the first block
1240 // of the sequence, no exporting is needed.
1241 if (CurBB == SwitchBB ||
1242 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
1243 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) {
1244 ISD::CondCode Condition;
1245 if (const ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
1246 Condition = getICmpCondCode(IC->getPredicate());
1247 } else if (const FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) {
1248 Condition = getFCmpCondCode(FC->getPredicate());
1250 Condition = ISD::SETEQ; // silence warning.
1251 llvm_unreachable("Unknown compare instruction");
1254 CaseBlock CB(Condition, BOp->getOperand(0),
1255 BOp->getOperand(1), NULL, TBB, FBB, CurBB);
1256 SwitchCases.push_back(CB);
1261 // Create a CaseBlock record representing this branch.
1262 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()),
1263 NULL, TBB, FBB, CurBB);
1264 SwitchCases.push_back(CB);
1267 /// FindMergedConditions - If Cond is an expression like
1268 void SelectionDAGBuilder::FindMergedConditions(const Value *Cond,
1269 MachineBasicBlock *TBB,
1270 MachineBasicBlock *FBB,
1271 MachineBasicBlock *CurBB,
1272 MachineBasicBlock *SwitchBB,
1274 // If this node is not part of the or/and tree, emit it as a branch.
1275 const Instruction *BOp = dyn_cast<Instruction>(Cond);
1276 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
1277 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
1278 BOp->getParent() != CurBB->getBasicBlock() ||
1279 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
1280 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
1281 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB);
1285 // Create TmpBB after CurBB.
1286 MachineFunction::iterator BBI = CurBB;
1287 MachineFunction &MF = DAG.getMachineFunction();
1288 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock());
1289 CurBB->getParent()->insert(++BBI, TmpBB);
1291 if (Opc == Instruction::Or) {
1292 // Codegen X | Y as:
1300 // Emit the LHS condition.
1301 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, SwitchBB, Opc);
1303 // Emit the RHS condition into TmpBB.
1304 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
1306 assert(Opc == Instruction::And && "Unknown merge op!");
1307 // Codegen X & Y as:
1314 // This requires creation of TmpBB after CurBB.
1316 // Emit the LHS condition.
1317 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, SwitchBB, Opc);
1319 // Emit the RHS condition into TmpBB.
1320 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
1324 /// If the set of cases should be emitted as a series of branches, return true.
1325 /// If we should emit this as a bunch of and/or'd together conditions, return
1328 SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases){
1329 if (Cases.size() != 2) return true;
1331 // If this is two comparisons of the same values or'd or and'd together, they
1332 // will get folded into a single comparison, so don't emit two blocks.
1333 if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
1334 Cases[0].CmpRHS == Cases[1].CmpRHS) ||
1335 (Cases[0].CmpRHS == Cases[1].CmpLHS &&
1336 Cases[0].CmpLHS == Cases[1].CmpRHS)) {
1340 // Handle: (X != null) | (Y != null) --> (X|Y) != 0
1341 // Handle: (X == null) & (Y == null) --> (X|Y) == 0
1342 if (Cases[0].CmpRHS == Cases[1].CmpRHS &&
1343 Cases[0].CC == Cases[1].CC &&
1344 isa<Constant>(Cases[0].CmpRHS) &&
1345 cast<Constant>(Cases[0].CmpRHS)->isNullValue()) {
1346 if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB)
1348 if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB)
1355 void SelectionDAGBuilder::visitBr(const BranchInst &I) {
1356 MachineBasicBlock *BrMBB = FuncInfo.MBB;
1358 // Update machine-CFG edges.
1359 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
1361 // Figure out which block is immediately after the current one.
1362 MachineBasicBlock *NextBlock = 0;
1363 MachineFunction::iterator BBI = BrMBB;
1364 if (++BBI != FuncInfo.MF->end())
1367 if (I.isUnconditional()) {
1368 // Update machine-CFG edges.
1369 BrMBB->addSuccessor(Succ0MBB);
1371 // If this is not a fall-through branch, emit the branch.
1372 if (Succ0MBB != NextBlock)
1373 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
1374 MVT::Other, getControlRoot(),
1375 DAG.getBasicBlock(Succ0MBB)));
1380 // If this condition is one of the special cases we handle, do special stuff
1382 const Value *CondVal = I.getCondition();
1383 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
1385 // If this is a series of conditions that are or'd or and'd together, emit
1386 // this as a sequence of branches instead of setcc's with and/or operations.
1387 // For example, instead of something like:
1400 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
1401 if (BOp->hasOneUse() &&
1402 (BOp->getOpcode() == Instruction::And ||
1403 BOp->getOpcode() == Instruction::Or)) {
1404 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB,
1406 // If the compares in later blocks need to use values not currently
1407 // exported from this block, export them now. This block should always
1408 // be the first entry.
1409 assert(SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!");
1411 // Allow some cases to be rejected.
1412 if (ShouldEmitAsBranches(SwitchCases)) {
1413 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1414 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1415 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1418 // Emit the branch for this block.
1419 visitSwitchCase(SwitchCases[0], BrMBB);
1420 SwitchCases.erase(SwitchCases.begin());
1424 // Okay, we decided not to do this, remove any inserted MBB's and clear
1426 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
1427 FuncInfo.MF->erase(SwitchCases[i].ThisBB);
1429 SwitchCases.clear();
1433 // Create a CaseBlock record representing this branch.
1434 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()),
1435 NULL, Succ0MBB, Succ1MBB, BrMBB);
1437 // Use visitSwitchCase to actually insert the fast branch sequence for this
1439 visitSwitchCase(CB, BrMBB);
1442 /// visitSwitchCase - Emits the necessary code to represent a single node in
1443 /// the binary search tree resulting from lowering a switch instruction.
1444 void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB,
1445 MachineBasicBlock *SwitchBB) {
1447 SDValue CondLHS = getValue(CB.CmpLHS);
1448 DebugLoc dl = getCurDebugLoc();
1450 // Build the setcc now.
1451 if (CB.CmpMHS == NULL) {
1452 // Fold "(X == true)" to X and "(X == false)" to !X to
1453 // handle common cases produced by branch lowering.
1454 if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) &&
1455 CB.CC == ISD::SETEQ)
1457 else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) &&
1458 CB.CC == ISD::SETEQ) {
1459 SDValue True = DAG.getConstant(1, CondLHS.getValueType());
1460 Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True);
1462 Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1464 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now");
1466 const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue();
1467 const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue();
1469 SDValue CmpOp = getValue(CB.CmpMHS);
1470 EVT VT = CmpOp.getValueType();
1472 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) {
1473 Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, VT),
1476 SDValue SUB = DAG.getNode(ISD::SUB, dl,
1477 VT, CmpOp, DAG.getConstant(Low, VT));
1478 Cond = DAG.getSetCC(dl, MVT::i1, SUB,
1479 DAG.getConstant(High-Low, VT), ISD::SETULE);
1483 // Update successor info
1484 SwitchBB->addSuccessor(CB.TrueBB);
1485 SwitchBB->addSuccessor(CB.FalseBB);
1487 // Set NextBlock to be the MBB immediately after the current one, if any.
1488 // This is used to avoid emitting unnecessary branches to the next block.
1489 MachineBasicBlock *NextBlock = 0;
1490 MachineFunction::iterator BBI = SwitchBB;
1491 if (++BBI != FuncInfo.MF->end())
1494 // If the lhs block is the next block, invert the condition so that we can
1495 // fall through to the lhs instead of the rhs block.
1496 if (CB.TrueBB == NextBlock) {
1497 std::swap(CB.TrueBB, CB.FalseBB);
1498 SDValue True = DAG.getConstant(1, Cond.getValueType());
1499 Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True);
1502 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1503 MVT::Other, getControlRoot(), Cond,
1504 DAG.getBasicBlock(CB.TrueBB));
1506 // Insert the false branch. Do this even if it's a fall through branch,
1507 // this makes it easier to do DAG optimizations which require inverting
1508 // the branch condition.
1509 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1510 DAG.getBasicBlock(CB.FalseBB));
1512 DAG.setRoot(BrCond);
1515 /// visitJumpTable - Emit JumpTable node in the current MBB
1516 void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) {
1517 // Emit the code for the jump table
1518 assert(JT.Reg != -1U && "Should lower JT Header first!");
1519 EVT PTy = TLI.getPointerTy();
1520 SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(),
1522 SDValue Table = DAG.getJumpTable(JT.JTI, PTy);
1523 SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurDebugLoc(),
1524 MVT::Other, Index.getValue(1),
1526 DAG.setRoot(BrJumpTable);
1529 /// visitJumpTableHeader - This function emits necessary code to produce index
1530 /// in the JumpTable from switch case.
1531 void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT,
1532 JumpTableHeader &JTH,
1533 MachineBasicBlock *SwitchBB) {
1534 // Subtract the lowest switch case value from the value being switched on and
1535 // conditional branch to default mbb if the result is greater than the
1536 // difference between smallest and largest cases.
1537 SDValue SwitchOp = getValue(JTH.SValue);
1538 EVT VT = SwitchOp.getValueType();
1539 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1540 DAG.getConstant(JTH.First, VT));
1542 // The SDNode we just created, which holds the value being switched on minus
1543 // the smallest case value, needs to be copied to a virtual register so it
1544 // can be used as an index into the jump table in a subsequent basic block.
1545 // This value may be smaller or larger than the target's pointer type, and
1546 // therefore require extension or truncating.
1547 SwitchOp = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(), TLI.getPointerTy());
1549 unsigned JumpTableReg = FuncInfo.CreateReg(TLI.getPointerTy());
1550 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1551 JumpTableReg, SwitchOp);
1552 JT.Reg = JumpTableReg;
1554 // Emit the range check for the jump table, and branch to the default block
1555 // for the switch statement if the value being switched on exceeds the largest
1556 // case in the switch.
1557 SDValue CMP = DAG.getSetCC(getCurDebugLoc(),
1558 TLI.getSetCCResultType(Sub.getValueType()), Sub,
1559 DAG.getConstant(JTH.Last-JTH.First,VT),
1562 // Set NextBlock to be the MBB immediately after the current one, if any.
1563 // This is used to avoid emitting unnecessary branches to the next block.
1564 MachineBasicBlock *NextBlock = 0;
1565 MachineFunction::iterator BBI = SwitchBB;
1567 if (++BBI != FuncInfo.MF->end())
1570 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1571 MVT::Other, CopyTo, CMP,
1572 DAG.getBasicBlock(JT.Default));
1574 if (JT.MBB != NextBlock)
1575 BrCond = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrCond,
1576 DAG.getBasicBlock(JT.MBB));
1578 DAG.setRoot(BrCond);
1581 /// visitBitTestHeader - This function emits necessary code to produce value
1582 /// suitable for "bit tests"
1583 void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B,
1584 MachineBasicBlock *SwitchBB) {
1585 // Subtract the minimum value
1586 SDValue SwitchOp = getValue(B.SValue);
1587 EVT VT = SwitchOp.getValueType();
1588 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1589 DAG.getConstant(B.First, VT));
1592 SDValue RangeCmp = DAG.getSetCC(getCurDebugLoc(),
1593 TLI.getSetCCResultType(Sub.getValueType()),
1594 Sub, DAG.getConstant(B.Range, VT),
1597 SDValue ShiftOp = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(),
1598 TLI.getPointerTy());
1600 B.Reg = FuncInfo.CreateReg(TLI.getPointerTy());
1601 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1604 // Set NextBlock to be the MBB immediately after the current one, if any.
1605 // This is used to avoid emitting unnecessary branches to the next block.
1606 MachineBasicBlock *NextBlock = 0;
1607 MachineFunction::iterator BBI = SwitchBB;
1608 if (++BBI != FuncInfo.MF->end())
1611 MachineBasicBlock* MBB = B.Cases[0].ThisBB;
1613 SwitchBB->addSuccessor(B.Default);
1614 SwitchBB->addSuccessor(MBB);
1616 SDValue BrRange = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1617 MVT::Other, CopyTo, RangeCmp,
1618 DAG.getBasicBlock(B.Default));
1620 if (MBB != NextBlock)
1621 BrRange = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, CopyTo,
1622 DAG.getBasicBlock(MBB));
1624 DAG.setRoot(BrRange);
1627 /// visitBitTestCase - this function produces one "bit test"
1628 void SelectionDAGBuilder::visitBitTestCase(MachineBasicBlock* NextMBB,
1631 MachineBasicBlock *SwitchBB) {
1632 SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(), Reg,
1633 TLI.getPointerTy());
1635 if (CountPopulation_64(B.Mask) == 1) {
1636 // Testing for a single bit; just compare the shift count with what it
1637 // would need to be to shift a 1 bit in that position.
1638 Cmp = DAG.getSetCC(getCurDebugLoc(),
1639 TLI.getSetCCResultType(ShiftOp.getValueType()),
1641 DAG.getConstant(CountTrailingZeros_64(B.Mask),
1642 TLI.getPointerTy()),
1645 // Make desired shift
1646 SDValue SwitchVal = DAG.getNode(ISD::SHL, getCurDebugLoc(),
1648 DAG.getConstant(1, TLI.getPointerTy()),
1651 // Emit bit tests and jumps
1652 SDValue AndOp = DAG.getNode(ISD::AND, getCurDebugLoc(),
1653 TLI.getPointerTy(), SwitchVal,
1654 DAG.getConstant(B.Mask, TLI.getPointerTy()));
1655 Cmp = DAG.getSetCC(getCurDebugLoc(),
1656 TLI.getSetCCResultType(AndOp.getValueType()),
1657 AndOp, DAG.getConstant(0, TLI.getPointerTy()),
1661 SwitchBB->addSuccessor(B.TargetBB);
1662 SwitchBB->addSuccessor(NextMBB);
1664 SDValue BrAnd = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1665 MVT::Other, getControlRoot(),
1666 Cmp, DAG.getBasicBlock(B.TargetBB));
1668 // Set NextBlock to be the MBB immediately after the current one, if any.
1669 // This is used to avoid emitting unnecessary branches to the next block.
1670 MachineBasicBlock *NextBlock = 0;
1671 MachineFunction::iterator BBI = SwitchBB;
1672 if (++BBI != FuncInfo.MF->end())
1675 if (NextMBB != NextBlock)
1676 BrAnd = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrAnd,
1677 DAG.getBasicBlock(NextMBB));
1682 void SelectionDAGBuilder::visitInvoke(const InvokeInst &I) {
1683 MachineBasicBlock *InvokeMBB = FuncInfo.MBB;
1685 // Retrieve successors.
1686 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
1687 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)];
1689 const Value *Callee(I.getCalledValue());
1690 if (isa<InlineAsm>(Callee))
1693 LowerCallTo(&I, getValue(Callee), false, LandingPad);
1695 // If the value of the invoke is used outside of its defining block, make it
1696 // available as a virtual register.
1697 CopyToExportRegsIfNeeded(&I);
1699 // Update successor info
1700 InvokeMBB->addSuccessor(Return);
1701 InvokeMBB->addSuccessor(LandingPad);
1703 // Drop into normal successor.
1704 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
1705 MVT::Other, getControlRoot(),
1706 DAG.getBasicBlock(Return)));
1709 void SelectionDAGBuilder::visitUnwind(const UnwindInst &I) {
1712 /// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for
1713 /// small case ranges).
1714 bool SelectionDAGBuilder::handleSmallSwitchRange(CaseRec& CR,
1715 CaseRecVector& WorkList,
1717 MachineBasicBlock *Default,
1718 MachineBasicBlock *SwitchBB) {
1719 Case& BackCase = *(CR.Range.second-1);
1721 // Size is the number of Cases represented by this range.
1722 size_t Size = CR.Range.second - CR.Range.first;
1726 // Get the MachineFunction which holds the current MBB. This is used when
1727 // inserting any additional MBBs necessary to represent the switch.
1728 MachineFunction *CurMF = FuncInfo.MF;
1730 // Figure out which block is immediately after the current one.
1731 MachineBasicBlock *NextBlock = 0;
1732 MachineFunction::iterator BBI = CR.CaseBB;
1734 if (++BBI != FuncInfo.MF->end())
1737 // TODO: If any two of the cases has the same destination, and if one value
1738 // is the same as the other, but has one bit unset that the other has set,
1739 // use bit manipulation to do two compares at once. For example:
1740 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
1742 // Rearrange the case blocks so that the last one falls through if possible.
1743 if (NextBlock && Default != NextBlock && BackCase.BB != NextBlock) {
1744 // The last case block won't fall through into 'NextBlock' if we emit the
1745 // branches in this order. See if rearranging a case value would help.
1746 for (CaseItr I = CR.Range.first, E = CR.Range.second-1; I != E; ++I) {
1747 if (I->BB == NextBlock) {
1748 std::swap(*I, BackCase);
1754 // Create a CaseBlock record representing a conditional branch to
1755 // the Case's target mbb if the value being switched on SV is equal
1757 MachineBasicBlock *CurBlock = CR.CaseBB;
1758 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
1759 MachineBasicBlock *FallThrough;
1761 FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock());
1762 CurMF->insert(BBI, FallThrough);
1764 // Put SV in a virtual register to make it available from the new blocks.
1765 ExportFromCurrentBlock(SV);
1767 // If the last case doesn't match, go to the default block.
1768 FallThrough = Default;
1771 const Value *RHS, *LHS, *MHS;
1773 if (I->High == I->Low) {
1774 // This is just small small case range :) containing exactly 1 case
1776 LHS = SV; RHS = I->High; MHS = NULL;
1779 LHS = I->Low; MHS = SV; RHS = I->High;
1781 CaseBlock CB(CC, LHS, RHS, MHS, I->BB, FallThrough, CurBlock);
1783 // If emitting the first comparison, just call visitSwitchCase to emit the
1784 // code into the current block. Otherwise, push the CaseBlock onto the
1785 // vector to be later processed by SDISel, and insert the node's MBB
1786 // before the next MBB.
1787 if (CurBlock == SwitchBB)
1788 visitSwitchCase(CB, SwitchBB);
1790 SwitchCases.push_back(CB);
1792 CurBlock = FallThrough;
1798 static inline bool areJTsAllowed(const TargetLowering &TLI) {
1799 return !DisableJumpTables &&
1800 (TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
1801 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other));
1804 static APInt ComputeRange(const APInt &First, const APInt &Last) {
1805 APInt LastExt(Last), FirstExt(First);
1806 uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1;
1807 LastExt.sext(BitWidth); FirstExt.sext(BitWidth);
1808 return (LastExt - FirstExt + 1ULL);
1811 /// handleJTSwitchCase - Emit jumptable for current switch case range
1812 bool SelectionDAGBuilder::handleJTSwitchCase(CaseRec& CR,
1813 CaseRecVector& WorkList,
1815 MachineBasicBlock* Default,
1816 MachineBasicBlock *SwitchBB) {
1817 Case& FrontCase = *CR.Range.first;
1818 Case& BackCase = *(CR.Range.second-1);
1820 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
1821 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
1823 APInt TSize(First.getBitWidth(), 0);
1824 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1828 if (!areJTsAllowed(TLI) || TSize.ult(4))
1831 APInt Range = ComputeRange(First, Last);
1832 double Density = TSize.roundToDouble() / Range.roundToDouble();
1836 DEBUG(dbgs() << "Lowering jump table\n"
1837 << "First entry: " << First << ". Last entry: " << Last << '\n'
1838 << "Range: " << Range
1839 << "Size: " << TSize << ". Density: " << Density << "\n\n");
1841 // Get the MachineFunction which holds the current MBB. This is used when
1842 // inserting any additional MBBs necessary to represent the switch.
1843 MachineFunction *CurMF = FuncInfo.MF;
1845 // Figure out which block is immediately after the current one.
1846 MachineFunction::iterator BBI = CR.CaseBB;
1849 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1851 // Create a new basic block to hold the code for loading the address
1852 // of the jump table, and jumping to it. Update successor information;
1853 // we will either branch to the default case for the switch, or the jump
1855 MachineBasicBlock *JumpTableBB = CurMF->CreateMachineBasicBlock(LLVMBB);
1856 CurMF->insert(BBI, JumpTableBB);
1857 CR.CaseBB->addSuccessor(Default);
1858 CR.CaseBB->addSuccessor(JumpTableBB);
1860 // Build a vector of destination BBs, corresponding to each target
1861 // of the jump table. If the value of the jump table slot corresponds to
1862 // a case statement, push the case's BB onto the vector, otherwise, push
1864 std::vector<MachineBasicBlock*> DestBBs;
1866 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) {
1867 const APInt &Low = cast<ConstantInt>(I->Low)->getValue();
1868 const APInt &High = cast<ConstantInt>(I->High)->getValue();
1870 if (Low.sle(TEI) && TEI.sle(High)) {
1871 DestBBs.push_back(I->BB);
1875 DestBBs.push_back(Default);
1879 // Update successor info. Add one edge to each unique successor.
1880 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs());
1881 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(),
1882 E = DestBBs.end(); I != E; ++I) {
1883 if (!SuccsHandled[(*I)->getNumber()]) {
1884 SuccsHandled[(*I)->getNumber()] = true;
1885 JumpTableBB->addSuccessor(*I);
1889 // Create a jump table index for this jump table.
1890 unsigned JTEncoding = TLI.getJumpTableEncoding();
1891 unsigned JTI = CurMF->getOrCreateJumpTableInfo(JTEncoding)
1892 ->createJumpTableIndex(DestBBs);
1894 // Set the jump table information so that we can codegen it as a second
1895 // MachineBasicBlock
1896 JumpTable JT(-1U, JTI, JumpTableBB, Default);
1897 JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == SwitchBB));
1898 if (CR.CaseBB == SwitchBB)
1899 visitJumpTableHeader(JT, JTH, SwitchBB);
1901 JTCases.push_back(JumpTableBlock(JTH, JT));
1906 /// handleBTSplitSwitchCase - emit comparison and split binary search tree into
1908 bool SelectionDAGBuilder::handleBTSplitSwitchCase(CaseRec& CR,
1909 CaseRecVector& WorkList,
1911 MachineBasicBlock *Default,
1912 MachineBasicBlock *SwitchBB) {
1913 // Get the MachineFunction which holds the current MBB. This is used when
1914 // inserting any additional MBBs necessary to represent the switch.
1915 MachineFunction *CurMF = FuncInfo.MF;
1917 // Figure out which block is immediately after the current one.
1918 MachineFunction::iterator BBI = CR.CaseBB;
1921 Case& FrontCase = *CR.Range.first;
1922 Case& BackCase = *(CR.Range.second-1);
1923 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1925 // Size is the number of Cases represented by this range.
1926 unsigned Size = CR.Range.second - CR.Range.first;
1928 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
1929 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
1931 CaseItr Pivot = CR.Range.first + Size/2;
1933 // Select optimal pivot, maximizing sum density of LHS and RHS. This will
1934 // (heuristically) allow us to emit JumpTable's later.
1935 APInt TSize(First.getBitWidth(), 0);
1936 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1940 APInt LSize = FrontCase.size();
1941 APInt RSize = TSize-LSize;
1942 DEBUG(dbgs() << "Selecting best pivot: \n"
1943 << "First: " << First << ", Last: " << Last <<'\n'
1944 << "LSize: " << LSize << ", RSize: " << RSize << '\n');
1945 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second;
1947 const APInt &LEnd = cast<ConstantInt>(I->High)->getValue();
1948 const APInt &RBegin = cast<ConstantInt>(J->Low)->getValue();
1949 APInt Range = ComputeRange(LEnd, RBegin);
1950 assert((Range - 2ULL).isNonNegative() &&
1951 "Invalid case distance");
1952 double LDensity = (double)LSize.roundToDouble() /
1953 (LEnd - First + 1ULL).roundToDouble();
1954 double RDensity = (double)RSize.roundToDouble() /
1955 (Last - RBegin + 1ULL).roundToDouble();
1956 double Metric = Range.logBase2()*(LDensity+RDensity);
1957 // Should always split in some non-trivial place
1958 DEBUG(dbgs() <<"=>Step\n"
1959 << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n'
1960 << "LDensity: " << LDensity
1961 << ", RDensity: " << RDensity << '\n'
1962 << "Metric: " << Metric << '\n');
1963 if (FMetric < Metric) {
1966 DEBUG(dbgs() << "Current metric set to: " << FMetric << '\n');
1972 if (areJTsAllowed(TLI)) {
1973 // If our case is dense we *really* should handle it earlier!
1974 assert((FMetric > 0) && "Should handle dense range earlier!");
1976 Pivot = CR.Range.first + Size/2;
1979 CaseRange LHSR(CR.Range.first, Pivot);
1980 CaseRange RHSR(Pivot, CR.Range.second);
1981 Constant *C = Pivot->Low;
1982 MachineBasicBlock *FalseBB = 0, *TrueBB = 0;
1984 // We know that we branch to the LHS if the Value being switched on is
1985 // less than the Pivot value, C. We use this to optimize our binary
1986 // tree a bit, by recognizing that if SV is greater than or equal to the
1987 // LHS's Case Value, and that Case Value is exactly one less than the
1988 // Pivot's Value, then we can branch directly to the LHS's Target,
1989 // rather than creating a leaf node for it.
1990 if ((LHSR.second - LHSR.first) == 1 &&
1991 LHSR.first->High == CR.GE &&
1992 cast<ConstantInt>(C)->getValue() ==
1993 (cast<ConstantInt>(CR.GE)->getValue() + 1LL)) {
1994 TrueBB = LHSR.first->BB;
1996 TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB);
1997 CurMF->insert(BBI, TrueBB);
1998 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR));
2000 // Put SV in a virtual register to make it available from the new blocks.
2001 ExportFromCurrentBlock(SV);
2004 // Similar to the optimization above, if the Value being switched on is
2005 // known to be less than the Constant CR.LT, and the current Case Value
2006 // is CR.LT - 1, then we can branch directly to the target block for
2007 // the current Case Value, rather than emitting a RHS leaf node for it.
2008 if ((RHSR.second - RHSR.first) == 1 && CR.LT &&
2009 cast<ConstantInt>(RHSR.first->Low)->getValue() ==
2010 (cast<ConstantInt>(CR.LT)->getValue() - 1LL)) {
2011 FalseBB = RHSR.first->BB;
2013 FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2014 CurMF->insert(BBI, FalseBB);
2015 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR));
2017 // Put SV in a virtual register to make it available from the new blocks.
2018 ExportFromCurrentBlock(SV);
2021 // Create a CaseBlock record representing a conditional branch to
2022 // the LHS node if the value being switched on SV is less than C.
2023 // Otherwise, branch to LHS.
2024 CaseBlock CB(ISD::SETLT, SV, C, NULL, TrueBB, FalseBB, CR.CaseBB);
2026 if (CR.CaseBB == SwitchBB)
2027 visitSwitchCase(CB, SwitchBB);
2029 SwitchCases.push_back(CB);
2034 /// handleBitTestsSwitchCase - if current case range has few destination and
2035 /// range span less, than machine word bitwidth, encode case range into series
2036 /// of masks and emit bit tests with these masks.
2037 bool SelectionDAGBuilder::handleBitTestsSwitchCase(CaseRec& CR,
2038 CaseRecVector& WorkList,
2040 MachineBasicBlock* Default,
2041 MachineBasicBlock *SwitchBB){
2042 EVT PTy = TLI.getPointerTy();
2043 unsigned IntPtrBits = PTy.getSizeInBits();
2045 Case& FrontCase = *CR.Range.first;
2046 Case& BackCase = *(CR.Range.second-1);
2048 // Get the MachineFunction which holds the current MBB. This is used when
2049 // inserting any additional MBBs necessary to represent the switch.
2050 MachineFunction *CurMF = FuncInfo.MF;
2052 // If target does not have legal shift left, do not emit bit tests at all.
2053 if (!TLI.isOperationLegal(ISD::SHL, TLI.getPointerTy()))
2057 for (CaseItr I = CR.Range.first, E = CR.Range.second;
2059 // Single case counts one, case range - two.
2060 numCmps += (I->Low == I->High ? 1 : 2);
2063 // Count unique destinations
2064 SmallSet<MachineBasicBlock*, 4> Dests;
2065 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2066 Dests.insert(I->BB);
2067 if (Dests.size() > 3)
2068 // Don't bother the code below, if there are too much unique destinations
2071 DEBUG(dbgs() << "Total number of unique destinations: "
2072 << Dests.size() << '\n'
2073 << "Total number of comparisons: " << numCmps << '\n');
2075 // Compute span of values.
2076 const APInt& minValue = cast<ConstantInt>(FrontCase.Low)->getValue();
2077 const APInt& maxValue = cast<ConstantInt>(BackCase.High)->getValue();
2078 APInt cmpRange = maxValue - minValue;
2080 DEBUG(dbgs() << "Compare range: " << cmpRange << '\n'
2081 << "Low bound: " << minValue << '\n'
2082 << "High bound: " << maxValue << '\n');
2084 if (cmpRange.uge(IntPtrBits) ||
2085 (!(Dests.size() == 1 && numCmps >= 3) &&
2086 !(Dests.size() == 2 && numCmps >= 5) &&
2087 !(Dests.size() >= 3 && numCmps >= 6)))
2090 DEBUG(dbgs() << "Emitting bit tests\n");
2091 APInt lowBound = APInt::getNullValue(cmpRange.getBitWidth());
2093 // Optimize the case where all the case values fit in a
2094 // word without having to subtract minValue. In this case,
2095 // we can optimize away the subtraction.
2096 if (minValue.isNonNegative() && maxValue.slt(IntPtrBits)) {
2097 cmpRange = maxValue;
2099 lowBound = minValue;
2102 CaseBitsVector CasesBits;
2103 unsigned i, count = 0;
2105 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2106 MachineBasicBlock* Dest = I->BB;
2107 for (i = 0; i < count; ++i)
2108 if (Dest == CasesBits[i].BB)
2112 assert((count < 3) && "Too much destinations to test!");
2113 CasesBits.push_back(CaseBits(0, Dest, 0));
2117 const APInt& lowValue = cast<ConstantInt>(I->Low)->getValue();
2118 const APInt& highValue = cast<ConstantInt>(I->High)->getValue();
2120 uint64_t lo = (lowValue - lowBound).getZExtValue();
2121 uint64_t hi = (highValue - lowBound).getZExtValue();
2123 for (uint64_t j = lo; j <= hi; j++) {
2124 CasesBits[i].Mask |= 1ULL << j;
2125 CasesBits[i].Bits++;
2129 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp());
2133 // Figure out which block is immediately after the current one.
2134 MachineFunction::iterator BBI = CR.CaseBB;
2137 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2139 DEBUG(dbgs() << "Cases:\n");
2140 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) {
2141 DEBUG(dbgs() << "Mask: " << CasesBits[i].Mask
2142 << ", Bits: " << CasesBits[i].Bits
2143 << ", BB: " << CasesBits[i].BB << '\n');
2145 MachineBasicBlock *CaseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2146 CurMF->insert(BBI, CaseBB);
2147 BTC.push_back(BitTestCase(CasesBits[i].Mask,
2151 // Put SV in a virtual register to make it available from the new blocks.
2152 ExportFromCurrentBlock(SV);
2155 BitTestBlock BTB(lowBound, cmpRange, SV,
2156 -1U, (CR.CaseBB == SwitchBB),
2157 CR.CaseBB, Default, BTC);
2159 if (CR.CaseBB == SwitchBB)
2160 visitBitTestHeader(BTB, SwitchBB);
2162 BitTestCases.push_back(BTB);
2167 /// Clusterify - Transform simple list of Cases into list of CaseRange's
2168 size_t SelectionDAGBuilder::Clusterify(CaseVector& Cases,
2169 const SwitchInst& SI) {
2172 // Start with "simple" cases
2173 for (size_t i = 1; i < SI.getNumSuccessors(); ++i) {
2174 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SI.getSuccessor(i)];
2175 Cases.push_back(Case(SI.getSuccessorValue(i),
2176 SI.getSuccessorValue(i),
2179 std::sort(Cases.begin(), Cases.end(), CaseCmp());
2181 // Merge case into clusters
2182 if (Cases.size() >= 2)
2183 // Must recompute end() each iteration because it may be
2184 // invalidated by erase if we hold on to it
2185 for (CaseItr I = Cases.begin(), J = ++(Cases.begin()); J != Cases.end(); ) {
2186 const APInt& nextValue = cast<ConstantInt>(J->Low)->getValue();
2187 const APInt& currentValue = cast<ConstantInt>(I->High)->getValue();
2188 MachineBasicBlock* nextBB = J->BB;
2189 MachineBasicBlock* currentBB = I->BB;
2191 // If the two neighboring cases go to the same destination, merge them
2192 // into a single case.
2193 if ((nextValue - currentValue == 1) && (currentBB == nextBB)) {
2201 for (CaseItr I=Cases.begin(), E=Cases.end(); I!=E; ++I, ++numCmps) {
2202 if (I->Low != I->High)
2203 // A range counts double, since it requires two compares.
2210 void SelectionDAGBuilder::UpdateSplitBlock(MachineBasicBlock *First,
2211 MachineBasicBlock *Last) {
2213 for (unsigned i = 0, e = JTCases.size(); i != e; ++i)
2214 if (JTCases[i].first.HeaderBB == First)
2215 JTCases[i].first.HeaderBB = Last;
2217 // Update BitTestCases.
2218 for (unsigned i = 0, e = BitTestCases.size(); i != e; ++i)
2219 if (BitTestCases[i].Parent == First)
2220 BitTestCases[i].Parent = Last;
2223 void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) {
2224 MachineBasicBlock *SwitchMBB = FuncInfo.MBB;
2226 // Figure out which block is immediately after the current one.
2227 MachineBasicBlock *NextBlock = 0;
2228 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()];
2230 // If there is only the default destination, branch to it if it is not the
2231 // next basic block. Otherwise, just fall through.
2232 if (SI.getNumOperands() == 2) {
2233 // Update machine-CFG edges.
2235 // If this is not a fall-through branch, emit the branch.
2236 SwitchMBB->addSuccessor(Default);
2237 if (Default != NextBlock)
2238 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
2239 MVT::Other, getControlRoot(),
2240 DAG.getBasicBlock(Default)));
2245 // If there are any non-default case statements, create a vector of Cases
2246 // representing each one, and sort the vector so that we can efficiently
2247 // create a binary search tree from them.
2249 size_t numCmps = Clusterify(Cases, SI);
2250 DEBUG(dbgs() << "Clusterify finished. Total clusters: " << Cases.size()
2251 << ". Total compares: " << numCmps << '\n');
2254 // Get the Value to be switched on and default basic blocks, which will be
2255 // inserted into CaseBlock records, representing basic blocks in the binary
2257 const Value *SV = SI.getOperand(0);
2259 // Push the initial CaseRec onto the worklist
2260 CaseRecVector WorkList;
2261 WorkList.push_back(CaseRec(SwitchMBB,0,0,
2262 CaseRange(Cases.begin(),Cases.end())));
2264 while (!WorkList.empty()) {
2265 // Grab a record representing a case range to process off the worklist
2266 CaseRec CR = WorkList.back();
2267 WorkList.pop_back();
2269 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2272 // If the range has few cases (two or less) emit a series of specific
2274 if (handleSmallSwitchRange(CR, WorkList, SV, Default, SwitchMBB))
2277 // If the switch has more than 5 blocks, and at least 40% dense, and the
2278 // target supports indirect branches, then emit a jump table rather than
2279 // lowering the switch to a binary tree of conditional branches.
2280 if (handleJTSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2283 // Emit binary tree. We need to pick a pivot, and push left and right ranges
2284 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call.
2285 handleBTSplitSwitchCase(CR, WorkList, SV, Default, SwitchMBB);
2289 void SelectionDAGBuilder::visitIndirectBr(const IndirectBrInst &I) {
2290 MachineBasicBlock *IndirectBrMBB = FuncInfo.MBB;
2292 // Update machine-CFG edges with unique successors.
2293 SmallVector<BasicBlock*, 32> succs;
2294 succs.reserve(I.getNumSuccessors());
2295 for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i)
2296 succs.push_back(I.getSuccessor(i));
2297 array_pod_sort(succs.begin(), succs.end());
2298 succs.erase(std::unique(succs.begin(), succs.end()), succs.end());
2299 for (unsigned i = 0, e = succs.size(); i != e; ++i)
2300 IndirectBrMBB->addSuccessor(FuncInfo.MBBMap[succs[i]]);
2302 DAG.setRoot(DAG.getNode(ISD::BRIND, getCurDebugLoc(),
2303 MVT::Other, getControlRoot(),
2304 getValue(I.getAddress())));
2307 void SelectionDAGBuilder::visitFSub(const User &I) {
2308 // -0.0 - X --> fneg
2309 const Type *Ty = I.getType();
2310 if (Ty->isVectorTy()) {
2311 if (ConstantVector *CV = dyn_cast<ConstantVector>(I.getOperand(0))) {
2312 const VectorType *DestTy = cast<VectorType>(I.getType());
2313 const Type *ElTy = DestTy->getElementType();
2314 unsigned VL = DestTy->getNumElements();
2315 std::vector<Constant*> NZ(VL, ConstantFP::getNegativeZero(ElTy));
2316 Constant *CNZ = ConstantVector::get(&NZ[0], NZ.size());
2318 SDValue Op2 = getValue(I.getOperand(1));
2319 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(),
2320 Op2.getValueType(), Op2));
2326 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I.getOperand(0)))
2327 if (CFP->isExactlyValue(ConstantFP::getNegativeZero(Ty)->getValueAPF())) {
2328 SDValue Op2 = getValue(I.getOperand(1));
2329 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(),
2330 Op2.getValueType(), Op2));
2334 visitBinary(I, ISD::FSUB);
2337 void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) {
2338 SDValue Op1 = getValue(I.getOperand(0));
2339 SDValue Op2 = getValue(I.getOperand(1));
2340 setValue(&I, DAG.getNode(OpCode, getCurDebugLoc(),
2341 Op1.getValueType(), Op1, Op2));
2344 void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) {
2345 SDValue Op1 = getValue(I.getOperand(0));
2346 SDValue Op2 = getValue(I.getOperand(1));
2347 if (!I.getType()->isVectorTy() &&
2348 Op2.getValueType() != TLI.getShiftAmountTy()) {
2349 // If the operand is smaller than the shift count type, promote it.
2350 EVT PTy = TLI.getPointerTy();
2351 EVT STy = TLI.getShiftAmountTy();
2352 if (STy.bitsGT(Op2.getValueType()))
2353 Op2 = DAG.getNode(ISD::ANY_EXTEND, getCurDebugLoc(),
2354 TLI.getShiftAmountTy(), Op2);
2355 // If the operand is larger than the shift count type but the shift
2356 // count type has enough bits to represent any shift value, truncate
2357 // it now. This is a common case and it exposes the truncate to
2358 // optimization early.
2359 else if (STy.getSizeInBits() >=
2360 Log2_32_Ceil(Op2.getValueType().getSizeInBits()))
2361 Op2 = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2362 TLI.getShiftAmountTy(), Op2);
2363 // Otherwise we'll need to temporarily settle for some other
2364 // convenient type; type legalization will make adjustments as
2366 else if (PTy.bitsLT(Op2.getValueType()))
2367 Op2 = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2368 TLI.getPointerTy(), Op2);
2369 else if (PTy.bitsGT(Op2.getValueType()))
2370 Op2 = DAG.getNode(ISD::ANY_EXTEND, getCurDebugLoc(),
2371 TLI.getPointerTy(), Op2);
2374 setValue(&I, DAG.getNode(Opcode, getCurDebugLoc(),
2375 Op1.getValueType(), Op1, Op2));
2378 void SelectionDAGBuilder::visitICmp(const User &I) {
2379 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2380 if (const ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2381 predicate = IC->getPredicate();
2382 else if (const ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2383 predicate = ICmpInst::Predicate(IC->getPredicate());
2384 SDValue Op1 = getValue(I.getOperand(0));
2385 SDValue Op2 = getValue(I.getOperand(1));
2386 ISD::CondCode Opcode = getICmpCondCode(predicate);
2388 EVT DestVT = TLI.getValueType(I.getType());
2389 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Opcode));
2392 void SelectionDAGBuilder::visitFCmp(const User &I) {
2393 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2394 if (const FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2395 predicate = FC->getPredicate();
2396 else if (const ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2397 predicate = FCmpInst::Predicate(FC->getPredicate());
2398 SDValue Op1 = getValue(I.getOperand(0));
2399 SDValue Op2 = getValue(I.getOperand(1));
2400 ISD::CondCode Condition = getFCmpCondCode(predicate);
2401 EVT DestVT = TLI.getValueType(I.getType());
2402 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Condition));
2405 void SelectionDAGBuilder::visitSelect(const User &I) {
2406 SmallVector<EVT, 4> ValueVTs;
2407 ComputeValueVTs(TLI, I.getType(), ValueVTs);
2408 unsigned NumValues = ValueVTs.size();
2409 if (NumValues == 0) return;
2411 SmallVector<SDValue, 4> Values(NumValues);
2412 SDValue Cond = getValue(I.getOperand(0));
2413 SDValue TrueVal = getValue(I.getOperand(1));
2414 SDValue FalseVal = getValue(I.getOperand(2));
2416 for (unsigned i = 0; i != NumValues; ++i)
2417 Values[i] = DAG.getNode(ISD::SELECT, getCurDebugLoc(),
2418 TrueVal.getNode()->getValueType(TrueVal.getResNo()+i),
2420 SDValue(TrueVal.getNode(),
2421 TrueVal.getResNo() + i),
2422 SDValue(FalseVal.getNode(),
2423 FalseVal.getResNo() + i));
2425 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2426 DAG.getVTList(&ValueVTs[0], NumValues),
2427 &Values[0], NumValues));
2430 void SelectionDAGBuilder::visitTrunc(const User &I) {
2431 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2432 SDValue N = getValue(I.getOperand(0));
2433 EVT DestVT = TLI.getValueType(I.getType());
2434 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), DestVT, N));
2437 void SelectionDAGBuilder::visitZExt(const User &I) {
2438 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2439 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2440 SDValue N = getValue(I.getOperand(0));
2441 EVT DestVT = TLI.getValueType(I.getType());
2442 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(), DestVT, N));
2445 void SelectionDAGBuilder::visitSExt(const User &I) {
2446 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2447 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2448 SDValue N = getValue(I.getOperand(0));
2449 EVT DestVT = TLI.getValueType(I.getType());
2450 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurDebugLoc(), DestVT, N));
2453 void SelectionDAGBuilder::visitFPTrunc(const User &I) {
2454 // FPTrunc is never a no-op cast, no need to check
2455 SDValue N = getValue(I.getOperand(0));
2456 EVT DestVT = TLI.getValueType(I.getType());
2457 setValue(&I, DAG.getNode(ISD::FP_ROUND, getCurDebugLoc(),
2458 DestVT, N, DAG.getIntPtrConstant(0)));
2461 void SelectionDAGBuilder::visitFPExt(const User &I){
2462 // FPTrunc is never a no-op cast, no need to check
2463 SDValue N = getValue(I.getOperand(0));
2464 EVT DestVT = TLI.getValueType(I.getType());
2465 setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurDebugLoc(), DestVT, N));
2468 void SelectionDAGBuilder::visitFPToUI(const User &I) {
2469 // FPToUI is never a no-op cast, no need to check
2470 SDValue N = getValue(I.getOperand(0));
2471 EVT DestVT = TLI.getValueType(I.getType());
2472 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurDebugLoc(), DestVT, N));
2475 void SelectionDAGBuilder::visitFPToSI(const User &I) {
2476 // FPToSI is never a no-op cast, no need to check
2477 SDValue N = getValue(I.getOperand(0));
2478 EVT DestVT = TLI.getValueType(I.getType());
2479 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurDebugLoc(), DestVT, N));
2482 void SelectionDAGBuilder::visitUIToFP(const User &I) {
2483 // UIToFP is never a no-op cast, no need to check
2484 SDValue N = getValue(I.getOperand(0));
2485 EVT DestVT = TLI.getValueType(I.getType());
2486 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurDebugLoc(), DestVT, N));
2489 void SelectionDAGBuilder::visitSIToFP(const User &I){
2490 // SIToFP is never a no-op cast, no need to check
2491 SDValue N = getValue(I.getOperand(0));
2492 EVT DestVT = TLI.getValueType(I.getType());
2493 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurDebugLoc(), DestVT, N));
2496 void SelectionDAGBuilder::visitPtrToInt(const User &I) {
2497 // What to do depends on the size of the integer and the size of the pointer.
2498 // We can either truncate, zero extend, or no-op, accordingly.
2499 SDValue N = getValue(I.getOperand(0));
2500 EVT DestVT = TLI.getValueType(I.getType());
2501 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT));
2504 void SelectionDAGBuilder::visitIntToPtr(const User &I) {
2505 // What to do depends on the size of the integer and the size of the pointer.
2506 // We can either truncate, zero extend, or no-op, accordingly.
2507 SDValue N = getValue(I.getOperand(0));
2508 EVT DestVT = TLI.getValueType(I.getType());
2509 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT));
2512 void SelectionDAGBuilder::visitBitCast(const User &I) {
2513 SDValue N = getValue(I.getOperand(0));
2514 EVT DestVT = TLI.getValueType(I.getType());
2516 // BitCast assures us that source and destination are the same size so this is
2517 // either a BIT_CONVERT or a no-op.
2518 if (DestVT != N.getValueType())
2519 setValue(&I, DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
2520 DestVT, N)); // convert types.
2522 setValue(&I, N); // noop cast.
2525 void SelectionDAGBuilder::visitInsertElement(const User &I) {
2526 SDValue InVec = getValue(I.getOperand(0));
2527 SDValue InVal = getValue(I.getOperand(1));
2528 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2530 getValue(I.getOperand(2)));
2531 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurDebugLoc(),
2532 TLI.getValueType(I.getType()),
2533 InVec, InVal, InIdx));
2536 void SelectionDAGBuilder::visitExtractElement(const User &I) {
2537 SDValue InVec = getValue(I.getOperand(0));
2538 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2540 getValue(I.getOperand(1)));
2541 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2542 TLI.getValueType(I.getType()), InVec, InIdx));
2545 // Utility for visitShuffleVector - Returns true if the mask is mask starting
2546 // from SIndx and increasing to the element length (undefs are allowed).
2547 static bool SequentialMask(SmallVectorImpl<int> &Mask, unsigned SIndx) {
2548 unsigned MaskNumElts = Mask.size();
2549 for (unsigned i = 0; i != MaskNumElts; ++i)
2550 if ((Mask[i] >= 0) && (Mask[i] != (int)(i + SIndx)))
2555 void SelectionDAGBuilder::visitShuffleVector(const User &I) {
2556 SmallVector<int, 8> Mask;
2557 SDValue Src1 = getValue(I.getOperand(0));
2558 SDValue Src2 = getValue(I.getOperand(1));
2560 // Convert the ConstantVector mask operand into an array of ints, with -1
2561 // representing undef values.
2562 SmallVector<Constant*, 8> MaskElts;
2563 cast<Constant>(I.getOperand(2))->getVectorElements(MaskElts);
2564 unsigned MaskNumElts = MaskElts.size();
2565 for (unsigned i = 0; i != MaskNumElts; ++i) {
2566 if (isa<UndefValue>(MaskElts[i]))
2569 Mask.push_back(cast<ConstantInt>(MaskElts[i])->getSExtValue());
2572 EVT VT = TLI.getValueType(I.getType());
2573 EVT SrcVT = Src1.getValueType();
2574 unsigned SrcNumElts = SrcVT.getVectorNumElements();
2576 if (SrcNumElts == MaskNumElts) {
2577 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2582 // Normalize the shuffle vector since mask and vector length don't match.
2583 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) {
2584 // Mask is longer than the source vectors and is a multiple of the source
2585 // vectors. We can use concatenate vector to make the mask and vectors
2587 if (SrcNumElts*2 == MaskNumElts && SequentialMask(Mask, 0)) {
2588 // The shuffle is concatenating two vectors together.
2589 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurDebugLoc(),
2594 // Pad both vectors with undefs to make them the same length as the mask.
2595 unsigned NumConcat = MaskNumElts / SrcNumElts;
2596 bool Src1U = Src1.getOpcode() == ISD::UNDEF;
2597 bool Src2U = Src2.getOpcode() == ISD::UNDEF;
2598 SDValue UndefVal = DAG.getUNDEF(SrcVT);
2600 SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal);
2601 SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal);
2605 Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2606 getCurDebugLoc(), VT,
2607 &MOps1[0], NumConcat);
2608 Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2609 getCurDebugLoc(), VT,
2610 &MOps2[0], NumConcat);
2612 // Readjust mask for new input vector length.
2613 SmallVector<int, 8> MappedOps;
2614 for (unsigned i = 0; i != MaskNumElts; ++i) {
2616 if (Idx < (int)SrcNumElts)
2617 MappedOps.push_back(Idx);
2619 MappedOps.push_back(Idx + MaskNumElts - SrcNumElts);
2622 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2627 if (SrcNumElts > MaskNumElts) {
2628 // Analyze the access pattern of the vector to see if we can extract
2629 // two subvectors and do the shuffle. The analysis is done by calculating
2630 // the range of elements the mask access on both vectors.
2631 int MinRange[2] = { SrcNumElts+1, SrcNumElts+1};
2632 int MaxRange[2] = {-1, -1};
2634 for (unsigned i = 0; i != MaskNumElts; ++i) {
2640 if (Idx >= (int)SrcNumElts) {
2644 if (Idx > MaxRange[Input])
2645 MaxRange[Input] = Idx;
2646 if (Idx < MinRange[Input])
2647 MinRange[Input] = Idx;
2650 // Check if the access is smaller than the vector size and can we find
2651 // a reasonable extract index.
2652 int RangeUse[2] = { 2, 2 }; // 0 = Unused, 1 = Extract, 2 = Can not
2654 int StartIdx[2]; // StartIdx to extract from
2655 for (int Input=0; Input < 2; ++Input) {
2656 if (MinRange[Input] == (int)(SrcNumElts+1) && MaxRange[Input] == -1) {
2657 RangeUse[Input] = 0; // Unused
2658 StartIdx[Input] = 0;
2659 } else if (MaxRange[Input] - MinRange[Input] < (int)MaskNumElts) {
2660 // Fits within range but we should see if we can find a good
2661 // start index that is a multiple of the mask length.
2662 if (MaxRange[Input] < (int)MaskNumElts) {
2663 RangeUse[Input] = 1; // Extract from beginning of the vector
2664 StartIdx[Input] = 0;
2666 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts;
2667 if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts &&
2668 StartIdx[Input] + MaskNumElts < SrcNumElts)
2669 RangeUse[Input] = 1; // Extract from a multiple of the mask length.
2674 if (RangeUse[0] == 0 && RangeUse[1] == 0) {
2675 setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used.
2678 else if (RangeUse[0] < 2 && RangeUse[1] < 2) {
2679 // Extract appropriate subvector and generate a vector shuffle
2680 for (int Input=0; Input < 2; ++Input) {
2681 SDValue &Src = Input == 0 ? Src1 : Src2;
2682 if (RangeUse[Input] == 0)
2683 Src = DAG.getUNDEF(VT);
2685 Src = DAG.getNode(ISD::EXTRACT_SUBVECTOR, getCurDebugLoc(), VT,
2686 Src, DAG.getIntPtrConstant(StartIdx[Input]));
2689 // Calculate new mask.
2690 SmallVector<int, 8> MappedOps;
2691 for (unsigned i = 0; i != MaskNumElts; ++i) {
2694 MappedOps.push_back(Idx);
2695 else if (Idx < (int)SrcNumElts)
2696 MappedOps.push_back(Idx - StartIdx[0]);
2698 MappedOps.push_back(Idx - SrcNumElts - StartIdx[1] + MaskNumElts);
2701 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2707 // We can't use either concat vectors or extract subvectors so fall back to
2708 // replacing the shuffle with extract and build vector.
2709 // to insert and build vector.
2710 EVT EltVT = VT.getVectorElementType();
2711 EVT PtrVT = TLI.getPointerTy();
2712 SmallVector<SDValue,8> Ops;
2713 for (unsigned i = 0; i != MaskNumElts; ++i) {
2715 Ops.push_back(DAG.getUNDEF(EltVT));
2720 if (Idx < (int)SrcNumElts)
2721 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2722 EltVT, Src1, DAG.getConstant(Idx, PtrVT));
2724 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2726 DAG.getConstant(Idx - SrcNumElts, PtrVT));
2732 setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
2733 VT, &Ops[0], Ops.size()));
2736 void SelectionDAGBuilder::visitInsertValue(const InsertValueInst &I) {
2737 const Value *Op0 = I.getOperand(0);
2738 const Value *Op1 = I.getOperand(1);
2739 const Type *AggTy = I.getType();
2740 const Type *ValTy = Op1->getType();
2741 bool IntoUndef = isa<UndefValue>(Op0);
2742 bool FromUndef = isa<UndefValue>(Op1);
2744 unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy,
2745 I.idx_begin(), I.idx_end());
2747 SmallVector<EVT, 4> AggValueVTs;
2748 ComputeValueVTs(TLI, AggTy, AggValueVTs);
2749 SmallVector<EVT, 4> ValValueVTs;
2750 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2752 unsigned NumAggValues = AggValueVTs.size();
2753 unsigned NumValValues = ValValueVTs.size();
2754 SmallVector<SDValue, 4> Values(NumAggValues);
2756 SDValue Agg = getValue(Op0);
2757 SDValue Val = getValue(Op1);
2759 // Copy the beginning value(s) from the original aggregate.
2760 for (; i != LinearIndex; ++i)
2761 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2762 SDValue(Agg.getNode(), Agg.getResNo() + i);
2763 // Copy values from the inserted value(s).
2764 for (; i != LinearIndex + NumValValues; ++i)
2765 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2766 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex);
2767 // Copy remaining value(s) from the original aggregate.
2768 for (; i != NumAggValues; ++i)
2769 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2770 SDValue(Agg.getNode(), Agg.getResNo() + i);
2772 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2773 DAG.getVTList(&AggValueVTs[0], NumAggValues),
2774 &Values[0], NumAggValues));
2777 void SelectionDAGBuilder::visitExtractValue(const ExtractValueInst &I) {
2778 const Value *Op0 = I.getOperand(0);
2779 const Type *AggTy = Op0->getType();
2780 const Type *ValTy = I.getType();
2781 bool OutOfUndef = isa<UndefValue>(Op0);
2783 unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy,
2784 I.idx_begin(), I.idx_end());
2786 SmallVector<EVT, 4> ValValueVTs;
2787 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2789 unsigned NumValValues = ValValueVTs.size();
2790 SmallVector<SDValue, 4> Values(NumValValues);
2792 SDValue Agg = getValue(Op0);
2793 // Copy out the selected value(s).
2794 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i)
2795 Values[i - LinearIndex] =
2797 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) :
2798 SDValue(Agg.getNode(), Agg.getResNo() + i);
2800 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2801 DAG.getVTList(&ValValueVTs[0], NumValValues),
2802 &Values[0], NumValValues));
2805 void SelectionDAGBuilder::visitGetElementPtr(const User &I) {
2806 SDValue N = getValue(I.getOperand(0));
2807 const Type *Ty = I.getOperand(0)->getType();
2809 for (GetElementPtrInst::const_op_iterator OI = I.op_begin()+1, E = I.op_end();
2811 const Value *Idx = *OI;
2812 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
2813 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
2816 uint64_t Offset = TD->getStructLayout(StTy)->getElementOffset(Field);
2817 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
2818 DAG.getIntPtrConstant(Offset));
2821 Ty = StTy->getElementType(Field);
2823 Ty = cast<SequentialType>(Ty)->getElementType();
2825 // If this is a constant subscript, handle it quickly.
2826 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
2827 if (CI->isZero()) continue;
2829 TD->getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
2831 EVT PTy = TLI.getPointerTy();
2832 unsigned PtrBits = PTy.getSizeInBits();
2834 OffsVal = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2836 DAG.getConstant(Offs, MVT::i64));
2838 OffsVal = DAG.getIntPtrConstant(Offs);
2840 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
2845 // N = N + Idx * ElementSize;
2846 APInt ElementSize = APInt(TLI.getPointerTy().getSizeInBits(),
2847 TD->getTypeAllocSize(Ty));
2848 SDValue IdxN = getValue(Idx);
2850 // If the index is smaller or larger than intptr_t, truncate or extend
2852 IdxN = DAG.getSExtOrTrunc(IdxN, getCurDebugLoc(), N.getValueType());
2854 // If this is a multiply by a power of two, turn it into a shl
2855 // immediately. This is a very common case.
2856 if (ElementSize != 1) {
2857 if (ElementSize.isPowerOf2()) {
2858 unsigned Amt = ElementSize.logBase2();
2859 IdxN = DAG.getNode(ISD::SHL, getCurDebugLoc(),
2860 N.getValueType(), IdxN,
2861 DAG.getConstant(Amt, TLI.getPointerTy()));
2863 SDValue Scale = DAG.getConstant(ElementSize, TLI.getPointerTy());
2864 IdxN = DAG.getNode(ISD::MUL, getCurDebugLoc(),
2865 N.getValueType(), IdxN, Scale);
2869 N = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2870 N.getValueType(), N, IdxN);
2877 void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) {
2878 // If this is a fixed sized alloca in the entry block of the function,
2879 // allocate it statically on the stack.
2880 if (FuncInfo.StaticAllocaMap.count(&I))
2881 return; // getValue will auto-populate this.
2883 const Type *Ty = I.getAllocatedType();
2884 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
2886 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty),
2889 SDValue AllocSize = getValue(I.getArraySize());
2891 EVT IntPtr = TLI.getPointerTy();
2892 if (AllocSize.getValueType() != IntPtr)
2893 AllocSize = DAG.getZExtOrTrunc(AllocSize, getCurDebugLoc(), IntPtr);
2895 AllocSize = DAG.getNode(ISD::MUL, getCurDebugLoc(), IntPtr,
2897 DAG.getConstant(TySize, IntPtr));
2899 // Handle alignment. If the requested alignment is less than or equal to
2900 // the stack alignment, ignore it. If the size is greater than or equal to
2901 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
2902 unsigned StackAlign = TM.getFrameInfo()->getStackAlignment();
2903 if (Align <= StackAlign)
2906 // Round the size of the allocation up to the stack alignment size
2907 // by add SA-1 to the size.
2908 AllocSize = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2909 AllocSize.getValueType(), AllocSize,
2910 DAG.getIntPtrConstant(StackAlign-1));
2912 // Mask out the low bits for alignment purposes.
2913 AllocSize = DAG.getNode(ISD::AND, getCurDebugLoc(),
2914 AllocSize.getValueType(), AllocSize,
2915 DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1)));
2917 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) };
2918 SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other);
2919 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, getCurDebugLoc(),
2922 DAG.setRoot(DSA.getValue(1));
2924 // Inform the Frame Information that we have just allocated a variable-sized
2926 FuncInfo.MF->getFrameInfo()->CreateVariableSizedObject(Align ? Align : 1);
2929 void SelectionDAGBuilder::visitLoad(const LoadInst &I) {
2930 const Value *SV = I.getOperand(0);
2931 SDValue Ptr = getValue(SV);
2933 const Type *Ty = I.getType();
2935 bool isVolatile = I.isVolatile();
2936 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
2937 unsigned Alignment = I.getAlignment();
2939 SmallVector<EVT, 4> ValueVTs;
2940 SmallVector<uint64_t, 4> Offsets;
2941 ComputeValueVTs(TLI, Ty, ValueVTs, &Offsets);
2942 unsigned NumValues = ValueVTs.size();
2947 bool ConstantMemory = false;
2949 // Serialize volatile loads with other side effects.
2951 else if (AA->pointsToConstantMemory(SV)) {
2952 // Do not serialize (non-volatile) loads of constant memory with anything.
2953 Root = DAG.getEntryNode();
2954 ConstantMemory = true;
2956 // Do not serialize non-volatile loads against each other.
2957 Root = DAG.getRoot();
2960 SmallVector<SDValue, 4> Values(NumValues);
2961 SmallVector<SDValue, 4> Chains(NumValues);
2962 EVT PtrVT = Ptr.getValueType();
2963 for (unsigned i = 0; i != NumValues; ++i) {
2964 SDValue A = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2966 DAG.getConstant(Offsets[i], PtrVT));
2967 SDValue L = DAG.getLoad(ValueVTs[i], getCurDebugLoc(), Root,
2968 A, MachinePointerInfo(SV, Offsets[i]), isVolatile,
2969 isNonTemporal, Alignment);
2972 Chains[i] = L.getValue(1);
2975 if (!ConstantMemory) {
2976 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
2977 MVT::Other, &Chains[0], NumValues);
2981 PendingLoads.push_back(Chain);
2984 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2985 DAG.getVTList(&ValueVTs[0], NumValues),
2986 &Values[0], NumValues));
2989 void SelectionDAGBuilder::visitStore(const StoreInst &I) {
2990 const Value *SrcV = I.getOperand(0);
2991 const Value *PtrV = I.getOperand(1);
2993 SmallVector<EVT, 4> ValueVTs;
2994 SmallVector<uint64_t, 4> Offsets;
2995 ComputeValueVTs(TLI, SrcV->getType(), ValueVTs, &Offsets);
2996 unsigned NumValues = ValueVTs.size();
3000 // Get the lowered operands. Note that we do this after
3001 // checking if NumResults is zero, because with zero results
3002 // the operands won't have values in the map.
3003 SDValue Src = getValue(SrcV);
3004 SDValue Ptr = getValue(PtrV);
3006 SDValue Root = getRoot();
3007 SmallVector<SDValue, 4> Chains(NumValues);
3008 EVT PtrVT = Ptr.getValueType();
3009 bool isVolatile = I.isVolatile();
3010 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
3011 unsigned Alignment = I.getAlignment();
3013 for (unsigned i = 0; i != NumValues; ++i) {
3014 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT, Ptr,
3015 DAG.getConstant(Offsets[i], PtrVT));
3016 Chains[i] = DAG.getStore(Root, getCurDebugLoc(),
3017 SDValue(Src.getNode(), Src.getResNo() + i),
3018 Add, MachinePointerInfo(PtrV, Offsets[i]),
3019 isVolatile, isNonTemporal, Alignment);
3022 DAG.setRoot(DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3023 MVT::Other, &Chains[0], NumValues));
3026 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
3028 void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I,
3029 unsigned Intrinsic) {
3030 bool HasChain = !I.doesNotAccessMemory();
3031 bool OnlyLoad = HasChain && I.onlyReadsMemory();
3033 // Build the operand list.
3034 SmallVector<SDValue, 8> Ops;
3035 if (HasChain) { // If this intrinsic has side-effects, chainify it.
3037 // We don't need to serialize loads against other loads.
3038 Ops.push_back(DAG.getRoot());
3040 Ops.push_back(getRoot());
3044 // Info is set by getTgtMemInstrinsic
3045 TargetLowering::IntrinsicInfo Info;
3046 bool IsTgtIntrinsic = TLI.getTgtMemIntrinsic(Info, I, Intrinsic);
3048 // Add the intrinsic ID as an integer operand if it's not a target intrinsic.
3049 if (!IsTgtIntrinsic || Info.opc == ISD::INTRINSIC_VOID ||
3050 Info.opc == ISD::INTRINSIC_W_CHAIN)
3051 Ops.push_back(DAG.getConstant(Intrinsic, TLI.getPointerTy()));
3053 // Add all operands of the call to the operand list.
3054 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) {
3055 SDValue Op = getValue(I.getArgOperand(i));
3056 assert(TLI.isTypeLegal(Op.getValueType()) &&
3057 "Intrinsic uses a non-legal type?");
3061 SmallVector<EVT, 4> ValueVTs;
3062 ComputeValueVTs(TLI, I.getType(), ValueVTs);
3064 for (unsigned Val = 0, E = ValueVTs.size(); Val != E; ++Val) {
3065 assert(TLI.isTypeLegal(ValueVTs[Val]) &&
3066 "Intrinsic uses a non-legal type?");
3071 ValueVTs.push_back(MVT::Other);
3073 SDVTList VTs = DAG.getVTList(ValueVTs.data(), ValueVTs.size());
3077 if (IsTgtIntrinsic) {
3078 // This is target intrinsic that touches memory
3079 Result = DAG.getMemIntrinsicNode(Info.opc, getCurDebugLoc(),
3080 VTs, &Ops[0], Ops.size(),
3082 MachinePointerInfo(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 const Argument *Arg = dyn_cast<Argument>(V);
3941 MachineFunction &MF = DAG.getMachineFunction();
3942 // Ignore inlined function arguments here.
3943 DIVariable DV(Variable);
3944 if (DV.isInlinedFnArgument(MF.getFunction()))
3947 MachineBasicBlock *MBB = FuncInfo.MBB;
3948 if (MBB != &MF.front())
3952 if (Arg->hasByValAttr()) {
3953 // Byval arguments' frame index is recorded during argument lowering.
3954 // Use this info directly.
3955 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
3956 Reg = TRI->getFrameRegister(MF);
3957 Offset = FuncInfo.getByValArgumentFrameIndex(Arg);
3960 if (N.getNode() && N.getOpcode() == ISD::CopyFromReg) {
3961 Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg();
3962 if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) {
3963 MachineRegisterInfo &RegInfo = MF.getRegInfo();
3964 unsigned PR = RegInfo.getLiveInPhysReg(Reg);
3971 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
3972 if (VMI == FuncInfo.ValueMap.end())
3977 const TargetInstrInfo *TII = DAG.getTarget().getInstrInfo();
3978 MachineInstrBuilder MIB = BuildMI(MF, getCurDebugLoc(),
3979 TII->get(TargetOpcode::DBG_VALUE))
3980 .addReg(Reg, RegState::Debug).addImm(Offset).addMetadata(Variable);
3981 FuncInfo.ArgDbgValues.push_back(&*MIB);
3985 // VisualStudio defines setjmp as _setjmp
3986 #if defined(_MSC_VER) && defined(setjmp) && \
3987 !defined(setjmp_undefined_for_msvc)
3988 # pragma push_macro("setjmp")
3990 # define setjmp_undefined_for_msvc
3993 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
3994 /// we want to emit this as a call to a named external function, return the name
3995 /// otherwise lower it and return null.
3997 SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) {
3998 DebugLoc dl = getCurDebugLoc();
4001 switch (Intrinsic) {
4003 // By default, turn this into a target intrinsic node.
4004 visitTargetIntrinsic(I, Intrinsic);
4006 case Intrinsic::vastart: visitVAStart(I); return 0;
4007 case Intrinsic::vaend: visitVAEnd(I); return 0;
4008 case Intrinsic::vacopy: visitVACopy(I); return 0;
4009 case Intrinsic::returnaddress:
4010 setValue(&I, DAG.getNode(ISD::RETURNADDR, dl, TLI.getPointerTy(),
4011 getValue(I.getArgOperand(0))));
4013 case Intrinsic::frameaddress:
4014 setValue(&I, DAG.getNode(ISD::FRAMEADDR, dl, TLI.getPointerTy(),
4015 getValue(I.getArgOperand(0))));
4017 case Intrinsic::setjmp:
4018 return "_setjmp"+!TLI.usesUnderscoreSetJmp();
4019 case Intrinsic::longjmp:
4020 return "_longjmp"+!TLI.usesUnderscoreLongJmp();
4021 case Intrinsic::memcpy: {
4022 // Assert for address < 256 since we support only user defined address
4024 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4026 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4028 "Unknown address space");
4029 SDValue Op1 = getValue(I.getArgOperand(0));
4030 SDValue Op2 = getValue(I.getArgOperand(1));
4031 SDValue Op3 = getValue(I.getArgOperand(2));
4032 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4033 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4034 DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, isVol, false,
4035 MachinePointerInfo(I.getArgOperand(0)),
4036 MachinePointerInfo(I.getArgOperand(1))));
4039 case Intrinsic::memset: {
4040 // Assert for address < 256 since we support only user defined address
4042 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4044 "Unknown address space");
4045 SDValue Op1 = getValue(I.getArgOperand(0));
4046 SDValue Op2 = getValue(I.getArgOperand(1));
4047 SDValue Op3 = getValue(I.getArgOperand(2));
4048 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4049 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4050 DAG.setRoot(DAG.getMemset(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4051 MachinePointerInfo(I.getArgOperand(0))));
4054 case Intrinsic::memmove: {
4055 // Assert for address < 256 since we support only user defined address
4057 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4059 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4061 "Unknown address space");
4062 SDValue Op1 = getValue(I.getArgOperand(0));
4063 SDValue Op2 = getValue(I.getArgOperand(1));
4064 SDValue Op3 = getValue(I.getArgOperand(2));
4065 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4066 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4068 // If the source and destination are known to not be aliases, we can
4069 // lower memmove as memcpy.
4070 uint64_t Size = -1ULL;
4071 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op3))
4072 Size = C->getZExtValue();
4073 if (AA->alias(I.getArgOperand(0), Size, I.getArgOperand(1), Size) ==
4074 AliasAnalysis::NoAlias) {
4075 DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4076 false, MachinePointerInfo(I.getArgOperand(0)),
4077 MachinePointerInfo(I.getArgOperand(1))));
4081 DAG.setRoot(DAG.getMemmove(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4082 MachinePointerInfo(I.getArgOperand(0)),
4083 MachinePointerInfo(I.getArgOperand(1))));
4086 case Intrinsic::dbg_declare: {
4087 const DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
4088 MDNode *Variable = DI.getVariable();
4089 const Value *Address = DI.getAddress();
4090 if (!Address || !DIVariable(DI.getVariable()).Verify())
4093 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder
4094 // but do not always have a corresponding SDNode built. The SDNodeOrder
4095 // absolute, but not relative, values are different depending on whether
4096 // debug info exists.
4099 // Check if address has undef value.
4100 if (isa<UndefValue>(Address) ||
4101 (Address->use_empty() && !isa<Argument>(Address))) {
4103 DAG.getDbgValue(Variable, UndefValue::get(Address->getType()),
4104 0, dl, SDNodeOrder);
4105 DAG.AddDbgValue(SDV, 0, false);
4109 SDValue &N = NodeMap[Address];
4110 if (!N.getNode() && isa<Argument>(Address))
4111 // Check unused arguments map.
4112 N = UnusedArgNodeMap[Address];
4115 // Parameters are handled specially.
4117 DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable;
4118 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
4119 Address = BCI->getOperand(0);
4120 const AllocaInst *AI = dyn_cast<AllocaInst>(Address);
4122 if (isParameter && !AI) {
4123 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode());
4125 // Byval parameter. We have a frame index at this point.
4126 SDV = DAG.getDbgValue(Variable, FINode->getIndex(),
4127 0, dl, SDNodeOrder);
4129 // Can't do anything with other non-AI cases yet. This might be a
4130 // parameter of a callee function that got inlined, for example.
4133 SDV = DAG.getDbgValue(Variable, N.getNode(), N.getResNo(),
4134 0, dl, SDNodeOrder);
4136 // Can't do anything with other non-AI cases yet.
4138 DAG.AddDbgValue(SDV, N.getNode(), isParameter);
4140 // If Address is an arugment then try to emits its dbg value using
4141 // virtual register info from the FuncInfo.ValueMap.
4142 if (!EmitFuncArgumentDbgValue(Address, Variable, 0, N)) {
4143 // If variable is pinned by a alloca in dominating bb then
4144 // use StaticAllocaMap.
4145 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Address)) {
4146 if (AI->getParent() != DI.getParent()) {
4147 DenseMap<const AllocaInst*, int>::iterator SI =
4148 FuncInfo.StaticAllocaMap.find(AI);
4149 if (SI != FuncInfo.StaticAllocaMap.end()) {
4150 SDV = DAG.getDbgValue(Variable, SI->second,
4151 0, dl, SDNodeOrder);
4152 DAG.AddDbgValue(SDV, 0, false);
4157 // Otherwise add undef to help track missing debug info.
4158 SDV = DAG.getDbgValue(Variable, UndefValue::get(Address->getType()),
4159 0, dl, SDNodeOrder);
4160 DAG.AddDbgValue(SDV, 0, false);
4165 case Intrinsic::dbg_value: {
4166 const DbgValueInst &DI = cast<DbgValueInst>(I);
4167 if (!DIVariable(DI.getVariable()).Verify())
4170 MDNode *Variable = DI.getVariable();
4171 uint64_t Offset = DI.getOffset();
4172 const Value *V = DI.getValue();
4176 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder
4177 // but do not always have a corresponding SDNode built. The SDNodeOrder
4178 // absolute, but not relative, values are different depending on whether
4179 // debug info exists.
4182 if (isa<ConstantInt>(V) || isa<ConstantFP>(V)) {
4183 SDV = DAG.getDbgValue(Variable, V, Offset, dl, SDNodeOrder);
4184 DAG.AddDbgValue(SDV, 0, false);
4186 // Do not use getValue() in here; we don't want to generate code at
4187 // this point if it hasn't been done yet.
4188 SDValue N = NodeMap[V];
4189 if (!N.getNode() && isa<Argument>(V))
4190 // Check unused arguments map.
4191 N = UnusedArgNodeMap[V];
4193 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, N)) {
4194 SDV = DAG.getDbgValue(Variable, N.getNode(),
4195 N.getResNo(), Offset, dl, SDNodeOrder);
4196 DAG.AddDbgValue(SDV, N.getNode(), false);
4198 } else if (isa<PHINode>(V) && !V->use_empty() ) {
4199 // Do not call getValue(V) yet, as we don't want to generate code.
4200 // Remember it for later.
4201 DanglingDebugInfo DDI(&DI, dl, SDNodeOrder);
4202 DanglingDebugInfoMap[V] = DDI;
4204 // We may expand this to cover more cases. One case where we have no
4205 // data available is an unreferenced parameter; we need this fallback.
4206 SDV = DAG.getDbgValue(Variable, UndefValue::get(V->getType()),
4207 Offset, dl, SDNodeOrder);
4208 DAG.AddDbgValue(SDV, 0, false);
4212 // Build a debug info table entry.
4213 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
4214 V = BCI->getOperand(0);
4215 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
4216 // Don't handle byval struct arguments or VLAs, for example.
4219 DenseMap<const AllocaInst*, int>::iterator SI =
4220 FuncInfo.StaticAllocaMap.find(AI);
4221 if (SI == FuncInfo.StaticAllocaMap.end())
4223 int FI = SI->second;
4225 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4226 if (!DI.getDebugLoc().isUnknown() && MMI.hasDebugInfo())
4227 MMI.setVariableDbgInfo(Variable, FI, DI.getDebugLoc());
4230 case Intrinsic::eh_exception: {
4231 // Insert the EXCEPTIONADDR instruction.
4232 assert(FuncInfo.MBB->isLandingPad() &&
4233 "Call to eh.exception not in landing pad!");
4234 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
4236 Ops[0] = DAG.getRoot();
4237 SDValue Op = DAG.getNode(ISD::EXCEPTIONADDR, dl, VTs, Ops, 1);
4239 DAG.setRoot(Op.getValue(1));
4243 case Intrinsic::eh_selector: {
4244 MachineBasicBlock *CallMBB = FuncInfo.MBB;
4245 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4246 if (CallMBB->isLandingPad())
4247 AddCatchInfo(I, &MMI, CallMBB);
4250 FuncInfo.CatchInfoLost.insert(&I);
4252 // FIXME: Mark exception selector register as live in. Hack for PR1508.
4253 unsigned Reg = TLI.getExceptionSelectorRegister();
4254 if (Reg) FuncInfo.MBB->addLiveIn(Reg);
4257 // Insert the EHSELECTION instruction.
4258 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
4260 Ops[0] = getValue(I.getArgOperand(0));
4262 SDValue Op = DAG.getNode(ISD::EHSELECTION, dl, VTs, Ops, 2);
4263 DAG.setRoot(Op.getValue(1));
4264 setValue(&I, DAG.getSExtOrTrunc(Op, dl, MVT::i32));
4268 case Intrinsic::eh_typeid_for: {
4269 // Find the type id for the given typeinfo.
4270 GlobalVariable *GV = ExtractTypeInfo(I.getArgOperand(0));
4271 unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(GV);
4272 Res = DAG.getConstant(TypeID, MVT::i32);
4277 case Intrinsic::eh_return_i32:
4278 case Intrinsic::eh_return_i64:
4279 DAG.getMachineFunction().getMMI().setCallsEHReturn(true);
4280 DAG.setRoot(DAG.getNode(ISD::EH_RETURN, dl,
4283 getValue(I.getArgOperand(0)),
4284 getValue(I.getArgOperand(1))));
4286 case Intrinsic::eh_unwind_init:
4287 DAG.getMachineFunction().getMMI().setCallsUnwindInit(true);
4289 case Intrinsic::eh_dwarf_cfa: {
4290 SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getArgOperand(0)), dl,
4291 TLI.getPointerTy());
4292 SDValue Offset = DAG.getNode(ISD::ADD, dl,
4294 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, dl,
4295 TLI.getPointerTy()),
4297 SDValue FA = DAG.getNode(ISD::FRAMEADDR, dl,
4299 DAG.getConstant(0, TLI.getPointerTy()));
4300 setValue(&I, DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
4304 case Intrinsic::eh_sjlj_callsite: {
4305 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4306 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(0));
4307 assert(CI && "Non-constant call site value in eh.sjlj.callsite!");
4308 assert(MMI.getCurrentCallSite() == 0 && "Overlapping call sites!");
4310 MMI.setCurrentCallSite(CI->getZExtValue());
4313 case Intrinsic::eh_sjlj_setjmp: {
4314 setValue(&I, DAG.getNode(ISD::EH_SJLJ_SETJMP, dl, MVT::i32, getRoot(),
4315 getValue(I.getArgOperand(0))));
4318 case Intrinsic::eh_sjlj_longjmp: {
4319 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, dl, MVT::Other,
4321 getValue(I.getArgOperand(0))));
4325 case Intrinsic::x86_mmx_pslli_w:
4326 case Intrinsic::x86_mmx_pslli_d:
4327 case Intrinsic::x86_mmx_pslli_q:
4328 case Intrinsic::x86_mmx_psrli_w:
4329 case Intrinsic::x86_mmx_psrli_d:
4330 case Intrinsic::x86_mmx_psrli_q:
4331 case Intrinsic::x86_mmx_psrai_w:
4332 case Intrinsic::x86_mmx_psrai_d: {
4333 SDValue ShAmt = getValue(I.getArgOperand(1));
4334 if (isa<ConstantSDNode>(ShAmt)) {
4335 visitTargetIntrinsic(I, Intrinsic);
4338 unsigned NewIntrinsic = 0;
4339 EVT ShAmtVT = MVT::v2i32;
4340 switch (Intrinsic) {
4341 case Intrinsic::x86_mmx_pslli_w:
4342 NewIntrinsic = Intrinsic::x86_mmx_psll_w;
4344 case Intrinsic::x86_mmx_pslli_d:
4345 NewIntrinsic = Intrinsic::x86_mmx_psll_d;
4347 case Intrinsic::x86_mmx_pslli_q:
4348 NewIntrinsic = Intrinsic::x86_mmx_psll_q;
4350 case Intrinsic::x86_mmx_psrli_w:
4351 NewIntrinsic = Intrinsic::x86_mmx_psrl_w;
4353 case Intrinsic::x86_mmx_psrli_d:
4354 NewIntrinsic = Intrinsic::x86_mmx_psrl_d;
4356 case Intrinsic::x86_mmx_psrli_q:
4357 NewIntrinsic = Intrinsic::x86_mmx_psrl_q;
4359 case Intrinsic::x86_mmx_psrai_w:
4360 NewIntrinsic = Intrinsic::x86_mmx_psra_w;
4362 case Intrinsic::x86_mmx_psrai_d:
4363 NewIntrinsic = Intrinsic::x86_mmx_psra_d;
4365 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4368 // The vector shift intrinsics with scalars uses 32b shift amounts but
4369 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
4371 // We must do this early because v2i32 is not a legal type.
4372 DebugLoc dl = getCurDebugLoc();
4375 ShOps[1] = DAG.getConstant(0, MVT::i32);
4376 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
4377 EVT DestVT = TLI.getValueType(I.getType());
4378 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, DestVT, ShAmt);
4379 Res = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4380 DAG.getConstant(NewIntrinsic, MVT::i32),
4381 getValue(I.getArgOperand(0)), ShAmt);
4385 case Intrinsic::convertff:
4386 case Intrinsic::convertfsi:
4387 case Intrinsic::convertfui:
4388 case Intrinsic::convertsif:
4389 case Intrinsic::convertuif:
4390 case Intrinsic::convertss:
4391 case Intrinsic::convertsu:
4392 case Intrinsic::convertus:
4393 case Intrinsic::convertuu: {
4394 ISD::CvtCode Code = ISD::CVT_INVALID;
4395 switch (Intrinsic) {
4396 case Intrinsic::convertff: Code = ISD::CVT_FF; break;
4397 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break;
4398 case Intrinsic::convertfui: Code = ISD::CVT_FU; break;
4399 case Intrinsic::convertsif: Code = ISD::CVT_SF; break;
4400 case Intrinsic::convertuif: Code = ISD::CVT_UF; break;
4401 case Intrinsic::convertss: Code = ISD::CVT_SS; break;
4402 case Intrinsic::convertsu: Code = ISD::CVT_SU; break;
4403 case Intrinsic::convertus: Code = ISD::CVT_US; break;
4404 case Intrinsic::convertuu: Code = ISD::CVT_UU; break;
4406 EVT DestVT = TLI.getValueType(I.getType());
4407 const Value *Op1 = I.getArgOperand(0);
4408 Res = DAG.getConvertRndSat(DestVT, getCurDebugLoc(), getValue(Op1),
4409 DAG.getValueType(DestVT),
4410 DAG.getValueType(getValue(Op1).getValueType()),
4411 getValue(I.getArgOperand(1)),
4412 getValue(I.getArgOperand(2)),
4417 case Intrinsic::sqrt:
4418 setValue(&I, DAG.getNode(ISD::FSQRT, dl,
4419 getValue(I.getArgOperand(0)).getValueType(),
4420 getValue(I.getArgOperand(0))));
4422 case Intrinsic::powi:
4423 setValue(&I, ExpandPowI(dl, getValue(I.getArgOperand(0)),
4424 getValue(I.getArgOperand(1)), DAG));
4426 case Intrinsic::sin:
4427 setValue(&I, DAG.getNode(ISD::FSIN, dl,
4428 getValue(I.getArgOperand(0)).getValueType(),
4429 getValue(I.getArgOperand(0))));
4431 case Intrinsic::cos:
4432 setValue(&I, DAG.getNode(ISD::FCOS, dl,
4433 getValue(I.getArgOperand(0)).getValueType(),
4434 getValue(I.getArgOperand(0))));
4436 case Intrinsic::log:
4439 case Intrinsic::log2:
4442 case Intrinsic::log10:
4445 case Intrinsic::exp:
4448 case Intrinsic::exp2:
4451 case Intrinsic::pow:
4454 case Intrinsic::convert_to_fp16:
4455 setValue(&I, DAG.getNode(ISD::FP32_TO_FP16, dl,
4456 MVT::i16, getValue(I.getArgOperand(0))));
4458 case Intrinsic::convert_from_fp16:
4459 setValue(&I, DAG.getNode(ISD::FP16_TO_FP32, dl,
4460 MVT::f32, getValue(I.getArgOperand(0))));
4462 case Intrinsic::pcmarker: {
4463 SDValue Tmp = getValue(I.getArgOperand(0));
4464 DAG.setRoot(DAG.getNode(ISD::PCMARKER, dl, MVT::Other, getRoot(), Tmp));
4467 case Intrinsic::readcyclecounter: {
4468 SDValue Op = getRoot();
4469 Res = DAG.getNode(ISD::READCYCLECOUNTER, dl,
4470 DAG.getVTList(MVT::i64, MVT::Other),
4473 DAG.setRoot(Res.getValue(1));
4476 case Intrinsic::bswap:
4477 setValue(&I, DAG.getNode(ISD::BSWAP, dl,
4478 getValue(I.getArgOperand(0)).getValueType(),
4479 getValue(I.getArgOperand(0))));
4481 case Intrinsic::cttz: {
4482 SDValue Arg = getValue(I.getArgOperand(0));
4483 EVT Ty = Arg.getValueType();
4484 setValue(&I, DAG.getNode(ISD::CTTZ, dl, Ty, Arg));
4487 case Intrinsic::ctlz: {
4488 SDValue Arg = getValue(I.getArgOperand(0));
4489 EVT Ty = Arg.getValueType();
4490 setValue(&I, DAG.getNode(ISD::CTLZ, dl, Ty, Arg));
4493 case Intrinsic::ctpop: {
4494 SDValue Arg = getValue(I.getArgOperand(0));
4495 EVT Ty = Arg.getValueType();
4496 setValue(&I, DAG.getNode(ISD::CTPOP, dl, Ty, Arg));
4499 case Intrinsic::stacksave: {
4500 SDValue Op = getRoot();
4501 Res = DAG.getNode(ISD::STACKSAVE, dl,
4502 DAG.getVTList(TLI.getPointerTy(), MVT::Other), &Op, 1);
4504 DAG.setRoot(Res.getValue(1));
4507 case Intrinsic::stackrestore: {
4508 Res = getValue(I.getArgOperand(0));
4509 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, dl, MVT::Other, getRoot(), Res));
4512 case Intrinsic::stackprotector: {
4513 // Emit code into the DAG to store the stack guard onto the stack.
4514 MachineFunction &MF = DAG.getMachineFunction();
4515 MachineFrameInfo *MFI = MF.getFrameInfo();
4516 EVT PtrTy = TLI.getPointerTy();
4518 SDValue Src = getValue(I.getArgOperand(0)); // The guard's value.
4519 AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
4521 int FI = FuncInfo.StaticAllocaMap[Slot];
4522 MFI->setStackProtectorIndex(FI);
4524 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
4526 // Store the stack protector onto the stack.
4527 Res = DAG.getStore(getRoot(), getCurDebugLoc(), Src, FIN,
4528 MachinePointerInfo::getFixedStack(FI),
4534 case Intrinsic::objectsize: {
4535 // If we don't know by now, we're never going to know.
4536 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1));
4538 assert(CI && "Non-constant type in __builtin_object_size?");
4540 SDValue Arg = getValue(I.getCalledValue());
4541 EVT Ty = Arg.getValueType();
4544 Res = DAG.getConstant(-1ULL, Ty);
4546 Res = DAG.getConstant(0, Ty);
4551 case Intrinsic::var_annotation:
4552 // Discard annotate attributes
4555 case Intrinsic::init_trampoline: {
4556 const Function *F = cast<Function>(I.getArgOperand(1)->stripPointerCasts());
4560 Ops[1] = getValue(I.getArgOperand(0));
4561 Ops[2] = getValue(I.getArgOperand(1));
4562 Ops[3] = getValue(I.getArgOperand(2));
4563 Ops[4] = DAG.getSrcValue(I.getArgOperand(0));
4564 Ops[5] = DAG.getSrcValue(F);
4566 Res = DAG.getNode(ISD::TRAMPOLINE, dl,
4567 DAG.getVTList(TLI.getPointerTy(), MVT::Other),
4571 DAG.setRoot(Res.getValue(1));
4574 case Intrinsic::gcroot:
4576 const Value *Alloca = I.getArgOperand(0);
4577 const Constant *TypeMap = cast<Constant>(I.getArgOperand(1));
4579 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode());
4580 GFI->addStackRoot(FI->getIndex(), TypeMap);
4583 case Intrinsic::gcread:
4584 case Intrinsic::gcwrite:
4585 llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!");
4587 case Intrinsic::flt_rounds:
4588 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, dl, MVT::i32));
4590 case Intrinsic::trap:
4591 DAG.setRoot(DAG.getNode(ISD::TRAP, dl,MVT::Other, getRoot()));
4593 case Intrinsic::uadd_with_overflow:
4594 return implVisitAluOverflow(I, ISD::UADDO);
4595 case Intrinsic::sadd_with_overflow:
4596 return implVisitAluOverflow(I, ISD::SADDO);
4597 case Intrinsic::usub_with_overflow:
4598 return implVisitAluOverflow(I, ISD::USUBO);
4599 case Intrinsic::ssub_with_overflow:
4600 return implVisitAluOverflow(I, ISD::SSUBO);
4601 case Intrinsic::umul_with_overflow:
4602 return implVisitAluOverflow(I, ISD::UMULO);
4603 case Intrinsic::smul_with_overflow:
4604 return implVisitAluOverflow(I, ISD::SMULO);
4606 case Intrinsic::prefetch: {
4609 Ops[1] = getValue(I.getArgOperand(0));
4610 Ops[2] = getValue(I.getArgOperand(1));
4611 Ops[3] = getValue(I.getArgOperand(2));
4612 DAG.setRoot(DAG.getNode(ISD::PREFETCH, dl, MVT::Other, &Ops[0], 4));
4616 case Intrinsic::memory_barrier: {
4619 for (int x = 1; x < 6; ++x)
4620 Ops[x] = getValue(I.getArgOperand(x - 1));
4622 DAG.setRoot(DAG.getNode(ISD::MEMBARRIER, dl, MVT::Other, &Ops[0], 6));
4625 case Intrinsic::atomic_cmp_swap: {
4626 SDValue Root = getRoot();
4628 DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, getCurDebugLoc(),
4629 getValue(I.getArgOperand(1)).getValueType().getSimpleVT(),
4631 getValue(I.getArgOperand(0)),
4632 getValue(I.getArgOperand(1)),
4633 getValue(I.getArgOperand(2)),
4634 MachinePointerInfo(I.getArgOperand(0)));
4636 DAG.setRoot(L.getValue(1));
4639 case Intrinsic::atomic_load_add:
4640 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_ADD);
4641 case Intrinsic::atomic_load_sub:
4642 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_SUB);
4643 case Intrinsic::atomic_load_or:
4644 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_OR);
4645 case Intrinsic::atomic_load_xor:
4646 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_XOR);
4647 case Intrinsic::atomic_load_and:
4648 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_AND);
4649 case Intrinsic::atomic_load_nand:
4650 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_NAND);
4651 case Intrinsic::atomic_load_max:
4652 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MAX);
4653 case Intrinsic::atomic_load_min:
4654 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MIN);
4655 case Intrinsic::atomic_load_umin:
4656 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMIN);
4657 case Intrinsic::atomic_load_umax:
4658 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMAX);
4659 case Intrinsic::atomic_swap:
4660 return implVisitBinaryAtomic(I, ISD::ATOMIC_SWAP);
4662 case Intrinsic::invariant_start:
4663 case Intrinsic::lifetime_start:
4664 // Discard region information.
4665 setValue(&I, DAG.getUNDEF(TLI.getPointerTy()));
4667 case Intrinsic::invariant_end:
4668 case Intrinsic::lifetime_end:
4669 // Discard region information.
4674 void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee,
4676 MachineBasicBlock *LandingPad) {
4677 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
4678 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
4679 const Type *RetTy = FTy->getReturnType();
4680 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4681 MCSymbol *BeginLabel = 0;
4683 TargetLowering::ArgListTy Args;
4684 TargetLowering::ArgListEntry Entry;
4685 Args.reserve(CS.arg_size());
4687 // Check whether the function can return without sret-demotion.
4688 SmallVector<ISD::OutputArg, 4> Outs;
4689 SmallVector<uint64_t, 4> Offsets;
4690 GetReturnInfo(RetTy, CS.getAttributes().getRetAttributes(),
4691 Outs, TLI, &Offsets);
4693 bool CanLowerReturn = TLI.CanLowerReturn(CS.getCallingConv(),
4694 FTy->isVarArg(), Outs, FTy->getContext());
4696 SDValue DemoteStackSlot;
4697 int DemoteStackIdx = -100;
4699 if (!CanLowerReturn) {
4700 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(
4701 FTy->getReturnType());
4702 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(
4703 FTy->getReturnType());
4704 MachineFunction &MF = DAG.getMachineFunction();
4705 DemoteStackIdx = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
4706 const Type *StackSlotPtrType = PointerType::getUnqual(FTy->getReturnType());
4708 DemoteStackSlot = DAG.getFrameIndex(DemoteStackIdx, TLI.getPointerTy());
4709 Entry.Node = DemoteStackSlot;
4710 Entry.Ty = StackSlotPtrType;
4711 Entry.isSExt = false;
4712 Entry.isZExt = false;
4713 Entry.isInReg = false;
4714 Entry.isSRet = true;
4715 Entry.isNest = false;
4716 Entry.isByVal = false;
4717 Entry.Alignment = Align;
4718 Args.push_back(Entry);
4719 RetTy = Type::getVoidTy(FTy->getContext());
4722 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
4724 SDValue ArgNode = getValue(*i);
4725 Entry.Node = ArgNode; Entry.Ty = (*i)->getType();
4727 unsigned attrInd = i - CS.arg_begin() + 1;
4728 Entry.isSExt = CS.paramHasAttr(attrInd, Attribute::SExt);
4729 Entry.isZExt = CS.paramHasAttr(attrInd, Attribute::ZExt);
4730 Entry.isInReg = CS.paramHasAttr(attrInd, Attribute::InReg);
4731 Entry.isSRet = CS.paramHasAttr(attrInd, Attribute::StructRet);
4732 Entry.isNest = CS.paramHasAttr(attrInd, Attribute::Nest);
4733 Entry.isByVal = CS.paramHasAttr(attrInd, Attribute::ByVal);
4734 Entry.Alignment = CS.getParamAlignment(attrInd);
4735 Args.push_back(Entry);
4739 // Insert a label before the invoke call to mark the try range. This can be
4740 // used to detect deletion of the invoke via the MachineModuleInfo.
4741 BeginLabel = MMI.getContext().CreateTempSymbol();
4743 // For SjLj, keep track of which landing pads go with which invokes
4744 // so as to maintain the ordering of pads in the LSDA.
4745 unsigned CallSiteIndex = MMI.getCurrentCallSite();
4746 if (CallSiteIndex) {
4747 MMI.setCallSiteBeginLabel(BeginLabel, CallSiteIndex);
4748 // Now that the call site is handled, stop tracking it.
4749 MMI.setCurrentCallSite(0);
4752 // Both PendingLoads and PendingExports must be flushed here;
4753 // this call might not return.
4755 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getControlRoot(), BeginLabel));
4758 // Check if target-independent constraints permit a tail call here.
4759 // Target-dependent constraints are checked within TLI.LowerCallTo.
4761 !isInTailCallPosition(CS, CS.getAttributes().getRetAttributes(), TLI))
4764 // If there's a possibility that fast-isel has already selected some amount
4765 // of the current basic block, don't emit a tail call.
4766 if (isTailCall && EnableFastISel)
4769 std::pair<SDValue,SDValue> Result =
4770 TLI.LowerCallTo(getRoot(), RetTy,
4771 CS.paramHasAttr(0, Attribute::SExt),
4772 CS.paramHasAttr(0, Attribute::ZExt), FTy->isVarArg(),
4773 CS.paramHasAttr(0, Attribute::InReg), FTy->getNumParams(),
4774 CS.getCallingConv(),
4776 !CS.getInstruction()->use_empty(),
4777 Callee, Args, DAG, getCurDebugLoc());
4778 assert((isTailCall || Result.second.getNode()) &&
4779 "Non-null chain expected with non-tail call!");
4780 assert((Result.second.getNode() || !Result.first.getNode()) &&
4781 "Null value expected with tail call!");
4782 if (Result.first.getNode()) {
4783 setValue(CS.getInstruction(), Result.first);
4784 } else if (!CanLowerReturn && Result.second.getNode()) {
4785 // The instruction result is the result of loading from the
4786 // hidden sret parameter.
4787 SmallVector<EVT, 1> PVTs;
4788 const Type *PtrRetTy = PointerType::getUnqual(FTy->getReturnType());
4790 ComputeValueVTs(TLI, PtrRetTy, PVTs);
4791 assert(PVTs.size() == 1 && "Pointers should fit in one register");
4792 EVT PtrVT = PVTs[0];
4793 unsigned NumValues = Outs.size();
4794 SmallVector<SDValue, 4> Values(NumValues);
4795 SmallVector<SDValue, 4> Chains(NumValues);
4797 for (unsigned i = 0; i < NumValues; ++i) {
4798 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT,
4800 DAG.getConstant(Offsets[i], PtrVT));
4801 SDValue L = DAG.getLoad(Outs[i].VT, getCurDebugLoc(), Result.second,
4803 MachinePointerInfo::getFixedStack(DemoteStackIdx, Offsets[i]),
4806 Chains[i] = L.getValue(1);
4809 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
4810 MVT::Other, &Chains[0], NumValues);
4811 PendingLoads.push_back(Chain);
4813 // Collect the legal value parts into potentially illegal values
4814 // that correspond to the original function's return values.
4815 SmallVector<EVT, 4> RetTys;
4816 RetTy = FTy->getReturnType();
4817 ComputeValueVTs(TLI, RetTy, RetTys);
4818 ISD::NodeType AssertOp = ISD::DELETED_NODE;
4819 SmallVector<SDValue, 4> ReturnValues;
4820 unsigned CurReg = 0;
4821 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
4823 EVT RegisterVT = TLI.getRegisterType(RetTy->getContext(), VT);
4824 unsigned NumRegs = TLI.getNumRegisters(RetTy->getContext(), VT);
4826 SDValue ReturnValue =
4827 getCopyFromParts(DAG, getCurDebugLoc(), &Values[CurReg], NumRegs,
4828 RegisterVT, VT, AssertOp);
4829 ReturnValues.push_back(ReturnValue);
4833 setValue(CS.getInstruction(),
4834 DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
4835 DAG.getVTList(&RetTys[0], RetTys.size()),
4836 &ReturnValues[0], ReturnValues.size()));
4840 // As a special case, a null chain means that a tail call has been emitted and
4841 // the DAG root is already updated.
4842 if (Result.second.getNode())
4843 DAG.setRoot(Result.second);
4848 // Insert a label at the end of the invoke call to mark the try range. This
4849 // can be used to detect deletion of the invoke via the MachineModuleInfo.
4850 MCSymbol *EndLabel = MMI.getContext().CreateTempSymbol();
4851 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getRoot(), EndLabel));
4853 // Inform MachineModuleInfo of range.
4854 MMI.addInvoke(LandingPad, BeginLabel, EndLabel);
4858 /// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the
4859 /// value is equal or not-equal to zero.
4860 static bool IsOnlyUsedInZeroEqualityComparison(const Value *V) {
4861 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end();
4863 if (const ICmpInst *IC = dyn_cast<ICmpInst>(*UI))
4864 if (IC->isEquality())
4865 if (const Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
4866 if (C->isNullValue())
4868 // Unknown instruction.
4874 static SDValue getMemCmpLoad(const Value *PtrVal, MVT LoadVT,
4876 SelectionDAGBuilder &Builder) {
4878 // Check to see if this load can be trivially constant folded, e.g. if the
4879 // input is from a string literal.
4880 if (const Constant *LoadInput = dyn_cast<Constant>(PtrVal)) {
4881 // Cast pointer to the type we really want to load.
4882 LoadInput = ConstantExpr::getBitCast(const_cast<Constant *>(LoadInput),
4883 PointerType::getUnqual(LoadTy));
4885 if (const Constant *LoadCst =
4886 ConstantFoldLoadFromConstPtr(const_cast<Constant *>(LoadInput),
4888 return Builder.getValue(LoadCst);
4891 // Otherwise, we have to emit the load. If the pointer is to unfoldable but
4892 // still constant memory, the input chain can be the entry node.
4894 bool ConstantMemory = false;
4896 // Do not serialize (non-volatile) loads of constant memory with anything.
4897 if (Builder.AA->pointsToConstantMemory(PtrVal)) {
4898 Root = Builder.DAG.getEntryNode();
4899 ConstantMemory = true;
4901 // Do not serialize non-volatile loads against each other.
4902 Root = Builder.DAG.getRoot();
4905 SDValue Ptr = Builder.getValue(PtrVal);
4906 SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurDebugLoc(), Root,
4907 Ptr, MachinePointerInfo(PtrVal),
4909 false /*nontemporal*/, 1 /* align=1 */);
4911 if (!ConstantMemory)
4912 Builder.PendingLoads.push_back(LoadVal.getValue(1));
4917 /// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form.
4918 /// If so, return true and lower it, otherwise return false and it will be
4919 /// lowered like a normal call.
4920 bool SelectionDAGBuilder::visitMemCmpCall(const CallInst &I) {
4921 // Verify that the prototype makes sense. int memcmp(void*,void*,size_t)
4922 if (I.getNumArgOperands() != 3)
4925 const Value *LHS = I.getArgOperand(0), *RHS = I.getArgOperand(1);
4926 if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() ||
4927 !I.getArgOperand(2)->getType()->isIntegerTy() ||
4928 !I.getType()->isIntegerTy())
4931 const ConstantInt *Size = dyn_cast<ConstantInt>(I.getArgOperand(2));
4933 // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0
4934 // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0
4935 if (Size && IsOnlyUsedInZeroEqualityComparison(&I)) {
4936 bool ActuallyDoIt = true;
4939 switch (Size->getZExtValue()) {
4941 LoadVT = MVT::Other;
4943 ActuallyDoIt = false;
4947 LoadTy = Type::getInt16Ty(Size->getContext());
4951 LoadTy = Type::getInt32Ty(Size->getContext());
4955 LoadTy = Type::getInt64Ty(Size->getContext());
4959 LoadVT = MVT::v4i32;
4960 LoadTy = Type::getInt32Ty(Size->getContext());
4961 LoadTy = VectorType::get(LoadTy, 4);
4966 // This turns into unaligned loads. We only do this if the target natively
4967 // supports the MVT we'll be loading or if it is small enough (<= 4) that
4968 // we'll only produce a small number of byte loads.
4970 // Require that we can find a legal MVT, and only do this if the target
4971 // supports unaligned loads of that type. Expanding into byte loads would
4973 if (ActuallyDoIt && Size->getZExtValue() > 4) {
4974 // TODO: Handle 5 byte compare as 4-byte + 1 byte.
4975 // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads.
4976 if (!TLI.isTypeLegal(LoadVT) ||!TLI.allowsUnalignedMemoryAccesses(LoadVT))
4977 ActuallyDoIt = false;
4981 SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this);
4982 SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this);
4984 SDValue Res = DAG.getSetCC(getCurDebugLoc(), MVT::i1, LHSVal, RHSVal,
4986 EVT CallVT = TLI.getValueType(I.getType(), true);
4987 setValue(&I, DAG.getZExtOrTrunc(Res, getCurDebugLoc(), CallVT));
4997 void SelectionDAGBuilder::visitCall(const CallInst &I) {
4998 // Handle inline assembly differently.
4999 if (isa<InlineAsm>(I.getCalledValue())) {
5004 const char *RenameFn = 0;
5005 if (Function *F = I.getCalledFunction()) {
5006 if (F->isDeclaration()) {
5007 if (const TargetIntrinsicInfo *II = TM.getIntrinsicInfo()) {
5008 if (unsigned IID = II->getIntrinsicID(F)) {
5009 RenameFn = visitIntrinsicCall(I, IID);
5014 if (unsigned IID = F->getIntrinsicID()) {
5015 RenameFn = visitIntrinsicCall(I, IID);
5021 // Check for well-known libc/libm calls. If the function is internal, it
5022 // can't be a library call.
5023 if (!F->hasLocalLinkage() && F->hasName()) {
5024 StringRef Name = F->getName();
5025 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl") {
5026 if (I.getNumArgOperands() == 2 && // Basic sanity checks.
5027 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5028 I.getType() == I.getArgOperand(0)->getType() &&
5029 I.getType() == I.getArgOperand(1)->getType()) {
5030 SDValue LHS = getValue(I.getArgOperand(0));
5031 SDValue RHS = getValue(I.getArgOperand(1));
5032 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurDebugLoc(),
5033 LHS.getValueType(), LHS, RHS));
5036 } else if (Name == "fabs" || Name == "fabsf" || Name == "fabsl") {
5037 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5038 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5039 I.getType() == I.getArgOperand(0)->getType()) {
5040 SDValue Tmp = getValue(I.getArgOperand(0));
5041 setValue(&I, DAG.getNode(ISD::FABS, getCurDebugLoc(),
5042 Tmp.getValueType(), Tmp));
5045 } else if (Name == "sin" || Name == "sinf" || Name == "sinl") {
5046 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5047 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5048 I.getType() == I.getArgOperand(0)->getType() &&
5049 I.onlyReadsMemory()) {
5050 SDValue Tmp = getValue(I.getArgOperand(0));
5051 setValue(&I, DAG.getNode(ISD::FSIN, getCurDebugLoc(),
5052 Tmp.getValueType(), Tmp));
5055 } else if (Name == "cos" || Name == "cosf" || Name == "cosl") {
5056 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5057 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5058 I.getType() == I.getArgOperand(0)->getType() &&
5059 I.onlyReadsMemory()) {
5060 SDValue Tmp = getValue(I.getArgOperand(0));
5061 setValue(&I, DAG.getNode(ISD::FCOS, getCurDebugLoc(),
5062 Tmp.getValueType(), Tmp));
5065 } else if (Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl") {
5066 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5067 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5068 I.getType() == I.getArgOperand(0)->getType() &&
5069 I.onlyReadsMemory()) {
5070 SDValue Tmp = getValue(I.getArgOperand(0));
5071 setValue(&I, DAG.getNode(ISD::FSQRT, getCurDebugLoc(),
5072 Tmp.getValueType(), Tmp));
5075 } else if (Name == "memcmp") {
5076 if (visitMemCmpCall(I))
5084 Callee = getValue(I.getCalledValue());
5086 Callee = DAG.getExternalSymbol(RenameFn, TLI.getPointerTy());
5088 // Check if we can potentially perform a tail call. More detailed checking is
5089 // be done within LowerCallTo, after more information about the call is known.
5090 LowerCallTo(&I, Callee, I.isTailCall());
5095 /// AsmOperandInfo - This contains information for each constraint that we are
5097 class LLVM_LIBRARY_VISIBILITY SDISelAsmOperandInfo :
5098 public TargetLowering::AsmOperandInfo {
5100 /// CallOperand - If this is the result output operand or a clobber
5101 /// this is null, otherwise it is the incoming operand to the CallInst.
5102 /// This gets modified as the asm is processed.
5103 SDValue CallOperand;
5105 /// AssignedRegs - If this is a register or register class operand, this
5106 /// contains the set of register corresponding to the operand.
5107 RegsForValue AssignedRegs;
5109 explicit SDISelAsmOperandInfo(const TargetLowering::AsmOperandInfo &info)
5110 : TargetLowering::AsmOperandInfo(info), CallOperand(0,0) {
5113 /// MarkAllocatedRegs - Once AssignedRegs is set, mark the assigned registers
5114 /// busy in OutputRegs/InputRegs.
5115 void MarkAllocatedRegs(bool isOutReg, bool isInReg,
5116 std::set<unsigned> &OutputRegs,
5117 std::set<unsigned> &InputRegs,
5118 const TargetRegisterInfo &TRI) const {
5120 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
5121 MarkRegAndAliases(AssignedRegs.Regs[i], OutputRegs, TRI);
5124 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
5125 MarkRegAndAliases(AssignedRegs.Regs[i], InputRegs, TRI);
5129 /// getCallOperandValEVT - Return the EVT of the Value* that this operand
5130 /// corresponds to. If there is no Value* for this operand, it returns
5132 EVT getCallOperandValEVT(LLVMContext &Context,
5133 const TargetLowering &TLI,
5134 const TargetData *TD) const {
5135 if (CallOperandVal == 0) return MVT::Other;
5137 if (isa<BasicBlock>(CallOperandVal))
5138 return TLI.getPointerTy();
5140 const llvm::Type *OpTy = CallOperandVal->getType();
5142 // If this is an indirect operand, the operand is a pointer to the
5145 const llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
5147 report_fatal_error("Indirect operand for inline asm not a pointer!");
5148 OpTy = PtrTy->getElementType();
5151 // If OpTy is not a single value, it may be a struct/union that we
5152 // can tile with integers.
5153 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
5154 unsigned BitSize = TD->getTypeSizeInBits(OpTy);
5163 OpTy = IntegerType::get(Context, BitSize);
5168 return TLI.getValueType(OpTy, true);
5172 /// MarkRegAndAliases - Mark the specified register and all aliases in the
5174 static void MarkRegAndAliases(unsigned Reg, std::set<unsigned> &Regs,
5175 const TargetRegisterInfo &TRI) {
5176 assert(TargetRegisterInfo::isPhysicalRegister(Reg) && "Isn't a physreg");
5178 if (const unsigned *Aliases = TRI.getAliasSet(Reg))
5179 for (; *Aliases; ++Aliases)
5180 Regs.insert(*Aliases);
5184 } // end llvm namespace.
5186 /// isAllocatableRegister - If the specified register is safe to allocate,
5187 /// i.e. it isn't a stack pointer or some other special register, return the
5188 /// register class for the register. Otherwise, return null.
5189 static const TargetRegisterClass *
5190 isAllocatableRegister(unsigned Reg, MachineFunction &MF,
5191 const TargetLowering &TLI,
5192 const TargetRegisterInfo *TRI) {
5193 EVT FoundVT = MVT::Other;
5194 const TargetRegisterClass *FoundRC = 0;
5195 for (TargetRegisterInfo::regclass_iterator RCI = TRI->regclass_begin(),
5196 E = TRI->regclass_end(); RCI != E; ++RCI) {
5197 EVT ThisVT = MVT::Other;
5199 const TargetRegisterClass *RC = *RCI;
5200 // If none of the value types for this register class are valid, we
5201 // can't use it. For example, 64-bit reg classes on 32-bit targets.
5202 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
5204 if (TLI.isTypeLegal(*I)) {
5205 // If we have already found this register in a different register class,
5206 // choose the one with the largest VT specified. For example, on
5207 // PowerPC, we favor f64 register classes over f32.
5208 if (FoundVT == MVT::Other || FoundVT.bitsLT(*I)) {
5215 if (ThisVT == MVT::Other) continue;
5217 // NOTE: This isn't ideal. In particular, this might allocate the
5218 // frame pointer in functions that need it (due to them not being taken
5219 // out of allocation, because a variable sized allocation hasn't been seen
5220 // yet). This is a slight code pessimization, but should still work.
5221 for (TargetRegisterClass::iterator I = RC->allocation_order_begin(MF),
5222 E = RC->allocation_order_end(MF); I != E; ++I)
5224 // We found a matching register class. Keep looking at others in case
5225 // we find one with larger registers that this physreg is also in.
5234 /// GetRegistersForValue - Assign registers (virtual or physical) for the
5235 /// specified operand. We prefer to assign virtual registers, to allow the
5236 /// register allocator to handle the assignment process. However, if the asm
5237 /// uses features that we can't model on machineinstrs, we have SDISel do the
5238 /// allocation. This produces generally horrible, but correct, code.
5240 /// OpInfo describes the operand.
5241 /// Input and OutputRegs are the set of already allocated physical registers.
5243 void SelectionDAGBuilder::
5244 GetRegistersForValue(SDISelAsmOperandInfo &OpInfo,
5245 std::set<unsigned> &OutputRegs,
5246 std::set<unsigned> &InputRegs) {
5247 LLVMContext &Context = FuncInfo.Fn->getContext();
5249 // Compute whether this value requires an input register, an output register,
5251 bool isOutReg = false;
5252 bool isInReg = false;
5253 switch (OpInfo.Type) {
5254 case InlineAsm::isOutput:
5257 // If there is an input constraint that matches this, we need to reserve
5258 // the input register so no other inputs allocate to it.
5259 isInReg = OpInfo.hasMatchingInput();
5261 case InlineAsm::isInput:
5265 case InlineAsm::isClobber:
5272 MachineFunction &MF = DAG.getMachineFunction();
5273 SmallVector<unsigned, 4> Regs;
5275 // If this is a constraint for a single physreg, or a constraint for a
5276 // register class, find it.
5277 std::pair<unsigned, const TargetRegisterClass*> PhysReg =
5278 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
5279 OpInfo.ConstraintVT);
5281 unsigned NumRegs = 1;
5282 if (OpInfo.ConstraintVT != MVT::Other) {
5283 // If this is a FP input in an integer register (or visa versa) insert a bit
5284 // cast of the input value. More generally, handle any case where the input
5285 // value disagrees with the register class we plan to stick this in.
5286 if (OpInfo.Type == InlineAsm::isInput &&
5287 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) {
5288 // Try to convert to the first EVT that the reg class contains. If the
5289 // types are identical size, use a bitcast to convert (e.g. two differing
5291 EVT RegVT = *PhysReg.second->vt_begin();
5292 if (RegVT.getSizeInBits() == OpInfo.ConstraintVT.getSizeInBits()) {
5293 OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
5294 RegVT, OpInfo.CallOperand);
5295 OpInfo.ConstraintVT = RegVT;
5296 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) {
5297 // If the input is a FP value and we want it in FP registers, do a
5298 // bitcast to the corresponding integer type. This turns an f64 value
5299 // into i64, which can be passed with two i32 values on a 32-bit
5301 RegVT = EVT::getIntegerVT(Context,
5302 OpInfo.ConstraintVT.getSizeInBits());
5303 OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
5304 RegVT, OpInfo.CallOperand);
5305 OpInfo.ConstraintVT = RegVT;
5309 NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT);
5313 EVT ValueVT = OpInfo.ConstraintVT;
5315 // If this is a constraint for a specific physical register, like {r17},
5317 if (unsigned AssignedReg = PhysReg.first) {
5318 const TargetRegisterClass *RC = PhysReg.second;
5319 if (OpInfo.ConstraintVT == MVT::Other)
5320 ValueVT = *RC->vt_begin();
5322 // Get the actual register value type. This is important, because the user
5323 // may have asked for (e.g.) the AX register in i32 type. We need to
5324 // remember that AX is actually i16 to get the right extension.
5325 RegVT = *RC->vt_begin();
5327 // This is a explicit reference to a physical register.
5328 Regs.push_back(AssignedReg);
5330 // If this is an expanded reference, add the rest of the regs to Regs.
5332 TargetRegisterClass::iterator I = RC->begin();
5333 for (; *I != AssignedReg; ++I)
5334 assert(I != RC->end() && "Didn't find reg!");
5336 // Already added the first reg.
5338 for (; NumRegs; --NumRegs, ++I) {
5339 assert(I != RC->end() && "Ran out of registers to allocate!");
5344 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5345 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
5346 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
5350 // Otherwise, if this was a reference to an LLVM register class, create vregs
5351 // for this reference.
5352 if (const TargetRegisterClass *RC = PhysReg.second) {
5353 RegVT = *RC->vt_begin();
5354 if (OpInfo.ConstraintVT == MVT::Other)
5357 // Create the appropriate number of virtual registers.
5358 MachineRegisterInfo &RegInfo = MF.getRegInfo();
5359 for (; NumRegs; --NumRegs)
5360 Regs.push_back(RegInfo.createVirtualRegister(RC));
5362 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5366 // This is a reference to a register class that doesn't directly correspond
5367 // to an LLVM register class. Allocate NumRegs consecutive, available,
5368 // registers from the class.
5369 std::vector<unsigned> RegClassRegs
5370 = TLI.getRegClassForInlineAsmConstraint(OpInfo.ConstraintCode,
5371 OpInfo.ConstraintVT);
5373 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
5374 unsigned NumAllocated = 0;
5375 for (unsigned i = 0, e = RegClassRegs.size(); i != e; ++i) {
5376 unsigned Reg = RegClassRegs[i];
5377 // See if this register is available.
5378 if ((isOutReg && OutputRegs.count(Reg)) || // Already used.
5379 (isInReg && InputRegs.count(Reg))) { // Already used.
5380 // Make sure we find consecutive registers.
5385 // Check to see if this register is allocatable (i.e. don't give out the
5387 const TargetRegisterClass *RC = isAllocatableRegister(Reg, MF, TLI, TRI);
5388 if (!RC) { // Couldn't allocate this register.
5389 // Reset NumAllocated to make sure we return consecutive registers.
5394 // Okay, this register is good, we can use it.
5397 // If we allocated enough consecutive registers, succeed.
5398 if (NumAllocated == NumRegs) {
5399 unsigned RegStart = (i-NumAllocated)+1;
5400 unsigned RegEnd = i+1;
5401 // Mark all of the allocated registers used.
5402 for (unsigned i = RegStart; i != RegEnd; ++i)
5403 Regs.push_back(RegClassRegs[i]);
5405 OpInfo.AssignedRegs = RegsForValue(Regs, *RC->vt_begin(),
5406 OpInfo.ConstraintVT);
5407 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
5412 // Otherwise, we couldn't allocate enough registers for this.
5415 /// visitInlineAsm - Handle a call to an InlineAsm object.
5417 void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) {
5418 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
5420 /// ConstraintOperands - Information about all of the constraints.
5421 std::vector<SDISelAsmOperandInfo> ConstraintOperands;
5423 std::set<unsigned> OutputRegs, InputRegs;
5425 std::vector<TargetLowering::AsmOperandInfo> TargetConstraints = TLI.ParseConstraints(CS);
5426 bool hasMemory = false;
5428 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
5429 unsigned ResNo = 0; // ResNo - The result number of the next output.
5430 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
5431 ConstraintOperands.push_back(SDISelAsmOperandInfo(TargetConstraints[i]));
5432 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
5434 EVT OpVT = MVT::Other;
5436 // Compute the value type for each operand.
5437 switch (OpInfo.Type) {
5438 case InlineAsm::isOutput:
5439 // Indirect outputs just consume an argument.
5440 if (OpInfo.isIndirect) {
5441 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5445 // The return value of the call is this value. As such, there is no
5446 // corresponding argument.
5447 assert(!CS.getType()->isVoidTy() &&
5449 if (const StructType *STy = dyn_cast<StructType>(CS.getType())) {
5450 OpVT = TLI.getValueType(STy->getElementType(ResNo));
5452 assert(ResNo == 0 && "Asm only has one result!");
5453 OpVT = TLI.getValueType(CS.getType());
5457 case InlineAsm::isInput:
5458 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5460 case InlineAsm::isClobber:
5465 // If this is an input or an indirect output, process the call argument.
5466 // BasicBlocks are labels, currently appearing only in asm's.
5467 if (OpInfo.CallOperandVal) {
5468 // Strip bitcasts, if any. This mostly comes up for functions.
5469 OpInfo.CallOperandVal = OpInfo.CallOperandVal->stripPointerCasts();
5471 if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) {
5472 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]);
5474 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
5477 OpVT = OpInfo.getCallOperandValEVT(*DAG.getContext(), TLI, TD);
5480 OpInfo.ConstraintVT = OpVT;
5482 // Indirect operand accesses access memory.
5483 if (OpInfo.isIndirect)
5486 for (unsigned j = 0, ee = OpInfo.Codes.size(); j != ee; ++j) {
5487 TargetLowering::ConstraintType CType = TLI.getConstraintType(OpInfo.Codes[j]);
5488 if (CType == TargetLowering::C_Memory) {
5496 SDValue Chain, Flag;
5498 // We won't need to flush pending loads if this asm doesn't touch
5499 // memory and is nonvolatile.
5500 if (hasMemory || IA->hasSideEffects())
5503 Chain = DAG.getRoot();
5505 // Second pass over the constraints: compute which constraint option to use
5506 // and assign registers to constraints that want a specific physreg.
5507 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5508 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5510 // If this is an output operand with a matching input operand, look up the
5511 // matching input. If their types mismatch, e.g. one is an integer, the
5512 // other is floating point, or their sizes are different, flag it as an
5514 if (OpInfo.hasMatchingInput()) {
5515 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
5517 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
5518 if ((OpInfo.ConstraintVT.isInteger() !=
5519 Input.ConstraintVT.isInteger()) ||
5520 (OpInfo.ConstraintVT.getSizeInBits() !=
5521 Input.ConstraintVT.getSizeInBits())) {
5522 report_fatal_error("Unsupported asm: input constraint"
5523 " with a matching output constraint of"
5524 " incompatible type!");
5526 Input.ConstraintVT = OpInfo.ConstraintVT;
5530 // Compute the constraint code and ConstraintType to use.
5531 TLI.ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, &DAG);
5533 // If this is a memory input, and if the operand is not indirect, do what we
5534 // need to to provide an address for the memory input.
5535 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5536 !OpInfo.isIndirect) {
5537 assert((OpInfo.isMultipleAlternative || (OpInfo.Type == InlineAsm::isInput)) &&
5538 "Can only indirectify direct input operands!");
5540 // Memory operands really want the address of the value. If we don't have
5541 // an indirect input, put it in the constpool if we can, otherwise spill
5542 // it to a stack slot.
5544 // If the operand is a float, integer, or vector constant, spill to a
5545 // constant pool entry to get its address.
5546 const Value *OpVal = OpInfo.CallOperandVal;
5547 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
5548 isa<ConstantVector>(OpVal)) {
5549 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal),
5550 TLI.getPointerTy());
5552 // Otherwise, create a stack slot and emit a store to it before the
5554 const Type *Ty = OpVal->getType();
5555 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
5556 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(Ty);
5557 MachineFunction &MF = DAG.getMachineFunction();
5558 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
5559 SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
5560 Chain = DAG.getStore(Chain, getCurDebugLoc(),
5561 OpInfo.CallOperand, StackSlot,
5562 MachinePointerInfo::getFixedStack(SSFI),
5564 OpInfo.CallOperand = StackSlot;
5567 // There is no longer a Value* corresponding to this operand.
5568 OpInfo.CallOperandVal = 0;
5570 // It is now an indirect operand.
5571 OpInfo.isIndirect = true;
5574 // If this constraint is for a specific register, allocate it before
5576 if (OpInfo.ConstraintType == TargetLowering::C_Register)
5577 GetRegistersForValue(OpInfo, OutputRegs, InputRegs);
5580 // Second pass - Loop over all of the operands, assigning virtual or physregs
5581 // to register class operands.
5582 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5583 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5585 // C_Register operands have already been allocated, Other/Memory don't need
5587 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
5588 GetRegistersForValue(OpInfo, OutputRegs, InputRegs);
5591 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
5592 std::vector<SDValue> AsmNodeOperands;
5593 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain
5594 AsmNodeOperands.push_back(
5595 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(),
5596 TLI.getPointerTy()));
5598 // If we have a !srcloc metadata node associated with it, we want to attach
5599 // this to the ultimately generated inline asm machineinstr. To do this, we
5600 // pass in the third operand as this (potentially null) inline asm MDNode.
5601 const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc");
5602 AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc));
5604 // Remember the AlignStack bit as operand 3.
5605 AsmNodeOperands.push_back(DAG.getTargetConstant(IA->isAlignStack() ? 1 : 0,
5608 // Loop over all of the inputs, copying the operand values into the
5609 // appropriate registers and processing the output regs.
5610 RegsForValue RetValRegs;
5612 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
5613 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
5615 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5616 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5618 switch (OpInfo.Type) {
5619 case InlineAsm::isOutput: {
5620 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
5621 OpInfo.ConstraintType != TargetLowering::C_Register) {
5622 // Memory output, or 'other' output (e.g. 'X' constraint).
5623 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
5625 // Add information to the INLINEASM node to know about this output.
5626 unsigned OpFlags = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
5627 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlags,
5628 TLI.getPointerTy()));
5629 AsmNodeOperands.push_back(OpInfo.CallOperand);
5633 // Otherwise, this is a register or register class output.
5635 // Copy the output from the appropriate register. Find a register that
5637 if (OpInfo.AssignedRegs.Regs.empty())
5638 report_fatal_error("Couldn't allocate output reg for constraint '" +
5639 Twine(OpInfo.ConstraintCode) + "'!");
5641 // If this is an indirect operand, store through the pointer after the
5643 if (OpInfo.isIndirect) {
5644 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
5645 OpInfo.CallOperandVal));
5647 // This is the result value of the call.
5648 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
5649 // Concatenate this output onto the outputs list.
5650 RetValRegs.append(OpInfo.AssignedRegs);
5653 // Add information to the INLINEASM node to know that this register is
5655 OpInfo.AssignedRegs.AddInlineAsmOperands(OpInfo.isEarlyClobber ?
5656 InlineAsm::Kind_RegDefEarlyClobber :
5657 InlineAsm::Kind_RegDef,
5664 case InlineAsm::isInput: {
5665 SDValue InOperandVal = OpInfo.CallOperand;
5667 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint?
5668 // If this is required to match an output register we have already set,
5669 // just use its register.
5670 unsigned OperandNo = OpInfo.getMatchedOperand();
5672 // Scan until we find the definition we already emitted of this operand.
5673 // When we find it, create a RegsForValue operand.
5674 unsigned CurOp = InlineAsm::Op_FirstOperand;
5675 for (; OperandNo; --OperandNo) {
5676 // Advance to the next operand.
5678 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
5679 assert((InlineAsm::isRegDefKind(OpFlag) ||
5680 InlineAsm::isRegDefEarlyClobberKind(OpFlag) ||
5681 InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?");
5682 CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1;
5686 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
5687 if (InlineAsm::isRegDefKind(OpFlag) ||
5688 InlineAsm::isRegDefEarlyClobberKind(OpFlag)) {
5689 // Add (OpFlag&0xffff)>>3 registers to MatchedRegs.
5690 if (OpInfo.isIndirect) {
5691 // This happens on gcc/testsuite/gcc.dg/pr8788-1.c
5692 LLVMContext &Ctx = *DAG.getContext();
5693 Ctx.emitError(CS.getInstruction(), "inline asm not supported yet:"
5694 " don't know how to handle tied "
5695 "indirect register inputs");
5698 RegsForValue MatchedRegs;
5699 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
5700 EVT RegVT = AsmNodeOperands[CurOp+1].getValueType();
5701 MatchedRegs.RegVTs.push_back(RegVT);
5702 MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo();
5703 for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag);
5705 MatchedRegs.Regs.push_back
5706 (RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT)));
5708 // Use the produced MatchedRegs object to
5709 MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
5711 MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse,
5712 true, OpInfo.getMatchedOperand(),
5713 DAG, AsmNodeOperands);
5717 assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!");
5718 assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 &&
5719 "Unexpected number of operands");
5720 // Add information to the INLINEASM node to know about this input.
5721 // See InlineAsm.h isUseOperandTiedToDef.
5722 OpFlag = InlineAsm::getFlagWordForMatchingOp(OpFlag,
5723 OpInfo.getMatchedOperand());
5724 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlag,
5725 TLI.getPointerTy()));
5726 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]);
5730 // Treat indirect 'X' constraint as memory.
5731 if (OpInfo.ConstraintType == TargetLowering::C_Other &&
5733 OpInfo.ConstraintType = TargetLowering::C_Memory;
5735 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
5736 std::vector<SDValue> Ops;
5737 TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode[0],
5740 report_fatal_error("Invalid operand for inline asm constraint '" +
5741 Twine(OpInfo.ConstraintCode) + "'!");
5743 // Add information to the INLINEASM node to know about this input.
5744 unsigned ResOpType =
5745 InlineAsm::getFlagWord(InlineAsm::Kind_Imm, Ops.size());
5746 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
5747 TLI.getPointerTy()));
5748 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
5752 if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
5753 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
5754 assert(InOperandVal.getValueType() == TLI.getPointerTy() &&
5755 "Memory operands expect pointer values");
5757 // Add information to the INLINEASM node to know about this input.
5758 unsigned ResOpType = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
5759 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
5760 TLI.getPointerTy()));
5761 AsmNodeOperands.push_back(InOperandVal);
5765 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
5766 OpInfo.ConstraintType == TargetLowering::C_Register) &&
5767 "Unknown constraint type!");
5768 assert(!OpInfo.isIndirect &&
5769 "Don't know how to handle indirect register inputs yet!");
5771 // Copy the input into the appropriate registers.
5772 if (OpInfo.AssignedRegs.Regs.empty() ||
5773 !OpInfo.AssignedRegs.areValueTypesLegal(TLI))
5774 report_fatal_error("Couldn't allocate input reg for constraint '" +
5775 Twine(OpInfo.ConstraintCode) + "'!");
5777 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
5780 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, false, 0,
5781 DAG, AsmNodeOperands);
5784 case InlineAsm::isClobber: {
5785 // Add the clobbered value to the operand list, so that the register
5786 // allocator is aware that the physreg got clobbered.
5787 if (!OpInfo.AssignedRegs.Regs.empty())
5788 OpInfo.AssignedRegs.AddInlineAsmOperands(
5789 InlineAsm::Kind_RegDefEarlyClobber,
5797 // Finish up input operands. Set the input chain and add the flag last.
5798 AsmNodeOperands[InlineAsm::Op_InputChain] = Chain;
5799 if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
5801 Chain = DAG.getNode(ISD::INLINEASM, getCurDebugLoc(),
5802 DAG.getVTList(MVT::Other, MVT::Flag),
5803 &AsmNodeOperands[0], AsmNodeOperands.size());
5804 Flag = Chain.getValue(1);
5806 // If this asm returns a register value, copy the result from that register
5807 // and set it as the value of the call.
5808 if (!RetValRegs.Regs.empty()) {
5809 SDValue Val = RetValRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(),
5812 // FIXME: Why don't we do this for inline asms with MRVs?
5813 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) {
5814 EVT ResultType = TLI.getValueType(CS.getType());
5816 // If any of the results of the inline asm is a vector, it may have the
5817 // wrong width/num elts. This can happen for register classes that can
5818 // contain multiple different value types. The preg or vreg allocated may
5819 // not have the same VT as was expected. Convert it to the right type
5820 // with bit_convert.
5821 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) {
5822 Val = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
5825 } else if (ResultType != Val.getValueType() &&
5826 ResultType.isInteger() && Val.getValueType().isInteger()) {
5827 // If a result value was tied to an input value, the computed result may
5828 // have a wider width than the expected result. Extract the relevant
5830 Val = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), ResultType, Val);
5833 assert(ResultType == Val.getValueType() && "Asm result value mismatch!");
5836 setValue(CS.getInstruction(), Val);
5837 // Don't need to use this as a chain in this case.
5838 if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty())
5842 std::vector<std::pair<SDValue, const Value *> > StoresToEmit;
5844 // Process indirect outputs, first output all of the flagged copies out of
5846 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
5847 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
5848 const Value *Ptr = IndirectStoresToEmit[i].second;
5849 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(),
5851 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
5854 // Emit the non-flagged stores from the physregs.
5855 SmallVector<SDValue, 8> OutChains;
5856 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) {
5857 SDValue Val = DAG.getStore(Chain, getCurDebugLoc(),
5858 StoresToEmit[i].first,
5859 getValue(StoresToEmit[i].second),
5860 MachinePointerInfo(StoresToEmit[i].second),
5862 OutChains.push_back(Val);
5865 if (!OutChains.empty())
5866 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
5867 &OutChains[0], OutChains.size());
5872 void SelectionDAGBuilder::visitVAStart(const CallInst &I) {
5873 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurDebugLoc(),
5874 MVT::Other, getRoot(),
5875 getValue(I.getArgOperand(0)),
5876 DAG.getSrcValue(I.getArgOperand(0))));
5879 void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) {
5880 const TargetData &TD = *TLI.getTargetData();
5881 SDValue V = DAG.getVAArg(TLI.getValueType(I.getType()), getCurDebugLoc(),
5882 getRoot(), getValue(I.getOperand(0)),
5883 DAG.getSrcValue(I.getOperand(0)),
5884 TD.getABITypeAlignment(I.getType()));
5886 DAG.setRoot(V.getValue(1));
5889 void SelectionDAGBuilder::visitVAEnd(const CallInst &I) {
5890 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurDebugLoc(),
5891 MVT::Other, getRoot(),
5892 getValue(I.getArgOperand(0)),
5893 DAG.getSrcValue(I.getArgOperand(0))));
5896 void SelectionDAGBuilder::visitVACopy(const CallInst &I) {
5897 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurDebugLoc(),
5898 MVT::Other, getRoot(),
5899 getValue(I.getArgOperand(0)),
5900 getValue(I.getArgOperand(1)),
5901 DAG.getSrcValue(I.getArgOperand(0)),
5902 DAG.getSrcValue(I.getArgOperand(1))));
5905 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
5906 /// implementation, which just calls LowerCall.
5907 /// FIXME: When all targets are
5908 /// migrated to using LowerCall, this hook should be integrated into SDISel.
5909 std::pair<SDValue, SDValue>
5910 TargetLowering::LowerCallTo(SDValue Chain, const Type *RetTy,
5911 bool RetSExt, bool RetZExt, bool isVarArg,
5912 bool isInreg, unsigned NumFixedArgs,
5913 CallingConv::ID CallConv, bool isTailCall,
5914 bool isReturnValueUsed,
5916 ArgListTy &Args, SelectionDAG &DAG,
5917 DebugLoc dl) const {
5918 // Handle all of the outgoing arguments.
5919 SmallVector<ISD::OutputArg, 32> Outs;
5920 SmallVector<SDValue, 32> OutVals;
5921 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
5922 SmallVector<EVT, 4> ValueVTs;
5923 ComputeValueVTs(*this, Args[i].Ty, ValueVTs);
5924 for (unsigned Value = 0, NumValues = ValueVTs.size();
5925 Value != NumValues; ++Value) {
5926 EVT VT = ValueVTs[Value];
5927 const Type *ArgTy = VT.getTypeForEVT(RetTy->getContext());
5928 SDValue Op = SDValue(Args[i].Node.getNode(),
5929 Args[i].Node.getResNo() + Value);
5930 ISD::ArgFlagsTy Flags;
5931 unsigned OriginalAlignment =
5932 getTargetData()->getABITypeAlignment(ArgTy);
5938 if (Args[i].isInReg)
5942 if (Args[i].isByVal) {
5944 const PointerType *Ty = cast<PointerType>(Args[i].Ty);
5945 const Type *ElementTy = Ty->getElementType();
5946 unsigned FrameAlign = getByValTypeAlignment(ElementTy);
5947 unsigned FrameSize = getTargetData()->getTypeAllocSize(ElementTy);
5948 // For ByVal, alignment should come from FE. BE will guess if this
5949 // info is not there but there are cases it cannot get right.
5950 if (Args[i].Alignment)
5951 FrameAlign = Args[i].Alignment;
5952 Flags.setByValAlign(FrameAlign);
5953 Flags.setByValSize(FrameSize);
5957 Flags.setOrigAlign(OriginalAlignment);
5959 EVT PartVT = getRegisterType(RetTy->getContext(), VT);
5960 unsigned NumParts = getNumRegisters(RetTy->getContext(), VT);
5961 SmallVector<SDValue, 4> Parts(NumParts);
5962 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
5965 ExtendKind = ISD::SIGN_EXTEND;
5966 else if (Args[i].isZExt)
5967 ExtendKind = ISD::ZERO_EXTEND;
5969 getCopyToParts(DAG, dl, Op, &Parts[0], NumParts,
5970 PartVT, ExtendKind);
5972 for (unsigned j = 0; j != NumParts; ++j) {
5973 // if it isn't first piece, alignment must be 1
5974 ISD::OutputArg MyFlags(Flags, Parts[j].getValueType(),
5976 if (NumParts > 1 && j == 0)
5977 MyFlags.Flags.setSplit();
5979 MyFlags.Flags.setOrigAlign(1);
5981 Outs.push_back(MyFlags);
5982 OutVals.push_back(Parts[j]);
5987 // Handle the incoming return values from the call.
5988 SmallVector<ISD::InputArg, 32> Ins;
5989 SmallVector<EVT, 4> RetTys;
5990 ComputeValueVTs(*this, RetTy, RetTys);
5991 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
5993 EVT RegisterVT = getRegisterType(RetTy->getContext(), VT);
5994 unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT);
5995 for (unsigned i = 0; i != NumRegs; ++i) {
5996 ISD::InputArg MyFlags;
5997 MyFlags.VT = RegisterVT;
5998 MyFlags.Used = isReturnValueUsed;
6000 MyFlags.Flags.setSExt();
6002 MyFlags.Flags.setZExt();
6004 MyFlags.Flags.setInReg();
6005 Ins.push_back(MyFlags);
6009 SmallVector<SDValue, 4> InVals;
6010 Chain = LowerCall(Chain, Callee, CallConv, isVarArg, isTailCall,
6011 Outs, OutVals, Ins, dl, DAG, InVals);
6013 // Verify that the target's LowerCall behaved as expected.
6014 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
6015 "LowerCall didn't return a valid chain!");
6016 assert((!isTailCall || InVals.empty()) &&
6017 "LowerCall emitted a return value for a tail call!");
6018 assert((isTailCall || InVals.size() == Ins.size()) &&
6019 "LowerCall didn't emit the correct number of values!");
6021 // For a tail call, the return value is merely live-out and there aren't
6022 // any nodes in the DAG representing it. Return a special value to
6023 // indicate that a tail call has been emitted and no more Instructions
6024 // should be processed in the current block.
6027 return std::make_pair(SDValue(), SDValue());
6030 DEBUG(for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
6031 assert(InVals[i].getNode() &&
6032 "LowerCall emitted a null value!");
6033 assert(Ins[i].VT == InVals[i].getValueType() &&
6034 "LowerCall emitted a value with the wrong type!");
6037 // Collect the legal value parts into potentially illegal values
6038 // that correspond to the original function's return values.
6039 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6041 AssertOp = ISD::AssertSext;
6043 AssertOp = ISD::AssertZext;
6044 SmallVector<SDValue, 4> ReturnValues;
6045 unsigned CurReg = 0;
6046 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
6048 EVT RegisterVT = getRegisterType(RetTy->getContext(), VT);
6049 unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT);
6051 ReturnValues.push_back(getCopyFromParts(DAG, dl, &InVals[CurReg],
6052 NumRegs, RegisterVT, VT,
6057 // For a function returning void, there is no return value. We can't create
6058 // such a node, so we just return a null return value in that case. In
6059 // that case, nothing will actualy look at the value.
6060 if (ReturnValues.empty())
6061 return std::make_pair(SDValue(), Chain);
6063 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, dl,
6064 DAG.getVTList(&RetTys[0], RetTys.size()),
6065 &ReturnValues[0], ReturnValues.size());
6066 return std::make_pair(Res, Chain);
6069 void TargetLowering::LowerOperationWrapper(SDNode *N,
6070 SmallVectorImpl<SDValue> &Results,
6071 SelectionDAG &DAG) const {
6072 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
6074 Results.push_back(Res);
6077 SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6078 llvm_unreachable("LowerOperation not implemented for this target!");
6083 SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) {
6084 SDValue Op = getNonRegisterValue(V);
6085 assert((Op.getOpcode() != ISD::CopyFromReg ||
6086 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
6087 "Copy from a reg to the same reg!");
6088 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg");
6090 RegsForValue RFV(V->getContext(), TLI, Reg, V->getType());
6091 SDValue Chain = DAG.getEntryNode();
6092 RFV.getCopyToRegs(Op, DAG, getCurDebugLoc(), Chain, 0);
6093 PendingExports.push_back(Chain);
6096 #include "llvm/CodeGen/SelectionDAGISel.h"
6098 void SelectionDAGISel::LowerArguments(const BasicBlock *LLVMBB) {
6099 // If this is the entry block, emit arguments.
6100 const Function &F = *LLVMBB->getParent();
6101 SelectionDAG &DAG = SDB->DAG;
6102 DebugLoc dl = SDB->getCurDebugLoc();
6103 const TargetData *TD = TLI.getTargetData();
6104 SmallVector<ISD::InputArg, 16> Ins;
6106 // Check whether the function can return without sret-demotion.
6107 SmallVector<ISD::OutputArg, 4> Outs;
6108 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
6111 if (!FuncInfo->CanLowerReturn) {
6112 // Put in an sret pointer parameter before all the other parameters.
6113 SmallVector<EVT, 1> ValueVTs;
6114 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
6116 // NOTE: Assuming that a pointer will never break down to more than one VT
6118 ISD::ArgFlagsTy Flags;
6120 EVT RegisterVT = TLI.getRegisterType(*DAG.getContext(), ValueVTs[0]);
6121 ISD::InputArg RetArg(Flags, RegisterVT, true);
6122 Ins.push_back(RetArg);
6125 // Set up the incoming argument description vector.
6127 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end();
6128 I != E; ++I, ++Idx) {
6129 SmallVector<EVT, 4> ValueVTs;
6130 ComputeValueVTs(TLI, I->getType(), ValueVTs);
6131 bool isArgValueUsed = !I->use_empty();
6132 for (unsigned Value = 0, NumValues = ValueVTs.size();
6133 Value != NumValues; ++Value) {
6134 EVT VT = ValueVTs[Value];
6135 const Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
6136 ISD::ArgFlagsTy Flags;
6137 unsigned OriginalAlignment =
6138 TD->getABITypeAlignment(ArgTy);
6140 if (F.paramHasAttr(Idx, Attribute::ZExt))
6142 if (F.paramHasAttr(Idx, Attribute::SExt))
6144 if (F.paramHasAttr(Idx, Attribute::InReg))
6146 if (F.paramHasAttr(Idx, Attribute::StructRet))
6148 if (F.paramHasAttr(Idx, Attribute::ByVal)) {
6150 const PointerType *Ty = cast<PointerType>(I->getType());
6151 const Type *ElementTy = Ty->getElementType();
6152 unsigned FrameAlign = TLI.getByValTypeAlignment(ElementTy);
6153 unsigned FrameSize = TD->getTypeAllocSize(ElementTy);
6154 // For ByVal, alignment should be passed from FE. BE will guess if
6155 // this info is not there but there are cases it cannot get right.
6156 if (F.getParamAlignment(Idx))
6157 FrameAlign = F.getParamAlignment(Idx);
6158 Flags.setByValAlign(FrameAlign);
6159 Flags.setByValSize(FrameSize);
6161 if (F.paramHasAttr(Idx, Attribute::Nest))
6163 Flags.setOrigAlign(OriginalAlignment);
6165 EVT RegisterVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6166 unsigned NumRegs = TLI.getNumRegisters(*CurDAG->getContext(), VT);
6167 for (unsigned i = 0; i != NumRegs; ++i) {
6168 ISD::InputArg MyFlags(Flags, RegisterVT, isArgValueUsed);
6169 if (NumRegs > 1 && i == 0)
6170 MyFlags.Flags.setSplit();
6171 // if it isn't first piece, alignment must be 1
6173 MyFlags.Flags.setOrigAlign(1);
6174 Ins.push_back(MyFlags);
6179 // Call the target to set up the argument values.
6180 SmallVector<SDValue, 8> InVals;
6181 SDValue NewRoot = TLI.LowerFormalArguments(DAG.getRoot(), F.getCallingConv(),
6185 // Verify that the target's LowerFormalArguments behaved as expected.
6186 assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other &&
6187 "LowerFormalArguments didn't return a valid chain!");
6188 assert(InVals.size() == Ins.size() &&
6189 "LowerFormalArguments didn't emit the correct number of values!");
6191 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
6192 assert(InVals[i].getNode() &&
6193 "LowerFormalArguments emitted a null value!");
6194 assert(Ins[i].VT == InVals[i].getValueType() &&
6195 "LowerFormalArguments emitted a value with the wrong type!");
6199 // Update the DAG with the new chain value resulting from argument lowering.
6200 DAG.setRoot(NewRoot);
6202 // Set up the argument values.
6205 if (!FuncInfo->CanLowerReturn) {
6206 // Create a virtual register for the sret pointer, and put in a copy
6207 // from the sret argument into it.
6208 SmallVector<EVT, 1> ValueVTs;
6209 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
6210 EVT VT = ValueVTs[0];
6211 EVT RegVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6212 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6213 SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1,
6214 RegVT, VT, AssertOp);
6216 MachineFunction& MF = SDB->DAG.getMachineFunction();
6217 MachineRegisterInfo& RegInfo = MF.getRegInfo();
6218 unsigned SRetReg = RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT));
6219 FuncInfo->DemoteRegister = SRetReg;
6220 NewRoot = SDB->DAG.getCopyToReg(NewRoot, SDB->getCurDebugLoc(),
6222 DAG.setRoot(NewRoot);
6224 // i indexes lowered arguments. Bump it past the hidden sret argument.
6225 // Idx indexes LLVM arguments. Don't touch it.
6229 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
6231 SmallVector<SDValue, 4> ArgValues;
6232 SmallVector<EVT, 4> ValueVTs;
6233 ComputeValueVTs(TLI, I->getType(), ValueVTs);
6234 unsigned NumValues = ValueVTs.size();
6236 // If this argument is unused then remember its value. It is used to generate
6237 // debugging information.
6238 if (I->use_empty() && NumValues)
6239 SDB->setUnusedArgValue(I, InVals[i]);
6241 for (unsigned Value = 0; Value != NumValues; ++Value) {
6242 EVT VT = ValueVTs[Value];
6243 EVT PartVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6244 unsigned NumParts = TLI.getNumRegisters(*CurDAG->getContext(), VT);
6246 if (!I->use_empty()) {
6247 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6248 if (F.paramHasAttr(Idx, Attribute::SExt))
6249 AssertOp = ISD::AssertSext;
6250 else if (F.paramHasAttr(Idx, Attribute::ZExt))
6251 AssertOp = ISD::AssertZext;
6253 ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i],
6254 NumParts, PartVT, VT,
6261 // Note down frame index for byval arguments.
6262 if (I->hasByValAttr() && !ArgValues.empty())
6263 if (FrameIndexSDNode *FI =
6264 dyn_cast<FrameIndexSDNode>(ArgValues[0].getNode()))
6265 FuncInfo->setByValArgumentFrameIndex(I, FI->getIndex());
6267 if (!I->use_empty()) {
6269 if (!ArgValues.empty())
6270 Res = DAG.getMergeValues(&ArgValues[0], NumValues,
6271 SDB->getCurDebugLoc());
6272 SDB->setValue(I, Res);
6274 // If this argument is live outside of the entry block, insert a copy from
6275 // whereever we got it to the vreg that other BB's will reference it as.
6276 SDB->CopyToExportRegsIfNeeded(I);
6280 assert(i == InVals.size() && "Argument register count mismatch!");
6282 // Finally, if the target has anything special to do, allow it to do so.
6283 // FIXME: this should insert code into the DAG!
6284 EmitFunctionEntryCode();
6287 /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
6288 /// ensure constants are generated when needed. Remember the virtual registers
6289 /// that need to be added to the Machine PHI nodes as input. We cannot just
6290 /// directly add them, because expansion might result in multiple MBB's for one
6291 /// BB. As such, the start of the BB might correspond to a different MBB than
6295 SelectionDAGBuilder::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
6296 const TerminatorInst *TI = LLVMBB->getTerminator();
6298 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
6300 // Check successor nodes' PHI nodes that expect a constant to be available
6302 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
6303 const BasicBlock *SuccBB = TI->getSuccessor(succ);
6304 if (!isa<PHINode>(SuccBB->begin())) continue;
6305 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
6307 // If this terminator has multiple identical successors (common for
6308 // switches), only handle each succ once.
6309 if (!SuccsHandled.insert(SuccMBB)) continue;
6311 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
6313 // At this point we know that there is a 1-1 correspondence between LLVM PHI
6314 // nodes and Machine PHI nodes, but the incoming operands have not been
6316 for (BasicBlock::const_iterator I = SuccBB->begin();
6317 const PHINode *PN = dyn_cast<PHINode>(I); ++I) {
6318 // Ignore dead phi's.
6319 if (PN->use_empty()) continue;
6322 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
6324 if (const Constant *C = dyn_cast<Constant>(PHIOp)) {
6325 unsigned &RegOut = ConstantsOut[C];
6327 RegOut = FuncInfo.CreateRegs(C->getType());
6328 CopyValueToVirtualRegister(C, RegOut);
6332 DenseMap<const Value *, unsigned>::iterator I =
6333 FuncInfo.ValueMap.find(PHIOp);
6334 if (I != FuncInfo.ValueMap.end())
6337 assert(isa<AllocaInst>(PHIOp) &&
6338 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
6339 "Didn't codegen value into a register!??");
6340 Reg = FuncInfo.CreateRegs(PHIOp->getType());
6341 CopyValueToVirtualRegister(PHIOp, Reg);
6345 // Remember that this register needs to added to the machine PHI node as
6346 // the input for this MBB.
6347 SmallVector<EVT, 4> ValueVTs;
6348 ComputeValueVTs(TLI, PN->getType(), ValueVTs);
6349 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
6350 EVT VT = ValueVTs[vti];
6351 unsigned NumRegisters = TLI.getNumRegisters(*DAG.getContext(), VT);
6352 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
6353 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
6354 Reg += NumRegisters;
6358 ConstantsOut.clear();