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 "SelectionDAGBuilder.h"
16 #include "FunctionLoweringInfo.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/Module.h"
31 #include "llvm/CodeGen/FastISel.h"
32 #include "llvm/CodeGen/GCStrategy.h"
33 #include "llvm/CodeGen/GCMetadata.h"
34 #include "llvm/CodeGen/MachineFunction.h"
35 #include "llvm/CodeGen/MachineFrameInfo.h"
36 #include "llvm/CodeGen/MachineInstrBuilder.h"
37 #include "llvm/CodeGen/MachineJumpTableInfo.h"
38 #include "llvm/CodeGen/MachineModuleInfo.h"
39 #include "llvm/CodeGen/MachineRegisterInfo.h"
40 #include "llvm/CodeGen/PseudoSourceValue.h"
41 #include "llvm/CodeGen/SelectionDAG.h"
42 #include "llvm/CodeGen/DwarfWriter.h"
43 #include "llvm/Analysis/DebugInfo.h"
44 #include "llvm/Target/TargetRegisterInfo.h"
45 #include "llvm/Target/TargetData.h"
46 #include "llvm/Target/TargetFrameInfo.h"
47 #include "llvm/Target/TargetInstrInfo.h"
48 #include "llvm/Target/TargetIntrinsicInfo.h"
49 #include "llvm/Target/TargetLowering.h"
50 #include "llvm/Target/TargetOptions.h"
51 #include "llvm/Support/Compiler.h"
52 #include "llvm/Support/CommandLine.h"
53 #include "llvm/Support/Debug.h"
54 #include "llvm/Support/ErrorHandling.h"
55 #include "llvm/Support/MathExtras.h"
56 #include "llvm/Support/raw_ostream.h"
60 /// LimitFloatPrecision - Generate low-precision inline sequences for
61 /// some float libcalls (6, 8 or 12 bits).
62 static unsigned LimitFloatPrecision;
64 static cl::opt<unsigned, true>
65 LimitFPPrecision("limit-float-precision",
66 cl::desc("Generate low-precision inline sequences "
67 "for some float libcalls"),
68 cl::location(LimitFloatPrecision),
72 /// RegsForValue - This struct represents the registers (physical or virtual)
73 /// that a particular set of values is assigned, and the type information
74 /// about the value. The most common situation is to represent one value at a
75 /// time, but struct or array values are handled element-wise as multiple
76 /// values. The splitting of aggregates is performed recursively, so that we
77 /// never have aggregate-typed registers. The values at this point do not
78 /// necessarily have legal types, so each value may require one or more
79 /// registers of some legal type.
82 /// TLI - The TargetLowering object.
84 const TargetLowering *TLI;
86 /// ValueVTs - The value types of the values, which may not be legal, and
87 /// may need be promoted or synthesized from one or more registers.
89 SmallVector<EVT, 4> ValueVTs;
91 /// RegVTs - The value types of the registers. This is the same size as
92 /// ValueVTs and it records, for each value, what the type of the assigned
93 /// register or registers are. (Individual values are never synthesized
94 /// from more than one type of register.)
96 /// With virtual registers, the contents of RegVTs is redundant with TLI's
97 /// getRegisterType member function, however when with physical registers
98 /// it is necessary to have a separate record of the types.
100 SmallVector<EVT, 4> RegVTs;
102 /// Regs - This list holds the registers assigned to the values.
103 /// Each legal or promoted value requires one register, and each
104 /// expanded value requires multiple registers.
106 SmallVector<unsigned, 4> Regs;
108 RegsForValue() : TLI(0) {}
110 RegsForValue(const TargetLowering &tli,
111 const SmallVector<unsigned, 4> ®s,
112 EVT regvt, EVT valuevt)
113 : TLI(&tli), ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {}
114 RegsForValue(const TargetLowering &tli,
115 const SmallVector<unsigned, 4> ®s,
116 const SmallVector<EVT, 4> ®vts,
117 const SmallVector<EVT, 4> &valuevts)
118 : TLI(&tli), ValueVTs(valuevts), RegVTs(regvts), Regs(regs) {}
119 RegsForValue(LLVMContext &Context, const TargetLowering &tli,
120 unsigned Reg, const Type *Ty) : TLI(&tli) {
121 ComputeValueVTs(tli, Ty, ValueVTs);
123 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
124 EVT ValueVT = ValueVTs[Value];
125 unsigned NumRegs = TLI->getNumRegisters(Context, ValueVT);
126 EVT RegisterVT = TLI->getRegisterType(Context, ValueVT);
127 for (unsigned i = 0; i != NumRegs; ++i)
128 Regs.push_back(Reg + i);
129 RegVTs.push_back(RegisterVT);
134 /// append - Add the specified values to this one.
135 void append(const RegsForValue &RHS) {
137 ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end());
138 RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end());
139 Regs.append(RHS.Regs.begin(), RHS.Regs.end());
143 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
144 /// this value and returns the result as a ValueVTs value. This uses
145 /// Chain/Flag as the input and updates them for the output Chain/Flag.
146 /// If the Flag pointer is NULL, no flag is used.
147 SDValue getCopyFromRegs(SelectionDAG &DAG, DebugLoc dl, unsigned Order,
148 SDValue &Chain, SDValue *Flag) const;
150 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
151 /// specified value into the registers specified by this object. This uses
152 /// Chain/Flag as the input and updates them for the output Chain/Flag.
153 /// If the Flag pointer is NULL, no flag is used.
154 void getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
155 unsigned Order, SDValue &Chain, SDValue *Flag) const;
157 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
158 /// operand list. This adds the code marker, matching input operand index
159 /// (if applicable), and includes the number of values added into it.
160 void AddInlineAsmOperands(unsigned Code,
161 bool HasMatching, unsigned MatchingIdx,
162 SelectionDAG &DAG, unsigned Order,
163 std::vector<SDValue> &Ops) const;
167 /// getCopyFromParts - Create a value that contains the specified legal parts
168 /// combined into the value they represent. If the parts combine to a type
169 /// larger then ValueVT then AssertOp can be used to specify whether the extra
170 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
171 /// (ISD::AssertSext).
172 static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc dl, unsigned Order,
173 const SDValue *Parts,
174 unsigned NumParts, EVT PartVT, EVT ValueVT,
175 ISD::NodeType AssertOp = ISD::DELETED_NODE) {
176 assert(NumParts > 0 && "No parts to assemble!");
177 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
178 SDValue Val = Parts[0];
179 DAG.AssignOrdering(Val.getNode(), Order);
182 // Assemble the value from multiple parts.
183 if (!ValueVT.isVector() && ValueVT.isInteger()) {
184 unsigned PartBits = PartVT.getSizeInBits();
185 unsigned ValueBits = ValueVT.getSizeInBits();
187 // Assemble the power of 2 part.
188 unsigned RoundParts = NumParts & (NumParts - 1) ?
189 1 << Log2_32(NumParts) : NumParts;
190 unsigned RoundBits = PartBits * RoundParts;
191 EVT RoundVT = RoundBits == ValueBits ?
192 ValueVT : EVT::getIntegerVT(*DAG.getContext(), RoundBits);
195 EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2);
197 if (RoundParts > 2) {
198 Lo = getCopyFromParts(DAG, dl, Order, Parts, RoundParts / 2,
200 Hi = getCopyFromParts(DAG, dl, Order, Parts + RoundParts / 2,
201 RoundParts / 2, PartVT, HalfVT);
203 Lo = DAG.getNode(ISD::BIT_CONVERT, dl, HalfVT, Parts[0]);
204 Hi = DAG.getNode(ISD::BIT_CONVERT, dl, HalfVT, Parts[1]);
207 if (TLI.isBigEndian())
210 Val = DAG.getNode(ISD::BUILD_PAIR, dl, RoundVT, Lo, Hi);
212 DAG.AssignOrdering(Lo.getNode(), Order);
213 DAG.AssignOrdering(Hi.getNode(), Order);
214 DAG.AssignOrdering(Val.getNode(), Order);
216 if (RoundParts < NumParts) {
217 // Assemble the trailing non-power-of-2 part.
218 unsigned OddParts = NumParts - RoundParts;
219 EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits);
220 Hi = getCopyFromParts(DAG, dl, Order,
221 Parts + RoundParts, OddParts, PartVT, OddVT);
223 // Combine the round and odd parts.
225 if (TLI.isBigEndian())
227 EVT TotalVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
228 Hi = DAG.getNode(ISD::ANY_EXTEND, dl, TotalVT, Hi);
229 DAG.AssignOrdering(Hi.getNode(), Order);
230 Hi = DAG.getNode(ISD::SHL, dl, TotalVT, Hi,
231 DAG.getConstant(Lo.getValueType().getSizeInBits(),
232 TLI.getPointerTy()));
233 DAG.AssignOrdering(Hi.getNode(), Order);
234 Lo = DAG.getNode(ISD::ZERO_EXTEND, dl, TotalVT, Lo);
235 DAG.AssignOrdering(Lo.getNode(), Order);
236 Val = DAG.getNode(ISD::OR, dl, TotalVT, Lo, Hi);
237 DAG.AssignOrdering(Val.getNode(), Order);
239 } else if (ValueVT.isVector()) {
240 // Handle a multi-element vector.
241 EVT IntermediateVT, RegisterVT;
242 unsigned NumIntermediates;
244 TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT,
245 NumIntermediates, RegisterVT);
246 assert(NumRegs == NumParts
247 && "Part count doesn't match vector breakdown!");
248 NumParts = NumRegs; // Silence a compiler warning.
249 assert(RegisterVT == PartVT
250 && "Part type doesn't match vector breakdown!");
251 assert(RegisterVT == Parts[0].getValueType() &&
252 "Part type doesn't match part!");
254 // Assemble the parts into intermediate operands.
255 SmallVector<SDValue, 8> Ops(NumIntermediates);
256 if (NumIntermediates == NumParts) {
257 // If the register was not expanded, truncate or copy the value,
259 for (unsigned i = 0; i != NumParts; ++i)
260 Ops[i] = getCopyFromParts(DAG, dl, Order, &Parts[i], 1,
261 PartVT, IntermediateVT);
262 } else if (NumParts > 0) {
263 // If the intermediate type was expanded, build the intermediate
264 // operands from the parts.
265 assert(NumParts % NumIntermediates == 0 &&
266 "Must expand into a divisible number of parts!");
267 unsigned Factor = NumParts / NumIntermediates;
268 for (unsigned i = 0; i != NumIntermediates; ++i)
269 Ops[i] = getCopyFromParts(DAG, dl, Order, &Parts[i * Factor], Factor,
270 PartVT, IntermediateVT);
273 // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the
274 // intermediate operands.
275 Val = DAG.getNode(IntermediateVT.isVector() ?
276 ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR, dl,
277 ValueVT, &Ops[0], NumIntermediates);
278 DAG.AssignOrdering(Val.getNode(), Order);
279 } else if (PartVT.isFloatingPoint()) {
280 // FP split into multiple FP parts (for ppcf128)
281 assert(ValueVT == EVT(MVT::ppcf128) && PartVT == EVT(MVT::f64) &&
284 Lo = DAG.getNode(ISD::BIT_CONVERT, dl, EVT(MVT::f64), Parts[0]);
285 Hi = DAG.getNode(ISD::BIT_CONVERT, dl, EVT(MVT::f64), Parts[1]);
286 if (TLI.isBigEndian())
288 Val = DAG.getNode(ISD::BUILD_PAIR, dl, ValueVT, Lo, Hi);
290 DAG.AssignOrdering(Hi.getNode(), Order);
291 DAG.AssignOrdering(Lo.getNode(), Order);
292 DAG.AssignOrdering(Val.getNode(), Order);
294 // FP split into integer parts (soft fp)
295 assert(ValueVT.isFloatingPoint() && PartVT.isInteger() &&
296 !PartVT.isVector() && "Unexpected split");
297 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits());
298 Val = getCopyFromParts(DAG, dl, Order, Parts, NumParts, PartVT, IntVT);
302 // There is now one part, held in Val. Correct it to match ValueVT.
303 PartVT = Val.getValueType();
305 if (PartVT == ValueVT)
308 if (PartVT.isVector()) {
309 assert(ValueVT.isVector() && "Unknown vector conversion!");
310 SDValue Res = DAG.getNode(ISD::BIT_CONVERT, dl, ValueVT, Val);
311 DAG.AssignOrdering(Res.getNode(), Order);
315 if (ValueVT.isVector()) {
316 assert(ValueVT.getVectorElementType() == PartVT &&
317 ValueVT.getVectorNumElements() == 1 &&
318 "Only trivial scalar-to-vector conversions should get here!");
319 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, ValueVT, Val);
320 DAG.AssignOrdering(Res.getNode(), Order);
324 if (PartVT.isInteger() &&
325 ValueVT.isInteger()) {
326 if (ValueVT.bitsLT(PartVT)) {
327 // For a truncate, see if we have any information to
328 // indicate whether the truncated bits will always be
329 // zero or sign-extension.
330 if (AssertOp != ISD::DELETED_NODE)
331 Val = DAG.getNode(AssertOp, dl, PartVT, Val,
332 DAG.getValueType(ValueVT));
333 DAG.AssignOrdering(Val.getNode(), Order);
334 Val = DAG.getNode(ISD::TRUNCATE, dl, ValueVT, Val);
335 DAG.AssignOrdering(Val.getNode(), Order);
338 Val = DAG.getNode(ISD::ANY_EXTEND, dl, ValueVT, Val);
339 DAG.AssignOrdering(Val.getNode(), Order);
344 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
345 if (ValueVT.bitsLT(Val.getValueType())) {
346 // FP_ROUND's are always exact here.
347 Val = DAG.getNode(ISD::FP_ROUND, dl, ValueVT, Val,
348 DAG.getIntPtrConstant(1));
349 DAG.AssignOrdering(Val.getNode(), Order);
353 Val = DAG.getNode(ISD::FP_EXTEND, dl, ValueVT, Val);
354 DAG.AssignOrdering(Val.getNode(), Order);
358 if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) {
359 Val = DAG.getNode(ISD::BIT_CONVERT, dl, ValueVT, Val);
360 DAG.AssignOrdering(Val.getNode(), Order);
364 llvm_unreachable("Unknown mismatch!");
368 /// getCopyToParts - Create a series of nodes that contain the specified value
369 /// split into legal parts. If the parts contain more bits than Val, then, for
370 /// integers, ExtendKind can be used to specify how to generate the extra bits.
371 static void getCopyToParts(SelectionDAG &DAG, DebugLoc dl, unsigned Order,
372 SDValue Val, SDValue *Parts, unsigned NumParts,
374 ISD::NodeType ExtendKind = ISD::ANY_EXTEND) {
375 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
376 EVT PtrVT = TLI.getPointerTy();
377 EVT ValueVT = Val.getValueType();
378 unsigned PartBits = PartVT.getSizeInBits();
379 unsigned OrigNumParts = NumParts;
380 assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!");
385 if (!ValueVT.isVector()) {
386 if (PartVT == ValueVT) {
387 assert(NumParts == 1 && "No-op copy with multiple parts!");
392 if (NumParts * PartBits > ValueVT.getSizeInBits()) {
393 // If the parts cover more bits than the value has, promote the value.
394 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
395 assert(NumParts == 1 && "Do not know what to promote to!");
396 Val = DAG.getNode(ISD::FP_EXTEND, dl, PartVT, Val);
397 } else if (PartVT.isInteger() && ValueVT.isInteger()) {
398 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
399 Val = DAG.getNode(ExtendKind, dl, ValueVT, Val);
401 llvm_unreachable("Unknown mismatch!");
403 } else if (PartBits == ValueVT.getSizeInBits()) {
404 // Different types of the same size.
405 assert(NumParts == 1 && PartVT != ValueVT);
406 Val = DAG.getNode(ISD::BIT_CONVERT, dl, PartVT, Val);
407 } else if (NumParts * PartBits < ValueVT.getSizeInBits()) {
408 // If the parts cover less bits than value has, truncate the value.
409 if (PartVT.isInteger() && ValueVT.isInteger()) {
410 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
411 Val = DAG.getNode(ISD::TRUNCATE, dl, ValueVT, Val);
413 llvm_unreachable("Unknown mismatch!");
417 DAG.AssignOrdering(Val.getNode(), Order);
419 // The value may have changed - recompute ValueVT.
420 ValueVT = Val.getValueType();
421 assert(NumParts * PartBits == ValueVT.getSizeInBits() &&
422 "Failed to tile the value with PartVT!");
425 assert(PartVT == ValueVT && "Type conversion failed!");
430 // Expand the value into multiple parts.
431 if (NumParts & (NumParts - 1)) {
432 // The number of parts is not a power of 2. Split off and copy the tail.
433 assert(PartVT.isInteger() && ValueVT.isInteger() &&
434 "Do not know what to expand to!");
435 unsigned RoundParts = 1 << Log2_32(NumParts);
436 unsigned RoundBits = RoundParts * PartBits;
437 unsigned OddParts = NumParts - RoundParts;
438 SDValue OddVal = DAG.getNode(ISD::SRL, dl, ValueVT, Val,
439 DAG.getConstant(RoundBits,
440 TLI.getPointerTy()));
441 getCopyToParts(DAG, dl, Order, OddVal, Parts + RoundParts,
444 if (TLI.isBigEndian())
445 // The odd parts were reversed by getCopyToParts - unreverse them.
446 std::reverse(Parts + RoundParts, Parts + NumParts);
448 NumParts = RoundParts;
449 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
450 Val = DAG.getNode(ISD::TRUNCATE, dl, ValueVT, Val);
452 DAG.AssignOrdering(OddVal.getNode(), Order);
453 DAG.AssignOrdering(Val.getNode(), Order);
456 // The number of parts is a power of 2. Repeatedly bisect the value using
458 Parts[0] = DAG.getNode(ISD::BIT_CONVERT, dl,
459 EVT::getIntegerVT(*DAG.getContext(),
460 ValueVT.getSizeInBits()),
463 DAG.AssignOrdering(Parts[0].getNode(), Order);
465 for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) {
466 for (unsigned i = 0; i < NumParts; i += StepSize) {
467 unsigned ThisBits = StepSize * PartBits / 2;
468 EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits);
469 SDValue &Part0 = Parts[i];
470 SDValue &Part1 = Parts[i+StepSize/2];
472 Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
474 DAG.getConstant(1, PtrVT));
475 Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
477 DAG.getConstant(0, PtrVT));
479 DAG.AssignOrdering(Part0.getNode(), Order);
480 DAG.AssignOrdering(Part1.getNode(), Order);
482 if (ThisBits == PartBits && ThisVT != PartVT) {
483 Part0 = DAG.getNode(ISD::BIT_CONVERT, dl,
485 Part1 = DAG.getNode(ISD::BIT_CONVERT, dl,
487 DAG.AssignOrdering(Part0.getNode(), Order);
488 DAG.AssignOrdering(Part1.getNode(), Order);
493 if (TLI.isBigEndian())
494 std::reverse(Parts, Parts + OrigNumParts);
501 if (PartVT != ValueVT) {
502 if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) {
503 Val = DAG.getNode(ISD::BIT_CONVERT, dl, PartVT, Val);
505 assert(ValueVT.getVectorElementType() == PartVT &&
506 ValueVT.getVectorNumElements() == 1 &&
507 "Only trivial vector-to-scalar conversions should get here!");
508 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
510 DAG.getConstant(0, PtrVT));
514 DAG.AssignOrdering(Val.getNode(), Order);
519 // Handle a multi-element vector.
520 EVT IntermediateVT, RegisterVT;
521 unsigned NumIntermediates;
522 unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT,
523 IntermediateVT, NumIntermediates, RegisterVT);
524 unsigned NumElements = ValueVT.getVectorNumElements();
526 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
527 NumParts = NumRegs; // Silence a compiler warning.
528 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
530 // Split the vector into intermediate operands.
531 SmallVector<SDValue, 8> Ops(NumIntermediates);
532 for (unsigned i = 0; i != NumIntermediates; ++i) {
533 if (IntermediateVT.isVector())
534 Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl,
536 DAG.getConstant(i * (NumElements / NumIntermediates),
539 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
541 DAG.getConstant(i, PtrVT));
543 DAG.AssignOrdering(Ops[i].getNode(), Order);
546 // Split the intermediate operands into legal parts.
547 if (NumParts == NumIntermediates) {
548 // If the register was not expanded, promote or copy the value,
550 for (unsigned i = 0; i != NumParts; ++i)
551 getCopyToParts(DAG, dl, Order, Ops[i], &Parts[i], 1, PartVT);
552 } else if (NumParts > 0) {
553 // If the intermediate type was expanded, split each the value into
555 assert(NumParts % NumIntermediates == 0 &&
556 "Must expand into a divisible number of parts!");
557 unsigned Factor = NumParts / NumIntermediates;
558 for (unsigned i = 0; i != NumIntermediates; ++i)
559 getCopyToParts(DAG, dl, Order, Ops[i], &Parts[i*Factor], Factor, PartVT);
564 void SelectionDAGBuilder::init(GCFunctionInfo *gfi, AliasAnalysis &aa) {
567 TD = DAG.getTarget().getTargetData();
570 /// clear - Clear out the curret SelectionDAG and the associated
571 /// state and prepare this SelectionDAGBuilder object to be used
572 /// for a new block. This doesn't clear out information about
573 /// additional blocks that are needed to complete switch lowering
574 /// or PHI node updating; that information is cleared out as it is
576 void SelectionDAGBuilder::clear() {
578 PendingLoads.clear();
579 PendingExports.clear();
582 CurDebugLoc = DebugLoc::getUnknownLoc();
586 /// getRoot - Return the current virtual root of the Selection DAG,
587 /// flushing any PendingLoad items. This must be done before emitting
588 /// a store or any other node that may need to be ordered after any
589 /// prior load instructions.
591 SDValue SelectionDAGBuilder::getRoot() {
592 if (PendingLoads.empty())
593 return DAG.getRoot();
595 if (PendingLoads.size() == 1) {
596 SDValue Root = PendingLoads[0];
598 PendingLoads.clear();
602 // Otherwise, we have to make a token factor node.
603 SDValue Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
604 &PendingLoads[0], PendingLoads.size());
605 PendingLoads.clear();
610 /// getControlRoot - Similar to getRoot, but instead of flushing all the
611 /// PendingLoad items, flush all the PendingExports items. It is necessary
612 /// to do this before emitting a terminator instruction.
614 SDValue SelectionDAGBuilder::getControlRoot() {
615 SDValue Root = DAG.getRoot();
617 if (PendingExports.empty())
620 // Turn all of the CopyToReg chains into one factored node.
621 if (Root.getOpcode() != ISD::EntryToken) {
622 unsigned i = 0, e = PendingExports.size();
623 for (; i != e; ++i) {
624 assert(PendingExports[i].getNode()->getNumOperands() > 1);
625 if (PendingExports[i].getNode()->getOperand(0) == Root)
626 break; // Don't add the root if we already indirectly depend on it.
630 PendingExports.push_back(Root);
633 Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
635 PendingExports.size());
636 PendingExports.clear();
641 void SelectionDAGBuilder::visit(Instruction &I) {
642 visit(I.getOpcode(), I);
645 void SelectionDAGBuilder::visit(unsigned Opcode, User &I) {
646 // We're processing a new instruction.
649 // Note: this doesn't use InstVisitor, because it has to work with
650 // ConstantExpr's in addition to instructions.
652 default: llvm_unreachable("Unknown instruction type encountered!");
653 // Build the switch statement using the Instruction.def file.
654 #define HANDLE_INST(NUM, OPCODE, CLASS) \
655 case Instruction::OPCODE: return visit##OPCODE((CLASS&)I);
656 #include "llvm/Instruction.def"
660 SDValue SelectionDAGBuilder::getValue(const Value *V) {
661 SDValue &N = NodeMap[V];
662 if (N.getNode()) return N;
664 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(V))) {
665 EVT VT = TLI.getValueType(V->getType(), true);
667 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
668 return N = DAG.getConstant(*CI, VT);
670 if (GlobalValue *GV = dyn_cast<GlobalValue>(C))
671 return N = DAG.getGlobalAddress(GV, VT);
673 if (isa<ConstantPointerNull>(C))
674 return N = DAG.getConstant(0, TLI.getPointerTy());
676 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C))
677 return N = DAG.getConstantFP(*CFP, VT);
679 if (isa<UndefValue>(C) && !V->getType()->isAggregateType())
680 return N = DAG.getUNDEF(VT);
682 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
683 visit(CE->getOpcode(), *CE);
684 SDValue N1 = NodeMap[V];
685 assert(N1.getNode() && "visit didn't populate the ValueMap!");
689 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) {
690 SmallVector<SDValue, 4> Constants;
691 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end();
693 SDNode *Val = getValue(*OI).getNode();
694 // If the operand is an empty aggregate, there are no values.
696 // Add each leaf value from the operand to the Constants list
697 // to form a flattened list of all the values.
698 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
699 Constants.push_back(SDValue(Val, i));
702 SDValue Res = DAG.getMergeValues(&Constants[0], Constants.size(),
704 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
708 if (isa<StructType>(C->getType()) || isa<ArrayType>(C->getType())) {
709 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
710 "Unknown struct or array constant!");
712 SmallVector<EVT, 4> ValueVTs;
713 ComputeValueVTs(TLI, C->getType(), ValueVTs);
714 unsigned NumElts = ValueVTs.size();
716 return SDValue(); // empty struct
717 SmallVector<SDValue, 4> Constants(NumElts);
718 for (unsigned i = 0; i != NumElts; ++i) {
719 EVT EltVT = ValueVTs[i];
720 if (isa<UndefValue>(C))
721 Constants[i] = DAG.getUNDEF(EltVT);
722 else if (EltVT.isFloatingPoint())
723 Constants[i] = DAG.getConstantFP(0, EltVT);
725 Constants[i] = DAG.getConstant(0, EltVT);
728 SDValue Res = DAG.getMergeValues(&Constants[0], NumElts,
730 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
734 if (BlockAddress *BA = dyn_cast<BlockAddress>(C))
735 return DAG.getBlockAddress(BA, VT);
737 const VectorType *VecTy = cast<VectorType>(V->getType());
738 unsigned NumElements = VecTy->getNumElements();
740 // Now that we know the number and type of the elements, get that number of
741 // elements into the Ops array based on what kind of constant it is.
742 SmallVector<SDValue, 16> Ops;
743 if (ConstantVector *CP = dyn_cast<ConstantVector>(C)) {
744 for (unsigned i = 0; i != NumElements; ++i)
745 Ops.push_back(getValue(CP->getOperand(i)));
747 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!");
748 EVT EltVT = TLI.getValueType(VecTy->getElementType());
751 if (EltVT.isFloatingPoint())
752 Op = DAG.getConstantFP(0, EltVT);
754 Op = DAG.getConstant(0, EltVT);
755 Ops.assign(NumElements, Op);
758 // Create a BUILD_VECTOR node.
759 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
760 VT, &Ops[0], Ops.size());
761 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
762 return NodeMap[V] = Res;
765 // If this is a static alloca, generate it as the frameindex instead of
767 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
768 DenseMap<const AllocaInst*, int>::iterator SI =
769 FuncInfo.StaticAllocaMap.find(AI);
770 if (SI != FuncInfo.StaticAllocaMap.end())
771 return DAG.getFrameIndex(SI->second, TLI.getPointerTy());
774 unsigned InReg = FuncInfo.ValueMap[V];
775 assert(InReg && "Value not in map!");
777 RegsForValue RFV(*DAG.getContext(), TLI, InReg, V->getType());
778 SDValue Chain = DAG.getEntryNode();
779 return RFV.getCopyFromRegs(DAG, getCurDebugLoc(),
780 SDNodeOrder, Chain, NULL);
783 /// Get the EVTs and ArgFlags collections that represent the legalized return
784 /// type of the given function. This does not require a DAG or a return value,
785 /// and is suitable for use before any DAGs for the function are constructed.
786 static void getReturnInfo(const Type* ReturnType,
787 Attributes attr, SmallVectorImpl<EVT> &OutVTs,
788 SmallVectorImpl<ISD::ArgFlagsTy> &OutFlags,
790 SmallVectorImpl<uint64_t> *Offsets = 0) {
791 SmallVector<EVT, 4> ValueVTs;
792 ComputeValueVTs(TLI, ReturnType, ValueVTs);
793 unsigned NumValues = ValueVTs.size();
794 if (NumValues == 0) return;
797 for (unsigned j = 0, f = NumValues; j != f; ++j) {
798 EVT VT = ValueVTs[j];
799 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
801 if (attr & Attribute::SExt)
802 ExtendKind = ISD::SIGN_EXTEND;
803 else if (attr & Attribute::ZExt)
804 ExtendKind = ISD::ZERO_EXTEND;
806 // FIXME: C calling convention requires the return type to be promoted to
807 // at least 32-bit. But this is not necessary for non-C calling
808 // conventions. The frontend should mark functions whose return values
809 // require promoting with signext or zeroext attributes.
810 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
811 EVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32);
812 if (VT.bitsLT(MinVT))
816 unsigned NumParts = TLI.getNumRegisters(ReturnType->getContext(), VT);
817 EVT PartVT = TLI.getRegisterType(ReturnType->getContext(), VT);
818 unsigned PartSize = TLI.getTargetData()->getTypeAllocSize(
819 PartVT.getTypeForEVT(ReturnType->getContext()));
821 // 'inreg' on function refers to return value
822 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
823 if (attr & Attribute::InReg)
826 // Propagate extension type if any
827 if (attr & Attribute::SExt)
829 else if (attr & Attribute::ZExt)
832 for (unsigned i = 0; i < NumParts; ++i) {
833 OutVTs.push_back(PartVT);
834 OutFlags.push_back(Flags);
837 Offsets->push_back(Offset);
844 void SelectionDAGBuilder::visitRet(ReturnInst &I) {
845 SDValue Chain = getControlRoot();
846 SmallVector<ISD::OutputArg, 8> Outs;
847 FunctionLoweringInfo &FLI = DAG.getFunctionLoweringInfo();
849 if (!FLI.CanLowerReturn) {
850 unsigned DemoteReg = FLI.DemoteRegister;
851 const Function *F = I.getParent()->getParent();
853 // Emit a store of the return value through the virtual register.
854 // Leave Outs empty so that LowerReturn won't try to load return
855 // registers the usual way.
856 SmallVector<EVT, 1> PtrValueVTs;
857 ComputeValueVTs(TLI, PointerType::getUnqual(F->getReturnType()),
860 SDValue RetPtr = DAG.getRegister(DemoteReg, PtrValueVTs[0]);
861 SDValue RetOp = getValue(I.getOperand(0));
863 SmallVector<EVT, 4> ValueVTs;
864 SmallVector<uint64_t, 4> Offsets;
865 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs, &Offsets);
866 unsigned NumValues = ValueVTs.size();
868 SmallVector<SDValue, 4> Chains(NumValues);
869 EVT PtrVT = PtrValueVTs[0];
870 for (unsigned i = 0; i != NumValues; ++i) {
871 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT, RetPtr,
872 DAG.getConstant(Offsets[i], PtrVT));
874 DAG.getStore(Chain, getCurDebugLoc(),
875 SDValue(RetOp.getNode(), RetOp.getResNo() + i),
876 Add, NULL, Offsets[i], false, 0);
878 DAG.AssignOrdering(Add.getNode(), SDNodeOrder);
879 DAG.AssignOrdering(Chains[i].getNode(), SDNodeOrder);
882 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
883 MVT::Other, &Chains[0], NumValues);
885 DAG.AssignOrdering(Chain.getNode(), SDNodeOrder);
887 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
888 SmallVector<EVT, 4> ValueVTs;
889 ComputeValueVTs(TLI, I.getOperand(i)->getType(), ValueVTs);
890 unsigned NumValues = ValueVTs.size();
891 if (NumValues == 0) continue;
893 SDValue RetOp = getValue(I.getOperand(i));
894 for (unsigned j = 0, f = NumValues; j != f; ++j) {
895 EVT VT = ValueVTs[j];
897 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
899 const Function *F = I.getParent()->getParent();
900 if (F->paramHasAttr(0, Attribute::SExt))
901 ExtendKind = ISD::SIGN_EXTEND;
902 else if (F->paramHasAttr(0, Attribute::ZExt))
903 ExtendKind = ISD::ZERO_EXTEND;
905 // FIXME: C calling convention requires the return type to be promoted
906 // to at least 32-bit. But this is not necessary for non-C calling
907 // conventions. The frontend should mark functions whose return values
908 // require promoting with signext or zeroext attributes.
909 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
910 EVT MinVT = TLI.getRegisterType(*DAG.getContext(), MVT::i32);
911 if (VT.bitsLT(MinVT))
915 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), VT);
916 EVT PartVT = TLI.getRegisterType(*DAG.getContext(), VT);
917 SmallVector<SDValue, 4> Parts(NumParts);
918 getCopyToParts(DAG, getCurDebugLoc(), SDNodeOrder,
919 SDValue(RetOp.getNode(), RetOp.getResNo() + j),
920 &Parts[0], NumParts, PartVT, ExtendKind);
922 // 'inreg' on function refers to return value
923 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
924 if (F->paramHasAttr(0, Attribute::InReg))
927 // Propagate extension type if any
928 if (F->paramHasAttr(0, Attribute::SExt))
930 else if (F->paramHasAttr(0, Attribute::ZExt))
933 for (unsigned i = 0; i < NumParts; ++i)
934 Outs.push_back(ISD::OutputArg(Flags, Parts[i], /*isfixed=*/true));
939 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
940 CallingConv::ID CallConv =
941 DAG.getMachineFunction().getFunction()->getCallingConv();
942 Chain = TLI.LowerReturn(Chain, CallConv, isVarArg,
943 Outs, getCurDebugLoc(), DAG);
945 // Verify that the target's LowerReturn behaved as expected.
946 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
947 "LowerReturn didn't return a valid chain!");
949 // Update the DAG with the new chain value resulting from return lowering.
951 DAG.AssignOrdering(Chain.getNode(), SDNodeOrder);
954 /// CopyToExportRegsIfNeeded - If the given value has virtual registers
955 /// created for it, emit nodes to copy the value into the virtual
957 void SelectionDAGBuilder::CopyToExportRegsIfNeeded(Value *V) {
958 if (!V->use_empty()) {
959 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
960 if (VMI != FuncInfo.ValueMap.end())
961 CopyValueToVirtualRegister(V, VMI->second);
965 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
966 /// the current basic block, add it to ValueMap now so that we'll get a
968 void SelectionDAGBuilder::ExportFromCurrentBlock(Value *V) {
969 // No need to export constants.
970 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
973 if (FuncInfo.isExportedInst(V)) return;
975 unsigned Reg = FuncInfo.InitializeRegForValue(V);
976 CopyValueToVirtualRegister(V, Reg);
979 bool SelectionDAGBuilder::isExportableFromCurrentBlock(Value *V,
980 const BasicBlock *FromBB) {
981 // The operands of the setcc have to be in this block. We don't know
982 // how to export them from some other block.
983 if (Instruction *VI = dyn_cast<Instruction>(V)) {
984 // Can export from current BB.
985 if (VI->getParent() == FromBB)
988 // Is already exported, noop.
989 return FuncInfo.isExportedInst(V);
992 // If this is an argument, we can export it if the BB is the entry block or
993 // if it is already exported.
994 if (isa<Argument>(V)) {
995 if (FromBB == &FromBB->getParent()->getEntryBlock())
998 // Otherwise, can only export this if it is already exported.
999 return FuncInfo.isExportedInst(V);
1002 // Otherwise, constants can always be exported.
1006 static bool InBlock(const Value *V, const BasicBlock *BB) {
1007 if (const Instruction *I = dyn_cast<Instruction>(V))
1008 return I->getParent() == BB;
1012 /// getFCmpCondCode - Return the ISD condition code corresponding to
1013 /// the given LLVM IR floating-point condition code. This includes
1014 /// consideration of global floating-point math flags.
1016 static ISD::CondCode getFCmpCondCode(FCmpInst::Predicate Pred) {
1017 ISD::CondCode FPC, FOC;
1019 case FCmpInst::FCMP_FALSE: FOC = FPC = ISD::SETFALSE; break;
1020 case FCmpInst::FCMP_OEQ: FOC = ISD::SETEQ; FPC = ISD::SETOEQ; break;
1021 case FCmpInst::FCMP_OGT: FOC = ISD::SETGT; FPC = ISD::SETOGT; break;
1022 case FCmpInst::FCMP_OGE: FOC = ISD::SETGE; FPC = ISD::SETOGE; break;
1023 case FCmpInst::FCMP_OLT: FOC = ISD::SETLT; FPC = ISD::SETOLT; break;
1024 case FCmpInst::FCMP_OLE: FOC = ISD::SETLE; FPC = ISD::SETOLE; break;
1025 case FCmpInst::FCMP_ONE: FOC = ISD::SETNE; FPC = ISD::SETONE; break;
1026 case FCmpInst::FCMP_ORD: FOC = FPC = ISD::SETO; break;
1027 case FCmpInst::FCMP_UNO: FOC = FPC = ISD::SETUO; break;
1028 case FCmpInst::FCMP_UEQ: FOC = ISD::SETEQ; FPC = ISD::SETUEQ; break;
1029 case FCmpInst::FCMP_UGT: FOC = ISD::SETGT; FPC = ISD::SETUGT; break;
1030 case FCmpInst::FCMP_UGE: FOC = ISD::SETGE; FPC = ISD::SETUGE; break;
1031 case FCmpInst::FCMP_ULT: FOC = ISD::SETLT; FPC = ISD::SETULT; break;
1032 case FCmpInst::FCMP_ULE: FOC = ISD::SETLE; FPC = ISD::SETULE; break;
1033 case FCmpInst::FCMP_UNE: FOC = ISD::SETNE; FPC = ISD::SETUNE; break;
1034 case FCmpInst::FCMP_TRUE: FOC = FPC = ISD::SETTRUE; break;
1036 llvm_unreachable("Invalid FCmp predicate opcode!");
1037 FOC = FPC = ISD::SETFALSE;
1040 if (FiniteOnlyFPMath())
1046 /// getICmpCondCode - Return the ISD condition code corresponding to
1047 /// the given LLVM IR integer condition code.
1049 static ISD::CondCode getICmpCondCode(ICmpInst::Predicate Pred) {
1051 case ICmpInst::ICMP_EQ: return ISD::SETEQ;
1052 case ICmpInst::ICMP_NE: return ISD::SETNE;
1053 case ICmpInst::ICMP_SLE: return ISD::SETLE;
1054 case ICmpInst::ICMP_ULE: return ISD::SETULE;
1055 case ICmpInst::ICMP_SGE: return ISD::SETGE;
1056 case ICmpInst::ICMP_UGE: return ISD::SETUGE;
1057 case ICmpInst::ICMP_SLT: return ISD::SETLT;
1058 case ICmpInst::ICMP_ULT: return ISD::SETULT;
1059 case ICmpInst::ICMP_SGT: return ISD::SETGT;
1060 case ICmpInst::ICMP_UGT: return ISD::SETUGT;
1062 llvm_unreachable("Invalid ICmp predicate opcode!");
1067 /// EmitBranchForMergedCondition - Helper method for FindMergedConditions.
1068 /// This function emits a branch and is used at the leaves of an OR or an
1069 /// AND operator tree.
1072 SelectionDAGBuilder::EmitBranchForMergedCondition(Value *Cond,
1073 MachineBasicBlock *TBB,
1074 MachineBasicBlock *FBB,
1075 MachineBasicBlock *CurBB) {
1076 const BasicBlock *BB = CurBB->getBasicBlock();
1078 // If the leaf of the tree is a comparison, merge the condition into
1080 if (CmpInst *BOp = dyn_cast<CmpInst>(Cond)) {
1081 // The operands of the cmp have to be in this block. We don't know
1082 // how to export them from some other block. If this is the first block
1083 // of the sequence, no exporting is needed.
1084 if (CurBB == CurMBB ||
1085 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
1086 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) {
1087 ISD::CondCode Condition;
1088 if (ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
1089 Condition = getICmpCondCode(IC->getPredicate());
1090 } else if (FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) {
1091 Condition = getFCmpCondCode(FC->getPredicate());
1093 Condition = ISD::SETEQ; // silence warning.
1094 llvm_unreachable("Unknown compare instruction");
1097 CaseBlock CB(Condition, BOp->getOperand(0),
1098 BOp->getOperand(1), NULL, TBB, FBB, CurBB);
1099 SwitchCases.push_back(CB);
1104 // Create a CaseBlock record representing this branch.
1105 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()),
1106 NULL, TBB, FBB, CurBB);
1107 SwitchCases.push_back(CB);
1110 /// FindMergedConditions - If Cond is an expression like
1111 void SelectionDAGBuilder::FindMergedConditions(Value *Cond,
1112 MachineBasicBlock *TBB,
1113 MachineBasicBlock *FBB,
1114 MachineBasicBlock *CurBB,
1116 // If this node is not part of the or/and tree, emit it as a branch.
1117 Instruction *BOp = dyn_cast<Instruction>(Cond);
1118 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
1119 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
1120 BOp->getParent() != CurBB->getBasicBlock() ||
1121 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
1122 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
1123 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB);
1127 // Create TmpBB after CurBB.
1128 MachineFunction::iterator BBI = CurBB;
1129 MachineFunction &MF = DAG.getMachineFunction();
1130 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock());
1131 CurBB->getParent()->insert(++BBI, TmpBB);
1133 if (Opc == Instruction::Or) {
1134 // Codegen X | Y as:
1142 // Emit the LHS condition.
1143 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, Opc);
1145 // Emit the RHS condition into TmpBB.
1146 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, Opc);
1148 assert(Opc == Instruction::And && "Unknown merge op!");
1149 // Codegen X & Y as:
1156 // This requires creation of TmpBB after CurBB.
1158 // Emit the LHS condition.
1159 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, Opc);
1161 // Emit the RHS condition into TmpBB.
1162 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, Opc);
1166 /// If the set of cases should be emitted as a series of branches, return true.
1167 /// If we should emit this as a bunch of and/or'd together conditions, return
1170 SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases){
1171 if (Cases.size() != 2) return true;
1173 // If this is two comparisons of the same values or'd or and'd together, they
1174 // will get folded into a single comparison, so don't emit two blocks.
1175 if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
1176 Cases[0].CmpRHS == Cases[1].CmpRHS) ||
1177 (Cases[0].CmpRHS == Cases[1].CmpLHS &&
1178 Cases[0].CmpLHS == Cases[1].CmpRHS)) {
1182 // Handle: (X != null) | (Y != null) --> (X|Y) != 0
1183 // Handle: (X == null) & (Y == null) --> (X|Y) == 0
1184 if (Cases[0].CmpRHS == Cases[1].CmpRHS &&
1185 Cases[0].CC == Cases[1].CC &&
1186 isa<Constant>(Cases[0].CmpRHS) &&
1187 cast<Constant>(Cases[0].CmpRHS)->isNullValue()) {
1188 if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB)
1190 if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB)
1197 void SelectionDAGBuilder::visitBr(BranchInst &I) {
1198 // Update machine-CFG edges.
1199 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
1201 // Figure out which block is immediately after the current one.
1202 MachineBasicBlock *NextBlock = 0;
1203 MachineFunction::iterator BBI = CurMBB;
1204 if (++BBI != FuncInfo.MF->end())
1207 if (I.isUnconditional()) {
1208 // Update machine-CFG edges.
1209 CurMBB->addSuccessor(Succ0MBB);
1211 // If this is not a fall-through branch, emit the branch.
1212 if (Succ0MBB != NextBlock) {
1213 SDValue V = DAG.getNode(ISD::BR, getCurDebugLoc(),
1214 MVT::Other, getControlRoot(),
1215 DAG.getBasicBlock(Succ0MBB));
1217 DAG.AssignOrdering(V.getNode(), SDNodeOrder);
1223 // If this condition is one of the special cases we handle, do special stuff
1225 Value *CondVal = I.getCondition();
1226 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
1228 // If this is a series of conditions that are or'd or and'd together, emit
1229 // this as a sequence of branches instead of setcc's with and/or operations.
1230 // For example, instead of something like:
1243 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
1244 if (BOp->hasOneUse() &&
1245 (BOp->getOpcode() == Instruction::And ||
1246 BOp->getOpcode() == Instruction::Or)) {
1247 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, CurMBB, BOp->getOpcode());
1248 // If the compares in later blocks need to use values not currently
1249 // exported from this block, export them now. This block should always
1250 // be the first entry.
1251 assert(SwitchCases[0].ThisBB == CurMBB && "Unexpected lowering!");
1253 // Allow some cases to be rejected.
1254 if (ShouldEmitAsBranches(SwitchCases)) {
1255 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1256 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1257 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1260 // Emit the branch for this block.
1261 visitSwitchCase(SwitchCases[0]);
1262 SwitchCases.erase(SwitchCases.begin());
1266 // Okay, we decided not to do this, remove any inserted MBB's and clear
1268 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
1269 FuncInfo.MF->erase(SwitchCases[i].ThisBB);
1271 SwitchCases.clear();
1275 // Create a CaseBlock record representing this branch.
1276 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()),
1277 NULL, Succ0MBB, Succ1MBB, CurMBB);
1279 // Use visitSwitchCase to actually insert the fast branch sequence for this
1281 visitSwitchCase(CB);
1284 /// visitSwitchCase - Emits the necessary code to represent a single node in
1285 /// the binary search tree resulting from lowering a switch instruction.
1286 void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB) {
1288 SDValue CondLHS = getValue(CB.CmpLHS);
1289 DebugLoc dl = getCurDebugLoc();
1291 // Build the setcc now.
1292 if (CB.CmpMHS == NULL) {
1293 // Fold "(X == true)" to X and "(X == false)" to !X to
1294 // handle common cases produced by branch lowering.
1295 if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) &&
1296 CB.CC == ISD::SETEQ)
1298 else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) &&
1299 CB.CC == ISD::SETEQ) {
1300 SDValue True = DAG.getConstant(1, CondLHS.getValueType());
1301 Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True);
1303 Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1305 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now");
1307 const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue();
1308 const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue();
1310 SDValue CmpOp = getValue(CB.CmpMHS);
1311 EVT VT = CmpOp.getValueType();
1313 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) {
1314 Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, VT),
1317 SDValue SUB = DAG.getNode(ISD::SUB, dl,
1318 VT, CmpOp, DAG.getConstant(Low, VT));
1319 Cond = DAG.getSetCC(dl, MVT::i1, SUB,
1320 DAG.getConstant(High-Low, VT), ISD::SETULE);
1324 DAG.AssignOrdering(Cond.getNode(), SDNodeOrder);
1326 // Update successor info
1327 CurMBB->addSuccessor(CB.TrueBB);
1328 CurMBB->addSuccessor(CB.FalseBB);
1330 // Set NextBlock to be the MBB immediately after the current one, if any.
1331 // This is used to avoid emitting unnecessary branches to the next block.
1332 MachineBasicBlock *NextBlock = 0;
1333 MachineFunction::iterator BBI = CurMBB;
1334 if (++BBI != FuncInfo.MF->end())
1337 // If the lhs block is the next block, invert the condition so that we can
1338 // fall through to the lhs instead of the rhs block.
1339 if (CB.TrueBB == NextBlock) {
1340 std::swap(CB.TrueBB, CB.FalseBB);
1341 SDValue True = DAG.getConstant(1, Cond.getValueType());
1342 Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True);
1343 DAG.AssignOrdering(Cond.getNode(), SDNodeOrder);
1346 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1347 MVT::Other, getControlRoot(), Cond,
1348 DAG.getBasicBlock(CB.TrueBB));
1349 DAG.AssignOrdering(BrCond.getNode(), SDNodeOrder);
1351 // If the branch was constant folded, fix up the CFG.
1352 if (BrCond.getOpcode() == ISD::BR) {
1353 CurMBB->removeSuccessor(CB.FalseBB);
1355 // Otherwise, go ahead and insert the false branch.
1356 if (BrCond == getControlRoot())
1357 CurMBB->removeSuccessor(CB.TrueBB);
1359 if (CB.FalseBB != NextBlock) {
1360 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1361 DAG.getBasicBlock(CB.FalseBB));
1363 DAG.AssignOrdering(BrCond.getNode(), SDNodeOrder);
1367 DAG.setRoot(BrCond);
1370 /// visitJumpTable - Emit JumpTable node in the current MBB
1371 void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) {
1372 // Emit the code for the jump table
1373 assert(JT.Reg != -1U && "Should lower JT Header first!");
1374 EVT PTy = TLI.getPointerTy();
1375 SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(),
1377 SDValue Table = DAG.getJumpTable(JT.JTI, PTy);
1378 SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurDebugLoc(),
1379 MVT::Other, Index.getValue(1),
1381 DAG.setRoot(BrJumpTable);
1383 DAG.AssignOrdering(Index.getNode(), SDNodeOrder);
1384 DAG.AssignOrdering(Table.getNode(), SDNodeOrder);
1385 DAG.AssignOrdering(BrJumpTable.getNode(), SDNodeOrder);
1388 /// visitJumpTableHeader - This function emits necessary code to produce index
1389 /// in the JumpTable from switch case.
1390 void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT,
1391 JumpTableHeader &JTH) {
1392 // Subtract the lowest switch case value from the value being switched on and
1393 // conditional branch to default mbb if the result is greater than the
1394 // difference between smallest and largest cases.
1395 SDValue SwitchOp = getValue(JTH.SValue);
1396 EVT VT = SwitchOp.getValueType();
1397 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1398 DAG.getConstant(JTH.First, VT));
1400 // The SDNode we just created, which holds the value being switched on minus
1401 // the the smallest case value, needs to be copied to a virtual register so it
1402 // can be used as an index into the jump table in a subsequent basic block.
1403 // This value may be smaller or larger than the target's pointer type, and
1404 // therefore require extension or truncating.
1405 SwitchOp = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(), TLI.getPointerTy());
1407 unsigned JumpTableReg = FuncInfo.MakeReg(TLI.getPointerTy());
1408 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1409 JumpTableReg, SwitchOp);
1410 JT.Reg = JumpTableReg;
1412 // Emit the range check for the jump table, and branch to the default block
1413 // for the switch statement if the value being switched on exceeds the largest
1414 // case in the switch.
1415 SDValue CMP = DAG.getSetCC(getCurDebugLoc(),
1416 TLI.getSetCCResultType(Sub.getValueType()), Sub,
1417 DAG.getConstant(JTH.Last-JTH.First,VT),
1420 DAG.AssignOrdering(Sub.getNode(), SDNodeOrder);
1421 DAG.AssignOrdering(SwitchOp.getNode(), SDNodeOrder);
1422 DAG.AssignOrdering(CopyTo.getNode(), SDNodeOrder);
1423 DAG.AssignOrdering(CMP.getNode(), SDNodeOrder);
1425 // Set NextBlock to be the MBB immediately after the current one, if any.
1426 // This is used to avoid emitting unnecessary branches to the next block.
1427 MachineBasicBlock *NextBlock = 0;
1428 MachineFunction::iterator BBI = CurMBB;
1430 if (++BBI != FuncInfo.MF->end())
1433 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1434 MVT::Other, CopyTo, CMP,
1435 DAG.getBasicBlock(JT.Default));
1437 DAG.AssignOrdering(BrCond.getNode(), SDNodeOrder);
1439 if (JT.MBB != NextBlock) {
1440 BrCond = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrCond,
1441 DAG.getBasicBlock(JT.MBB));
1442 DAG.AssignOrdering(BrCond.getNode(), SDNodeOrder);
1445 DAG.setRoot(BrCond);
1448 /// visitBitTestHeader - This function emits necessary code to produce value
1449 /// suitable for "bit tests"
1450 void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B) {
1451 // Subtract the minimum value
1452 SDValue SwitchOp = getValue(B.SValue);
1453 EVT VT = SwitchOp.getValueType();
1454 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1455 DAG.getConstant(B.First, VT));
1458 SDValue RangeCmp = DAG.getSetCC(getCurDebugLoc(),
1459 TLI.getSetCCResultType(Sub.getValueType()),
1460 Sub, DAG.getConstant(B.Range, VT),
1463 SDValue ShiftOp = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(),
1464 TLI.getPointerTy());
1466 B.Reg = FuncInfo.MakeReg(TLI.getPointerTy());
1467 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1470 DAG.AssignOrdering(Sub.getNode(), SDNodeOrder);
1471 DAG.AssignOrdering(RangeCmp.getNode(), SDNodeOrder);
1472 DAG.AssignOrdering(ShiftOp.getNode(), SDNodeOrder);
1473 DAG.AssignOrdering(CopyTo.getNode(), SDNodeOrder);
1475 // Set NextBlock to be the MBB immediately after the current one, if any.
1476 // This is used to avoid emitting unnecessary branches to the next block.
1477 MachineBasicBlock *NextBlock = 0;
1478 MachineFunction::iterator BBI = CurMBB;
1479 if (++BBI != FuncInfo.MF->end())
1482 MachineBasicBlock* MBB = B.Cases[0].ThisBB;
1484 CurMBB->addSuccessor(B.Default);
1485 CurMBB->addSuccessor(MBB);
1487 SDValue BrRange = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1488 MVT::Other, CopyTo, RangeCmp,
1489 DAG.getBasicBlock(B.Default));
1491 DAG.AssignOrdering(BrRange.getNode(), SDNodeOrder);
1493 if (MBB != NextBlock) {
1494 BrRange = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, CopyTo,
1495 DAG.getBasicBlock(MBB));
1496 DAG.AssignOrdering(BrRange.getNode(), SDNodeOrder);
1499 DAG.setRoot(BrRange);
1502 /// visitBitTestCase - this function produces one "bit test"
1503 void SelectionDAGBuilder::visitBitTestCase(MachineBasicBlock* NextMBB,
1506 // Make desired shift
1507 SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(), Reg,
1508 TLI.getPointerTy());
1509 SDValue SwitchVal = DAG.getNode(ISD::SHL, getCurDebugLoc(),
1511 DAG.getConstant(1, TLI.getPointerTy()),
1514 // Emit bit tests and jumps
1515 SDValue AndOp = DAG.getNode(ISD::AND, getCurDebugLoc(),
1516 TLI.getPointerTy(), SwitchVal,
1517 DAG.getConstant(B.Mask, TLI.getPointerTy()));
1518 SDValue AndCmp = DAG.getSetCC(getCurDebugLoc(),
1519 TLI.getSetCCResultType(AndOp.getValueType()),
1520 AndOp, DAG.getConstant(0, TLI.getPointerTy()),
1523 DAG.AssignOrdering(ShiftOp.getNode(), SDNodeOrder);
1524 DAG.AssignOrdering(SwitchVal.getNode(), SDNodeOrder);
1525 DAG.AssignOrdering(AndOp.getNode(), SDNodeOrder);
1526 DAG.AssignOrdering(AndCmp.getNode(), SDNodeOrder);
1528 CurMBB->addSuccessor(B.TargetBB);
1529 CurMBB->addSuccessor(NextMBB);
1531 SDValue BrAnd = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1532 MVT::Other, getControlRoot(),
1533 AndCmp, DAG.getBasicBlock(B.TargetBB));
1535 DAG.AssignOrdering(BrAnd.getNode(), SDNodeOrder);
1537 // Set NextBlock to be the MBB immediately after the current one, if any.
1538 // This is used to avoid emitting unnecessary branches to the next block.
1539 MachineBasicBlock *NextBlock = 0;
1540 MachineFunction::iterator BBI = CurMBB;
1541 if (++BBI != FuncInfo.MF->end())
1544 if (NextMBB != NextBlock) {
1545 BrAnd = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrAnd,
1546 DAG.getBasicBlock(NextMBB));
1547 DAG.AssignOrdering(BrAnd.getNode(), SDNodeOrder);
1553 void SelectionDAGBuilder::visitInvoke(InvokeInst &I) {
1554 // Retrieve successors.
1555 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
1556 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)];
1558 const Value *Callee(I.getCalledValue());
1559 if (isa<InlineAsm>(Callee))
1562 LowerCallTo(&I, getValue(Callee), false, LandingPad);
1564 // If the value of the invoke is used outside of its defining block, make it
1565 // available as a virtual register.
1566 CopyToExportRegsIfNeeded(&I);
1568 // Update successor info
1569 CurMBB->addSuccessor(Return);
1570 CurMBB->addSuccessor(LandingPad);
1572 // Drop into normal successor.
1573 SDValue Branch = DAG.getNode(ISD::BR, getCurDebugLoc(),
1574 MVT::Other, getControlRoot(),
1575 DAG.getBasicBlock(Return));
1576 DAG.setRoot(Branch);
1577 DAG.AssignOrdering(Branch.getNode(), SDNodeOrder);
1580 void SelectionDAGBuilder::visitUnwind(UnwindInst &I) {
1583 /// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for
1584 /// small case ranges).
1585 bool SelectionDAGBuilder::handleSmallSwitchRange(CaseRec& CR,
1586 CaseRecVector& WorkList,
1588 MachineBasicBlock* Default) {
1589 Case& BackCase = *(CR.Range.second-1);
1591 // Size is the number of Cases represented by this range.
1592 size_t Size = CR.Range.second - CR.Range.first;
1596 // Get the MachineFunction which holds the current MBB. This is used when
1597 // inserting any additional MBBs necessary to represent the switch.
1598 MachineFunction *CurMF = FuncInfo.MF;
1600 // Figure out which block is immediately after the current one.
1601 MachineBasicBlock *NextBlock = 0;
1602 MachineFunction::iterator BBI = CR.CaseBB;
1604 if (++BBI != FuncInfo.MF->end())
1607 // TODO: If any two of the cases has the same destination, and if one value
1608 // is the same as the other, but has one bit unset that the other has set,
1609 // use bit manipulation to do two compares at once. For example:
1610 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
1612 // Rearrange the case blocks so that the last one falls through if possible.
1613 if (NextBlock && Default != NextBlock && BackCase.BB != NextBlock) {
1614 // The last case block won't fall through into 'NextBlock' if we emit the
1615 // branches in this order. See if rearranging a case value would help.
1616 for (CaseItr I = CR.Range.first, E = CR.Range.second-1; I != E; ++I) {
1617 if (I->BB == NextBlock) {
1618 std::swap(*I, BackCase);
1624 // Create a CaseBlock record representing a conditional branch to
1625 // the Case's target mbb if the value being switched on SV is equal
1627 MachineBasicBlock *CurBlock = CR.CaseBB;
1628 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
1629 MachineBasicBlock *FallThrough;
1631 FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock());
1632 CurMF->insert(BBI, FallThrough);
1634 // Put SV in a virtual register to make it available from the new blocks.
1635 ExportFromCurrentBlock(SV);
1637 // If the last case doesn't match, go to the default block.
1638 FallThrough = Default;
1641 Value *RHS, *LHS, *MHS;
1643 if (I->High == I->Low) {
1644 // This is just small small case range :) containing exactly 1 case
1646 LHS = SV; RHS = I->High; MHS = NULL;
1649 LHS = I->Low; MHS = SV; RHS = I->High;
1651 CaseBlock CB(CC, LHS, RHS, MHS, I->BB, FallThrough, CurBlock);
1653 // If emitting the first comparison, just call visitSwitchCase to emit the
1654 // code into the current block. Otherwise, push the CaseBlock onto the
1655 // vector to be later processed by SDISel, and insert the node's MBB
1656 // before the next MBB.
1657 if (CurBlock == CurMBB)
1658 visitSwitchCase(CB);
1660 SwitchCases.push_back(CB);
1662 CurBlock = FallThrough;
1668 static inline bool areJTsAllowed(const TargetLowering &TLI) {
1669 return !DisableJumpTables &&
1670 (TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
1671 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other));
1674 static APInt ComputeRange(const APInt &First, const APInt &Last) {
1675 APInt LastExt(Last), FirstExt(First);
1676 uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1;
1677 LastExt.sext(BitWidth); FirstExt.sext(BitWidth);
1678 return (LastExt - FirstExt + 1ULL);
1681 /// handleJTSwitchCase - Emit jumptable for current switch case range
1682 bool SelectionDAGBuilder::handleJTSwitchCase(CaseRec& CR,
1683 CaseRecVector& WorkList,
1685 MachineBasicBlock* Default) {
1686 Case& FrontCase = *CR.Range.first;
1687 Case& BackCase = *(CR.Range.second-1);
1689 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
1690 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
1692 APInt TSize(First.getBitWidth(), 0);
1693 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1697 if (!areJTsAllowed(TLI) || TSize.ult(APInt(First.getBitWidth(), 4)))
1700 APInt Range = ComputeRange(First, Last);
1701 double Density = TSize.roundToDouble() / Range.roundToDouble();
1705 DEBUG(dbgs() << "Lowering jump table\n"
1706 << "First entry: " << First << ". Last entry: " << Last << '\n'
1707 << "Range: " << Range
1708 << "Size: " << TSize << ". Density: " << Density << "\n\n");
1710 // Get the MachineFunction which holds the current MBB. This is used when
1711 // inserting any additional MBBs necessary to represent the switch.
1712 MachineFunction *CurMF = FuncInfo.MF;
1714 // Figure out which block is immediately after the current one.
1715 MachineFunction::iterator BBI = CR.CaseBB;
1718 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1720 // Create a new basic block to hold the code for loading the address
1721 // of the jump table, and jumping to it. Update successor information;
1722 // we will either branch to the default case for the switch, or the jump
1724 MachineBasicBlock *JumpTableBB = CurMF->CreateMachineBasicBlock(LLVMBB);
1725 CurMF->insert(BBI, JumpTableBB);
1726 CR.CaseBB->addSuccessor(Default);
1727 CR.CaseBB->addSuccessor(JumpTableBB);
1729 // Build a vector of destination BBs, corresponding to each target
1730 // of the jump table. If the value of the jump table slot corresponds to
1731 // a case statement, push the case's BB onto the vector, otherwise, push
1733 std::vector<MachineBasicBlock*> DestBBs;
1735 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) {
1736 const APInt &Low = cast<ConstantInt>(I->Low)->getValue();
1737 const APInt &High = cast<ConstantInt>(I->High)->getValue();
1739 if (Low.sle(TEI) && TEI.sle(High)) {
1740 DestBBs.push_back(I->BB);
1744 DestBBs.push_back(Default);
1748 // Update successor info. Add one edge to each unique successor.
1749 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs());
1750 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(),
1751 E = DestBBs.end(); I != E; ++I) {
1752 if (!SuccsHandled[(*I)->getNumber()]) {
1753 SuccsHandled[(*I)->getNumber()] = true;
1754 JumpTableBB->addSuccessor(*I);
1758 // Create a jump table index for this jump table, or return an existing
1760 unsigned JTEncoding = TLI.getJumpTableEncoding();
1761 unsigned JTI = CurMF->getOrCreateJumpTableInfo(JTEncoding)
1762 ->getJumpTableIndex(DestBBs);
1764 // Set the jump table information so that we can codegen it as a second
1765 // MachineBasicBlock
1766 JumpTable JT(-1U, JTI, JumpTableBB, Default);
1767 JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == CurMBB));
1768 if (CR.CaseBB == CurMBB)
1769 visitJumpTableHeader(JT, JTH);
1771 JTCases.push_back(JumpTableBlock(JTH, JT));
1776 /// handleBTSplitSwitchCase - emit comparison and split binary search tree into
1778 bool SelectionDAGBuilder::handleBTSplitSwitchCase(CaseRec& CR,
1779 CaseRecVector& WorkList,
1781 MachineBasicBlock* Default) {
1782 // Get the MachineFunction which holds the current MBB. This is used when
1783 // inserting any additional MBBs necessary to represent the switch.
1784 MachineFunction *CurMF = FuncInfo.MF;
1786 // Figure out which block is immediately after the current one.
1787 MachineFunction::iterator BBI = CR.CaseBB;
1790 Case& FrontCase = *CR.Range.first;
1791 Case& BackCase = *(CR.Range.second-1);
1792 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1794 // Size is the number of Cases represented by this range.
1795 unsigned Size = CR.Range.second - CR.Range.first;
1797 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
1798 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
1800 CaseItr Pivot = CR.Range.first + Size/2;
1802 // Select optimal pivot, maximizing sum density of LHS and RHS. This will
1803 // (heuristically) allow us to emit JumpTable's later.
1804 APInt TSize(First.getBitWidth(), 0);
1805 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1809 APInt LSize = FrontCase.size();
1810 APInt RSize = TSize-LSize;
1811 DEBUG(dbgs() << "Selecting best pivot: \n"
1812 << "First: " << First << ", Last: " << Last <<'\n'
1813 << "LSize: " << LSize << ", RSize: " << RSize << '\n');
1814 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second;
1816 const APInt &LEnd = cast<ConstantInt>(I->High)->getValue();
1817 const APInt &RBegin = cast<ConstantInt>(J->Low)->getValue();
1818 APInt Range = ComputeRange(LEnd, RBegin);
1819 assert((Range - 2ULL).isNonNegative() &&
1820 "Invalid case distance");
1821 double LDensity = (double)LSize.roundToDouble() /
1822 (LEnd - First + 1ULL).roundToDouble();
1823 double RDensity = (double)RSize.roundToDouble() /
1824 (Last - RBegin + 1ULL).roundToDouble();
1825 double Metric = Range.logBase2()*(LDensity+RDensity);
1826 // Should always split in some non-trivial place
1827 DEBUG(dbgs() <<"=>Step\n"
1828 << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n'
1829 << "LDensity: " << LDensity
1830 << ", RDensity: " << RDensity << '\n'
1831 << "Metric: " << Metric << '\n');
1832 if (FMetric < Metric) {
1835 DEBUG(dbgs() << "Current metric set to: " << FMetric << '\n');
1841 if (areJTsAllowed(TLI)) {
1842 // If our case is dense we *really* should handle it earlier!
1843 assert((FMetric > 0) && "Should handle dense range earlier!");
1845 Pivot = CR.Range.first + Size/2;
1848 CaseRange LHSR(CR.Range.first, Pivot);
1849 CaseRange RHSR(Pivot, CR.Range.second);
1850 Constant *C = Pivot->Low;
1851 MachineBasicBlock *FalseBB = 0, *TrueBB = 0;
1853 // We know that we branch to the LHS if the Value being switched on is
1854 // less than the Pivot value, C. We use this to optimize our binary
1855 // tree a bit, by recognizing that if SV is greater than or equal to the
1856 // LHS's Case Value, and that Case Value is exactly one less than the
1857 // Pivot's Value, then we can branch directly to the LHS's Target,
1858 // rather than creating a leaf node for it.
1859 if ((LHSR.second - LHSR.first) == 1 &&
1860 LHSR.first->High == CR.GE &&
1861 cast<ConstantInt>(C)->getValue() ==
1862 (cast<ConstantInt>(CR.GE)->getValue() + 1LL)) {
1863 TrueBB = LHSR.first->BB;
1865 TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB);
1866 CurMF->insert(BBI, TrueBB);
1867 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR));
1869 // Put SV in a virtual register to make it available from the new blocks.
1870 ExportFromCurrentBlock(SV);
1873 // Similar to the optimization above, if the Value being switched on is
1874 // known to be less than the Constant CR.LT, and the current Case Value
1875 // is CR.LT - 1, then we can branch directly to the target block for
1876 // the current Case Value, rather than emitting a RHS leaf node for it.
1877 if ((RHSR.second - RHSR.first) == 1 && CR.LT &&
1878 cast<ConstantInt>(RHSR.first->Low)->getValue() ==
1879 (cast<ConstantInt>(CR.LT)->getValue() - 1LL)) {
1880 FalseBB = RHSR.first->BB;
1882 FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
1883 CurMF->insert(BBI, FalseBB);
1884 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR));
1886 // Put SV in a virtual register to make it available from the new blocks.
1887 ExportFromCurrentBlock(SV);
1890 // Create a CaseBlock record representing a conditional branch to
1891 // the LHS node if the value being switched on SV is less than C.
1892 // Otherwise, branch to LHS.
1893 CaseBlock CB(ISD::SETLT, SV, C, NULL, TrueBB, FalseBB, CR.CaseBB);
1895 if (CR.CaseBB == CurMBB)
1896 visitSwitchCase(CB);
1898 SwitchCases.push_back(CB);
1903 /// handleBitTestsSwitchCase - if current case range has few destination and
1904 /// range span less, than machine word bitwidth, encode case range into series
1905 /// of masks and emit bit tests with these masks.
1906 bool SelectionDAGBuilder::handleBitTestsSwitchCase(CaseRec& CR,
1907 CaseRecVector& WorkList,
1909 MachineBasicBlock* Default){
1910 EVT PTy = TLI.getPointerTy();
1911 unsigned IntPtrBits = PTy.getSizeInBits();
1913 Case& FrontCase = *CR.Range.first;
1914 Case& BackCase = *(CR.Range.second-1);
1916 // Get the MachineFunction which holds the current MBB. This is used when
1917 // inserting any additional MBBs necessary to represent the switch.
1918 MachineFunction *CurMF = FuncInfo.MF;
1920 // If target does not have legal shift left, do not emit bit tests at all.
1921 if (!TLI.isOperationLegal(ISD::SHL, TLI.getPointerTy()))
1925 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1927 // Single case counts one, case range - two.
1928 numCmps += (I->Low == I->High ? 1 : 2);
1931 // Count unique destinations
1932 SmallSet<MachineBasicBlock*, 4> Dests;
1933 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
1934 Dests.insert(I->BB);
1935 if (Dests.size() > 3)
1936 // Don't bother the code below, if there are too much unique destinations
1939 DEBUG(dbgs() << "Total number of unique destinations: "
1940 << Dests.size() << '\n'
1941 << "Total number of comparisons: " << numCmps << '\n');
1943 // Compute span of values.
1944 const APInt& minValue = cast<ConstantInt>(FrontCase.Low)->getValue();
1945 const APInt& maxValue = cast<ConstantInt>(BackCase.High)->getValue();
1946 APInt cmpRange = maxValue - minValue;
1948 DEBUG(dbgs() << "Compare range: " << cmpRange << '\n'
1949 << "Low bound: " << minValue << '\n'
1950 << "High bound: " << maxValue << '\n');
1952 if (cmpRange.uge(APInt(cmpRange.getBitWidth(), IntPtrBits)) ||
1953 (!(Dests.size() == 1 && numCmps >= 3) &&
1954 !(Dests.size() == 2 && numCmps >= 5) &&
1955 !(Dests.size() >= 3 && numCmps >= 6)))
1958 DEBUG(dbgs() << "Emitting bit tests\n");
1959 APInt lowBound = APInt::getNullValue(cmpRange.getBitWidth());
1961 // Optimize the case where all the case values fit in a
1962 // word without having to subtract minValue. In this case,
1963 // we can optimize away the subtraction.
1964 if (minValue.isNonNegative() &&
1965 maxValue.slt(APInt(maxValue.getBitWidth(), IntPtrBits))) {
1966 cmpRange = maxValue;
1968 lowBound = minValue;
1971 CaseBitsVector CasesBits;
1972 unsigned i, count = 0;
1974 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
1975 MachineBasicBlock* Dest = I->BB;
1976 for (i = 0; i < count; ++i)
1977 if (Dest == CasesBits[i].BB)
1981 assert((count < 3) && "Too much destinations to test!");
1982 CasesBits.push_back(CaseBits(0, Dest, 0));
1986 const APInt& lowValue = cast<ConstantInt>(I->Low)->getValue();
1987 const APInt& highValue = cast<ConstantInt>(I->High)->getValue();
1989 uint64_t lo = (lowValue - lowBound).getZExtValue();
1990 uint64_t hi = (highValue - lowBound).getZExtValue();
1992 for (uint64_t j = lo; j <= hi; j++) {
1993 CasesBits[i].Mask |= 1ULL << j;
1994 CasesBits[i].Bits++;
1998 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp());
2002 // Figure out which block is immediately after the current one.
2003 MachineFunction::iterator BBI = CR.CaseBB;
2006 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2008 DEBUG(dbgs() << "Cases:\n");
2009 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) {
2010 DEBUG(dbgs() << "Mask: " << CasesBits[i].Mask
2011 << ", Bits: " << CasesBits[i].Bits
2012 << ", BB: " << CasesBits[i].BB << '\n');
2014 MachineBasicBlock *CaseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2015 CurMF->insert(BBI, CaseBB);
2016 BTC.push_back(BitTestCase(CasesBits[i].Mask,
2020 // Put SV in a virtual register to make it available from the new blocks.
2021 ExportFromCurrentBlock(SV);
2024 BitTestBlock BTB(lowBound, cmpRange, SV,
2025 -1U, (CR.CaseBB == CurMBB),
2026 CR.CaseBB, Default, BTC);
2028 if (CR.CaseBB == CurMBB)
2029 visitBitTestHeader(BTB);
2031 BitTestCases.push_back(BTB);
2036 /// Clusterify - Transform simple list of Cases into list of CaseRange's
2037 size_t SelectionDAGBuilder::Clusterify(CaseVector& Cases,
2038 const SwitchInst& SI) {
2041 // Start with "simple" cases
2042 for (size_t i = 1; i < SI.getNumSuccessors(); ++i) {
2043 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SI.getSuccessor(i)];
2044 Cases.push_back(Case(SI.getSuccessorValue(i),
2045 SI.getSuccessorValue(i),
2048 std::sort(Cases.begin(), Cases.end(), CaseCmp());
2050 // Merge case into clusters
2051 if (Cases.size() >= 2)
2052 // Must recompute end() each iteration because it may be
2053 // invalidated by erase if we hold on to it
2054 for (CaseItr I = Cases.begin(), J = ++(Cases.begin()); J != Cases.end(); ) {
2055 const APInt& nextValue = cast<ConstantInt>(J->Low)->getValue();
2056 const APInt& currentValue = cast<ConstantInt>(I->High)->getValue();
2057 MachineBasicBlock* nextBB = J->BB;
2058 MachineBasicBlock* currentBB = I->BB;
2060 // If the two neighboring cases go to the same destination, merge them
2061 // into a single case.
2062 if ((nextValue - currentValue == 1) && (currentBB == nextBB)) {
2070 for (CaseItr I=Cases.begin(), E=Cases.end(); I!=E; ++I, ++numCmps) {
2071 if (I->Low != I->High)
2072 // A range counts double, since it requires two compares.
2079 void SelectionDAGBuilder::visitSwitch(SwitchInst &SI) {
2080 // Figure out which block is immediately after the current one.
2081 MachineBasicBlock *NextBlock = 0;
2082 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()];
2084 // If there is only the default destination, branch to it if it is not the
2085 // next basic block. Otherwise, just fall through.
2086 if (SI.getNumOperands() == 2) {
2087 // Update machine-CFG edges.
2089 // If this is not a fall-through branch, emit the branch.
2090 CurMBB->addSuccessor(Default);
2091 if (Default != NextBlock) {
2092 SDValue Res = DAG.getNode(ISD::BR, getCurDebugLoc(),
2093 MVT::Other, getControlRoot(),
2094 DAG.getBasicBlock(Default));
2096 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2102 // If there are any non-default case statements, create a vector of Cases
2103 // representing each one, and sort the vector so that we can efficiently
2104 // create a binary search tree from them.
2106 size_t numCmps = Clusterify(Cases, SI);
2107 DEBUG(dbgs() << "Clusterify finished. Total clusters: " << Cases.size()
2108 << ". Total compares: " << numCmps << '\n');
2111 // Get the Value to be switched on and default basic blocks, which will be
2112 // inserted into CaseBlock records, representing basic blocks in the binary
2114 Value *SV = SI.getOperand(0);
2116 // Push the initial CaseRec onto the worklist
2117 CaseRecVector WorkList;
2118 WorkList.push_back(CaseRec(CurMBB,0,0,CaseRange(Cases.begin(),Cases.end())));
2120 while (!WorkList.empty()) {
2121 // Grab a record representing a case range to process off the worklist
2122 CaseRec CR = WorkList.back();
2123 WorkList.pop_back();
2125 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default))
2128 // If the range has few cases (two or less) emit a series of specific
2130 if (handleSmallSwitchRange(CR, WorkList, SV, Default))
2133 // If the switch has more than 5 blocks, and at least 40% dense, and the
2134 // target supports indirect branches, then emit a jump table rather than
2135 // lowering the switch to a binary tree of conditional branches.
2136 if (handleJTSwitchCase(CR, WorkList, SV, Default))
2139 // Emit binary tree. We need to pick a pivot, and push left and right ranges
2140 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call.
2141 handleBTSplitSwitchCase(CR, WorkList, SV, Default);
2145 void SelectionDAGBuilder::visitIndirectBr(IndirectBrInst &I) {
2146 // Update machine-CFG edges.
2147 for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i)
2148 CurMBB->addSuccessor(FuncInfo.MBBMap[I.getSuccessor(i)]);
2150 SDValue Res = DAG.getNode(ISD::BRIND, getCurDebugLoc(),
2151 MVT::Other, getControlRoot(),
2152 getValue(I.getAddress()));
2154 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2157 void SelectionDAGBuilder::visitFSub(User &I) {
2158 // -0.0 - X --> fneg
2159 const Type *Ty = I.getType();
2160 if (isa<VectorType>(Ty)) {
2161 if (ConstantVector *CV = dyn_cast<ConstantVector>(I.getOperand(0))) {
2162 const VectorType *DestTy = cast<VectorType>(I.getType());
2163 const Type *ElTy = DestTy->getElementType();
2164 unsigned VL = DestTy->getNumElements();
2165 std::vector<Constant*> NZ(VL, ConstantFP::getNegativeZero(ElTy));
2166 Constant *CNZ = ConstantVector::get(&NZ[0], NZ.size());
2168 SDValue Op2 = getValue(I.getOperand(1));
2169 SDValue Res = DAG.getNode(ISD::FNEG, getCurDebugLoc(),
2170 Op2.getValueType(), Op2);
2172 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2178 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I.getOperand(0)))
2179 if (CFP->isExactlyValue(ConstantFP::getNegativeZero(Ty)->getValueAPF())) {
2180 SDValue Op2 = getValue(I.getOperand(1));
2181 SDValue Res = DAG.getNode(ISD::FNEG, getCurDebugLoc(),
2182 Op2.getValueType(), Op2);
2184 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2188 visitBinary(I, ISD::FSUB);
2191 void SelectionDAGBuilder::visitBinary(User &I, unsigned OpCode) {
2192 SDValue Op1 = getValue(I.getOperand(0));
2193 SDValue Op2 = getValue(I.getOperand(1));
2194 SDValue Res = DAG.getNode(OpCode, getCurDebugLoc(),
2195 Op1.getValueType(), Op1, Op2);
2197 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2200 void SelectionDAGBuilder::visitShift(User &I, unsigned Opcode) {
2201 SDValue Op1 = getValue(I.getOperand(0));
2202 SDValue Op2 = getValue(I.getOperand(1));
2203 if (!isa<VectorType>(I.getType()) &&
2204 Op2.getValueType() != TLI.getShiftAmountTy()) {
2205 // If the operand is smaller than the shift count type, promote it.
2206 EVT PTy = TLI.getPointerTy();
2207 EVT STy = TLI.getShiftAmountTy();
2208 if (STy.bitsGT(Op2.getValueType()))
2209 Op2 = DAG.getNode(ISD::ANY_EXTEND, getCurDebugLoc(),
2210 TLI.getShiftAmountTy(), Op2);
2211 // If the operand is larger than the shift count type but the shift
2212 // count type has enough bits to represent any shift value, truncate
2213 // it now. This is a common case and it exposes the truncate to
2214 // optimization early.
2215 else if (STy.getSizeInBits() >=
2216 Log2_32_Ceil(Op2.getValueType().getSizeInBits()))
2217 Op2 = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2218 TLI.getShiftAmountTy(), Op2);
2219 // Otherwise we'll need to temporarily settle for some other
2220 // convenient type; type legalization will make adjustments as
2222 else if (PTy.bitsLT(Op2.getValueType()))
2223 Op2 = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2224 TLI.getPointerTy(), Op2);
2225 else if (PTy.bitsGT(Op2.getValueType()))
2226 Op2 = DAG.getNode(ISD::ANY_EXTEND, getCurDebugLoc(),
2227 TLI.getPointerTy(), Op2);
2230 SDValue Res = DAG.getNode(Opcode, getCurDebugLoc(),
2231 Op1.getValueType(), Op1, Op2);
2233 DAG.AssignOrdering(Op1.getNode(), SDNodeOrder);
2234 DAG.AssignOrdering(Op2.getNode(), SDNodeOrder);
2235 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2238 void SelectionDAGBuilder::visitICmp(User &I) {
2239 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2240 if (ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2241 predicate = IC->getPredicate();
2242 else if (ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2243 predicate = ICmpInst::Predicate(IC->getPredicate());
2244 SDValue Op1 = getValue(I.getOperand(0));
2245 SDValue Op2 = getValue(I.getOperand(1));
2246 ISD::CondCode Opcode = getICmpCondCode(predicate);
2248 EVT DestVT = TLI.getValueType(I.getType());
2249 SDValue Res = DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Opcode);
2251 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2254 void SelectionDAGBuilder::visitFCmp(User &I) {
2255 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2256 if (FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2257 predicate = FC->getPredicate();
2258 else if (ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2259 predicate = FCmpInst::Predicate(FC->getPredicate());
2260 SDValue Op1 = getValue(I.getOperand(0));
2261 SDValue Op2 = getValue(I.getOperand(1));
2262 ISD::CondCode Condition = getFCmpCondCode(predicate);
2263 EVT DestVT = TLI.getValueType(I.getType());
2264 SDValue Res = DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Condition);
2266 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2269 void SelectionDAGBuilder::visitSelect(User &I) {
2270 SmallVector<EVT, 4> ValueVTs;
2271 ComputeValueVTs(TLI, I.getType(), ValueVTs);
2272 unsigned NumValues = ValueVTs.size();
2273 if (NumValues == 0) return;
2275 SmallVector<SDValue, 4> Values(NumValues);
2276 SDValue Cond = getValue(I.getOperand(0));
2277 SDValue TrueVal = getValue(I.getOperand(1));
2278 SDValue FalseVal = getValue(I.getOperand(2));
2280 for (unsigned i = 0; i != NumValues; ++i) {
2281 Values[i] = DAG.getNode(ISD::SELECT, getCurDebugLoc(),
2282 TrueVal.getNode()->getValueType(i), Cond,
2283 SDValue(TrueVal.getNode(),
2284 TrueVal.getResNo() + i),
2285 SDValue(FalseVal.getNode(),
2286 FalseVal.getResNo() + i));
2288 DAG.AssignOrdering(Values[i].getNode(), SDNodeOrder);
2291 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2292 DAG.getVTList(&ValueVTs[0], NumValues),
2293 &Values[0], NumValues);
2295 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2298 void SelectionDAGBuilder::visitTrunc(User &I) {
2299 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2300 SDValue N = getValue(I.getOperand(0));
2301 EVT DestVT = TLI.getValueType(I.getType());
2302 SDValue Res = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), DestVT, N);
2304 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2307 void SelectionDAGBuilder::visitZExt(User &I) {
2308 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2309 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2310 SDValue N = getValue(I.getOperand(0));
2311 EVT DestVT = TLI.getValueType(I.getType());
2312 SDValue Res = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(), DestVT, N);
2314 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2317 void SelectionDAGBuilder::visitSExt(User &I) {
2318 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2319 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2320 SDValue N = getValue(I.getOperand(0));
2321 EVT DestVT = TLI.getValueType(I.getType());
2322 SDValue Res = DAG.getNode(ISD::SIGN_EXTEND, getCurDebugLoc(), DestVT, N);
2324 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2327 void SelectionDAGBuilder::visitFPTrunc(User &I) {
2328 // FPTrunc is never a no-op cast, no need to check
2329 SDValue N = getValue(I.getOperand(0));
2330 EVT DestVT = TLI.getValueType(I.getType());
2331 SDValue Res = DAG.getNode(ISD::FP_ROUND, getCurDebugLoc(),
2332 DestVT, N, DAG.getIntPtrConstant(0));
2334 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2337 void SelectionDAGBuilder::visitFPExt(User &I){
2338 // FPTrunc is never a no-op cast, no need to check
2339 SDValue N = getValue(I.getOperand(0));
2340 EVT DestVT = TLI.getValueType(I.getType());
2341 SDValue Res = DAG.getNode(ISD::FP_EXTEND, getCurDebugLoc(), DestVT, N);
2343 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2346 void SelectionDAGBuilder::visitFPToUI(User &I) {
2347 // FPToUI is never a no-op cast, no need to check
2348 SDValue N = getValue(I.getOperand(0));
2349 EVT DestVT = TLI.getValueType(I.getType());
2350 SDValue Res = DAG.getNode(ISD::FP_TO_UINT, getCurDebugLoc(), DestVT, N);
2352 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2355 void SelectionDAGBuilder::visitFPToSI(User &I) {
2356 // FPToSI is never a no-op cast, no need to check
2357 SDValue N = getValue(I.getOperand(0));
2358 EVT DestVT = TLI.getValueType(I.getType());
2359 SDValue Res = DAG.getNode(ISD::FP_TO_SINT, getCurDebugLoc(), DestVT, N);
2361 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2364 void SelectionDAGBuilder::visitUIToFP(User &I) {
2365 // UIToFP is never a no-op cast, no need to check
2366 SDValue N = getValue(I.getOperand(0));
2367 EVT DestVT = TLI.getValueType(I.getType());
2368 SDValue Res = DAG.getNode(ISD::UINT_TO_FP, getCurDebugLoc(), DestVT, N);
2370 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2373 void SelectionDAGBuilder::visitSIToFP(User &I){
2374 // SIToFP is never a no-op cast, no need to check
2375 SDValue N = getValue(I.getOperand(0));
2376 EVT DestVT = TLI.getValueType(I.getType());
2377 SDValue Res = DAG.getNode(ISD::SINT_TO_FP, getCurDebugLoc(), DestVT, N);
2379 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2382 void SelectionDAGBuilder::visitPtrToInt(User &I) {
2383 // What to do depends on the size of the integer and the size of the pointer.
2384 // We can either truncate, zero extend, or no-op, accordingly.
2385 SDValue N = getValue(I.getOperand(0));
2386 EVT SrcVT = N.getValueType();
2387 EVT DestVT = TLI.getValueType(I.getType());
2388 SDValue Res = DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT);
2390 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2393 void SelectionDAGBuilder::visitIntToPtr(User &I) {
2394 // What to do depends on the size of the integer and the size of the pointer.
2395 // We can either truncate, zero extend, or no-op, accordingly.
2396 SDValue N = getValue(I.getOperand(0));
2397 EVT SrcVT = N.getValueType();
2398 EVT DestVT = TLI.getValueType(I.getType());
2399 SDValue Res = DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT);
2401 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2404 void SelectionDAGBuilder::visitBitCast(User &I) {
2405 SDValue N = getValue(I.getOperand(0));
2406 EVT DestVT = TLI.getValueType(I.getType());
2408 // BitCast assures us that source and destination are the same size so this is
2409 // either a BIT_CONVERT or a no-op.
2410 if (DestVT != N.getValueType()) {
2411 SDValue Res = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
2412 DestVT, N); // convert types.
2414 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2416 setValue(&I, N); // noop cast.
2420 void SelectionDAGBuilder::visitInsertElement(User &I) {
2421 SDValue InVec = getValue(I.getOperand(0));
2422 SDValue InVal = getValue(I.getOperand(1));
2423 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2425 getValue(I.getOperand(2)));
2426 SDValue Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurDebugLoc(),
2427 TLI.getValueType(I.getType()),
2428 InVec, InVal, InIdx);
2431 DAG.AssignOrdering(InIdx.getNode(), SDNodeOrder);
2432 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2435 void SelectionDAGBuilder::visitExtractElement(User &I) {
2436 SDValue InVec = getValue(I.getOperand(0));
2437 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2439 getValue(I.getOperand(1)));
2440 SDValue Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2441 TLI.getValueType(I.getType()), InVec, InIdx);
2444 DAG.AssignOrdering(InIdx.getNode(), SDNodeOrder);
2445 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2449 // Utility for visitShuffleVector - Returns true if the mask is mask starting
2450 // from SIndx and increasing to the element length (undefs are allowed).
2451 static bool SequentialMask(SmallVectorImpl<int> &Mask, unsigned SIndx) {
2452 unsigned MaskNumElts = Mask.size();
2453 for (unsigned i = 0; i != MaskNumElts; ++i)
2454 if ((Mask[i] >= 0) && (Mask[i] != (int)(i + SIndx)))
2459 void SelectionDAGBuilder::visitShuffleVector(User &I) {
2460 SmallVector<int, 8> Mask;
2461 SDValue Src1 = getValue(I.getOperand(0));
2462 SDValue Src2 = getValue(I.getOperand(1));
2464 // Convert the ConstantVector mask operand into an array of ints, with -1
2465 // representing undef values.
2466 SmallVector<Constant*, 8> MaskElts;
2467 cast<Constant>(I.getOperand(2))->getVectorElements(*DAG.getContext(),
2469 unsigned MaskNumElts = MaskElts.size();
2470 for (unsigned i = 0; i != MaskNumElts; ++i) {
2471 if (isa<UndefValue>(MaskElts[i]))
2474 Mask.push_back(cast<ConstantInt>(MaskElts[i])->getSExtValue());
2477 EVT VT = TLI.getValueType(I.getType());
2478 EVT SrcVT = Src1.getValueType();
2479 unsigned SrcNumElts = SrcVT.getVectorNumElements();
2481 if (SrcNumElts == MaskNumElts) {
2482 SDValue Res = DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2485 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2489 // Normalize the shuffle vector since mask and vector length don't match.
2490 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) {
2491 // Mask is longer than the source vectors and is a multiple of the source
2492 // vectors. We can use concatenate vector to make the mask and vectors
2494 if (SrcNumElts*2 == MaskNumElts && SequentialMask(Mask, 0)) {
2495 // The shuffle is concatenating two vectors together.
2496 SDValue Res = DAG.getNode(ISD::CONCAT_VECTORS, getCurDebugLoc(),
2499 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2503 // Pad both vectors with undefs to make them the same length as the mask.
2504 unsigned NumConcat = MaskNumElts / SrcNumElts;
2505 bool Src1U = Src1.getOpcode() == ISD::UNDEF;
2506 bool Src2U = Src2.getOpcode() == ISD::UNDEF;
2507 SDValue UndefVal = DAG.getUNDEF(SrcVT);
2509 SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal);
2510 SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal);
2514 Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2515 getCurDebugLoc(), VT,
2516 &MOps1[0], NumConcat);
2517 Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2518 getCurDebugLoc(), VT,
2519 &MOps2[0], NumConcat);
2521 // Readjust mask for new input vector length.
2522 SmallVector<int, 8> MappedOps;
2523 for (unsigned i = 0; i != MaskNumElts; ++i) {
2525 if (Idx < (int)SrcNumElts)
2526 MappedOps.push_back(Idx);
2528 MappedOps.push_back(Idx + MaskNumElts - SrcNumElts);
2531 SDValue Res = DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2534 DAG.AssignOrdering(Src1.getNode(), SDNodeOrder);
2535 DAG.AssignOrdering(Src2.getNode(), SDNodeOrder);
2536 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2540 if (SrcNumElts > MaskNumElts) {
2541 // Analyze the access pattern of the vector to see if we can extract
2542 // two subvectors and do the shuffle. The analysis is done by calculating
2543 // the range of elements the mask access on both vectors.
2544 int MinRange[2] = { SrcNumElts+1, SrcNumElts+1};
2545 int MaxRange[2] = {-1, -1};
2547 for (unsigned i = 0; i != MaskNumElts; ++i) {
2553 if (Idx >= (int)SrcNumElts) {
2557 if (Idx > MaxRange[Input])
2558 MaxRange[Input] = Idx;
2559 if (Idx < MinRange[Input])
2560 MinRange[Input] = Idx;
2563 // Check if the access is smaller than the vector size and can we find
2564 // a reasonable extract index.
2565 int RangeUse[2] = { 2, 2 }; // 0 = Unused, 1 = Extract, 2 = Can not
2567 int StartIdx[2]; // StartIdx to extract from
2568 for (int Input=0; Input < 2; ++Input) {
2569 if (MinRange[Input] == (int)(SrcNumElts+1) && MaxRange[Input] == -1) {
2570 RangeUse[Input] = 0; // Unused
2571 StartIdx[Input] = 0;
2572 } else if (MaxRange[Input] - MinRange[Input] < (int)MaskNumElts) {
2573 // Fits within range but we should see if we can find a good
2574 // start index that is a multiple of the mask length.
2575 if (MaxRange[Input] < (int)MaskNumElts) {
2576 RangeUse[Input] = 1; // Extract from beginning of the vector
2577 StartIdx[Input] = 0;
2579 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts;
2580 if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts &&
2581 StartIdx[Input] + MaskNumElts < SrcNumElts)
2582 RangeUse[Input] = 1; // Extract from a multiple of the mask length.
2587 if (RangeUse[0] == 0 && RangeUse[1] == 0) {
2588 SDValue Res = DAG.getUNDEF(VT);
2589 setValue(&I, Res); // Vectors are not used.
2590 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2593 else if (RangeUse[0] < 2 && RangeUse[1] < 2) {
2594 // Extract appropriate subvector and generate a vector shuffle
2595 for (int Input=0; Input < 2; ++Input) {
2596 SDValue &Src = Input == 0 ? Src1 : Src2;
2597 if (RangeUse[Input] == 0)
2598 Src = DAG.getUNDEF(VT);
2600 Src = DAG.getNode(ISD::EXTRACT_SUBVECTOR, getCurDebugLoc(), VT,
2601 Src, DAG.getIntPtrConstant(StartIdx[Input]));
2603 DAG.AssignOrdering(Src.getNode(), SDNodeOrder);
2606 // Calculate new mask.
2607 SmallVector<int, 8> MappedOps;
2608 for (unsigned i = 0; i != MaskNumElts; ++i) {
2611 MappedOps.push_back(Idx);
2612 else if (Idx < (int)SrcNumElts)
2613 MappedOps.push_back(Idx - StartIdx[0]);
2615 MappedOps.push_back(Idx - SrcNumElts - StartIdx[1] + MaskNumElts);
2618 SDValue Res = DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2621 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2626 // We can't use either concat vectors or extract subvectors so fall back to
2627 // replacing the shuffle with extract and build vector.
2628 // to insert and build vector.
2629 EVT EltVT = VT.getVectorElementType();
2630 EVT PtrVT = TLI.getPointerTy();
2631 SmallVector<SDValue,8> Ops;
2632 for (unsigned i = 0; i != MaskNumElts; ++i) {
2634 Ops.push_back(DAG.getUNDEF(EltVT));
2639 if (Idx < (int)SrcNumElts)
2640 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2641 EltVT, Src1, DAG.getConstant(Idx, PtrVT));
2643 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2645 DAG.getConstant(Idx - SrcNumElts, PtrVT));
2648 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2652 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
2653 VT, &Ops[0], Ops.size());
2655 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2658 void SelectionDAGBuilder::visitInsertValue(InsertValueInst &I) {
2659 const Value *Op0 = I.getOperand(0);
2660 const Value *Op1 = I.getOperand(1);
2661 const Type *AggTy = I.getType();
2662 const Type *ValTy = Op1->getType();
2663 bool IntoUndef = isa<UndefValue>(Op0);
2664 bool FromUndef = isa<UndefValue>(Op1);
2666 unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy,
2667 I.idx_begin(), I.idx_end());
2669 SmallVector<EVT, 4> AggValueVTs;
2670 ComputeValueVTs(TLI, AggTy, AggValueVTs);
2671 SmallVector<EVT, 4> ValValueVTs;
2672 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2674 unsigned NumAggValues = AggValueVTs.size();
2675 unsigned NumValValues = ValValueVTs.size();
2676 SmallVector<SDValue, 4> Values(NumAggValues);
2678 SDValue Agg = getValue(Op0);
2679 SDValue Val = getValue(Op1);
2681 // Copy the beginning value(s) from the original aggregate.
2682 for (; i != LinearIndex; ++i)
2683 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2684 SDValue(Agg.getNode(), Agg.getResNo() + i);
2685 // Copy values from the inserted value(s).
2686 for (; i != LinearIndex + NumValValues; ++i)
2687 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2688 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex);
2689 // Copy remaining value(s) from the original aggregate.
2690 for (; i != NumAggValues; ++i)
2691 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2692 SDValue(Agg.getNode(), Agg.getResNo() + i);
2694 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2695 DAG.getVTList(&AggValueVTs[0], NumAggValues),
2696 &Values[0], NumAggValues);
2698 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2701 void SelectionDAGBuilder::visitExtractValue(ExtractValueInst &I) {
2702 const Value *Op0 = I.getOperand(0);
2703 const Type *AggTy = Op0->getType();
2704 const Type *ValTy = I.getType();
2705 bool OutOfUndef = isa<UndefValue>(Op0);
2707 unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy,
2708 I.idx_begin(), I.idx_end());
2710 SmallVector<EVT, 4> ValValueVTs;
2711 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2713 unsigned NumValValues = ValValueVTs.size();
2714 SmallVector<SDValue, 4> Values(NumValValues);
2716 SDValue Agg = getValue(Op0);
2717 // Copy out the selected value(s).
2718 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i)
2719 Values[i - LinearIndex] =
2721 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) :
2722 SDValue(Agg.getNode(), Agg.getResNo() + i);
2724 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2725 DAG.getVTList(&ValValueVTs[0], NumValValues),
2726 &Values[0], NumValValues);
2728 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2731 void SelectionDAGBuilder::visitGetElementPtr(User &I) {
2732 SDValue N = getValue(I.getOperand(0));
2733 const Type *Ty = I.getOperand(0)->getType();
2735 for (GetElementPtrInst::op_iterator OI = I.op_begin()+1, E = I.op_end();
2738 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
2739 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
2742 uint64_t Offset = TD->getStructLayout(StTy)->getElementOffset(Field);
2743 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
2744 DAG.getIntPtrConstant(Offset));
2745 DAG.AssignOrdering(N.getNode(), SDNodeOrder);
2748 Ty = StTy->getElementType(Field);
2750 Ty = cast<SequentialType>(Ty)->getElementType();
2752 // If this is a constant subscript, handle it quickly.
2753 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
2754 if (CI->getZExtValue() == 0) continue;
2756 TD->getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
2758 EVT PTy = TLI.getPointerTy();
2759 unsigned PtrBits = PTy.getSizeInBits();
2761 OffsVal = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2763 DAG.getConstant(Offs, MVT::i64));
2765 OffsVal = DAG.getIntPtrConstant(Offs);
2767 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
2770 DAG.AssignOrdering(OffsVal.getNode(), SDNodeOrder);
2771 DAG.AssignOrdering(N.getNode(), SDNodeOrder);
2775 // N = N + Idx * ElementSize;
2776 APInt ElementSize = APInt(TLI.getPointerTy().getSizeInBits(),
2777 TD->getTypeAllocSize(Ty));
2778 SDValue IdxN = getValue(Idx);
2780 // If the index is smaller or larger than intptr_t, truncate or extend
2782 IdxN = DAG.getSExtOrTrunc(IdxN, getCurDebugLoc(), N.getValueType());
2784 // If this is a multiply by a power of two, turn it into a shl
2785 // immediately. This is a very common case.
2786 if (ElementSize != 1) {
2787 if (ElementSize.isPowerOf2()) {
2788 unsigned Amt = ElementSize.logBase2();
2789 IdxN = DAG.getNode(ISD::SHL, getCurDebugLoc(),
2790 N.getValueType(), IdxN,
2791 DAG.getConstant(Amt, TLI.getPointerTy()));
2793 SDValue Scale = DAG.getConstant(ElementSize, TLI.getPointerTy());
2794 IdxN = DAG.getNode(ISD::MUL, getCurDebugLoc(),
2795 N.getValueType(), IdxN, Scale);
2798 DAG.AssignOrdering(IdxN.getNode(), SDNodeOrder);
2801 N = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2802 N.getValueType(), N, IdxN);
2803 DAG.AssignOrdering(N.getNode(), SDNodeOrder);
2810 void SelectionDAGBuilder::visitAlloca(AllocaInst &I) {
2811 // If this is a fixed sized alloca in the entry block of the function,
2812 // allocate it statically on the stack.
2813 if (FuncInfo.StaticAllocaMap.count(&I))
2814 return; // getValue will auto-populate this.
2816 const Type *Ty = I.getAllocatedType();
2817 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
2819 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty),
2822 SDValue AllocSize = getValue(I.getArraySize());
2824 AllocSize = DAG.getNode(ISD::MUL, getCurDebugLoc(), AllocSize.getValueType(),
2826 DAG.getConstant(TySize, AllocSize.getValueType()));
2828 DAG.AssignOrdering(AllocSize.getNode(), SDNodeOrder);
2830 EVT IntPtr = TLI.getPointerTy();
2831 AllocSize = DAG.getZExtOrTrunc(AllocSize, getCurDebugLoc(), IntPtr);
2832 DAG.AssignOrdering(AllocSize.getNode(), SDNodeOrder);
2834 // Handle alignment. If the requested alignment is less than or equal to
2835 // the stack alignment, ignore it. If the size is greater than or equal to
2836 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
2837 unsigned StackAlign =
2838 TLI.getTargetMachine().getFrameInfo()->getStackAlignment();
2839 if (Align <= StackAlign)
2842 // Round the size of the allocation up to the stack alignment size
2843 // by add SA-1 to the size.
2844 AllocSize = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2845 AllocSize.getValueType(), AllocSize,
2846 DAG.getIntPtrConstant(StackAlign-1));
2847 DAG.AssignOrdering(AllocSize.getNode(), SDNodeOrder);
2849 // Mask out the low bits for alignment purposes.
2850 AllocSize = DAG.getNode(ISD::AND, getCurDebugLoc(),
2851 AllocSize.getValueType(), AllocSize,
2852 DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1)));
2853 DAG.AssignOrdering(AllocSize.getNode(), SDNodeOrder);
2855 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) };
2856 SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other);
2857 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, getCurDebugLoc(),
2860 DAG.setRoot(DSA.getValue(1));
2861 DAG.AssignOrdering(DSA.getNode(), SDNodeOrder);
2863 // Inform the Frame Information that we have just allocated a variable-sized
2865 FuncInfo.MF->getFrameInfo()->CreateVariableSizedObject();
2868 void SelectionDAGBuilder::visitLoad(LoadInst &I) {
2869 const Value *SV = I.getOperand(0);
2870 SDValue Ptr = getValue(SV);
2872 const Type *Ty = I.getType();
2873 bool isVolatile = I.isVolatile();
2874 unsigned Alignment = I.getAlignment();
2876 SmallVector<EVT, 4> ValueVTs;
2877 SmallVector<uint64_t, 4> Offsets;
2878 ComputeValueVTs(TLI, Ty, ValueVTs, &Offsets);
2879 unsigned NumValues = ValueVTs.size();
2884 bool ConstantMemory = false;
2886 // Serialize volatile loads with other side effects.
2888 else if (AA->pointsToConstantMemory(SV)) {
2889 // Do not serialize (non-volatile) loads of constant memory with anything.
2890 Root = DAG.getEntryNode();
2891 ConstantMemory = true;
2893 // Do not serialize non-volatile loads against each other.
2894 Root = DAG.getRoot();
2897 SmallVector<SDValue, 4> Values(NumValues);
2898 SmallVector<SDValue, 4> Chains(NumValues);
2899 EVT PtrVT = Ptr.getValueType();
2900 for (unsigned i = 0; i != NumValues; ++i) {
2901 SDValue A = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2903 DAG.getConstant(Offsets[i], PtrVT));
2904 SDValue L = DAG.getLoad(ValueVTs[i], getCurDebugLoc(), Root,
2905 A, SV, Offsets[i], isVolatile, Alignment);
2908 Chains[i] = L.getValue(1);
2910 DAG.AssignOrdering(A.getNode(), SDNodeOrder);
2911 DAG.AssignOrdering(L.getNode(), SDNodeOrder);
2914 if (!ConstantMemory) {
2915 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
2916 MVT::Other, &Chains[0], NumValues);
2920 PendingLoads.push_back(Chain);
2922 DAG.AssignOrdering(Chain.getNode(), SDNodeOrder);
2925 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2926 DAG.getVTList(&ValueVTs[0], NumValues),
2927 &Values[0], NumValues);
2929 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2932 void SelectionDAGBuilder::visitStore(StoreInst &I) {
2933 Value *SrcV = I.getOperand(0);
2934 Value *PtrV = I.getOperand(1);
2936 SmallVector<EVT, 4> ValueVTs;
2937 SmallVector<uint64_t, 4> Offsets;
2938 ComputeValueVTs(TLI, SrcV->getType(), ValueVTs, &Offsets);
2939 unsigned NumValues = ValueVTs.size();
2943 // Get the lowered operands. Note that we do this after
2944 // checking if NumResults is zero, because with zero results
2945 // the operands won't have values in the map.
2946 SDValue Src = getValue(SrcV);
2947 SDValue Ptr = getValue(PtrV);
2949 SDValue Root = getRoot();
2950 SmallVector<SDValue, 4> Chains(NumValues);
2951 EVT PtrVT = Ptr.getValueType();
2952 bool isVolatile = I.isVolatile();
2953 unsigned Alignment = I.getAlignment();
2955 for (unsigned i = 0; i != NumValues; ++i) {
2956 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT, Ptr,
2957 DAG.getConstant(Offsets[i], PtrVT));
2958 Chains[i] = DAG.getStore(Root, getCurDebugLoc(),
2959 SDValue(Src.getNode(), Src.getResNo() + i),
2960 Add, PtrV, Offsets[i], isVolatile, Alignment);
2962 DAG.AssignOrdering(Add.getNode(), SDNodeOrder);
2963 DAG.AssignOrdering(Chains[i].getNode(), SDNodeOrder);
2966 SDValue Res = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
2967 MVT::Other, &Chains[0], NumValues);
2969 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
2972 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
2974 void SelectionDAGBuilder::visitTargetIntrinsic(CallInst &I,
2975 unsigned Intrinsic) {
2976 bool HasChain = !I.doesNotAccessMemory();
2977 bool OnlyLoad = HasChain && I.onlyReadsMemory();
2979 // Build the operand list.
2980 SmallVector<SDValue, 8> Ops;
2981 if (HasChain) { // If this intrinsic has side-effects, chainify it.
2983 // We don't need to serialize loads against other loads.
2984 Ops.push_back(DAG.getRoot());
2986 Ops.push_back(getRoot());
2990 // Info is set by getTgtMemInstrinsic
2991 TargetLowering::IntrinsicInfo Info;
2992 bool IsTgtIntrinsic = TLI.getTgtMemIntrinsic(Info, I, Intrinsic);
2994 // Add the intrinsic ID as an integer operand if it's not a target intrinsic.
2995 if (!IsTgtIntrinsic)
2996 Ops.push_back(DAG.getConstant(Intrinsic, TLI.getPointerTy()));
2998 // Add all operands of the call to the operand list.
2999 for (unsigned i = 1, e = I.getNumOperands(); i != e; ++i) {
3000 SDValue Op = getValue(I.getOperand(i));
3001 assert(TLI.isTypeLegal(Op.getValueType()) &&
3002 "Intrinsic uses a non-legal type?");
3006 SmallVector<EVT, 4> ValueVTs;
3007 ComputeValueVTs(TLI, I.getType(), ValueVTs);
3009 for (unsigned Val = 0, E = ValueVTs.size(); Val != E; ++Val) {
3010 assert(TLI.isTypeLegal(ValueVTs[Val]) &&
3011 "Intrinsic uses a non-legal type?");
3016 ValueVTs.push_back(MVT::Other);
3018 SDVTList VTs = DAG.getVTList(ValueVTs.data(), ValueVTs.size());
3022 if (IsTgtIntrinsic) {
3023 // This is target intrinsic that touches memory
3024 Result = DAG.getMemIntrinsicNode(Info.opc, getCurDebugLoc(),
3025 VTs, &Ops[0], Ops.size(),
3026 Info.memVT, Info.ptrVal, Info.offset,
3027 Info.align, Info.vol,
3028 Info.readMem, Info.writeMem);
3029 } else if (!HasChain) {
3030 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurDebugLoc(),
3031 VTs, &Ops[0], Ops.size());
3032 } else if (!I.getType()->isVoidTy()) {
3033 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurDebugLoc(),
3034 VTs, &Ops[0], Ops.size());
3036 Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurDebugLoc(),
3037 VTs, &Ops[0], Ops.size());
3040 DAG.AssignOrdering(Result.getNode(), SDNodeOrder);
3043 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1);
3045 PendingLoads.push_back(Chain);
3050 if (!I.getType()->isVoidTy()) {
3051 if (const VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
3052 EVT VT = TLI.getValueType(PTy);
3053 Result = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(), VT, Result);
3054 DAG.AssignOrdering(Result.getNode(), SDNodeOrder);
3057 setValue(&I, Result);
3061 /// GetSignificand - Get the significand and build it into a floating-point
3062 /// number with exponent of 1:
3064 /// Op = (Op & 0x007fffff) | 0x3f800000;
3066 /// where Op is the hexidecimal representation of floating point value.
3068 GetSignificand(SelectionDAG &DAG, SDValue Op, DebugLoc dl, unsigned Order) {
3069 SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3070 DAG.getConstant(0x007fffff, MVT::i32));
3071 SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1,
3072 DAG.getConstant(0x3f800000, MVT::i32));
3073 SDValue Res = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t2);
3075 DAG.AssignOrdering(t1.getNode(), Order);
3076 DAG.AssignOrdering(t2.getNode(), Order);
3077 DAG.AssignOrdering(Res.getNode(), Order);
3081 /// GetExponent - Get the exponent:
3083 /// (float)(int)(((Op & 0x7f800000) >> 23) - 127);
3085 /// where Op is the hexidecimal representation of floating point value.
3087 GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI,
3088 DebugLoc dl, unsigned Order) {
3089 SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3090 DAG.getConstant(0x7f800000, MVT::i32));
3091 SDValue t1 = DAG.getNode(ISD::SRL, dl, MVT::i32, t0,
3092 DAG.getConstant(23, TLI.getPointerTy()));
3093 SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1,
3094 DAG.getConstant(127, MVT::i32));
3095 SDValue Res = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2);
3097 DAG.AssignOrdering(t0.getNode(), Order);
3098 DAG.AssignOrdering(t1.getNode(), Order);
3099 DAG.AssignOrdering(t2.getNode(), Order);
3100 DAG.AssignOrdering(Res.getNode(), Order);
3104 /// getF32Constant - Get 32-bit floating point constant.
3106 getF32Constant(SelectionDAG &DAG, unsigned Flt) {
3107 return DAG.getConstantFP(APFloat(APInt(32, Flt)), MVT::f32);
3110 /// Inlined utility function to implement binary input atomic intrinsics for
3111 /// visitIntrinsicCall: I is a call instruction
3112 /// Op is the associated NodeType for I
3114 SelectionDAGBuilder::implVisitBinaryAtomic(CallInst& I, ISD::NodeType Op) {
3115 SDValue Root = getRoot();
3117 DAG.getAtomic(Op, getCurDebugLoc(),
3118 getValue(I.getOperand(2)).getValueType().getSimpleVT(),
3120 getValue(I.getOperand(1)),
3121 getValue(I.getOperand(2)),
3124 DAG.setRoot(L.getValue(1));
3125 DAG.AssignOrdering(L.getNode(), SDNodeOrder);
3129 // implVisitAluOverflow - Lower arithmetic overflow instrinsics.
3131 SelectionDAGBuilder::implVisitAluOverflow(CallInst &I, ISD::NodeType Op) {
3132 SDValue Op1 = getValue(I.getOperand(1));
3133 SDValue Op2 = getValue(I.getOperand(2));
3135 SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1);
3136 SDValue Result = DAG.getNode(Op, getCurDebugLoc(), VTs, Op1, Op2);
3138 setValue(&I, Result);
3139 DAG.AssignOrdering(Result.getNode(), SDNodeOrder);
3143 /// visitExp - Lower an exp intrinsic. Handles the special sequences for
3144 /// limited-precision mode.
3146 SelectionDAGBuilder::visitExp(CallInst &I) {
3148 DebugLoc dl = getCurDebugLoc();
3150 if (getValue(I.getOperand(1)).getValueType() == MVT::f32 &&
3151 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3152 SDValue Op = getValue(I.getOperand(1));
3154 // Put the exponent in the right bit position for later addition to the
3157 // #define LOG2OFe 1.4426950f
3158 // IntegerPartOfX = ((int32_t)(X * LOG2OFe));
3159 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3160 getF32Constant(DAG, 0x3fb8aa3b));
3161 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3163 // FractionalPartOfX = (X * LOG2OFe) - (float)IntegerPartOfX;
3164 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3165 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3167 DAG.AssignOrdering(t0.getNode(), SDNodeOrder);
3168 DAG.AssignOrdering(IntegerPartOfX.getNode(), SDNodeOrder);
3169 DAG.AssignOrdering(t1.getNode(), SDNodeOrder);
3170 DAG.AssignOrdering(X.getNode(), SDNodeOrder);
3172 // IntegerPartOfX <<= 23;
3173 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3174 DAG.getConstant(23, TLI.getPointerTy()));
3175 DAG.AssignOrdering(IntegerPartOfX.getNode(), SDNodeOrder);
3177 if (LimitFloatPrecision <= 6) {
3178 // For floating-point precision of 6:
3180 // TwoToFractionalPartOfX =
3182 // (0.735607626f + 0.252464424f * x) * x;
3184 // error 0.0144103317, which is 6 bits
3185 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3186 getF32Constant(DAG, 0x3e814304));
3187 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3188 getF32Constant(DAG, 0x3f3c50c8));
3189 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3190 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3191 getF32Constant(DAG, 0x3f7f5e7e));
3192 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,MVT::i32, t5);
3194 // Add the exponent into the result in integer domain.
3195 SDValue t6 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3196 TwoToFracPartOfX, IntegerPartOfX);
3198 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t6);
3200 DAG.AssignOrdering(t2.getNode(), SDNodeOrder);
3201 DAG.AssignOrdering(t3.getNode(), SDNodeOrder);
3202 DAG.AssignOrdering(t4.getNode(), SDNodeOrder);
3203 DAG.AssignOrdering(t5.getNode(), SDNodeOrder);
3204 DAG.AssignOrdering(t6.getNode(), SDNodeOrder);
3205 DAG.AssignOrdering(TwoToFracPartOfX.getNode(), SDNodeOrder);
3206 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
3207 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3208 // For floating-point precision of 12:
3210 // TwoToFractionalPartOfX =
3213 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3215 // 0.000107046256 error, which is 13 to 14 bits
3216 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3217 getF32Constant(DAG, 0x3da235e3));
3218 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3219 getF32Constant(DAG, 0x3e65b8f3));
3220 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3221 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3222 getF32Constant(DAG, 0x3f324b07));
3223 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3224 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3225 getF32Constant(DAG, 0x3f7ff8fd));
3226 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,MVT::i32, t7);
3228 // Add the exponent into the result in integer domain.
3229 SDValue t8 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3230 TwoToFracPartOfX, IntegerPartOfX);
3232 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t8);
3234 DAG.AssignOrdering(t2.getNode(), SDNodeOrder);
3235 DAG.AssignOrdering(t3.getNode(), SDNodeOrder);
3236 DAG.AssignOrdering(t4.getNode(), SDNodeOrder);
3237 DAG.AssignOrdering(t5.getNode(), SDNodeOrder);
3238 DAG.AssignOrdering(t6.getNode(), SDNodeOrder);
3239 DAG.AssignOrdering(t7.getNode(), SDNodeOrder);
3240 DAG.AssignOrdering(t8.getNode(), SDNodeOrder);
3241 DAG.AssignOrdering(TwoToFracPartOfX.getNode(), SDNodeOrder);
3242 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
3243 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3244 // For floating-point precision of 18:
3246 // TwoToFractionalPartOfX =
3250 // (0.554906021e-1f +
3251 // (0.961591928e-2f +
3252 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3254 // error 2.47208000*10^(-7), which is better than 18 bits
3255 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3256 getF32Constant(DAG, 0x3924b03e));
3257 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3258 getF32Constant(DAG, 0x3ab24b87));
3259 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3260 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3261 getF32Constant(DAG, 0x3c1d8c17));
3262 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3263 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3264 getF32Constant(DAG, 0x3d634a1d));
3265 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3266 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3267 getF32Constant(DAG, 0x3e75fe14));
3268 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3269 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3270 getF32Constant(DAG, 0x3f317234));
3271 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3272 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3273 getF32Constant(DAG, 0x3f800000));
3274 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,
3277 // Add the exponent into the result in integer domain.
3278 SDValue t14 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3279 TwoToFracPartOfX, IntegerPartOfX);
3281 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t14);
3283 DAG.AssignOrdering(t2.getNode(), SDNodeOrder);
3284 DAG.AssignOrdering(t3.getNode(), SDNodeOrder);
3285 DAG.AssignOrdering(t4.getNode(), SDNodeOrder);
3286 DAG.AssignOrdering(t5.getNode(), SDNodeOrder);
3287 DAG.AssignOrdering(t6.getNode(), SDNodeOrder);
3288 DAG.AssignOrdering(t7.getNode(), SDNodeOrder);
3289 DAG.AssignOrdering(t8.getNode(), SDNodeOrder);
3290 DAG.AssignOrdering(t9.getNode(), SDNodeOrder);
3291 DAG.AssignOrdering(t10.getNode(), SDNodeOrder);
3292 DAG.AssignOrdering(t11.getNode(), SDNodeOrder);
3293 DAG.AssignOrdering(t12.getNode(), SDNodeOrder);
3294 DAG.AssignOrdering(t13.getNode(), SDNodeOrder);
3295 DAG.AssignOrdering(t14.getNode(), SDNodeOrder);
3296 DAG.AssignOrdering(TwoToFracPartOfX.getNode(), SDNodeOrder);
3297 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
3300 // No special expansion.
3301 result = DAG.getNode(ISD::FEXP, dl,
3302 getValue(I.getOperand(1)).getValueType(),
3303 getValue(I.getOperand(1)));
3304 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
3307 setValue(&I, result);
3310 /// visitLog - Lower a log intrinsic. Handles the special sequences for
3311 /// limited-precision mode.
3313 SelectionDAGBuilder::visitLog(CallInst &I) {
3315 DebugLoc dl = getCurDebugLoc();
3317 if (getValue(I.getOperand(1)).getValueType() == MVT::f32 &&
3318 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3319 SDValue Op = getValue(I.getOperand(1));
3320 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
3322 DAG.AssignOrdering(Op1.getNode(), SDNodeOrder);
3324 // Scale the exponent by log(2) [0.69314718f].
3325 SDValue Exp = GetExponent(DAG, Op1, TLI, dl, SDNodeOrder);
3326 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3327 getF32Constant(DAG, 0x3f317218));
3329 DAG.AssignOrdering(LogOfExponent.getNode(), SDNodeOrder);
3331 // Get the significand and build it into a floating-point number with
3333 SDValue X = GetSignificand(DAG, Op1, dl, SDNodeOrder);
3335 if (LimitFloatPrecision <= 6) {
3336 // For floating-point precision of 6:
3340 // (1.4034025f - 0.23903021f * x) * x;
3342 // error 0.0034276066, which is better than 8 bits
3343 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3344 getF32Constant(DAG, 0xbe74c456));
3345 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3346 getF32Constant(DAG, 0x3fb3a2b1));
3347 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3348 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3349 getF32Constant(DAG, 0x3f949a29));
3351 result = DAG.getNode(ISD::FADD, dl,
3352 MVT::f32, LogOfExponent, LogOfMantissa);
3354 DAG.AssignOrdering(t0.getNode(), SDNodeOrder);
3355 DAG.AssignOrdering(t1.getNode(), SDNodeOrder);
3356 DAG.AssignOrdering(t2.getNode(), SDNodeOrder);
3357 DAG.AssignOrdering(LogOfMantissa.getNode(), SDNodeOrder);
3358 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
3359 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3360 // For floating-point precision of 12:
3366 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x;
3368 // error 0.000061011436, which is 14 bits
3369 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3370 getF32Constant(DAG, 0xbd67b6d6));
3371 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3372 getF32Constant(DAG, 0x3ee4f4b8));
3373 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3374 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3375 getF32Constant(DAG, 0x3fbc278b));
3376 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3377 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3378 getF32Constant(DAG, 0x40348e95));
3379 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3380 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3381 getF32Constant(DAG, 0x3fdef31a));
3383 result = DAG.getNode(ISD::FADD, dl,
3384 MVT::f32, LogOfExponent, LogOfMantissa);
3386 DAG.AssignOrdering(t0.getNode(), SDNodeOrder);
3387 DAG.AssignOrdering(t1.getNode(), SDNodeOrder);
3388 DAG.AssignOrdering(t2.getNode(), SDNodeOrder);
3389 DAG.AssignOrdering(t3.getNode(), SDNodeOrder);
3390 DAG.AssignOrdering(t4.getNode(), SDNodeOrder);
3391 DAG.AssignOrdering(t5.getNode(), SDNodeOrder);
3392 DAG.AssignOrdering(t6.getNode(), SDNodeOrder);
3393 DAG.AssignOrdering(LogOfMantissa.getNode(), SDNodeOrder);
3394 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
3395 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3396 // For floating-point precision of 18:
3404 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x;
3406 // error 0.0000023660568, which is better than 18 bits
3407 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3408 getF32Constant(DAG, 0xbc91e5ac));
3409 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3410 getF32Constant(DAG, 0x3e4350aa));
3411 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3412 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3413 getF32Constant(DAG, 0x3f60d3e3));
3414 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3415 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3416 getF32Constant(DAG, 0x4011cdf0));
3417 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3418 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3419 getF32Constant(DAG, 0x406cfd1c));
3420 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3421 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3422 getF32Constant(DAG, 0x408797cb));
3423 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3424 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3425 getF32Constant(DAG, 0x4006dcab));
3427 result = DAG.getNode(ISD::FADD, dl,
3428 MVT::f32, LogOfExponent, LogOfMantissa);
3430 DAG.AssignOrdering(t0.getNode(), SDNodeOrder);
3431 DAG.AssignOrdering(t1.getNode(), SDNodeOrder);
3432 DAG.AssignOrdering(t2.getNode(), SDNodeOrder);
3433 DAG.AssignOrdering(t3.getNode(), SDNodeOrder);
3434 DAG.AssignOrdering(t4.getNode(), SDNodeOrder);
3435 DAG.AssignOrdering(t5.getNode(), SDNodeOrder);
3436 DAG.AssignOrdering(t6.getNode(), SDNodeOrder);
3437 DAG.AssignOrdering(t7.getNode(), SDNodeOrder);
3438 DAG.AssignOrdering(t8.getNode(), SDNodeOrder);
3439 DAG.AssignOrdering(t9.getNode(), SDNodeOrder);
3440 DAG.AssignOrdering(t10.getNode(), SDNodeOrder);
3441 DAG.AssignOrdering(LogOfMantissa.getNode(), SDNodeOrder);
3442 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
3445 // No special expansion.
3446 result = DAG.getNode(ISD::FLOG, dl,
3447 getValue(I.getOperand(1)).getValueType(),
3448 getValue(I.getOperand(1)));
3449 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
3452 setValue(&I, result);
3455 /// visitLog2 - Lower a log2 intrinsic. Handles the special sequences for
3456 /// limited-precision mode.
3458 SelectionDAGBuilder::visitLog2(CallInst &I) {
3460 DebugLoc dl = getCurDebugLoc();
3462 if (getValue(I.getOperand(1)).getValueType() == MVT::f32 &&
3463 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3464 SDValue Op = getValue(I.getOperand(1));
3465 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
3467 DAG.AssignOrdering(Op1.getNode(), SDNodeOrder);
3469 // Get the exponent.
3470 SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl, SDNodeOrder);
3472 DAG.AssignOrdering(LogOfExponent.getNode(), SDNodeOrder);
3474 // Get the significand and build it into a floating-point number with
3476 SDValue X = GetSignificand(DAG, Op1, dl, SDNodeOrder);
3478 // Different possible minimax approximations of significand in
3479 // floating-point for various degrees of accuracy over [1,2].
3480 if (LimitFloatPrecision <= 6) {
3481 // For floating-point precision of 6:
3483 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x;
3485 // error 0.0049451742, which is more than 7 bits
3486 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3487 getF32Constant(DAG, 0xbeb08fe0));
3488 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3489 getF32Constant(DAG, 0x40019463));
3490 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3491 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3492 getF32Constant(DAG, 0x3fd6633d));
3494 result = DAG.getNode(ISD::FADD, dl,
3495 MVT::f32, LogOfExponent, Log2ofMantissa);
3497 DAG.AssignOrdering(t0.getNode(), SDNodeOrder);
3498 DAG.AssignOrdering(t1.getNode(), SDNodeOrder);
3499 DAG.AssignOrdering(t2.getNode(), SDNodeOrder);
3500 DAG.AssignOrdering(Log2ofMantissa.getNode(), SDNodeOrder);
3501 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
3502 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3503 // For floating-point precision of 12:
3509 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x;
3511 // error 0.0000876136000, which is better than 13 bits
3512 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3513 getF32Constant(DAG, 0xbda7262e));
3514 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3515 getF32Constant(DAG, 0x3f25280b));
3516 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3517 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3518 getF32Constant(DAG, 0x4007b923));
3519 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3520 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3521 getF32Constant(DAG, 0x40823e2f));
3522 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3523 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3524 getF32Constant(DAG, 0x4020d29c));
3526 result = DAG.getNode(ISD::FADD, dl,
3527 MVT::f32, LogOfExponent, Log2ofMantissa);
3529 DAG.AssignOrdering(t0.getNode(), SDNodeOrder);
3530 DAG.AssignOrdering(t1.getNode(), SDNodeOrder);
3531 DAG.AssignOrdering(t2.getNode(), SDNodeOrder);
3532 DAG.AssignOrdering(t3.getNode(), SDNodeOrder);
3533 DAG.AssignOrdering(t4.getNode(), SDNodeOrder);
3534 DAG.AssignOrdering(t5.getNode(), SDNodeOrder);
3535 DAG.AssignOrdering(t6.getNode(), SDNodeOrder);
3536 DAG.AssignOrdering(Log2ofMantissa.getNode(), SDNodeOrder);
3537 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
3538 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3539 // For floating-point precision of 18:
3548 // 0.25691327e-1f * x) * x) * x) * x) * x) * x;
3550 // error 0.0000018516, which is better than 18 bits
3551 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3552 getF32Constant(DAG, 0xbcd2769e));
3553 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3554 getF32Constant(DAG, 0x3e8ce0b9));
3555 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3556 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3557 getF32Constant(DAG, 0x3fa22ae7));
3558 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3559 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3560 getF32Constant(DAG, 0x40525723));
3561 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3562 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3563 getF32Constant(DAG, 0x40aaf200));
3564 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3565 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3566 getF32Constant(DAG, 0x40c39dad));
3567 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3568 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3569 getF32Constant(DAG, 0x4042902c));
3571 result = DAG.getNode(ISD::FADD, dl,
3572 MVT::f32, LogOfExponent, Log2ofMantissa);
3574 DAG.AssignOrdering(t0.getNode(), SDNodeOrder);
3575 DAG.AssignOrdering(t1.getNode(), SDNodeOrder);
3576 DAG.AssignOrdering(t2.getNode(), SDNodeOrder);
3577 DAG.AssignOrdering(t3.getNode(), SDNodeOrder);
3578 DAG.AssignOrdering(t4.getNode(), SDNodeOrder);
3579 DAG.AssignOrdering(t5.getNode(), SDNodeOrder);
3580 DAG.AssignOrdering(t6.getNode(), SDNodeOrder);
3581 DAG.AssignOrdering(t7.getNode(), SDNodeOrder);
3582 DAG.AssignOrdering(t8.getNode(), SDNodeOrder);
3583 DAG.AssignOrdering(t9.getNode(), SDNodeOrder);
3584 DAG.AssignOrdering(t10.getNode(), SDNodeOrder);
3585 DAG.AssignOrdering(Log2ofMantissa.getNode(), SDNodeOrder);
3586 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
3589 // No special expansion.
3590 result = DAG.getNode(ISD::FLOG2, dl,
3591 getValue(I.getOperand(1)).getValueType(),
3592 getValue(I.getOperand(1)));
3593 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
3596 setValue(&I, result);
3599 /// visitLog10 - Lower a log10 intrinsic. Handles the special sequences for
3600 /// limited-precision mode.
3602 SelectionDAGBuilder::visitLog10(CallInst &I) {
3604 DebugLoc dl = getCurDebugLoc();
3606 if (getValue(I.getOperand(1)).getValueType() == MVT::f32 &&
3607 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3608 SDValue Op = getValue(I.getOperand(1));
3609 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
3611 DAG.AssignOrdering(Op1.getNode(), SDNodeOrder);
3613 // Scale the exponent by log10(2) [0.30102999f].
3614 SDValue Exp = GetExponent(DAG, Op1, TLI, dl, SDNodeOrder);
3615 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3616 getF32Constant(DAG, 0x3e9a209a));
3618 DAG.AssignOrdering(LogOfExponent.getNode(), SDNodeOrder);
3620 // Get the significand and build it into a floating-point number with
3622 SDValue X = GetSignificand(DAG, Op1, dl, SDNodeOrder);
3624 if (LimitFloatPrecision <= 6) {
3625 // For floating-point precision of 6:
3627 // Log10ofMantissa =
3629 // (0.60948995f - 0.10380950f * x) * x;
3631 // error 0.0014886165, which is 6 bits
3632 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3633 getF32Constant(DAG, 0xbdd49a13));
3634 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3635 getF32Constant(DAG, 0x3f1c0789));
3636 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3637 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3638 getF32Constant(DAG, 0x3f011300));
3640 result = DAG.getNode(ISD::FADD, dl,
3641 MVT::f32, LogOfExponent, Log10ofMantissa);
3643 DAG.AssignOrdering(t0.getNode(), SDNodeOrder);
3644 DAG.AssignOrdering(t1.getNode(), SDNodeOrder);
3645 DAG.AssignOrdering(t2.getNode(), SDNodeOrder);
3646 DAG.AssignOrdering(Log10ofMantissa.getNode(), SDNodeOrder);
3647 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
3648 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3649 // For floating-point precision of 12:
3651 // Log10ofMantissa =
3654 // (-0.31664806f + 0.47637168e-1f * x) * x) * x;
3656 // error 0.00019228036, which is better than 12 bits
3657 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3658 getF32Constant(DAG, 0x3d431f31));
3659 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
3660 getF32Constant(DAG, 0x3ea21fb2));
3661 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3662 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3663 getF32Constant(DAG, 0x3f6ae232));
3664 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3665 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
3666 getF32Constant(DAG, 0x3f25f7c3));
3668 result = DAG.getNode(ISD::FADD, dl,
3669 MVT::f32, LogOfExponent, Log10ofMantissa);
3671 DAG.AssignOrdering(t0.getNode(), SDNodeOrder);
3672 DAG.AssignOrdering(t1.getNode(), SDNodeOrder);
3673 DAG.AssignOrdering(t2.getNode(), SDNodeOrder);
3674 DAG.AssignOrdering(t3.getNode(), SDNodeOrder);
3675 DAG.AssignOrdering(t4.getNode(), SDNodeOrder);
3676 DAG.AssignOrdering(Log10ofMantissa.getNode(), SDNodeOrder);
3677 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
3678 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3679 // For floating-point precision of 18:
3681 // Log10ofMantissa =
3686 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x;
3688 // error 0.0000037995730, which is better than 18 bits
3689 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3690 getF32Constant(DAG, 0x3c5d51ce));
3691 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
3692 getF32Constant(DAG, 0x3e00685a));
3693 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3694 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3695 getF32Constant(DAG, 0x3efb6798));
3696 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3697 SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
3698 getF32Constant(DAG, 0x3f88d192));
3699 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3700 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3701 getF32Constant(DAG, 0x3fc4316c));
3702 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3703 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8,
3704 getF32Constant(DAG, 0x3f57ce70));
3706 result = DAG.getNode(ISD::FADD, dl,
3707 MVT::f32, LogOfExponent, Log10ofMantissa);
3709 DAG.AssignOrdering(t0.getNode(), SDNodeOrder);
3710 DAG.AssignOrdering(t1.getNode(), SDNodeOrder);
3711 DAG.AssignOrdering(t2.getNode(), SDNodeOrder);
3712 DAG.AssignOrdering(t3.getNode(), SDNodeOrder);
3713 DAG.AssignOrdering(t4.getNode(), SDNodeOrder);
3714 DAG.AssignOrdering(t5.getNode(), SDNodeOrder);
3715 DAG.AssignOrdering(t6.getNode(), SDNodeOrder);
3716 DAG.AssignOrdering(t7.getNode(), SDNodeOrder);
3717 DAG.AssignOrdering(t8.getNode(), SDNodeOrder);
3718 DAG.AssignOrdering(Log10ofMantissa.getNode(), SDNodeOrder);
3719 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
3722 // No special expansion.
3723 result = DAG.getNode(ISD::FLOG10, dl,
3724 getValue(I.getOperand(1)).getValueType(),
3725 getValue(I.getOperand(1)));
3726 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
3729 setValue(&I, result);
3732 /// visitExp2 - Lower an exp2 intrinsic. Handles the special sequences for
3733 /// limited-precision mode.
3735 SelectionDAGBuilder::visitExp2(CallInst &I) {
3737 DebugLoc dl = getCurDebugLoc();
3739 if (getValue(I.getOperand(1)).getValueType() == MVT::f32 &&
3740 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3741 SDValue Op = getValue(I.getOperand(1));
3743 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Op);
3745 DAG.AssignOrdering(IntegerPartOfX.getNode(), SDNodeOrder);
3747 // FractionalPartOfX = x - (float)IntegerPartOfX;
3748 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3749 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, Op, t1);
3751 // IntegerPartOfX <<= 23;
3752 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3753 DAG.getConstant(23, TLI.getPointerTy()));
3755 DAG.AssignOrdering(t1.getNode(), SDNodeOrder);
3756 DAG.AssignOrdering(X.getNode(), SDNodeOrder);
3757 DAG.AssignOrdering(IntegerPartOfX.getNode(), SDNodeOrder);
3759 if (LimitFloatPrecision <= 6) {
3760 // For floating-point precision of 6:
3762 // TwoToFractionalPartOfX =
3764 // (0.735607626f + 0.252464424f * x) * x;
3766 // error 0.0144103317, which is 6 bits
3767 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3768 getF32Constant(DAG, 0x3e814304));
3769 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3770 getF32Constant(DAG, 0x3f3c50c8));
3771 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3772 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3773 getF32Constant(DAG, 0x3f7f5e7e));
3774 SDValue t6 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t5);
3775 SDValue TwoToFractionalPartOfX =
3776 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
3778 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3779 MVT::f32, TwoToFractionalPartOfX);
3781 DAG.AssignOrdering(t2.getNode(), SDNodeOrder);
3782 DAG.AssignOrdering(t3.getNode(), SDNodeOrder);
3783 DAG.AssignOrdering(t4.getNode(), SDNodeOrder);
3784 DAG.AssignOrdering(t5.getNode(), SDNodeOrder);
3785 DAG.AssignOrdering(t6.getNode(), SDNodeOrder);
3786 DAG.AssignOrdering(TwoToFractionalPartOfX.getNode(), SDNodeOrder);
3787 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
3788 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3789 // For floating-point precision of 12:
3791 // TwoToFractionalPartOfX =
3794 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3796 // error 0.000107046256, which is 13 to 14 bits
3797 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3798 getF32Constant(DAG, 0x3da235e3));
3799 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3800 getF32Constant(DAG, 0x3e65b8f3));
3801 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3802 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3803 getF32Constant(DAG, 0x3f324b07));
3804 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3805 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3806 getF32Constant(DAG, 0x3f7ff8fd));
3807 SDValue t8 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t7);
3808 SDValue TwoToFractionalPartOfX =
3809 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
3811 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3812 MVT::f32, TwoToFractionalPartOfX);
3814 DAG.AssignOrdering(t2.getNode(), SDNodeOrder);
3815 DAG.AssignOrdering(t3.getNode(), SDNodeOrder);
3816 DAG.AssignOrdering(t4.getNode(), SDNodeOrder);
3817 DAG.AssignOrdering(t5.getNode(), SDNodeOrder);
3818 DAG.AssignOrdering(t6.getNode(), SDNodeOrder);
3819 DAG.AssignOrdering(t7.getNode(), SDNodeOrder);
3820 DAG.AssignOrdering(t8.getNode(), SDNodeOrder);
3821 DAG.AssignOrdering(TwoToFractionalPartOfX.getNode(), SDNodeOrder);
3822 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
3823 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3824 // For floating-point precision of 18:
3826 // TwoToFractionalPartOfX =
3830 // (0.554906021e-1f +
3831 // (0.961591928e-2f +
3832 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3833 // error 2.47208000*10^(-7), which is better than 18 bits
3834 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3835 getF32Constant(DAG, 0x3924b03e));
3836 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3837 getF32Constant(DAG, 0x3ab24b87));
3838 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3839 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3840 getF32Constant(DAG, 0x3c1d8c17));
3841 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3842 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3843 getF32Constant(DAG, 0x3d634a1d));
3844 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3845 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3846 getF32Constant(DAG, 0x3e75fe14));
3847 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3848 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3849 getF32Constant(DAG, 0x3f317234));
3850 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3851 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3852 getF32Constant(DAG, 0x3f800000));
3853 SDValue t14 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t13);
3854 SDValue TwoToFractionalPartOfX =
3855 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
3857 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3858 MVT::f32, TwoToFractionalPartOfX);
3860 DAG.AssignOrdering(t2.getNode(), SDNodeOrder);
3861 DAG.AssignOrdering(t3.getNode(), SDNodeOrder);
3862 DAG.AssignOrdering(t4.getNode(), SDNodeOrder);
3863 DAG.AssignOrdering(t5.getNode(), SDNodeOrder);
3864 DAG.AssignOrdering(t6.getNode(), SDNodeOrder);
3865 DAG.AssignOrdering(t7.getNode(), SDNodeOrder);
3866 DAG.AssignOrdering(t8.getNode(), SDNodeOrder);
3867 DAG.AssignOrdering(t9.getNode(), SDNodeOrder);
3868 DAG.AssignOrdering(t10.getNode(), SDNodeOrder);
3869 DAG.AssignOrdering(t11.getNode(), SDNodeOrder);
3870 DAG.AssignOrdering(t12.getNode(), SDNodeOrder);
3871 DAG.AssignOrdering(t13.getNode(), SDNodeOrder);
3872 DAG.AssignOrdering(t14.getNode(), SDNodeOrder);
3873 DAG.AssignOrdering(TwoToFractionalPartOfX.getNode(), SDNodeOrder);
3874 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
3877 // No special expansion.
3878 result = DAG.getNode(ISD::FEXP2, dl,
3879 getValue(I.getOperand(1)).getValueType(),
3880 getValue(I.getOperand(1)));
3881 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
3884 setValue(&I, result);
3887 /// visitPow - Lower a pow intrinsic. Handles the special sequences for
3888 /// limited-precision mode with x == 10.0f.
3890 SelectionDAGBuilder::visitPow(CallInst &I) {
3892 Value *Val = I.getOperand(1);
3893 DebugLoc dl = getCurDebugLoc();
3894 bool IsExp10 = false;
3896 if (getValue(Val).getValueType() == MVT::f32 &&
3897 getValue(I.getOperand(2)).getValueType() == MVT::f32 &&
3898 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3899 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(Val))) {
3900 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
3902 IsExp10 = CFP->getValueAPF().bitwiseIsEqual(Ten);
3907 if (IsExp10 && LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3908 SDValue Op = getValue(I.getOperand(2));
3910 // Put the exponent in the right bit position for later addition to the
3913 // #define LOG2OF10 3.3219281f
3914 // IntegerPartOfX = (int32_t)(x * LOG2OF10);
3915 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3916 getF32Constant(DAG, 0x40549a78));
3917 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3919 // FractionalPartOfX = x - (float)IntegerPartOfX;
3920 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3921 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3923 DAG.AssignOrdering(t0.getNode(), SDNodeOrder);
3924 DAG.AssignOrdering(t1.getNode(), SDNodeOrder);
3925 DAG.AssignOrdering(IntegerPartOfX.getNode(), SDNodeOrder);
3926 DAG.AssignOrdering(X.getNode(), SDNodeOrder);
3928 // IntegerPartOfX <<= 23;
3929 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3930 DAG.getConstant(23, TLI.getPointerTy()));
3932 DAG.AssignOrdering(IntegerPartOfX.getNode(), SDNodeOrder);
3934 if (LimitFloatPrecision <= 6) {
3935 // For floating-point precision of 6:
3937 // twoToFractionalPartOfX =
3939 // (0.735607626f + 0.252464424f * x) * x;
3941 // error 0.0144103317, which is 6 bits
3942 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3943 getF32Constant(DAG, 0x3e814304));
3944 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3945 getF32Constant(DAG, 0x3f3c50c8));
3946 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3947 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3948 getF32Constant(DAG, 0x3f7f5e7e));
3949 SDValue t6 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t5);
3950 SDValue TwoToFractionalPartOfX =
3951 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
3953 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3954 MVT::f32, TwoToFractionalPartOfX);
3956 DAG.AssignOrdering(t2.getNode(), SDNodeOrder);
3957 DAG.AssignOrdering(t3.getNode(), SDNodeOrder);
3958 DAG.AssignOrdering(t4.getNode(), SDNodeOrder);
3959 DAG.AssignOrdering(t5.getNode(), SDNodeOrder);
3960 DAG.AssignOrdering(t6.getNode(), SDNodeOrder);
3961 DAG.AssignOrdering(TwoToFractionalPartOfX.getNode(), SDNodeOrder);
3962 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
3963 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3964 // For floating-point precision of 12:
3966 // TwoToFractionalPartOfX =
3969 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3971 // error 0.000107046256, which is 13 to 14 bits
3972 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3973 getF32Constant(DAG, 0x3da235e3));
3974 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3975 getF32Constant(DAG, 0x3e65b8f3));
3976 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3977 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3978 getF32Constant(DAG, 0x3f324b07));
3979 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3980 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3981 getF32Constant(DAG, 0x3f7ff8fd));
3982 SDValue t8 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t7);
3983 SDValue TwoToFractionalPartOfX =
3984 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
3986 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3987 MVT::f32, TwoToFractionalPartOfX);
3989 DAG.AssignOrdering(t2.getNode(), SDNodeOrder);
3990 DAG.AssignOrdering(t3.getNode(), SDNodeOrder);
3991 DAG.AssignOrdering(t4.getNode(), SDNodeOrder);
3992 DAG.AssignOrdering(t5.getNode(), SDNodeOrder);
3993 DAG.AssignOrdering(t6.getNode(), SDNodeOrder);
3994 DAG.AssignOrdering(t7.getNode(), SDNodeOrder);
3995 DAG.AssignOrdering(t8.getNode(), SDNodeOrder);
3996 DAG.AssignOrdering(TwoToFractionalPartOfX.getNode(), SDNodeOrder);
3997 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
3998 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3999 // For floating-point precision of 18:
4001 // TwoToFractionalPartOfX =
4005 // (0.554906021e-1f +
4006 // (0.961591928e-2f +
4007 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
4008 // error 2.47208000*10^(-7), which is better than 18 bits
4009 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4010 getF32Constant(DAG, 0x3924b03e));
4011 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4012 getF32Constant(DAG, 0x3ab24b87));
4013 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4014 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4015 getF32Constant(DAG, 0x3c1d8c17));
4016 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4017 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4018 getF32Constant(DAG, 0x3d634a1d));
4019 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4020 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4021 getF32Constant(DAG, 0x3e75fe14));
4022 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4023 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
4024 getF32Constant(DAG, 0x3f317234));
4025 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
4026 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
4027 getF32Constant(DAG, 0x3f800000));
4028 SDValue t14 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t13);
4029 SDValue TwoToFractionalPartOfX =
4030 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
4032 result = DAG.getNode(ISD::BIT_CONVERT, dl,
4033 MVT::f32, TwoToFractionalPartOfX);
4035 DAG.AssignOrdering(t2.getNode(), SDNodeOrder);
4036 DAG.AssignOrdering(t3.getNode(), SDNodeOrder);
4037 DAG.AssignOrdering(t4.getNode(), SDNodeOrder);
4038 DAG.AssignOrdering(t5.getNode(), SDNodeOrder);
4039 DAG.AssignOrdering(t6.getNode(), SDNodeOrder);
4040 DAG.AssignOrdering(t7.getNode(), SDNodeOrder);
4041 DAG.AssignOrdering(t8.getNode(), SDNodeOrder);
4042 DAG.AssignOrdering(t9.getNode(), SDNodeOrder);
4043 DAG.AssignOrdering(t10.getNode(), SDNodeOrder);
4044 DAG.AssignOrdering(t11.getNode(), SDNodeOrder);
4045 DAG.AssignOrdering(t12.getNode(), SDNodeOrder);
4046 DAG.AssignOrdering(t13.getNode(), SDNodeOrder);
4047 DAG.AssignOrdering(t14.getNode(), SDNodeOrder);
4048 DAG.AssignOrdering(TwoToFractionalPartOfX.getNode(), SDNodeOrder);
4049 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
4052 // No special expansion.
4053 result = DAG.getNode(ISD::FPOW, dl,
4054 getValue(I.getOperand(1)).getValueType(),
4055 getValue(I.getOperand(1)),
4056 getValue(I.getOperand(2)));
4057 DAG.AssignOrdering(result.getNode(), SDNodeOrder);
4060 setValue(&I, result);
4064 /// ExpandPowI - Expand a llvm.powi intrinsic.
4065 static SDValue ExpandPowI(DebugLoc DL, SDValue LHS, SDValue RHS,
4066 SelectionDAG &DAG) {
4067 // If RHS is a constant, we can expand this out to a multiplication tree,
4068 // otherwise we end up lowering to a call to __powidf2 (for example). When
4069 // optimizing for size, we only want to do this if the expansion would produce
4070 // a small number of multiplies, otherwise we do the full expansion.
4071 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
4072 // Get the exponent as a positive value.
4073 unsigned Val = RHSC->getSExtValue();
4074 if ((int)Val < 0) Val = -Val;
4076 // powi(x, 0) -> 1.0
4078 return DAG.getConstantFP(1.0, LHS.getValueType());
4080 Function *F = DAG.getMachineFunction().getFunction();
4081 if (!F->hasFnAttr(Attribute::OptimizeForSize) ||
4082 // If optimizing for size, don't insert too many multiplies. This
4083 // inserts up to 5 multiplies.
4084 CountPopulation_32(Val)+Log2_32(Val) < 7) {
4085 // We use the simple binary decomposition method to generate the multiply
4086 // sequence. There are more optimal ways to do this (for example,
4087 // powi(x,15) generates one more multiply than it should), but this has
4088 // the benefit of being both really simple and much better than a libcall.
4089 SDValue Res; // Logically starts equal to 1.0
4090 SDValue CurSquare = LHS;
4094 Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare);
4096 Res = CurSquare; // 1.0*CurSquare.
4099 CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(),
4100 CurSquare, CurSquare);
4104 // If the original was negative, invert the result, producing 1/(x*x*x).
4105 if (RHSC->getSExtValue() < 0)
4106 Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(),
4107 DAG.getConstantFP(1.0, LHS.getValueType()), Res);
4112 // Otherwise, expand to a libcall.
4113 return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS);
4117 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
4118 /// we want to emit this as a call to a named external function, return the name
4119 /// otherwise lower it and return null.
4121 SelectionDAGBuilder::visitIntrinsicCall(CallInst &I, unsigned Intrinsic) {
4122 DebugLoc dl = getCurDebugLoc();
4125 switch (Intrinsic) {
4127 // By default, turn this into a target intrinsic node.
4128 visitTargetIntrinsic(I, Intrinsic);
4130 case Intrinsic::vastart: visitVAStart(I); return 0;
4131 case Intrinsic::vaend: visitVAEnd(I); return 0;
4132 case Intrinsic::vacopy: visitVACopy(I); return 0;
4133 case Intrinsic::returnaddress:
4134 Res = DAG.getNode(ISD::RETURNADDR, dl, TLI.getPointerTy(),
4135 getValue(I.getOperand(1)));
4137 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4139 case Intrinsic::frameaddress:
4140 Res = DAG.getNode(ISD::FRAMEADDR, dl, TLI.getPointerTy(),
4141 getValue(I.getOperand(1)));
4143 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4145 case Intrinsic::setjmp:
4146 return "_setjmp"+!TLI.usesUnderscoreSetJmp();
4147 case Intrinsic::longjmp:
4148 return "_longjmp"+!TLI.usesUnderscoreLongJmp();
4149 case Intrinsic::memcpy: {
4150 SDValue Op1 = getValue(I.getOperand(1));
4151 SDValue Op2 = getValue(I.getOperand(2));
4152 SDValue Op3 = getValue(I.getOperand(3));
4153 unsigned Align = cast<ConstantInt>(I.getOperand(4))->getZExtValue();
4154 Res = DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, false,
4155 I.getOperand(1), 0, I.getOperand(2), 0);
4157 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4160 case Intrinsic::memset: {
4161 SDValue Op1 = getValue(I.getOperand(1));
4162 SDValue Op2 = getValue(I.getOperand(2));
4163 SDValue Op3 = getValue(I.getOperand(3));
4164 unsigned Align = cast<ConstantInt>(I.getOperand(4))->getZExtValue();
4165 Res = DAG.getMemset(getRoot(), dl, Op1, Op2, Op3, Align,
4166 I.getOperand(1), 0);
4168 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4171 case Intrinsic::memmove: {
4172 SDValue Op1 = getValue(I.getOperand(1));
4173 SDValue Op2 = getValue(I.getOperand(2));
4174 SDValue Op3 = getValue(I.getOperand(3));
4175 unsigned Align = cast<ConstantInt>(I.getOperand(4))->getZExtValue();
4177 // If the source and destination are known to not be aliases, we can
4178 // lower memmove as memcpy.
4179 uint64_t Size = -1ULL;
4180 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op3))
4181 Size = C->getZExtValue();
4182 if (AA->alias(I.getOperand(1), Size, I.getOperand(2), Size) ==
4183 AliasAnalysis::NoAlias) {
4184 Res = DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, false,
4185 I.getOperand(1), 0, I.getOperand(2), 0);
4187 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4191 Res = DAG.getMemmove(getRoot(), dl, Op1, Op2, Op3, Align,
4192 I.getOperand(1), 0, I.getOperand(2), 0);
4194 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4197 case Intrinsic::dbg_declare: {
4198 if (OptLevel != CodeGenOpt::None)
4199 // FIXME: Variable debug info is not supported here.
4201 DwarfWriter *DW = DAG.getDwarfWriter();
4204 DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
4205 if (!DIDescriptor::ValidDebugInfo(DI.getVariable(), CodeGenOpt::None))
4208 MDNode *Variable = DI.getVariable();
4209 Value *Address = DI.getAddress();
4210 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
4211 Address = BCI->getOperand(0);
4212 AllocaInst *AI = dyn_cast<AllocaInst>(Address);
4213 // Don't handle byval struct arguments or VLAs, for example.
4216 DenseMap<const AllocaInst*, int>::iterator SI =
4217 FuncInfo.StaticAllocaMap.find(AI);
4218 if (SI == FuncInfo.StaticAllocaMap.end())
4220 int FI = SI->second;
4222 if (MachineModuleInfo *MMI = DAG.getMachineModuleInfo())
4223 if (MDNode *Dbg = DI.getMetadata("dbg"))
4224 MMI->setVariableDbgInfo(Variable, FI, Dbg);
4227 case Intrinsic::eh_exception: {
4228 // Insert the EXCEPTIONADDR instruction.
4229 assert(CurMBB->isLandingPad() &&"Call to eh.exception not in landing pad!");
4230 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
4232 Ops[0] = DAG.getRoot();
4233 SDValue Op = DAG.getNode(ISD::EXCEPTIONADDR, dl, VTs, Ops, 1);
4235 DAG.setRoot(Op.getValue(1));
4236 DAG.AssignOrdering(Op.getNode(), SDNodeOrder);
4240 case Intrinsic::eh_selector: {
4241 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
4243 if (CurMBB->isLandingPad())
4244 AddCatchInfo(I, MMI, CurMBB);
4247 FuncInfo.CatchInfoLost.insert(&I);
4249 // FIXME: Mark exception selector register as live in. Hack for PR1508.
4250 unsigned Reg = TLI.getExceptionSelectorRegister();
4251 if (Reg) CurMBB->addLiveIn(Reg);
4254 // Insert the EHSELECTION instruction.
4255 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
4257 Ops[0] = getValue(I.getOperand(1));
4259 SDValue Op = DAG.getNode(ISD::EHSELECTION, dl, VTs, Ops, 2);
4261 DAG.setRoot(Op.getValue(1));
4263 Res = DAG.getSExtOrTrunc(Op, dl, MVT::i32);
4265 DAG.AssignOrdering(Op.getNode(), SDNodeOrder);
4266 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4270 case Intrinsic::eh_typeid_for: {
4271 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
4274 // Find the type id for the given typeinfo.
4275 GlobalVariable *GV = ExtractTypeInfo(I.getOperand(1));
4276 unsigned TypeID = MMI->getTypeIDFor(GV);
4277 Res = DAG.getConstant(TypeID, MVT::i32);
4279 // Return something different to eh_selector.
4280 Res = DAG.getConstant(1, MVT::i32);
4284 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4288 case Intrinsic::eh_return_i32:
4289 case Intrinsic::eh_return_i64:
4290 if (MachineModuleInfo *MMI = DAG.getMachineModuleInfo()) {
4291 MMI->setCallsEHReturn(true);
4292 Res = DAG.getNode(ISD::EH_RETURN, dl,
4295 getValue(I.getOperand(1)),
4296 getValue(I.getOperand(2)));
4298 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4300 setValue(&I, DAG.getConstant(0, TLI.getPointerTy()));
4304 case Intrinsic::eh_unwind_init:
4305 if (MachineModuleInfo *MMI = DAG.getMachineModuleInfo()) {
4306 MMI->setCallsUnwindInit(true);
4309 case Intrinsic::eh_dwarf_cfa: {
4310 EVT VT = getValue(I.getOperand(1)).getValueType();
4311 SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getOperand(1)), dl,
4312 TLI.getPointerTy());
4313 SDValue Offset = DAG.getNode(ISD::ADD, dl,
4315 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, dl,
4316 TLI.getPointerTy()),
4318 SDValue FA = DAG.getNode(ISD::FRAMEADDR, dl,
4320 DAG.getConstant(0, TLI.getPointerTy()));
4321 Res = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
4324 DAG.AssignOrdering(CfaArg.getNode(), SDNodeOrder);
4325 DAG.AssignOrdering(Offset.getNode(), SDNodeOrder);
4326 DAG.AssignOrdering(FA.getNode(), SDNodeOrder);
4327 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4330 case Intrinsic::eh_sjlj_callsite: {
4331 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
4332 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
4333 assert(CI && "Non-constant call site value in eh.sjlj.callsite!");
4334 assert(MMI->getCurrentCallSite() == 0 && "Overlapping call sites!");
4336 MMI->setCurrentCallSite(CI->getZExtValue());
4340 case Intrinsic::convertff:
4341 case Intrinsic::convertfsi:
4342 case Intrinsic::convertfui:
4343 case Intrinsic::convertsif:
4344 case Intrinsic::convertuif:
4345 case Intrinsic::convertss:
4346 case Intrinsic::convertsu:
4347 case Intrinsic::convertus:
4348 case Intrinsic::convertuu: {
4349 ISD::CvtCode Code = ISD::CVT_INVALID;
4350 switch (Intrinsic) {
4351 case Intrinsic::convertff: Code = ISD::CVT_FF; break;
4352 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break;
4353 case Intrinsic::convertfui: Code = ISD::CVT_FU; break;
4354 case Intrinsic::convertsif: Code = ISD::CVT_SF; break;
4355 case Intrinsic::convertuif: Code = ISD::CVT_UF; break;
4356 case Intrinsic::convertss: Code = ISD::CVT_SS; break;
4357 case Intrinsic::convertsu: Code = ISD::CVT_SU; break;
4358 case Intrinsic::convertus: Code = ISD::CVT_US; break;
4359 case Intrinsic::convertuu: Code = ISD::CVT_UU; break;
4361 EVT DestVT = TLI.getValueType(I.getType());
4362 Value *Op1 = I.getOperand(1);
4363 Res = DAG.getConvertRndSat(DestVT, getCurDebugLoc(), getValue(Op1),
4364 DAG.getValueType(DestVT),
4365 DAG.getValueType(getValue(Op1).getValueType()),
4366 getValue(I.getOperand(2)),
4367 getValue(I.getOperand(3)),
4370 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4373 case Intrinsic::sqrt:
4374 Res = DAG.getNode(ISD::FSQRT, dl,
4375 getValue(I.getOperand(1)).getValueType(),
4376 getValue(I.getOperand(1)));
4378 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4380 case Intrinsic::powi:
4381 Res = ExpandPowI(dl, getValue(I.getOperand(1)), getValue(I.getOperand(2)),
4384 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4386 case Intrinsic::sin:
4387 Res = DAG.getNode(ISD::FSIN, dl,
4388 getValue(I.getOperand(1)).getValueType(),
4389 getValue(I.getOperand(1)));
4391 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4393 case Intrinsic::cos:
4394 Res = DAG.getNode(ISD::FCOS, dl,
4395 getValue(I.getOperand(1)).getValueType(),
4396 getValue(I.getOperand(1)));
4398 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4400 case Intrinsic::log:
4403 case Intrinsic::log2:
4406 case Intrinsic::log10:
4409 case Intrinsic::exp:
4412 case Intrinsic::exp2:
4415 case Intrinsic::pow:
4418 case Intrinsic::pcmarker: {
4419 SDValue Tmp = getValue(I.getOperand(1));
4420 Res = DAG.getNode(ISD::PCMARKER, dl, MVT::Other, getRoot(), Tmp);
4422 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4425 case Intrinsic::readcyclecounter: {
4426 SDValue Op = getRoot();
4427 Res = DAG.getNode(ISD::READCYCLECOUNTER, dl,
4428 DAG.getVTList(MVT::i64, MVT::Other),
4431 DAG.setRoot(Res.getValue(1));
4432 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4435 case Intrinsic::bswap:
4436 Res = DAG.getNode(ISD::BSWAP, dl,
4437 getValue(I.getOperand(1)).getValueType(),
4438 getValue(I.getOperand(1)));
4440 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4442 case Intrinsic::cttz: {
4443 SDValue Arg = getValue(I.getOperand(1));
4444 EVT Ty = Arg.getValueType();
4445 Res = DAG.getNode(ISD::CTTZ, dl, Ty, Arg);
4447 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4450 case Intrinsic::ctlz: {
4451 SDValue Arg = getValue(I.getOperand(1));
4452 EVT Ty = Arg.getValueType();
4453 Res = DAG.getNode(ISD::CTLZ, dl, Ty, Arg);
4455 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4458 case Intrinsic::ctpop: {
4459 SDValue Arg = getValue(I.getOperand(1));
4460 EVT Ty = Arg.getValueType();
4461 Res = DAG.getNode(ISD::CTPOP, dl, Ty, Arg);
4463 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4466 case Intrinsic::stacksave: {
4467 SDValue Op = getRoot();
4468 Res = DAG.getNode(ISD::STACKSAVE, dl,
4469 DAG.getVTList(TLI.getPointerTy(), MVT::Other), &Op, 1);
4471 DAG.setRoot(Res.getValue(1));
4472 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4475 case Intrinsic::stackrestore: {
4476 Res = getValue(I.getOperand(1));
4477 Res = DAG.getNode(ISD::STACKRESTORE, dl, MVT::Other, getRoot(), Res);
4479 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4482 case Intrinsic::stackprotector: {
4483 // Emit code into the DAG to store the stack guard onto the stack.
4484 MachineFunction &MF = DAG.getMachineFunction();
4485 MachineFrameInfo *MFI = MF.getFrameInfo();
4486 EVT PtrTy = TLI.getPointerTy();
4488 SDValue Src = getValue(I.getOperand(1)); // The guard's value.
4489 AllocaInst *Slot = cast<AllocaInst>(I.getOperand(2));
4491 int FI = FuncInfo.StaticAllocaMap[Slot];
4492 MFI->setStackProtectorIndex(FI);
4494 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
4496 // Store the stack protector onto the stack.
4497 Res = DAG.getStore(getRoot(), getCurDebugLoc(), Src, FIN,
4498 PseudoSourceValue::getFixedStack(FI),
4502 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4505 case Intrinsic::objectsize: {
4506 // If we don't know by now, we're never going to know.
4507 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(2));
4509 assert(CI && "Non-constant type in __builtin_object_size?");
4511 SDValue Arg = getValue(I.getOperand(0));
4512 EVT Ty = Arg.getValueType();
4514 if (CI->getZExtValue() == 0)
4515 Res = DAG.getConstant(-1ULL, Ty);
4517 Res = DAG.getConstant(0, Ty);
4520 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4523 case Intrinsic::var_annotation:
4524 // Discard annotate attributes
4527 case Intrinsic::init_trampoline: {
4528 const Function *F = cast<Function>(I.getOperand(2)->stripPointerCasts());
4532 Ops[1] = getValue(I.getOperand(1));
4533 Ops[2] = getValue(I.getOperand(2));
4534 Ops[3] = getValue(I.getOperand(3));
4535 Ops[4] = DAG.getSrcValue(I.getOperand(1));
4536 Ops[5] = DAG.getSrcValue(F);
4538 Res = DAG.getNode(ISD::TRAMPOLINE, dl,
4539 DAG.getVTList(TLI.getPointerTy(), MVT::Other),
4543 DAG.setRoot(Res.getValue(1));
4544 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4547 case Intrinsic::gcroot:
4549 Value *Alloca = I.getOperand(1);
4550 Constant *TypeMap = cast<Constant>(I.getOperand(2));
4552 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode());
4553 GFI->addStackRoot(FI->getIndex(), TypeMap);
4556 case Intrinsic::gcread:
4557 case Intrinsic::gcwrite:
4558 llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!");
4560 case Intrinsic::flt_rounds:
4561 Res = DAG.getNode(ISD::FLT_ROUNDS_, dl, MVT::i32);
4563 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4565 case Intrinsic::trap:
4566 Res = DAG.getNode(ISD::TRAP, dl,MVT::Other, getRoot());
4568 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4570 case Intrinsic::uadd_with_overflow:
4571 return implVisitAluOverflow(I, ISD::UADDO);
4572 case Intrinsic::sadd_with_overflow:
4573 return implVisitAluOverflow(I, ISD::SADDO);
4574 case Intrinsic::usub_with_overflow:
4575 return implVisitAluOverflow(I, ISD::USUBO);
4576 case Intrinsic::ssub_with_overflow:
4577 return implVisitAluOverflow(I, ISD::SSUBO);
4578 case Intrinsic::umul_with_overflow:
4579 return implVisitAluOverflow(I, ISD::UMULO);
4580 case Intrinsic::smul_with_overflow:
4581 return implVisitAluOverflow(I, ISD::SMULO);
4583 case Intrinsic::prefetch: {
4586 Ops[1] = getValue(I.getOperand(1));
4587 Ops[2] = getValue(I.getOperand(2));
4588 Ops[3] = getValue(I.getOperand(3));
4589 Res = DAG.getNode(ISD::PREFETCH, dl, MVT::Other, &Ops[0], 4);
4591 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4595 case Intrinsic::memory_barrier: {
4598 for (int x = 1; x < 6; ++x)
4599 Ops[x] = getValue(I.getOperand(x));
4601 Res = DAG.getNode(ISD::MEMBARRIER, dl, MVT::Other, &Ops[0], 6);
4603 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4606 case Intrinsic::atomic_cmp_swap: {
4607 SDValue Root = getRoot();
4609 DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, getCurDebugLoc(),
4610 getValue(I.getOperand(2)).getValueType().getSimpleVT(),
4612 getValue(I.getOperand(1)),
4613 getValue(I.getOperand(2)),
4614 getValue(I.getOperand(3)),
4617 DAG.setRoot(L.getValue(1));
4618 DAG.AssignOrdering(L.getNode(), SDNodeOrder);
4621 case Intrinsic::atomic_load_add:
4622 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_ADD);
4623 case Intrinsic::atomic_load_sub:
4624 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_SUB);
4625 case Intrinsic::atomic_load_or:
4626 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_OR);
4627 case Intrinsic::atomic_load_xor:
4628 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_XOR);
4629 case Intrinsic::atomic_load_and:
4630 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_AND);
4631 case Intrinsic::atomic_load_nand:
4632 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_NAND);
4633 case Intrinsic::atomic_load_max:
4634 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MAX);
4635 case Intrinsic::atomic_load_min:
4636 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MIN);
4637 case Intrinsic::atomic_load_umin:
4638 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMIN);
4639 case Intrinsic::atomic_load_umax:
4640 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMAX);
4641 case Intrinsic::atomic_swap:
4642 return implVisitBinaryAtomic(I, ISD::ATOMIC_SWAP);
4644 case Intrinsic::invariant_start:
4645 case Intrinsic::lifetime_start:
4646 // Discard region information.
4647 Res = DAG.getUNDEF(TLI.getPointerTy());
4649 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4651 case Intrinsic::invariant_end:
4652 case Intrinsic::lifetime_end:
4653 // Discard region information.
4658 /// Test if the given instruction is in a position to be optimized
4659 /// with a tail-call. This roughly means that it's in a block with
4660 /// a return and there's nothing that needs to be scheduled
4661 /// between it and the return.
4663 /// This function only tests target-independent requirements.
4665 isInTailCallPosition(const Instruction *I, Attributes CalleeRetAttr,
4666 const TargetLowering &TLI) {
4667 const BasicBlock *ExitBB = I->getParent();
4668 const TerminatorInst *Term = ExitBB->getTerminator();
4669 const ReturnInst *Ret = dyn_cast<ReturnInst>(Term);
4670 const Function *F = ExitBB->getParent();
4672 // The block must end in a return statement or an unreachable.
4673 if (!Ret && !isa<UnreachableInst>(Term)) return false;
4675 // If I will have a chain, make sure no other instruction that will have a
4676 // chain interposes between I and the return.
4677 if (I->mayHaveSideEffects() || I->mayReadFromMemory() ||
4678 !I->isSafeToSpeculativelyExecute())
4679 for (BasicBlock::const_iterator BBI = prior(prior(ExitBB->end())); ;
4683 if (BBI->mayHaveSideEffects() || BBI->mayReadFromMemory() ||
4684 !BBI->isSafeToSpeculativelyExecute())
4688 // If the block ends with a void return or unreachable, it doesn't matter
4689 // what the call's return type is.
4690 if (!Ret || Ret->getNumOperands() == 0) return true;
4692 // If the return value is undef, it doesn't matter what the call's
4694 if (isa<UndefValue>(Ret->getOperand(0))) return true;
4696 // Conservatively require the attributes of the call to match those of
4697 // the return. Ignore noalias because it doesn't affect the call sequence.
4698 unsigned CallerRetAttr = F->getAttributes().getRetAttributes();
4699 if ((CalleeRetAttr ^ CallerRetAttr) & ~Attribute::NoAlias)
4702 // Otherwise, make sure the unmodified return value of I is the return value.
4703 for (const Instruction *U = dyn_cast<Instruction>(Ret->getOperand(0)); ;
4704 U = dyn_cast<Instruction>(U->getOperand(0))) {
4707 if (!U->hasOneUse())
4711 // Check for a truly no-op truncate.
4712 if (isa<TruncInst>(U) &&
4713 TLI.isTruncateFree(U->getOperand(0)->getType(), U->getType()))
4715 // Check for a truly no-op bitcast.
4716 if (isa<BitCastInst>(U) &&
4717 (U->getOperand(0)->getType() == U->getType() ||
4718 (isa<PointerType>(U->getOperand(0)->getType()) &&
4719 isa<PointerType>(U->getType()))))
4721 // Otherwise it's not a true no-op.
4728 void SelectionDAGBuilder::LowerCallTo(CallSite CS, SDValue Callee,
4730 MachineBasicBlock *LandingPad) {
4731 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
4732 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
4733 const Type *RetTy = FTy->getReturnType();
4734 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
4735 unsigned BeginLabel = 0, EndLabel = 0;
4737 TargetLowering::ArgListTy Args;
4738 TargetLowering::ArgListEntry Entry;
4739 Args.reserve(CS.arg_size());
4741 // Check whether the function can return without sret-demotion.
4742 SmallVector<EVT, 4> OutVTs;
4743 SmallVector<ISD::ArgFlagsTy, 4> OutsFlags;
4744 SmallVector<uint64_t, 4> Offsets;
4745 getReturnInfo(RetTy, CS.getAttributes().getRetAttributes(),
4746 OutVTs, OutsFlags, TLI, &Offsets);
4748 bool CanLowerReturn = TLI.CanLowerReturn(CS.getCallingConv(),
4749 FTy->isVarArg(), OutVTs, OutsFlags, DAG);
4751 SDValue DemoteStackSlot;
4753 if (!CanLowerReturn) {
4754 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(
4755 FTy->getReturnType());
4756 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(
4757 FTy->getReturnType());
4758 MachineFunction &MF = DAG.getMachineFunction();
4759 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
4760 const Type *StackSlotPtrType = PointerType::getUnqual(FTy->getReturnType());
4762 DemoteStackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
4763 Entry.Node = DemoteStackSlot;
4764 Entry.Ty = StackSlotPtrType;
4765 Entry.isSExt = false;
4766 Entry.isZExt = false;
4767 Entry.isInReg = false;
4768 Entry.isSRet = true;
4769 Entry.isNest = false;
4770 Entry.isByVal = false;
4771 Entry.Alignment = Align;
4772 Args.push_back(Entry);
4773 RetTy = Type::getVoidTy(FTy->getContext());
4776 for (CallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
4778 SDValue ArgNode = getValue(*i);
4779 Entry.Node = ArgNode; Entry.Ty = (*i)->getType();
4781 unsigned attrInd = i - CS.arg_begin() + 1;
4782 Entry.isSExt = CS.paramHasAttr(attrInd, Attribute::SExt);
4783 Entry.isZExt = CS.paramHasAttr(attrInd, Attribute::ZExt);
4784 Entry.isInReg = CS.paramHasAttr(attrInd, Attribute::InReg);
4785 Entry.isSRet = CS.paramHasAttr(attrInd, Attribute::StructRet);
4786 Entry.isNest = CS.paramHasAttr(attrInd, Attribute::Nest);
4787 Entry.isByVal = CS.paramHasAttr(attrInd, Attribute::ByVal);
4788 Entry.Alignment = CS.getParamAlignment(attrInd);
4789 Args.push_back(Entry);
4792 if (LandingPad && MMI) {
4793 // Insert a label before the invoke call to mark the try range. This can be
4794 // used to detect deletion of the invoke via the MachineModuleInfo.
4795 BeginLabel = MMI->NextLabelID();
4797 // For SjLj, keep track of which landing pads go with which invokes
4798 // so as to maintain the ordering of pads in the LSDA.
4799 unsigned CallSiteIndex = MMI->getCurrentCallSite();
4800 if (CallSiteIndex) {
4801 MMI->setCallSiteBeginLabel(BeginLabel, CallSiteIndex);
4802 // Now that the call site is handled, stop tracking it.
4803 MMI->setCurrentCallSite(0);
4806 // Both PendingLoads and PendingExports must be flushed here;
4807 // this call might not return.
4809 DAG.setRoot(DAG.getLabel(ISD::EH_LABEL, getCurDebugLoc(),
4810 getControlRoot(), BeginLabel));
4813 // Check if target-independent constraints permit a tail call here.
4814 // Target-dependent constraints are checked within TLI.LowerCallTo.
4816 !isInTailCallPosition(CS.getInstruction(),
4817 CS.getAttributes().getRetAttributes(),
4821 std::pair<SDValue,SDValue> Result =
4822 TLI.LowerCallTo(getRoot(), RetTy,
4823 CS.paramHasAttr(0, Attribute::SExt),
4824 CS.paramHasAttr(0, Attribute::ZExt), FTy->isVarArg(),
4825 CS.paramHasAttr(0, Attribute::InReg), FTy->getNumParams(),
4826 CS.getCallingConv(),
4828 !CS.getInstruction()->use_empty(),
4829 Callee, Args, DAG, getCurDebugLoc(), SDNodeOrder);
4830 assert((isTailCall || Result.second.getNode()) &&
4831 "Non-null chain expected with non-tail call!");
4832 assert((Result.second.getNode() || !Result.first.getNode()) &&
4833 "Null value expected with tail call!");
4834 if (Result.first.getNode()) {
4835 setValue(CS.getInstruction(), Result.first);
4836 DAG.AssignOrdering(Result.first.getNode(), SDNodeOrder);
4837 } else if (!CanLowerReturn && Result.second.getNode()) {
4838 // The instruction result is the result of loading from the
4839 // hidden sret parameter.
4840 SmallVector<EVT, 1> PVTs;
4841 const Type *PtrRetTy = PointerType::getUnqual(FTy->getReturnType());
4843 ComputeValueVTs(TLI, PtrRetTy, PVTs);
4844 assert(PVTs.size() == 1 && "Pointers should fit in one register");
4845 EVT PtrVT = PVTs[0];
4846 unsigned NumValues = OutVTs.size();
4847 SmallVector<SDValue, 4> Values(NumValues);
4848 SmallVector<SDValue, 4> Chains(NumValues);
4850 for (unsigned i = 0; i < NumValues; ++i) {
4851 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT,
4853 DAG.getConstant(Offsets[i], PtrVT));
4854 SDValue L = DAG.getLoad(OutVTs[i], getCurDebugLoc(), Result.second,
4855 Add, NULL, Offsets[i], false, 1);
4857 Chains[i] = L.getValue(1);
4860 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
4861 MVT::Other, &Chains[0], NumValues);
4862 PendingLoads.push_back(Chain);
4864 // Collect the legal value parts into potentially illegal values
4865 // that correspond to the original function's return values.
4866 SmallVector<EVT, 4> RetTys;
4867 RetTy = FTy->getReturnType();
4868 ComputeValueVTs(TLI, RetTy, RetTys);
4869 ISD::NodeType AssertOp = ISD::DELETED_NODE;
4870 SmallVector<SDValue, 4> ReturnValues;
4871 unsigned CurReg = 0;
4872 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
4874 EVT RegisterVT = TLI.getRegisterType(RetTy->getContext(), VT);
4875 unsigned NumRegs = TLI.getNumRegisters(RetTy->getContext(), VT);
4877 SDValue ReturnValue =
4878 getCopyFromParts(DAG, getCurDebugLoc(), SDNodeOrder, &Values[CurReg], NumRegs,
4879 RegisterVT, VT, AssertOp);
4880 ReturnValues.push_back(ReturnValue);
4881 DAG.AssignOrdering(ReturnValue.getNode(), SDNodeOrder);
4884 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
4885 DAG.getVTList(&RetTys[0], RetTys.size()),
4886 &ReturnValues[0], ReturnValues.size());
4888 setValue(CS.getInstruction(), Res);
4890 DAG.AssignOrdering(Chain.getNode(), SDNodeOrder);
4891 DAG.AssignOrdering(Res.getNode(), SDNodeOrder);
4894 // As a special case, a null chain means that a tail call has been emitted and
4895 // the DAG root is already updated.
4896 if (Result.second.getNode()) {
4897 DAG.setRoot(Result.second);
4898 DAG.AssignOrdering(Result.second.getNode(), SDNodeOrder);
4903 if (LandingPad && MMI) {
4904 // Insert a label at the end of the invoke call to mark the try range. This
4905 // can be used to detect deletion of the invoke via the MachineModuleInfo.
4906 EndLabel = MMI->NextLabelID();
4907 DAG.setRoot(DAG.getLabel(ISD::EH_LABEL, getCurDebugLoc(),
4908 getRoot(), EndLabel));
4910 // Inform MachineModuleInfo of range.
4911 MMI->addInvoke(LandingPad, BeginLabel, EndLabel);
4915 /// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the
4916 /// value is equal or not-equal to zero.
4917 static bool IsOnlyUsedInZeroEqualityComparison(Value *V) {
4918 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
4920 if (ICmpInst *IC = dyn_cast<ICmpInst>(*UI))
4921 if (IC->isEquality())
4922 if (Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
4923 if (C->isNullValue())
4925 // Unknown instruction.
4931 static SDValue getMemCmpLoad(Value *PtrVal, MVT LoadVT, const Type *LoadTy,
4932 SelectionDAGBuilder &Builder) {
4934 // Check to see if this load can be trivially constant folded, e.g. if the
4935 // input is from a string literal.
4936 if (Constant *LoadInput = dyn_cast<Constant>(PtrVal)) {
4937 // Cast pointer to the type we really want to load.
4938 LoadInput = ConstantExpr::getBitCast(LoadInput,
4939 PointerType::getUnqual(LoadTy));
4941 if (Constant *LoadCst = ConstantFoldLoadFromConstPtr(LoadInput, Builder.TD))
4942 return Builder.getValue(LoadCst);
4945 // Otherwise, we have to emit the load. If the pointer is to unfoldable but
4946 // still constant memory, the input chain can be the entry node.
4948 bool ConstantMemory = false;
4950 // Do not serialize (non-volatile) loads of constant memory with anything.
4951 if (Builder.AA->pointsToConstantMemory(PtrVal)) {
4952 Root = Builder.DAG.getEntryNode();
4953 ConstantMemory = true;
4955 // Do not serialize non-volatile loads against each other.
4956 Root = Builder.DAG.getRoot();
4959 SDValue Ptr = Builder.getValue(PtrVal);
4960 SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurDebugLoc(), Root,
4961 Ptr, PtrVal /*SrcValue*/, 0/*SVOffset*/,
4962 false /*volatile*/, 1 /* align=1 */);
4964 if (!ConstantMemory)
4965 Builder.PendingLoads.push_back(LoadVal.getValue(1));
4970 /// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form.
4971 /// If so, return true and lower it, otherwise return false and it will be
4972 /// lowered like a normal call.
4973 bool SelectionDAGBuilder::visitMemCmpCall(CallInst &I) {
4974 // Verify that the prototype makes sense. int memcmp(void*,void*,size_t)
4975 if (I.getNumOperands() != 4)
4978 Value *LHS = I.getOperand(1), *RHS = I.getOperand(2);
4979 if (!isa<PointerType>(LHS->getType()) || !isa<PointerType>(RHS->getType()) ||
4980 !isa<IntegerType>(I.getOperand(3)->getType()) ||
4981 !isa<IntegerType>(I.getType()))
4984 ConstantInt *Size = dyn_cast<ConstantInt>(I.getOperand(3));
4986 // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0
4987 // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0
4988 if (Size && IsOnlyUsedInZeroEqualityComparison(&I)) {
4989 bool ActuallyDoIt = true;
4992 switch (Size->getZExtValue()) {
4994 LoadVT = MVT::Other;
4996 ActuallyDoIt = false;
5000 LoadTy = Type::getInt16Ty(Size->getContext());
5004 LoadTy = Type::getInt32Ty(Size->getContext());
5008 LoadTy = Type::getInt64Ty(Size->getContext());
5012 LoadVT = MVT::v4i32;
5013 LoadTy = Type::getInt32Ty(Size->getContext());
5014 LoadTy = VectorType::get(LoadTy, 4);
5019 // This turns into unaligned loads. We only do this if the target natively
5020 // supports the MVT we'll be loading or if it is small enough (<= 4) that
5021 // we'll only produce a small number of byte loads.
5023 // Require that we can find a legal MVT, and only do this if the target
5024 // supports unaligned loads of that type. Expanding into byte loads would
5026 if (ActuallyDoIt && Size->getZExtValue() > 4) {
5027 // TODO: Handle 5 byte compare as 4-byte + 1 byte.
5028 // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads.
5029 if (!TLI.isTypeLegal(LoadVT) ||!TLI.allowsUnalignedMemoryAccesses(LoadVT))
5030 ActuallyDoIt = false;
5034 SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this);
5035 SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this);
5037 SDValue Res = DAG.getSetCC(getCurDebugLoc(), MVT::i1, LHSVal, RHSVal,
5039 EVT CallVT = TLI.getValueType(I.getType(), true);
5040 setValue(&I, DAG.getZExtOrTrunc(Res, getCurDebugLoc(), CallVT));
5050 void SelectionDAGBuilder::visitCall(CallInst &I) {
5051 const char *RenameFn = 0;
5052 if (Function *F = I.getCalledFunction()) {
5053 if (F->isDeclaration()) {
5054 const TargetIntrinsicInfo *II = TLI.getTargetMachine().getIntrinsicInfo();
5056 if (unsigned IID = II->getIntrinsicID(F)) {
5057 RenameFn = visitIntrinsicCall(I, IID);
5062 if (unsigned IID = F->getIntrinsicID()) {
5063 RenameFn = visitIntrinsicCall(I, IID);
5069 // Check for well-known libc/libm calls. If the function is internal, it
5070 // can't be a library call.
5071 if (!F->hasLocalLinkage() && F->hasName()) {
5072 StringRef Name = F->getName();
5073 if (Name == "copysign" || Name == "copysignf") {
5074 if (I.getNumOperands() == 3 && // Basic sanity checks.
5075 I.getOperand(1)->getType()->isFloatingPoint() &&
5076 I.getType() == I.getOperand(1)->getType() &&
5077 I.getType() == I.getOperand(2)->getType()) {
5078 SDValue LHS = getValue(I.getOperand(1));
5079 SDValue RHS = getValue(I.getOperand(2));
5080 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurDebugLoc(),
5081 LHS.getValueType(), LHS, RHS));
5084 } else if (Name == "fabs" || Name == "fabsf" || Name == "fabsl") {
5085 if (I.getNumOperands() == 2 && // Basic sanity checks.
5086 I.getOperand(1)->getType()->isFloatingPoint() &&
5087 I.getType() == I.getOperand(1)->getType()) {
5088 SDValue Tmp = getValue(I.getOperand(1));
5089 setValue(&I, DAG.getNode(ISD::FABS, getCurDebugLoc(),
5090 Tmp.getValueType(), Tmp));
5093 } else if (Name == "sin" || Name == "sinf" || Name == "sinl") {
5094 if (I.getNumOperands() == 2 && // Basic sanity checks.
5095 I.getOperand(1)->getType()->isFloatingPoint() &&
5096 I.getType() == I.getOperand(1)->getType() &&
5097 I.onlyReadsMemory()) {
5098 SDValue Tmp = getValue(I.getOperand(1));
5099 setValue(&I, DAG.getNode(ISD::FSIN, getCurDebugLoc(),
5100 Tmp.getValueType(), Tmp));
5103 } else if (Name == "cos" || Name == "cosf" || Name == "cosl") {
5104 if (I.getNumOperands() == 2 && // Basic sanity checks.
5105 I.getOperand(1)->getType()->isFloatingPoint() &&
5106 I.getType() == I.getOperand(1)->getType() &&
5107 I.onlyReadsMemory()) {
5108 SDValue Tmp = getValue(I.getOperand(1));
5109 setValue(&I, DAG.getNode(ISD::FCOS, getCurDebugLoc(),
5110 Tmp.getValueType(), Tmp));
5113 } else if (Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl") {
5114 if (I.getNumOperands() == 2 && // Basic sanity checks.
5115 I.getOperand(1)->getType()->isFloatingPoint() &&
5116 I.getType() == I.getOperand(1)->getType() &&
5117 I.onlyReadsMemory()) {
5118 SDValue Tmp = getValue(I.getOperand(1));
5119 setValue(&I, DAG.getNode(ISD::FSQRT, getCurDebugLoc(),
5120 Tmp.getValueType(), Tmp));
5123 } else if (Name == "memcmp") {
5124 if (visitMemCmpCall(I))
5128 } else if (isa<InlineAsm>(I.getOperand(0))) {
5135 Callee = getValue(I.getOperand(0));
5137 Callee = DAG.getExternalSymbol(RenameFn, TLI.getPointerTy());
5139 // Check if we can potentially perform a tail call. More detailed checking is
5140 // be done within LowerCallTo, after more information about the call is known.
5141 LowerCallTo(&I, Callee, I.isTailCall());
5144 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
5145 /// this value and returns the result as a ValueVT value. This uses
5146 /// Chain/Flag as the input and updates them for the output Chain/Flag.
5147 /// If the Flag pointer is NULL, no flag is used.
5148 SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG, DebugLoc dl,
5149 unsigned Order, SDValue &Chain,
5150 SDValue *Flag) const {
5151 // Assemble the legal parts into the final values.
5152 SmallVector<SDValue, 4> Values(ValueVTs.size());
5153 SmallVector<SDValue, 8> Parts;
5154 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
5155 // Copy the legal parts from the registers.
5156 EVT ValueVT = ValueVTs[Value];
5157 unsigned NumRegs = TLI->getNumRegisters(*DAG.getContext(), ValueVT);
5158 EVT RegisterVT = RegVTs[Value];
5160 Parts.resize(NumRegs);
5161 for (unsigned i = 0; i != NumRegs; ++i) {
5164 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT);
5166 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag);
5167 *Flag = P.getValue(2);
5170 Chain = P.getValue(1);
5171 DAG.AssignOrdering(P.getNode(), Order);
5173 // If the source register was virtual and if we know something about it,
5174 // add an assert node.
5175 if (TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) &&
5176 RegisterVT.isInteger() && !RegisterVT.isVector()) {
5177 unsigned SlotNo = Regs[Part+i]-TargetRegisterInfo::FirstVirtualRegister;
5178 FunctionLoweringInfo &FLI = DAG.getFunctionLoweringInfo();
5179 if (FLI.LiveOutRegInfo.size() > SlotNo) {
5180 FunctionLoweringInfo::LiveOutInfo &LOI = FLI.LiveOutRegInfo[SlotNo];
5182 unsigned RegSize = RegisterVT.getSizeInBits();
5183 unsigned NumSignBits = LOI.NumSignBits;
5184 unsigned NumZeroBits = LOI.KnownZero.countLeadingOnes();
5186 // FIXME: We capture more information than the dag can represent. For
5187 // now, just use the tightest assertzext/assertsext possible.
5189 EVT FromVT(MVT::Other);
5190 if (NumSignBits == RegSize)
5191 isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1
5192 else if (NumZeroBits >= RegSize-1)
5193 isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1
5194 else if (NumSignBits > RegSize-8)
5195 isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8
5196 else if (NumZeroBits >= RegSize-8)
5197 isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8
5198 else if (NumSignBits > RegSize-16)
5199 isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16
5200 else if (NumZeroBits >= RegSize-16)
5201 isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16
5202 else if (NumSignBits > RegSize-32)
5203 isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32
5204 else if (NumZeroBits >= RegSize-32)
5205 isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32
5207 if (FromVT != MVT::Other) {
5208 P = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl,
5209 RegisterVT, P, DAG.getValueType(FromVT));
5210 DAG.AssignOrdering(P.getNode(), Order);
5218 Values[Value] = getCopyFromParts(DAG, dl, Order, Parts.begin(),
5219 NumRegs, RegisterVT, ValueVT);
5220 DAG.AssignOrdering(Values[Value].getNode(), Order);
5225 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, dl,
5226 DAG.getVTList(&ValueVTs[0], ValueVTs.size()),
5227 &Values[0], ValueVTs.size());
5228 DAG.AssignOrdering(Res.getNode(), Order);
5232 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
5233 /// specified value into the registers specified by this object. This uses
5234 /// Chain/Flag as the input and updates them for the output Chain/Flag.
5235 /// If the Flag pointer is NULL, no flag is used.
5236 void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
5237 unsigned Order, SDValue &Chain,
5238 SDValue *Flag) const {
5239 // Get the list of the values's legal parts.
5240 unsigned NumRegs = Regs.size();
5241 SmallVector<SDValue, 8> Parts(NumRegs);
5242 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
5243 EVT ValueVT = ValueVTs[Value];
5244 unsigned NumParts = TLI->getNumRegisters(*DAG.getContext(), ValueVT);
5245 EVT RegisterVT = RegVTs[Value];
5247 getCopyToParts(DAG, dl, Order,
5248 Val.getValue(Val.getResNo() + Value),
5249 &Parts[Part], NumParts, RegisterVT);
5253 // Copy the parts into the registers.
5254 SmallVector<SDValue, 8> Chains(NumRegs);
5255 for (unsigned i = 0; i != NumRegs; ++i) {
5258 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]);
5260 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag);
5261 *Flag = Part.getValue(1);
5264 Chains[i] = Part.getValue(0);
5265 DAG.AssignOrdering(Part.getNode(), Order);
5268 if (NumRegs == 1 || Flag)
5269 // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is
5270 // flagged to it. That is the CopyToReg nodes and the user are considered
5271 // a single scheduling unit. If we create a TokenFactor and return it as
5272 // chain, then the TokenFactor is both a predecessor (operand) of the
5273 // user as well as a successor (the TF operands are flagged to the user).
5274 // c1, f1 = CopyToReg
5275 // c2, f2 = CopyToReg
5276 // c3 = TokenFactor c1, c2
5279 Chain = Chains[NumRegs-1];
5281 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0], NumRegs);
5283 DAG.AssignOrdering(Chain.getNode(), Order);
5286 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
5287 /// operand list. This adds the code marker and includes the number of
5288 /// values added into it.
5289 void RegsForValue::AddInlineAsmOperands(unsigned Code,
5290 bool HasMatching,unsigned MatchingIdx,
5291 SelectionDAG &DAG, unsigned Order,
5292 std::vector<SDValue> &Ops) const {
5293 assert(Regs.size() < (1 << 13) && "Too many inline asm outputs!");
5294 unsigned Flag = Code | (Regs.size() << 3);
5296 Flag |= 0x80000000 | (MatchingIdx << 16);
5297 SDValue Res = DAG.getTargetConstant(Flag, MVT::i32);
5300 DAG.AssignOrdering(Res.getNode(), Order);
5302 for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) {
5303 unsigned NumRegs = TLI->getNumRegisters(*DAG.getContext(), ValueVTs[Value]);
5304 EVT RegisterVT = RegVTs[Value];
5305 for (unsigned i = 0; i != NumRegs; ++i) {
5306 assert(Reg < Regs.size() && "Mismatch in # registers expected");
5307 SDValue Res = DAG.getRegister(Regs[Reg++], RegisterVT);
5309 DAG.AssignOrdering(Res.getNode(), Order);
5314 /// isAllocatableRegister - If the specified register is safe to allocate,
5315 /// i.e. it isn't a stack pointer or some other special register, return the
5316 /// register class for the register. Otherwise, return null.
5317 static const TargetRegisterClass *
5318 isAllocatableRegister(unsigned Reg, MachineFunction &MF,
5319 const TargetLowering &TLI,
5320 const TargetRegisterInfo *TRI) {
5321 EVT FoundVT = MVT::Other;
5322 const TargetRegisterClass *FoundRC = 0;
5323 for (TargetRegisterInfo::regclass_iterator RCI = TRI->regclass_begin(),
5324 E = TRI->regclass_end(); RCI != E; ++RCI) {
5325 EVT ThisVT = MVT::Other;
5327 const TargetRegisterClass *RC = *RCI;
5328 // If none of the the value types for this register class are valid, we
5329 // can't use it. For example, 64-bit reg classes on 32-bit targets.
5330 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
5332 if (TLI.isTypeLegal(*I)) {
5333 // If we have already found this register in a different register class,
5334 // choose the one with the largest VT specified. For example, on
5335 // PowerPC, we favor f64 register classes over f32.
5336 if (FoundVT == MVT::Other || FoundVT.bitsLT(*I)) {
5343 if (ThisVT == MVT::Other) continue;
5345 // NOTE: This isn't ideal. In particular, this might allocate the
5346 // frame pointer in functions that need it (due to them not being taken
5347 // out of allocation, because a variable sized allocation hasn't been seen
5348 // yet). This is a slight code pessimization, but should still work.
5349 for (TargetRegisterClass::iterator I = RC->allocation_order_begin(MF),
5350 E = RC->allocation_order_end(MF); I != E; ++I)
5352 // We found a matching register class. Keep looking at others in case
5353 // we find one with larger registers that this physreg is also in.
5364 /// AsmOperandInfo - This contains information for each constraint that we are
5366 class VISIBILITY_HIDDEN SDISelAsmOperandInfo :
5367 public TargetLowering::AsmOperandInfo {
5369 /// CallOperand - If this is the result output operand or a clobber
5370 /// this is null, otherwise it is the incoming operand to the CallInst.
5371 /// This gets modified as the asm is processed.
5372 SDValue CallOperand;
5374 /// AssignedRegs - If this is a register or register class operand, this
5375 /// contains the set of register corresponding to the operand.
5376 RegsForValue AssignedRegs;
5378 explicit SDISelAsmOperandInfo(const InlineAsm::ConstraintInfo &info)
5379 : TargetLowering::AsmOperandInfo(info), CallOperand(0,0) {
5382 /// MarkAllocatedRegs - Once AssignedRegs is set, mark the assigned registers
5383 /// busy in OutputRegs/InputRegs.
5384 void MarkAllocatedRegs(bool isOutReg, bool isInReg,
5385 std::set<unsigned> &OutputRegs,
5386 std::set<unsigned> &InputRegs,
5387 const TargetRegisterInfo &TRI) const {
5389 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
5390 MarkRegAndAliases(AssignedRegs.Regs[i], OutputRegs, TRI);
5393 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
5394 MarkRegAndAliases(AssignedRegs.Regs[i], InputRegs, TRI);
5398 /// getCallOperandValEVT - Return the EVT of the Value* that this operand
5399 /// corresponds to. If there is no Value* for this operand, it returns
5401 EVT getCallOperandValEVT(LLVMContext &Context,
5402 const TargetLowering &TLI,
5403 const TargetData *TD) const {
5404 if (CallOperandVal == 0) return MVT::Other;
5406 if (isa<BasicBlock>(CallOperandVal))
5407 return TLI.getPointerTy();
5409 const llvm::Type *OpTy = CallOperandVal->getType();
5411 // If this is an indirect operand, the operand is a pointer to the
5414 const llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
5416 llvm_report_error("Indirect operand for inline asm not a pointer!");
5417 OpTy = PtrTy->getElementType();
5420 // If OpTy is not a single value, it may be a struct/union that we
5421 // can tile with integers.
5422 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
5423 unsigned BitSize = TD->getTypeSizeInBits(OpTy);
5432 OpTy = IntegerType::get(Context, BitSize);
5437 return TLI.getValueType(OpTy, true);
5441 /// MarkRegAndAliases - Mark the specified register and all aliases in the
5443 static void MarkRegAndAliases(unsigned Reg, std::set<unsigned> &Regs,
5444 const TargetRegisterInfo &TRI) {
5445 assert(TargetRegisterInfo::isPhysicalRegister(Reg) && "Isn't a physreg");
5447 if (const unsigned *Aliases = TRI.getAliasSet(Reg))
5448 for (; *Aliases; ++Aliases)
5449 Regs.insert(*Aliases);
5452 } // end llvm namespace.
5455 /// GetRegistersForValue - Assign registers (virtual or physical) for the
5456 /// specified operand. We prefer to assign virtual registers, to allow the
5457 /// register allocator to handle the assignment process. However, if the asm
5458 /// uses features that we can't model on machineinstrs, we have SDISel do the
5459 /// allocation. This produces generally horrible, but correct, code.
5461 /// OpInfo describes the operand.
5462 /// Input and OutputRegs are the set of already allocated physical registers.
5464 void SelectionDAGBuilder::
5465 GetRegistersForValue(SDISelAsmOperandInfo &OpInfo,
5466 std::set<unsigned> &OutputRegs,
5467 std::set<unsigned> &InputRegs) {
5468 LLVMContext &Context = FuncInfo.Fn->getContext();
5470 // Compute whether this value requires an input register, an output register,
5472 bool isOutReg = false;
5473 bool isInReg = false;
5474 switch (OpInfo.Type) {
5475 case InlineAsm::isOutput:
5478 // If there is an input constraint that matches this, we need to reserve
5479 // the input register so no other inputs allocate to it.
5480 isInReg = OpInfo.hasMatchingInput();
5482 case InlineAsm::isInput:
5486 case InlineAsm::isClobber:
5493 MachineFunction &MF = DAG.getMachineFunction();
5494 SmallVector<unsigned, 4> Regs;
5496 // If this is a constraint for a single physreg, or a constraint for a
5497 // register class, find it.
5498 std::pair<unsigned, const TargetRegisterClass*> PhysReg =
5499 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
5500 OpInfo.ConstraintVT);
5502 unsigned NumRegs = 1;
5503 if (OpInfo.ConstraintVT != MVT::Other) {
5504 // If this is a FP input in an integer register (or visa versa) insert a bit
5505 // cast of the input value. More generally, handle any case where the input
5506 // value disagrees with the register class we plan to stick this in.
5507 if (OpInfo.Type == InlineAsm::isInput &&
5508 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) {
5509 // Try to convert to the first EVT that the reg class contains. If the
5510 // types are identical size, use a bitcast to convert (e.g. two differing
5512 EVT RegVT = *PhysReg.second->vt_begin();
5513 if (RegVT.getSizeInBits() == OpInfo.ConstraintVT.getSizeInBits()) {
5514 OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
5515 RegVT, OpInfo.CallOperand);
5516 OpInfo.ConstraintVT = RegVT;
5517 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) {
5518 // If the input is a FP value and we want it in FP registers, do a
5519 // bitcast to the corresponding integer type. This turns an f64 value
5520 // into i64, which can be passed with two i32 values on a 32-bit
5522 RegVT = EVT::getIntegerVT(Context,
5523 OpInfo.ConstraintVT.getSizeInBits());
5524 OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
5525 RegVT, OpInfo.CallOperand);
5526 OpInfo.ConstraintVT = RegVT;
5529 DAG.AssignOrdering(OpInfo.CallOperand.getNode(), SDNodeOrder);
5532 NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT);
5536 EVT ValueVT = OpInfo.ConstraintVT;
5538 // If this is a constraint for a specific physical register, like {r17},
5540 if (unsigned AssignedReg = PhysReg.first) {
5541 const TargetRegisterClass *RC = PhysReg.second;
5542 if (OpInfo.ConstraintVT == MVT::Other)
5543 ValueVT = *RC->vt_begin();
5545 // Get the actual register value type. This is important, because the user
5546 // may have asked for (e.g.) the AX register in i32 type. We need to
5547 // remember that AX is actually i16 to get the right extension.
5548 RegVT = *RC->vt_begin();
5550 // This is a explicit reference to a physical register.
5551 Regs.push_back(AssignedReg);
5553 // If this is an expanded reference, add the rest of the regs to Regs.
5555 TargetRegisterClass::iterator I = RC->begin();
5556 for (; *I != AssignedReg; ++I)
5557 assert(I != RC->end() && "Didn't find reg!");
5559 // Already added the first reg.
5561 for (; NumRegs; --NumRegs, ++I) {
5562 assert(I != RC->end() && "Ran out of registers to allocate!");
5567 OpInfo.AssignedRegs = RegsForValue(TLI, Regs, RegVT, ValueVT);
5568 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
5569 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
5573 // Otherwise, if this was a reference to an LLVM register class, create vregs
5574 // for this reference.
5575 if (const TargetRegisterClass *RC = PhysReg.second) {
5576 RegVT = *RC->vt_begin();
5577 if (OpInfo.ConstraintVT == MVT::Other)
5580 // Create the appropriate number of virtual registers.
5581 MachineRegisterInfo &RegInfo = MF.getRegInfo();
5582 for (; NumRegs; --NumRegs)
5583 Regs.push_back(RegInfo.createVirtualRegister(RC));
5585 OpInfo.AssignedRegs = RegsForValue(TLI, Regs, RegVT, ValueVT);
5589 // This is a reference to a register class that doesn't directly correspond
5590 // to an LLVM register class. Allocate NumRegs consecutive, available,
5591 // registers from the class.
5592 std::vector<unsigned> RegClassRegs
5593 = TLI.getRegClassForInlineAsmConstraint(OpInfo.ConstraintCode,
5594 OpInfo.ConstraintVT);
5596 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
5597 unsigned NumAllocated = 0;
5598 for (unsigned i = 0, e = RegClassRegs.size(); i != e; ++i) {
5599 unsigned Reg = RegClassRegs[i];
5600 // See if this register is available.
5601 if ((isOutReg && OutputRegs.count(Reg)) || // Already used.
5602 (isInReg && InputRegs.count(Reg))) { // Already used.
5603 // Make sure we find consecutive registers.
5608 // Check to see if this register is allocatable (i.e. don't give out the
5610 const TargetRegisterClass *RC = isAllocatableRegister(Reg, MF, TLI, TRI);
5611 if (!RC) { // Couldn't allocate this register.
5612 // Reset NumAllocated to make sure we return consecutive registers.
5617 // Okay, this register is good, we can use it.
5620 // If we allocated enough consecutive registers, succeed.
5621 if (NumAllocated == NumRegs) {
5622 unsigned RegStart = (i-NumAllocated)+1;
5623 unsigned RegEnd = i+1;
5624 // Mark all of the allocated registers used.
5625 for (unsigned i = RegStart; i != RegEnd; ++i)
5626 Regs.push_back(RegClassRegs[i]);
5628 OpInfo.AssignedRegs = RegsForValue(TLI, Regs, *RC->vt_begin(),
5629 OpInfo.ConstraintVT);
5630 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
5635 // Otherwise, we couldn't allocate enough registers for this.
5638 /// hasInlineAsmMemConstraint - Return true if the inline asm instruction being
5639 /// processed uses a memory 'm' constraint.
5641 hasInlineAsmMemConstraint(std::vector<InlineAsm::ConstraintInfo> &CInfos,
5642 const TargetLowering &TLI) {
5643 for (unsigned i = 0, e = CInfos.size(); i != e; ++i) {
5644 InlineAsm::ConstraintInfo &CI = CInfos[i];
5645 for (unsigned j = 0, ee = CI.Codes.size(); j != ee; ++j) {
5646 TargetLowering::ConstraintType CType = TLI.getConstraintType(CI.Codes[j]);
5647 if (CType == TargetLowering::C_Memory)
5651 // Indirect operand accesses access memory.
5659 /// visitInlineAsm - Handle a call to an InlineAsm object.
5661 void SelectionDAGBuilder::visitInlineAsm(CallSite CS) {
5662 InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
5664 /// ConstraintOperands - Information about all of the constraints.
5665 std::vector<SDISelAsmOperandInfo> ConstraintOperands;
5667 std::set<unsigned> OutputRegs, InputRegs;
5669 // Do a prepass over the constraints, canonicalizing them, and building up the
5670 // ConstraintOperands list.
5671 std::vector<InlineAsm::ConstraintInfo>
5672 ConstraintInfos = IA->ParseConstraints();
5674 bool hasMemory = hasInlineAsmMemConstraint(ConstraintInfos, TLI);
5676 SDValue Chain, Flag;
5678 // We won't need to flush pending loads if this asm doesn't touch
5679 // memory and is nonvolatile.
5680 if (hasMemory || IA->hasSideEffects())
5683 Chain = DAG.getRoot();
5685 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
5686 unsigned ResNo = 0; // ResNo - The result number of the next output.
5687 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
5688 ConstraintOperands.push_back(SDISelAsmOperandInfo(ConstraintInfos[i]));
5689 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
5691 EVT OpVT = MVT::Other;
5693 // Compute the value type for each operand.
5694 switch (OpInfo.Type) {
5695 case InlineAsm::isOutput:
5696 // Indirect outputs just consume an argument.
5697 if (OpInfo.isIndirect) {
5698 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
5702 // The return value of the call is this value. As such, there is no
5703 // corresponding argument.
5704 assert(!CS.getType()->isVoidTy() &&
5706 if (const StructType *STy = dyn_cast<StructType>(CS.getType())) {
5707 OpVT = TLI.getValueType(STy->getElementType(ResNo));
5709 assert(ResNo == 0 && "Asm only has one result!");
5710 OpVT = TLI.getValueType(CS.getType());
5714 case InlineAsm::isInput:
5715 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
5717 case InlineAsm::isClobber:
5722 // If this is an input or an indirect output, process the call argument.
5723 // BasicBlocks are labels, currently appearing only in asm's.
5724 if (OpInfo.CallOperandVal) {
5725 // Strip bitcasts, if any. This mostly comes up for functions.
5726 OpInfo.CallOperandVal = OpInfo.CallOperandVal->stripPointerCasts();
5728 if (BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) {
5729 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]);
5731 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
5734 OpVT = OpInfo.getCallOperandValEVT(*DAG.getContext(), TLI, TD);
5737 OpInfo.ConstraintVT = OpVT;
5740 // Second pass over the constraints: compute which constraint option to use
5741 // and assign registers to constraints that want a specific physreg.
5742 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
5743 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5745 // If this is an output operand with a matching input operand, look up the
5746 // matching input. If their types mismatch, e.g. one is an integer, the
5747 // other is floating point, or their sizes are different, flag it as an
5749 if (OpInfo.hasMatchingInput()) {
5750 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
5751 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
5752 if ((OpInfo.ConstraintVT.isInteger() !=
5753 Input.ConstraintVT.isInteger()) ||
5754 (OpInfo.ConstraintVT.getSizeInBits() !=
5755 Input.ConstraintVT.getSizeInBits())) {
5756 llvm_report_error("Unsupported asm: input constraint"
5757 " with a matching output constraint of incompatible"
5760 Input.ConstraintVT = OpInfo.ConstraintVT;
5764 // Compute the constraint code and ConstraintType to use.
5765 TLI.ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, hasMemory, &DAG);
5767 // If this is a memory input, and if the operand is not indirect, do what we
5768 // need to to provide an address for the memory input.
5769 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5770 !OpInfo.isIndirect) {
5771 assert(OpInfo.Type == InlineAsm::isInput &&
5772 "Can only indirectify direct input operands!");
5774 // Memory operands really want the address of the value. If we don't have
5775 // an indirect input, put it in the constpool if we can, otherwise spill
5776 // it to a stack slot.
5778 // If the operand is a float, integer, or vector constant, spill to a
5779 // constant pool entry to get its address.
5780 Value *OpVal = OpInfo.CallOperandVal;
5781 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
5782 isa<ConstantVector>(OpVal)) {
5783 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal),
5784 TLI.getPointerTy());
5786 // Otherwise, create a stack slot and emit a store to it before the
5788 const Type *Ty = OpVal->getType();
5789 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
5790 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(Ty);
5791 MachineFunction &MF = DAG.getMachineFunction();
5792 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
5793 SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
5794 Chain = DAG.getStore(Chain, getCurDebugLoc(),
5795 OpInfo.CallOperand, StackSlot, NULL, 0);
5796 OpInfo.CallOperand = StackSlot;
5799 // There is no longer a Value* corresponding to this operand.
5800 OpInfo.CallOperandVal = 0;
5802 // It is now an indirect operand.
5803 OpInfo.isIndirect = true;
5806 // If this constraint is for a specific register, allocate it before
5808 if (OpInfo.ConstraintType == TargetLowering::C_Register)
5809 GetRegistersForValue(OpInfo, OutputRegs, InputRegs);
5812 ConstraintInfos.clear();
5814 // Second pass - Loop over all of the operands, assigning virtual or physregs
5815 // to register class operands.
5816 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5817 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5819 // C_Register operands have already been allocated, Other/Memory don't need
5821 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
5822 GetRegistersForValue(OpInfo, OutputRegs, InputRegs);
5825 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
5826 std::vector<SDValue> AsmNodeOperands;
5827 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain
5828 AsmNodeOperands.push_back(
5829 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(),
5830 TLI.getPointerTy()));
5833 // Loop over all of the inputs, copying the operand values into the
5834 // appropriate registers and processing the output regs.
5835 RegsForValue RetValRegs;
5837 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
5838 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
5840 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5841 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5843 switch (OpInfo.Type) {
5844 case InlineAsm::isOutput: {
5845 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
5846 OpInfo.ConstraintType != TargetLowering::C_Register) {
5847 // Memory output, or 'other' output (e.g. 'X' constraint).
5848 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
5850 // Add information to the INLINEASM node to know about this output.
5851 unsigned ResOpType = 4/*MEM*/ | (1<<3);
5852 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
5853 TLI.getPointerTy()));
5854 AsmNodeOperands.push_back(OpInfo.CallOperand);
5858 // Otherwise, this is a register or register class output.
5860 // Copy the output from the appropriate register. Find a register that
5862 if (OpInfo.AssignedRegs.Regs.empty()) {
5863 llvm_report_error("Couldn't allocate output reg for"
5864 " constraint '" + OpInfo.ConstraintCode + "'!");
5867 // If this is an indirect operand, store through the pointer after the
5869 if (OpInfo.isIndirect) {
5870 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
5871 OpInfo.CallOperandVal));
5873 // This is the result value of the call.
5874 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
5875 // Concatenate this output onto the outputs list.
5876 RetValRegs.append(OpInfo.AssignedRegs);
5879 // Add information to the INLINEASM node to know that this register is
5881 OpInfo.AssignedRegs.AddInlineAsmOperands(OpInfo.isEarlyClobber ?
5882 6 /* EARLYCLOBBER REGDEF */ :
5890 case InlineAsm::isInput: {
5891 SDValue InOperandVal = OpInfo.CallOperand;
5893 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint?
5894 // If this is required to match an output register we have already set,
5895 // just use its register.
5896 unsigned OperandNo = OpInfo.getMatchedOperand();
5898 // Scan until we find the definition we already emitted of this operand.
5899 // When we find it, create a RegsForValue operand.
5900 unsigned CurOp = 2; // The first operand.
5901 for (; OperandNo; --OperandNo) {
5902 // Advance to the next operand.
5904 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
5905 assert(((OpFlag & 7) == 2 /*REGDEF*/ ||
5906 (OpFlag & 7) == 6 /*EARLYCLOBBER REGDEF*/ ||
5907 (OpFlag & 7) == 4 /*MEM*/) &&
5908 "Skipped past definitions?");
5909 CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1;
5913 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
5914 if ((OpFlag & 7) == 2 /*REGDEF*/
5915 || (OpFlag & 7) == 6 /* EARLYCLOBBER REGDEF */) {
5916 // Add (OpFlag&0xffff)>>3 registers to MatchedRegs.
5917 if (OpInfo.isIndirect) {
5918 llvm_report_error("Don't know how to handle tied indirect "
5919 "register inputs yet!");
5921 RegsForValue MatchedRegs;
5922 MatchedRegs.TLI = &TLI;
5923 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
5924 EVT RegVT = AsmNodeOperands[CurOp+1].getValueType();
5925 MatchedRegs.RegVTs.push_back(RegVT);
5926 MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo();
5927 for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag);
5929 MatchedRegs.Regs.push_back
5930 (RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT)));
5932 // Use the produced MatchedRegs object to
5933 MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
5934 SDNodeOrder, Chain, &Flag);
5935 MatchedRegs.AddInlineAsmOperands(1 /*REGUSE*/,
5936 true, OpInfo.getMatchedOperand(),
5937 DAG, SDNodeOrder, AsmNodeOperands);
5940 assert(((OpFlag & 7) == 4) && "Unknown matching constraint!");
5941 assert((InlineAsm::getNumOperandRegisters(OpFlag)) == 1 &&
5942 "Unexpected number of operands");
5943 // Add information to the INLINEASM node to know about this input.
5944 // See InlineAsm.h isUseOperandTiedToDef.
5945 OpFlag |= 0x80000000 | (OpInfo.getMatchedOperand() << 16);
5946 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlag,
5947 TLI.getPointerTy()));
5948 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]);
5953 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
5954 assert(!OpInfo.isIndirect &&
5955 "Don't know how to handle indirect other inputs yet!");
5957 std::vector<SDValue> Ops;
5958 TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode[0],
5959 hasMemory, Ops, DAG);
5961 llvm_report_error("Invalid operand for inline asm"
5962 " constraint '" + OpInfo.ConstraintCode + "'!");
5965 // Add information to the INLINEASM node to know about this input.
5966 unsigned ResOpType = 3 /*IMM*/ | (Ops.size() << 3);
5967 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
5968 TLI.getPointerTy()));
5969 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
5971 } else if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
5972 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
5973 assert(InOperandVal.getValueType() == TLI.getPointerTy() &&
5974 "Memory operands expect pointer values");
5976 // Add information to the INLINEASM node to know about this input.
5977 unsigned ResOpType = 4/*MEM*/ | (1<<3);
5978 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
5979 TLI.getPointerTy()));
5980 AsmNodeOperands.push_back(InOperandVal);
5984 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
5985 OpInfo.ConstraintType == TargetLowering::C_Register) &&
5986 "Unknown constraint type!");
5987 assert(!OpInfo.isIndirect &&
5988 "Don't know how to handle indirect register inputs yet!");
5990 // Copy the input into the appropriate registers.
5991 if (OpInfo.AssignedRegs.Regs.empty()) {
5992 llvm_report_error("Couldn't allocate input reg for"
5993 " constraint '"+ OpInfo.ConstraintCode +"'!");
5996 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
5997 SDNodeOrder, Chain, &Flag);
5999 OpInfo.AssignedRegs.AddInlineAsmOperands(1/*REGUSE*/, false, 0,
6004 case InlineAsm::isClobber: {
6005 // Add the clobbered value to the operand list, so that the register
6006 // allocator is aware that the physreg got clobbered.
6007 if (!OpInfo.AssignedRegs.Regs.empty())
6008 OpInfo.AssignedRegs.AddInlineAsmOperands(6 /* EARLYCLOBBER REGDEF */,
6009 false, 0, DAG, SDNodeOrder,
6016 // Finish up input operands.
6017 AsmNodeOperands[0] = Chain;
6018 if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
6020 Chain = DAG.getNode(ISD::INLINEASM, getCurDebugLoc(),
6021 DAG.getVTList(MVT::Other, MVT::Flag),
6022 &AsmNodeOperands[0], AsmNodeOperands.size());
6023 Flag = Chain.getValue(1);
6025 // If this asm returns a register value, copy the result from that register
6026 // and set it as the value of the call.
6027 if (!RetValRegs.Regs.empty()) {
6028 SDValue Val = RetValRegs.getCopyFromRegs(DAG, getCurDebugLoc(),
6029 SDNodeOrder, Chain, &Flag);
6031 // FIXME: Why don't we do this for inline asms with MRVs?
6032 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) {
6033 EVT ResultType = TLI.getValueType(CS.getType());
6035 // If any of the results of the inline asm is a vector, it may have the
6036 // wrong width/num elts. This can happen for register classes that can
6037 // contain multiple different value types. The preg or vreg allocated may
6038 // not have the same VT as was expected. Convert it to the right type
6039 // with bit_convert.
6040 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) {
6041 Val = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
6044 } else if (ResultType != Val.getValueType() &&
6045 ResultType.isInteger() && Val.getValueType().isInteger()) {
6046 // If a result value was tied to an input value, the computed result may
6047 // have a wider width than the expected result. Extract the relevant
6049 Val = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), ResultType, Val);
6052 assert(ResultType == Val.getValueType() && "Asm result value mismatch!");
6055 setValue(CS.getInstruction(), Val);
6056 // Don't need to use this as a chain in this case.
6057 if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty())
6061 std::vector<std::pair<SDValue, Value*> > StoresToEmit;
6063 // Process indirect outputs, first output all of the flagged copies out of
6065 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
6066 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
6067 Value *Ptr = IndirectStoresToEmit[i].second;
6068 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, getCurDebugLoc(),
6069 SDNodeOrder, Chain, &Flag);
6070 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
6074 // Emit the non-flagged stores from the physregs.
6075 SmallVector<SDValue, 8> OutChains;
6076 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) {
6077 SDValue Val = DAG.getStore(Chain, getCurDebugLoc(),
6078 StoresToEmit[i].first,
6079 getValue(StoresToEmit[i].second),
6080 StoresToEmit[i].second, 0);
6081 OutChains.push_back(Val);
6084 if (!OutChains.empty())
6085 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
6086 &OutChains[0], OutChains.size());
6091 void SelectionDAGBuilder::visitVAStart(CallInst &I) {
6092 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurDebugLoc(),
6093 MVT::Other, getRoot(),
6094 getValue(I.getOperand(1)),
6095 DAG.getSrcValue(I.getOperand(1))));
6098 void SelectionDAGBuilder::visitVAArg(VAArgInst &I) {
6099 SDValue V = DAG.getVAArg(TLI.getValueType(I.getType()), getCurDebugLoc(),
6100 getRoot(), getValue(I.getOperand(0)),
6101 DAG.getSrcValue(I.getOperand(0)));
6103 DAG.setRoot(V.getValue(1));
6106 void SelectionDAGBuilder::visitVAEnd(CallInst &I) {
6107 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurDebugLoc(),
6108 MVT::Other, getRoot(),
6109 getValue(I.getOperand(1)),
6110 DAG.getSrcValue(I.getOperand(1))));
6113 void SelectionDAGBuilder::visitVACopy(CallInst &I) {
6114 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurDebugLoc(),
6115 MVT::Other, getRoot(),
6116 getValue(I.getOperand(1)),
6117 getValue(I.getOperand(2)),
6118 DAG.getSrcValue(I.getOperand(1)),
6119 DAG.getSrcValue(I.getOperand(2))));
6122 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
6123 /// implementation, which just calls LowerCall.
6124 /// FIXME: When all targets are
6125 /// migrated to using LowerCall, this hook should be integrated into SDISel.
6126 std::pair<SDValue, SDValue>
6127 TargetLowering::LowerCallTo(SDValue Chain, const Type *RetTy,
6128 bool RetSExt, bool RetZExt, bool isVarArg,
6129 bool isInreg, unsigned NumFixedArgs,
6130 CallingConv::ID CallConv, bool isTailCall,
6131 bool isReturnValueUsed,
6133 ArgListTy &Args, SelectionDAG &DAG, DebugLoc dl,
6135 // Handle all of the outgoing arguments.
6136 SmallVector<ISD::OutputArg, 32> Outs;
6137 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
6138 SmallVector<EVT, 4> ValueVTs;
6139 ComputeValueVTs(*this, Args[i].Ty, ValueVTs);
6140 for (unsigned Value = 0, NumValues = ValueVTs.size();
6141 Value != NumValues; ++Value) {
6142 EVT VT = ValueVTs[Value];
6143 const Type *ArgTy = VT.getTypeForEVT(RetTy->getContext());
6144 SDValue Op = SDValue(Args[i].Node.getNode(),
6145 Args[i].Node.getResNo() + Value);
6146 ISD::ArgFlagsTy Flags;
6147 unsigned OriginalAlignment =
6148 getTargetData()->getABITypeAlignment(ArgTy);
6154 if (Args[i].isInReg)
6158 if (Args[i].isByVal) {
6160 const PointerType *Ty = cast<PointerType>(Args[i].Ty);
6161 const Type *ElementTy = Ty->getElementType();
6162 unsigned FrameAlign = getByValTypeAlignment(ElementTy);
6163 unsigned FrameSize = getTargetData()->getTypeAllocSize(ElementTy);
6164 // For ByVal, alignment should come from FE. BE will guess if this
6165 // info is not there but there are cases it cannot get right.
6166 if (Args[i].Alignment)
6167 FrameAlign = Args[i].Alignment;
6168 Flags.setByValAlign(FrameAlign);
6169 Flags.setByValSize(FrameSize);
6173 Flags.setOrigAlign(OriginalAlignment);
6175 EVT PartVT = getRegisterType(RetTy->getContext(), VT);
6176 unsigned NumParts = getNumRegisters(RetTy->getContext(), VT);
6177 SmallVector<SDValue, 4> Parts(NumParts);
6178 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
6181 ExtendKind = ISD::SIGN_EXTEND;
6182 else if (Args[i].isZExt)
6183 ExtendKind = ISD::ZERO_EXTEND;
6185 getCopyToParts(DAG, dl, Order, Op, &Parts[0], NumParts,
6186 PartVT, ExtendKind);
6188 for (unsigned j = 0; j != NumParts; ++j) {
6189 // if it isn't first piece, alignment must be 1
6190 ISD::OutputArg MyFlags(Flags, Parts[j], i < NumFixedArgs);
6191 if (NumParts > 1 && j == 0)
6192 MyFlags.Flags.setSplit();
6194 MyFlags.Flags.setOrigAlign(1);
6196 Outs.push_back(MyFlags);
6201 // Handle the incoming return values from the call.
6202 SmallVector<ISD::InputArg, 32> Ins;
6203 SmallVector<EVT, 4> RetTys;
6204 ComputeValueVTs(*this, RetTy, RetTys);
6205 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
6207 EVT RegisterVT = getRegisterType(RetTy->getContext(), VT);
6208 unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT);
6209 for (unsigned i = 0; i != NumRegs; ++i) {
6210 ISD::InputArg MyFlags;
6211 MyFlags.VT = RegisterVT;
6212 MyFlags.Used = isReturnValueUsed;
6214 MyFlags.Flags.setSExt();
6216 MyFlags.Flags.setZExt();
6218 MyFlags.Flags.setInReg();
6219 Ins.push_back(MyFlags);
6223 SmallVector<SDValue, 4> InVals;
6224 Chain = LowerCall(Chain, Callee, CallConv, isVarArg, isTailCall,
6225 Outs, Ins, dl, DAG, InVals);
6227 // Verify that the target's LowerCall behaved as expected.
6228 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
6229 "LowerCall didn't return a valid chain!");
6230 assert((!isTailCall || InVals.empty()) &&
6231 "LowerCall emitted a return value for a tail call!");
6232 assert((isTailCall || InVals.size() == Ins.size()) &&
6233 "LowerCall didn't emit the correct number of values!");
6234 DEBUG(for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
6235 assert(InVals[i].getNode() &&
6236 "LowerCall emitted a null value!");
6237 assert(Ins[i].VT == InVals[i].getValueType() &&
6238 "LowerCall emitted a value with the wrong type!");
6241 DAG.AssignOrdering(Chain.getNode(), Order);
6243 // For a tail call, the return value is merely live-out and there aren't
6244 // any nodes in the DAG representing it. Return a special value to
6245 // indicate that a tail call has been emitted and no more Instructions
6246 // should be processed in the current block.
6249 return std::make_pair(SDValue(), SDValue());
6252 // Collect the legal value parts into potentially illegal values
6253 // that correspond to the original function's return values.
6254 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6256 AssertOp = ISD::AssertSext;
6258 AssertOp = ISD::AssertZext;
6259 SmallVector<SDValue, 4> ReturnValues;
6260 unsigned CurReg = 0;
6261 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
6263 EVT RegisterVT = getRegisterType(RetTy->getContext(), VT);
6264 unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT);
6266 SDValue ReturnValue =
6267 getCopyFromParts(DAG, dl, Order, &InVals[CurReg], NumRegs,
6268 RegisterVT, VT, AssertOp);
6269 ReturnValues.push_back(ReturnValue);
6270 DAG.AssignOrdering(ReturnValue.getNode(), Order);
6274 // For a function returning void, there is no return value. We can't create
6275 // such a node, so we just return a null return value in that case. In
6276 // that case, nothing will actualy look at the value.
6277 if (ReturnValues.empty())
6278 return std::make_pair(SDValue(), Chain);
6280 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, dl,
6281 DAG.getVTList(&RetTys[0], RetTys.size()),
6282 &ReturnValues[0], ReturnValues.size());
6283 DAG.AssignOrdering(Res.getNode(), Order);
6284 return std::make_pair(Res, Chain);
6287 void TargetLowering::LowerOperationWrapper(SDNode *N,
6288 SmallVectorImpl<SDValue> &Results,
6289 SelectionDAG &DAG) {
6290 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
6292 Results.push_back(Res);
6295 SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
6296 llvm_unreachable("LowerOperation not implemented for this target!");
6300 void SelectionDAGBuilder::CopyValueToVirtualRegister(Value *V, unsigned Reg) {
6301 SDValue Op = getValue(V);
6302 assert((Op.getOpcode() != ISD::CopyFromReg ||
6303 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
6304 "Copy from a reg to the same reg!");
6305 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg");
6307 RegsForValue RFV(V->getContext(), TLI, Reg, V->getType());
6308 SDValue Chain = DAG.getEntryNode();
6309 RFV.getCopyToRegs(Op, DAG, getCurDebugLoc(), SDNodeOrder, Chain, 0);
6310 PendingExports.push_back(Chain);
6313 #include "llvm/CodeGen/SelectionDAGISel.h"
6315 void SelectionDAGISel::LowerArguments(BasicBlock *LLVMBB) {
6316 // If this is the entry block, emit arguments.
6317 Function &F = *LLVMBB->getParent();
6318 SelectionDAG &DAG = SDB->DAG;
6319 SDValue OldRoot = DAG.getRoot();
6320 DebugLoc dl = SDB->getCurDebugLoc();
6321 const TargetData *TD = TLI.getTargetData();
6322 SmallVector<ISD::InputArg, 16> Ins;
6324 // Check whether the function can return without sret-demotion.
6325 SmallVector<EVT, 4> OutVTs;
6326 SmallVector<ISD::ArgFlagsTy, 4> OutsFlags;
6327 getReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
6328 OutVTs, OutsFlags, TLI);
6329 FunctionLoweringInfo &FLI = DAG.getFunctionLoweringInfo();
6331 FLI.CanLowerReturn = TLI.CanLowerReturn(F.getCallingConv(), F.isVarArg(),
6332 OutVTs, OutsFlags, DAG);
6333 if (!FLI.CanLowerReturn) {
6334 // Put in an sret pointer parameter before all the other parameters.
6335 SmallVector<EVT, 1> ValueVTs;
6336 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
6338 // NOTE: Assuming that a pointer will never break down to more than one VT
6340 ISD::ArgFlagsTy Flags;
6342 EVT RegisterVT = TLI.getRegisterType(*CurDAG->getContext(), ValueVTs[0]);
6343 ISD::InputArg RetArg(Flags, RegisterVT, true);
6344 Ins.push_back(RetArg);
6347 // Set up the incoming argument description vector.
6349 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end();
6350 I != E; ++I, ++Idx) {
6351 SmallVector<EVT, 4> ValueVTs;
6352 ComputeValueVTs(TLI, I->getType(), ValueVTs);
6353 bool isArgValueUsed = !I->use_empty();
6354 for (unsigned Value = 0, NumValues = ValueVTs.size();
6355 Value != NumValues; ++Value) {
6356 EVT VT = ValueVTs[Value];
6357 const Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
6358 ISD::ArgFlagsTy Flags;
6359 unsigned OriginalAlignment =
6360 TD->getABITypeAlignment(ArgTy);
6362 if (F.paramHasAttr(Idx, Attribute::ZExt))
6364 if (F.paramHasAttr(Idx, Attribute::SExt))
6366 if (F.paramHasAttr(Idx, Attribute::InReg))
6368 if (F.paramHasAttr(Idx, Attribute::StructRet))
6370 if (F.paramHasAttr(Idx, Attribute::ByVal)) {
6372 const PointerType *Ty = cast<PointerType>(I->getType());
6373 const Type *ElementTy = Ty->getElementType();
6374 unsigned FrameAlign = TLI.getByValTypeAlignment(ElementTy);
6375 unsigned FrameSize = TD->getTypeAllocSize(ElementTy);
6376 // For ByVal, alignment should be passed from FE. BE will guess if
6377 // this info is not there but there are cases it cannot get right.
6378 if (F.getParamAlignment(Idx))
6379 FrameAlign = F.getParamAlignment(Idx);
6380 Flags.setByValAlign(FrameAlign);
6381 Flags.setByValSize(FrameSize);
6383 if (F.paramHasAttr(Idx, Attribute::Nest))
6385 Flags.setOrigAlign(OriginalAlignment);
6387 EVT RegisterVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6388 unsigned NumRegs = TLI.getNumRegisters(*CurDAG->getContext(), VT);
6389 for (unsigned i = 0; i != NumRegs; ++i) {
6390 ISD::InputArg MyFlags(Flags, RegisterVT, isArgValueUsed);
6391 if (NumRegs > 1 && i == 0)
6392 MyFlags.Flags.setSplit();
6393 // if it isn't first piece, alignment must be 1
6395 MyFlags.Flags.setOrigAlign(1);
6396 Ins.push_back(MyFlags);
6401 // Call the target to set up the argument values.
6402 SmallVector<SDValue, 8> InVals;
6403 SDValue NewRoot = TLI.LowerFormalArguments(DAG.getRoot(), F.getCallingConv(),
6407 // Verify that the target's LowerFormalArguments behaved as expected.
6408 assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other &&
6409 "LowerFormalArguments didn't return a valid chain!");
6410 assert(InVals.size() == Ins.size() &&
6411 "LowerFormalArguments didn't emit the correct number of values!");
6413 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
6414 assert(InVals[i].getNode() &&
6415 "LowerFormalArguments emitted a null value!");
6416 assert(Ins[i].VT == InVals[i].getValueType() &&
6417 "LowerFormalArguments emitted a value with the wrong type!");
6421 // Update the DAG with the new chain value resulting from argument lowering.
6422 DAG.setRoot(NewRoot);
6424 // Set up the argument values.
6427 if (!FLI.CanLowerReturn) {
6428 // Create a virtual register for the sret pointer, and put in a copy
6429 // from the sret argument into it.
6430 SmallVector<EVT, 1> ValueVTs;
6431 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
6432 EVT VT = ValueVTs[0];
6433 EVT RegVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6434 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6435 SDValue ArgValue = getCopyFromParts(DAG, dl, 0, &InVals[0], 1,
6436 RegVT, VT, AssertOp);
6438 MachineFunction& MF = SDB->DAG.getMachineFunction();
6439 MachineRegisterInfo& RegInfo = MF.getRegInfo();
6440 unsigned SRetReg = RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT));
6441 FLI.DemoteRegister = SRetReg;
6442 NewRoot = SDB->DAG.getCopyToReg(NewRoot, SDB->getCurDebugLoc(),
6444 DAG.setRoot(NewRoot);
6446 // i indexes lowered arguments. Bump it past the hidden sret argument.
6447 // Idx indexes LLVM arguments. Don't touch it.
6451 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
6453 SmallVector<SDValue, 4> ArgValues;
6454 SmallVector<EVT, 4> ValueVTs;
6455 ComputeValueVTs(TLI, I->getType(), ValueVTs);
6456 unsigned NumValues = ValueVTs.size();
6457 for (unsigned Value = 0; Value != NumValues; ++Value) {
6458 EVT VT = ValueVTs[Value];
6459 EVT PartVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6460 unsigned NumParts = TLI.getNumRegisters(*CurDAG->getContext(), VT);
6462 if (!I->use_empty()) {
6463 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6464 if (F.paramHasAttr(Idx, Attribute::SExt))
6465 AssertOp = ISD::AssertSext;
6466 else if (F.paramHasAttr(Idx, Attribute::ZExt))
6467 AssertOp = ISD::AssertZext;
6469 ArgValues.push_back(getCopyFromParts(DAG, dl, 0, &InVals[i],
6470 NumParts, PartVT, VT,
6477 if (!I->use_empty()) {
6478 SDValue Res = DAG.getMergeValues(&ArgValues[0], NumValues,
6479 SDB->getCurDebugLoc());
6480 SDB->setValue(I, Res);
6482 // If this argument is live outside of the entry block, insert a copy from
6483 // whereever we got it to the vreg that other BB's will reference it as.
6484 SDB->CopyToExportRegsIfNeeded(I);
6488 assert(i == InVals.size() && "Argument register count mismatch!");
6490 // Finally, if the target has anything special to do, allow it to do so.
6491 // FIXME: this should insert code into the DAG!
6492 EmitFunctionEntryCode(F, SDB->DAG.getMachineFunction());
6495 /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
6496 /// ensure constants are generated when needed. Remember the virtual registers
6497 /// that need to be added to the Machine PHI nodes as input. We cannot just
6498 /// directly add them, because expansion might result in multiple MBB's for one
6499 /// BB. As such, the start of the BB might correspond to a different MBB than
6503 SelectionDAGISel::HandlePHINodesInSuccessorBlocks(BasicBlock *LLVMBB) {
6504 TerminatorInst *TI = LLVMBB->getTerminator();
6506 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
6508 // Check successor nodes' PHI nodes that expect a constant to be available
6510 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
6511 BasicBlock *SuccBB = TI->getSuccessor(succ);
6512 if (!isa<PHINode>(SuccBB->begin())) continue;
6513 MachineBasicBlock *SuccMBB = FuncInfo->MBBMap[SuccBB];
6515 // If this terminator has multiple identical successors (common for
6516 // switches), only handle each succ once.
6517 if (!SuccsHandled.insert(SuccMBB)) continue;
6519 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
6522 // At this point we know that there is a 1-1 correspondence between LLVM PHI
6523 // nodes and Machine PHI nodes, but the incoming operands have not been
6525 for (BasicBlock::iterator I = SuccBB->begin();
6526 (PN = dyn_cast<PHINode>(I)); ++I) {
6527 // Ignore dead phi's.
6528 if (PN->use_empty()) continue;
6531 Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
6533 if (Constant *C = dyn_cast<Constant>(PHIOp)) {
6534 unsigned &RegOut = SDB->ConstantsOut[C];
6536 RegOut = FuncInfo->CreateRegForValue(C);
6537 SDB->CopyValueToVirtualRegister(C, RegOut);
6541 Reg = FuncInfo->ValueMap[PHIOp];
6543 assert(isa<AllocaInst>(PHIOp) &&
6544 FuncInfo->StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
6545 "Didn't codegen value into a register!??");
6546 Reg = FuncInfo->CreateRegForValue(PHIOp);
6547 SDB->CopyValueToVirtualRegister(PHIOp, Reg);
6551 // Remember that this register needs to added to the machine PHI node as
6552 // the input for this MBB.
6553 SmallVector<EVT, 4> ValueVTs;
6554 ComputeValueVTs(TLI, PN->getType(), ValueVTs);
6555 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
6556 EVT VT = ValueVTs[vti];
6557 unsigned NumRegisters = TLI.getNumRegisters(*CurDAG->getContext(), VT);
6558 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
6559 SDB->PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
6560 Reg += NumRegisters;
6564 SDB->ConstantsOut.clear();
6567 /// This is the Fast-ISel version of HandlePHINodesInSuccessorBlocks. It only
6568 /// supports legal types, and it emits MachineInstrs directly instead of
6569 /// creating SelectionDAG nodes.
6572 SelectionDAGISel::HandlePHINodesInSuccessorBlocksFast(BasicBlock *LLVMBB,
6574 TerminatorInst *TI = LLVMBB->getTerminator();
6576 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
6577 unsigned OrigNumPHINodesToUpdate = SDB->PHINodesToUpdate.size();
6579 // Check successor nodes' PHI nodes that expect a constant to be available
6581 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
6582 BasicBlock *SuccBB = TI->getSuccessor(succ);
6583 if (!isa<PHINode>(SuccBB->begin())) continue;
6584 MachineBasicBlock *SuccMBB = FuncInfo->MBBMap[SuccBB];
6586 // If this terminator has multiple identical successors (common for
6587 // switches), only handle each succ once.
6588 if (!SuccsHandled.insert(SuccMBB)) continue;
6590 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
6593 // At this point we know that there is a 1-1 correspondence between LLVM PHI
6594 // nodes and Machine PHI nodes, but the incoming operands have not been
6596 for (BasicBlock::iterator I = SuccBB->begin();
6597 (PN = dyn_cast<PHINode>(I)); ++I) {
6598 // Ignore dead phi's.
6599 if (PN->use_empty()) continue;
6601 // Only handle legal types. Two interesting things to note here. First,
6602 // by bailing out early, we may leave behind some dead instructions,
6603 // since SelectionDAG's HandlePHINodesInSuccessorBlocks will insert its
6604 // own moves. Second, this check is necessary becuase FastISel doesn't
6605 // use CreateRegForValue to create registers, so it always creates
6606 // exactly one register for each non-void instruction.
6607 EVT VT = TLI.getValueType(PN->getType(), /*AllowUnknown=*/true);
6608 if (VT == MVT::Other || !TLI.isTypeLegal(VT)) {
6611 VT = TLI.getTypeToTransformTo(*CurDAG->getContext(), VT);
6613 SDB->PHINodesToUpdate.resize(OrigNumPHINodesToUpdate);
6618 Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
6620 unsigned Reg = F->getRegForValue(PHIOp);
6622 SDB->PHINodesToUpdate.resize(OrigNumPHINodesToUpdate);
6625 SDB->PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg));