1 //===-- SelectionDAGBuilder.cpp - Selection-DAG building ------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements routines for translating from LLVM IR into SelectionDAG IR.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "isel"
15 #include "SDNodeDbgValue.h"
16 #include "SelectionDAGBuilder.h"
17 #include "llvm/ADT/BitVector.h"
18 #include "llvm/ADT/PostOrderIterator.h"
19 #include "llvm/ADT/SmallSet.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/Analysis/ConstantFolding.h"
22 #include "llvm/Constants.h"
23 #include "llvm/CallingConv.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/Function.h"
26 #include "llvm/GlobalVariable.h"
27 #include "llvm/InlineAsm.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/Intrinsics.h"
30 #include "llvm/IntrinsicInst.h"
31 #include "llvm/LLVMContext.h"
32 #include "llvm/Module.h"
33 #include "llvm/CodeGen/Analysis.h"
34 #include "llvm/CodeGen/FastISel.h"
35 #include "llvm/CodeGen/FunctionLoweringInfo.h"
36 #include "llvm/CodeGen/GCStrategy.h"
37 #include "llvm/CodeGen/GCMetadata.h"
38 #include "llvm/CodeGen/MachineFunction.h"
39 #include "llvm/CodeGen/MachineFrameInfo.h"
40 #include "llvm/CodeGen/MachineInstrBuilder.h"
41 #include "llvm/CodeGen/MachineJumpTableInfo.h"
42 #include "llvm/CodeGen/MachineModuleInfo.h"
43 #include "llvm/CodeGen/MachineRegisterInfo.h"
44 #include "llvm/CodeGen/SelectionDAG.h"
45 #include "llvm/Analysis/DebugInfo.h"
46 #include "llvm/Target/TargetData.h"
47 #include "llvm/Target/TargetFrameLowering.h"
48 #include "llvm/Target/TargetInstrInfo.h"
49 #include "llvm/Target/TargetIntrinsicInfo.h"
50 #include "llvm/Target/TargetLibraryInfo.h"
51 #include "llvm/Target/TargetLowering.h"
52 #include "llvm/Target/TargetOptions.h"
53 #include "llvm/Support/CommandLine.h"
54 #include "llvm/Support/Debug.h"
55 #include "llvm/Support/ErrorHandling.h"
56 #include "llvm/Support/MathExtras.h"
57 #include "llvm/Support/raw_ostream.h"
61 /// LimitFloatPrecision - Generate low-precision inline sequences for
62 /// some float libcalls (6, 8 or 12 bits).
63 static unsigned LimitFloatPrecision;
65 static cl::opt<unsigned, true>
66 LimitFPPrecision("limit-float-precision",
67 cl::desc("Generate low-precision inline sequences "
68 "for some float libcalls"),
69 cl::location(LimitFloatPrecision),
72 // Limit the width of DAG chains. This is important in general to prevent
73 // prevent DAG-based analysis from blowing up. For example, alias analysis and
74 // load clustering may not complete in reasonable time. It is difficult to
75 // recognize and avoid this situation within each individual analysis, and
76 // future analyses are likely to have the same behavior. Limiting DAG width is
77 // the safe approach, and will be especially important with global DAGs.
79 // MaxParallelChains default is arbitrarily high to avoid affecting
80 // optimization, but could be lowered to improve compile time. Any ld-ld-st-st
81 // sequence over this should have been converted to llvm.memcpy by the
82 // frontend. It easy to induce this behavior with .ll code such as:
83 // %buffer = alloca [4096 x i8]
84 // %data = load [4096 x i8]* %argPtr
85 // store [4096 x i8] %data, [4096 x i8]* %buffer
86 static const unsigned MaxParallelChains = 64;
88 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, DebugLoc DL,
89 const SDValue *Parts, unsigned NumParts,
90 EVT PartVT, EVT ValueVT);
92 /// getCopyFromParts - Create a value that contains the specified legal parts
93 /// combined into the value they represent. If the parts combine to a type
94 /// larger then ValueVT then AssertOp can be used to specify whether the extra
95 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
96 /// (ISD::AssertSext).
97 static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc DL,
99 unsigned NumParts, EVT PartVT, EVT ValueVT,
100 ISD::NodeType AssertOp = ISD::DELETED_NODE) {
101 if (ValueVT.isVector())
102 return getCopyFromPartsVector(DAG, DL, Parts, NumParts, PartVT, ValueVT);
104 assert(NumParts > 0 && "No parts to assemble!");
105 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
106 SDValue Val = Parts[0];
109 // Assemble the value from multiple parts.
110 if (ValueVT.isInteger()) {
111 unsigned PartBits = PartVT.getSizeInBits();
112 unsigned ValueBits = ValueVT.getSizeInBits();
114 // Assemble the power of 2 part.
115 unsigned RoundParts = NumParts & (NumParts - 1) ?
116 1 << Log2_32(NumParts) : NumParts;
117 unsigned RoundBits = PartBits * RoundParts;
118 EVT RoundVT = RoundBits == ValueBits ?
119 ValueVT : EVT::getIntegerVT(*DAG.getContext(), RoundBits);
122 EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2);
124 if (RoundParts > 2) {
125 Lo = getCopyFromParts(DAG, DL, Parts, RoundParts / 2,
127 Hi = getCopyFromParts(DAG, DL, Parts + RoundParts / 2,
128 RoundParts / 2, PartVT, HalfVT);
130 Lo = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[0]);
131 Hi = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[1]);
134 if (TLI.isBigEndian())
137 Val = DAG.getNode(ISD::BUILD_PAIR, DL, RoundVT, Lo, Hi);
139 if (RoundParts < NumParts) {
140 // Assemble the trailing non-power-of-2 part.
141 unsigned OddParts = NumParts - RoundParts;
142 EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits);
143 Hi = getCopyFromParts(DAG, DL,
144 Parts + RoundParts, OddParts, PartVT, OddVT);
146 // Combine the round and odd parts.
148 if (TLI.isBigEndian())
150 EVT TotalVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
151 Hi = DAG.getNode(ISD::ANY_EXTEND, DL, TotalVT, Hi);
152 Hi = DAG.getNode(ISD::SHL, DL, TotalVT, Hi,
153 DAG.getConstant(Lo.getValueType().getSizeInBits(),
154 TLI.getPointerTy()));
155 Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, TotalVT, Lo);
156 Val = DAG.getNode(ISD::OR, DL, TotalVT, Lo, Hi);
158 } else if (PartVT.isFloatingPoint()) {
159 // FP split into multiple FP parts (for ppcf128)
160 assert(ValueVT == EVT(MVT::ppcf128) && PartVT == EVT(MVT::f64) &&
163 Lo = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[0]);
164 Hi = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[1]);
165 if (TLI.isBigEndian())
167 Val = DAG.getNode(ISD::BUILD_PAIR, DL, ValueVT, Lo, Hi);
169 // FP split into integer parts (soft fp)
170 assert(ValueVT.isFloatingPoint() && PartVT.isInteger() &&
171 !PartVT.isVector() && "Unexpected split");
172 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits());
173 Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT);
177 // There is now one part, held in Val. Correct it to match ValueVT.
178 PartVT = Val.getValueType();
180 if (PartVT == ValueVT)
183 if (PartVT.isInteger() && ValueVT.isInteger()) {
184 if (ValueVT.bitsLT(PartVT)) {
185 // For a truncate, see if we have any information to
186 // indicate whether the truncated bits will always be
187 // zero or sign-extension.
188 if (AssertOp != ISD::DELETED_NODE)
189 Val = DAG.getNode(AssertOp, DL, PartVT, Val,
190 DAG.getValueType(ValueVT));
191 return DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
193 return DAG.getNode(ISD::ANY_EXTEND, DL, ValueVT, Val);
196 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
197 // FP_ROUND's are always exact here.
198 if (ValueVT.bitsLT(Val.getValueType()))
199 return DAG.getNode(ISD::FP_ROUND, DL, ValueVT, Val,
200 DAG.getTargetConstant(1, TLI.getPointerTy()));
202 return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val);
205 if (PartVT.getSizeInBits() == ValueVT.getSizeInBits())
206 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
208 llvm_unreachable("Unknown mismatch!");
211 /// getCopyFromParts - Create a value that contains the specified legal parts
212 /// combined into the value they represent. If the parts combine to a type
213 /// larger then ValueVT then AssertOp can be used to specify whether the extra
214 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
215 /// (ISD::AssertSext).
216 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, DebugLoc DL,
217 const SDValue *Parts, unsigned NumParts,
218 EVT PartVT, EVT ValueVT) {
219 assert(ValueVT.isVector() && "Not a vector value");
220 assert(NumParts > 0 && "No parts to assemble!");
221 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
222 SDValue Val = Parts[0];
224 // Handle a multi-element vector.
226 EVT IntermediateVT, RegisterVT;
227 unsigned NumIntermediates;
229 TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT,
230 NumIntermediates, RegisterVT);
231 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
232 NumParts = NumRegs; // Silence a compiler warning.
233 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
234 assert(RegisterVT == Parts[0].getValueType() &&
235 "Part type doesn't match part!");
237 // Assemble the parts into intermediate operands.
238 SmallVector<SDValue, 8> Ops(NumIntermediates);
239 if (NumIntermediates == NumParts) {
240 // If the register was not expanded, truncate or copy the value,
242 for (unsigned i = 0; i != NumParts; ++i)
243 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i], 1,
244 PartVT, IntermediateVT);
245 } else if (NumParts > 0) {
246 // If the intermediate type was expanded, build the intermediate
247 // operands from the parts.
248 assert(NumParts % NumIntermediates == 0 &&
249 "Must expand into a divisible number of parts!");
250 unsigned Factor = NumParts / NumIntermediates;
251 for (unsigned i = 0; i != NumIntermediates; ++i)
252 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor,
253 PartVT, IntermediateVT);
256 // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the
257 // intermediate operands.
258 Val = DAG.getNode(IntermediateVT.isVector() ?
259 ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR, DL,
260 ValueVT, &Ops[0], NumIntermediates);
263 // There is now one part, held in Val. Correct it to match ValueVT.
264 PartVT = Val.getValueType();
266 if (PartVT == ValueVT)
269 if (PartVT.isVector()) {
270 // If the element type of the source/dest vectors are the same, but the
271 // parts vector has more elements than the value vector, then we have a
272 // vector widening case (e.g. <2 x float> -> <4 x float>). Extract the
274 if (PartVT.getVectorElementType() == ValueVT.getVectorElementType()) {
275 assert(PartVT.getVectorNumElements() > ValueVT.getVectorNumElements() &&
276 "Cannot narrow, it would be a lossy transformation");
277 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val,
278 DAG.getIntPtrConstant(0));
281 // Vector/Vector bitcast.
282 if (ValueVT.getSizeInBits() == PartVT.getSizeInBits())
283 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
285 assert(PartVT.getVectorNumElements() == ValueVT.getVectorNumElements() &&
286 "Cannot handle this kind of promotion");
287 // Promoted vector extract
288 bool Smaller = ValueVT.bitsLE(PartVT);
289 return DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
294 // Trivial bitcast if the types are the same size and the destination
295 // vector type is legal.
296 if (PartVT.getSizeInBits() == ValueVT.getSizeInBits() &&
297 TLI.isTypeLegal(ValueVT))
298 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
300 // Handle cases such as i8 -> <1 x i1>
301 assert(ValueVT.getVectorNumElements() == 1 &&
302 "Only trivial scalar-to-vector conversions should get here!");
304 if (ValueVT.getVectorNumElements() == 1 &&
305 ValueVT.getVectorElementType() != PartVT) {
306 bool Smaller = ValueVT.bitsLE(PartVT);
307 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
308 DL, ValueVT.getScalarType(), Val);
311 return DAG.getNode(ISD::BUILD_VECTOR, DL, ValueVT, Val);
317 static void getCopyToPartsVector(SelectionDAG &DAG, DebugLoc dl,
318 SDValue Val, SDValue *Parts, unsigned NumParts,
321 /// getCopyToParts - Create a series of nodes that contain the specified value
322 /// split into legal parts. If the parts contain more bits than Val, then, for
323 /// integers, ExtendKind can be used to specify how to generate the extra bits.
324 static void getCopyToParts(SelectionDAG &DAG, DebugLoc DL,
325 SDValue Val, SDValue *Parts, unsigned NumParts,
327 ISD::NodeType ExtendKind = ISD::ANY_EXTEND) {
328 EVT ValueVT = Val.getValueType();
330 // Handle the vector case separately.
331 if (ValueVT.isVector())
332 return getCopyToPartsVector(DAG, DL, Val, Parts, NumParts, PartVT);
334 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
335 unsigned PartBits = PartVT.getSizeInBits();
336 unsigned OrigNumParts = NumParts;
337 assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!");
342 assert(!ValueVT.isVector() && "Vector case handled elsewhere");
343 if (PartVT == ValueVT) {
344 assert(NumParts == 1 && "No-op copy with multiple parts!");
349 if (NumParts * PartBits > ValueVT.getSizeInBits()) {
350 // If the parts cover more bits than the value has, promote the value.
351 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
352 assert(NumParts == 1 && "Do not know what to promote to!");
353 Val = DAG.getNode(ISD::FP_EXTEND, DL, PartVT, Val);
355 assert(PartVT.isInteger() && ValueVT.isInteger() &&
356 "Unknown mismatch!");
357 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
358 Val = DAG.getNode(ExtendKind, DL, ValueVT, Val);
360 } else if (PartBits == ValueVT.getSizeInBits()) {
361 // Different types of the same size.
362 assert(NumParts == 1 && PartVT != ValueVT);
363 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
364 } else if (NumParts * PartBits < ValueVT.getSizeInBits()) {
365 // If the parts cover less bits than value has, truncate the value.
366 assert(PartVT.isInteger() && ValueVT.isInteger() &&
367 "Unknown mismatch!");
368 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
369 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
372 // The value may have changed - recompute ValueVT.
373 ValueVT = Val.getValueType();
374 assert(NumParts * PartBits == ValueVT.getSizeInBits() &&
375 "Failed to tile the value with PartVT!");
378 assert(PartVT == ValueVT && "Type conversion failed!");
383 // Expand the value into multiple parts.
384 if (NumParts & (NumParts - 1)) {
385 // The number of parts is not a power of 2. Split off and copy the tail.
386 assert(PartVT.isInteger() && ValueVT.isInteger() &&
387 "Do not know what to expand to!");
388 unsigned RoundParts = 1 << Log2_32(NumParts);
389 unsigned RoundBits = RoundParts * PartBits;
390 unsigned OddParts = NumParts - RoundParts;
391 SDValue OddVal = DAG.getNode(ISD::SRL, DL, ValueVT, Val,
392 DAG.getIntPtrConstant(RoundBits));
393 getCopyToParts(DAG, DL, OddVal, Parts + RoundParts, OddParts, PartVT);
395 if (TLI.isBigEndian())
396 // The odd parts were reversed by getCopyToParts - unreverse them.
397 std::reverse(Parts + RoundParts, Parts + NumParts);
399 NumParts = RoundParts;
400 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
401 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
404 // The number of parts is a power of 2. Repeatedly bisect the value using
406 Parts[0] = DAG.getNode(ISD::BITCAST, DL,
407 EVT::getIntegerVT(*DAG.getContext(),
408 ValueVT.getSizeInBits()),
411 for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) {
412 for (unsigned i = 0; i < NumParts; i += StepSize) {
413 unsigned ThisBits = StepSize * PartBits / 2;
414 EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits);
415 SDValue &Part0 = Parts[i];
416 SDValue &Part1 = Parts[i+StepSize/2];
418 Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
419 ThisVT, Part0, DAG.getIntPtrConstant(1));
420 Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
421 ThisVT, Part0, DAG.getIntPtrConstant(0));
423 if (ThisBits == PartBits && ThisVT != PartVT) {
424 Part0 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part0);
425 Part1 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part1);
430 if (TLI.isBigEndian())
431 std::reverse(Parts, Parts + OrigNumParts);
435 /// getCopyToPartsVector - Create a series of nodes that contain the specified
436 /// value split into legal parts.
437 static void getCopyToPartsVector(SelectionDAG &DAG, DebugLoc DL,
438 SDValue Val, SDValue *Parts, unsigned NumParts,
440 EVT ValueVT = Val.getValueType();
441 assert(ValueVT.isVector() && "Not a vector");
442 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
445 if (PartVT == ValueVT) {
447 } else if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) {
448 // Bitconvert vector->vector case.
449 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
450 } else if (PartVT.isVector() &&
451 PartVT.getVectorElementType() == ValueVT.getVectorElementType() &&
452 PartVT.getVectorNumElements() > ValueVT.getVectorNumElements()) {
453 EVT ElementVT = PartVT.getVectorElementType();
454 // Vector widening case, e.g. <2 x float> -> <4 x float>. Shuffle in
456 SmallVector<SDValue, 16> Ops;
457 for (unsigned i = 0, e = ValueVT.getVectorNumElements(); i != e; ++i)
458 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
459 ElementVT, Val, DAG.getIntPtrConstant(i)));
461 for (unsigned i = ValueVT.getVectorNumElements(),
462 e = PartVT.getVectorNumElements(); i != e; ++i)
463 Ops.push_back(DAG.getUNDEF(ElementVT));
465 Val = DAG.getNode(ISD::BUILD_VECTOR, DL, PartVT, &Ops[0], Ops.size());
467 // FIXME: Use CONCAT for 2x -> 4x.
469 //SDValue UndefElts = DAG.getUNDEF(VectorTy);
470 //Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts);
471 } else if (PartVT.isVector() &&
472 PartVT.getVectorElementType().bitsGE(
473 ValueVT.getVectorElementType()) &&
474 PartVT.getVectorNumElements() == ValueVT.getVectorNumElements()) {
476 // Promoted vector extract
477 bool Smaller = PartVT.bitsLE(ValueVT);
478 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
481 // Vector -> scalar conversion.
482 assert(ValueVT.getVectorNumElements() == 1 &&
483 "Only trivial vector-to-scalar conversions should get here!");
484 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
485 PartVT, Val, DAG.getIntPtrConstant(0));
487 bool Smaller = ValueVT.bitsLE(PartVT);
488 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
496 // Handle a multi-element vector.
497 EVT IntermediateVT, RegisterVT;
498 unsigned NumIntermediates;
499 unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT,
501 NumIntermediates, RegisterVT);
502 unsigned NumElements = ValueVT.getVectorNumElements();
504 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
505 NumParts = NumRegs; // Silence a compiler warning.
506 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
508 // Split the vector into intermediate operands.
509 SmallVector<SDValue, 8> Ops(NumIntermediates);
510 for (unsigned i = 0; i != NumIntermediates; ++i) {
511 if (IntermediateVT.isVector())
512 Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL,
514 DAG.getIntPtrConstant(i * (NumElements / NumIntermediates)));
516 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
517 IntermediateVT, Val, DAG.getIntPtrConstant(i));
520 // Split the intermediate operands into legal parts.
521 if (NumParts == NumIntermediates) {
522 // If the register was not expanded, promote or copy the value,
524 for (unsigned i = 0; i != NumParts; ++i)
525 getCopyToParts(DAG, DL, Ops[i], &Parts[i], 1, PartVT);
526 } else if (NumParts > 0) {
527 // If the intermediate type was expanded, split each the value into
529 assert(NumParts % NumIntermediates == 0 &&
530 "Must expand into a divisible number of parts!");
531 unsigned Factor = NumParts / NumIntermediates;
532 for (unsigned i = 0; i != NumIntermediates; ++i)
533 getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT);
541 /// RegsForValue - This struct represents the registers (physical or virtual)
542 /// that a particular set of values is assigned, and the type information
543 /// about the value. The most common situation is to represent one value at a
544 /// time, but struct or array values are handled element-wise as multiple
545 /// values. The splitting of aggregates is performed recursively, so that we
546 /// never have aggregate-typed registers. The values at this point do not
547 /// necessarily have legal types, so each value may require one or more
548 /// registers of some legal type.
550 struct RegsForValue {
551 /// ValueVTs - The value types of the values, which may not be legal, and
552 /// may need be promoted or synthesized from one or more registers.
554 SmallVector<EVT, 4> ValueVTs;
556 /// RegVTs - The value types of the registers. This is the same size as
557 /// ValueVTs and it records, for each value, what the type of the assigned
558 /// register or registers are. (Individual values are never synthesized
559 /// from more than one type of register.)
561 /// With virtual registers, the contents of RegVTs is redundant with TLI's
562 /// getRegisterType member function, however when with physical registers
563 /// it is necessary to have a separate record of the types.
565 SmallVector<EVT, 4> RegVTs;
567 /// Regs - This list holds the registers assigned to the values.
568 /// Each legal or promoted value requires one register, and each
569 /// expanded value requires multiple registers.
571 SmallVector<unsigned, 4> Regs;
575 RegsForValue(const SmallVector<unsigned, 4> ®s,
576 EVT regvt, EVT valuevt)
577 : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {}
579 RegsForValue(LLVMContext &Context, const TargetLowering &tli,
580 unsigned Reg, Type *Ty) {
581 ComputeValueVTs(tli, Ty, ValueVTs);
583 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
584 EVT ValueVT = ValueVTs[Value];
585 unsigned NumRegs = tli.getNumRegisters(Context, ValueVT);
586 EVT RegisterVT = tli.getRegisterType(Context, ValueVT);
587 for (unsigned i = 0; i != NumRegs; ++i)
588 Regs.push_back(Reg + i);
589 RegVTs.push_back(RegisterVT);
594 /// areValueTypesLegal - Return true if types of all the values are legal.
595 bool areValueTypesLegal(const TargetLowering &TLI) {
596 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
597 EVT RegisterVT = RegVTs[Value];
598 if (!TLI.isTypeLegal(RegisterVT))
604 /// append - Add the specified values to this one.
605 void append(const RegsForValue &RHS) {
606 ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end());
607 RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end());
608 Regs.append(RHS.Regs.begin(), RHS.Regs.end());
611 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
612 /// this value and returns the result as a ValueVTs value. This uses
613 /// Chain/Flag as the input and updates them for the output Chain/Flag.
614 /// If the Flag pointer is NULL, no flag is used.
615 SDValue getCopyFromRegs(SelectionDAG &DAG, FunctionLoweringInfo &FuncInfo,
617 SDValue &Chain, SDValue *Flag) const;
619 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
620 /// specified value into the registers specified by this object. This uses
621 /// Chain/Flag as the input and updates them for the output Chain/Flag.
622 /// If the Flag pointer is NULL, no flag is used.
623 void getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
624 SDValue &Chain, SDValue *Flag) const;
626 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
627 /// operand list. This adds the code marker, matching input operand index
628 /// (if applicable), and includes the number of values added into it.
629 void AddInlineAsmOperands(unsigned Kind,
630 bool HasMatching, unsigned MatchingIdx,
632 std::vector<SDValue> &Ops) const;
636 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
637 /// this value and returns the result as a ValueVT value. This uses
638 /// Chain/Flag as the input and updates them for the output Chain/Flag.
639 /// If the Flag pointer is NULL, no flag is used.
640 SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
641 FunctionLoweringInfo &FuncInfo,
643 SDValue &Chain, SDValue *Flag) const {
644 // A Value with type {} or [0 x %t] needs no registers.
645 if (ValueVTs.empty())
648 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
650 // Assemble the legal parts into the final values.
651 SmallVector<SDValue, 4> Values(ValueVTs.size());
652 SmallVector<SDValue, 8> Parts;
653 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
654 // Copy the legal parts from the registers.
655 EVT ValueVT = ValueVTs[Value];
656 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
657 EVT RegisterVT = RegVTs[Value];
659 Parts.resize(NumRegs);
660 for (unsigned i = 0; i != NumRegs; ++i) {
663 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT);
665 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag);
666 *Flag = P.getValue(2);
669 Chain = P.getValue(1);
672 // If the source register was virtual and if we know something about it,
673 // add an assert node.
674 if (!TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) ||
675 !RegisterVT.isInteger() || RegisterVT.isVector())
678 const FunctionLoweringInfo::LiveOutInfo *LOI =
679 FuncInfo.GetLiveOutRegInfo(Regs[Part+i]);
683 unsigned RegSize = RegisterVT.getSizeInBits();
684 unsigned NumSignBits = LOI->NumSignBits;
685 unsigned NumZeroBits = LOI->KnownZero.countLeadingOnes();
687 // FIXME: We capture more information than the dag can represent. For
688 // now, just use the tightest assertzext/assertsext possible.
690 EVT FromVT(MVT::Other);
691 if (NumSignBits == RegSize)
692 isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1
693 else if (NumZeroBits >= RegSize-1)
694 isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1
695 else if (NumSignBits > RegSize-8)
696 isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8
697 else if (NumZeroBits >= RegSize-8)
698 isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8
699 else if (NumSignBits > RegSize-16)
700 isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16
701 else if (NumZeroBits >= RegSize-16)
702 isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16
703 else if (NumSignBits > RegSize-32)
704 isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32
705 else if (NumZeroBits >= RegSize-32)
706 isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32
710 // Add an assertion node.
711 assert(FromVT != MVT::Other);
712 Parts[i] = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl,
713 RegisterVT, P, DAG.getValueType(FromVT));
716 Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(),
717 NumRegs, RegisterVT, ValueVT);
722 return DAG.getNode(ISD::MERGE_VALUES, dl,
723 DAG.getVTList(&ValueVTs[0], ValueVTs.size()),
724 &Values[0], ValueVTs.size());
727 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
728 /// specified value into the registers specified by this object. This uses
729 /// Chain/Flag as the input and updates them for the output Chain/Flag.
730 /// If the Flag pointer is NULL, no flag is used.
731 void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
732 SDValue &Chain, SDValue *Flag) const {
733 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
735 // Get the list of the values's legal parts.
736 unsigned NumRegs = Regs.size();
737 SmallVector<SDValue, 8> Parts(NumRegs);
738 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
739 EVT ValueVT = ValueVTs[Value];
740 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
741 EVT RegisterVT = RegVTs[Value];
743 getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value),
744 &Parts[Part], NumParts, RegisterVT);
748 // Copy the parts into the registers.
749 SmallVector<SDValue, 8> Chains(NumRegs);
750 for (unsigned i = 0; i != NumRegs; ++i) {
753 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]);
755 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag);
756 *Flag = Part.getValue(1);
759 Chains[i] = Part.getValue(0);
762 if (NumRegs == 1 || Flag)
763 // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is
764 // flagged to it. That is the CopyToReg nodes and the user are considered
765 // a single scheduling unit. If we create a TokenFactor and return it as
766 // chain, then the TokenFactor is both a predecessor (operand) of the
767 // user as well as a successor (the TF operands are flagged to the user).
768 // c1, f1 = CopyToReg
769 // c2, f2 = CopyToReg
770 // c3 = TokenFactor c1, c2
773 Chain = Chains[NumRegs-1];
775 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0], NumRegs);
778 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
779 /// operand list. This adds the code marker and includes the number of
780 /// values added into it.
781 void RegsForValue::AddInlineAsmOperands(unsigned Code, bool HasMatching,
782 unsigned MatchingIdx,
784 std::vector<SDValue> &Ops) const {
785 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
787 unsigned Flag = InlineAsm::getFlagWord(Code, Regs.size());
789 Flag = InlineAsm::getFlagWordForMatchingOp(Flag, MatchingIdx);
790 else if (!Regs.empty() &&
791 TargetRegisterInfo::isVirtualRegister(Regs.front())) {
792 // Put the register class of the virtual registers in the flag word. That
793 // way, later passes can recompute register class constraints for inline
794 // assembly as well as normal instructions.
795 // Don't do this for tied operands that can use the regclass information
797 const MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
798 const TargetRegisterClass *RC = MRI.getRegClass(Regs.front());
799 Flag = InlineAsm::getFlagWordForRegClass(Flag, RC->getID());
802 SDValue Res = DAG.getTargetConstant(Flag, MVT::i32);
805 for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) {
806 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVTs[Value]);
807 EVT RegisterVT = RegVTs[Value];
808 for (unsigned i = 0; i != NumRegs; ++i) {
809 assert(Reg < Regs.size() && "Mismatch in # registers expected");
810 Ops.push_back(DAG.getRegister(Regs[Reg++], RegisterVT));
815 void SelectionDAGBuilder::init(GCFunctionInfo *gfi, AliasAnalysis &aa,
816 const TargetLibraryInfo *li) {
820 TD = DAG.getTarget().getTargetData();
821 LPadToCallSiteMap.clear();
824 /// clear - Clear out the current SelectionDAG and the associated
825 /// state and prepare this SelectionDAGBuilder object to be used
826 /// for a new block. This doesn't clear out information about
827 /// additional blocks that are needed to complete switch lowering
828 /// or PHI node updating; that information is cleared out as it is
830 void SelectionDAGBuilder::clear() {
832 UnusedArgNodeMap.clear();
833 PendingLoads.clear();
834 PendingExports.clear();
835 CurDebugLoc = DebugLoc();
839 /// clearDanglingDebugInfo - Clear the dangling debug information
840 /// map. This function is seperated from the clear so that debug
841 /// information that is dangling in a basic block can be properly
842 /// resolved in a different basic block. This allows the
843 /// SelectionDAG to resolve dangling debug information attached
845 void SelectionDAGBuilder::clearDanglingDebugInfo() {
846 DanglingDebugInfoMap.clear();
849 /// getRoot - Return the current virtual root of the Selection DAG,
850 /// flushing any PendingLoad items. This must be done before emitting
851 /// a store or any other node that may need to be ordered after any
852 /// prior load instructions.
854 SDValue SelectionDAGBuilder::getRoot() {
855 if (PendingLoads.empty())
856 return DAG.getRoot();
858 if (PendingLoads.size() == 1) {
859 SDValue Root = PendingLoads[0];
861 PendingLoads.clear();
865 // Otherwise, we have to make a token factor node.
866 SDValue Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
867 &PendingLoads[0], PendingLoads.size());
868 PendingLoads.clear();
873 /// getControlRoot - Similar to getRoot, but instead of flushing all the
874 /// PendingLoad items, flush all the PendingExports items. It is necessary
875 /// to do this before emitting a terminator instruction.
877 SDValue SelectionDAGBuilder::getControlRoot() {
878 SDValue Root = DAG.getRoot();
880 if (PendingExports.empty())
883 // Turn all of the CopyToReg chains into one factored node.
884 if (Root.getOpcode() != ISD::EntryToken) {
885 unsigned i = 0, e = PendingExports.size();
886 for (; i != e; ++i) {
887 assert(PendingExports[i].getNode()->getNumOperands() > 1);
888 if (PendingExports[i].getNode()->getOperand(0) == Root)
889 break; // Don't add the root if we already indirectly depend on it.
893 PendingExports.push_back(Root);
896 Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
898 PendingExports.size());
899 PendingExports.clear();
904 void SelectionDAGBuilder::AssignOrderingToNode(const SDNode *Node) {
905 if (DAG.GetOrdering(Node) != 0) return; // Already has ordering.
906 DAG.AssignOrdering(Node, SDNodeOrder);
908 for (unsigned I = 0, E = Node->getNumOperands(); I != E; ++I)
909 AssignOrderingToNode(Node->getOperand(I).getNode());
912 void SelectionDAGBuilder::visit(const Instruction &I) {
913 // Set up outgoing PHI node register values before emitting the terminator.
914 if (isa<TerminatorInst>(&I))
915 HandlePHINodesInSuccessorBlocks(I.getParent());
917 CurDebugLoc = I.getDebugLoc();
919 visit(I.getOpcode(), I);
921 if (!isa<TerminatorInst>(&I) && !HasTailCall)
922 CopyToExportRegsIfNeeded(&I);
924 CurDebugLoc = DebugLoc();
927 void SelectionDAGBuilder::visitPHI(const PHINode &) {
928 llvm_unreachable("SelectionDAGBuilder shouldn't visit PHI nodes!");
931 void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) {
932 // Note: this doesn't use InstVisitor, because it has to work with
933 // ConstantExpr's in addition to instructions.
935 default: llvm_unreachable("Unknown instruction type encountered!");
936 // Build the switch statement using the Instruction.def file.
937 #define HANDLE_INST(NUM, OPCODE, CLASS) \
938 case Instruction::OPCODE: visit##OPCODE((CLASS&)I); break;
939 #include "llvm/Instruction.def"
942 // Assign the ordering to the freshly created DAG nodes.
943 if (NodeMap.count(&I)) {
945 AssignOrderingToNode(getValue(&I).getNode());
949 // resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V,
950 // generate the debug data structures now that we've seen its definition.
951 void SelectionDAGBuilder::resolveDanglingDebugInfo(const Value *V,
953 DanglingDebugInfo &DDI = DanglingDebugInfoMap[V];
955 const DbgValueInst *DI = DDI.getDI();
956 DebugLoc dl = DDI.getdl();
957 unsigned DbgSDNodeOrder = DDI.getSDNodeOrder();
958 MDNode *Variable = DI->getVariable();
959 uint64_t Offset = DI->getOffset();
962 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, Val)) {
963 SDV = DAG.getDbgValue(Variable, Val.getNode(),
964 Val.getResNo(), Offset, dl, DbgSDNodeOrder);
965 DAG.AddDbgValue(SDV, Val.getNode(), false);
968 DEBUG(dbgs() << "Dropping debug info for " << DI);
969 DanglingDebugInfoMap[V] = DanglingDebugInfo();
973 /// getValue - Return an SDValue for the given Value.
974 SDValue SelectionDAGBuilder::getValue(const Value *V) {
975 // If we already have an SDValue for this value, use it. It's important
976 // to do this first, so that we don't create a CopyFromReg if we already
977 // have a regular SDValue.
978 SDValue &N = NodeMap[V];
979 if (N.getNode()) return N;
981 // If there's a virtual register allocated and initialized for this
983 DenseMap<const Value *, unsigned>::iterator It = FuncInfo.ValueMap.find(V);
984 if (It != FuncInfo.ValueMap.end()) {
985 unsigned InReg = It->second;
986 RegsForValue RFV(*DAG.getContext(), TLI, InReg, V->getType());
987 SDValue Chain = DAG.getEntryNode();
988 N = RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain, NULL);
989 resolveDanglingDebugInfo(V, N);
993 // Otherwise create a new SDValue and remember it.
994 SDValue Val = getValueImpl(V);
996 resolveDanglingDebugInfo(V, Val);
1000 /// getNonRegisterValue - Return an SDValue for the given Value, but
1001 /// don't look in FuncInfo.ValueMap for a virtual register.
1002 SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) {
1003 // If we already have an SDValue for this value, use it.
1004 SDValue &N = NodeMap[V];
1005 if (N.getNode()) return N;
1007 // Otherwise create a new SDValue and remember it.
1008 SDValue Val = getValueImpl(V);
1010 resolveDanglingDebugInfo(V, Val);
1014 /// getValueImpl - Helper function for getValue and getNonRegisterValue.
1015 /// Create an SDValue for the given value.
1016 SDValue SelectionDAGBuilder::getValueImpl(const Value *V) {
1017 if (const Constant *C = dyn_cast<Constant>(V)) {
1018 EVT VT = TLI.getValueType(V->getType(), true);
1020 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C))
1021 return DAG.getConstant(*CI, VT);
1023 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
1024 return DAG.getGlobalAddress(GV, getCurDebugLoc(), VT);
1026 if (isa<ConstantPointerNull>(C))
1027 return DAG.getConstant(0, TLI.getPointerTy());
1029 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
1030 return DAG.getConstantFP(*CFP, VT);
1032 if (isa<UndefValue>(C) && !V->getType()->isAggregateType())
1033 return DAG.getUNDEF(VT);
1035 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
1036 visit(CE->getOpcode(), *CE);
1037 SDValue N1 = NodeMap[V];
1038 assert(N1.getNode() && "visit didn't populate the NodeMap!");
1042 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) {
1043 SmallVector<SDValue, 4> Constants;
1044 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end();
1046 SDNode *Val = getValue(*OI).getNode();
1047 // If the operand is an empty aggregate, there are no values.
1049 // Add each leaf value from the operand to the Constants list
1050 // to form a flattened list of all the values.
1051 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
1052 Constants.push_back(SDValue(Val, i));
1055 return DAG.getMergeValues(&Constants[0], Constants.size(),
1059 if (const ConstantDataSequential *CDS =
1060 dyn_cast<ConstantDataSequential>(C)) {
1061 SmallVector<SDValue, 4> Ops;
1062 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1063 SDNode *Val = getValue(CDS->getElementAsConstant(i)).getNode();
1064 // Add each leaf value from the operand to the Constants list
1065 // to form a flattened list of all the values.
1066 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
1067 Ops.push_back(SDValue(Val, i));
1070 if (isa<ArrayType>(CDS->getType()))
1071 return DAG.getMergeValues(&Ops[0], Ops.size(), getCurDebugLoc());
1072 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
1073 VT, &Ops[0], Ops.size());
1076 if (C->getType()->isStructTy() || C->getType()->isArrayTy()) {
1077 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
1078 "Unknown struct or array constant!");
1080 SmallVector<EVT, 4> ValueVTs;
1081 ComputeValueVTs(TLI, C->getType(), ValueVTs);
1082 unsigned NumElts = ValueVTs.size();
1084 return SDValue(); // empty struct
1085 SmallVector<SDValue, 4> Constants(NumElts);
1086 for (unsigned i = 0; i != NumElts; ++i) {
1087 EVT EltVT = ValueVTs[i];
1088 if (isa<UndefValue>(C))
1089 Constants[i] = DAG.getUNDEF(EltVT);
1090 else if (EltVT.isFloatingPoint())
1091 Constants[i] = DAG.getConstantFP(0, EltVT);
1093 Constants[i] = DAG.getConstant(0, EltVT);
1096 return DAG.getMergeValues(&Constants[0], NumElts,
1100 if (const BlockAddress *BA = dyn_cast<BlockAddress>(C))
1101 return DAG.getBlockAddress(BA, VT);
1103 VectorType *VecTy = cast<VectorType>(V->getType());
1104 unsigned NumElements = VecTy->getNumElements();
1106 // Now that we know the number and type of the elements, get that number of
1107 // elements into the Ops array based on what kind of constant it is.
1108 SmallVector<SDValue, 16> Ops;
1109 if (const ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
1110 for (unsigned i = 0; i != NumElements; ++i)
1111 Ops.push_back(getValue(CV->getOperand(i)));
1113 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!");
1114 EVT EltVT = TLI.getValueType(VecTy->getElementType());
1117 if (EltVT.isFloatingPoint())
1118 Op = DAG.getConstantFP(0, EltVT);
1120 Op = DAG.getConstant(0, EltVT);
1121 Ops.assign(NumElements, Op);
1124 // Create a BUILD_VECTOR node.
1125 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
1126 VT, &Ops[0], Ops.size());
1129 // If this is a static alloca, generate it as the frameindex instead of
1131 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
1132 DenseMap<const AllocaInst*, int>::iterator SI =
1133 FuncInfo.StaticAllocaMap.find(AI);
1134 if (SI != FuncInfo.StaticAllocaMap.end())
1135 return DAG.getFrameIndex(SI->second, TLI.getPointerTy());
1138 // If this is an instruction which fast-isel has deferred, select it now.
1139 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
1140 unsigned InReg = FuncInfo.InitializeRegForValue(Inst);
1141 RegsForValue RFV(*DAG.getContext(), TLI, InReg, Inst->getType());
1142 SDValue Chain = DAG.getEntryNode();
1143 return RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain, NULL);
1146 llvm_unreachable("Can't get register for value!");
1149 void SelectionDAGBuilder::visitRet(const ReturnInst &I) {
1150 SDValue Chain = getControlRoot();
1151 SmallVector<ISD::OutputArg, 8> Outs;
1152 SmallVector<SDValue, 8> OutVals;
1154 if (!FuncInfo.CanLowerReturn) {
1155 unsigned DemoteReg = FuncInfo.DemoteRegister;
1156 const Function *F = I.getParent()->getParent();
1158 // Emit a store of the return value through the virtual register.
1159 // Leave Outs empty so that LowerReturn won't try to load return
1160 // registers the usual way.
1161 SmallVector<EVT, 1> PtrValueVTs;
1162 ComputeValueVTs(TLI, PointerType::getUnqual(F->getReturnType()),
1165 SDValue RetPtr = DAG.getRegister(DemoteReg, PtrValueVTs[0]);
1166 SDValue RetOp = getValue(I.getOperand(0));
1168 SmallVector<EVT, 4> ValueVTs;
1169 SmallVector<uint64_t, 4> Offsets;
1170 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs, &Offsets);
1171 unsigned NumValues = ValueVTs.size();
1173 SmallVector<SDValue, 4> Chains(NumValues);
1174 for (unsigned i = 0; i != NumValues; ++i) {
1175 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(),
1176 RetPtr.getValueType(), RetPtr,
1177 DAG.getIntPtrConstant(Offsets[i]));
1179 DAG.getStore(Chain, getCurDebugLoc(),
1180 SDValue(RetOp.getNode(), RetOp.getResNo() + i),
1181 // FIXME: better loc info would be nice.
1182 Add, MachinePointerInfo(), false, false, 0);
1185 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
1186 MVT::Other, &Chains[0], NumValues);
1187 } else if (I.getNumOperands() != 0) {
1188 SmallVector<EVT, 4> ValueVTs;
1189 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs);
1190 unsigned NumValues = ValueVTs.size();
1192 SDValue RetOp = getValue(I.getOperand(0));
1193 for (unsigned j = 0, f = NumValues; j != f; ++j) {
1194 EVT VT = ValueVTs[j];
1196 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1198 const Function *F = I.getParent()->getParent();
1199 if (F->paramHasAttr(0, Attribute::SExt))
1200 ExtendKind = ISD::SIGN_EXTEND;
1201 else if (F->paramHasAttr(0, Attribute::ZExt))
1202 ExtendKind = ISD::ZERO_EXTEND;
1204 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger())
1205 VT = TLI.getTypeForExtArgOrReturn(*DAG.getContext(), VT, ExtendKind);
1207 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), VT);
1208 EVT PartVT = TLI.getRegisterType(*DAG.getContext(), VT);
1209 SmallVector<SDValue, 4> Parts(NumParts);
1210 getCopyToParts(DAG, getCurDebugLoc(),
1211 SDValue(RetOp.getNode(), RetOp.getResNo() + j),
1212 &Parts[0], NumParts, PartVT, ExtendKind);
1214 // 'inreg' on function refers to return value
1215 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1216 if (F->paramHasAttr(0, Attribute::InReg))
1219 // Propagate extension type if any
1220 if (ExtendKind == ISD::SIGN_EXTEND)
1222 else if (ExtendKind == ISD::ZERO_EXTEND)
1225 for (unsigned i = 0; i < NumParts; ++i) {
1226 Outs.push_back(ISD::OutputArg(Flags, Parts[i].getValueType(),
1228 OutVals.push_back(Parts[i]);
1234 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1235 CallingConv::ID CallConv =
1236 DAG.getMachineFunction().getFunction()->getCallingConv();
1237 Chain = TLI.LowerReturn(Chain, CallConv, isVarArg,
1238 Outs, OutVals, getCurDebugLoc(), DAG);
1240 // Verify that the target's LowerReturn behaved as expected.
1241 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
1242 "LowerReturn didn't return a valid chain!");
1244 // Update the DAG with the new chain value resulting from return lowering.
1248 /// CopyToExportRegsIfNeeded - If the given value has virtual registers
1249 /// created for it, emit nodes to copy the value into the virtual
1251 void SelectionDAGBuilder::CopyToExportRegsIfNeeded(const Value *V) {
1253 if (V->getType()->isEmptyTy())
1256 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
1257 if (VMI != FuncInfo.ValueMap.end()) {
1258 assert(!V->use_empty() && "Unused value assigned virtual registers!");
1259 CopyValueToVirtualRegister(V, VMI->second);
1263 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
1264 /// the current basic block, add it to ValueMap now so that we'll get a
1266 void SelectionDAGBuilder::ExportFromCurrentBlock(const Value *V) {
1267 // No need to export constants.
1268 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
1270 // Already exported?
1271 if (FuncInfo.isExportedInst(V)) return;
1273 unsigned Reg = FuncInfo.InitializeRegForValue(V);
1274 CopyValueToVirtualRegister(V, Reg);
1277 bool SelectionDAGBuilder::isExportableFromCurrentBlock(const Value *V,
1278 const BasicBlock *FromBB) {
1279 // The operands of the setcc have to be in this block. We don't know
1280 // how to export them from some other block.
1281 if (const Instruction *VI = dyn_cast<Instruction>(V)) {
1282 // Can export from current BB.
1283 if (VI->getParent() == FromBB)
1286 // Is already exported, noop.
1287 return FuncInfo.isExportedInst(V);
1290 // If this is an argument, we can export it if the BB is the entry block or
1291 // if it is already exported.
1292 if (isa<Argument>(V)) {
1293 if (FromBB == &FromBB->getParent()->getEntryBlock())
1296 // Otherwise, can only export this if it is already exported.
1297 return FuncInfo.isExportedInst(V);
1300 // Otherwise, constants can always be exported.
1304 /// Return branch probability calculated by BranchProbabilityInfo for IR blocks.
1305 uint32_t SelectionDAGBuilder::getEdgeWeight(const MachineBasicBlock *Src,
1306 const MachineBasicBlock *Dst) const {
1307 BranchProbabilityInfo *BPI = FuncInfo.BPI;
1310 const BasicBlock *SrcBB = Src->getBasicBlock();
1311 const BasicBlock *DstBB = Dst->getBasicBlock();
1312 return BPI->getEdgeWeight(SrcBB, DstBB);
1315 void SelectionDAGBuilder::
1316 addSuccessorWithWeight(MachineBasicBlock *Src, MachineBasicBlock *Dst,
1317 uint32_t Weight /* = 0 */) {
1319 Weight = getEdgeWeight(Src, Dst);
1320 Src->addSuccessor(Dst, Weight);
1324 static bool InBlock(const Value *V, const BasicBlock *BB) {
1325 if (const Instruction *I = dyn_cast<Instruction>(V))
1326 return I->getParent() == BB;
1330 /// EmitBranchForMergedCondition - Helper method for FindMergedConditions.
1331 /// This function emits a branch and is used at the leaves of an OR or an
1332 /// AND operator tree.
1335 SelectionDAGBuilder::EmitBranchForMergedCondition(const Value *Cond,
1336 MachineBasicBlock *TBB,
1337 MachineBasicBlock *FBB,
1338 MachineBasicBlock *CurBB,
1339 MachineBasicBlock *SwitchBB) {
1340 const BasicBlock *BB = CurBB->getBasicBlock();
1342 // If the leaf of the tree is a comparison, merge the condition into
1344 if (const CmpInst *BOp = dyn_cast<CmpInst>(Cond)) {
1345 // The operands of the cmp have to be in this block. We don't know
1346 // how to export them from some other block. If this is the first block
1347 // of the sequence, no exporting is needed.
1348 if (CurBB == SwitchBB ||
1349 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
1350 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) {
1351 ISD::CondCode Condition;
1352 if (const ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
1353 Condition = getICmpCondCode(IC->getPredicate());
1354 } else if (const FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) {
1355 Condition = getFCmpCondCode(FC->getPredicate());
1356 if (TM.Options.NoNaNsFPMath)
1357 Condition = getFCmpCodeWithoutNaN(Condition);
1359 Condition = ISD::SETEQ; // silence warning.
1360 llvm_unreachable("Unknown compare instruction");
1363 CaseBlock CB(Condition, BOp->getOperand(0),
1364 BOp->getOperand(1), NULL, TBB, FBB, CurBB);
1365 SwitchCases.push_back(CB);
1370 // Create a CaseBlock record representing this branch.
1371 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()),
1372 NULL, TBB, FBB, CurBB);
1373 SwitchCases.push_back(CB);
1376 /// FindMergedConditions - If Cond is an expression like
1377 void SelectionDAGBuilder::FindMergedConditions(const Value *Cond,
1378 MachineBasicBlock *TBB,
1379 MachineBasicBlock *FBB,
1380 MachineBasicBlock *CurBB,
1381 MachineBasicBlock *SwitchBB,
1383 // If this node is not part of the or/and tree, emit it as a branch.
1384 const Instruction *BOp = dyn_cast<Instruction>(Cond);
1385 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
1386 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
1387 BOp->getParent() != CurBB->getBasicBlock() ||
1388 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
1389 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
1390 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB);
1394 // Create TmpBB after CurBB.
1395 MachineFunction::iterator BBI = CurBB;
1396 MachineFunction &MF = DAG.getMachineFunction();
1397 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock());
1398 CurBB->getParent()->insert(++BBI, TmpBB);
1400 if (Opc == Instruction::Or) {
1401 // Codegen X | Y as:
1409 // Emit the LHS condition.
1410 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, SwitchBB, Opc);
1412 // Emit the RHS condition into TmpBB.
1413 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
1415 assert(Opc == Instruction::And && "Unknown merge op!");
1416 // Codegen X & Y as:
1423 // This requires creation of TmpBB after CurBB.
1425 // Emit the LHS condition.
1426 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, SwitchBB, Opc);
1428 // Emit the RHS condition into TmpBB.
1429 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
1433 /// If the set of cases should be emitted as a series of branches, return true.
1434 /// If we should emit this as a bunch of and/or'd together conditions, return
1437 SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases){
1438 if (Cases.size() != 2) return true;
1440 // If this is two comparisons of the same values or'd or and'd together, they
1441 // will get folded into a single comparison, so don't emit two blocks.
1442 if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
1443 Cases[0].CmpRHS == Cases[1].CmpRHS) ||
1444 (Cases[0].CmpRHS == Cases[1].CmpLHS &&
1445 Cases[0].CmpLHS == Cases[1].CmpRHS)) {
1449 // Handle: (X != null) | (Y != null) --> (X|Y) != 0
1450 // Handle: (X == null) & (Y == null) --> (X|Y) == 0
1451 if (Cases[0].CmpRHS == Cases[1].CmpRHS &&
1452 Cases[0].CC == Cases[1].CC &&
1453 isa<Constant>(Cases[0].CmpRHS) &&
1454 cast<Constant>(Cases[0].CmpRHS)->isNullValue()) {
1455 if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB)
1457 if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB)
1464 void SelectionDAGBuilder::visitBr(const BranchInst &I) {
1465 MachineBasicBlock *BrMBB = FuncInfo.MBB;
1467 // Update machine-CFG edges.
1468 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
1470 // Figure out which block is immediately after the current one.
1471 MachineBasicBlock *NextBlock = 0;
1472 MachineFunction::iterator BBI = BrMBB;
1473 if (++BBI != FuncInfo.MF->end())
1476 if (I.isUnconditional()) {
1477 // Update machine-CFG edges.
1478 BrMBB->addSuccessor(Succ0MBB);
1480 // If this is not a fall-through branch, emit the branch.
1481 if (Succ0MBB != NextBlock)
1482 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
1483 MVT::Other, getControlRoot(),
1484 DAG.getBasicBlock(Succ0MBB)));
1489 // If this condition is one of the special cases we handle, do special stuff
1491 const Value *CondVal = I.getCondition();
1492 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
1494 // If this is a series of conditions that are or'd or and'd together, emit
1495 // this as a sequence of branches instead of setcc's with and/or operations.
1496 // As long as jumps are not expensive, this should improve performance.
1497 // For example, instead of something like:
1510 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
1511 if (!TLI.isJumpExpensive() &&
1513 (BOp->getOpcode() == Instruction::And ||
1514 BOp->getOpcode() == Instruction::Or)) {
1515 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB,
1517 // If the compares in later blocks need to use values not currently
1518 // exported from this block, export them now. This block should always
1519 // be the first entry.
1520 assert(SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!");
1522 // Allow some cases to be rejected.
1523 if (ShouldEmitAsBranches(SwitchCases)) {
1524 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1525 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1526 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1529 // Emit the branch for this block.
1530 visitSwitchCase(SwitchCases[0], BrMBB);
1531 SwitchCases.erase(SwitchCases.begin());
1535 // Okay, we decided not to do this, remove any inserted MBB's and clear
1537 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
1538 FuncInfo.MF->erase(SwitchCases[i].ThisBB);
1540 SwitchCases.clear();
1544 // Create a CaseBlock record representing this branch.
1545 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()),
1546 NULL, Succ0MBB, Succ1MBB, BrMBB);
1548 // Use visitSwitchCase to actually insert the fast branch sequence for this
1550 visitSwitchCase(CB, BrMBB);
1553 /// visitSwitchCase - Emits the necessary code to represent a single node in
1554 /// the binary search tree resulting from lowering a switch instruction.
1555 void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB,
1556 MachineBasicBlock *SwitchBB) {
1558 SDValue CondLHS = getValue(CB.CmpLHS);
1559 DebugLoc dl = getCurDebugLoc();
1561 // Build the setcc now.
1562 if (CB.CmpMHS == NULL) {
1563 // Fold "(X == true)" to X and "(X == false)" to !X to
1564 // handle common cases produced by branch lowering.
1565 if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) &&
1566 CB.CC == ISD::SETEQ)
1568 else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) &&
1569 CB.CC == ISD::SETEQ) {
1570 SDValue True = DAG.getConstant(1, CondLHS.getValueType());
1571 Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True);
1573 Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1575 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now");
1577 const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue();
1578 const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue();
1580 SDValue CmpOp = getValue(CB.CmpMHS);
1581 EVT VT = CmpOp.getValueType();
1583 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) {
1584 Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, VT),
1587 SDValue SUB = DAG.getNode(ISD::SUB, dl,
1588 VT, CmpOp, DAG.getConstant(Low, VT));
1589 Cond = DAG.getSetCC(dl, MVT::i1, SUB,
1590 DAG.getConstant(High-Low, VT), ISD::SETULE);
1594 // Update successor info
1595 addSuccessorWithWeight(SwitchBB, CB.TrueBB, CB.TrueWeight);
1596 addSuccessorWithWeight(SwitchBB, CB.FalseBB, CB.FalseWeight);
1598 // Set NextBlock to be the MBB immediately after the current one, if any.
1599 // This is used to avoid emitting unnecessary branches to the next block.
1600 MachineBasicBlock *NextBlock = 0;
1601 MachineFunction::iterator BBI = SwitchBB;
1602 if (++BBI != FuncInfo.MF->end())
1605 // If the lhs block is the next block, invert the condition so that we can
1606 // fall through to the lhs instead of the rhs block.
1607 if (CB.TrueBB == NextBlock) {
1608 std::swap(CB.TrueBB, CB.FalseBB);
1609 SDValue True = DAG.getConstant(1, Cond.getValueType());
1610 Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True);
1613 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1614 MVT::Other, getControlRoot(), Cond,
1615 DAG.getBasicBlock(CB.TrueBB));
1617 // Insert the false branch. Do this even if it's a fall through branch,
1618 // this makes it easier to do DAG optimizations which require inverting
1619 // the branch condition.
1620 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1621 DAG.getBasicBlock(CB.FalseBB));
1623 DAG.setRoot(BrCond);
1626 /// visitJumpTable - Emit JumpTable node in the current MBB
1627 void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) {
1628 // Emit the code for the jump table
1629 assert(JT.Reg != -1U && "Should lower JT Header first!");
1630 EVT PTy = TLI.getPointerTy();
1631 SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(),
1633 SDValue Table = DAG.getJumpTable(JT.JTI, PTy);
1634 SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurDebugLoc(),
1635 MVT::Other, Index.getValue(1),
1637 DAG.setRoot(BrJumpTable);
1640 /// visitJumpTableHeader - This function emits necessary code to produce index
1641 /// in the JumpTable from switch case.
1642 void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT,
1643 JumpTableHeader &JTH,
1644 MachineBasicBlock *SwitchBB) {
1645 // Subtract the lowest switch case value from the value being switched on and
1646 // conditional branch to default mbb if the result is greater than the
1647 // difference between smallest and largest cases.
1648 SDValue SwitchOp = getValue(JTH.SValue);
1649 EVT VT = SwitchOp.getValueType();
1650 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1651 DAG.getConstant(JTH.First, VT));
1653 // The SDNode we just created, which holds the value being switched on minus
1654 // the smallest case value, needs to be copied to a virtual register so it
1655 // can be used as an index into the jump table in a subsequent basic block.
1656 // This value may be smaller or larger than the target's pointer type, and
1657 // therefore require extension or truncating.
1658 SwitchOp = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(), TLI.getPointerTy());
1660 unsigned JumpTableReg = FuncInfo.CreateReg(TLI.getPointerTy());
1661 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1662 JumpTableReg, SwitchOp);
1663 JT.Reg = JumpTableReg;
1665 // Emit the range check for the jump table, and branch to the default block
1666 // for the switch statement if the value being switched on exceeds the largest
1667 // case in the switch.
1668 SDValue CMP = DAG.getSetCC(getCurDebugLoc(),
1669 TLI.getSetCCResultType(Sub.getValueType()), Sub,
1670 DAG.getConstant(JTH.Last-JTH.First,VT),
1673 // Set NextBlock to be the MBB immediately after the current one, if any.
1674 // This is used to avoid emitting unnecessary branches to the next block.
1675 MachineBasicBlock *NextBlock = 0;
1676 MachineFunction::iterator BBI = SwitchBB;
1678 if (++BBI != FuncInfo.MF->end())
1681 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1682 MVT::Other, CopyTo, CMP,
1683 DAG.getBasicBlock(JT.Default));
1685 if (JT.MBB != NextBlock)
1686 BrCond = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrCond,
1687 DAG.getBasicBlock(JT.MBB));
1689 DAG.setRoot(BrCond);
1692 /// visitBitTestHeader - This function emits necessary code to produce value
1693 /// suitable for "bit tests"
1694 void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B,
1695 MachineBasicBlock *SwitchBB) {
1696 // Subtract the minimum value
1697 SDValue SwitchOp = getValue(B.SValue);
1698 EVT VT = SwitchOp.getValueType();
1699 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1700 DAG.getConstant(B.First, VT));
1703 SDValue RangeCmp = DAG.getSetCC(getCurDebugLoc(),
1704 TLI.getSetCCResultType(Sub.getValueType()),
1705 Sub, DAG.getConstant(B.Range, VT),
1708 // Determine the type of the test operands.
1709 bool UsePtrType = false;
1710 if (!TLI.isTypeLegal(VT))
1713 for (unsigned i = 0, e = B.Cases.size(); i != e; ++i)
1714 if (!isUIntN(VT.getSizeInBits(), B.Cases[i].Mask)) {
1715 // Switch table case range are encoded into series of masks.
1716 // Just use pointer type, it's guaranteed to fit.
1722 VT = TLI.getPointerTy();
1723 Sub = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(), VT);
1727 B.Reg = FuncInfo.CreateReg(VT);
1728 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1731 // Set NextBlock to be the MBB immediately after the current one, if any.
1732 // This is used to avoid emitting unnecessary branches to the next block.
1733 MachineBasicBlock *NextBlock = 0;
1734 MachineFunction::iterator BBI = SwitchBB;
1735 if (++BBI != FuncInfo.MF->end())
1738 MachineBasicBlock* MBB = B.Cases[0].ThisBB;
1740 addSuccessorWithWeight(SwitchBB, B.Default);
1741 addSuccessorWithWeight(SwitchBB, MBB);
1743 SDValue BrRange = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1744 MVT::Other, CopyTo, RangeCmp,
1745 DAG.getBasicBlock(B.Default));
1747 if (MBB != NextBlock)
1748 BrRange = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, CopyTo,
1749 DAG.getBasicBlock(MBB));
1751 DAG.setRoot(BrRange);
1754 /// visitBitTestCase - this function produces one "bit test"
1755 void SelectionDAGBuilder::visitBitTestCase(BitTestBlock &BB,
1756 MachineBasicBlock* NextMBB,
1759 MachineBasicBlock *SwitchBB) {
1761 SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(),
1764 unsigned PopCount = CountPopulation_64(B.Mask);
1765 if (PopCount == 1) {
1766 // Testing for a single bit; just compare the shift count with what it
1767 // would need to be to shift a 1 bit in that position.
1768 Cmp = DAG.getSetCC(getCurDebugLoc(),
1769 TLI.getSetCCResultType(VT),
1771 DAG.getConstant(CountTrailingZeros_64(B.Mask), VT),
1773 } else if (PopCount == BB.Range) {
1774 // There is only one zero bit in the range, test for it directly.
1775 Cmp = DAG.getSetCC(getCurDebugLoc(),
1776 TLI.getSetCCResultType(VT),
1778 DAG.getConstant(CountTrailingOnes_64(B.Mask), VT),
1781 // Make desired shift
1782 SDValue SwitchVal = DAG.getNode(ISD::SHL, getCurDebugLoc(), VT,
1783 DAG.getConstant(1, VT), ShiftOp);
1785 // Emit bit tests and jumps
1786 SDValue AndOp = DAG.getNode(ISD::AND, getCurDebugLoc(),
1787 VT, SwitchVal, DAG.getConstant(B.Mask, VT));
1788 Cmp = DAG.getSetCC(getCurDebugLoc(),
1789 TLI.getSetCCResultType(VT),
1790 AndOp, DAG.getConstant(0, VT),
1794 addSuccessorWithWeight(SwitchBB, B.TargetBB);
1795 addSuccessorWithWeight(SwitchBB, NextMBB);
1797 SDValue BrAnd = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1798 MVT::Other, getControlRoot(),
1799 Cmp, DAG.getBasicBlock(B.TargetBB));
1801 // Set NextBlock to be the MBB immediately after the current one, if any.
1802 // This is used to avoid emitting unnecessary branches to the next block.
1803 MachineBasicBlock *NextBlock = 0;
1804 MachineFunction::iterator BBI = SwitchBB;
1805 if (++BBI != FuncInfo.MF->end())
1808 if (NextMBB != NextBlock)
1809 BrAnd = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrAnd,
1810 DAG.getBasicBlock(NextMBB));
1815 void SelectionDAGBuilder::visitInvoke(const InvokeInst &I) {
1816 MachineBasicBlock *InvokeMBB = FuncInfo.MBB;
1818 // Retrieve successors.
1819 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
1820 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)];
1822 const Value *Callee(I.getCalledValue());
1823 if (isa<InlineAsm>(Callee))
1826 LowerCallTo(&I, getValue(Callee), false, LandingPad);
1828 // If the value of the invoke is used outside of its defining block, make it
1829 // available as a virtual register.
1830 CopyToExportRegsIfNeeded(&I);
1832 // Update successor info
1833 addSuccessorWithWeight(InvokeMBB, Return);
1834 addSuccessorWithWeight(InvokeMBB, LandingPad);
1836 // Drop into normal successor.
1837 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
1838 MVT::Other, getControlRoot(),
1839 DAG.getBasicBlock(Return)));
1842 void SelectionDAGBuilder::visitResume(const ResumeInst &RI) {
1843 llvm_unreachable("SelectionDAGBuilder shouldn't visit resume instructions!");
1846 void SelectionDAGBuilder::visitLandingPad(const LandingPadInst &LP) {
1847 assert(FuncInfo.MBB->isLandingPad() &&
1848 "Call to landingpad not in landing pad!");
1850 MachineBasicBlock *MBB = FuncInfo.MBB;
1851 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
1852 AddLandingPadInfo(LP, MMI, MBB);
1854 // If there aren't registers to copy the values into (e.g., during SjLj
1855 // exceptions), then don't bother to create these DAG nodes.
1856 if (TLI.getExceptionPointerRegister() == 0 &&
1857 TLI.getExceptionSelectorRegister() == 0)
1860 SmallVector<EVT, 2> ValueVTs;
1861 ComputeValueVTs(TLI, LP.getType(), ValueVTs);
1863 // Insert the EXCEPTIONADDR instruction.
1864 assert(FuncInfo.MBB->isLandingPad() &&
1865 "Call to eh.exception not in landing pad!");
1866 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
1868 Ops[0] = DAG.getRoot();
1869 SDValue Op1 = DAG.getNode(ISD::EXCEPTIONADDR, getCurDebugLoc(), VTs, Ops, 1);
1870 SDValue Chain = Op1.getValue(1);
1872 // Insert the EHSELECTION instruction.
1873 VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
1876 SDValue Op2 = DAG.getNode(ISD::EHSELECTION, getCurDebugLoc(), VTs, Ops, 2);
1877 Chain = Op2.getValue(1);
1878 Op2 = DAG.getSExtOrTrunc(Op2, getCurDebugLoc(), MVT::i32);
1882 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
1883 DAG.getVTList(&ValueVTs[0], ValueVTs.size()),
1886 std::pair<SDValue, SDValue> RetPair = std::make_pair(Res, Chain);
1887 setValue(&LP, RetPair.first);
1888 DAG.setRoot(RetPair.second);
1891 /// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for
1892 /// small case ranges).
1893 bool SelectionDAGBuilder::handleSmallSwitchRange(CaseRec& CR,
1894 CaseRecVector& WorkList,
1896 MachineBasicBlock *Default,
1897 MachineBasicBlock *SwitchBB) {
1898 Case& BackCase = *(CR.Range.second-1);
1900 // Size is the number of Cases represented by this range.
1901 size_t Size = CR.Range.second - CR.Range.first;
1905 // Get the MachineFunction which holds the current MBB. This is used when
1906 // inserting any additional MBBs necessary to represent the switch.
1907 MachineFunction *CurMF = FuncInfo.MF;
1909 // Figure out which block is immediately after the current one.
1910 MachineBasicBlock *NextBlock = 0;
1911 MachineFunction::iterator BBI = CR.CaseBB;
1913 if (++BBI != FuncInfo.MF->end())
1916 // If any two of the cases has the same destination, and if one value
1917 // is the same as the other, but has one bit unset that the other has set,
1918 // use bit manipulation to do two compares at once. For example:
1919 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
1920 // TODO: This could be extended to merge any 2 cases in switches with 3 cases.
1921 // TODO: Handle cases where CR.CaseBB != SwitchBB.
1922 if (Size == 2 && CR.CaseBB == SwitchBB) {
1923 Case &Small = *CR.Range.first;
1924 Case &Big = *(CR.Range.second-1);
1926 if (Small.Low == Small.High && Big.Low == Big.High && Small.BB == Big.BB) {
1927 const APInt& SmallValue = cast<ConstantInt>(Small.Low)->getValue();
1928 const APInt& BigValue = cast<ConstantInt>(Big.Low)->getValue();
1930 // Check that there is only one bit different.
1931 if (BigValue.countPopulation() == SmallValue.countPopulation() + 1 &&
1932 (SmallValue | BigValue) == BigValue) {
1933 // Isolate the common bit.
1934 APInt CommonBit = BigValue & ~SmallValue;
1935 assert((SmallValue | CommonBit) == BigValue &&
1936 CommonBit.countPopulation() == 1 && "Not a common bit?");
1938 SDValue CondLHS = getValue(SV);
1939 EVT VT = CondLHS.getValueType();
1940 DebugLoc DL = getCurDebugLoc();
1942 SDValue Or = DAG.getNode(ISD::OR, DL, VT, CondLHS,
1943 DAG.getConstant(CommonBit, VT));
1944 SDValue Cond = DAG.getSetCC(DL, MVT::i1,
1945 Or, DAG.getConstant(BigValue, VT),
1948 // Update successor info.
1949 addSuccessorWithWeight(SwitchBB, Small.BB);
1950 addSuccessorWithWeight(SwitchBB, Default);
1952 // Insert the true branch.
1953 SDValue BrCond = DAG.getNode(ISD::BRCOND, DL, MVT::Other,
1954 getControlRoot(), Cond,
1955 DAG.getBasicBlock(Small.BB));
1957 // Insert the false branch.
1958 BrCond = DAG.getNode(ISD::BR, DL, MVT::Other, BrCond,
1959 DAG.getBasicBlock(Default));
1961 DAG.setRoot(BrCond);
1967 // Rearrange the case blocks so that the last one falls through if possible.
1968 if (NextBlock && Default != NextBlock && BackCase.BB != NextBlock) {
1969 // The last case block won't fall through into 'NextBlock' if we emit the
1970 // branches in this order. See if rearranging a case value would help.
1971 for (CaseItr I = CR.Range.first, E = CR.Range.second-1; I != E; ++I) {
1972 if (I->BB == NextBlock) {
1973 std::swap(*I, BackCase);
1979 // Create a CaseBlock record representing a conditional branch to
1980 // the Case's target mbb if the value being switched on SV is equal
1982 MachineBasicBlock *CurBlock = CR.CaseBB;
1983 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
1984 MachineBasicBlock *FallThrough;
1986 FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock());
1987 CurMF->insert(BBI, FallThrough);
1989 // Put SV in a virtual register to make it available from the new blocks.
1990 ExportFromCurrentBlock(SV);
1992 // If the last case doesn't match, go to the default block.
1993 FallThrough = Default;
1996 const Value *RHS, *LHS, *MHS;
1998 if (I->High == I->Low) {
1999 // This is just small small case range :) containing exactly 1 case
2001 LHS = SV; RHS = I->High; MHS = NULL;
2004 LHS = I->Low; MHS = SV; RHS = I->High;
2007 uint32_t ExtraWeight = I->ExtraWeight;
2008 CaseBlock CB(CC, LHS, RHS, MHS, /* truebb */ I->BB, /* falsebb */ FallThrough,
2010 /* trueweight */ ExtraWeight / 2, /* falseweight */ ExtraWeight / 2);
2012 // If emitting the first comparison, just call visitSwitchCase to emit the
2013 // code into the current block. Otherwise, push the CaseBlock onto the
2014 // vector to be later processed by SDISel, and insert the node's MBB
2015 // before the next MBB.
2016 if (CurBlock == SwitchBB)
2017 visitSwitchCase(CB, SwitchBB);
2019 SwitchCases.push_back(CB);
2021 CurBlock = FallThrough;
2027 static inline bool areJTsAllowed(const TargetLowering &TLI) {
2028 return !TLI.getTargetMachine().Options.DisableJumpTables &&
2029 (TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
2030 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other));
2033 static APInt ComputeRange(const APInt &First, const APInt &Last) {
2034 uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1;
2035 APInt LastExt = Last.sext(BitWidth), FirstExt = First.sext(BitWidth);
2036 return (LastExt - FirstExt + 1ULL);
2039 /// handleJTSwitchCase - Emit jumptable for current switch case range
2040 bool SelectionDAGBuilder::handleJTSwitchCase(CaseRec &CR,
2041 CaseRecVector &WorkList,
2043 MachineBasicBlock *Default,
2044 MachineBasicBlock *SwitchBB) {
2045 Case& FrontCase = *CR.Range.first;
2046 Case& BackCase = *(CR.Range.second-1);
2048 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
2049 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
2051 APInt TSize(First.getBitWidth(), 0);
2052 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I)
2055 if (!areJTsAllowed(TLI) || TSize.ult(4))
2058 APInt Range = ComputeRange(First, Last);
2059 // The density is TSize / Range. Require at least 40%.
2060 // It should not be possible for IntTSize to saturate for sane code, but make
2061 // sure we handle Range saturation correctly.
2062 uint64_t IntRange = Range.getLimitedValue(UINT64_MAX/10);
2063 uint64_t IntTSize = TSize.getLimitedValue(UINT64_MAX/10);
2064 if (IntTSize * 10 < IntRange * 4)
2067 DEBUG(dbgs() << "Lowering jump table\n"
2068 << "First entry: " << First << ". Last entry: " << Last << '\n'
2069 << "Range: " << Range << ". Size: " << TSize << ".\n\n");
2071 // Get the MachineFunction which holds the current MBB. This is used when
2072 // inserting any additional MBBs necessary to represent the switch.
2073 MachineFunction *CurMF = FuncInfo.MF;
2075 // Figure out which block is immediately after the current one.
2076 MachineFunction::iterator BBI = CR.CaseBB;
2079 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2081 // Create a new basic block to hold the code for loading the address
2082 // of the jump table, and jumping to it. Update successor information;
2083 // we will either branch to the default case for the switch, or the jump
2085 MachineBasicBlock *JumpTableBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2086 CurMF->insert(BBI, JumpTableBB);
2088 addSuccessorWithWeight(CR.CaseBB, Default);
2089 addSuccessorWithWeight(CR.CaseBB, JumpTableBB);
2091 // Build a vector of destination BBs, corresponding to each target
2092 // of the jump table. If the value of the jump table slot corresponds to
2093 // a case statement, push the case's BB onto the vector, otherwise, push
2095 std::vector<MachineBasicBlock*> DestBBs;
2097 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) {
2098 const APInt &Low = cast<ConstantInt>(I->Low)->getValue();
2099 const APInt &High = cast<ConstantInt>(I->High)->getValue();
2101 if (Low.sle(TEI) && TEI.sle(High)) {
2102 DestBBs.push_back(I->BB);
2106 DestBBs.push_back(Default);
2110 // Update successor info. Add one edge to each unique successor.
2111 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs());
2112 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(),
2113 E = DestBBs.end(); I != E; ++I) {
2114 if (!SuccsHandled[(*I)->getNumber()]) {
2115 SuccsHandled[(*I)->getNumber()] = true;
2116 addSuccessorWithWeight(JumpTableBB, *I);
2120 // Create a jump table index for this jump table.
2121 unsigned JTEncoding = TLI.getJumpTableEncoding();
2122 unsigned JTI = CurMF->getOrCreateJumpTableInfo(JTEncoding)
2123 ->createJumpTableIndex(DestBBs);
2125 // Set the jump table information so that we can codegen it as a second
2126 // MachineBasicBlock
2127 JumpTable JT(-1U, JTI, JumpTableBB, Default);
2128 JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == SwitchBB));
2129 if (CR.CaseBB == SwitchBB)
2130 visitJumpTableHeader(JT, JTH, SwitchBB);
2132 JTCases.push_back(JumpTableBlock(JTH, JT));
2136 /// handleBTSplitSwitchCase - emit comparison and split binary search tree into
2138 bool SelectionDAGBuilder::handleBTSplitSwitchCase(CaseRec& CR,
2139 CaseRecVector& WorkList,
2141 MachineBasicBlock *Default,
2142 MachineBasicBlock *SwitchBB) {
2143 // Get the MachineFunction which holds the current MBB. This is used when
2144 // inserting any additional MBBs necessary to represent the switch.
2145 MachineFunction *CurMF = FuncInfo.MF;
2147 // Figure out which block is immediately after the current one.
2148 MachineFunction::iterator BBI = CR.CaseBB;
2151 Case& FrontCase = *CR.Range.first;
2152 Case& BackCase = *(CR.Range.second-1);
2153 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2155 // Size is the number of Cases represented by this range.
2156 unsigned Size = CR.Range.second - CR.Range.first;
2158 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
2159 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
2161 CaseItr Pivot = CR.Range.first + Size/2;
2163 // Select optimal pivot, maximizing sum density of LHS and RHS. This will
2164 // (heuristically) allow us to emit JumpTable's later.
2165 APInt TSize(First.getBitWidth(), 0);
2166 for (CaseItr I = CR.Range.first, E = CR.Range.second;
2170 APInt LSize = FrontCase.size();
2171 APInt RSize = TSize-LSize;
2172 DEBUG(dbgs() << "Selecting best pivot: \n"
2173 << "First: " << First << ", Last: " << Last <<'\n'
2174 << "LSize: " << LSize << ", RSize: " << RSize << '\n');
2175 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second;
2177 const APInt &LEnd = cast<ConstantInt>(I->High)->getValue();
2178 const APInt &RBegin = cast<ConstantInt>(J->Low)->getValue();
2179 APInt Range = ComputeRange(LEnd, RBegin);
2180 assert((Range - 2ULL).isNonNegative() &&
2181 "Invalid case distance");
2182 // Use volatile double here to avoid excess precision issues on some hosts,
2183 // e.g. that use 80-bit X87 registers.
2184 volatile double LDensity =
2185 (double)LSize.roundToDouble() /
2186 (LEnd - First + 1ULL).roundToDouble();
2187 volatile double RDensity =
2188 (double)RSize.roundToDouble() /
2189 (Last - RBegin + 1ULL).roundToDouble();
2190 double Metric = Range.logBase2()*(LDensity+RDensity);
2191 // Should always split in some non-trivial place
2192 DEBUG(dbgs() <<"=>Step\n"
2193 << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n'
2194 << "LDensity: " << LDensity
2195 << ", RDensity: " << RDensity << '\n'
2196 << "Metric: " << Metric << '\n');
2197 if (FMetric < Metric) {
2200 DEBUG(dbgs() << "Current metric set to: " << FMetric << '\n');
2206 if (areJTsAllowed(TLI)) {
2207 // If our case is dense we *really* should handle it earlier!
2208 assert((FMetric > 0) && "Should handle dense range earlier!");
2210 Pivot = CR.Range.first + Size/2;
2213 CaseRange LHSR(CR.Range.first, Pivot);
2214 CaseRange RHSR(Pivot, CR.Range.second);
2215 const Constant *C = Pivot->Low;
2216 MachineBasicBlock *FalseBB = 0, *TrueBB = 0;
2218 // We know that we branch to the LHS if the Value being switched on is
2219 // less than the Pivot value, C. We use this to optimize our binary
2220 // tree a bit, by recognizing that if SV is greater than or equal to the
2221 // LHS's Case Value, and that Case Value is exactly one less than the
2222 // Pivot's Value, then we can branch directly to the LHS's Target,
2223 // rather than creating a leaf node for it.
2224 if ((LHSR.second - LHSR.first) == 1 &&
2225 LHSR.first->High == CR.GE &&
2226 cast<ConstantInt>(C)->getValue() ==
2227 (cast<ConstantInt>(CR.GE)->getValue() + 1LL)) {
2228 TrueBB = LHSR.first->BB;
2230 TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2231 CurMF->insert(BBI, TrueBB);
2232 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR));
2234 // Put SV in a virtual register to make it available from the new blocks.
2235 ExportFromCurrentBlock(SV);
2238 // Similar to the optimization above, if the Value being switched on is
2239 // known to be less than the Constant CR.LT, and the current Case Value
2240 // is CR.LT - 1, then we can branch directly to the target block for
2241 // the current Case Value, rather than emitting a RHS leaf node for it.
2242 if ((RHSR.second - RHSR.first) == 1 && CR.LT &&
2243 cast<ConstantInt>(RHSR.first->Low)->getValue() ==
2244 (cast<ConstantInt>(CR.LT)->getValue() - 1LL)) {
2245 FalseBB = RHSR.first->BB;
2247 FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2248 CurMF->insert(BBI, FalseBB);
2249 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR));
2251 // Put SV in a virtual register to make it available from the new blocks.
2252 ExportFromCurrentBlock(SV);
2255 // Create a CaseBlock record representing a conditional branch to
2256 // the LHS node if the value being switched on SV is less than C.
2257 // Otherwise, branch to LHS.
2258 CaseBlock CB(ISD::SETLT, SV, C, NULL, TrueBB, FalseBB, CR.CaseBB);
2260 if (CR.CaseBB == SwitchBB)
2261 visitSwitchCase(CB, SwitchBB);
2263 SwitchCases.push_back(CB);
2268 /// handleBitTestsSwitchCase - if current case range has few destination and
2269 /// range span less, than machine word bitwidth, encode case range into series
2270 /// of masks and emit bit tests with these masks.
2271 bool SelectionDAGBuilder::handleBitTestsSwitchCase(CaseRec& CR,
2272 CaseRecVector& WorkList,
2274 MachineBasicBlock* Default,
2275 MachineBasicBlock *SwitchBB){
2276 EVT PTy = TLI.getPointerTy();
2277 unsigned IntPtrBits = PTy.getSizeInBits();
2279 Case& FrontCase = *CR.Range.first;
2280 Case& BackCase = *(CR.Range.second-1);
2282 // Get the MachineFunction which holds the current MBB. This is used when
2283 // inserting any additional MBBs necessary to represent the switch.
2284 MachineFunction *CurMF = FuncInfo.MF;
2286 // If target does not have legal shift left, do not emit bit tests at all.
2287 if (!TLI.isOperationLegal(ISD::SHL, TLI.getPointerTy()))
2291 for (CaseItr I = CR.Range.first, E = CR.Range.second;
2293 // Single case counts one, case range - two.
2294 numCmps += (I->Low == I->High ? 1 : 2);
2297 // Count unique destinations
2298 SmallSet<MachineBasicBlock*, 4> Dests;
2299 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2300 Dests.insert(I->BB);
2301 if (Dests.size() > 3)
2302 // Don't bother the code below, if there are too much unique destinations
2305 DEBUG(dbgs() << "Total number of unique destinations: "
2306 << Dests.size() << '\n'
2307 << "Total number of comparisons: " << numCmps << '\n');
2309 // Compute span of values.
2310 const APInt& minValue = cast<ConstantInt>(FrontCase.Low)->getValue();
2311 const APInt& maxValue = cast<ConstantInt>(BackCase.High)->getValue();
2312 APInt cmpRange = maxValue - minValue;
2314 DEBUG(dbgs() << "Compare range: " << cmpRange << '\n'
2315 << "Low bound: " << minValue << '\n'
2316 << "High bound: " << maxValue << '\n');
2318 if (cmpRange.uge(IntPtrBits) ||
2319 (!(Dests.size() == 1 && numCmps >= 3) &&
2320 !(Dests.size() == 2 && numCmps >= 5) &&
2321 !(Dests.size() >= 3 && numCmps >= 6)))
2324 DEBUG(dbgs() << "Emitting bit tests\n");
2325 APInt lowBound = APInt::getNullValue(cmpRange.getBitWidth());
2327 // Optimize the case where all the case values fit in a
2328 // word without having to subtract minValue. In this case,
2329 // we can optimize away the subtraction.
2330 if (minValue.isNonNegative() && maxValue.slt(IntPtrBits)) {
2331 cmpRange = maxValue;
2333 lowBound = minValue;
2336 CaseBitsVector CasesBits;
2337 unsigned i, count = 0;
2339 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2340 MachineBasicBlock* Dest = I->BB;
2341 for (i = 0; i < count; ++i)
2342 if (Dest == CasesBits[i].BB)
2346 assert((count < 3) && "Too much destinations to test!");
2347 CasesBits.push_back(CaseBits(0, Dest, 0));
2351 const APInt& lowValue = cast<ConstantInt>(I->Low)->getValue();
2352 const APInt& highValue = cast<ConstantInt>(I->High)->getValue();
2354 uint64_t lo = (lowValue - lowBound).getZExtValue();
2355 uint64_t hi = (highValue - lowBound).getZExtValue();
2357 for (uint64_t j = lo; j <= hi; j++) {
2358 CasesBits[i].Mask |= 1ULL << j;
2359 CasesBits[i].Bits++;
2363 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp());
2367 // Figure out which block is immediately after the current one.
2368 MachineFunction::iterator BBI = CR.CaseBB;
2371 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2373 DEBUG(dbgs() << "Cases:\n");
2374 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) {
2375 DEBUG(dbgs() << "Mask: " << CasesBits[i].Mask
2376 << ", Bits: " << CasesBits[i].Bits
2377 << ", BB: " << CasesBits[i].BB << '\n');
2379 MachineBasicBlock *CaseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2380 CurMF->insert(BBI, CaseBB);
2381 BTC.push_back(BitTestCase(CasesBits[i].Mask,
2385 // Put SV in a virtual register to make it available from the new blocks.
2386 ExportFromCurrentBlock(SV);
2389 BitTestBlock BTB(lowBound, cmpRange, SV,
2390 -1U, MVT::Other, (CR.CaseBB == SwitchBB),
2391 CR.CaseBB, Default, BTC);
2393 if (CR.CaseBB == SwitchBB)
2394 visitBitTestHeader(BTB, SwitchBB);
2396 BitTestCases.push_back(BTB);
2401 /// Clusterify - Transform simple list of Cases into list of CaseRange's
2402 size_t SelectionDAGBuilder::Clusterify(CaseVector& Cases,
2403 const SwitchInst& SI) {
2406 BranchProbabilityInfo *BPI = FuncInfo.BPI;
2407 // Start with "simple" cases
2408 for (size_t i = 0; i < SI.getNumCases(); ++i) {
2409 BasicBlock *SuccBB = SI.getCaseSuccessor(i);
2410 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SuccBB];
2412 uint32_t ExtraWeight = BPI ? BPI->getEdgeWeight(SI.getParent(), SuccBB) : 0;
2414 Cases.push_back(Case(SI.getCaseValue(i),
2416 SMBB, ExtraWeight));
2418 std::sort(Cases.begin(), Cases.end(), CaseCmp());
2420 // Merge case into clusters
2421 if (Cases.size() >= 2)
2422 // Must recompute end() each iteration because it may be
2423 // invalidated by erase if we hold on to it
2424 for (CaseItr I = Cases.begin(), J = llvm::next(Cases.begin());
2425 J != Cases.end(); ) {
2426 const APInt& nextValue = cast<ConstantInt>(J->Low)->getValue();
2427 const APInt& currentValue = cast<ConstantInt>(I->High)->getValue();
2428 MachineBasicBlock* nextBB = J->BB;
2429 MachineBasicBlock* currentBB = I->BB;
2431 // If the two neighboring cases go to the same destination, merge them
2432 // into a single case.
2433 if ((nextValue - currentValue == 1) && (currentBB == nextBB)) {
2437 if (BranchProbabilityInfo *BPI = FuncInfo.BPI) {
2438 uint32_t CurWeight = currentBB->getBasicBlock() ?
2439 BPI->getEdgeWeight(SI.getParent(), currentBB->getBasicBlock()) : 16;
2440 uint32_t NextWeight = nextBB->getBasicBlock() ?
2441 BPI->getEdgeWeight(SI.getParent(), nextBB->getBasicBlock()) : 16;
2443 BPI->setEdgeWeight(SI.getParent(), currentBB->getBasicBlock(),
2444 CurWeight + NextWeight);
2451 for (CaseItr I=Cases.begin(), E=Cases.end(); I!=E; ++I, ++numCmps) {
2452 if (I->Low != I->High)
2453 // A range counts double, since it requires two compares.
2460 void SelectionDAGBuilder::UpdateSplitBlock(MachineBasicBlock *First,
2461 MachineBasicBlock *Last) {
2463 for (unsigned i = 0, e = JTCases.size(); i != e; ++i)
2464 if (JTCases[i].first.HeaderBB == First)
2465 JTCases[i].first.HeaderBB = Last;
2467 // Update BitTestCases.
2468 for (unsigned i = 0, e = BitTestCases.size(); i != e; ++i)
2469 if (BitTestCases[i].Parent == First)
2470 BitTestCases[i].Parent = Last;
2473 void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) {
2474 MachineBasicBlock *SwitchMBB = FuncInfo.MBB;
2476 // Figure out which block is immediately after the current one.
2477 MachineBasicBlock *NextBlock = 0;
2478 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()];
2480 // If there is only the default destination, branch to it if it is not the
2481 // next basic block. Otherwise, just fall through.
2482 if (!SI.getNumCases()) {
2483 // Update machine-CFG edges.
2485 // If this is not a fall-through branch, emit the branch.
2486 SwitchMBB->addSuccessor(Default);
2487 if (Default != NextBlock)
2488 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
2489 MVT::Other, getControlRoot(),
2490 DAG.getBasicBlock(Default)));
2495 // If there are any non-default case statements, create a vector of Cases
2496 // representing each one, and sort the vector so that we can efficiently
2497 // create a binary search tree from them.
2499 size_t numCmps = Clusterify(Cases, SI);
2500 DEBUG(dbgs() << "Clusterify finished. Total clusters: " << Cases.size()
2501 << ". Total compares: " << numCmps << '\n');
2504 // Get the Value to be switched on and default basic blocks, which will be
2505 // inserted into CaseBlock records, representing basic blocks in the binary
2507 const Value *SV = SI.getCondition();
2509 // Push the initial CaseRec onto the worklist
2510 CaseRecVector WorkList;
2511 WorkList.push_back(CaseRec(SwitchMBB,0,0,
2512 CaseRange(Cases.begin(),Cases.end())));
2514 while (!WorkList.empty()) {
2515 // Grab a record representing a case range to process off the worklist
2516 CaseRec CR = WorkList.back();
2517 WorkList.pop_back();
2519 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2522 // If the range has few cases (two or less) emit a series of specific
2524 if (handleSmallSwitchRange(CR, WorkList, SV, Default, SwitchMBB))
2527 // If the switch has more than 5 blocks, and at least 40% dense, and the
2528 // target supports indirect branches, then emit a jump table rather than
2529 // lowering the switch to a binary tree of conditional branches.
2530 if (handleJTSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2533 // Emit binary tree. We need to pick a pivot, and push left and right ranges
2534 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call.
2535 handleBTSplitSwitchCase(CR, WorkList, SV, Default, SwitchMBB);
2539 void SelectionDAGBuilder::visitIndirectBr(const IndirectBrInst &I) {
2540 MachineBasicBlock *IndirectBrMBB = FuncInfo.MBB;
2542 // Update machine-CFG edges with unique successors.
2543 SmallVector<BasicBlock*, 32> succs;
2544 succs.reserve(I.getNumSuccessors());
2545 for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i)
2546 succs.push_back(I.getSuccessor(i));
2547 array_pod_sort(succs.begin(), succs.end());
2548 succs.erase(std::unique(succs.begin(), succs.end()), succs.end());
2549 for (unsigned i = 0, e = succs.size(); i != e; ++i) {
2550 MachineBasicBlock *Succ = FuncInfo.MBBMap[succs[i]];
2551 addSuccessorWithWeight(IndirectBrMBB, Succ);
2554 DAG.setRoot(DAG.getNode(ISD::BRIND, getCurDebugLoc(),
2555 MVT::Other, getControlRoot(),
2556 getValue(I.getAddress())));
2559 void SelectionDAGBuilder::visitFSub(const User &I) {
2560 // -0.0 - X --> fneg
2561 Type *Ty = I.getType();
2562 if (isa<Constant>(I.getOperand(0)) &&
2563 I.getOperand(0) == ConstantFP::getZeroValueForNegation(Ty)) {
2564 SDValue Op2 = getValue(I.getOperand(1));
2565 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(),
2566 Op2.getValueType(), Op2));
2570 visitBinary(I, ISD::FSUB);
2573 void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) {
2574 SDValue Op1 = getValue(I.getOperand(0));
2575 SDValue Op2 = getValue(I.getOperand(1));
2576 setValue(&I, DAG.getNode(OpCode, getCurDebugLoc(),
2577 Op1.getValueType(), Op1, Op2));
2580 void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) {
2581 SDValue Op1 = getValue(I.getOperand(0));
2582 SDValue Op2 = getValue(I.getOperand(1));
2584 MVT ShiftTy = TLI.getShiftAmountTy(Op2.getValueType());
2586 // Coerce the shift amount to the right type if we can.
2587 if (!I.getType()->isVectorTy() && Op2.getValueType() != ShiftTy) {
2588 unsigned ShiftSize = ShiftTy.getSizeInBits();
2589 unsigned Op2Size = Op2.getValueType().getSizeInBits();
2590 DebugLoc DL = getCurDebugLoc();
2592 // If the operand is smaller than the shift count type, promote it.
2593 if (ShiftSize > Op2Size)
2594 Op2 = DAG.getNode(ISD::ZERO_EXTEND, DL, ShiftTy, Op2);
2596 // If the operand is larger than the shift count type but the shift
2597 // count type has enough bits to represent any shift value, truncate
2598 // it now. This is a common case and it exposes the truncate to
2599 // optimization early.
2600 else if (ShiftSize >= Log2_32_Ceil(Op2.getValueType().getSizeInBits()))
2601 Op2 = DAG.getNode(ISD::TRUNCATE, DL, ShiftTy, Op2);
2602 // Otherwise we'll need to temporarily settle for some other convenient
2603 // type. Type legalization will make adjustments once the shiftee is split.
2605 Op2 = DAG.getZExtOrTrunc(Op2, DL, MVT::i32);
2608 setValue(&I, DAG.getNode(Opcode, getCurDebugLoc(),
2609 Op1.getValueType(), Op1, Op2));
2612 void SelectionDAGBuilder::visitSDiv(const User &I) {
2613 SDValue Op1 = getValue(I.getOperand(0));
2614 SDValue Op2 = getValue(I.getOperand(1));
2616 // Turn exact SDivs into multiplications.
2617 // FIXME: This should be in DAGCombiner, but it doesn't have access to the
2619 if (isa<BinaryOperator>(&I) && cast<BinaryOperator>(&I)->isExact() &&
2620 !isa<ConstantSDNode>(Op1) &&
2621 isa<ConstantSDNode>(Op2) && !cast<ConstantSDNode>(Op2)->isNullValue())
2622 setValue(&I, TLI.BuildExactSDIV(Op1, Op2, getCurDebugLoc(), DAG));
2624 setValue(&I, DAG.getNode(ISD::SDIV, getCurDebugLoc(), Op1.getValueType(),
2628 void SelectionDAGBuilder::visitICmp(const User &I) {
2629 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2630 if (const ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2631 predicate = IC->getPredicate();
2632 else if (const ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2633 predicate = ICmpInst::Predicate(IC->getPredicate());
2634 SDValue Op1 = getValue(I.getOperand(0));
2635 SDValue Op2 = getValue(I.getOperand(1));
2636 ISD::CondCode Opcode = getICmpCondCode(predicate);
2638 EVT DestVT = TLI.getValueType(I.getType());
2639 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Opcode));
2642 void SelectionDAGBuilder::visitFCmp(const User &I) {
2643 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2644 if (const FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2645 predicate = FC->getPredicate();
2646 else if (const ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2647 predicate = FCmpInst::Predicate(FC->getPredicate());
2648 SDValue Op1 = getValue(I.getOperand(0));
2649 SDValue Op2 = getValue(I.getOperand(1));
2650 ISD::CondCode Condition = getFCmpCondCode(predicate);
2651 if (TM.Options.NoNaNsFPMath)
2652 Condition = getFCmpCodeWithoutNaN(Condition);
2653 EVT DestVT = TLI.getValueType(I.getType());
2654 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Condition));
2657 void SelectionDAGBuilder::visitSelect(const User &I) {
2658 SmallVector<EVT, 4> ValueVTs;
2659 ComputeValueVTs(TLI, I.getType(), ValueVTs);
2660 unsigned NumValues = ValueVTs.size();
2661 if (NumValues == 0) return;
2663 SmallVector<SDValue, 4> Values(NumValues);
2664 SDValue Cond = getValue(I.getOperand(0));
2665 SDValue TrueVal = getValue(I.getOperand(1));
2666 SDValue FalseVal = getValue(I.getOperand(2));
2667 ISD::NodeType OpCode = Cond.getValueType().isVector() ?
2668 ISD::VSELECT : ISD::SELECT;
2670 for (unsigned i = 0; i != NumValues; ++i)
2671 Values[i] = DAG.getNode(OpCode, getCurDebugLoc(),
2672 TrueVal.getNode()->getValueType(TrueVal.getResNo()+i),
2674 SDValue(TrueVal.getNode(),
2675 TrueVal.getResNo() + i),
2676 SDValue(FalseVal.getNode(),
2677 FalseVal.getResNo() + i));
2679 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2680 DAG.getVTList(&ValueVTs[0], NumValues),
2681 &Values[0], NumValues));
2684 void SelectionDAGBuilder::visitTrunc(const User &I) {
2685 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2686 SDValue N = getValue(I.getOperand(0));
2687 EVT DestVT = TLI.getValueType(I.getType());
2688 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), DestVT, N));
2691 void SelectionDAGBuilder::visitZExt(const User &I) {
2692 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2693 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2694 SDValue N = getValue(I.getOperand(0));
2695 EVT DestVT = TLI.getValueType(I.getType());
2696 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(), DestVT, N));
2699 void SelectionDAGBuilder::visitSExt(const User &I) {
2700 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2701 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2702 SDValue N = getValue(I.getOperand(0));
2703 EVT DestVT = TLI.getValueType(I.getType());
2704 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurDebugLoc(), DestVT, N));
2707 void SelectionDAGBuilder::visitFPTrunc(const User &I) {
2708 // FPTrunc is never a no-op cast, no need to check
2709 SDValue N = getValue(I.getOperand(0));
2710 EVT DestVT = TLI.getValueType(I.getType());
2711 setValue(&I, DAG.getNode(ISD::FP_ROUND, getCurDebugLoc(),
2713 DAG.getTargetConstant(0, TLI.getPointerTy())));
2716 void SelectionDAGBuilder::visitFPExt(const User &I){
2717 // FPExt is never a no-op cast, no need to check
2718 SDValue N = getValue(I.getOperand(0));
2719 EVT DestVT = TLI.getValueType(I.getType());
2720 setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurDebugLoc(), DestVT, N));
2723 void SelectionDAGBuilder::visitFPToUI(const User &I) {
2724 // FPToUI is never a no-op cast, no need to check
2725 SDValue N = getValue(I.getOperand(0));
2726 EVT DestVT = TLI.getValueType(I.getType());
2727 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurDebugLoc(), DestVT, N));
2730 void SelectionDAGBuilder::visitFPToSI(const User &I) {
2731 // FPToSI is never a no-op cast, no need to check
2732 SDValue N = getValue(I.getOperand(0));
2733 EVT DestVT = TLI.getValueType(I.getType());
2734 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurDebugLoc(), DestVT, N));
2737 void SelectionDAGBuilder::visitUIToFP(const User &I) {
2738 // UIToFP is never a no-op cast, no need to check
2739 SDValue N = getValue(I.getOperand(0));
2740 EVT DestVT = TLI.getValueType(I.getType());
2741 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurDebugLoc(), DestVT, N));
2744 void SelectionDAGBuilder::visitSIToFP(const User &I){
2745 // SIToFP is never a no-op cast, no need to check
2746 SDValue N = getValue(I.getOperand(0));
2747 EVT DestVT = TLI.getValueType(I.getType());
2748 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurDebugLoc(), DestVT, N));
2751 void SelectionDAGBuilder::visitPtrToInt(const User &I) {
2752 // What to do depends on the size of the integer and the size of the pointer.
2753 // We can either truncate, zero extend, or no-op, accordingly.
2754 SDValue N = getValue(I.getOperand(0));
2755 EVT DestVT = TLI.getValueType(I.getType());
2756 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT));
2759 void SelectionDAGBuilder::visitIntToPtr(const User &I) {
2760 // What to do depends on the size of the integer and the size of the pointer.
2761 // We can either truncate, zero extend, or no-op, accordingly.
2762 SDValue N = getValue(I.getOperand(0));
2763 EVT DestVT = TLI.getValueType(I.getType());
2764 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT));
2767 void SelectionDAGBuilder::visitBitCast(const User &I) {
2768 SDValue N = getValue(I.getOperand(0));
2769 EVT DestVT = TLI.getValueType(I.getType());
2771 // BitCast assures us that source and destination are the same size so this is
2772 // either a BITCAST or a no-op.
2773 if (DestVT != N.getValueType())
2774 setValue(&I, DAG.getNode(ISD::BITCAST, getCurDebugLoc(),
2775 DestVT, N)); // convert types.
2777 setValue(&I, N); // noop cast.
2780 void SelectionDAGBuilder::visitInsertElement(const User &I) {
2781 SDValue InVec = getValue(I.getOperand(0));
2782 SDValue InVal = getValue(I.getOperand(1));
2783 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2785 getValue(I.getOperand(2)));
2786 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurDebugLoc(),
2787 TLI.getValueType(I.getType()),
2788 InVec, InVal, InIdx));
2791 void SelectionDAGBuilder::visitExtractElement(const User &I) {
2792 SDValue InVec = getValue(I.getOperand(0));
2793 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2795 getValue(I.getOperand(1)));
2796 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2797 TLI.getValueType(I.getType()), InVec, InIdx));
2800 // Utility for visitShuffleVector - Return true if every element in Mask,
2801 // begining // from position Pos and ending in Pos+Size, falls within the
2802 // specified sequential range [L, L+Pos). or is undef.
2803 static bool isSequentialInRange(const SmallVectorImpl<int> &Mask,
2804 int Pos, int Size, int Low) {
2805 for (int i = Pos, e = Pos+Size; i != e; ++i, ++Low)
2806 if (Mask[i] >= 0 && Mask[i] != Low)
2811 void SelectionDAGBuilder::visitShuffleVector(const User &I) {
2812 SDValue Src1 = getValue(I.getOperand(0));
2813 SDValue Src2 = getValue(I.getOperand(1));
2815 SmallVector<int, 8> Mask;
2816 ShuffleVectorInst::getShuffleMask(cast<Constant>(I.getOperand(2)), Mask);
2817 unsigned MaskNumElts = Mask.size();
2819 EVT VT = TLI.getValueType(I.getType());
2820 EVT SrcVT = Src1.getValueType();
2821 unsigned SrcNumElts = SrcVT.getVectorNumElements();
2823 if (SrcNumElts == MaskNumElts) {
2824 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2829 // Normalize the shuffle vector since mask and vector length don't match.
2830 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) {
2831 // Mask is longer than the source vectors and is a multiple of the source
2832 // vectors. We can use concatenate vector to make the mask and vectors
2834 if (SrcNumElts*2 == MaskNumElts) {
2835 // First check for Src1 in low and Src2 in high
2836 if (isSequentialInRange(Mask, 0, SrcNumElts, 0) &&
2837 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, SrcNumElts)) {
2838 // The shuffle is concatenating two vectors together.
2839 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurDebugLoc(),
2843 // Then check for Src2 in low and Src1 in high
2844 if (isSequentialInRange(Mask, 0, SrcNumElts, SrcNumElts) &&
2845 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, 0)) {
2846 // The shuffle is concatenating two vectors together.
2847 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurDebugLoc(),
2853 // Pad both vectors with undefs to make them the same length as the mask.
2854 unsigned NumConcat = MaskNumElts / SrcNumElts;
2855 bool Src1U = Src1.getOpcode() == ISD::UNDEF;
2856 bool Src2U = Src2.getOpcode() == ISD::UNDEF;
2857 SDValue UndefVal = DAG.getUNDEF(SrcVT);
2859 SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal);
2860 SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal);
2864 Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2865 getCurDebugLoc(), VT,
2866 &MOps1[0], NumConcat);
2867 Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2868 getCurDebugLoc(), VT,
2869 &MOps2[0], NumConcat);
2871 // Readjust mask for new input vector length.
2872 SmallVector<int, 8> MappedOps;
2873 for (unsigned i = 0; i != MaskNumElts; ++i) {
2875 if (Idx < (int)SrcNumElts)
2876 MappedOps.push_back(Idx);
2878 MappedOps.push_back(Idx + MaskNumElts - SrcNumElts);
2881 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2886 if (SrcNumElts > MaskNumElts) {
2887 // Analyze the access pattern of the vector to see if we can extract
2888 // two subvectors and do the shuffle. The analysis is done by calculating
2889 // the range of elements the mask access on both vectors.
2890 int MinRange[2] = { static_cast<int>(SrcNumElts+1),
2891 static_cast<int>(SrcNumElts+1)};
2892 int MaxRange[2] = {-1, -1};
2894 for (unsigned i = 0; i != MaskNumElts; ++i) {
2900 if (Idx >= (int)SrcNumElts) {
2904 if (Idx > MaxRange[Input])
2905 MaxRange[Input] = Idx;
2906 if (Idx < MinRange[Input])
2907 MinRange[Input] = Idx;
2910 // Check if the access is smaller than the vector size and can we find
2911 // a reasonable extract index.
2912 int RangeUse[2] = { 2, 2 }; // 0 = Unused, 1 = Extract, 2 = Can not
2914 int StartIdx[2]; // StartIdx to extract from
2915 for (int Input=0; Input < 2; ++Input) {
2916 if (MinRange[Input] == (int)(SrcNumElts+1) && MaxRange[Input] == -1) {
2917 RangeUse[Input] = 0; // Unused
2918 StartIdx[Input] = 0;
2919 } else if (MaxRange[Input] - MinRange[Input] < (int)MaskNumElts) {
2920 // Fits within range but we should see if we can find a good
2921 // start index that is a multiple of the mask length.
2922 if (MaxRange[Input] < (int)MaskNumElts) {
2923 RangeUse[Input] = 1; // Extract from beginning of the vector
2924 StartIdx[Input] = 0;
2926 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts;
2927 if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts &&
2928 StartIdx[Input] + MaskNumElts <= SrcNumElts)
2929 RangeUse[Input] = 1; // Extract from a multiple of the mask length.
2934 if (RangeUse[0] == 0 && RangeUse[1] == 0) {
2935 setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used.
2938 else if (RangeUse[0] < 2 && RangeUse[1] < 2) {
2939 // Extract appropriate subvector and generate a vector shuffle
2940 for (int Input=0; Input < 2; ++Input) {
2941 SDValue &Src = Input == 0 ? Src1 : Src2;
2942 if (RangeUse[Input] == 0)
2943 Src = DAG.getUNDEF(VT);
2945 Src = DAG.getNode(ISD::EXTRACT_SUBVECTOR, getCurDebugLoc(), VT,
2946 Src, DAG.getIntPtrConstant(StartIdx[Input]));
2949 // Calculate new mask.
2950 SmallVector<int, 8> MappedOps;
2951 for (unsigned i = 0; i != MaskNumElts; ++i) {
2954 MappedOps.push_back(Idx);
2955 else if (Idx < (int)SrcNumElts)
2956 MappedOps.push_back(Idx - StartIdx[0]);
2958 MappedOps.push_back(Idx - SrcNumElts - StartIdx[1] + MaskNumElts);
2961 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2967 // We can't use either concat vectors or extract subvectors so fall back to
2968 // replacing the shuffle with extract and build vector.
2969 // to insert and build vector.
2970 EVT EltVT = VT.getVectorElementType();
2971 EVT PtrVT = TLI.getPointerTy();
2972 SmallVector<SDValue,8> Ops;
2973 for (unsigned i = 0; i != MaskNumElts; ++i) {
2975 Ops.push_back(DAG.getUNDEF(EltVT));
2980 if (Idx < (int)SrcNumElts)
2981 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2982 EltVT, Src1, DAG.getConstant(Idx, PtrVT));
2984 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2986 DAG.getConstant(Idx - SrcNumElts, PtrVT));
2992 setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
2993 VT, &Ops[0], Ops.size()));
2996 void SelectionDAGBuilder::visitInsertValue(const InsertValueInst &I) {
2997 const Value *Op0 = I.getOperand(0);
2998 const Value *Op1 = I.getOperand(1);
2999 Type *AggTy = I.getType();
3000 Type *ValTy = Op1->getType();
3001 bool IntoUndef = isa<UndefValue>(Op0);
3002 bool FromUndef = isa<UndefValue>(Op1);
3004 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
3006 SmallVector<EVT, 4> AggValueVTs;
3007 ComputeValueVTs(TLI, AggTy, AggValueVTs);
3008 SmallVector<EVT, 4> ValValueVTs;
3009 ComputeValueVTs(TLI, ValTy, ValValueVTs);
3011 unsigned NumAggValues = AggValueVTs.size();
3012 unsigned NumValValues = ValValueVTs.size();
3013 SmallVector<SDValue, 4> Values(NumAggValues);
3015 SDValue Agg = getValue(Op0);
3017 // Copy the beginning value(s) from the original aggregate.
3018 for (; i != LinearIndex; ++i)
3019 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
3020 SDValue(Agg.getNode(), Agg.getResNo() + i);
3021 // Copy values from the inserted value(s).
3023 SDValue Val = getValue(Op1);
3024 for (; i != LinearIndex + NumValValues; ++i)
3025 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) :
3026 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex);
3028 // Copy remaining value(s) from the original aggregate.
3029 for (; i != NumAggValues; ++i)
3030 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
3031 SDValue(Agg.getNode(), Agg.getResNo() + i);
3033 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
3034 DAG.getVTList(&AggValueVTs[0], NumAggValues),
3035 &Values[0], NumAggValues));
3038 void SelectionDAGBuilder::visitExtractValue(const ExtractValueInst &I) {
3039 const Value *Op0 = I.getOperand(0);
3040 Type *AggTy = Op0->getType();
3041 Type *ValTy = I.getType();
3042 bool OutOfUndef = isa<UndefValue>(Op0);
3044 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
3046 SmallVector<EVT, 4> ValValueVTs;
3047 ComputeValueVTs(TLI, ValTy, ValValueVTs);
3049 unsigned NumValValues = ValValueVTs.size();
3051 // Ignore a extractvalue that produces an empty object
3052 if (!NumValValues) {
3053 setValue(&I, DAG.getUNDEF(MVT(MVT::Other)));
3057 SmallVector<SDValue, 4> Values(NumValValues);
3059 SDValue Agg = getValue(Op0);
3060 // Copy out the selected value(s).
3061 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i)
3062 Values[i - LinearIndex] =
3064 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) :
3065 SDValue(Agg.getNode(), Agg.getResNo() + i);
3067 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
3068 DAG.getVTList(&ValValueVTs[0], NumValValues),
3069 &Values[0], NumValValues));
3072 void SelectionDAGBuilder::visitGetElementPtr(const User &I) {
3073 SDValue N = getValue(I.getOperand(0));
3074 Type *Ty = I.getOperand(0)->getType();
3076 for (GetElementPtrInst::const_op_iterator OI = I.op_begin()+1, E = I.op_end();
3078 const Value *Idx = *OI;
3079 if (StructType *StTy = dyn_cast<StructType>(Ty)) {
3080 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
3083 uint64_t Offset = TD->getStructLayout(StTy)->getElementOffset(Field);
3084 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
3085 DAG.getIntPtrConstant(Offset));
3088 Ty = StTy->getElementType(Field);
3090 Ty = cast<SequentialType>(Ty)->getElementType();
3092 // If this is a constant subscript, handle it quickly.
3093 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
3094 if (CI->isZero()) continue;
3096 TD->getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
3098 EVT PTy = TLI.getPointerTy();
3099 unsigned PtrBits = PTy.getSizeInBits();
3101 OffsVal = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
3103 DAG.getConstant(Offs, MVT::i64));
3105 OffsVal = DAG.getIntPtrConstant(Offs);
3107 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
3112 // N = N + Idx * ElementSize;
3113 APInt ElementSize = APInt(TLI.getPointerTy().getSizeInBits(),
3114 TD->getTypeAllocSize(Ty));
3115 SDValue IdxN = getValue(Idx);
3117 // If the index is smaller or larger than intptr_t, truncate or extend
3119 IdxN = DAG.getSExtOrTrunc(IdxN, getCurDebugLoc(), N.getValueType());
3121 // If this is a multiply by a power of two, turn it into a shl
3122 // immediately. This is a very common case.
3123 if (ElementSize != 1) {
3124 if (ElementSize.isPowerOf2()) {
3125 unsigned Amt = ElementSize.logBase2();
3126 IdxN = DAG.getNode(ISD::SHL, getCurDebugLoc(),
3127 N.getValueType(), IdxN,
3128 DAG.getConstant(Amt, IdxN.getValueType()));
3130 SDValue Scale = DAG.getConstant(ElementSize, TLI.getPointerTy());
3131 IdxN = DAG.getNode(ISD::MUL, getCurDebugLoc(),
3132 N.getValueType(), IdxN, Scale);
3136 N = DAG.getNode(ISD::ADD, getCurDebugLoc(),
3137 N.getValueType(), N, IdxN);
3144 void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) {
3145 // If this is a fixed sized alloca in the entry block of the function,
3146 // allocate it statically on the stack.
3147 if (FuncInfo.StaticAllocaMap.count(&I))
3148 return; // getValue will auto-populate this.
3150 Type *Ty = I.getAllocatedType();
3151 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
3153 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty),
3156 SDValue AllocSize = getValue(I.getArraySize());
3158 EVT IntPtr = TLI.getPointerTy();
3159 if (AllocSize.getValueType() != IntPtr)
3160 AllocSize = DAG.getZExtOrTrunc(AllocSize, getCurDebugLoc(), IntPtr);
3162 AllocSize = DAG.getNode(ISD::MUL, getCurDebugLoc(), IntPtr,
3164 DAG.getConstant(TySize, IntPtr));
3166 // Handle alignment. If the requested alignment is less than or equal to
3167 // the stack alignment, ignore it. If the size is greater than or equal to
3168 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
3169 unsigned StackAlign = TM.getFrameLowering()->getStackAlignment();
3170 if (Align <= StackAlign)
3173 // Round the size of the allocation up to the stack alignment size
3174 // by add SA-1 to the size.
3175 AllocSize = DAG.getNode(ISD::ADD, getCurDebugLoc(),
3176 AllocSize.getValueType(), AllocSize,
3177 DAG.getIntPtrConstant(StackAlign-1));
3179 // Mask out the low bits for alignment purposes.
3180 AllocSize = DAG.getNode(ISD::AND, getCurDebugLoc(),
3181 AllocSize.getValueType(), AllocSize,
3182 DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1)));
3184 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) };
3185 SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other);
3186 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, getCurDebugLoc(),
3189 DAG.setRoot(DSA.getValue(1));
3191 // Inform the Frame Information that we have just allocated a variable-sized
3193 FuncInfo.MF->getFrameInfo()->CreateVariableSizedObject(Align ? Align : 1);
3196 void SelectionDAGBuilder::visitLoad(const LoadInst &I) {
3198 return visitAtomicLoad(I);
3200 const Value *SV = I.getOperand(0);
3201 SDValue Ptr = getValue(SV);
3203 Type *Ty = I.getType();
3205 bool isVolatile = I.isVolatile();
3206 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
3207 bool isInvariant = I.getMetadata("invariant.load") != 0;
3208 unsigned Alignment = I.getAlignment();
3209 const MDNode *TBAAInfo = I.getMetadata(LLVMContext::MD_tbaa);
3211 SmallVector<EVT, 4> ValueVTs;
3212 SmallVector<uint64_t, 4> Offsets;
3213 ComputeValueVTs(TLI, Ty, ValueVTs, &Offsets);
3214 unsigned NumValues = ValueVTs.size();
3219 bool ConstantMemory = false;
3220 if (I.isVolatile() || NumValues > MaxParallelChains)
3221 // Serialize volatile loads with other side effects.
3223 else if (AA->pointsToConstantMemory(
3224 AliasAnalysis::Location(SV, AA->getTypeStoreSize(Ty), TBAAInfo))) {
3225 // Do not serialize (non-volatile) loads of constant memory with anything.
3226 Root = DAG.getEntryNode();
3227 ConstantMemory = true;
3229 // Do not serialize non-volatile loads against each other.
3230 Root = DAG.getRoot();
3233 SmallVector<SDValue, 4> Values(NumValues);
3234 SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains),
3236 EVT PtrVT = Ptr.getValueType();
3237 unsigned ChainI = 0;
3238 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3239 // Serializing loads here may result in excessive register pressure, and
3240 // TokenFactor places arbitrary choke points on the scheduler. SD scheduling
3241 // could recover a bit by hoisting nodes upward in the chain by recognizing
3242 // they are side-effect free or do not alias. The optimizer should really
3243 // avoid this case by converting large object/array copies to llvm.memcpy
3244 // (MaxParallelChains should always remain as failsafe).
3245 if (ChainI == MaxParallelChains) {
3246 assert(PendingLoads.empty() && "PendingLoads must be serialized first");
3247 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3248 MVT::Other, &Chains[0], ChainI);
3252 SDValue A = DAG.getNode(ISD::ADD, getCurDebugLoc(),
3254 DAG.getConstant(Offsets[i], PtrVT));
3255 SDValue L = DAG.getLoad(ValueVTs[i], getCurDebugLoc(), Root,
3256 A, MachinePointerInfo(SV, Offsets[i]), isVolatile,
3257 isNonTemporal, isInvariant, Alignment, TBAAInfo);
3260 Chains[ChainI] = L.getValue(1);
3263 if (!ConstantMemory) {
3264 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3265 MVT::Other, &Chains[0], ChainI);
3269 PendingLoads.push_back(Chain);
3272 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
3273 DAG.getVTList(&ValueVTs[0], NumValues),
3274 &Values[0], NumValues));
3277 void SelectionDAGBuilder::visitStore(const StoreInst &I) {
3279 return visitAtomicStore(I);
3281 const Value *SrcV = I.getOperand(0);
3282 const Value *PtrV = I.getOperand(1);
3284 SmallVector<EVT, 4> ValueVTs;
3285 SmallVector<uint64_t, 4> Offsets;
3286 ComputeValueVTs(TLI, SrcV->getType(), ValueVTs, &Offsets);
3287 unsigned NumValues = ValueVTs.size();
3291 // Get the lowered operands. Note that we do this after
3292 // checking if NumResults is zero, because with zero results
3293 // the operands won't have values in the map.
3294 SDValue Src = getValue(SrcV);
3295 SDValue Ptr = getValue(PtrV);
3297 SDValue Root = getRoot();
3298 SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains),
3300 EVT PtrVT = Ptr.getValueType();
3301 bool isVolatile = I.isVolatile();
3302 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
3303 unsigned Alignment = I.getAlignment();
3304 const MDNode *TBAAInfo = I.getMetadata(LLVMContext::MD_tbaa);
3306 unsigned ChainI = 0;
3307 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3308 // See visitLoad comments.
3309 if (ChainI == MaxParallelChains) {
3310 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3311 MVT::Other, &Chains[0], ChainI);
3315 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT, Ptr,
3316 DAG.getConstant(Offsets[i], PtrVT));
3317 SDValue St = DAG.getStore(Root, getCurDebugLoc(),
3318 SDValue(Src.getNode(), Src.getResNo() + i),
3319 Add, MachinePointerInfo(PtrV, Offsets[i]),
3320 isVolatile, isNonTemporal, Alignment, TBAAInfo);
3321 Chains[ChainI] = St;
3324 SDValue StoreNode = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3325 MVT::Other, &Chains[0], ChainI);
3327 AssignOrderingToNode(StoreNode.getNode());
3328 DAG.setRoot(StoreNode);
3331 static SDValue InsertFenceForAtomic(SDValue Chain, AtomicOrdering Order,
3332 SynchronizationScope Scope,
3333 bool Before, DebugLoc dl,
3335 const TargetLowering &TLI) {
3336 // Fence, if necessary
3338 if (Order == AcquireRelease || Order == SequentiallyConsistent)
3340 else if (Order == Acquire || Order == Monotonic)
3343 if (Order == AcquireRelease)
3345 else if (Order == Release || Order == Monotonic)
3350 Ops[1] = DAG.getConstant(Order, TLI.getPointerTy());
3351 Ops[2] = DAG.getConstant(Scope, TLI.getPointerTy());
3352 return DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops, 3);
3355 void SelectionDAGBuilder::visitAtomicCmpXchg(const AtomicCmpXchgInst &I) {
3356 DebugLoc dl = getCurDebugLoc();
3357 AtomicOrdering Order = I.getOrdering();
3358 SynchronizationScope Scope = I.getSynchScope();
3360 SDValue InChain = getRoot();
3362 if (TLI.getInsertFencesForAtomic())
3363 InChain = InsertFenceForAtomic(InChain, Order, Scope, true, dl,
3367 DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, dl,
3368 getValue(I.getCompareOperand()).getValueType().getSimpleVT(),
3370 getValue(I.getPointerOperand()),
3371 getValue(I.getCompareOperand()),
3372 getValue(I.getNewValOperand()),
3373 MachinePointerInfo(I.getPointerOperand()), 0 /* Alignment */,
3374 TLI.getInsertFencesForAtomic() ? Monotonic : Order,
3377 SDValue OutChain = L.getValue(1);
3379 if (TLI.getInsertFencesForAtomic())
3380 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3384 DAG.setRoot(OutChain);
3387 void SelectionDAGBuilder::visitAtomicRMW(const AtomicRMWInst &I) {
3388 DebugLoc dl = getCurDebugLoc();
3390 switch (I.getOperation()) {
3391 default: llvm_unreachable("Unknown atomicrmw operation");
3392 case AtomicRMWInst::Xchg: NT = ISD::ATOMIC_SWAP; break;
3393 case AtomicRMWInst::Add: NT = ISD::ATOMIC_LOAD_ADD; break;
3394 case AtomicRMWInst::Sub: NT = ISD::ATOMIC_LOAD_SUB; break;
3395 case AtomicRMWInst::And: NT = ISD::ATOMIC_LOAD_AND; break;
3396 case AtomicRMWInst::Nand: NT = ISD::ATOMIC_LOAD_NAND; break;
3397 case AtomicRMWInst::Or: NT = ISD::ATOMIC_LOAD_OR; break;
3398 case AtomicRMWInst::Xor: NT = ISD::ATOMIC_LOAD_XOR; break;
3399 case AtomicRMWInst::Max: NT = ISD::ATOMIC_LOAD_MAX; break;
3400 case AtomicRMWInst::Min: NT = ISD::ATOMIC_LOAD_MIN; break;
3401 case AtomicRMWInst::UMax: NT = ISD::ATOMIC_LOAD_UMAX; break;
3402 case AtomicRMWInst::UMin: NT = ISD::ATOMIC_LOAD_UMIN; break;
3404 AtomicOrdering Order = I.getOrdering();
3405 SynchronizationScope Scope = I.getSynchScope();
3407 SDValue InChain = getRoot();
3409 if (TLI.getInsertFencesForAtomic())
3410 InChain = InsertFenceForAtomic(InChain, Order, Scope, true, dl,
3414 DAG.getAtomic(NT, dl,
3415 getValue(I.getValOperand()).getValueType().getSimpleVT(),
3417 getValue(I.getPointerOperand()),
3418 getValue(I.getValOperand()),
3419 I.getPointerOperand(), 0 /* Alignment */,
3420 TLI.getInsertFencesForAtomic() ? Monotonic : Order,
3423 SDValue OutChain = L.getValue(1);
3425 if (TLI.getInsertFencesForAtomic())
3426 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3430 DAG.setRoot(OutChain);
3433 void SelectionDAGBuilder::visitFence(const FenceInst &I) {
3434 DebugLoc dl = getCurDebugLoc();
3437 Ops[1] = DAG.getConstant(I.getOrdering(), TLI.getPointerTy());
3438 Ops[2] = DAG.getConstant(I.getSynchScope(), TLI.getPointerTy());
3439 DAG.setRoot(DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops, 3));
3442 void SelectionDAGBuilder::visitAtomicLoad(const LoadInst &I) {
3443 DebugLoc dl = getCurDebugLoc();
3444 AtomicOrdering Order = I.getOrdering();
3445 SynchronizationScope Scope = I.getSynchScope();
3447 SDValue InChain = getRoot();
3449 EVT VT = EVT::getEVT(I.getType());
3451 if (I.getAlignment() * 8 < VT.getSizeInBits())
3452 report_fatal_error("Cannot generate unaligned atomic load");
3455 DAG.getAtomic(ISD::ATOMIC_LOAD, dl, VT, VT, InChain,
3456 getValue(I.getPointerOperand()),
3457 I.getPointerOperand(), I.getAlignment(),
3458 TLI.getInsertFencesForAtomic() ? Monotonic : Order,
3461 SDValue OutChain = L.getValue(1);
3463 if (TLI.getInsertFencesForAtomic())
3464 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3468 DAG.setRoot(OutChain);
3471 void SelectionDAGBuilder::visitAtomicStore(const StoreInst &I) {
3472 DebugLoc dl = getCurDebugLoc();
3474 AtomicOrdering Order = I.getOrdering();
3475 SynchronizationScope Scope = I.getSynchScope();
3477 SDValue InChain = getRoot();
3479 EVT VT = EVT::getEVT(I.getValueOperand()->getType());
3481 if (I.getAlignment() * 8 < VT.getSizeInBits())
3482 report_fatal_error("Cannot generate unaligned atomic store");
3484 if (TLI.getInsertFencesForAtomic())
3485 InChain = InsertFenceForAtomic(InChain, Order, Scope, true, dl,
3489 DAG.getAtomic(ISD::ATOMIC_STORE, dl, VT,
3491 getValue(I.getPointerOperand()),
3492 getValue(I.getValueOperand()),
3493 I.getPointerOperand(), I.getAlignment(),
3494 TLI.getInsertFencesForAtomic() ? Monotonic : Order,
3497 if (TLI.getInsertFencesForAtomic())
3498 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3501 DAG.setRoot(OutChain);
3504 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
3506 void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I,
3507 unsigned Intrinsic) {
3508 bool HasChain = !I.doesNotAccessMemory();
3509 bool OnlyLoad = HasChain && I.onlyReadsMemory();
3511 // Build the operand list.
3512 SmallVector<SDValue, 8> Ops;
3513 if (HasChain) { // If this intrinsic has side-effects, chainify it.
3515 // We don't need to serialize loads against other loads.
3516 Ops.push_back(DAG.getRoot());
3518 Ops.push_back(getRoot());
3522 // Info is set by getTgtMemInstrinsic
3523 TargetLowering::IntrinsicInfo Info;
3524 bool IsTgtIntrinsic = TLI.getTgtMemIntrinsic(Info, I, Intrinsic);
3526 // Add the intrinsic ID as an integer operand if it's not a target intrinsic.
3527 if (!IsTgtIntrinsic || Info.opc == ISD::INTRINSIC_VOID ||
3528 Info.opc == ISD::INTRINSIC_W_CHAIN)
3529 Ops.push_back(DAG.getTargetConstant(Intrinsic, TLI.getPointerTy()));
3531 // Add all operands of the call to the operand list.
3532 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) {
3533 SDValue Op = getValue(I.getArgOperand(i));
3537 SmallVector<EVT, 4> ValueVTs;
3538 ComputeValueVTs(TLI, I.getType(), ValueVTs);
3541 ValueVTs.push_back(MVT::Other);
3543 SDVTList VTs = DAG.getVTList(ValueVTs.data(), ValueVTs.size());
3547 if (IsTgtIntrinsic) {
3548 // This is target intrinsic that touches memory
3549 Result = DAG.getMemIntrinsicNode(Info.opc, getCurDebugLoc(),
3550 VTs, &Ops[0], Ops.size(),
3552 MachinePointerInfo(Info.ptrVal, Info.offset),
3553 Info.align, Info.vol,
3554 Info.readMem, Info.writeMem);
3555 } else if (!HasChain) {
3556 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurDebugLoc(),
3557 VTs, &Ops[0], Ops.size());
3558 } else if (!I.getType()->isVoidTy()) {
3559 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurDebugLoc(),
3560 VTs, &Ops[0], Ops.size());
3562 Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurDebugLoc(),
3563 VTs, &Ops[0], Ops.size());
3567 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1);
3569 PendingLoads.push_back(Chain);
3574 if (!I.getType()->isVoidTy()) {
3575 if (VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
3576 EVT VT = TLI.getValueType(PTy);
3577 Result = DAG.getNode(ISD::BITCAST, getCurDebugLoc(), VT, Result);
3580 setValue(&I, Result);
3584 /// GetSignificand - Get the significand and build it into a floating-point
3585 /// number with exponent of 1:
3587 /// Op = (Op & 0x007fffff) | 0x3f800000;
3589 /// where Op is the hexidecimal representation of floating point value.
3591 GetSignificand(SelectionDAG &DAG, SDValue Op, DebugLoc dl) {
3592 SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3593 DAG.getConstant(0x007fffff, MVT::i32));
3594 SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1,
3595 DAG.getConstant(0x3f800000, MVT::i32));
3596 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, t2);
3599 /// GetExponent - Get the exponent:
3601 /// (float)(int)(((Op & 0x7f800000) >> 23) - 127);
3603 /// where Op is the hexidecimal representation of floating point value.
3605 GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI,
3607 SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3608 DAG.getConstant(0x7f800000, MVT::i32));
3609 SDValue t1 = DAG.getNode(ISD::SRL, dl, MVT::i32, t0,
3610 DAG.getConstant(23, TLI.getPointerTy()));
3611 SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1,
3612 DAG.getConstant(127, MVT::i32));
3613 return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2);
3616 /// getF32Constant - Get 32-bit floating point constant.
3618 getF32Constant(SelectionDAG &DAG, unsigned Flt) {
3619 return DAG.getConstantFP(APFloat(APInt(32, Flt)), MVT::f32);
3622 // implVisitAluOverflow - Lower arithmetic overflow instrinsics.
3624 SelectionDAGBuilder::implVisitAluOverflow(const CallInst &I, ISD::NodeType Op) {
3625 SDValue Op1 = getValue(I.getArgOperand(0));
3626 SDValue Op2 = getValue(I.getArgOperand(1));
3628 SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1);
3629 setValue(&I, DAG.getNode(Op, getCurDebugLoc(), VTs, Op1, Op2));
3633 /// visitExp - Lower an exp intrinsic. Handles the special sequences for
3634 /// limited-precision mode.
3636 SelectionDAGBuilder::visitExp(const CallInst &I) {
3638 DebugLoc dl = getCurDebugLoc();
3640 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3641 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3642 SDValue Op = getValue(I.getArgOperand(0));
3644 // Put the exponent in the right bit position for later addition to the
3647 // #define LOG2OFe 1.4426950f
3648 // IntegerPartOfX = ((int32_t)(X * LOG2OFe));
3649 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3650 getF32Constant(DAG, 0x3fb8aa3b));
3651 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3653 // FractionalPartOfX = (X * LOG2OFe) - (float)IntegerPartOfX;
3654 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3655 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3657 // IntegerPartOfX <<= 23;
3658 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3659 DAG.getConstant(23, TLI.getPointerTy()));
3661 if (LimitFloatPrecision <= 6) {
3662 // For floating-point precision of 6:
3664 // TwoToFractionalPartOfX =
3666 // (0.735607626f + 0.252464424f * x) * x;
3668 // error 0.0144103317, which is 6 bits
3669 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3670 getF32Constant(DAG, 0x3e814304));
3671 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3672 getF32Constant(DAG, 0x3f3c50c8));
3673 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3674 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3675 getF32Constant(DAG, 0x3f7f5e7e));
3676 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BITCAST, dl,MVT::i32, t5);
3678 // Add the exponent into the result in integer domain.
3679 SDValue t6 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3680 TwoToFracPartOfX, IntegerPartOfX);
3682 result = DAG.getNode(ISD::BITCAST, dl, MVT::f32, t6);
3683 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3684 // For floating-point precision of 12:
3686 // TwoToFractionalPartOfX =
3689 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3691 // 0.000107046256 error, which is 13 to 14 bits
3692 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3693 getF32Constant(DAG, 0x3da235e3));
3694 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3695 getF32Constant(DAG, 0x3e65b8f3));
3696 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3697 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3698 getF32Constant(DAG, 0x3f324b07));
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, 0x3f7ff8fd));
3702 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BITCAST, dl,MVT::i32, t7);
3704 // Add the exponent into the result in integer domain.
3705 SDValue t8 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3706 TwoToFracPartOfX, IntegerPartOfX);
3708 result = DAG.getNode(ISD::BITCAST, dl, MVT::f32, t8);
3709 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3710 // For floating-point precision of 18:
3712 // TwoToFractionalPartOfX =
3716 // (0.554906021e-1f +
3717 // (0.961591928e-2f +
3718 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3720 // error 2.47208000*10^(-7), which is better than 18 bits
3721 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3722 getF32Constant(DAG, 0x3924b03e));
3723 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3724 getF32Constant(DAG, 0x3ab24b87));
3725 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3726 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3727 getF32Constant(DAG, 0x3c1d8c17));
3728 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3729 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3730 getF32Constant(DAG, 0x3d634a1d));
3731 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3732 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3733 getF32Constant(DAG, 0x3e75fe14));
3734 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3735 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3736 getF32Constant(DAG, 0x3f317234));
3737 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3738 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3739 getF32Constant(DAG, 0x3f800000));
3740 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BITCAST, dl,
3743 // Add the exponent into the result in integer domain.
3744 SDValue t14 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3745 TwoToFracPartOfX, IntegerPartOfX);
3747 result = DAG.getNode(ISD::BITCAST, dl, MVT::f32, t14);
3750 // No special expansion.
3751 result = DAG.getNode(ISD::FEXP, dl,
3752 getValue(I.getArgOperand(0)).getValueType(),
3753 getValue(I.getArgOperand(0)));
3756 setValue(&I, result);
3759 /// visitLog - Lower a log intrinsic. Handles the special sequences for
3760 /// limited-precision mode.
3762 SelectionDAGBuilder::visitLog(const CallInst &I) {
3764 DebugLoc dl = getCurDebugLoc();
3766 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3767 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3768 SDValue Op = getValue(I.getArgOperand(0));
3769 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3771 // Scale the exponent by log(2) [0.69314718f].
3772 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3773 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3774 getF32Constant(DAG, 0x3f317218));
3776 // Get the significand and build it into a floating-point number with
3778 SDValue X = GetSignificand(DAG, Op1, dl);
3780 if (LimitFloatPrecision <= 6) {
3781 // For floating-point precision of 6:
3785 // (1.4034025f - 0.23903021f * x) * x;
3787 // error 0.0034276066, which is better than 8 bits
3788 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3789 getF32Constant(DAG, 0xbe74c456));
3790 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3791 getF32Constant(DAG, 0x3fb3a2b1));
3792 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3793 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3794 getF32Constant(DAG, 0x3f949a29));
3796 result = DAG.getNode(ISD::FADD, dl,
3797 MVT::f32, LogOfExponent, LogOfMantissa);
3798 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3799 // For floating-point precision of 12:
3805 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x;
3807 // error 0.000061011436, which is 14 bits
3808 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3809 getF32Constant(DAG, 0xbd67b6d6));
3810 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3811 getF32Constant(DAG, 0x3ee4f4b8));
3812 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3813 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3814 getF32Constant(DAG, 0x3fbc278b));
3815 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3816 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3817 getF32Constant(DAG, 0x40348e95));
3818 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3819 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3820 getF32Constant(DAG, 0x3fdef31a));
3822 result = DAG.getNode(ISD::FADD, dl,
3823 MVT::f32, LogOfExponent, LogOfMantissa);
3824 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3825 // For floating-point precision of 18:
3833 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x;
3835 // error 0.0000023660568, which is better than 18 bits
3836 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3837 getF32Constant(DAG, 0xbc91e5ac));
3838 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3839 getF32Constant(DAG, 0x3e4350aa));
3840 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3841 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3842 getF32Constant(DAG, 0x3f60d3e3));
3843 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3844 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3845 getF32Constant(DAG, 0x4011cdf0));
3846 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3847 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3848 getF32Constant(DAG, 0x406cfd1c));
3849 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3850 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3851 getF32Constant(DAG, 0x408797cb));
3852 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3853 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3854 getF32Constant(DAG, 0x4006dcab));
3856 result = DAG.getNode(ISD::FADD, dl,
3857 MVT::f32, LogOfExponent, LogOfMantissa);
3860 // No special expansion.
3861 result = DAG.getNode(ISD::FLOG, dl,
3862 getValue(I.getArgOperand(0)).getValueType(),
3863 getValue(I.getArgOperand(0)));
3866 setValue(&I, result);
3869 /// visitLog2 - Lower a log2 intrinsic. Handles the special sequences for
3870 /// limited-precision mode.
3872 SelectionDAGBuilder::visitLog2(const CallInst &I) {
3874 DebugLoc dl = getCurDebugLoc();
3876 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3877 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3878 SDValue Op = getValue(I.getArgOperand(0));
3879 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3881 // Get the exponent.
3882 SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl);
3884 // Get the significand and build it into a floating-point number with
3886 SDValue X = GetSignificand(DAG, Op1, dl);
3888 // Different possible minimax approximations of significand in
3889 // floating-point for various degrees of accuracy over [1,2].
3890 if (LimitFloatPrecision <= 6) {
3891 // For floating-point precision of 6:
3893 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x;
3895 // error 0.0049451742, which is more than 7 bits
3896 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3897 getF32Constant(DAG, 0xbeb08fe0));
3898 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3899 getF32Constant(DAG, 0x40019463));
3900 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3901 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3902 getF32Constant(DAG, 0x3fd6633d));
3904 result = DAG.getNode(ISD::FADD, dl,
3905 MVT::f32, LogOfExponent, Log2ofMantissa);
3906 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3907 // For floating-point precision of 12:
3913 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x;
3915 // error 0.0000876136000, which is better than 13 bits
3916 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3917 getF32Constant(DAG, 0xbda7262e));
3918 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3919 getF32Constant(DAG, 0x3f25280b));
3920 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3921 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3922 getF32Constant(DAG, 0x4007b923));
3923 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3924 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3925 getF32Constant(DAG, 0x40823e2f));
3926 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3927 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3928 getF32Constant(DAG, 0x4020d29c));
3930 result = DAG.getNode(ISD::FADD, dl,
3931 MVT::f32, LogOfExponent, Log2ofMantissa);
3932 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3933 // For floating-point precision of 18:
3942 // 0.25691327e-1f * x) * x) * x) * x) * x) * x;
3944 // error 0.0000018516, which is better than 18 bits
3945 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3946 getF32Constant(DAG, 0xbcd2769e));
3947 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3948 getF32Constant(DAG, 0x3e8ce0b9));
3949 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3950 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3951 getF32Constant(DAG, 0x3fa22ae7));
3952 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3953 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3954 getF32Constant(DAG, 0x40525723));
3955 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3956 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3957 getF32Constant(DAG, 0x40aaf200));
3958 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3959 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3960 getF32Constant(DAG, 0x40c39dad));
3961 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3962 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3963 getF32Constant(DAG, 0x4042902c));
3965 result = DAG.getNode(ISD::FADD, dl,
3966 MVT::f32, LogOfExponent, Log2ofMantissa);
3969 // No special expansion.
3970 result = DAG.getNode(ISD::FLOG2, dl,
3971 getValue(I.getArgOperand(0)).getValueType(),
3972 getValue(I.getArgOperand(0)));
3975 setValue(&I, result);
3978 /// visitLog10 - Lower a log10 intrinsic. Handles the special sequences for
3979 /// limited-precision mode.
3981 SelectionDAGBuilder::visitLog10(const CallInst &I) {
3983 DebugLoc dl = getCurDebugLoc();
3985 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3986 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3987 SDValue Op = getValue(I.getArgOperand(0));
3988 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3990 // Scale the exponent by log10(2) [0.30102999f].
3991 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3992 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3993 getF32Constant(DAG, 0x3e9a209a));
3995 // Get the significand and build it into a floating-point number with
3997 SDValue X = GetSignificand(DAG, Op1, dl);
3999 if (LimitFloatPrecision <= 6) {
4000 // For floating-point precision of 6:
4002 // Log10ofMantissa =
4004 // (0.60948995f - 0.10380950f * x) * x;
4006 // error 0.0014886165, which is 6 bits
4007 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4008 getF32Constant(DAG, 0xbdd49a13));
4009 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4010 getF32Constant(DAG, 0x3f1c0789));
4011 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4012 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4013 getF32Constant(DAG, 0x3f011300));
4015 result = DAG.getNode(ISD::FADD, dl,
4016 MVT::f32, LogOfExponent, Log10ofMantissa);
4017 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
4018 // For floating-point precision of 12:
4020 // Log10ofMantissa =
4023 // (-0.31664806f + 0.47637168e-1f * x) * x) * x;
4025 // error 0.00019228036, which is better than 12 bits
4026 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4027 getF32Constant(DAG, 0x3d431f31));
4028 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
4029 getF32Constant(DAG, 0x3ea21fb2));
4030 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4031 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4032 getF32Constant(DAG, 0x3f6ae232));
4033 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4034 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
4035 getF32Constant(DAG, 0x3f25f7c3));
4037 result = DAG.getNode(ISD::FADD, dl,
4038 MVT::f32, LogOfExponent, Log10ofMantissa);
4039 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
4040 // For floating-point precision of 18:
4042 // Log10ofMantissa =
4047 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x;
4049 // error 0.0000037995730, which is better than 18 bits
4050 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4051 getF32Constant(DAG, 0x3c5d51ce));
4052 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
4053 getF32Constant(DAG, 0x3e00685a));
4054 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4055 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4056 getF32Constant(DAG, 0x3efb6798));
4057 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4058 SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
4059 getF32Constant(DAG, 0x3f88d192));
4060 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4061 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4062 getF32Constant(DAG, 0x3fc4316c));
4063 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4064 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8,
4065 getF32Constant(DAG, 0x3f57ce70));
4067 result = DAG.getNode(ISD::FADD, dl,
4068 MVT::f32, LogOfExponent, Log10ofMantissa);
4071 // No special expansion.
4072 result = DAG.getNode(ISD::FLOG10, dl,
4073 getValue(I.getArgOperand(0)).getValueType(),
4074 getValue(I.getArgOperand(0)));
4077 setValue(&I, result);
4080 /// visitExp2 - Lower an exp2 intrinsic. Handles the special sequences for
4081 /// limited-precision mode.
4083 SelectionDAGBuilder::visitExp2(const CallInst &I) {
4085 DebugLoc dl = getCurDebugLoc();
4087 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
4088 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4089 SDValue Op = getValue(I.getArgOperand(0));
4091 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Op);
4093 // FractionalPartOfX = x - (float)IntegerPartOfX;
4094 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
4095 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, Op, t1);
4097 // IntegerPartOfX <<= 23;
4098 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
4099 DAG.getConstant(23, TLI.getPointerTy()));
4101 if (LimitFloatPrecision <= 6) {
4102 // For floating-point precision of 6:
4104 // TwoToFractionalPartOfX =
4106 // (0.735607626f + 0.252464424f * x) * x;
4108 // error 0.0144103317, which is 6 bits
4109 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4110 getF32Constant(DAG, 0x3e814304));
4111 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4112 getF32Constant(DAG, 0x3f3c50c8));
4113 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4114 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4115 getF32Constant(DAG, 0x3f7f5e7e));
4116 SDValue t6 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t5);
4117 SDValue TwoToFractionalPartOfX =
4118 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
4120 result = DAG.getNode(ISD::BITCAST, dl,
4121 MVT::f32, TwoToFractionalPartOfX);
4122 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
4123 // For floating-point precision of 12:
4125 // TwoToFractionalPartOfX =
4128 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
4130 // error 0.000107046256, which is 13 to 14 bits
4131 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4132 getF32Constant(DAG, 0x3da235e3));
4133 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4134 getF32Constant(DAG, 0x3e65b8f3));
4135 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4136 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4137 getF32Constant(DAG, 0x3f324b07));
4138 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4139 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4140 getF32Constant(DAG, 0x3f7ff8fd));
4141 SDValue t8 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t7);
4142 SDValue TwoToFractionalPartOfX =
4143 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
4145 result = DAG.getNode(ISD::BITCAST, dl,
4146 MVT::f32, TwoToFractionalPartOfX);
4147 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
4148 // For floating-point precision of 18:
4150 // TwoToFractionalPartOfX =
4154 // (0.554906021e-1f +
4155 // (0.961591928e-2f +
4156 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
4157 // error 2.47208000*10^(-7), which is better than 18 bits
4158 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4159 getF32Constant(DAG, 0x3924b03e));
4160 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4161 getF32Constant(DAG, 0x3ab24b87));
4162 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4163 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4164 getF32Constant(DAG, 0x3c1d8c17));
4165 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4166 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4167 getF32Constant(DAG, 0x3d634a1d));
4168 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4169 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4170 getF32Constant(DAG, 0x3e75fe14));
4171 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4172 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
4173 getF32Constant(DAG, 0x3f317234));
4174 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
4175 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
4176 getF32Constant(DAG, 0x3f800000));
4177 SDValue t14 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t13);
4178 SDValue TwoToFractionalPartOfX =
4179 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
4181 result = DAG.getNode(ISD::BITCAST, dl,
4182 MVT::f32, TwoToFractionalPartOfX);
4185 // No special expansion.
4186 result = DAG.getNode(ISD::FEXP2, dl,
4187 getValue(I.getArgOperand(0)).getValueType(),
4188 getValue(I.getArgOperand(0)));
4191 setValue(&I, result);
4194 /// visitPow - Lower a pow intrinsic. Handles the special sequences for
4195 /// limited-precision mode with x == 10.0f.
4197 SelectionDAGBuilder::visitPow(const CallInst &I) {
4199 const Value *Val = I.getArgOperand(0);
4200 DebugLoc dl = getCurDebugLoc();
4201 bool IsExp10 = false;
4203 if (getValue(Val).getValueType() == MVT::f32 &&
4204 getValue(I.getArgOperand(1)).getValueType() == MVT::f32 &&
4205 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4206 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(Val))) {
4207 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
4209 IsExp10 = CFP->getValueAPF().bitwiseIsEqual(Ten);
4214 if (IsExp10 && LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4215 SDValue Op = getValue(I.getArgOperand(1));
4217 // Put the exponent in the right bit position for later addition to the
4220 // #define LOG2OF10 3.3219281f
4221 // IntegerPartOfX = (int32_t)(x * LOG2OF10);
4222 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
4223 getF32Constant(DAG, 0x40549a78));
4224 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
4226 // FractionalPartOfX = x - (float)IntegerPartOfX;
4227 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
4228 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
4230 // IntegerPartOfX <<= 23;
4231 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
4232 DAG.getConstant(23, TLI.getPointerTy()));
4234 if (LimitFloatPrecision <= 6) {
4235 // For floating-point precision of 6:
4237 // twoToFractionalPartOfX =
4239 // (0.735607626f + 0.252464424f * x) * x;
4241 // error 0.0144103317, which is 6 bits
4242 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4243 getF32Constant(DAG, 0x3e814304));
4244 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4245 getF32Constant(DAG, 0x3f3c50c8));
4246 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4247 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4248 getF32Constant(DAG, 0x3f7f5e7e));
4249 SDValue t6 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t5);
4250 SDValue TwoToFractionalPartOfX =
4251 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
4253 result = DAG.getNode(ISD::BITCAST, dl,
4254 MVT::f32, TwoToFractionalPartOfX);
4255 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
4256 // For floating-point precision of 12:
4258 // TwoToFractionalPartOfX =
4261 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
4263 // error 0.000107046256, which is 13 to 14 bits
4264 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4265 getF32Constant(DAG, 0x3da235e3));
4266 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4267 getF32Constant(DAG, 0x3e65b8f3));
4268 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4269 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4270 getF32Constant(DAG, 0x3f324b07));
4271 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4272 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4273 getF32Constant(DAG, 0x3f7ff8fd));
4274 SDValue t8 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t7);
4275 SDValue TwoToFractionalPartOfX =
4276 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
4278 result = DAG.getNode(ISD::BITCAST, dl,
4279 MVT::f32, TwoToFractionalPartOfX);
4280 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
4281 // For floating-point precision of 18:
4283 // TwoToFractionalPartOfX =
4287 // (0.554906021e-1f +
4288 // (0.961591928e-2f +
4289 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
4290 // error 2.47208000*10^(-7), which is better than 18 bits
4291 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4292 getF32Constant(DAG, 0x3924b03e));
4293 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4294 getF32Constant(DAG, 0x3ab24b87));
4295 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4296 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4297 getF32Constant(DAG, 0x3c1d8c17));
4298 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4299 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4300 getF32Constant(DAG, 0x3d634a1d));
4301 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4302 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4303 getF32Constant(DAG, 0x3e75fe14));
4304 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4305 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
4306 getF32Constant(DAG, 0x3f317234));
4307 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
4308 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
4309 getF32Constant(DAG, 0x3f800000));
4310 SDValue t14 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t13);
4311 SDValue TwoToFractionalPartOfX =
4312 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
4314 result = DAG.getNode(ISD::BITCAST, dl,
4315 MVT::f32, TwoToFractionalPartOfX);
4318 // No special expansion.
4319 result = DAG.getNode(ISD::FPOW, dl,
4320 getValue(I.getArgOperand(0)).getValueType(),
4321 getValue(I.getArgOperand(0)),
4322 getValue(I.getArgOperand(1)));
4325 setValue(&I, result);
4329 /// ExpandPowI - Expand a llvm.powi intrinsic.
4330 static SDValue ExpandPowI(DebugLoc DL, SDValue LHS, SDValue RHS,
4331 SelectionDAG &DAG) {
4332 // If RHS is a constant, we can expand this out to a multiplication tree,
4333 // otherwise we end up lowering to a call to __powidf2 (for example). When
4334 // optimizing for size, we only want to do this if the expansion would produce
4335 // a small number of multiplies, otherwise we do the full expansion.
4336 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
4337 // Get the exponent as a positive value.
4338 unsigned Val = RHSC->getSExtValue();
4339 if ((int)Val < 0) Val = -Val;
4341 // powi(x, 0) -> 1.0
4343 return DAG.getConstantFP(1.0, LHS.getValueType());
4345 const Function *F = DAG.getMachineFunction().getFunction();
4346 if (!F->hasFnAttr(Attribute::OptimizeForSize) ||
4347 // If optimizing for size, don't insert too many multiplies. This
4348 // inserts up to 5 multiplies.
4349 CountPopulation_32(Val)+Log2_32(Val) < 7) {
4350 // We use the simple binary decomposition method to generate the multiply
4351 // sequence. There are more optimal ways to do this (for example,
4352 // powi(x,15) generates one more multiply than it should), but this has
4353 // the benefit of being both really simple and much better than a libcall.
4354 SDValue Res; // Logically starts equal to 1.0
4355 SDValue CurSquare = LHS;
4359 Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare);
4361 Res = CurSquare; // 1.0*CurSquare.
4364 CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(),
4365 CurSquare, CurSquare);
4369 // If the original was negative, invert the result, producing 1/(x*x*x).
4370 if (RHSC->getSExtValue() < 0)
4371 Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(),
4372 DAG.getConstantFP(1.0, LHS.getValueType()), Res);
4377 // Otherwise, expand to a libcall.
4378 return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS);
4381 // getTruncatedArgReg - Find underlying register used for an truncated
4383 static unsigned getTruncatedArgReg(const SDValue &N) {
4384 if (N.getOpcode() != ISD::TRUNCATE)
4387 const SDValue &Ext = N.getOperand(0);
4388 if (Ext.getOpcode() == ISD::AssertZext || Ext.getOpcode() == ISD::AssertSext){
4389 const SDValue &CFR = Ext.getOperand(0);
4390 if (CFR.getOpcode() == ISD::CopyFromReg)
4391 return cast<RegisterSDNode>(CFR.getOperand(1))->getReg();
4393 if (CFR.getOpcode() == ISD::TRUNCATE)
4394 return getTruncatedArgReg(CFR);
4399 /// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a function
4400 /// argument, create the corresponding DBG_VALUE machine instruction for it now.
4401 /// At the end of instruction selection, they will be inserted to the entry BB.
4403 SelectionDAGBuilder::EmitFuncArgumentDbgValue(const Value *V, MDNode *Variable,
4406 const Argument *Arg = dyn_cast<Argument>(V);
4410 MachineFunction &MF = DAG.getMachineFunction();
4411 const TargetInstrInfo *TII = DAG.getTarget().getInstrInfo();
4412 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
4414 // Ignore inlined function arguments here.
4415 DIVariable DV(Variable);
4416 if (DV.isInlinedFnArgument(MF.getFunction()))
4420 // Some arguments' frame index is recorded during argument lowering.
4421 Offset = FuncInfo.getArgumentFrameIndex(Arg);
4423 Reg = TRI->getFrameRegister(MF);
4425 if (!Reg && N.getNode()) {
4426 if (N.getOpcode() == ISD::CopyFromReg)
4427 Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg();
4429 Reg = getTruncatedArgReg(N);
4430 if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) {
4431 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4432 unsigned PR = RegInfo.getLiveInPhysReg(Reg);
4439 // Check if ValueMap has reg number.
4440 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
4441 if (VMI != FuncInfo.ValueMap.end())
4445 if (!Reg && N.getNode()) {
4446 // Check if frame index is available.
4447 if (LoadSDNode *LNode = dyn_cast<LoadSDNode>(N.getNode()))
4448 if (FrameIndexSDNode *FINode =
4449 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode())) {
4450 Reg = TRI->getFrameRegister(MF);
4451 Offset = FINode->getIndex();
4458 MachineInstrBuilder MIB = BuildMI(MF, getCurDebugLoc(),
4459 TII->get(TargetOpcode::DBG_VALUE))
4460 .addReg(Reg, RegState::Debug).addImm(Offset).addMetadata(Variable);
4461 FuncInfo.ArgDbgValues.push_back(&*MIB);
4465 // VisualStudio defines setjmp as _setjmp
4466 #if defined(_MSC_VER) && defined(setjmp) && \
4467 !defined(setjmp_undefined_for_msvc)
4468 # pragma push_macro("setjmp")
4470 # define setjmp_undefined_for_msvc
4473 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
4474 /// we want to emit this as a call to a named external function, return the name
4475 /// otherwise lower it and return null.
4477 SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) {
4478 DebugLoc dl = getCurDebugLoc();
4481 switch (Intrinsic) {
4483 // By default, turn this into a target intrinsic node.
4484 visitTargetIntrinsic(I, Intrinsic);
4486 case Intrinsic::vastart: visitVAStart(I); return 0;
4487 case Intrinsic::vaend: visitVAEnd(I); return 0;
4488 case Intrinsic::vacopy: visitVACopy(I); return 0;
4489 case Intrinsic::returnaddress:
4490 setValue(&I, DAG.getNode(ISD::RETURNADDR, dl, TLI.getPointerTy(),
4491 getValue(I.getArgOperand(0))));
4493 case Intrinsic::frameaddress:
4494 setValue(&I, DAG.getNode(ISD::FRAMEADDR, dl, TLI.getPointerTy(),
4495 getValue(I.getArgOperand(0))));
4497 case Intrinsic::setjmp:
4498 return "_setjmp"+!TLI.usesUnderscoreSetJmp();
4499 case Intrinsic::longjmp:
4500 return "_longjmp"+!TLI.usesUnderscoreLongJmp();
4501 case Intrinsic::memcpy: {
4502 // Assert for address < 256 since we support only user defined address
4504 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4506 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4508 "Unknown address space");
4509 SDValue Op1 = getValue(I.getArgOperand(0));
4510 SDValue Op2 = getValue(I.getArgOperand(1));
4511 SDValue Op3 = getValue(I.getArgOperand(2));
4512 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4513 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4514 DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, isVol, false,
4515 MachinePointerInfo(I.getArgOperand(0)),
4516 MachinePointerInfo(I.getArgOperand(1))));
4519 case Intrinsic::memset: {
4520 // Assert for address < 256 since we support only user defined address
4522 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4524 "Unknown address space");
4525 SDValue Op1 = getValue(I.getArgOperand(0));
4526 SDValue Op2 = getValue(I.getArgOperand(1));
4527 SDValue Op3 = getValue(I.getArgOperand(2));
4528 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4529 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4530 DAG.setRoot(DAG.getMemset(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4531 MachinePointerInfo(I.getArgOperand(0))));
4534 case Intrinsic::memmove: {
4535 // Assert for address < 256 since we support only user defined address
4537 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4539 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4541 "Unknown address space");
4542 SDValue Op1 = getValue(I.getArgOperand(0));
4543 SDValue Op2 = getValue(I.getArgOperand(1));
4544 SDValue Op3 = getValue(I.getArgOperand(2));
4545 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4546 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4547 DAG.setRoot(DAG.getMemmove(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4548 MachinePointerInfo(I.getArgOperand(0)),
4549 MachinePointerInfo(I.getArgOperand(1))));
4552 case Intrinsic::dbg_declare: {
4553 const DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
4554 MDNode *Variable = DI.getVariable();
4555 const Value *Address = DI.getAddress();
4556 if (!Address || !DIVariable(Variable).Verify())
4559 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder
4560 // but do not always have a corresponding SDNode built. The SDNodeOrder
4561 // absolute, but not relative, values are different depending on whether
4562 // debug info exists.
4565 // Check if address has undef value.
4566 if (isa<UndefValue>(Address) ||
4567 (Address->use_empty() && !isa<Argument>(Address))) {
4568 DEBUG(dbgs() << "Dropping debug info for " << DI);
4572 SDValue &N = NodeMap[Address];
4573 if (!N.getNode() && isa<Argument>(Address))
4574 // Check unused arguments map.
4575 N = UnusedArgNodeMap[Address];
4578 // Parameters are handled specially.
4580 DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable;
4581 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
4582 Address = BCI->getOperand(0);
4583 const AllocaInst *AI = dyn_cast<AllocaInst>(Address);
4585 if (isParameter && !AI) {
4586 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode());
4588 // Byval parameter. We have a frame index at this point.
4589 SDV = DAG.getDbgValue(Variable, FINode->getIndex(),
4590 0, dl, SDNodeOrder);
4592 // Address is an argument, so try to emit its dbg value using
4593 // virtual register info from the FuncInfo.ValueMap.
4594 EmitFuncArgumentDbgValue(Address, Variable, 0, N);
4598 SDV = DAG.getDbgValue(Variable, N.getNode(), N.getResNo(),
4599 0, dl, SDNodeOrder);
4601 // Can't do anything with other non-AI cases yet.
4602 DEBUG(dbgs() << "Dropping debug info for " << DI);
4605 DAG.AddDbgValue(SDV, N.getNode(), isParameter);
4607 // If Address is an argument then try to emit its dbg value using
4608 // virtual register info from the FuncInfo.ValueMap.
4609 if (!EmitFuncArgumentDbgValue(Address, Variable, 0, N)) {
4610 // If variable is pinned by a alloca in dominating bb then
4611 // use StaticAllocaMap.
4612 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Address)) {
4613 if (AI->getParent() != DI.getParent()) {
4614 DenseMap<const AllocaInst*, int>::iterator SI =
4615 FuncInfo.StaticAllocaMap.find(AI);
4616 if (SI != FuncInfo.StaticAllocaMap.end()) {
4617 SDV = DAG.getDbgValue(Variable, SI->second,
4618 0, dl, SDNodeOrder);
4619 DAG.AddDbgValue(SDV, 0, false);
4624 DEBUG(dbgs() << "Dropping debug info for " << DI);
4629 case Intrinsic::dbg_value: {
4630 const DbgValueInst &DI = cast<DbgValueInst>(I);
4631 if (!DIVariable(DI.getVariable()).Verify())
4634 MDNode *Variable = DI.getVariable();
4635 uint64_t Offset = DI.getOffset();
4636 const Value *V = DI.getValue();
4640 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder
4641 // but do not always have a corresponding SDNode built. The SDNodeOrder
4642 // absolute, but not relative, values are different depending on whether
4643 // debug info exists.
4646 if (isa<ConstantInt>(V) || isa<ConstantFP>(V) || isa<UndefValue>(V)) {
4647 SDV = DAG.getDbgValue(Variable, V, Offset, dl, SDNodeOrder);
4648 DAG.AddDbgValue(SDV, 0, false);
4650 // Do not use getValue() in here; we don't want to generate code at
4651 // this point if it hasn't been done yet.
4652 SDValue N = NodeMap[V];
4653 if (!N.getNode() && isa<Argument>(V))
4654 // Check unused arguments map.
4655 N = UnusedArgNodeMap[V];
4657 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, N)) {
4658 SDV = DAG.getDbgValue(Variable, N.getNode(),
4659 N.getResNo(), Offset, dl, SDNodeOrder);
4660 DAG.AddDbgValue(SDV, N.getNode(), false);
4662 } else if (!V->use_empty() ) {
4663 // Do not call getValue(V) yet, as we don't want to generate code.
4664 // Remember it for later.
4665 DanglingDebugInfo DDI(&DI, dl, SDNodeOrder);
4666 DanglingDebugInfoMap[V] = DDI;
4668 // We may expand this to cover more cases. One case where we have no
4669 // data available is an unreferenced parameter.
4670 DEBUG(dbgs() << "Dropping debug info for " << DI);
4674 // Build a debug info table entry.
4675 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
4676 V = BCI->getOperand(0);
4677 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
4678 // Don't handle byval struct arguments or VLAs, for example.
4681 DenseMap<const AllocaInst*, int>::iterator SI =
4682 FuncInfo.StaticAllocaMap.find(AI);
4683 if (SI == FuncInfo.StaticAllocaMap.end())
4685 int FI = SI->second;
4687 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4688 if (!DI.getDebugLoc().isUnknown() && MMI.hasDebugInfo())
4689 MMI.setVariableDbgInfo(Variable, FI, DI.getDebugLoc());
4693 case Intrinsic::eh_typeid_for: {
4694 // Find the type id for the given typeinfo.
4695 GlobalVariable *GV = ExtractTypeInfo(I.getArgOperand(0));
4696 unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(GV);
4697 Res = DAG.getConstant(TypeID, MVT::i32);
4702 case Intrinsic::eh_return_i32:
4703 case Intrinsic::eh_return_i64:
4704 DAG.getMachineFunction().getMMI().setCallsEHReturn(true);
4705 DAG.setRoot(DAG.getNode(ISD::EH_RETURN, dl,
4708 getValue(I.getArgOperand(0)),
4709 getValue(I.getArgOperand(1))));
4711 case Intrinsic::eh_unwind_init:
4712 DAG.getMachineFunction().getMMI().setCallsUnwindInit(true);
4714 case Intrinsic::eh_dwarf_cfa: {
4715 SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getArgOperand(0)), dl,
4716 TLI.getPointerTy());
4717 SDValue Offset = DAG.getNode(ISD::ADD, dl,
4719 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, dl,
4720 TLI.getPointerTy()),
4722 SDValue FA = DAG.getNode(ISD::FRAMEADDR, dl,
4724 DAG.getConstant(0, TLI.getPointerTy()));
4725 setValue(&I, DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
4729 case Intrinsic::eh_sjlj_callsite: {
4730 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4731 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(0));
4732 assert(CI && "Non-constant call site value in eh.sjlj.callsite!");
4733 assert(MMI.getCurrentCallSite() == 0 && "Overlapping call sites!");
4735 MMI.setCurrentCallSite(CI->getZExtValue());
4738 case Intrinsic::eh_sjlj_functioncontext: {
4739 // Get and store the index of the function context.
4740 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4742 cast<AllocaInst>(I.getArgOperand(0)->stripPointerCasts());
4743 int FI = FuncInfo.StaticAllocaMap[FnCtx];
4744 MFI->setFunctionContextIndex(FI);
4747 case Intrinsic::eh_sjlj_setjmp: {
4750 Ops[1] = getValue(I.getArgOperand(0));
4751 SDValue Op = DAG.getNode(ISD::EH_SJLJ_SETJMP, dl,
4752 DAG.getVTList(MVT::i32, MVT::Other),
4754 setValue(&I, Op.getValue(0));
4755 DAG.setRoot(Op.getValue(1));
4758 case Intrinsic::eh_sjlj_longjmp: {
4759 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, dl, MVT::Other,
4760 getRoot(), getValue(I.getArgOperand(0))));
4764 case Intrinsic::x86_mmx_pslli_w:
4765 case Intrinsic::x86_mmx_pslli_d:
4766 case Intrinsic::x86_mmx_pslli_q:
4767 case Intrinsic::x86_mmx_psrli_w:
4768 case Intrinsic::x86_mmx_psrli_d:
4769 case Intrinsic::x86_mmx_psrli_q:
4770 case Intrinsic::x86_mmx_psrai_w:
4771 case Intrinsic::x86_mmx_psrai_d: {
4772 SDValue ShAmt = getValue(I.getArgOperand(1));
4773 if (isa<ConstantSDNode>(ShAmt)) {
4774 visitTargetIntrinsic(I, Intrinsic);
4777 unsigned NewIntrinsic = 0;
4778 EVT ShAmtVT = MVT::v2i32;
4779 switch (Intrinsic) {
4780 case Intrinsic::x86_mmx_pslli_w:
4781 NewIntrinsic = Intrinsic::x86_mmx_psll_w;
4783 case Intrinsic::x86_mmx_pslli_d:
4784 NewIntrinsic = Intrinsic::x86_mmx_psll_d;
4786 case Intrinsic::x86_mmx_pslli_q:
4787 NewIntrinsic = Intrinsic::x86_mmx_psll_q;
4789 case Intrinsic::x86_mmx_psrli_w:
4790 NewIntrinsic = Intrinsic::x86_mmx_psrl_w;
4792 case Intrinsic::x86_mmx_psrli_d:
4793 NewIntrinsic = Intrinsic::x86_mmx_psrl_d;
4795 case Intrinsic::x86_mmx_psrli_q:
4796 NewIntrinsic = Intrinsic::x86_mmx_psrl_q;
4798 case Intrinsic::x86_mmx_psrai_w:
4799 NewIntrinsic = Intrinsic::x86_mmx_psra_w;
4801 case Intrinsic::x86_mmx_psrai_d:
4802 NewIntrinsic = Intrinsic::x86_mmx_psra_d;
4804 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4807 // The vector shift intrinsics with scalars uses 32b shift amounts but
4808 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
4810 // We must do this early because v2i32 is not a legal type.
4811 DebugLoc dl = getCurDebugLoc();
4814 ShOps[1] = DAG.getConstant(0, MVT::i32);
4815 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
4816 EVT DestVT = TLI.getValueType(I.getType());
4817 ShAmt = DAG.getNode(ISD::BITCAST, dl, DestVT, ShAmt);
4818 Res = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4819 DAG.getConstant(NewIntrinsic, MVT::i32),
4820 getValue(I.getArgOperand(0)), ShAmt);
4824 case Intrinsic::convertff:
4825 case Intrinsic::convertfsi:
4826 case Intrinsic::convertfui:
4827 case Intrinsic::convertsif:
4828 case Intrinsic::convertuif:
4829 case Intrinsic::convertss:
4830 case Intrinsic::convertsu:
4831 case Intrinsic::convertus:
4832 case Intrinsic::convertuu: {
4833 ISD::CvtCode Code = ISD::CVT_INVALID;
4834 switch (Intrinsic) {
4835 case Intrinsic::convertff: Code = ISD::CVT_FF; break;
4836 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break;
4837 case Intrinsic::convertfui: Code = ISD::CVT_FU; break;
4838 case Intrinsic::convertsif: Code = ISD::CVT_SF; break;
4839 case Intrinsic::convertuif: Code = ISD::CVT_UF; break;
4840 case Intrinsic::convertss: Code = ISD::CVT_SS; break;
4841 case Intrinsic::convertsu: Code = ISD::CVT_SU; break;
4842 case Intrinsic::convertus: Code = ISD::CVT_US; break;
4843 case Intrinsic::convertuu: Code = ISD::CVT_UU; break;
4845 EVT DestVT = TLI.getValueType(I.getType());
4846 const Value *Op1 = I.getArgOperand(0);
4847 Res = DAG.getConvertRndSat(DestVT, getCurDebugLoc(), getValue(Op1),
4848 DAG.getValueType(DestVT),
4849 DAG.getValueType(getValue(Op1).getValueType()),
4850 getValue(I.getArgOperand(1)),
4851 getValue(I.getArgOperand(2)),
4856 case Intrinsic::sqrt:
4857 setValue(&I, DAG.getNode(ISD::FSQRT, dl,
4858 getValue(I.getArgOperand(0)).getValueType(),
4859 getValue(I.getArgOperand(0))));
4861 case Intrinsic::powi:
4862 setValue(&I, ExpandPowI(dl, getValue(I.getArgOperand(0)),
4863 getValue(I.getArgOperand(1)), DAG));
4865 case Intrinsic::sin:
4866 setValue(&I, DAG.getNode(ISD::FSIN, dl,
4867 getValue(I.getArgOperand(0)).getValueType(),
4868 getValue(I.getArgOperand(0))));
4870 case Intrinsic::cos:
4871 setValue(&I, DAG.getNode(ISD::FCOS, dl,
4872 getValue(I.getArgOperand(0)).getValueType(),
4873 getValue(I.getArgOperand(0))));
4875 case Intrinsic::log:
4878 case Intrinsic::log2:
4881 case Intrinsic::log10:
4884 case Intrinsic::exp:
4887 case Intrinsic::exp2:
4890 case Intrinsic::pow:
4893 case Intrinsic::fma:
4894 setValue(&I, DAG.getNode(ISD::FMA, dl,
4895 getValue(I.getArgOperand(0)).getValueType(),
4896 getValue(I.getArgOperand(0)),
4897 getValue(I.getArgOperand(1)),
4898 getValue(I.getArgOperand(2))));
4900 case Intrinsic::convert_to_fp16:
4901 setValue(&I, DAG.getNode(ISD::FP32_TO_FP16, dl,
4902 MVT::i16, getValue(I.getArgOperand(0))));
4904 case Intrinsic::convert_from_fp16:
4905 setValue(&I, DAG.getNode(ISD::FP16_TO_FP32, dl,
4906 MVT::f32, getValue(I.getArgOperand(0))));
4908 case Intrinsic::pcmarker: {
4909 SDValue Tmp = getValue(I.getArgOperand(0));
4910 DAG.setRoot(DAG.getNode(ISD::PCMARKER, dl, MVT::Other, getRoot(), Tmp));
4913 case Intrinsic::readcyclecounter: {
4914 SDValue Op = getRoot();
4915 Res = DAG.getNode(ISD::READCYCLECOUNTER, dl,
4916 DAG.getVTList(MVT::i64, MVT::Other),
4919 DAG.setRoot(Res.getValue(1));
4922 case Intrinsic::bswap:
4923 setValue(&I, DAG.getNode(ISD::BSWAP, dl,
4924 getValue(I.getArgOperand(0)).getValueType(),
4925 getValue(I.getArgOperand(0))));
4927 case Intrinsic::cttz: {
4928 SDValue Arg = getValue(I.getArgOperand(0));
4929 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1));
4930 EVT Ty = Arg.getValueType();
4931 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTTZ : ISD::CTTZ_ZERO_UNDEF,
4935 case Intrinsic::ctlz: {
4936 SDValue Arg = getValue(I.getArgOperand(0));
4937 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1));
4938 EVT Ty = Arg.getValueType();
4939 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTLZ : ISD::CTLZ_ZERO_UNDEF,
4943 case Intrinsic::ctpop: {
4944 SDValue Arg = getValue(I.getArgOperand(0));
4945 EVT Ty = Arg.getValueType();
4946 setValue(&I, DAG.getNode(ISD::CTPOP, dl, Ty, Arg));
4949 case Intrinsic::stacksave: {
4950 SDValue Op = getRoot();
4951 Res = DAG.getNode(ISD::STACKSAVE, dl,
4952 DAG.getVTList(TLI.getPointerTy(), MVT::Other), &Op, 1);
4954 DAG.setRoot(Res.getValue(1));
4957 case Intrinsic::stackrestore: {
4958 Res = getValue(I.getArgOperand(0));
4959 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, dl, MVT::Other, getRoot(), Res));
4962 case Intrinsic::stackprotector: {
4963 // Emit code into the DAG to store the stack guard onto the stack.
4964 MachineFunction &MF = DAG.getMachineFunction();
4965 MachineFrameInfo *MFI = MF.getFrameInfo();
4966 EVT PtrTy = TLI.getPointerTy();
4968 SDValue Src = getValue(I.getArgOperand(0)); // The guard's value.
4969 AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
4971 int FI = FuncInfo.StaticAllocaMap[Slot];
4972 MFI->setStackProtectorIndex(FI);
4974 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
4976 // Store the stack protector onto the stack.
4977 Res = DAG.getStore(getRoot(), getCurDebugLoc(), Src, FIN,
4978 MachinePointerInfo::getFixedStack(FI),
4984 case Intrinsic::objectsize: {
4985 // If we don't know by now, we're never going to know.
4986 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1));
4988 assert(CI && "Non-constant type in __builtin_object_size?");
4990 SDValue Arg = getValue(I.getCalledValue());
4991 EVT Ty = Arg.getValueType();
4994 Res = DAG.getConstant(-1ULL, Ty);
4996 Res = DAG.getConstant(0, Ty);
5001 case Intrinsic::var_annotation:
5002 // Discard annotate attributes
5005 case Intrinsic::init_trampoline: {
5006 const Function *F = cast<Function>(I.getArgOperand(1)->stripPointerCasts());
5010 Ops[1] = getValue(I.getArgOperand(0));
5011 Ops[2] = getValue(I.getArgOperand(1));
5012 Ops[3] = getValue(I.getArgOperand(2));
5013 Ops[4] = DAG.getSrcValue(I.getArgOperand(0));
5014 Ops[5] = DAG.getSrcValue(F);
5016 Res = DAG.getNode(ISD::INIT_TRAMPOLINE, dl, MVT::Other, Ops, 6);
5021 case Intrinsic::adjust_trampoline: {
5022 setValue(&I, DAG.getNode(ISD::ADJUST_TRAMPOLINE, dl,
5024 getValue(I.getArgOperand(0))));
5027 case Intrinsic::gcroot:
5029 const Value *Alloca = I.getArgOperand(0);
5030 const Constant *TypeMap = cast<Constant>(I.getArgOperand(1));
5032 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode());
5033 GFI->addStackRoot(FI->getIndex(), TypeMap);
5036 case Intrinsic::gcread:
5037 case Intrinsic::gcwrite:
5038 llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!");
5039 case Intrinsic::flt_rounds:
5040 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, dl, MVT::i32));
5043 case Intrinsic::expect: {
5044 // Just replace __builtin_expect(exp, c) with EXP.
5045 setValue(&I, getValue(I.getArgOperand(0)));
5049 case Intrinsic::trap: {
5050 StringRef TrapFuncName = TM.Options.getTrapFunctionName();
5051 if (TrapFuncName.empty()) {
5052 DAG.setRoot(DAG.getNode(ISD::TRAP, dl,MVT::Other, getRoot()));
5055 TargetLowering::ArgListTy Args;
5056 std::pair<SDValue, SDValue> Result =
5057 TLI.LowerCallTo(getRoot(), I.getType(),
5058 false, false, false, false, 0, CallingConv::C,
5059 /*isTailCall=*/false, /*isReturnValueUsed=*/true,
5060 DAG.getExternalSymbol(TrapFuncName.data(), TLI.getPointerTy()),
5061 Args, DAG, getCurDebugLoc());
5062 DAG.setRoot(Result.second);
5065 case Intrinsic::uadd_with_overflow:
5066 return implVisitAluOverflow(I, ISD::UADDO);
5067 case Intrinsic::sadd_with_overflow:
5068 return implVisitAluOverflow(I, ISD::SADDO);
5069 case Intrinsic::usub_with_overflow:
5070 return implVisitAluOverflow(I, ISD::USUBO);
5071 case Intrinsic::ssub_with_overflow:
5072 return implVisitAluOverflow(I, ISD::SSUBO);
5073 case Intrinsic::umul_with_overflow:
5074 return implVisitAluOverflow(I, ISD::UMULO);
5075 case Intrinsic::smul_with_overflow:
5076 return implVisitAluOverflow(I, ISD::SMULO);
5078 case Intrinsic::prefetch: {
5080 unsigned rw = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
5082 Ops[1] = getValue(I.getArgOperand(0));
5083 Ops[2] = getValue(I.getArgOperand(1));
5084 Ops[3] = getValue(I.getArgOperand(2));
5085 Ops[4] = getValue(I.getArgOperand(3));
5086 DAG.setRoot(DAG.getMemIntrinsicNode(ISD::PREFETCH, dl,
5087 DAG.getVTList(MVT::Other),
5089 EVT::getIntegerVT(*Context, 8),
5090 MachinePointerInfo(I.getArgOperand(0)),
5092 false, /* volatile */
5094 rw==1)); /* write */
5098 case Intrinsic::invariant_start:
5099 case Intrinsic::lifetime_start:
5100 // Discard region information.
5101 setValue(&I, DAG.getUNDEF(TLI.getPointerTy()));
5103 case Intrinsic::invariant_end:
5104 case Intrinsic::lifetime_end:
5105 // Discard region information.
5110 void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee,
5112 MachineBasicBlock *LandingPad) {
5113 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
5114 FunctionType *FTy = cast<FunctionType>(PT->getElementType());
5115 Type *RetTy = FTy->getReturnType();
5116 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5117 MCSymbol *BeginLabel = 0;
5119 TargetLowering::ArgListTy Args;
5120 TargetLowering::ArgListEntry Entry;
5121 Args.reserve(CS.arg_size());
5123 // Check whether the function can return without sret-demotion.
5124 SmallVector<ISD::OutputArg, 4> Outs;
5125 SmallVector<uint64_t, 4> Offsets;
5126 GetReturnInfo(RetTy, CS.getAttributes().getRetAttributes(),
5127 Outs, TLI, &Offsets);
5129 bool CanLowerReturn = TLI.CanLowerReturn(CS.getCallingConv(),
5130 DAG.getMachineFunction(),
5131 FTy->isVarArg(), Outs,
5134 SDValue DemoteStackSlot;
5135 int DemoteStackIdx = -100;
5137 if (!CanLowerReturn) {
5138 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(
5139 FTy->getReturnType());
5140 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(
5141 FTy->getReturnType());
5142 MachineFunction &MF = DAG.getMachineFunction();
5143 DemoteStackIdx = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
5144 Type *StackSlotPtrType = PointerType::getUnqual(FTy->getReturnType());
5146 DemoteStackSlot = DAG.getFrameIndex(DemoteStackIdx, TLI.getPointerTy());
5147 Entry.Node = DemoteStackSlot;
5148 Entry.Ty = StackSlotPtrType;
5149 Entry.isSExt = false;
5150 Entry.isZExt = false;
5151 Entry.isInReg = false;
5152 Entry.isSRet = true;
5153 Entry.isNest = false;
5154 Entry.isByVal = false;
5155 Entry.Alignment = Align;
5156 Args.push_back(Entry);
5157 RetTy = Type::getVoidTy(FTy->getContext());
5160 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
5162 const Value *V = *i;
5165 if (V->getType()->isEmptyTy())
5168 SDValue ArgNode = getValue(V);
5169 Entry.Node = ArgNode; Entry.Ty = V->getType();
5171 unsigned attrInd = i - CS.arg_begin() + 1;
5172 Entry.isSExt = CS.paramHasAttr(attrInd, Attribute::SExt);
5173 Entry.isZExt = CS.paramHasAttr(attrInd, Attribute::ZExt);
5174 Entry.isInReg = CS.paramHasAttr(attrInd, Attribute::InReg);
5175 Entry.isSRet = CS.paramHasAttr(attrInd, Attribute::StructRet);
5176 Entry.isNest = CS.paramHasAttr(attrInd, Attribute::Nest);
5177 Entry.isByVal = CS.paramHasAttr(attrInd, Attribute::ByVal);
5178 Entry.Alignment = CS.getParamAlignment(attrInd);
5179 Args.push_back(Entry);
5183 // Insert a label before the invoke call to mark the try range. This can be
5184 // used to detect deletion of the invoke via the MachineModuleInfo.
5185 BeginLabel = MMI.getContext().CreateTempSymbol();
5187 // For SjLj, keep track of which landing pads go with which invokes
5188 // so as to maintain the ordering of pads in the LSDA.
5189 unsigned CallSiteIndex = MMI.getCurrentCallSite();
5190 if (CallSiteIndex) {
5191 MMI.setCallSiteBeginLabel(BeginLabel, CallSiteIndex);
5192 LPadToCallSiteMap[LandingPad].push_back(CallSiteIndex);
5194 // Now that the call site is handled, stop tracking it.
5195 MMI.setCurrentCallSite(0);
5198 // Both PendingLoads and PendingExports must be flushed here;
5199 // this call might not return.
5201 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getControlRoot(), BeginLabel));
5204 // Check if target-independent constraints permit a tail call here.
5205 // Target-dependent constraints are checked within TLI.LowerCallTo.
5207 !isInTailCallPosition(CS, CS.getAttributes().getRetAttributes(), TLI))
5210 // If there's a possibility that fast-isel has already selected some amount
5211 // of the current basic block, don't emit a tail call.
5212 if (isTailCall && TM.Options.EnableFastISel)
5215 std::pair<SDValue,SDValue> Result =
5216 TLI.LowerCallTo(getRoot(), RetTy,
5217 CS.paramHasAttr(0, Attribute::SExt),
5218 CS.paramHasAttr(0, Attribute::ZExt), FTy->isVarArg(),
5219 CS.paramHasAttr(0, Attribute::InReg), FTy->getNumParams(),
5220 CS.getCallingConv(),
5222 !CS.getInstruction()->use_empty(),
5223 Callee, Args, DAG, getCurDebugLoc());
5224 assert((isTailCall || Result.second.getNode()) &&
5225 "Non-null chain expected with non-tail call!");
5226 assert((Result.second.getNode() || !Result.first.getNode()) &&
5227 "Null value expected with tail call!");
5228 if (Result.first.getNode()) {
5229 setValue(CS.getInstruction(), Result.first);
5230 } else if (!CanLowerReturn && Result.second.getNode()) {
5231 // The instruction result is the result of loading from the
5232 // hidden sret parameter.
5233 SmallVector<EVT, 1> PVTs;
5234 Type *PtrRetTy = PointerType::getUnqual(FTy->getReturnType());
5236 ComputeValueVTs(TLI, PtrRetTy, PVTs);
5237 assert(PVTs.size() == 1 && "Pointers should fit in one register");
5238 EVT PtrVT = PVTs[0];
5239 unsigned NumValues = Outs.size();
5240 SmallVector<SDValue, 4> Values(NumValues);
5241 SmallVector<SDValue, 4> Chains(NumValues);
5243 for (unsigned i = 0; i < NumValues; ++i) {
5244 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT,
5246 DAG.getConstant(Offsets[i], PtrVT));
5247 SDValue L = DAG.getLoad(Outs[i].VT, getCurDebugLoc(), Result.second,
5249 MachinePointerInfo::getFixedStack(DemoteStackIdx, Offsets[i]),
5250 false, false, false, 1);
5252 Chains[i] = L.getValue(1);
5255 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
5256 MVT::Other, &Chains[0], NumValues);
5257 PendingLoads.push_back(Chain);
5259 // Collect the legal value parts into potentially illegal values
5260 // that correspond to the original function's return values.
5261 SmallVector<EVT, 4> RetTys;
5262 RetTy = FTy->getReturnType();
5263 ComputeValueVTs(TLI, RetTy, RetTys);
5264 ISD::NodeType AssertOp = ISD::DELETED_NODE;
5265 SmallVector<SDValue, 4> ReturnValues;
5266 unsigned CurReg = 0;
5267 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
5269 EVT RegisterVT = TLI.getRegisterType(RetTy->getContext(), VT);
5270 unsigned NumRegs = TLI.getNumRegisters(RetTy->getContext(), VT);
5272 SDValue ReturnValue =
5273 getCopyFromParts(DAG, getCurDebugLoc(), &Values[CurReg], NumRegs,
5274 RegisterVT, VT, AssertOp);
5275 ReturnValues.push_back(ReturnValue);
5279 setValue(CS.getInstruction(),
5280 DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
5281 DAG.getVTList(&RetTys[0], RetTys.size()),
5282 &ReturnValues[0], ReturnValues.size()));
5285 // Assign order to nodes here. If the call does not produce a result, it won't
5286 // be mapped to a SDNode and visit() will not assign it an order number.
5287 if (!Result.second.getNode()) {
5288 // As a special case, a null chain means that a tail call has been emitted and
5289 // the DAG root is already updated.
5292 AssignOrderingToNode(DAG.getRoot().getNode());
5294 DAG.setRoot(Result.second);
5296 AssignOrderingToNode(Result.second.getNode());
5300 // Insert a label at the end of the invoke call to mark the try range. This
5301 // can be used to detect deletion of the invoke via the MachineModuleInfo.
5302 MCSymbol *EndLabel = MMI.getContext().CreateTempSymbol();
5303 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getRoot(), EndLabel));
5305 // Inform MachineModuleInfo of range.
5306 MMI.addInvoke(LandingPad, BeginLabel, EndLabel);
5310 /// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the
5311 /// value is equal or not-equal to zero.
5312 static bool IsOnlyUsedInZeroEqualityComparison(const Value *V) {
5313 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end();
5315 if (const ICmpInst *IC = dyn_cast<ICmpInst>(*UI))
5316 if (IC->isEquality())
5317 if (const Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
5318 if (C->isNullValue())
5320 // Unknown instruction.
5326 static SDValue getMemCmpLoad(const Value *PtrVal, MVT LoadVT,
5328 SelectionDAGBuilder &Builder) {
5330 // Check to see if this load can be trivially constant folded, e.g. if the
5331 // input is from a string literal.
5332 if (const Constant *LoadInput = dyn_cast<Constant>(PtrVal)) {
5333 // Cast pointer to the type we really want to load.
5334 LoadInput = ConstantExpr::getBitCast(const_cast<Constant *>(LoadInput),
5335 PointerType::getUnqual(LoadTy));
5337 if (const Constant *LoadCst =
5338 ConstantFoldLoadFromConstPtr(const_cast<Constant *>(LoadInput),
5340 return Builder.getValue(LoadCst);
5343 // Otherwise, we have to emit the load. If the pointer is to unfoldable but
5344 // still constant memory, the input chain can be the entry node.
5346 bool ConstantMemory = false;
5348 // Do not serialize (non-volatile) loads of constant memory with anything.
5349 if (Builder.AA->pointsToConstantMemory(PtrVal)) {
5350 Root = Builder.DAG.getEntryNode();
5351 ConstantMemory = true;
5353 // Do not serialize non-volatile loads against each other.
5354 Root = Builder.DAG.getRoot();
5357 SDValue Ptr = Builder.getValue(PtrVal);
5358 SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurDebugLoc(), Root,
5359 Ptr, MachinePointerInfo(PtrVal),
5361 false /*nontemporal*/,
5362 false /*isinvariant*/, 1 /* align=1 */);
5364 if (!ConstantMemory)
5365 Builder.PendingLoads.push_back(LoadVal.getValue(1));
5370 /// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form.
5371 /// If so, return true and lower it, otherwise return false and it will be
5372 /// lowered like a normal call.
5373 bool SelectionDAGBuilder::visitMemCmpCall(const CallInst &I) {
5374 // Verify that the prototype makes sense. int memcmp(void*,void*,size_t)
5375 if (I.getNumArgOperands() != 3)
5378 const Value *LHS = I.getArgOperand(0), *RHS = I.getArgOperand(1);
5379 if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() ||
5380 !I.getArgOperand(2)->getType()->isIntegerTy() ||
5381 !I.getType()->isIntegerTy())
5384 const ConstantInt *Size = dyn_cast<ConstantInt>(I.getArgOperand(2));
5386 // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0
5387 // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0
5388 if (Size && IsOnlyUsedInZeroEqualityComparison(&I)) {
5389 bool ActuallyDoIt = true;
5392 switch (Size->getZExtValue()) {
5394 LoadVT = MVT::Other;
5396 ActuallyDoIt = false;
5400 LoadTy = Type::getInt16Ty(Size->getContext());
5404 LoadTy = Type::getInt32Ty(Size->getContext());
5408 LoadTy = Type::getInt64Ty(Size->getContext());
5412 LoadVT = MVT::v4i32;
5413 LoadTy = Type::getInt32Ty(Size->getContext());
5414 LoadTy = VectorType::get(LoadTy, 4);
5419 // This turns into unaligned loads. We only do this if the target natively
5420 // supports the MVT we'll be loading or if it is small enough (<= 4) that
5421 // we'll only produce a small number of byte loads.
5423 // Require that we can find a legal MVT, and only do this if the target
5424 // supports unaligned loads of that type. Expanding into byte loads would
5426 if (ActuallyDoIt && Size->getZExtValue() > 4) {
5427 // TODO: Handle 5 byte compare as 4-byte + 1 byte.
5428 // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads.
5429 if (!TLI.isTypeLegal(LoadVT) ||!TLI.allowsUnalignedMemoryAccesses(LoadVT))
5430 ActuallyDoIt = false;
5434 SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this);
5435 SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this);
5437 SDValue Res = DAG.getSetCC(getCurDebugLoc(), MVT::i1, LHSVal, RHSVal,
5439 EVT CallVT = TLI.getValueType(I.getType(), true);
5440 setValue(&I, DAG.getZExtOrTrunc(Res, getCurDebugLoc(), CallVT));
5450 void SelectionDAGBuilder::visitCall(const CallInst &I) {
5451 // Handle inline assembly differently.
5452 if (isa<InlineAsm>(I.getCalledValue())) {
5457 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5458 ComputeUsesVAFloatArgument(I, &MMI);
5460 const char *RenameFn = 0;
5461 if (Function *F = I.getCalledFunction()) {
5462 if (F->isDeclaration()) {
5463 if (const TargetIntrinsicInfo *II = TM.getIntrinsicInfo()) {
5464 if (unsigned IID = II->getIntrinsicID(F)) {
5465 RenameFn = visitIntrinsicCall(I, IID);
5470 if (unsigned IID = F->getIntrinsicID()) {
5471 RenameFn = visitIntrinsicCall(I, IID);
5477 // Check for well-known libc/libm calls. If the function is internal, it
5478 // can't be a library call.
5479 if (!F->hasLocalLinkage() && F->hasName()) {
5480 StringRef Name = F->getName();
5481 if ((LibInfo->has(LibFunc::copysign) && Name == "copysign") ||
5482 (LibInfo->has(LibFunc::copysignf) && Name == "copysignf") ||
5483 (LibInfo->has(LibFunc::copysignl) && Name == "copysignl")) {
5484 if (I.getNumArgOperands() == 2 && // Basic sanity checks.
5485 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5486 I.getType() == I.getArgOperand(0)->getType() &&
5487 I.getType() == I.getArgOperand(1)->getType()) {
5488 SDValue LHS = getValue(I.getArgOperand(0));
5489 SDValue RHS = getValue(I.getArgOperand(1));
5490 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurDebugLoc(),
5491 LHS.getValueType(), LHS, RHS));
5494 } else if ((LibInfo->has(LibFunc::fabs) && Name == "fabs") ||
5495 (LibInfo->has(LibFunc::fabsf) && Name == "fabsf") ||
5496 (LibInfo->has(LibFunc::fabsl) && Name == "fabsl")) {
5497 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5498 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5499 I.getType() == I.getArgOperand(0)->getType()) {
5500 SDValue Tmp = getValue(I.getArgOperand(0));
5501 setValue(&I, DAG.getNode(ISD::FABS, getCurDebugLoc(),
5502 Tmp.getValueType(), Tmp));
5505 } else if ((LibInfo->has(LibFunc::sin) && Name == "sin") ||
5506 (LibInfo->has(LibFunc::sinf) && Name == "sinf") ||
5507 (LibInfo->has(LibFunc::sinl) && Name == "sinl")) {
5508 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5509 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5510 I.getType() == I.getArgOperand(0)->getType() &&
5511 I.onlyReadsMemory()) {
5512 SDValue Tmp = getValue(I.getArgOperand(0));
5513 setValue(&I, DAG.getNode(ISD::FSIN, getCurDebugLoc(),
5514 Tmp.getValueType(), Tmp));
5517 } else if ((LibInfo->has(LibFunc::cos) && Name == "cos") ||
5518 (LibInfo->has(LibFunc::cosf) && Name == "cosf") ||
5519 (LibInfo->has(LibFunc::cosl) && Name == "cosl")) {
5520 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5521 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5522 I.getType() == I.getArgOperand(0)->getType() &&
5523 I.onlyReadsMemory()) {
5524 SDValue Tmp = getValue(I.getArgOperand(0));
5525 setValue(&I, DAG.getNode(ISD::FCOS, getCurDebugLoc(),
5526 Tmp.getValueType(), Tmp));
5529 } else if ((LibInfo->has(LibFunc::sqrt) && Name == "sqrt") ||
5530 (LibInfo->has(LibFunc::sqrtf) && Name == "sqrtf") ||
5531 (LibInfo->has(LibFunc::sqrtl) && Name == "sqrtl")) {
5532 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5533 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5534 I.getType() == I.getArgOperand(0)->getType() &&
5535 I.onlyReadsMemory()) {
5536 SDValue Tmp = getValue(I.getArgOperand(0));
5537 setValue(&I, DAG.getNode(ISD::FSQRT, getCurDebugLoc(),
5538 Tmp.getValueType(), Tmp));
5541 } else if ((LibInfo->has(LibFunc::floor) && Name == "floor") ||
5542 (LibInfo->has(LibFunc::floorf) && Name == "floorf") ||
5543 (LibInfo->has(LibFunc::floorl) && Name == "floorl")) {
5544 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5545 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5546 I.getType() == I.getArgOperand(0)->getType()) {
5547 SDValue Tmp = getValue(I.getArgOperand(0));
5548 setValue(&I, DAG.getNode(ISD::FFLOOR, getCurDebugLoc(),
5549 Tmp.getValueType(), Tmp));
5552 } else if ((LibInfo->has(LibFunc::nearbyint) && Name == "nearbyint") ||
5553 (LibInfo->has(LibFunc::nearbyintf) && Name == "nearbyintf") ||
5554 (LibInfo->has(LibFunc::nearbyintl) && Name == "nearbyintl")) {
5555 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5556 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5557 I.getType() == I.getArgOperand(0)->getType()) {
5558 SDValue Tmp = getValue(I.getArgOperand(0));
5559 setValue(&I, DAG.getNode(ISD::FNEARBYINT, getCurDebugLoc(),
5560 Tmp.getValueType(), Tmp));
5563 } else if ((LibInfo->has(LibFunc::ceil) && Name == "ceil") ||
5564 (LibInfo->has(LibFunc::ceilf) && Name == "ceilf") ||
5565 (LibInfo->has(LibFunc::ceill) && Name == "ceill")) {
5566 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5567 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5568 I.getType() == I.getArgOperand(0)->getType()) {
5569 SDValue Tmp = getValue(I.getArgOperand(0));
5570 setValue(&I, DAG.getNode(ISD::FCEIL, getCurDebugLoc(),
5571 Tmp.getValueType(), Tmp));
5574 } else if ((LibInfo->has(LibFunc::rint) && Name == "rint") ||
5575 (LibInfo->has(LibFunc::rintf) && Name == "rintf") ||
5576 (LibInfo->has(LibFunc::rintl) && Name == "rintl")) {
5577 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5578 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5579 I.getType() == I.getArgOperand(0)->getType()) {
5580 SDValue Tmp = getValue(I.getArgOperand(0));
5581 setValue(&I, DAG.getNode(ISD::FRINT, getCurDebugLoc(),
5582 Tmp.getValueType(), Tmp));
5585 } else if ((LibInfo->has(LibFunc::trunc) && Name == "trunc") ||
5586 (LibInfo->has(LibFunc::truncf) && Name == "truncf") ||
5587 (LibInfo->has(LibFunc::truncl) && Name == "truncl")) {
5588 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5589 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5590 I.getType() == I.getArgOperand(0)->getType()) {
5591 SDValue Tmp = getValue(I.getArgOperand(0));
5592 setValue(&I, DAG.getNode(ISD::FTRUNC, getCurDebugLoc(),
5593 Tmp.getValueType(), Tmp));
5596 } else if ((LibInfo->has(LibFunc::log2) && Name == "log2") ||
5597 (LibInfo->has(LibFunc::log2f) && Name == "log2f") ||
5598 (LibInfo->has(LibFunc::log2l) && Name == "log2l")) {
5599 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5600 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5601 I.getType() == I.getArgOperand(0)->getType()) {
5602 SDValue Tmp = getValue(I.getArgOperand(0));
5603 setValue(&I, DAG.getNode(ISD::FLOG2, getCurDebugLoc(),
5604 Tmp.getValueType(), Tmp));
5607 } else if ((LibInfo->has(LibFunc::exp2) && Name == "exp2") ||
5608 (LibInfo->has(LibFunc::exp2f) && Name == "exp2f") ||
5609 (LibInfo->has(LibFunc::exp2l) && Name == "exp2l")) {
5610 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5611 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5612 I.getType() == I.getArgOperand(0)->getType()) {
5613 SDValue Tmp = getValue(I.getArgOperand(0));
5614 setValue(&I, DAG.getNode(ISD::FEXP2, getCurDebugLoc(),
5615 Tmp.getValueType(), Tmp));
5618 } else if (Name == "memcmp") {
5619 if (visitMemCmpCall(I))
5627 Callee = getValue(I.getCalledValue());
5629 Callee = DAG.getExternalSymbol(RenameFn, TLI.getPointerTy());
5631 // Check if we can potentially perform a tail call. More detailed checking is
5632 // be done within LowerCallTo, after more information about the call is known.
5633 LowerCallTo(&I, Callee, I.isTailCall());
5638 /// AsmOperandInfo - This contains information for each constraint that we are
5640 class SDISelAsmOperandInfo : public TargetLowering::AsmOperandInfo {
5642 /// CallOperand - If this is the result output operand or a clobber
5643 /// this is null, otherwise it is the incoming operand to the CallInst.
5644 /// This gets modified as the asm is processed.
5645 SDValue CallOperand;
5647 /// AssignedRegs - If this is a register or register class operand, this
5648 /// contains the set of register corresponding to the operand.
5649 RegsForValue AssignedRegs;
5651 explicit SDISelAsmOperandInfo(const TargetLowering::AsmOperandInfo &info)
5652 : TargetLowering::AsmOperandInfo(info), CallOperand(0,0) {
5655 /// MarkAllocatedRegs - Once AssignedRegs is set, mark the assigned registers
5656 /// busy in OutputRegs/InputRegs.
5657 void MarkAllocatedRegs(bool isOutReg, bool isInReg,
5658 std::set<unsigned> &OutputRegs,
5659 std::set<unsigned> &InputRegs,
5660 const TargetRegisterInfo &TRI) const {
5662 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
5663 MarkRegAndAliases(AssignedRegs.Regs[i], OutputRegs, TRI);
5666 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
5667 MarkRegAndAliases(AssignedRegs.Regs[i], InputRegs, TRI);
5671 /// getCallOperandValEVT - Return the EVT of the Value* that this operand
5672 /// corresponds to. If there is no Value* for this operand, it returns
5674 EVT getCallOperandValEVT(LLVMContext &Context,
5675 const TargetLowering &TLI,
5676 const TargetData *TD) const {
5677 if (CallOperandVal == 0) return MVT::Other;
5679 if (isa<BasicBlock>(CallOperandVal))
5680 return TLI.getPointerTy();
5682 llvm::Type *OpTy = CallOperandVal->getType();
5684 // FIXME: code duplicated from TargetLowering::ParseConstraints().
5685 // If this is an indirect operand, the operand is a pointer to the
5688 llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
5690 report_fatal_error("Indirect operand for inline asm not a pointer!");
5691 OpTy = PtrTy->getElementType();
5694 // Look for vector wrapped in a struct. e.g. { <16 x i8> }.
5695 if (StructType *STy = dyn_cast<StructType>(OpTy))
5696 if (STy->getNumElements() == 1)
5697 OpTy = STy->getElementType(0);
5699 // If OpTy is not a single value, it may be a struct/union that we
5700 // can tile with integers.
5701 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
5702 unsigned BitSize = TD->getTypeSizeInBits(OpTy);
5711 OpTy = IntegerType::get(Context, BitSize);
5716 return TLI.getValueType(OpTy, true);
5720 /// MarkRegAndAliases - Mark the specified register and all aliases in the
5722 static void MarkRegAndAliases(unsigned Reg, std::set<unsigned> &Regs,
5723 const TargetRegisterInfo &TRI) {
5724 assert(TargetRegisterInfo::isPhysicalRegister(Reg) && "Isn't a physreg");
5726 if (const unsigned *Aliases = TRI.getAliasSet(Reg))
5727 for (; *Aliases; ++Aliases)
5728 Regs.insert(*Aliases);
5732 typedef SmallVector<SDISelAsmOperandInfo,16> SDISelAsmOperandInfoVector;
5734 } // end anonymous namespace
5736 /// GetRegistersForValue - Assign registers (virtual or physical) for the
5737 /// specified operand. We prefer to assign virtual registers, to allow the
5738 /// register allocator to handle the assignment process. However, if the asm
5739 /// uses features that we can't model on machineinstrs, we have SDISel do the
5740 /// allocation. This produces generally horrible, but correct, code.
5742 /// OpInfo describes the operand.
5743 /// Input and OutputRegs are the set of already allocated physical registers.
5745 static void GetRegistersForValue(SelectionDAG &DAG,
5746 const TargetLowering &TLI,
5748 SDISelAsmOperandInfo &OpInfo,
5749 std::set<unsigned> &OutputRegs,
5750 std::set<unsigned> &InputRegs) {
5751 LLVMContext &Context = *DAG.getContext();
5753 // Compute whether this value requires an input register, an output register,
5755 bool isOutReg = false;
5756 bool isInReg = false;
5757 switch (OpInfo.Type) {
5758 case InlineAsm::isOutput:
5761 // If there is an input constraint that matches this, we need to reserve
5762 // the input register so no other inputs allocate to it.
5763 isInReg = OpInfo.hasMatchingInput();
5765 case InlineAsm::isInput:
5769 case InlineAsm::isClobber:
5776 MachineFunction &MF = DAG.getMachineFunction();
5777 SmallVector<unsigned, 4> Regs;
5779 // If this is a constraint for a single physreg, or a constraint for a
5780 // register class, find it.
5781 std::pair<unsigned, const TargetRegisterClass*> PhysReg =
5782 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
5783 OpInfo.ConstraintVT);
5785 unsigned NumRegs = 1;
5786 if (OpInfo.ConstraintVT != MVT::Other) {
5787 // If this is a FP input in an integer register (or visa versa) insert a bit
5788 // cast of the input value. More generally, handle any case where the input
5789 // value disagrees with the register class we plan to stick this in.
5790 if (OpInfo.Type == InlineAsm::isInput &&
5791 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) {
5792 // Try to convert to the first EVT that the reg class contains. If the
5793 // types are identical size, use a bitcast to convert (e.g. two differing
5795 EVT RegVT = *PhysReg.second->vt_begin();
5796 if (RegVT.getSizeInBits() == OpInfo.ConstraintVT.getSizeInBits()) {
5797 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
5798 RegVT, OpInfo.CallOperand);
5799 OpInfo.ConstraintVT = RegVT;
5800 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) {
5801 // If the input is a FP value and we want it in FP registers, do a
5802 // bitcast to the corresponding integer type. This turns an f64 value
5803 // into i64, which can be passed with two i32 values on a 32-bit
5805 RegVT = EVT::getIntegerVT(Context,
5806 OpInfo.ConstraintVT.getSizeInBits());
5807 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
5808 RegVT, OpInfo.CallOperand);
5809 OpInfo.ConstraintVT = RegVT;
5813 NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT);
5817 EVT ValueVT = OpInfo.ConstraintVT;
5819 // If this is a constraint for a specific physical register, like {r17},
5821 if (unsigned AssignedReg = PhysReg.first) {
5822 const TargetRegisterClass *RC = PhysReg.second;
5823 if (OpInfo.ConstraintVT == MVT::Other)
5824 ValueVT = *RC->vt_begin();
5826 // Get the actual register value type. This is important, because the user
5827 // may have asked for (e.g.) the AX register in i32 type. We need to
5828 // remember that AX is actually i16 to get the right extension.
5829 RegVT = *RC->vt_begin();
5831 // This is a explicit reference to a physical register.
5832 Regs.push_back(AssignedReg);
5834 // If this is an expanded reference, add the rest of the regs to Regs.
5836 TargetRegisterClass::iterator I = RC->begin();
5837 for (; *I != AssignedReg; ++I)
5838 assert(I != RC->end() && "Didn't find reg!");
5840 // Already added the first reg.
5842 for (; NumRegs; --NumRegs, ++I) {
5843 assert(I != RC->end() && "Ran out of registers to allocate!");
5848 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5849 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
5850 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
5854 // Otherwise, if this was a reference to an LLVM register class, create vregs
5855 // for this reference.
5856 if (const TargetRegisterClass *RC = PhysReg.second) {
5857 RegVT = *RC->vt_begin();
5858 if (OpInfo.ConstraintVT == MVT::Other)
5861 // Create the appropriate number of virtual registers.
5862 MachineRegisterInfo &RegInfo = MF.getRegInfo();
5863 for (; NumRegs; --NumRegs)
5864 Regs.push_back(RegInfo.createVirtualRegister(RC));
5866 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5870 // Otherwise, we couldn't allocate enough registers for this.
5873 /// visitInlineAsm - Handle a call to an InlineAsm object.
5875 void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) {
5876 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
5878 /// ConstraintOperands - Information about all of the constraints.
5879 SDISelAsmOperandInfoVector ConstraintOperands;
5881 std::set<unsigned> OutputRegs, InputRegs;
5883 TargetLowering::AsmOperandInfoVector
5884 TargetConstraints = TLI.ParseConstraints(CS);
5886 bool hasMemory = false;
5888 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
5889 unsigned ResNo = 0; // ResNo - The result number of the next output.
5890 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
5891 ConstraintOperands.push_back(SDISelAsmOperandInfo(TargetConstraints[i]));
5892 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
5894 EVT OpVT = MVT::Other;
5896 // Compute the value type for each operand.
5897 switch (OpInfo.Type) {
5898 case InlineAsm::isOutput:
5899 // Indirect outputs just consume an argument.
5900 if (OpInfo.isIndirect) {
5901 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5905 // The return value of the call is this value. As such, there is no
5906 // corresponding argument.
5907 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
5908 if (StructType *STy = dyn_cast<StructType>(CS.getType())) {
5909 OpVT = TLI.getValueType(STy->getElementType(ResNo));
5911 assert(ResNo == 0 && "Asm only has one result!");
5912 OpVT = TLI.getValueType(CS.getType());
5916 case InlineAsm::isInput:
5917 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5919 case InlineAsm::isClobber:
5924 // If this is an input or an indirect output, process the call argument.
5925 // BasicBlocks are labels, currently appearing only in asm's.
5926 if (OpInfo.CallOperandVal) {
5927 if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) {
5928 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]);
5930 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
5933 OpVT = OpInfo.getCallOperandValEVT(*DAG.getContext(), TLI, TD);
5936 OpInfo.ConstraintVT = OpVT;
5938 // Indirect operand accesses access memory.
5939 if (OpInfo.isIndirect)
5942 for (unsigned j = 0, ee = OpInfo.Codes.size(); j != ee; ++j) {
5943 TargetLowering::ConstraintType
5944 CType = TLI.getConstraintType(OpInfo.Codes[j]);
5945 if (CType == TargetLowering::C_Memory) {
5953 SDValue Chain, Flag;
5955 // We won't need to flush pending loads if this asm doesn't touch
5956 // memory and is nonvolatile.
5957 if (hasMemory || IA->hasSideEffects())
5960 Chain = DAG.getRoot();
5962 // Second pass over the constraints: compute which constraint option to use
5963 // and assign registers to constraints that want a specific physreg.
5964 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5965 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5967 // If this is an output operand with a matching input operand, look up the
5968 // matching input. If their types mismatch, e.g. one is an integer, the
5969 // other is floating point, or their sizes are different, flag it as an
5971 if (OpInfo.hasMatchingInput()) {
5972 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
5974 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
5975 std::pair<unsigned, const TargetRegisterClass*> MatchRC =
5976 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
5977 OpInfo.ConstraintVT);
5978 std::pair<unsigned, const TargetRegisterClass*> InputRC =
5979 TLI.getRegForInlineAsmConstraint(Input.ConstraintCode,
5980 Input.ConstraintVT);
5981 if ((OpInfo.ConstraintVT.isInteger() !=
5982 Input.ConstraintVT.isInteger()) ||
5983 (MatchRC.second != InputRC.second)) {
5984 report_fatal_error("Unsupported asm: input constraint"
5985 " with a matching output constraint of"
5986 " incompatible type!");
5988 Input.ConstraintVT = OpInfo.ConstraintVT;
5992 // Compute the constraint code and ConstraintType to use.
5993 TLI.ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, &DAG);
5995 // If this is a memory input, and if the operand is not indirect, do what we
5996 // need to to provide an address for the memory input.
5997 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5998 !OpInfo.isIndirect) {
5999 assert((OpInfo.isMultipleAlternative ||
6000 (OpInfo.Type == InlineAsm::isInput)) &&
6001 "Can only indirectify direct input operands!");
6003 // Memory operands really want the address of the value. If we don't have
6004 // an indirect input, put it in the constpool if we can, otherwise spill
6005 // it to a stack slot.
6006 // TODO: This isn't quite right. We need to handle these according to
6007 // the addressing mode that the constraint wants. Also, this may take
6008 // an additional register for the computation and we don't want that
6011 // If the operand is a float, integer, or vector constant, spill to a
6012 // constant pool entry to get its address.
6013 const Value *OpVal = OpInfo.CallOperandVal;
6014 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
6015 isa<ConstantVector>(OpVal) || isa<ConstantDataVector>(OpVal)) {
6016 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal),
6017 TLI.getPointerTy());
6019 // Otherwise, create a stack slot and emit a store to it before the
6021 Type *Ty = OpVal->getType();
6022 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
6023 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(Ty);
6024 MachineFunction &MF = DAG.getMachineFunction();
6025 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
6026 SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
6027 Chain = DAG.getStore(Chain, getCurDebugLoc(),
6028 OpInfo.CallOperand, StackSlot,
6029 MachinePointerInfo::getFixedStack(SSFI),
6031 OpInfo.CallOperand = StackSlot;
6034 // There is no longer a Value* corresponding to this operand.
6035 OpInfo.CallOperandVal = 0;
6037 // It is now an indirect operand.
6038 OpInfo.isIndirect = true;
6041 // If this constraint is for a specific register, allocate it before
6043 if (OpInfo.ConstraintType == TargetLowering::C_Register)
6044 GetRegistersForValue(DAG, TLI, getCurDebugLoc(), OpInfo, OutputRegs,
6048 // Second pass - Loop over all of the operands, assigning virtual or physregs
6049 // to register class operands.
6050 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6051 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6053 // C_Register operands have already been allocated, Other/Memory don't need
6055 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
6056 GetRegistersForValue(DAG, TLI, getCurDebugLoc(), OpInfo, OutputRegs,
6060 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
6061 std::vector<SDValue> AsmNodeOperands;
6062 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain
6063 AsmNodeOperands.push_back(
6064 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(),
6065 TLI.getPointerTy()));
6067 // If we have a !srcloc metadata node associated with it, we want to attach
6068 // this to the ultimately generated inline asm machineinstr. To do this, we
6069 // pass in the third operand as this (potentially null) inline asm MDNode.
6070 const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc");
6071 AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc));
6073 // Remember the HasSideEffect and AlignStack bits as operand 3.
6074 unsigned ExtraInfo = 0;
6075 if (IA->hasSideEffects())
6076 ExtraInfo |= InlineAsm::Extra_HasSideEffects;
6077 if (IA->isAlignStack())
6078 ExtraInfo |= InlineAsm::Extra_IsAlignStack;
6079 AsmNodeOperands.push_back(DAG.getTargetConstant(ExtraInfo,
6080 TLI.getPointerTy()));
6082 // Loop over all of the inputs, copying the operand values into the
6083 // appropriate registers and processing the output regs.
6084 RegsForValue RetValRegs;
6086 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
6087 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
6089 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6090 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6092 switch (OpInfo.Type) {
6093 case InlineAsm::isOutput: {
6094 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
6095 OpInfo.ConstraintType != TargetLowering::C_Register) {
6096 // Memory output, or 'other' output (e.g. 'X' constraint).
6097 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
6099 // Add information to the INLINEASM node to know about this output.
6100 unsigned OpFlags = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
6101 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlags,
6102 TLI.getPointerTy()));
6103 AsmNodeOperands.push_back(OpInfo.CallOperand);
6107 // Otherwise, this is a register or register class output.
6109 // Copy the output from the appropriate register. Find a register that
6111 if (OpInfo.AssignedRegs.Regs.empty()) {
6112 LLVMContext &Ctx = *DAG.getContext();
6113 Ctx.emitError(CS.getInstruction(),
6114 "couldn't allocate output register for constraint '" +
6115 Twine(OpInfo.ConstraintCode) + "'");
6119 // If this is an indirect operand, store through the pointer after the
6121 if (OpInfo.isIndirect) {
6122 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
6123 OpInfo.CallOperandVal));
6125 // This is the result value of the call.
6126 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
6127 // Concatenate this output onto the outputs list.
6128 RetValRegs.append(OpInfo.AssignedRegs);
6131 // Add information to the INLINEASM node to know that this register is
6133 OpInfo.AssignedRegs.AddInlineAsmOperands(OpInfo.isEarlyClobber ?
6134 InlineAsm::Kind_RegDefEarlyClobber :
6135 InlineAsm::Kind_RegDef,
6142 case InlineAsm::isInput: {
6143 SDValue InOperandVal = OpInfo.CallOperand;
6145 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint?
6146 // If this is required to match an output register we have already set,
6147 // just use its register.
6148 unsigned OperandNo = OpInfo.getMatchedOperand();
6150 // Scan until we find the definition we already emitted of this operand.
6151 // When we find it, create a RegsForValue operand.
6152 unsigned CurOp = InlineAsm::Op_FirstOperand;
6153 for (; OperandNo; --OperandNo) {
6154 // Advance to the next operand.
6156 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6157 assert((InlineAsm::isRegDefKind(OpFlag) ||
6158 InlineAsm::isRegDefEarlyClobberKind(OpFlag) ||
6159 InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?");
6160 CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1;
6164 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6165 if (InlineAsm::isRegDefKind(OpFlag) ||
6166 InlineAsm::isRegDefEarlyClobberKind(OpFlag)) {
6167 // Add (OpFlag&0xffff)>>3 registers to MatchedRegs.
6168 if (OpInfo.isIndirect) {
6169 // This happens on gcc/testsuite/gcc.dg/pr8788-1.c
6170 LLVMContext &Ctx = *DAG.getContext();
6171 Ctx.emitError(CS.getInstruction(), "inline asm not supported yet:"
6172 " don't know how to handle tied "
6173 "indirect register inputs");
6176 RegsForValue MatchedRegs;
6177 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
6178 EVT RegVT = AsmNodeOperands[CurOp+1].getValueType();
6179 MatchedRegs.RegVTs.push_back(RegVT);
6180 MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo();
6181 for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag);
6183 MatchedRegs.Regs.push_back
6184 (RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT)));
6186 // Use the produced MatchedRegs object to
6187 MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
6189 MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse,
6190 true, OpInfo.getMatchedOperand(),
6191 DAG, AsmNodeOperands);
6195 assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!");
6196 assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 &&
6197 "Unexpected number of operands");
6198 // Add information to the INLINEASM node to know about this input.
6199 // See InlineAsm.h isUseOperandTiedToDef.
6200 OpFlag = InlineAsm::getFlagWordForMatchingOp(OpFlag,
6201 OpInfo.getMatchedOperand());
6202 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlag,
6203 TLI.getPointerTy()));
6204 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]);
6208 // Treat indirect 'X' constraint as memory.
6209 if (OpInfo.ConstraintType == TargetLowering::C_Other &&
6211 OpInfo.ConstraintType = TargetLowering::C_Memory;
6213 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
6214 std::vector<SDValue> Ops;
6215 TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode,
6218 LLVMContext &Ctx = *DAG.getContext();
6219 Ctx.emitError(CS.getInstruction(),
6220 "invalid operand for inline asm constraint '" +
6221 Twine(OpInfo.ConstraintCode) + "'");
6225 // Add information to the INLINEASM node to know about this input.
6226 unsigned ResOpType =
6227 InlineAsm::getFlagWord(InlineAsm::Kind_Imm, Ops.size());
6228 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
6229 TLI.getPointerTy()));
6230 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
6234 if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
6235 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
6236 assert(InOperandVal.getValueType() == TLI.getPointerTy() &&
6237 "Memory operands expect pointer values");
6239 // Add information to the INLINEASM node to know about this input.
6240 unsigned ResOpType = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
6241 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
6242 TLI.getPointerTy()));
6243 AsmNodeOperands.push_back(InOperandVal);
6247 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
6248 OpInfo.ConstraintType == TargetLowering::C_Register) &&
6249 "Unknown constraint type!");
6250 assert(!OpInfo.isIndirect &&
6251 "Don't know how to handle indirect register inputs yet!");
6253 // Copy the input into the appropriate registers.
6254 if (OpInfo.AssignedRegs.Regs.empty()) {
6255 LLVMContext &Ctx = *DAG.getContext();
6256 Ctx.emitError(CS.getInstruction(),
6257 "couldn't allocate input reg for constraint '" +
6258 Twine(OpInfo.ConstraintCode) + "'");
6262 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
6265 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, false, 0,
6266 DAG, AsmNodeOperands);
6269 case InlineAsm::isClobber: {
6270 // Add the clobbered value to the operand list, so that the register
6271 // allocator is aware that the physreg got clobbered.
6272 if (!OpInfo.AssignedRegs.Regs.empty())
6273 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_Clobber,
6281 // Finish up input operands. Set the input chain and add the flag last.
6282 AsmNodeOperands[InlineAsm::Op_InputChain] = Chain;
6283 if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
6285 Chain = DAG.getNode(ISD::INLINEASM, getCurDebugLoc(),
6286 DAG.getVTList(MVT::Other, MVT::Glue),
6287 &AsmNodeOperands[0], AsmNodeOperands.size());
6288 Flag = Chain.getValue(1);
6290 // If this asm returns a register value, copy the result from that register
6291 // and set it as the value of the call.
6292 if (!RetValRegs.Regs.empty()) {
6293 SDValue Val = RetValRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(),
6296 // FIXME: Why don't we do this for inline asms with MRVs?
6297 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) {
6298 EVT ResultType = TLI.getValueType(CS.getType());
6300 // If any of the results of the inline asm is a vector, it may have the
6301 // wrong width/num elts. This can happen for register classes that can
6302 // contain multiple different value types. The preg or vreg allocated may
6303 // not have the same VT as was expected. Convert it to the right type
6304 // with bit_convert.
6305 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) {
6306 Val = DAG.getNode(ISD::BITCAST, getCurDebugLoc(),
6309 } else if (ResultType != Val.getValueType() &&
6310 ResultType.isInteger() && Val.getValueType().isInteger()) {
6311 // If a result value was tied to an input value, the computed result may
6312 // have a wider width than the expected result. Extract the relevant
6314 Val = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), ResultType, Val);
6317 assert(ResultType == Val.getValueType() && "Asm result value mismatch!");
6320 setValue(CS.getInstruction(), Val);
6321 // Don't need to use this as a chain in this case.
6322 if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty())
6326 std::vector<std::pair<SDValue, const Value *> > StoresToEmit;
6328 // Process indirect outputs, first output all of the flagged copies out of
6330 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
6331 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
6332 const Value *Ptr = IndirectStoresToEmit[i].second;
6333 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(),
6335 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
6338 // Emit the non-flagged stores from the physregs.
6339 SmallVector<SDValue, 8> OutChains;
6340 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) {
6341 SDValue Val = DAG.getStore(Chain, getCurDebugLoc(),
6342 StoresToEmit[i].first,
6343 getValue(StoresToEmit[i].second),
6344 MachinePointerInfo(StoresToEmit[i].second),
6346 OutChains.push_back(Val);
6349 if (!OutChains.empty())
6350 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
6351 &OutChains[0], OutChains.size());
6356 void SelectionDAGBuilder::visitVAStart(const CallInst &I) {
6357 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurDebugLoc(),
6358 MVT::Other, getRoot(),
6359 getValue(I.getArgOperand(0)),
6360 DAG.getSrcValue(I.getArgOperand(0))));
6363 void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) {
6364 const TargetData &TD = *TLI.getTargetData();
6365 SDValue V = DAG.getVAArg(TLI.getValueType(I.getType()), getCurDebugLoc(),
6366 getRoot(), getValue(I.getOperand(0)),
6367 DAG.getSrcValue(I.getOperand(0)),
6368 TD.getABITypeAlignment(I.getType()));
6370 DAG.setRoot(V.getValue(1));
6373 void SelectionDAGBuilder::visitVAEnd(const CallInst &I) {
6374 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurDebugLoc(),
6375 MVT::Other, getRoot(),
6376 getValue(I.getArgOperand(0)),
6377 DAG.getSrcValue(I.getArgOperand(0))));
6380 void SelectionDAGBuilder::visitVACopy(const CallInst &I) {
6381 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurDebugLoc(),
6382 MVT::Other, getRoot(),
6383 getValue(I.getArgOperand(0)),
6384 getValue(I.getArgOperand(1)),
6385 DAG.getSrcValue(I.getArgOperand(0)),
6386 DAG.getSrcValue(I.getArgOperand(1))));
6389 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
6390 /// implementation, which just calls LowerCall.
6391 /// FIXME: When all targets are
6392 /// migrated to using LowerCall, this hook should be integrated into SDISel.
6393 std::pair<SDValue, SDValue>
6394 TargetLowering::LowerCallTo(SDValue Chain, Type *RetTy,
6395 bool RetSExt, bool RetZExt, bool isVarArg,
6396 bool isInreg, unsigned NumFixedArgs,
6397 CallingConv::ID CallConv, bool isTailCall,
6398 bool isReturnValueUsed,
6400 ArgListTy &Args, SelectionDAG &DAG,
6401 DebugLoc dl) const {
6402 // Handle all of the outgoing arguments.
6403 SmallVector<ISD::OutputArg, 32> Outs;
6404 SmallVector<SDValue, 32> OutVals;
6405 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
6406 SmallVector<EVT, 4> ValueVTs;
6407 ComputeValueVTs(*this, Args[i].Ty, ValueVTs);
6408 for (unsigned Value = 0, NumValues = ValueVTs.size();
6409 Value != NumValues; ++Value) {
6410 EVT VT = ValueVTs[Value];
6411 Type *ArgTy = VT.getTypeForEVT(RetTy->getContext());
6412 SDValue Op = SDValue(Args[i].Node.getNode(),
6413 Args[i].Node.getResNo() + Value);
6414 ISD::ArgFlagsTy Flags;
6415 unsigned OriginalAlignment =
6416 getTargetData()->getABITypeAlignment(ArgTy);
6422 if (Args[i].isInReg)
6426 if (Args[i].isByVal) {
6428 PointerType *Ty = cast<PointerType>(Args[i].Ty);
6429 Type *ElementTy = Ty->getElementType();
6430 Flags.setByValSize(getTargetData()->getTypeAllocSize(ElementTy));
6431 // For ByVal, alignment should come from FE. BE will guess if this
6432 // info is not there but there are cases it cannot get right.
6433 unsigned FrameAlign;
6434 if (Args[i].Alignment)
6435 FrameAlign = Args[i].Alignment;
6437 FrameAlign = getByValTypeAlignment(ElementTy);
6438 Flags.setByValAlign(FrameAlign);
6442 Flags.setOrigAlign(OriginalAlignment);
6444 EVT PartVT = getRegisterType(RetTy->getContext(), VT);
6445 unsigned NumParts = getNumRegisters(RetTy->getContext(), VT);
6446 SmallVector<SDValue, 4> Parts(NumParts);
6447 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
6450 ExtendKind = ISD::SIGN_EXTEND;
6451 else if (Args[i].isZExt)
6452 ExtendKind = ISD::ZERO_EXTEND;
6454 getCopyToParts(DAG, dl, Op, &Parts[0], NumParts,
6455 PartVT, ExtendKind);
6457 for (unsigned j = 0; j != NumParts; ++j) {
6458 // if it isn't first piece, alignment must be 1
6459 ISD::OutputArg MyFlags(Flags, Parts[j].getValueType(),
6461 if (NumParts > 1 && j == 0)
6462 MyFlags.Flags.setSplit();
6464 MyFlags.Flags.setOrigAlign(1);
6466 Outs.push_back(MyFlags);
6467 OutVals.push_back(Parts[j]);
6472 // Handle the incoming return values from the call.
6473 SmallVector<ISD::InputArg, 32> Ins;
6474 SmallVector<EVT, 4> RetTys;
6475 ComputeValueVTs(*this, RetTy, RetTys);
6476 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
6478 EVT RegisterVT = getRegisterType(RetTy->getContext(), VT);
6479 unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT);
6480 for (unsigned i = 0; i != NumRegs; ++i) {
6481 ISD::InputArg MyFlags;
6482 MyFlags.VT = RegisterVT.getSimpleVT();
6483 MyFlags.Used = isReturnValueUsed;
6485 MyFlags.Flags.setSExt();
6487 MyFlags.Flags.setZExt();
6489 MyFlags.Flags.setInReg();
6490 Ins.push_back(MyFlags);
6494 SmallVector<SDValue, 4> InVals;
6495 Chain = LowerCall(Chain, Callee, CallConv, isVarArg, isTailCall,
6496 Outs, OutVals, Ins, dl, DAG, InVals);
6498 // Verify that the target's LowerCall behaved as expected.
6499 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
6500 "LowerCall didn't return a valid chain!");
6501 assert((!isTailCall || InVals.empty()) &&
6502 "LowerCall emitted a return value for a tail call!");
6503 assert((isTailCall || InVals.size() == Ins.size()) &&
6504 "LowerCall didn't emit the correct number of values!");
6506 // For a tail call, the return value is merely live-out and there aren't
6507 // any nodes in the DAG representing it. Return a special value to
6508 // indicate that a tail call has been emitted and no more Instructions
6509 // should be processed in the current block.
6512 return std::make_pair(SDValue(), SDValue());
6515 DEBUG(for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
6516 assert(InVals[i].getNode() &&
6517 "LowerCall emitted a null value!");
6518 assert(EVT(Ins[i].VT) == InVals[i].getValueType() &&
6519 "LowerCall emitted a value with the wrong type!");
6522 // Collect the legal value parts into potentially illegal values
6523 // that correspond to the original function's return values.
6524 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6526 AssertOp = ISD::AssertSext;
6528 AssertOp = ISD::AssertZext;
6529 SmallVector<SDValue, 4> ReturnValues;
6530 unsigned CurReg = 0;
6531 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
6533 EVT RegisterVT = getRegisterType(RetTy->getContext(), VT);
6534 unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT);
6536 ReturnValues.push_back(getCopyFromParts(DAG, dl, &InVals[CurReg],
6537 NumRegs, RegisterVT, VT,
6542 // For a function returning void, there is no return value. We can't create
6543 // such a node, so we just return a null return value in that case. In
6544 // that case, nothing will actually look at the value.
6545 if (ReturnValues.empty())
6546 return std::make_pair(SDValue(), Chain);
6548 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, dl,
6549 DAG.getVTList(&RetTys[0], RetTys.size()),
6550 &ReturnValues[0], ReturnValues.size());
6551 return std::make_pair(Res, Chain);
6554 void TargetLowering::LowerOperationWrapper(SDNode *N,
6555 SmallVectorImpl<SDValue> &Results,
6556 SelectionDAG &DAG) const {
6557 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
6559 Results.push_back(Res);
6562 SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6563 llvm_unreachable("LowerOperation not implemented for this target!");
6567 SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) {
6568 SDValue Op = getNonRegisterValue(V);
6569 assert((Op.getOpcode() != ISD::CopyFromReg ||
6570 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
6571 "Copy from a reg to the same reg!");
6572 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg");
6574 RegsForValue RFV(V->getContext(), TLI, Reg, V->getType());
6575 SDValue Chain = DAG.getEntryNode();
6576 RFV.getCopyToRegs(Op, DAG, getCurDebugLoc(), Chain, 0);
6577 PendingExports.push_back(Chain);
6580 #include "llvm/CodeGen/SelectionDAGISel.h"
6582 /// isOnlyUsedInEntryBlock - If the specified argument is only used in the
6583 /// entry block, return true. This includes arguments used by switches, since
6584 /// the switch may expand into multiple basic blocks.
6585 static bool isOnlyUsedInEntryBlock(const Argument *A, bool FastISel) {
6586 // With FastISel active, we may be splitting blocks, so force creation
6587 // of virtual registers for all non-dead arguments.
6589 return A->use_empty();
6591 const BasicBlock *Entry = A->getParent()->begin();
6592 for (Value::const_use_iterator UI = A->use_begin(), E = A->use_end();
6594 const User *U = *UI;
6595 if (cast<Instruction>(U)->getParent() != Entry || isa<SwitchInst>(U))
6596 return false; // Use not in entry block.
6601 void SelectionDAGISel::LowerArguments(const BasicBlock *LLVMBB) {
6602 // If this is the entry block, emit arguments.
6603 const Function &F = *LLVMBB->getParent();
6604 SelectionDAG &DAG = SDB->DAG;
6605 DebugLoc dl = SDB->getCurDebugLoc();
6606 const TargetData *TD = TLI.getTargetData();
6607 SmallVector<ISD::InputArg, 16> Ins;
6609 // Check whether the function can return without sret-demotion.
6610 SmallVector<ISD::OutputArg, 4> Outs;
6611 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
6614 if (!FuncInfo->CanLowerReturn) {
6615 // Put in an sret pointer parameter before all the other parameters.
6616 SmallVector<EVT, 1> ValueVTs;
6617 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
6619 // NOTE: Assuming that a pointer will never break down to more than one VT
6621 ISD::ArgFlagsTy Flags;
6623 EVT RegisterVT = TLI.getRegisterType(*DAG.getContext(), ValueVTs[0]);
6624 ISD::InputArg RetArg(Flags, RegisterVT, true);
6625 Ins.push_back(RetArg);
6628 // Set up the incoming argument description vector.
6630 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end();
6631 I != E; ++I, ++Idx) {
6632 SmallVector<EVT, 4> ValueVTs;
6633 ComputeValueVTs(TLI, I->getType(), ValueVTs);
6634 bool isArgValueUsed = !I->use_empty();
6635 for (unsigned Value = 0, NumValues = ValueVTs.size();
6636 Value != NumValues; ++Value) {
6637 EVT VT = ValueVTs[Value];
6638 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
6639 ISD::ArgFlagsTy Flags;
6640 unsigned OriginalAlignment =
6641 TD->getABITypeAlignment(ArgTy);
6643 if (F.paramHasAttr(Idx, Attribute::ZExt))
6645 if (F.paramHasAttr(Idx, Attribute::SExt))
6647 if (F.paramHasAttr(Idx, Attribute::InReg))
6649 if (F.paramHasAttr(Idx, Attribute::StructRet))
6651 if (F.paramHasAttr(Idx, Attribute::ByVal)) {
6653 PointerType *Ty = cast<PointerType>(I->getType());
6654 Type *ElementTy = Ty->getElementType();
6655 Flags.setByValSize(TD->getTypeAllocSize(ElementTy));
6656 // For ByVal, alignment should be passed from FE. BE will guess if
6657 // this info is not there but there are cases it cannot get right.
6658 unsigned FrameAlign;
6659 if (F.getParamAlignment(Idx))
6660 FrameAlign = F.getParamAlignment(Idx);
6662 FrameAlign = TLI.getByValTypeAlignment(ElementTy);
6663 Flags.setByValAlign(FrameAlign);
6665 if (F.paramHasAttr(Idx, Attribute::Nest))
6667 Flags.setOrigAlign(OriginalAlignment);
6669 EVT RegisterVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6670 unsigned NumRegs = TLI.getNumRegisters(*CurDAG->getContext(), VT);
6671 for (unsigned i = 0; i != NumRegs; ++i) {
6672 ISD::InputArg MyFlags(Flags, RegisterVT, isArgValueUsed);
6673 if (NumRegs > 1 && i == 0)
6674 MyFlags.Flags.setSplit();
6675 // if it isn't first piece, alignment must be 1
6677 MyFlags.Flags.setOrigAlign(1);
6678 Ins.push_back(MyFlags);
6683 // Call the target to set up the argument values.
6684 SmallVector<SDValue, 8> InVals;
6685 SDValue NewRoot = TLI.LowerFormalArguments(DAG.getRoot(), F.getCallingConv(),
6689 // Verify that the target's LowerFormalArguments behaved as expected.
6690 assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other &&
6691 "LowerFormalArguments didn't return a valid chain!");
6692 assert(InVals.size() == Ins.size() &&
6693 "LowerFormalArguments didn't emit the correct number of values!");
6695 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
6696 assert(InVals[i].getNode() &&
6697 "LowerFormalArguments emitted a null value!");
6698 assert(EVT(Ins[i].VT) == InVals[i].getValueType() &&
6699 "LowerFormalArguments emitted a value with the wrong type!");
6703 // Update the DAG with the new chain value resulting from argument lowering.
6704 DAG.setRoot(NewRoot);
6706 // Set up the argument values.
6709 if (!FuncInfo->CanLowerReturn) {
6710 // Create a virtual register for the sret pointer, and put in a copy
6711 // from the sret argument into it.
6712 SmallVector<EVT, 1> ValueVTs;
6713 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
6714 EVT VT = ValueVTs[0];
6715 EVT RegVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6716 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6717 SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1,
6718 RegVT, VT, AssertOp);
6720 MachineFunction& MF = SDB->DAG.getMachineFunction();
6721 MachineRegisterInfo& RegInfo = MF.getRegInfo();
6722 unsigned SRetReg = RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT));
6723 FuncInfo->DemoteRegister = SRetReg;
6724 NewRoot = SDB->DAG.getCopyToReg(NewRoot, SDB->getCurDebugLoc(),
6726 DAG.setRoot(NewRoot);
6728 // i indexes lowered arguments. Bump it past the hidden sret argument.
6729 // Idx indexes LLVM arguments. Don't touch it.
6733 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
6735 SmallVector<SDValue, 4> ArgValues;
6736 SmallVector<EVT, 4> ValueVTs;
6737 ComputeValueVTs(TLI, I->getType(), ValueVTs);
6738 unsigned NumValues = ValueVTs.size();
6740 // If this argument is unused then remember its value. It is used to generate
6741 // debugging information.
6742 if (I->use_empty() && NumValues)
6743 SDB->setUnusedArgValue(I, InVals[i]);
6745 for (unsigned Val = 0; Val != NumValues; ++Val) {
6746 EVT VT = ValueVTs[Val];
6747 EVT PartVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6748 unsigned NumParts = TLI.getNumRegisters(*CurDAG->getContext(), VT);
6750 if (!I->use_empty()) {
6751 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6752 if (F.paramHasAttr(Idx, Attribute::SExt))
6753 AssertOp = ISD::AssertSext;
6754 else if (F.paramHasAttr(Idx, Attribute::ZExt))
6755 AssertOp = ISD::AssertZext;
6757 ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i],
6758 NumParts, PartVT, VT,
6765 // We don't need to do anything else for unused arguments.
6766 if (ArgValues.empty())
6769 // Note down frame index.
6770 if (FrameIndexSDNode *FI =
6771 dyn_cast<FrameIndexSDNode>(ArgValues[0].getNode()))
6772 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
6774 SDValue Res = DAG.getMergeValues(&ArgValues[0], NumValues,
6775 SDB->getCurDebugLoc());
6777 SDB->setValue(I, Res);
6778 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::BUILD_PAIR) {
6779 if (LoadSDNode *LNode =
6780 dyn_cast<LoadSDNode>(Res.getOperand(0).getNode()))
6781 if (FrameIndexSDNode *FI =
6782 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode()))
6783 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
6786 // If this argument is live outside of the entry block, insert a copy from
6787 // wherever we got it to the vreg that other BB's will reference it as.
6788 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::CopyFromReg) {
6789 // If we can, though, try to skip creating an unnecessary vreg.
6790 // FIXME: This isn't very clean... it would be nice to make this more
6791 // general. It's also subtly incompatible with the hacks FastISel
6793 unsigned Reg = cast<RegisterSDNode>(Res.getOperand(1))->getReg();
6794 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
6795 FuncInfo->ValueMap[I] = Reg;
6799 if (!isOnlyUsedInEntryBlock(I, TM.Options.EnableFastISel)) {
6800 FuncInfo->InitializeRegForValue(I);
6801 SDB->CopyToExportRegsIfNeeded(I);
6805 assert(i == InVals.size() && "Argument register count mismatch!");
6807 // Finally, if the target has anything special to do, allow it to do so.
6808 // FIXME: this should insert code into the DAG!
6809 EmitFunctionEntryCode();
6812 /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
6813 /// ensure constants are generated when needed. Remember the virtual registers
6814 /// that need to be added to the Machine PHI nodes as input. We cannot just
6815 /// directly add them, because expansion might result in multiple MBB's for one
6816 /// BB. As such, the start of the BB might correspond to a different MBB than
6820 SelectionDAGBuilder::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
6821 const TerminatorInst *TI = LLVMBB->getTerminator();
6823 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
6825 // Check successor nodes' PHI nodes that expect a constant to be available
6827 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
6828 const BasicBlock *SuccBB = TI->getSuccessor(succ);
6829 if (!isa<PHINode>(SuccBB->begin())) continue;
6830 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
6832 // If this terminator has multiple identical successors (common for
6833 // switches), only handle each succ once.
6834 if (!SuccsHandled.insert(SuccMBB)) continue;
6836 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
6838 // At this point we know that there is a 1-1 correspondence between LLVM PHI
6839 // nodes and Machine PHI nodes, but the incoming operands have not been
6841 for (BasicBlock::const_iterator I = SuccBB->begin();
6842 const PHINode *PN = dyn_cast<PHINode>(I); ++I) {
6843 // Ignore dead phi's.
6844 if (PN->use_empty()) continue;
6847 if (PN->getType()->isEmptyTy())
6851 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
6853 if (const Constant *C = dyn_cast<Constant>(PHIOp)) {
6854 unsigned &RegOut = ConstantsOut[C];
6856 RegOut = FuncInfo.CreateRegs(C->getType());
6857 CopyValueToVirtualRegister(C, RegOut);
6861 DenseMap<const Value *, unsigned>::iterator I =
6862 FuncInfo.ValueMap.find(PHIOp);
6863 if (I != FuncInfo.ValueMap.end())
6866 assert(isa<AllocaInst>(PHIOp) &&
6867 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
6868 "Didn't codegen value into a register!??");
6869 Reg = FuncInfo.CreateRegs(PHIOp->getType());
6870 CopyValueToVirtualRegister(PHIOp, Reg);
6874 // Remember that this register needs to added to the machine PHI node as
6875 // the input for this MBB.
6876 SmallVector<EVT, 4> ValueVTs;
6877 ComputeValueVTs(TLI, PN->getType(), ValueVTs);
6878 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
6879 EVT VT = ValueVTs[vti];
6880 unsigned NumRegisters = TLI.getNumRegisters(*DAG.getContext(), VT);
6881 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
6882 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
6883 Reg += NumRegisters;
6887 ConstantsOut.clear();