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/PseudoSourceValue.h"
45 #include "llvm/CodeGen/SelectionDAG.h"
46 #include "llvm/Analysis/DebugInfo.h"
47 #include "llvm/Target/TargetData.h"
48 #include "llvm/Target/TargetFrameLowering.h"
49 #include "llvm/Target/TargetInstrInfo.h"
50 #include "llvm/Target/TargetIntrinsicInfo.h"
51 #include "llvm/Target/TargetLowering.h"
52 #include "llvm/Target/TargetOptions.h"
53 #include "llvm/Support/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.getIntPtrConstant(1));
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!");
212 /// getCopyFromParts - Create a value that contains the specified legal parts
213 /// combined into the value they represent. If the parts combine to a type
214 /// larger then ValueVT then AssertOp can be used to specify whether the extra
215 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
216 /// (ISD::AssertSext).
217 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, DebugLoc DL,
218 const SDValue *Parts, unsigned NumParts,
219 EVT PartVT, EVT ValueVT) {
220 assert(ValueVT.isVector() && "Not a vector value");
221 assert(NumParts > 0 && "No parts to assemble!");
222 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
223 SDValue Val = Parts[0];
225 // Handle a multi-element vector.
227 EVT IntermediateVT, RegisterVT;
228 unsigned NumIntermediates;
230 TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT,
231 NumIntermediates, RegisterVT);
232 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
233 NumParts = NumRegs; // Silence a compiler warning.
234 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
235 assert(RegisterVT == Parts[0].getValueType() &&
236 "Part type doesn't match part!");
238 // Assemble the parts into intermediate operands.
239 SmallVector<SDValue, 8> Ops(NumIntermediates);
240 if (NumIntermediates == NumParts) {
241 // If the register was not expanded, truncate or copy the value,
243 for (unsigned i = 0; i != NumParts; ++i)
244 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i], 1,
245 PartVT, IntermediateVT);
246 } else if (NumParts > 0) {
247 // If the intermediate type was expanded, build the intermediate
248 // operands from the parts.
249 assert(NumParts % NumIntermediates == 0 &&
250 "Must expand into a divisible number of parts!");
251 unsigned Factor = NumParts / NumIntermediates;
252 for (unsigned i = 0; i != NumIntermediates; ++i)
253 Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor,
254 PartVT, IntermediateVT);
257 // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the
258 // intermediate operands.
259 Val = DAG.getNode(IntermediateVT.isVector() ?
260 ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR, DL,
261 ValueVT, &Ops[0], NumIntermediates);
264 // There is now one part, held in Val. Correct it to match ValueVT.
265 PartVT = Val.getValueType();
267 if (PartVT == ValueVT)
270 if (PartVT.isVector()) {
271 // If the element type of the source/dest vectors are the same, but the
272 // parts vector has more elements than the value vector, then we have a
273 // vector widening case (e.g. <2 x float> -> <4 x float>). Extract the
275 if (PartVT.getVectorElementType() == ValueVT.getVectorElementType()) {
276 assert(PartVT.getVectorNumElements() > ValueVT.getVectorNumElements() &&
277 "Cannot narrow, it would be a lossy transformation");
278 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val,
279 DAG.getIntPtrConstant(0));
282 // Vector/Vector bitcast.
283 if (ValueVT.getSizeInBits() == PartVT.getSizeInBits())
284 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
286 assert(PartVT.getVectorNumElements() == ValueVT.getVectorNumElements() &&
287 "Cannot handle this kind of promotion");
288 // Promoted vector extract
289 bool Smaller = ValueVT.bitsLE(PartVT);
290 return DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
295 // Trivial bitcast if the types are the same size and the destination
296 // vector type is legal.
297 if (PartVT.getSizeInBits() == ValueVT.getSizeInBits() &&
298 TLI.isTypeLegal(ValueVT))
299 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
301 // Handle cases such as i8 -> <1 x i1>
302 assert(ValueVT.getVectorNumElements() == 1 &&
303 "Only trivial scalar-to-vector conversions should get here!");
305 if (ValueVT.getVectorNumElements() == 1 &&
306 ValueVT.getVectorElementType() != PartVT) {
307 bool Smaller = ValueVT.bitsLE(PartVT);
308 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
309 DL, ValueVT.getScalarType(), Val);
312 return DAG.getNode(ISD::BUILD_VECTOR, DL, ValueVT, Val);
318 static void getCopyToPartsVector(SelectionDAG &DAG, DebugLoc dl,
319 SDValue Val, SDValue *Parts, unsigned NumParts,
322 /// getCopyToParts - Create a series of nodes that contain the specified value
323 /// split into legal parts. If the parts contain more bits than Val, then, for
324 /// integers, ExtendKind can be used to specify how to generate the extra bits.
325 static void getCopyToParts(SelectionDAG &DAG, DebugLoc DL,
326 SDValue Val, SDValue *Parts, unsigned NumParts,
328 ISD::NodeType ExtendKind = ISD::ANY_EXTEND) {
329 EVT ValueVT = Val.getValueType();
331 // Handle the vector case separately.
332 if (ValueVT.isVector())
333 return getCopyToPartsVector(DAG, DL, Val, Parts, NumParts, PartVT);
335 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
336 unsigned PartBits = PartVT.getSizeInBits();
337 unsigned OrigNumParts = NumParts;
338 assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!");
343 assert(!ValueVT.isVector() && "Vector case handled elsewhere");
344 if (PartVT == ValueVT) {
345 assert(NumParts == 1 && "No-op copy with multiple parts!");
350 if (NumParts * PartBits > ValueVT.getSizeInBits()) {
351 // If the parts cover more bits than the value has, promote the value.
352 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
353 assert(NumParts == 1 && "Do not know what to promote to!");
354 Val = DAG.getNode(ISD::FP_EXTEND, DL, PartVT, Val);
356 assert(PartVT.isInteger() && ValueVT.isInteger() &&
357 "Unknown mismatch!");
358 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
359 Val = DAG.getNode(ExtendKind, DL, ValueVT, Val);
361 } else if (PartBits == ValueVT.getSizeInBits()) {
362 // Different types of the same size.
363 assert(NumParts == 1 && PartVT != ValueVT);
364 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
365 } else if (NumParts * PartBits < ValueVT.getSizeInBits()) {
366 // If the parts cover less bits than value has, truncate the value.
367 assert(PartVT.isInteger() && ValueVT.isInteger() &&
368 "Unknown mismatch!");
369 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
370 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
373 // The value may have changed - recompute ValueVT.
374 ValueVT = Val.getValueType();
375 assert(NumParts * PartBits == ValueVT.getSizeInBits() &&
376 "Failed to tile the value with PartVT!");
379 assert(PartVT == ValueVT && "Type conversion failed!");
384 // Expand the value into multiple parts.
385 if (NumParts & (NumParts - 1)) {
386 // The number of parts is not a power of 2. Split off and copy the tail.
387 assert(PartVT.isInteger() && ValueVT.isInteger() &&
388 "Do not know what to expand to!");
389 unsigned RoundParts = 1 << Log2_32(NumParts);
390 unsigned RoundBits = RoundParts * PartBits;
391 unsigned OddParts = NumParts - RoundParts;
392 SDValue OddVal = DAG.getNode(ISD::SRL, DL, ValueVT, Val,
393 DAG.getIntPtrConstant(RoundBits));
394 getCopyToParts(DAG, DL, OddVal, Parts + RoundParts, OddParts, PartVT);
396 if (TLI.isBigEndian())
397 // The odd parts were reversed by getCopyToParts - unreverse them.
398 std::reverse(Parts + RoundParts, Parts + NumParts);
400 NumParts = RoundParts;
401 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
402 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
405 // The number of parts is a power of 2. Repeatedly bisect the value using
407 Parts[0] = DAG.getNode(ISD::BITCAST, DL,
408 EVT::getIntegerVT(*DAG.getContext(),
409 ValueVT.getSizeInBits()),
412 for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) {
413 for (unsigned i = 0; i < NumParts; i += StepSize) {
414 unsigned ThisBits = StepSize * PartBits / 2;
415 EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits);
416 SDValue &Part0 = Parts[i];
417 SDValue &Part1 = Parts[i+StepSize/2];
419 Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
420 ThisVT, Part0, DAG.getIntPtrConstant(1));
421 Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
422 ThisVT, Part0, DAG.getIntPtrConstant(0));
424 if (ThisBits == PartBits && ThisVT != PartVT) {
425 Part0 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part0);
426 Part1 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part1);
431 if (TLI.isBigEndian())
432 std::reverse(Parts, Parts + OrigNumParts);
436 /// getCopyToPartsVector - Create a series of nodes that contain the specified
437 /// value split into legal parts.
438 static void getCopyToPartsVector(SelectionDAG &DAG, DebugLoc DL,
439 SDValue Val, SDValue *Parts, unsigned NumParts,
441 EVT ValueVT = Val.getValueType();
442 assert(ValueVT.isVector() && "Not a vector");
443 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
446 if (PartVT == ValueVT) {
448 } else if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) {
449 // Bitconvert vector->vector case.
450 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
451 } else if (PartVT.isVector() &&
452 PartVT.getVectorElementType() == ValueVT.getVectorElementType() &&
453 PartVT.getVectorNumElements() > ValueVT.getVectorNumElements()) {
454 EVT ElementVT = PartVT.getVectorElementType();
455 // Vector widening case, e.g. <2 x float> -> <4 x float>. Shuffle in
457 SmallVector<SDValue, 16> Ops;
458 for (unsigned i = 0, e = ValueVT.getVectorNumElements(); i != e; ++i)
459 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
460 ElementVT, Val, DAG.getIntPtrConstant(i)));
462 for (unsigned i = ValueVT.getVectorNumElements(),
463 e = PartVT.getVectorNumElements(); i != e; ++i)
464 Ops.push_back(DAG.getUNDEF(ElementVT));
466 Val = DAG.getNode(ISD::BUILD_VECTOR, DL, PartVT, &Ops[0], Ops.size());
468 // FIXME: Use CONCAT for 2x -> 4x.
470 //SDValue UndefElts = DAG.getUNDEF(VectorTy);
471 //Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts);
472 } else if (PartVT.isVector() &&
473 PartVT.getVectorElementType().bitsGE(
474 ValueVT.getVectorElementType()) &&
475 PartVT.getVectorNumElements() == ValueVT.getVectorNumElements()) {
477 // Promoted vector extract
478 bool Smaller = PartVT.bitsLE(ValueVT);
479 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
482 // Vector -> scalar conversion.
483 assert(ValueVT.getVectorNumElements() == 1 &&
484 "Only trivial vector-to-scalar conversions should get here!");
485 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
486 PartVT, Val, DAG.getIntPtrConstant(0));
488 bool Smaller = ValueVT.bitsLE(PartVT);
489 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
497 // Handle a multi-element vector.
498 EVT IntermediateVT, RegisterVT;
499 unsigned NumIntermediates;
500 unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT,
502 NumIntermediates, RegisterVT);
503 unsigned NumElements = ValueVT.getVectorNumElements();
505 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
506 NumParts = NumRegs; // Silence a compiler warning.
507 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
509 // Split the vector into intermediate operands.
510 SmallVector<SDValue, 8> Ops(NumIntermediates);
511 for (unsigned i = 0; i != NumIntermediates; ++i) {
512 if (IntermediateVT.isVector())
513 Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL,
515 DAG.getIntPtrConstant(i * (NumElements / NumIntermediates)));
517 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
518 IntermediateVT, Val, DAG.getIntPtrConstant(i));
521 // Split the intermediate operands into legal parts.
522 if (NumParts == NumIntermediates) {
523 // If the register was not expanded, promote or copy the value,
525 for (unsigned i = 0; i != NumParts; ++i)
526 getCopyToParts(DAG, DL, Ops[i], &Parts[i], 1, PartVT);
527 } else if (NumParts > 0) {
528 // If the intermediate type was expanded, split each the value into
530 assert(NumParts % NumIntermediates == 0 &&
531 "Must expand into a divisible number of parts!");
532 unsigned Factor = NumParts / NumIntermediates;
533 for (unsigned i = 0; i != NumIntermediates; ++i)
534 getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT);
542 /// RegsForValue - This struct represents the registers (physical or virtual)
543 /// that a particular set of values is assigned, and the type information
544 /// about the value. The most common situation is to represent one value at a
545 /// time, but struct or array values are handled element-wise as multiple
546 /// values. The splitting of aggregates is performed recursively, so that we
547 /// never have aggregate-typed registers. The values at this point do not
548 /// necessarily have legal types, so each value may require one or more
549 /// registers of some legal type.
551 struct RegsForValue {
552 /// ValueVTs - The value types of the values, which may not be legal, and
553 /// may need be promoted or synthesized from one or more registers.
555 SmallVector<EVT, 4> ValueVTs;
557 /// RegVTs - The value types of the registers. This is the same size as
558 /// ValueVTs and it records, for each value, what the type of the assigned
559 /// register or registers are. (Individual values are never synthesized
560 /// from more than one type of register.)
562 /// With virtual registers, the contents of RegVTs is redundant with TLI's
563 /// getRegisterType member function, however when with physical registers
564 /// it is necessary to have a separate record of the types.
566 SmallVector<EVT, 4> RegVTs;
568 /// Regs - This list holds the registers assigned to the values.
569 /// Each legal or promoted value requires one register, and each
570 /// expanded value requires multiple registers.
572 SmallVector<unsigned, 4> Regs;
576 RegsForValue(const SmallVector<unsigned, 4> ®s,
577 EVT regvt, EVT valuevt)
578 : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {}
580 RegsForValue(LLVMContext &Context, const TargetLowering &tli,
581 unsigned Reg, Type *Ty) {
582 ComputeValueVTs(tli, Ty, ValueVTs);
584 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
585 EVT ValueVT = ValueVTs[Value];
586 unsigned NumRegs = tli.getNumRegisters(Context, ValueVT);
587 EVT RegisterVT = tli.getRegisterType(Context, ValueVT);
588 for (unsigned i = 0; i != NumRegs; ++i)
589 Regs.push_back(Reg + i);
590 RegVTs.push_back(RegisterVT);
595 /// areValueTypesLegal - Return true if types of all the values are legal.
596 bool areValueTypesLegal(const TargetLowering &TLI) {
597 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
598 EVT RegisterVT = RegVTs[Value];
599 if (!TLI.isTypeLegal(RegisterVT))
605 /// append - Add the specified values to this one.
606 void append(const RegsForValue &RHS) {
607 ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end());
608 RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end());
609 Regs.append(RHS.Regs.begin(), RHS.Regs.end());
612 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
613 /// this value and returns the result as a ValueVTs value. This uses
614 /// Chain/Flag as the input and updates them for the output Chain/Flag.
615 /// If the Flag pointer is NULL, no flag is used.
616 SDValue getCopyFromRegs(SelectionDAG &DAG, FunctionLoweringInfo &FuncInfo,
618 SDValue &Chain, SDValue *Flag) const;
620 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
621 /// specified value into the registers specified by this object. This uses
622 /// Chain/Flag as the input and updates them for the output Chain/Flag.
623 /// If the Flag pointer is NULL, no flag is used.
624 void getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
625 SDValue &Chain, SDValue *Flag) const;
627 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
628 /// operand list. This adds the code marker, matching input operand index
629 /// (if applicable), and includes the number of values added into it.
630 void AddInlineAsmOperands(unsigned Kind,
631 bool HasMatching, unsigned MatchingIdx,
633 std::vector<SDValue> &Ops) const;
637 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
638 /// this value and returns the result as a ValueVT value. This uses
639 /// Chain/Flag as the input and updates them for the output Chain/Flag.
640 /// If the Flag pointer is NULL, no flag is used.
641 SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
642 FunctionLoweringInfo &FuncInfo,
644 SDValue &Chain, SDValue *Flag) const {
645 // A Value with type {} or [0 x %t] needs no registers.
646 if (ValueVTs.empty())
649 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
651 // Assemble the legal parts into the final values.
652 SmallVector<SDValue, 4> Values(ValueVTs.size());
653 SmallVector<SDValue, 8> Parts;
654 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
655 // Copy the legal parts from the registers.
656 EVT ValueVT = ValueVTs[Value];
657 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
658 EVT RegisterVT = RegVTs[Value];
660 Parts.resize(NumRegs);
661 for (unsigned i = 0; i != NumRegs; ++i) {
664 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT);
666 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag);
667 *Flag = P.getValue(2);
670 Chain = P.getValue(1);
673 // If the source register was virtual and if we know something about it,
674 // add an assert node.
675 if (!TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) ||
676 !RegisterVT.isInteger() || RegisterVT.isVector())
679 const FunctionLoweringInfo::LiveOutInfo *LOI =
680 FuncInfo.GetLiveOutRegInfo(Regs[Part+i]);
684 unsigned RegSize = RegisterVT.getSizeInBits();
685 unsigned NumSignBits = LOI->NumSignBits;
686 unsigned NumZeroBits = LOI->KnownZero.countLeadingOnes();
688 // FIXME: We capture more information than the dag can represent. For
689 // now, just use the tightest assertzext/assertsext possible.
691 EVT FromVT(MVT::Other);
692 if (NumSignBits == RegSize)
693 isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1
694 else if (NumZeroBits >= RegSize-1)
695 isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1
696 else if (NumSignBits > RegSize-8)
697 isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8
698 else if (NumZeroBits >= RegSize-8)
699 isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8
700 else if (NumSignBits > RegSize-16)
701 isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16
702 else if (NumZeroBits >= RegSize-16)
703 isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16
704 else if (NumSignBits > RegSize-32)
705 isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32
706 else if (NumZeroBits >= RegSize-32)
707 isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32
711 // Add an assertion node.
712 assert(FromVT != MVT::Other);
713 Parts[i] = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl,
714 RegisterVT, P, DAG.getValueType(FromVT));
717 Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(),
718 NumRegs, RegisterVT, ValueVT);
723 return DAG.getNode(ISD::MERGE_VALUES, dl,
724 DAG.getVTList(&ValueVTs[0], ValueVTs.size()),
725 &Values[0], ValueVTs.size());
728 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
729 /// specified value into the registers specified by this object. This uses
730 /// Chain/Flag as the input and updates them for the output Chain/Flag.
731 /// If the Flag pointer is NULL, no flag is used.
732 void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
733 SDValue &Chain, SDValue *Flag) const {
734 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
736 // Get the list of the values's legal parts.
737 unsigned NumRegs = Regs.size();
738 SmallVector<SDValue, 8> Parts(NumRegs);
739 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
740 EVT ValueVT = ValueVTs[Value];
741 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
742 EVT RegisterVT = RegVTs[Value];
744 getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value),
745 &Parts[Part], NumParts, RegisterVT);
749 // Copy the parts into the registers.
750 SmallVector<SDValue, 8> Chains(NumRegs);
751 for (unsigned i = 0; i != NumRegs; ++i) {
754 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]);
756 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag);
757 *Flag = Part.getValue(1);
760 Chains[i] = Part.getValue(0);
763 if (NumRegs == 1 || Flag)
764 // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is
765 // flagged to it. That is the CopyToReg nodes and the user are considered
766 // a single scheduling unit. If we create a TokenFactor and return it as
767 // chain, then the TokenFactor is both a predecessor (operand) of the
768 // user as well as a successor (the TF operands are flagged to the user).
769 // c1, f1 = CopyToReg
770 // c2, f2 = CopyToReg
771 // c3 = TokenFactor c1, c2
774 Chain = Chains[NumRegs-1];
776 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0], NumRegs);
779 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
780 /// operand list. This adds the code marker and includes the number of
781 /// values added into it.
782 void RegsForValue::AddInlineAsmOperands(unsigned Code, bool HasMatching,
783 unsigned MatchingIdx,
785 std::vector<SDValue> &Ops) const {
786 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
788 unsigned Flag = InlineAsm::getFlagWord(Code, Regs.size());
790 Flag = InlineAsm::getFlagWordForMatchingOp(Flag, MatchingIdx);
791 SDValue Res = DAG.getTargetConstant(Flag, MVT::i32);
794 for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) {
795 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVTs[Value]);
796 EVT RegisterVT = RegVTs[Value];
797 for (unsigned i = 0; i != NumRegs; ++i) {
798 assert(Reg < Regs.size() && "Mismatch in # registers expected");
799 Ops.push_back(DAG.getRegister(Regs[Reg++], RegisterVT));
804 void SelectionDAGBuilder::init(GCFunctionInfo *gfi, AliasAnalysis &aa) {
807 TD = DAG.getTarget().getTargetData();
810 /// clear - Clear out the current SelectionDAG and the associated
811 /// state and prepare this SelectionDAGBuilder object to be used
812 /// for a new block. This doesn't clear out information about
813 /// additional blocks that are needed to complete switch lowering
814 /// or PHI node updating; that information is cleared out as it is
816 void SelectionDAGBuilder::clear() {
818 UnusedArgNodeMap.clear();
819 PendingLoads.clear();
820 PendingExports.clear();
821 CurDebugLoc = DebugLoc();
825 /// clearDanglingDebugInfo - Clear the dangling debug information
826 /// map. This function is seperated from the clear so that debug
827 /// information that is dangling in a basic block can be properly
828 /// resolved in a different basic block. This allows the
829 /// SelectionDAG to resolve dangling debug information attached
831 void SelectionDAGBuilder::clearDanglingDebugInfo() {
832 DanglingDebugInfoMap.clear();
835 /// getRoot - Return the current virtual root of the Selection DAG,
836 /// flushing any PendingLoad items. This must be done before emitting
837 /// a store or any other node that may need to be ordered after any
838 /// prior load instructions.
840 SDValue SelectionDAGBuilder::getRoot() {
841 if (PendingLoads.empty())
842 return DAG.getRoot();
844 if (PendingLoads.size() == 1) {
845 SDValue Root = PendingLoads[0];
847 PendingLoads.clear();
851 // Otherwise, we have to make a token factor node.
852 SDValue Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
853 &PendingLoads[0], PendingLoads.size());
854 PendingLoads.clear();
859 /// getControlRoot - Similar to getRoot, but instead of flushing all the
860 /// PendingLoad items, flush all the PendingExports items. It is necessary
861 /// to do this before emitting a terminator instruction.
863 SDValue SelectionDAGBuilder::getControlRoot() {
864 SDValue Root = DAG.getRoot();
866 if (PendingExports.empty())
869 // Turn all of the CopyToReg chains into one factored node.
870 if (Root.getOpcode() != ISD::EntryToken) {
871 unsigned i = 0, e = PendingExports.size();
872 for (; i != e; ++i) {
873 assert(PendingExports[i].getNode()->getNumOperands() > 1);
874 if (PendingExports[i].getNode()->getOperand(0) == Root)
875 break; // Don't add the root if we already indirectly depend on it.
879 PendingExports.push_back(Root);
882 Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
884 PendingExports.size());
885 PendingExports.clear();
890 void SelectionDAGBuilder::AssignOrderingToNode(const SDNode *Node) {
891 if (DAG.GetOrdering(Node) != 0) return; // Already has ordering.
892 DAG.AssignOrdering(Node, SDNodeOrder);
894 for (unsigned I = 0, E = Node->getNumOperands(); I != E; ++I)
895 AssignOrderingToNode(Node->getOperand(I).getNode());
898 void SelectionDAGBuilder::visit(const Instruction &I) {
899 // Set up outgoing PHI node register values before emitting the terminator.
900 if (isa<TerminatorInst>(&I))
901 HandlePHINodesInSuccessorBlocks(I.getParent());
903 CurDebugLoc = I.getDebugLoc();
905 visit(I.getOpcode(), I);
907 if (!isa<TerminatorInst>(&I) && !HasTailCall)
908 CopyToExportRegsIfNeeded(&I);
910 CurDebugLoc = DebugLoc();
913 void SelectionDAGBuilder::visitPHI(const PHINode &) {
914 llvm_unreachable("SelectionDAGBuilder shouldn't visit PHI nodes!");
917 void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) {
918 // Note: this doesn't use InstVisitor, because it has to work with
919 // ConstantExpr's in addition to instructions.
921 default: llvm_unreachable("Unknown instruction type encountered!");
922 // Build the switch statement using the Instruction.def file.
923 #define HANDLE_INST(NUM, OPCODE, CLASS) \
924 case Instruction::OPCODE: visit##OPCODE((CLASS&)I); break;
925 #include "llvm/Instruction.def"
928 // Assign the ordering to the freshly created DAG nodes.
929 if (NodeMap.count(&I)) {
931 AssignOrderingToNode(getValue(&I).getNode());
935 // resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V,
936 // generate the debug data structures now that we've seen its definition.
937 void SelectionDAGBuilder::resolveDanglingDebugInfo(const Value *V,
939 DanglingDebugInfo &DDI = DanglingDebugInfoMap[V];
941 const DbgValueInst *DI = DDI.getDI();
942 DebugLoc dl = DDI.getdl();
943 unsigned DbgSDNodeOrder = DDI.getSDNodeOrder();
944 MDNode *Variable = DI->getVariable();
945 uint64_t Offset = DI->getOffset();
948 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, Val)) {
949 SDV = DAG.getDbgValue(Variable, Val.getNode(),
950 Val.getResNo(), Offset, dl, DbgSDNodeOrder);
951 DAG.AddDbgValue(SDV, Val.getNode(), false);
954 DEBUG(dbgs() << "Dropping debug info for " << DI);
955 DanglingDebugInfoMap[V] = DanglingDebugInfo();
959 // getValue - Return an SDValue for the given Value.
960 SDValue SelectionDAGBuilder::getValue(const Value *V) {
961 // If we already have an SDValue for this value, use it. It's important
962 // to do this first, so that we don't create a CopyFromReg if we already
963 // have a regular SDValue.
964 SDValue &N = NodeMap[V];
965 if (N.getNode()) return N;
967 // If there's a virtual register allocated and initialized for this
969 DenseMap<const Value *, unsigned>::iterator It = FuncInfo.ValueMap.find(V);
970 if (It != FuncInfo.ValueMap.end()) {
971 unsigned InReg = It->second;
972 RegsForValue RFV(*DAG.getContext(), TLI, InReg, V->getType());
973 SDValue Chain = DAG.getEntryNode();
974 N = RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain,NULL);
975 resolveDanglingDebugInfo(V, N);
979 // Otherwise create a new SDValue and remember it.
980 SDValue Val = getValueImpl(V);
982 resolveDanglingDebugInfo(V, Val);
986 /// getNonRegisterValue - Return an SDValue for the given Value, but
987 /// don't look in FuncInfo.ValueMap for a virtual register.
988 SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) {
989 // If we already have an SDValue for this value, use it.
990 SDValue &N = NodeMap[V];
991 if (N.getNode()) return N;
993 // Otherwise create a new SDValue and remember it.
994 SDValue Val = getValueImpl(V);
996 resolveDanglingDebugInfo(V, Val);
1000 /// getValueImpl - Helper function for getValue and getNonRegisterValue.
1001 /// Create an SDValue for the given value.
1002 SDValue SelectionDAGBuilder::getValueImpl(const Value *V) {
1003 if (const Constant *C = dyn_cast<Constant>(V)) {
1004 EVT VT = TLI.getValueType(V->getType(), true);
1006 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C))
1007 return DAG.getConstant(*CI, VT);
1009 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
1010 return DAG.getGlobalAddress(GV, getCurDebugLoc(), VT);
1012 if (isa<ConstantPointerNull>(C))
1013 return DAG.getConstant(0, TLI.getPointerTy());
1015 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
1016 return DAG.getConstantFP(*CFP, VT);
1018 if (isa<UndefValue>(C) && !V->getType()->isAggregateType())
1019 return DAG.getUNDEF(VT);
1021 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
1022 visit(CE->getOpcode(), *CE);
1023 SDValue N1 = NodeMap[V];
1024 assert(N1.getNode() && "visit didn't populate the NodeMap!");
1028 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) {
1029 SmallVector<SDValue, 4> Constants;
1030 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end();
1032 SDNode *Val = getValue(*OI).getNode();
1033 // If the operand is an empty aggregate, there are no values.
1035 // Add each leaf value from the operand to the Constants list
1036 // to form a flattened list of all the values.
1037 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
1038 Constants.push_back(SDValue(Val, i));
1041 return DAG.getMergeValues(&Constants[0], Constants.size(),
1045 if (C->getType()->isStructTy() || C->getType()->isArrayTy()) {
1046 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
1047 "Unknown struct or array constant!");
1049 SmallVector<EVT, 4> ValueVTs;
1050 ComputeValueVTs(TLI, C->getType(), ValueVTs);
1051 unsigned NumElts = ValueVTs.size();
1053 return SDValue(); // empty struct
1054 SmallVector<SDValue, 4> Constants(NumElts);
1055 for (unsigned i = 0; i != NumElts; ++i) {
1056 EVT EltVT = ValueVTs[i];
1057 if (isa<UndefValue>(C))
1058 Constants[i] = DAG.getUNDEF(EltVT);
1059 else if (EltVT.isFloatingPoint())
1060 Constants[i] = DAG.getConstantFP(0, EltVT);
1062 Constants[i] = DAG.getConstant(0, EltVT);
1065 return DAG.getMergeValues(&Constants[0], NumElts,
1069 if (const BlockAddress *BA = dyn_cast<BlockAddress>(C))
1070 return DAG.getBlockAddress(BA, VT);
1072 VectorType *VecTy = cast<VectorType>(V->getType());
1073 unsigned NumElements = VecTy->getNumElements();
1075 // Now that we know the number and type of the elements, get that number of
1076 // elements into the Ops array based on what kind of constant it is.
1077 SmallVector<SDValue, 16> Ops;
1078 if (const ConstantVector *CP = dyn_cast<ConstantVector>(C)) {
1079 for (unsigned i = 0; i != NumElements; ++i)
1080 Ops.push_back(getValue(CP->getOperand(i)));
1082 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!");
1083 EVT EltVT = TLI.getValueType(VecTy->getElementType());
1086 if (EltVT.isFloatingPoint())
1087 Op = DAG.getConstantFP(0, EltVT);
1089 Op = DAG.getConstant(0, EltVT);
1090 Ops.assign(NumElements, Op);
1093 // Create a BUILD_VECTOR node.
1094 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
1095 VT, &Ops[0], Ops.size());
1098 // If this is a static alloca, generate it as the frameindex instead of
1100 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
1101 DenseMap<const AllocaInst*, int>::iterator SI =
1102 FuncInfo.StaticAllocaMap.find(AI);
1103 if (SI != FuncInfo.StaticAllocaMap.end())
1104 return DAG.getFrameIndex(SI->second, TLI.getPointerTy());
1107 // If this is an instruction which fast-isel has deferred, select it now.
1108 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
1109 unsigned InReg = FuncInfo.InitializeRegForValue(Inst);
1110 RegsForValue RFV(*DAG.getContext(), TLI, InReg, Inst->getType());
1111 SDValue Chain = DAG.getEntryNode();
1112 return RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain, NULL);
1115 llvm_unreachable("Can't get register for value!");
1119 void SelectionDAGBuilder::visitRet(const ReturnInst &I) {
1120 SDValue Chain = getControlRoot();
1121 SmallVector<ISD::OutputArg, 8> Outs;
1122 SmallVector<SDValue, 8> OutVals;
1124 if (!FuncInfo.CanLowerReturn) {
1125 unsigned DemoteReg = FuncInfo.DemoteRegister;
1126 const Function *F = I.getParent()->getParent();
1128 // Emit a store of the return value through the virtual register.
1129 // Leave Outs empty so that LowerReturn won't try to load return
1130 // registers the usual way.
1131 SmallVector<EVT, 1> PtrValueVTs;
1132 ComputeValueVTs(TLI, PointerType::getUnqual(F->getReturnType()),
1135 SDValue RetPtr = DAG.getRegister(DemoteReg, PtrValueVTs[0]);
1136 SDValue RetOp = getValue(I.getOperand(0));
1138 SmallVector<EVT, 4> ValueVTs;
1139 SmallVector<uint64_t, 4> Offsets;
1140 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs, &Offsets);
1141 unsigned NumValues = ValueVTs.size();
1143 SmallVector<SDValue, 4> Chains(NumValues);
1144 for (unsigned i = 0; i != NumValues; ++i) {
1145 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(),
1146 RetPtr.getValueType(), RetPtr,
1147 DAG.getIntPtrConstant(Offsets[i]));
1149 DAG.getStore(Chain, getCurDebugLoc(),
1150 SDValue(RetOp.getNode(), RetOp.getResNo() + i),
1151 // FIXME: better loc info would be nice.
1152 Add, MachinePointerInfo(), false, false, 0);
1155 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
1156 MVT::Other, &Chains[0], NumValues);
1157 } else if (I.getNumOperands() != 0) {
1158 SmallVector<EVT, 4> ValueVTs;
1159 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs);
1160 unsigned NumValues = ValueVTs.size();
1162 SDValue RetOp = getValue(I.getOperand(0));
1163 for (unsigned j = 0, f = NumValues; j != f; ++j) {
1164 EVT VT = ValueVTs[j];
1166 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1168 const Function *F = I.getParent()->getParent();
1169 if (F->paramHasAttr(0, Attribute::SExt))
1170 ExtendKind = ISD::SIGN_EXTEND;
1171 else if (F->paramHasAttr(0, Attribute::ZExt))
1172 ExtendKind = ISD::ZERO_EXTEND;
1174 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger())
1175 VT = TLI.getTypeForExtArgOrReturn(*DAG.getContext(), VT, ExtendKind);
1177 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), VT);
1178 EVT PartVT = TLI.getRegisterType(*DAG.getContext(), VT);
1179 SmallVector<SDValue, 4> Parts(NumParts);
1180 getCopyToParts(DAG, getCurDebugLoc(),
1181 SDValue(RetOp.getNode(), RetOp.getResNo() + j),
1182 &Parts[0], NumParts, PartVT, ExtendKind);
1184 // 'inreg' on function refers to return value
1185 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1186 if (F->paramHasAttr(0, Attribute::InReg))
1189 // Propagate extension type if any
1190 if (ExtendKind == ISD::SIGN_EXTEND)
1192 else if (ExtendKind == ISD::ZERO_EXTEND)
1195 for (unsigned i = 0; i < NumParts; ++i) {
1196 Outs.push_back(ISD::OutputArg(Flags, Parts[i].getValueType(),
1198 OutVals.push_back(Parts[i]);
1204 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1205 CallingConv::ID CallConv =
1206 DAG.getMachineFunction().getFunction()->getCallingConv();
1207 Chain = TLI.LowerReturn(Chain, CallConv, isVarArg,
1208 Outs, OutVals, getCurDebugLoc(), DAG);
1210 // Verify that the target's LowerReturn behaved as expected.
1211 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
1212 "LowerReturn didn't return a valid chain!");
1214 // Update the DAG with the new chain value resulting from return lowering.
1218 /// CopyToExportRegsIfNeeded - If the given value has virtual registers
1219 /// created for it, emit nodes to copy the value into the virtual
1221 void SelectionDAGBuilder::CopyToExportRegsIfNeeded(const Value *V) {
1223 if (V->getType()->isEmptyTy())
1226 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
1227 if (VMI != FuncInfo.ValueMap.end()) {
1228 assert(!V->use_empty() && "Unused value assigned virtual registers!");
1229 CopyValueToVirtualRegister(V, VMI->second);
1233 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
1234 /// the current basic block, add it to ValueMap now so that we'll get a
1236 void SelectionDAGBuilder::ExportFromCurrentBlock(const Value *V) {
1237 // No need to export constants.
1238 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
1240 // Already exported?
1241 if (FuncInfo.isExportedInst(V)) return;
1243 unsigned Reg = FuncInfo.InitializeRegForValue(V);
1244 CopyValueToVirtualRegister(V, Reg);
1247 bool SelectionDAGBuilder::isExportableFromCurrentBlock(const Value *V,
1248 const BasicBlock *FromBB) {
1249 // The operands of the setcc have to be in this block. We don't know
1250 // how to export them from some other block.
1251 if (const Instruction *VI = dyn_cast<Instruction>(V)) {
1252 // Can export from current BB.
1253 if (VI->getParent() == FromBB)
1256 // Is already exported, noop.
1257 return FuncInfo.isExportedInst(V);
1260 // If this is an argument, we can export it if the BB is the entry block or
1261 // if it is already exported.
1262 if (isa<Argument>(V)) {
1263 if (FromBB == &FromBB->getParent()->getEntryBlock())
1266 // Otherwise, can only export this if it is already exported.
1267 return FuncInfo.isExportedInst(V);
1270 // Otherwise, constants can always be exported.
1274 /// Return branch probability calculated by BranchProbabilityInfo for IR blocks.
1275 uint32_t SelectionDAGBuilder::getEdgeWeight(MachineBasicBlock *Src,
1276 MachineBasicBlock *Dst) {
1277 BranchProbabilityInfo *BPI = FuncInfo.BPI;
1280 const BasicBlock *SrcBB = Src->getBasicBlock();
1281 const BasicBlock *DstBB = Dst->getBasicBlock();
1282 return BPI->getEdgeWeight(SrcBB, DstBB);
1285 void SelectionDAGBuilder::
1286 addSuccessorWithWeight(MachineBasicBlock *Src, MachineBasicBlock *Dst,
1287 uint32_t Weight /* = 0 */) {
1289 Weight = getEdgeWeight(Src, Dst);
1290 Src->addSuccessor(Dst, Weight);
1294 static bool InBlock(const Value *V, const BasicBlock *BB) {
1295 if (const Instruction *I = dyn_cast<Instruction>(V))
1296 return I->getParent() == BB;
1300 /// EmitBranchForMergedCondition - Helper method for FindMergedConditions.
1301 /// This function emits a branch and is used at the leaves of an OR or an
1302 /// AND operator tree.
1305 SelectionDAGBuilder::EmitBranchForMergedCondition(const Value *Cond,
1306 MachineBasicBlock *TBB,
1307 MachineBasicBlock *FBB,
1308 MachineBasicBlock *CurBB,
1309 MachineBasicBlock *SwitchBB) {
1310 const BasicBlock *BB = CurBB->getBasicBlock();
1312 // If the leaf of the tree is a comparison, merge the condition into
1314 if (const CmpInst *BOp = dyn_cast<CmpInst>(Cond)) {
1315 // The operands of the cmp have to be in this block. We don't know
1316 // how to export them from some other block. If this is the first block
1317 // of the sequence, no exporting is needed.
1318 if (CurBB == SwitchBB ||
1319 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
1320 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) {
1321 ISD::CondCode Condition;
1322 if (const ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
1323 Condition = getICmpCondCode(IC->getPredicate());
1324 } else if (const FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) {
1325 Condition = getFCmpCondCode(FC->getPredicate());
1327 Condition = ISD::SETEQ; // silence warning.
1328 llvm_unreachable("Unknown compare instruction");
1331 CaseBlock CB(Condition, BOp->getOperand(0),
1332 BOp->getOperand(1), NULL, TBB, FBB, CurBB);
1333 SwitchCases.push_back(CB);
1338 // Create a CaseBlock record representing this branch.
1339 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()),
1340 NULL, TBB, FBB, CurBB);
1341 SwitchCases.push_back(CB);
1344 /// FindMergedConditions - If Cond is an expression like
1345 void SelectionDAGBuilder::FindMergedConditions(const Value *Cond,
1346 MachineBasicBlock *TBB,
1347 MachineBasicBlock *FBB,
1348 MachineBasicBlock *CurBB,
1349 MachineBasicBlock *SwitchBB,
1351 // If this node is not part of the or/and tree, emit it as a branch.
1352 const Instruction *BOp = dyn_cast<Instruction>(Cond);
1353 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
1354 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
1355 BOp->getParent() != CurBB->getBasicBlock() ||
1356 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
1357 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
1358 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB);
1362 // Create TmpBB after CurBB.
1363 MachineFunction::iterator BBI = CurBB;
1364 MachineFunction &MF = DAG.getMachineFunction();
1365 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock());
1366 CurBB->getParent()->insert(++BBI, TmpBB);
1368 if (Opc == Instruction::Or) {
1369 // Codegen X | Y as:
1377 // Emit the LHS condition.
1378 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, SwitchBB, Opc);
1380 // Emit the RHS condition into TmpBB.
1381 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
1383 assert(Opc == Instruction::And && "Unknown merge op!");
1384 // Codegen X & Y as:
1391 // This requires creation of TmpBB after CurBB.
1393 // Emit the LHS condition.
1394 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, SwitchBB, Opc);
1396 // Emit the RHS condition into TmpBB.
1397 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
1401 /// If the set of cases should be emitted as a series of branches, return true.
1402 /// If we should emit this as a bunch of and/or'd together conditions, return
1405 SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases){
1406 if (Cases.size() != 2) return true;
1408 // If this is two comparisons of the same values or'd or and'd together, they
1409 // will get folded into a single comparison, so don't emit two blocks.
1410 if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
1411 Cases[0].CmpRHS == Cases[1].CmpRHS) ||
1412 (Cases[0].CmpRHS == Cases[1].CmpLHS &&
1413 Cases[0].CmpLHS == Cases[1].CmpRHS)) {
1417 // Handle: (X != null) | (Y != null) --> (X|Y) != 0
1418 // Handle: (X == null) & (Y == null) --> (X|Y) == 0
1419 if (Cases[0].CmpRHS == Cases[1].CmpRHS &&
1420 Cases[0].CC == Cases[1].CC &&
1421 isa<Constant>(Cases[0].CmpRHS) &&
1422 cast<Constant>(Cases[0].CmpRHS)->isNullValue()) {
1423 if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB)
1425 if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB)
1432 void SelectionDAGBuilder::visitBr(const BranchInst &I) {
1433 MachineBasicBlock *BrMBB = FuncInfo.MBB;
1435 // Update machine-CFG edges.
1436 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
1438 // Figure out which block is immediately after the current one.
1439 MachineBasicBlock *NextBlock = 0;
1440 MachineFunction::iterator BBI = BrMBB;
1441 if (++BBI != FuncInfo.MF->end())
1444 if (I.isUnconditional()) {
1445 // Update machine-CFG edges.
1446 BrMBB->addSuccessor(Succ0MBB);
1448 // If this is not a fall-through branch, emit the branch.
1449 if (Succ0MBB != NextBlock)
1450 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
1451 MVT::Other, getControlRoot(),
1452 DAG.getBasicBlock(Succ0MBB)));
1457 // If this condition is one of the special cases we handle, do special stuff
1459 const Value *CondVal = I.getCondition();
1460 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
1462 // If this is a series of conditions that are or'd or and'd together, emit
1463 // this as a sequence of branches instead of setcc's with and/or operations.
1464 // As long as jumps are not expensive, this should improve performance.
1465 // For example, instead of something like:
1478 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
1479 if (!TLI.isJumpExpensive() &&
1481 (BOp->getOpcode() == Instruction::And ||
1482 BOp->getOpcode() == Instruction::Or)) {
1483 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB,
1485 // If the compares in later blocks need to use values not currently
1486 // exported from this block, export them now. This block should always
1487 // be the first entry.
1488 assert(SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!");
1490 // Allow some cases to be rejected.
1491 if (ShouldEmitAsBranches(SwitchCases)) {
1492 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1493 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1494 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1497 // Emit the branch for this block.
1498 visitSwitchCase(SwitchCases[0], BrMBB);
1499 SwitchCases.erase(SwitchCases.begin());
1503 // Okay, we decided not to do this, remove any inserted MBB's and clear
1505 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
1506 FuncInfo.MF->erase(SwitchCases[i].ThisBB);
1508 SwitchCases.clear();
1512 // Create a CaseBlock record representing this branch.
1513 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()),
1514 NULL, Succ0MBB, Succ1MBB, BrMBB);
1516 // Use visitSwitchCase to actually insert the fast branch sequence for this
1518 visitSwitchCase(CB, BrMBB);
1521 /// visitSwitchCase - Emits the necessary code to represent a single node in
1522 /// the binary search tree resulting from lowering a switch instruction.
1523 void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB,
1524 MachineBasicBlock *SwitchBB) {
1526 SDValue CondLHS = getValue(CB.CmpLHS);
1527 DebugLoc dl = getCurDebugLoc();
1529 // Build the setcc now.
1530 if (CB.CmpMHS == NULL) {
1531 // Fold "(X == true)" to X and "(X == false)" to !X to
1532 // handle common cases produced by branch lowering.
1533 if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) &&
1534 CB.CC == ISD::SETEQ)
1536 else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) &&
1537 CB.CC == ISD::SETEQ) {
1538 SDValue True = DAG.getConstant(1, CondLHS.getValueType());
1539 Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True);
1541 Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1543 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now");
1545 const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue();
1546 const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue();
1548 SDValue CmpOp = getValue(CB.CmpMHS);
1549 EVT VT = CmpOp.getValueType();
1551 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) {
1552 Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, VT),
1555 SDValue SUB = DAG.getNode(ISD::SUB, dl,
1556 VT, CmpOp, DAG.getConstant(Low, VT));
1557 Cond = DAG.getSetCC(dl, MVT::i1, SUB,
1558 DAG.getConstant(High-Low, VT), ISD::SETULE);
1562 // Update successor info
1563 addSuccessorWithWeight(SwitchBB, CB.TrueBB, CB.TrueWeight);
1564 addSuccessorWithWeight(SwitchBB, CB.FalseBB, CB.FalseWeight);
1566 // Set NextBlock to be the MBB immediately after the current one, if any.
1567 // This is used to avoid emitting unnecessary branches to the next block.
1568 MachineBasicBlock *NextBlock = 0;
1569 MachineFunction::iterator BBI = SwitchBB;
1570 if (++BBI != FuncInfo.MF->end())
1573 // If the lhs block is the next block, invert the condition so that we can
1574 // fall through to the lhs instead of the rhs block.
1575 if (CB.TrueBB == NextBlock) {
1576 std::swap(CB.TrueBB, CB.FalseBB);
1577 SDValue True = DAG.getConstant(1, Cond.getValueType());
1578 Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True);
1581 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1582 MVT::Other, getControlRoot(), Cond,
1583 DAG.getBasicBlock(CB.TrueBB));
1585 // Insert the false branch. Do this even if it's a fall through branch,
1586 // this makes it easier to do DAG optimizations which require inverting
1587 // the branch condition.
1588 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1589 DAG.getBasicBlock(CB.FalseBB));
1591 DAG.setRoot(BrCond);
1594 /// visitJumpTable - Emit JumpTable node in the current MBB
1595 void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) {
1596 // Emit the code for the jump table
1597 assert(JT.Reg != -1U && "Should lower JT Header first!");
1598 EVT PTy = TLI.getPointerTy();
1599 SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(),
1601 SDValue Table = DAG.getJumpTable(JT.JTI, PTy);
1602 SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurDebugLoc(),
1603 MVT::Other, Index.getValue(1),
1605 DAG.setRoot(BrJumpTable);
1608 /// visitJumpTableHeader - This function emits necessary code to produce index
1609 /// in the JumpTable from switch case.
1610 void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT,
1611 JumpTableHeader &JTH,
1612 MachineBasicBlock *SwitchBB) {
1613 // Subtract the lowest switch case value from the value being switched on and
1614 // conditional branch to default mbb if the result is greater than the
1615 // difference between smallest and largest cases.
1616 SDValue SwitchOp = getValue(JTH.SValue);
1617 EVT VT = SwitchOp.getValueType();
1618 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1619 DAG.getConstant(JTH.First, VT));
1621 // The SDNode we just created, which holds the value being switched on minus
1622 // the smallest case value, needs to be copied to a virtual register so it
1623 // can be used as an index into the jump table in a subsequent basic block.
1624 // This value may be smaller or larger than the target's pointer type, and
1625 // therefore require extension or truncating.
1626 SwitchOp = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(), TLI.getPointerTy());
1628 unsigned JumpTableReg = FuncInfo.CreateReg(TLI.getPointerTy());
1629 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1630 JumpTableReg, SwitchOp);
1631 JT.Reg = JumpTableReg;
1633 // Emit the range check for the jump table, and branch to the default block
1634 // for the switch statement if the value being switched on exceeds the largest
1635 // case in the switch.
1636 SDValue CMP = DAG.getSetCC(getCurDebugLoc(),
1637 TLI.getSetCCResultType(Sub.getValueType()), Sub,
1638 DAG.getConstant(JTH.Last-JTH.First,VT),
1641 // Set NextBlock to be the MBB immediately after the current one, if any.
1642 // This is used to avoid emitting unnecessary branches to the next block.
1643 MachineBasicBlock *NextBlock = 0;
1644 MachineFunction::iterator BBI = SwitchBB;
1646 if (++BBI != FuncInfo.MF->end())
1649 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1650 MVT::Other, CopyTo, CMP,
1651 DAG.getBasicBlock(JT.Default));
1653 if (JT.MBB != NextBlock)
1654 BrCond = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrCond,
1655 DAG.getBasicBlock(JT.MBB));
1657 DAG.setRoot(BrCond);
1660 /// visitBitTestHeader - This function emits necessary code to produce value
1661 /// suitable for "bit tests"
1662 void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B,
1663 MachineBasicBlock *SwitchBB) {
1664 // Subtract the minimum value
1665 SDValue SwitchOp = getValue(B.SValue);
1666 EVT VT = SwitchOp.getValueType();
1667 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1668 DAG.getConstant(B.First, VT));
1671 SDValue RangeCmp = DAG.getSetCC(getCurDebugLoc(),
1672 TLI.getSetCCResultType(Sub.getValueType()),
1673 Sub, DAG.getConstant(B.Range, VT),
1676 // Determine the type of the test operands.
1677 bool UsePtrType = false;
1678 if (!TLI.isTypeLegal(VT))
1681 for (unsigned i = 0, e = B.Cases.size(); i != e; ++i)
1682 if ((uint64_t)((int64_t)B.Cases[i].Mask >> VT.getSizeInBits()) + 1 >= 2) {
1683 // Switch table case range are encoded into series of masks.
1684 // Just use pointer type, it's guaranteed to fit.
1690 VT = TLI.getPointerTy();
1691 Sub = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(), VT);
1695 B.Reg = FuncInfo.CreateReg(VT);
1696 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1699 // Set NextBlock to be the MBB immediately after the current one, if any.
1700 // This is used to avoid emitting unnecessary branches to the next block.
1701 MachineBasicBlock *NextBlock = 0;
1702 MachineFunction::iterator BBI = SwitchBB;
1703 if (++BBI != FuncInfo.MF->end())
1706 MachineBasicBlock* MBB = B.Cases[0].ThisBB;
1708 addSuccessorWithWeight(SwitchBB, B.Default);
1709 addSuccessorWithWeight(SwitchBB, MBB);
1711 SDValue BrRange = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1712 MVT::Other, CopyTo, RangeCmp,
1713 DAG.getBasicBlock(B.Default));
1715 if (MBB != NextBlock)
1716 BrRange = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, CopyTo,
1717 DAG.getBasicBlock(MBB));
1719 DAG.setRoot(BrRange);
1722 /// visitBitTestCase - this function produces one "bit test"
1723 void SelectionDAGBuilder::visitBitTestCase(BitTestBlock &BB,
1724 MachineBasicBlock* NextMBB,
1727 MachineBasicBlock *SwitchBB) {
1729 SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(),
1732 unsigned PopCount = CountPopulation_64(B.Mask);
1733 if (PopCount == 1) {
1734 // Testing for a single bit; just compare the shift count with what it
1735 // would need to be to shift a 1 bit in that position.
1736 Cmp = DAG.getSetCC(getCurDebugLoc(),
1737 TLI.getSetCCResultType(VT),
1739 DAG.getConstant(CountTrailingZeros_64(B.Mask), VT),
1741 } else if (PopCount == BB.Range) {
1742 // There is only one zero bit in the range, test for it directly.
1743 Cmp = DAG.getSetCC(getCurDebugLoc(),
1744 TLI.getSetCCResultType(VT),
1746 DAG.getConstant(CountTrailingOnes_64(B.Mask), VT),
1749 // Make desired shift
1750 SDValue SwitchVal = DAG.getNode(ISD::SHL, getCurDebugLoc(), VT,
1751 DAG.getConstant(1, VT), ShiftOp);
1753 // Emit bit tests and jumps
1754 SDValue AndOp = DAG.getNode(ISD::AND, getCurDebugLoc(),
1755 VT, SwitchVal, DAG.getConstant(B.Mask, VT));
1756 Cmp = DAG.getSetCC(getCurDebugLoc(),
1757 TLI.getSetCCResultType(VT),
1758 AndOp, DAG.getConstant(0, VT),
1762 addSuccessorWithWeight(SwitchBB, B.TargetBB);
1763 addSuccessorWithWeight(SwitchBB, NextMBB);
1765 SDValue BrAnd = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1766 MVT::Other, getControlRoot(),
1767 Cmp, DAG.getBasicBlock(B.TargetBB));
1769 // Set NextBlock to be the MBB immediately after the current one, if any.
1770 // This is used to avoid emitting unnecessary branches to the next block.
1771 MachineBasicBlock *NextBlock = 0;
1772 MachineFunction::iterator BBI = SwitchBB;
1773 if (++BBI != FuncInfo.MF->end())
1776 if (NextMBB != NextBlock)
1777 BrAnd = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrAnd,
1778 DAG.getBasicBlock(NextMBB));
1783 void SelectionDAGBuilder::visitInvoke(const InvokeInst &I) {
1784 MachineBasicBlock *InvokeMBB = FuncInfo.MBB;
1786 // Retrieve successors.
1787 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
1788 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)];
1790 const Value *Callee(I.getCalledValue());
1791 if (isa<InlineAsm>(Callee))
1794 LowerCallTo(&I, getValue(Callee), false, LandingPad);
1796 // If the value of the invoke is used outside of its defining block, make it
1797 // available as a virtual register.
1798 CopyToExportRegsIfNeeded(&I);
1800 // Update successor info
1801 InvokeMBB->addSuccessor(Return);
1802 InvokeMBB->addSuccessor(LandingPad);
1804 // Drop into normal successor.
1805 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
1806 MVT::Other, getControlRoot(),
1807 DAG.getBasicBlock(Return)));
1810 void SelectionDAGBuilder::visitUnwind(const UnwindInst &I) {
1813 void SelectionDAGBuilder::visitResume(const ResumeInst &RI) {
1814 llvm_unreachable("SelectionDAGBuilder shouldn't visit resume instructions!");
1817 /// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for
1818 /// small case ranges).
1819 bool SelectionDAGBuilder::handleSmallSwitchRange(CaseRec& CR,
1820 CaseRecVector& WorkList,
1822 MachineBasicBlock *Default,
1823 MachineBasicBlock *SwitchBB) {
1824 Case& BackCase = *(CR.Range.second-1);
1826 // Size is the number of Cases represented by this range.
1827 size_t Size = CR.Range.second - CR.Range.first;
1831 // Get the MachineFunction which holds the current MBB. This is used when
1832 // inserting any additional MBBs necessary to represent the switch.
1833 MachineFunction *CurMF = FuncInfo.MF;
1835 // Figure out which block is immediately after the current one.
1836 MachineBasicBlock *NextBlock = 0;
1837 MachineFunction::iterator BBI = CR.CaseBB;
1839 if (++BBI != FuncInfo.MF->end())
1842 // If any two of the cases has the same destination, and if one value
1843 // is the same as the other, but has one bit unset that the other has set,
1844 // use bit manipulation to do two compares at once. For example:
1845 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
1846 // TODO: This could be extended to merge any 2 cases in switches with 3 cases.
1847 // TODO: Handle cases where CR.CaseBB != SwitchBB.
1848 if (Size == 2 && CR.CaseBB == SwitchBB) {
1849 Case &Small = *CR.Range.first;
1850 Case &Big = *(CR.Range.second-1);
1852 if (Small.Low == Small.High && Big.Low == Big.High && Small.BB == Big.BB) {
1853 const APInt& SmallValue = cast<ConstantInt>(Small.Low)->getValue();
1854 const APInt& BigValue = cast<ConstantInt>(Big.Low)->getValue();
1856 // Check that there is only one bit different.
1857 if (BigValue.countPopulation() == SmallValue.countPopulation() + 1 &&
1858 (SmallValue | BigValue) == BigValue) {
1859 // Isolate the common bit.
1860 APInt CommonBit = BigValue & ~SmallValue;
1861 assert((SmallValue | CommonBit) == BigValue &&
1862 CommonBit.countPopulation() == 1 && "Not a common bit?");
1864 SDValue CondLHS = getValue(SV);
1865 EVT VT = CondLHS.getValueType();
1866 DebugLoc DL = getCurDebugLoc();
1868 SDValue Or = DAG.getNode(ISD::OR, DL, VT, CondLHS,
1869 DAG.getConstant(CommonBit, VT));
1870 SDValue Cond = DAG.getSetCC(DL, MVT::i1,
1871 Or, DAG.getConstant(BigValue, VT),
1874 // Update successor info.
1875 addSuccessorWithWeight(SwitchBB, Small.BB);
1876 addSuccessorWithWeight(SwitchBB, Default);
1878 // Insert the true branch.
1879 SDValue BrCond = DAG.getNode(ISD::BRCOND, DL, MVT::Other,
1880 getControlRoot(), Cond,
1881 DAG.getBasicBlock(Small.BB));
1883 // Insert the false branch.
1884 BrCond = DAG.getNode(ISD::BR, DL, MVT::Other, BrCond,
1885 DAG.getBasicBlock(Default));
1887 DAG.setRoot(BrCond);
1893 // Rearrange the case blocks so that the last one falls through if possible.
1894 if (NextBlock && Default != NextBlock && BackCase.BB != NextBlock) {
1895 // The last case block won't fall through into 'NextBlock' if we emit the
1896 // branches in this order. See if rearranging a case value would help.
1897 for (CaseItr I = CR.Range.first, E = CR.Range.second-1; I != E; ++I) {
1898 if (I->BB == NextBlock) {
1899 std::swap(*I, BackCase);
1905 // Create a CaseBlock record representing a conditional branch to
1906 // the Case's target mbb if the value being switched on SV is equal
1908 MachineBasicBlock *CurBlock = CR.CaseBB;
1909 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
1910 MachineBasicBlock *FallThrough;
1912 FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock());
1913 CurMF->insert(BBI, FallThrough);
1915 // Put SV in a virtual register to make it available from the new blocks.
1916 ExportFromCurrentBlock(SV);
1918 // If the last case doesn't match, go to the default block.
1919 FallThrough = Default;
1922 const Value *RHS, *LHS, *MHS;
1924 if (I->High == I->Low) {
1925 // This is just small small case range :) containing exactly 1 case
1927 LHS = SV; RHS = I->High; MHS = NULL;
1930 LHS = I->Low; MHS = SV; RHS = I->High;
1933 uint32_t ExtraWeight = I->ExtraWeight;
1934 CaseBlock CB(CC, LHS, RHS, MHS, /* truebb */ I->BB, /* falsebb */ FallThrough,
1936 /* trueweight */ ExtraWeight / 2, /* falseweight */ ExtraWeight / 2);
1938 // If emitting the first comparison, just call visitSwitchCase to emit the
1939 // code into the current block. Otherwise, push the CaseBlock onto the
1940 // vector to be later processed by SDISel, and insert the node's MBB
1941 // before the next MBB.
1942 if (CurBlock == SwitchBB)
1943 visitSwitchCase(CB, SwitchBB);
1945 SwitchCases.push_back(CB);
1947 CurBlock = FallThrough;
1953 static inline bool areJTsAllowed(const TargetLowering &TLI) {
1954 return !DisableJumpTables &&
1955 (TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
1956 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other));
1959 static APInt ComputeRange(const APInt &First, const APInt &Last) {
1960 uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1;
1961 APInt LastExt = Last.sext(BitWidth), FirstExt = First.sext(BitWidth);
1962 return (LastExt - FirstExt + 1ULL);
1965 /// handleJTSwitchCase - Emit jumptable for current switch case range
1966 bool SelectionDAGBuilder::handleJTSwitchCase(CaseRec& CR,
1967 CaseRecVector& WorkList,
1969 MachineBasicBlock* Default,
1970 MachineBasicBlock *SwitchBB) {
1971 Case& FrontCase = *CR.Range.first;
1972 Case& BackCase = *(CR.Range.second-1);
1974 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
1975 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
1977 APInt TSize(First.getBitWidth(), 0);
1978 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1982 if (!areJTsAllowed(TLI) || TSize.ult(4))
1985 APInt Range = ComputeRange(First, Last);
1986 double Density = TSize.roundToDouble() / Range.roundToDouble();
1990 DEBUG(dbgs() << "Lowering jump table\n"
1991 << "First entry: " << First << ". Last entry: " << Last << '\n'
1992 << "Range: " << Range
1993 << ". Size: " << TSize << ". Density: " << Density << "\n\n");
1995 // Get the MachineFunction which holds the current MBB. This is used when
1996 // inserting any additional MBBs necessary to represent the switch.
1997 MachineFunction *CurMF = FuncInfo.MF;
1999 // Figure out which block is immediately after the current one.
2000 MachineFunction::iterator BBI = CR.CaseBB;
2003 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2005 // Create a new basic block to hold the code for loading the address
2006 // of the jump table, and jumping to it. Update successor information;
2007 // we will either branch to the default case for the switch, or the jump
2009 MachineBasicBlock *JumpTableBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2010 CurMF->insert(BBI, JumpTableBB);
2012 addSuccessorWithWeight(CR.CaseBB, Default);
2013 addSuccessorWithWeight(CR.CaseBB, JumpTableBB);
2015 // Build a vector of destination BBs, corresponding to each target
2016 // of the jump table. If the value of the jump table slot corresponds to
2017 // a case statement, push the case's BB onto the vector, otherwise, push
2019 std::vector<MachineBasicBlock*> DestBBs;
2021 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) {
2022 const APInt &Low = cast<ConstantInt>(I->Low)->getValue();
2023 const APInt &High = cast<ConstantInt>(I->High)->getValue();
2025 if (Low.sle(TEI) && TEI.sle(High)) {
2026 DestBBs.push_back(I->BB);
2030 DestBBs.push_back(Default);
2034 // Update successor info. Add one edge to each unique successor.
2035 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs());
2036 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(),
2037 E = DestBBs.end(); I != E; ++I) {
2038 if (!SuccsHandled[(*I)->getNumber()]) {
2039 SuccsHandled[(*I)->getNumber()] = true;
2040 addSuccessorWithWeight(JumpTableBB, *I);
2044 // Create a jump table index for this jump table.
2045 unsigned JTEncoding = TLI.getJumpTableEncoding();
2046 unsigned JTI = CurMF->getOrCreateJumpTableInfo(JTEncoding)
2047 ->createJumpTableIndex(DestBBs);
2049 // Set the jump table information so that we can codegen it as a second
2050 // MachineBasicBlock
2051 JumpTable JT(-1U, JTI, JumpTableBB, Default);
2052 JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == SwitchBB));
2053 if (CR.CaseBB == SwitchBB)
2054 visitJumpTableHeader(JT, JTH, SwitchBB);
2056 JTCases.push_back(JumpTableBlock(JTH, JT));
2061 /// handleBTSplitSwitchCase - emit comparison and split binary search tree into
2063 bool SelectionDAGBuilder::handleBTSplitSwitchCase(CaseRec& CR,
2064 CaseRecVector& WorkList,
2066 MachineBasicBlock *Default,
2067 MachineBasicBlock *SwitchBB) {
2068 // Get the MachineFunction which holds the current MBB. This is used when
2069 // inserting any additional MBBs necessary to represent the switch.
2070 MachineFunction *CurMF = FuncInfo.MF;
2072 // Figure out which block is immediately after the current one.
2073 MachineFunction::iterator BBI = CR.CaseBB;
2076 Case& FrontCase = *CR.Range.first;
2077 Case& BackCase = *(CR.Range.second-1);
2078 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2080 // Size is the number of Cases represented by this range.
2081 unsigned Size = CR.Range.second - CR.Range.first;
2083 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
2084 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
2086 CaseItr Pivot = CR.Range.first + Size/2;
2088 // Select optimal pivot, maximizing sum density of LHS and RHS. This will
2089 // (heuristically) allow us to emit JumpTable's later.
2090 APInt TSize(First.getBitWidth(), 0);
2091 for (CaseItr I = CR.Range.first, E = CR.Range.second;
2095 APInt LSize = FrontCase.size();
2096 APInt RSize = TSize-LSize;
2097 DEBUG(dbgs() << "Selecting best pivot: \n"
2098 << "First: " << First << ", Last: " << Last <<'\n'
2099 << "LSize: " << LSize << ", RSize: " << RSize << '\n');
2100 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second;
2102 const APInt &LEnd = cast<ConstantInt>(I->High)->getValue();
2103 const APInt &RBegin = cast<ConstantInt>(J->Low)->getValue();
2104 APInt Range = ComputeRange(LEnd, RBegin);
2105 assert((Range - 2ULL).isNonNegative() &&
2106 "Invalid case distance");
2107 // Use volatile double here to avoid excess precision issues on some hosts,
2108 // e.g. that use 80-bit X87 registers.
2109 volatile double LDensity =
2110 (double)LSize.roundToDouble() /
2111 (LEnd - First + 1ULL).roundToDouble();
2112 volatile double RDensity =
2113 (double)RSize.roundToDouble() /
2114 (Last - RBegin + 1ULL).roundToDouble();
2115 double Metric = Range.logBase2()*(LDensity+RDensity);
2116 // Should always split in some non-trivial place
2117 DEBUG(dbgs() <<"=>Step\n"
2118 << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n'
2119 << "LDensity: " << LDensity
2120 << ", RDensity: " << RDensity << '\n'
2121 << "Metric: " << Metric << '\n');
2122 if (FMetric < Metric) {
2125 DEBUG(dbgs() << "Current metric set to: " << FMetric << '\n');
2131 if (areJTsAllowed(TLI)) {
2132 // If our case is dense we *really* should handle it earlier!
2133 assert((FMetric > 0) && "Should handle dense range earlier!");
2135 Pivot = CR.Range.first + Size/2;
2138 CaseRange LHSR(CR.Range.first, Pivot);
2139 CaseRange RHSR(Pivot, CR.Range.second);
2140 Constant *C = Pivot->Low;
2141 MachineBasicBlock *FalseBB = 0, *TrueBB = 0;
2143 // We know that we branch to the LHS if the Value being switched on is
2144 // less than the Pivot value, C. We use this to optimize our binary
2145 // tree a bit, by recognizing that if SV is greater than or equal to the
2146 // LHS's Case Value, and that Case Value is exactly one less than the
2147 // Pivot's Value, then we can branch directly to the LHS's Target,
2148 // rather than creating a leaf node for it.
2149 if ((LHSR.second - LHSR.first) == 1 &&
2150 LHSR.first->High == CR.GE &&
2151 cast<ConstantInt>(C)->getValue() ==
2152 (cast<ConstantInt>(CR.GE)->getValue() + 1LL)) {
2153 TrueBB = LHSR.first->BB;
2155 TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2156 CurMF->insert(BBI, TrueBB);
2157 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR));
2159 // Put SV in a virtual register to make it available from the new blocks.
2160 ExportFromCurrentBlock(SV);
2163 // Similar to the optimization above, if the Value being switched on is
2164 // known to be less than the Constant CR.LT, and the current Case Value
2165 // is CR.LT - 1, then we can branch directly to the target block for
2166 // the current Case Value, rather than emitting a RHS leaf node for it.
2167 if ((RHSR.second - RHSR.first) == 1 && CR.LT &&
2168 cast<ConstantInt>(RHSR.first->Low)->getValue() ==
2169 (cast<ConstantInt>(CR.LT)->getValue() - 1LL)) {
2170 FalseBB = RHSR.first->BB;
2172 FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2173 CurMF->insert(BBI, FalseBB);
2174 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR));
2176 // Put SV in a virtual register to make it available from the new blocks.
2177 ExportFromCurrentBlock(SV);
2180 // Create a CaseBlock record representing a conditional branch to
2181 // the LHS node if the value being switched on SV is less than C.
2182 // Otherwise, branch to LHS.
2183 CaseBlock CB(ISD::SETLT, SV, C, NULL, TrueBB, FalseBB, CR.CaseBB);
2185 if (CR.CaseBB == SwitchBB)
2186 visitSwitchCase(CB, SwitchBB);
2188 SwitchCases.push_back(CB);
2193 /// handleBitTestsSwitchCase - if current case range has few destination and
2194 /// range span less, than machine word bitwidth, encode case range into series
2195 /// of masks and emit bit tests with these masks.
2196 bool SelectionDAGBuilder::handleBitTestsSwitchCase(CaseRec& CR,
2197 CaseRecVector& WorkList,
2199 MachineBasicBlock* Default,
2200 MachineBasicBlock *SwitchBB){
2201 EVT PTy = TLI.getPointerTy();
2202 unsigned IntPtrBits = PTy.getSizeInBits();
2204 Case& FrontCase = *CR.Range.first;
2205 Case& BackCase = *(CR.Range.second-1);
2207 // Get the MachineFunction which holds the current MBB. This is used when
2208 // inserting any additional MBBs necessary to represent the switch.
2209 MachineFunction *CurMF = FuncInfo.MF;
2211 // If target does not have legal shift left, do not emit bit tests at all.
2212 if (!TLI.isOperationLegal(ISD::SHL, TLI.getPointerTy()))
2216 for (CaseItr I = CR.Range.first, E = CR.Range.second;
2218 // Single case counts one, case range - two.
2219 numCmps += (I->Low == I->High ? 1 : 2);
2222 // Count unique destinations
2223 SmallSet<MachineBasicBlock*, 4> Dests;
2224 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2225 Dests.insert(I->BB);
2226 if (Dests.size() > 3)
2227 // Don't bother the code below, if there are too much unique destinations
2230 DEBUG(dbgs() << "Total number of unique destinations: "
2231 << Dests.size() << '\n'
2232 << "Total number of comparisons: " << numCmps << '\n');
2234 // Compute span of values.
2235 const APInt& minValue = cast<ConstantInt>(FrontCase.Low)->getValue();
2236 const APInt& maxValue = cast<ConstantInt>(BackCase.High)->getValue();
2237 APInt cmpRange = maxValue - minValue;
2239 DEBUG(dbgs() << "Compare range: " << cmpRange << '\n'
2240 << "Low bound: " << minValue << '\n'
2241 << "High bound: " << maxValue << '\n');
2243 if (cmpRange.uge(IntPtrBits) ||
2244 (!(Dests.size() == 1 && numCmps >= 3) &&
2245 !(Dests.size() == 2 && numCmps >= 5) &&
2246 !(Dests.size() >= 3 && numCmps >= 6)))
2249 DEBUG(dbgs() << "Emitting bit tests\n");
2250 APInt lowBound = APInt::getNullValue(cmpRange.getBitWidth());
2252 // Optimize the case where all the case values fit in a
2253 // word without having to subtract minValue. In this case,
2254 // we can optimize away the subtraction.
2255 if (minValue.isNonNegative() && maxValue.slt(IntPtrBits)) {
2256 cmpRange = maxValue;
2258 lowBound = minValue;
2261 CaseBitsVector CasesBits;
2262 unsigned i, count = 0;
2264 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2265 MachineBasicBlock* Dest = I->BB;
2266 for (i = 0; i < count; ++i)
2267 if (Dest == CasesBits[i].BB)
2271 assert((count < 3) && "Too much destinations to test!");
2272 CasesBits.push_back(CaseBits(0, Dest, 0));
2276 const APInt& lowValue = cast<ConstantInt>(I->Low)->getValue();
2277 const APInt& highValue = cast<ConstantInt>(I->High)->getValue();
2279 uint64_t lo = (lowValue - lowBound).getZExtValue();
2280 uint64_t hi = (highValue - lowBound).getZExtValue();
2282 for (uint64_t j = lo; j <= hi; j++) {
2283 CasesBits[i].Mask |= 1ULL << j;
2284 CasesBits[i].Bits++;
2288 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp());
2292 // Figure out which block is immediately after the current one.
2293 MachineFunction::iterator BBI = CR.CaseBB;
2296 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2298 DEBUG(dbgs() << "Cases:\n");
2299 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) {
2300 DEBUG(dbgs() << "Mask: " << CasesBits[i].Mask
2301 << ", Bits: " << CasesBits[i].Bits
2302 << ", BB: " << CasesBits[i].BB << '\n');
2304 MachineBasicBlock *CaseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2305 CurMF->insert(BBI, CaseBB);
2306 BTC.push_back(BitTestCase(CasesBits[i].Mask,
2310 // Put SV in a virtual register to make it available from the new blocks.
2311 ExportFromCurrentBlock(SV);
2314 BitTestBlock BTB(lowBound, cmpRange, SV,
2315 -1U, MVT::Other, (CR.CaseBB == SwitchBB),
2316 CR.CaseBB, Default, BTC);
2318 if (CR.CaseBB == SwitchBB)
2319 visitBitTestHeader(BTB, SwitchBB);
2321 BitTestCases.push_back(BTB);
2326 /// Clusterify - Transform simple list of Cases into list of CaseRange's
2327 size_t SelectionDAGBuilder::Clusterify(CaseVector& Cases,
2328 const SwitchInst& SI) {
2331 BranchProbabilityInfo *BPI = FuncInfo.BPI;
2332 // Start with "simple" cases
2333 for (size_t i = 1; i < SI.getNumSuccessors(); ++i) {
2334 BasicBlock *SuccBB = SI.getSuccessor(i);
2335 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SuccBB];
2337 uint32_t ExtraWeight = BPI ? BPI->getEdgeWeight(SI.getParent(), SuccBB) : 0;
2339 Cases.push_back(Case(SI.getSuccessorValue(i),
2340 SI.getSuccessorValue(i),
2341 SMBB, ExtraWeight));
2343 std::sort(Cases.begin(), Cases.end(), CaseCmp());
2345 // Merge case into clusters
2346 if (Cases.size() >= 2)
2347 // Must recompute end() each iteration because it may be
2348 // invalidated by erase if we hold on to it
2349 for (CaseItr I = Cases.begin(), J = llvm::next(Cases.begin());
2350 J != Cases.end(); ) {
2351 const APInt& nextValue = cast<ConstantInt>(J->Low)->getValue();
2352 const APInt& currentValue = cast<ConstantInt>(I->High)->getValue();
2353 MachineBasicBlock* nextBB = J->BB;
2354 MachineBasicBlock* currentBB = I->BB;
2356 // If the two neighboring cases go to the same destination, merge them
2357 // into a single case.
2358 if ((nextValue - currentValue == 1) && (currentBB == nextBB)) {
2362 if (BranchProbabilityInfo *BPI = FuncInfo.BPI) {
2363 uint32_t CurWeight = currentBB->getBasicBlock() ?
2364 BPI->getEdgeWeight(SI.getParent(), currentBB->getBasicBlock()) : 16;
2365 uint32_t NextWeight = nextBB->getBasicBlock() ?
2366 BPI->getEdgeWeight(SI.getParent(), nextBB->getBasicBlock()) : 16;
2368 BPI->setEdgeWeight(SI.getParent(), currentBB->getBasicBlock(),
2369 CurWeight + NextWeight);
2376 for (CaseItr I=Cases.begin(), E=Cases.end(); I!=E; ++I, ++numCmps) {
2377 if (I->Low != I->High)
2378 // A range counts double, since it requires two compares.
2385 void SelectionDAGBuilder::UpdateSplitBlock(MachineBasicBlock *First,
2386 MachineBasicBlock *Last) {
2388 for (unsigned i = 0, e = JTCases.size(); i != e; ++i)
2389 if (JTCases[i].first.HeaderBB == First)
2390 JTCases[i].first.HeaderBB = Last;
2392 // Update BitTestCases.
2393 for (unsigned i = 0, e = BitTestCases.size(); i != e; ++i)
2394 if (BitTestCases[i].Parent == First)
2395 BitTestCases[i].Parent = Last;
2398 void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) {
2399 MachineBasicBlock *SwitchMBB = FuncInfo.MBB;
2401 // Figure out which block is immediately after the current one.
2402 MachineBasicBlock *NextBlock = 0;
2403 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()];
2405 // If there is only the default destination, branch to it if it is not the
2406 // next basic block. Otherwise, just fall through.
2407 if (SI.getNumOperands() == 2) {
2408 // Update machine-CFG edges.
2410 // If this is not a fall-through branch, emit the branch.
2411 SwitchMBB->addSuccessor(Default);
2412 if (Default != NextBlock)
2413 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
2414 MVT::Other, getControlRoot(),
2415 DAG.getBasicBlock(Default)));
2420 // If there are any non-default case statements, create a vector of Cases
2421 // representing each one, and sort the vector so that we can efficiently
2422 // create a binary search tree from them.
2424 size_t numCmps = Clusterify(Cases, SI);
2425 DEBUG(dbgs() << "Clusterify finished. Total clusters: " << Cases.size()
2426 << ". Total compares: " << numCmps << '\n');
2429 // Get the Value to be switched on and default basic blocks, which will be
2430 // inserted into CaseBlock records, representing basic blocks in the binary
2432 const Value *SV = SI.getOperand(0);
2434 // Push the initial CaseRec onto the worklist
2435 CaseRecVector WorkList;
2436 WorkList.push_back(CaseRec(SwitchMBB,0,0,
2437 CaseRange(Cases.begin(),Cases.end())));
2439 while (!WorkList.empty()) {
2440 // Grab a record representing a case range to process off the worklist
2441 CaseRec CR = WorkList.back();
2442 WorkList.pop_back();
2444 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2447 // If the range has few cases (two or less) emit a series of specific
2449 if (handleSmallSwitchRange(CR, WorkList, SV, Default, SwitchMBB))
2452 // If the switch has more than 5 blocks, and at least 40% dense, and the
2453 // target supports indirect branches, then emit a jump table rather than
2454 // lowering the switch to a binary tree of conditional branches.
2455 if (handleJTSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2458 // Emit binary tree. We need to pick a pivot, and push left and right ranges
2459 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call.
2460 handleBTSplitSwitchCase(CR, WorkList, SV, Default, SwitchMBB);
2464 void SelectionDAGBuilder::visitIndirectBr(const IndirectBrInst &I) {
2465 MachineBasicBlock *IndirectBrMBB = FuncInfo.MBB;
2467 // Update machine-CFG edges with unique successors.
2468 SmallVector<BasicBlock*, 32> succs;
2469 succs.reserve(I.getNumSuccessors());
2470 for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i)
2471 succs.push_back(I.getSuccessor(i));
2472 array_pod_sort(succs.begin(), succs.end());
2473 succs.erase(std::unique(succs.begin(), succs.end()), succs.end());
2474 for (unsigned i = 0, e = succs.size(); i != e; ++i) {
2475 MachineBasicBlock *Succ = FuncInfo.MBBMap[succs[i]];
2476 addSuccessorWithWeight(IndirectBrMBB, Succ);
2479 DAG.setRoot(DAG.getNode(ISD::BRIND, getCurDebugLoc(),
2480 MVT::Other, getControlRoot(),
2481 getValue(I.getAddress())));
2484 void SelectionDAGBuilder::visitFSub(const User &I) {
2485 // -0.0 - X --> fneg
2486 Type *Ty = I.getType();
2487 if (isa<Constant>(I.getOperand(0)) &&
2488 I.getOperand(0) == ConstantFP::getZeroValueForNegation(Ty)) {
2489 SDValue Op2 = getValue(I.getOperand(1));
2490 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(),
2491 Op2.getValueType(), Op2));
2495 visitBinary(I, ISD::FSUB);
2498 void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) {
2499 SDValue Op1 = getValue(I.getOperand(0));
2500 SDValue Op2 = getValue(I.getOperand(1));
2501 setValue(&I, DAG.getNode(OpCode, getCurDebugLoc(),
2502 Op1.getValueType(), Op1, Op2));
2505 void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) {
2506 SDValue Op1 = getValue(I.getOperand(0));
2507 SDValue Op2 = getValue(I.getOperand(1));
2509 MVT ShiftTy = TLI.getShiftAmountTy(Op2.getValueType());
2511 // Coerce the shift amount to the right type if we can.
2512 if (!I.getType()->isVectorTy() && Op2.getValueType() != ShiftTy) {
2513 unsigned ShiftSize = ShiftTy.getSizeInBits();
2514 unsigned Op2Size = Op2.getValueType().getSizeInBits();
2515 DebugLoc DL = getCurDebugLoc();
2517 // If the operand is smaller than the shift count type, promote it.
2518 if (ShiftSize > Op2Size)
2519 Op2 = DAG.getNode(ISD::ZERO_EXTEND, DL, ShiftTy, Op2);
2521 // If the operand is larger than the shift count type but the shift
2522 // count type has enough bits to represent any shift value, truncate
2523 // it now. This is a common case and it exposes the truncate to
2524 // optimization early.
2525 else if (ShiftSize >= Log2_32_Ceil(Op2.getValueType().getSizeInBits()))
2526 Op2 = DAG.getNode(ISD::TRUNCATE, DL, ShiftTy, Op2);
2527 // Otherwise we'll need to temporarily settle for some other convenient
2528 // type. Type legalization will make adjustments once the shiftee is split.
2530 Op2 = DAG.getZExtOrTrunc(Op2, DL, MVT::i32);
2533 setValue(&I, DAG.getNode(Opcode, getCurDebugLoc(),
2534 Op1.getValueType(), Op1, Op2));
2537 void SelectionDAGBuilder::visitSDiv(const User &I) {
2538 SDValue Op1 = getValue(I.getOperand(0));
2539 SDValue Op2 = getValue(I.getOperand(1));
2541 // Turn exact SDivs into multiplications.
2542 // FIXME: This should be in DAGCombiner, but it doesn't have access to the
2544 if (isa<BinaryOperator>(&I) && cast<BinaryOperator>(&I)->isExact() &&
2545 !isa<ConstantSDNode>(Op1) &&
2546 isa<ConstantSDNode>(Op2) && !cast<ConstantSDNode>(Op2)->isNullValue())
2547 setValue(&I, TLI.BuildExactSDIV(Op1, Op2, getCurDebugLoc(), DAG));
2549 setValue(&I, DAG.getNode(ISD::SDIV, getCurDebugLoc(), Op1.getValueType(),
2553 void SelectionDAGBuilder::visitICmp(const User &I) {
2554 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2555 if (const ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2556 predicate = IC->getPredicate();
2557 else if (const ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2558 predicate = ICmpInst::Predicate(IC->getPredicate());
2559 SDValue Op1 = getValue(I.getOperand(0));
2560 SDValue Op2 = getValue(I.getOperand(1));
2561 ISD::CondCode Opcode = getICmpCondCode(predicate);
2563 EVT DestVT = TLI.getValueType(I.getType());
2564 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Opcode));
2567 void SelectionDAGBuilder::visitFCmp(const User &I) {
2568 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2569 if (const FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2570 predicate = FC->getPredicate();
2571 else if (const ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2572 predicate = FCmpInst::Predicate(FC->getPredicate());
2573 SDValue Op1 = getValue(I.getOperand(0));
2574 SDValue Op2 = getValue(I.getOperand(1));
2575 ISD::CondCode Condition = getFCmpCondCode(predicate);
2576 EVT DestVT = TLI.getValueType(I.getType());
2577 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Condition));
2580 void SelectionDAGBuilder::visitSelect(const User &I) {
2581 SmallVector<EVT, 4> ValueVTs;
2582 ComputeValueVTs(TLI, I.getType(), ValueVTs);
2583 unsigned NumValues = ValueVTs.size();
2584 if (NumValues == 0) return;
2586 SmallVector<SDValue, 4> Values(NumValues);
2587 SDValue Cond = getValue(I.getOperand(0));
2588 SDValue TrueVal = getValue(I.getOperand(1));
2589 SDValue FalseVal = getValue(I.getOperand(2));
2591 for (unsigned i = 0; i != NumValues; ++i)
2592 Values[i] = DAG.getNode(ISD::SELECT, getCurDebugLoc(),
2593 TrueVal.getNode()->getValueType(TrueVal.getResNo()+i),
2595 SDValue(TrueVal.getNode(),
2596 TrueVal.getResNo() + i),
2597 SDValue(FalseVal.getNode(),
2598 FalseVal.getResNo() + i));
2600 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2601 DAG.getVTList(&ValueVTs[0], NumValues),
2602 &Values[0], NumValues));
2605 void SelectionDAGBuilder::visitTrunc(const User &I) {
2606 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2607 SDValue N = getValue(I.getOperand(0));
2608 EVT DestVT = TLI.getValueType(I.getType());
2609 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), DestVT, N));
2612 void SelectionDAGBuilder::visitZExt(const User &I) {
2613 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2614 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2615 SDValue N = getValue(I.getOperand(0));
2616 EVT DestVT = TLI.getValueType(I.getType());
2617 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(), DestVT, N));
2620 void SelectionDAGBuilder::visitSExt(const User &I) {
2621 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2622 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2623 SDValue N = getValue(I.getOperand(0));
2624 EVT DestVT = TLI.getValueType(I.getType());
2625 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurDebugLoc(), DestVT, N));
2628 void SelectionDAGBuilder::visitFPTrunc(const User &I) {
2629 // FPTrunc is never a no-op cast, no need to check
2630 SDValue N = getValue(I.getOperand(0));
2631 EVT DestVT = TLI.getValueType(I.getType());
2632 setValue(&I, DAG.getNode(ISD::FP_ROUND, getCurDebugLoc(),
2633 DestVT, N, DAG.getIntPtrConstant(0)));
2636 void SelectionDAGBuilder::visitFPExt(const User &I){
2637 // FPTrunc is never a no-op cast, no need to check
2638 SDValue N = getValue(I.getOperand(0));
2639 EVT DestVT = TLI.getValueType(I.getType());
2640 setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurDebugLoc(), DestVT, N));
2643 void SelectionDAGBuilder::visitFPToUI(const User &I) {
2644 // FPToUI is never a no-op cast, no need to check
2645 SDValue N = getValue(I.getOperand(0));
2646 EVT DestVT = TLI.getValueType(I.getType());
2647 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurDebugLoc(), DestVT, N));
2650 void SelectionDAGBuilder::visitFPToSI(const User &I) {
2651 // FPToSI is never a no-op cast, no need to check
2652 SDValue N = getValue(I.getOperand(0));
2653 EVT DestVT = TLI.getValueType(I.getType());
2654 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurDebugLoc(), DestVT, N));
2657 void SelectionDAGBuilder::visitUIToFP(const User &I) {
2658 // UIToFP is never a no-op cast, no need to check
2659 SDValue N = getValue(I.getOperand(0));
2660 EVT DestVT = TLI.getValueType(I.getType());
2661 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurDebugLoc(), DestVT, N));
2664 void SelectionDAGBuilder::visitSIToFP(const User &I){
2665 // SIToFP is never a no-op cast, no need to check
2666 SDValue N = getValue(I.getOperand(0));
2667 EVT DestVT = TLI.getValueType(I.getType());
2668 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurDebugLoc(), DestVT, N));
2671 void SelectionDAGBuilder::visitPtrToInt(const User &I) {
2672 // What to do depends on the size of the integer and the size of the pointer.
2673 // We can either truncate, zero extend, or no-op, accordingly.
2674 SDValue N = getValue(I.getOperand(0));
2675 EVT DestVT = TLI.getValueType(I.getType());
2676 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT));
2679 void SelectionDAGBuilder::visitIntToPtr(const User &I) {
2680 // What to do depends on the size of the integer and the size of the pointer.
2681 // We can either truncate, zero extend, or no-op, accordingly.
2682 SDValue N = getValue(I.getOperand(0));
2683 EVT DestVT = TLI.getValueType(I.getType());
2684 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT));
2687 void SelectionDAGBuilder::visitBitCast(const User &I) {
2688 SDValue N = getValue(I.getOperand(0));
2689 EVT DestVT = TLI.getValueType(I.getType());
2691 // BitCast assures us that source and destination are the same size so this is
2692 // either a BITCAST or a no-op.
2693 if (DestVT != N.getValueType())
2694 setValue(&I, DAG.getNode(ISD::BITCAST, getCurDebugLoc(),
2695 DestVT, N)); // convert types.
2697 setValue(&I, N); // noop cast.
2700 void SelectionDAGBuilder::visitInsertElement(const User &I) {
2701 SDValue InVec = getValue(I.getOperand(0));
2702 SDValue InVal = getValue(I.getOperand(1));
2703 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2705 getValue(I.getOperand(2)));
2706 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurDebugLoc(),
2707 TLI.getValueType(I.getType()),
2708 InVec, InVal, InIdx));
2711 void SelectionDAGBuilder::visitExtractElement(const User &I) {
2712 SDValue InVec = getValue(I.getOperand(0));
2713 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2715 getValue(I.getOperand(1)));
2716 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2717 TLI.getValueType(I.getType()), InVec, InIdx));
2720 // Utility for visitShuffleVector - Returns true if the mask is mask starting
2721 // from SIndx and increasing to the element length (undefs are allowed).
2722 static bool SequentialMask(SmallVectorImpl<int> &Mask, unsigned SIndx) {
2723 unsigned MaskNumElts = Mask.size();
2724 for (unsigned i = 0; i != MaskNumElts; ++i)
2725 if ((Mask[i] >= 0) && (Mask[i] != (int)(i + SIndx)))
2730 void SelectionDAGBuilder::visitShuffleVector(const User &I) {
2731 SmallVector<int, 8> Mask;
2732 SDValue Src1 = getValue(I.getOperand(0));
2733 SDValue Src2 = getValue(I.getOperand(1));
2735 // Convert the ConstantVector mask operand into an array of ints, with -1
2736 // representing undef values.
2737 SmallVector<Constant*, 8> MaskElts;
2738 cast<Constant>(I.getOperand(2))->getVectorElements(MaskElts);
2739 unsigned MaskNumElts = MaskElts.size();
2740 for (unsigned i = 0; i != MaskNumElts; ++i) {
2741 if (isa<UndefValue>(MaskElts[i]))
2744 Mask.push_back(cast<ConstantInt>(MaskElts[i])->getSExtValue());
2747 EVT VT = TLI.getValueType(I.getType());
2748 EVT SrcVT = Src1.getValueType();
2749 unsigned SrcNumElts = SrcVT.getVectorNumElements();
2751 if (SrcNumElts == MaskNumElts) {
2752 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2757 // Normalize the shuffle vector since mask and vector length don't match.
2758 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) {
2759 // Mask is longer than the source vectors and is a multiple of the source
2760 // vectors. We can use concatenate vector to make the mask and vectors
2762 if (SrcNumElts*2 == MaskNumElts && SequentialMask(Mask, 0)) {
2763 // The shuffle is concatenating two vectors together.
2764 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurDebugLoc(),
2769 // Pad both vectors with undefs to make them the same length as the mask.
2770 unsigned NumConcat = MaskNumElts / SrcNumElts;
2771 bool Src1U = Src1.getOpcode() == ISD::UNDEF;
2772 bool Src2U = Src2.getOpcode() == ISD::UNDEF;
2773 SDValue UndefVal = DAG.getUNDEF(SrcVT);
2775 SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal);
2776 SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal);
2780 Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2781 getCurDebugLoc(), VT,
2782 &MOps1[0], NumConcat);
2783 Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2784 getCurDebugLoc(), VT,
2785 &MOps2[0], NumConcat);
2787 // Readjust mask for new input vector length.
2788 SmallVector<int, 8> MappedOps;
2789 for (unsigned i = 0; i != MaskNumElts; ++i) {
2791 if (Idx < (int)SrcNumElts)
2792 MappedOps.push_back(Idx);
2794 MappedOps.push_back(Idx + MaskNumElts - SrcNumElts);
2797 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2802 if (SrcNumElts > MaskNumElts) {
2803 // Analyze the access pattern of the vector to see if we can extract
2804 // two subvectors and do the shuffle. The analysis is done by calculating
2805 // the range of elements the mask access on both vectors.
2806 int MinRange[2] = { static_cast<int>(SrcNumElts+1),
2807 static_cast<int>(SrcNumElts+1)};
2808 int MaxRange[2] = {-1, -1};
2810 for (unsigned i = 0; i != MaskNumElts; ++i) {
2816 if (Idx >= (int)SrcNumElts) {
2820 if (Idx > MaxRange[Input])
2821 MaxRange[Input] = Idx;
2822 if (Idx < MinRange[Input])
2823 MinRange[Input] = Idx;
2826 // Check if the access is smaller than the vector size and can we find
2827 // a reasonable extract index.
2828 int RangeUse[2] = { 2, 2 }; // 0 = Unused, 1 = Extract, 2 = Can not
2830 int StartIdx[2]; // StartIdx to extract from
2831 for (int Input=0; Input < 2; ++Input) {
2832 if (MinRange[Input] == (int)(SrcNumElts+1) && MaxRange[Input] == -1) {
2833 RangeUse[Input] = 0; // Unused
2834 StartIdx[Input] = 0;
2835 } else if (MaxRange[Input] - MinRange[Input] < (int)MaskNumElts) {
2836 // Fits within range but we should see if we can find a good
2837 // start index that is a multiple of the mask length.
2838 if (MaxRange[Input] < (int)MaskNumElts) {
2839 RangeUse[Input] = 1; // Extract from beginning of the vector
2840 StartIdx[Input] = 0;
2842 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts;
2843 if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts &&
2844 StartIdx[Input] + MaskNumElts <= SrcNumElts)
2845 RangeUse[Input] = 1; // Extract from a multiple of the mask length.
2850 if (RangeUse[0] == 0 && RangeUse[1] == 0) {
2851 setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used.
2854 else if (RangeUse[0] < 2 && RangeUse[1] < 2) {
2855 // Extract appropriate subvector and generate a vector shuffle
2856 for (int Input=0; Input < 2; ++Input) {
2857 SDValue &Src = Input == 0 ? Src1 : Src2;
2858 if (RangeUse[Input] == 0)
2859 Src = DAG.getUNDEF(VT);
2861 Src = DAG.getNode(ISD::EXTRACT_SUBVECTOR, getCurDebugLoc(), VT,
2862 Src, DAG.getIntPtrConstant(StartIdx[Input]));
2865 // Calculate new mask.
2866 SmallVector<int, 8> MappedOps;
2867 for (unsigned i = 0; i != MaskNumElts; ++i) {
2870 MappedOps.push_back(Idx);
2871 else if (Idx < (int)SrcNumElts)
2872 MappedOps.push_back(Idx - StartIdx[0]);
2874 MappedOps.push_back(Idx - SrcNumElts - StartIdx[1] + MaskNumElts);
2877 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2883 // We can't use either concat vectors or extract subvectors so fall back to
2884 // replacing the shuffle with extract and build vector.
2885 // to insert and build vector.
2886 EVT EltVT = VT.getVectorElementType();
2887 EVT PtrVT = TLI.getPointerTy();
2888 SmallVector<SDValue,8> Ops;
2889 for (unsigned i = 0; i != MaskNumElts; ++i) {
2891 Ops.push_back(DAG.getUNDEF(EltVT));
2896 if (Idx < (int)SrcNumElts)
2897 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2898 EltVT, Src1, DAG.getConstant(Idx, PtrVT));
2900 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2902 DAG.getConstant(Idx - SrcNumElts, PtrVT));
2908 setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
2909 VT, &Ops[0], Ops.size()));
2912 void SelectionDAGBuilder::visitInsertValue(const InsertValueInst &I) {
2913 const Value *Op0 = I.getOperand(0);
2914 const Value *Op1 = I.getOperand(1);
2915 Type *AggTy = I.getType();
2916 Type *ValTy = Op1->getType();
2917 bool IntoUndef = isa<UndefValue>(Op0);
2918 bool FromUndef = isa<UndefValue>(Op1);
2920 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
2922 SmallVector<EVT, 4> AggValueVTs;
2923 ComputeValueVTs(TLI, AggTy, AggValueVTs);
2924 SmallVector<EVT, 4> ValValueVTs;
2925 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2927 unsigned NumAggValues = AggValueVTs.size();
2928 unsigned NumValValues = ValValueVTs.size();
2929 SmallVector<SDValue, 4> Values(NumAggValues);
2931 SDValue Agg = getValue(Op0);
2933 // Copy the beginning value(s) from the original aggregate.
2934 for (; i != LinearIndex; ++i)
2935 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2936 SDValue(Agg.getNode(), Agg.getResNo() + i);
2937 // Copy values from the inserted value(s).
2939 SDValue Val = getValue(Op1);
2940 for (; i != LinearIndex + NumValValues; ++i)
2941 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2942 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex);
2944 // Copy remaining value(s) from the original aggregate.
2945 for (; i != NumAggValues; ++i)
2946 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2947 SDValue(Agg.getNode(), Agg.getResNo() + i);
2949 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2950 DAG.getVTList(&AggValueVTs[0], NumAggValues),
2951 &Values[0], NumAggValues));
2954 void SelectionDAGBuilder::visitExtractValue(const ExtractValueInst &I) {
2955 const Value *Op0 = I.getOperand(0);
2956 Type *AggTy = Op0->getType();
2957 Type *ValTy = I.getType();
2958 bool OutOfUndef = isa<UndefValue>(Op0);
2960 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
2962 SmallVector<EVT, 4> ValValueVTs;
2963 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2965 unsigned NumValValues = ValValueVTs.size();
2967 // Ignore a extractvalue that produces an empty object
2968 if (!NumValValues) {
2969 setValue(&I, DAG.getUNDEF(MVT(MVT::Other)));
2973 SmallVector<SDValue, 4> Values(NumValValues);
2975 SDValue Agg = getValue(Op0);
2976 // Copy out the selected value(s).
2977 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i)
2978 Values[i - LinearIndex] =
2980 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) :
2981 SDValue(Agg.getNode(), Agg.getResNo() + i);
2983 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2984 DAG.getVTList(&ValValueVTs[0], NumValValues),
2985 &Values[0], NumValValues));
2988 void SelectionDAGBuilder::visitLandingPad(const LandingPadInst &I) {
2991 void SelectionDAGBuilder::visitGetElementPtr(const User &I) {
2992 SDValue N = getValue(I.getOperand(0));
2993 Type *Ty = I.getOperand(0)->getType();
2995 for (GetElementPtrInst::const_op_iterator OI = I.op_begin()+1, E = I.op_end();
2997 const Value *Idx = *OI;
2998 if (StructType *StTy = dyn_cast<StructType>(Ty)) {
2999 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
3002 uint64_t Offset = TD->getStructLayout(StTy)->getElementOffset(Field);
3003 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
3004 DAG.getIntPtrConstant(Offset));
3007 Ty = StTy->getElementType(Field);
3009 Ty = cast<SequentialType>(Ty)->getElementType();
3011 // If this is a constant subscript, handle it quickly.
3012 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
3013 if (CI->isZero()) continue;
3015 TD->getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
3017 EVT PTy = TLI.getPointerTy();
3018 unsigned PtrBits = PTy.getSizeInBits();
3020 OffsVal = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
3022 DAG.getConstant(Offs, MVT::i64));
3024 OffsVal = DAG.getIntPtrConstant(Offs);
3026 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
3031 // N = N + Idx * ElementSize;
3032 APInt ElementSize = APInt(TLI.getPointerTy().getSizeInBits(),
3033 TD->getTypeAllocSize(Ty));
3034 SDValue IdxN = getValue(Idx);
3036 // If the index is smaller or larger than intptr_t, truncate or extend
3038 IdxN = DAG.getSExtOrTrunc(IdxN, getCurDebugLoc(), N.getValueType());
3040 // If this is a multiply by a power of two, turn it into a shl
3041 // immediately. This is a very common case.
3042 if (ElementSize != 1) {
3043 if (ElementSize.isPowerOf2()) {
3044 unsigned Amt = ElementSize.logBase2();
3045 IdxN = DAG.getNode(ISD::SHL, getCurDebugLoc(),
3046 N.getValueType(), IdxN,
3047 DAG.getConstant(Amt, TLI.getPointerTy()));
3049 SDValue Scale = DAG.getConstant(ElementSize, TLI.getPointerTy());
3050 IdxN = DAG.getNode(ISD::MUL, getCurDebugLoc(),
3051 N.getValueType(), IdxN, Scale);
3055 N = DAG.getNode(ISD::ADD, getCurDebugLoc(),
3056 N.getValueType(), N, IdxN);
3063 void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) {
3064 // If this is a fixed sized alloca in the entry block of the function,
3065 // allocate it statically on the stack.
3066 if (FuncInfo.StaticAllocaMap.count(&I))
3067 return; // getValue will auto-populate this.
3069 Type *Ty = I.getAllocatedType();
3070 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
3072 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty),
3075 SDValue AllocSize = getValue(I.getArraySize());
3077 EVT IntPtr = TLI.getPointerTy();
3078 if (AllocSize.getValueType() != IntPtr)
3079 AllocSize = DAG.getZExtOrTrunc(AllocSize, getCurDebugLoc(), IntPtr);
3081 AllocSize = DAG.getNode(ISD::MUL, getCurDebugLoc(), IntPtr,
3083 DAG.getConstant(TySize, IntPtr));
3085 // Handle alignment. If the requested alignment is less than or equal to
3086 // the stack alignment, ignore it. If the size is greater than or equal to
3087 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
3088 unsigned StackAlign = TM.getFrameLowering()->getStackAlignment();
3089 if (Align <= StackAlign)
3092 // Round the size of the allocation up to the stack alignment size
3093 // by add SA-1 to the size.
3094 AllocSize = DAG.getNode(ISD::ADD, getCurDebugLoc(),
3095 AllocSize.getValueType(), AllocSize,
3096 DAG.getIntPtrConstant(StackAlign-1));
3098 // Mask out the low bits for alignment purposes.
3099 AllocSize = DAG.getNode(ISD::AND, getCurDebugLoc(),
3100 AllocSize.getValueType(), AllocSize,
3101 DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1)));
3103 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) };
3104 SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other);
3105 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, getCurDebugLoc(),
3108 DAG.setRoot(DSA.getValue(1));
3110 // Inform the Frame Information that we have just allocated a variable-sized
3112 FuncInfo.MF->getFrameInfo()->CreateVariableSizedObject(Align ? Align : 1);
3115 void SelectionDAGBuilder::visitLoad(const LoadInst &I) {
3116 const Value *SV = I.getOperand(0);
3117 SDValue Ptr = getValue(SV);
3119 Type *Ty = I.getType();
3121 bool isVolatile = I.isVolatile();
3122 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
3123 unsigned Alignment = I.getAlignment();
3124 const MDNode *TBAAInfo = I.getMetadata(LLVMContext::MD_tbaa);
3126 SmallVector<EVT, 4> ValueVTs;
3127 SmallVector<uint64_t, 4> Offsets;
3128 ComputeValueVTs(TLI, Ty, ValueVTs, &Offsets);
3129 unsigned NumValues = ValueVTs.size();
3134 bool ConstantMemory = false;
3135 if (I.isVolatile() || NumValues > MaxParallelChains)
3136 // Serialize volatile loads with other side effects.
3138 else if (AA->pointsToConstantMemory(
3139 AliasAnalysis::Location(SV, AA->getTypeStoreSize(Ty), TBAAInfo))) {
3140 // Do not serialize (non-volatile) loads of constant memory with anything.
3141 Root = DAG.getEntryNode();
3142 ConstantMemory = true;
3144 // Do not serialize non-volatile loads against each other.
3145 Root = DAG.getRoot();
3148 SmallVector<SDValue, 4> Values(NumValues);
3149 SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains),
3151 EVT PtrVT = Ptr.getValueType();
3152 unsigned ChainI = 0;
3153 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3154 // Serializing loads here may result in excessive register pressure, and
3155 // TokenFactor places arbitrary choke points on the scheduler. SD scheduling
3156 // could recover a bit by hoisting nodes upward in the chain by recognizing
3157 // they are side-effect free or do not alias. The optimizer should really
3158 // avoid this case by converting large object/array copies to llvm.memcpy
3159 // (MaxParallelChains should always remain as failsafe).
3160 if (ChainI == MaxParallelChains) {
3161 assert(PendingLoads.empty() && "PendingLoads must be serialized first");
3162 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3163 MVT::Other, &Chains[0], ChainI);
3167 SDValue A = DAG.getNode(ISD::ADD, getCurDebugLoc(),
3169 DAG.getConstant(Offsets[i], PtrVT));
3170 SDValue L = DAG.getLoad(ValueVTs[i], getCurDebugLoc(), Root,
3171 A, MachinePointerInfo(SV, Offsets[i]), isVolatile,
3172 isNonTemporal, Alignment, TBAAInfo);
3175 Chains[ChainI] = L.getValue(1);
3178 if (!ConstantMemory) {
3179 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3180 MVT::Other, &Chains[0], ChainI);
3184 PendingLoads.push_back(Chain);
3187 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
3188 DAG.getVTList(&ValueVTs[0], NumValues),
3189 &Values[0], NumValues));
3192 void SelectionDAGBuilder::visitStore(const StoreInst &I) {
3193 const Value *SrcV = I.getOperand(0);
3194 const Value *PtrV = I.getOperand(1);
3196 SmallVector<EVT, 4> ValueVTs;
3197 SmallVector<uint64_t, 4> Offsets;
3198 ComputeValueVTs(TLI, SrcV->getType(), ValueVTs, &Offsets);
3199 unsigned NumValues = ValueVTs.size();
3203 // Get the lowered operands. Note that we do this after
3204 // checking if NumResults is zero, because with zero results
3205 // the operands won't have values in the map.
3206 SDValue Src = getValue(SrcV);
3207 SDValue Ptr = getValue(PtrV);
3209 SDValue Root = getRoot();
3210 SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains),
3212 EVT PtrVT = Ptr.getValueType();
3213 bool isVolatile = I.isVolatile();
3214 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
3215 unsigned Alignment = I.getAlignment();
3216 const MDNode *TBAAInfo = I.getMetadata(LLVMContext::MD_tbaa);
3218 unsigned ChainI = 0;
3219 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3220 // See visitLoad comments.
3221 if (ChainI == MaxParallelChains) {
3222 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3223 MVT::Other, &Chains[0], ChainI);
3227 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT, Ptr,
3228 DAG.getConstant(Offsets[i], PtrVT));
3229 SDValue St = DAG.getStore(Root, getCurDebugLoc(),
3230 SDValue(Src.getNode(), Src.getResNo() + i),
3231 Add, MachinePointerInfo(PtrV, Offsets[i]),
3232 isVolatile, isNonTemporal, Alignment, TBAAInfo);
3233 Chains[ChainI] = St;
3236 SDValue StoreNode = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3237 MVT::Other, &Chains[0], ChainI);
3239 AssignOrderingToNode(StoreNode.getNode());
3240 DAG.setRoot(StoreNode);
3243 static SDValue InsertFenceForAtomic(SDValue Chain, AtomicOrdering Order,
3244 bool Before, DebugLoc dl,
3246 const TargetLowering &TLI) {
3247 // Fence, if necessary
3249 if (Order == AcquireRelease)
3251 else if (Order == Acquire || Order == Monotonic)
3254 if (Order == AcquireRelease)
3256 else if (Order == Release || Order == Monotonic)
3261 Ops[1] = DAG.getConstant(SequentiallyConsistent, TLI.getPointerTy());
3262 Ops[2] = DAG.getConstant(Order, TLI.getPointerTy());
3263 return DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops, 3);
3266 void SelectionDAGBuilder::visitAtomicCmpXchg(const AtomicCmpXchgInst &I) {
3267 DebugLoc dl = getCurDebugLoc();
3268 AtomicOrdering Order = I.getOrdering();
3270 SDValue InChain = getRoot();
3272 if (TLI.getInsertFencesForAtomic())
3273 InChain = InsertFenceForAtomic(InChain, Order, true, dl, DAG, TLI);
3276 DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, dl,
3277 getValue(I.getCompareOperand()).getValueType().getSimpleVT(),
3279 getValue(I.getPointerOperand()),
3280 getValue(I.getCompareOperand()),
3281 getValue(I.getNewValOperand()),
3282 MachinePointerInfo(I.getPointerOperand()), 0 /* Alignment */,
3283 I.getOrdering(), I.getSynchScope());
3285 SDValue OutChain = L.getValue(1);
3287 if (TLI.getInsertFencesForAtomic())
3288 OutChain = InsertFenceForAtomic(OutChain, Order, false, dl, DAG, TLI);
3291 DAG.setRoot(OutChain);
3294 void SelectionDAGBuilder::visitAtomicRMW(const AtomicRMWInst &I) {
3295 DebugLoc dl = getCurDebugLoc();
3297 switch (I.getOperation()) {
3298 default: llvm_unreachable("Unknown atomicrmw operation"); return;
3299 case AtomicRMWInst::Xchg: NT = ISD::ATOMIC_SWAP; break;
3300 case AtomicRMWInst::Add: NT = ISD::ATOMIC_LOAD_ADD; break;
3301 case AtomicRMWInst::Sub: NT = ISD::ATOMIC_LOAD_SUB; break;
3302 case AtomicRMWInst::And: NT = ISD::ATOMIC_LOAD_AND; break;
3303 case AtomicRMWInst::Nand: NT = ISD::ATOMIC_LOAD_NAND; break;
3304 case AtomicRMWInst::Or: NT = ISD::ATOMIC_LOAD_OR; break;
3305 case AtomicRMWInst::Xor: NT = ISD::ATOMIC_LOAD_XOR; break;
3306 case AtomicRMWInst::Max: NT = ISD::ATOMIC_LOAD_MAX; break;
3307 case AtomicRMWInst::Min: NT = ISD::ATOMIC_LOAD_MIN; break;
3308 case AtomicRMWInst::UMax: NT = ISD::ATOMIC_LOAD_UMAX; break;
3309 case AtomicRMWInst::UMin: NT = ISD::ATOMIC_LOAD_UMIN; break;
3311 AtomicOrdering Order = I.getOrdering();
3313 SDValue InChain = getRoot();
3315 if (TLI.getInsertFencesForAtomic())
3316 InChain = InsertFenceForAtomic(InChain, Order, true, dl, DAG, TLI);
3319 DAG.getAtomic(NT, dl,
3320 getValue(I.getValOperand()).getValueType().getSimpleVT(),
3322 getValue(I.getPointerOperand()),
3323 getValue(I.getValOperand()),
3324 I.getPointerOperand(), 0 /* Alignment */,
3325 TLI.getInsertFencesForAtomic() ? Monotonic : Order,
3328 SDValue OutChain = L.getValue(1);
3330 if (TLI.getInsertFencesForAtomic())
3331 OutChain = InsertFenceForAtomic(OutChain, Order, false, dl, DAG, TLI);
3334 DAG.setRoot(OutChain);
3337 void SelectionDAGBuilder::visitFence(const FenceInst &I) {
3338 DebugLoc dl = getCurDebugLoc();
3341 Ops[1] = DAG.getConstant(I.getOrdering(), TLI.getPointerTy());
3342 Ops[2] = DAG.getConstant(I.getSynchScope(), TLI.getPointerTy());
3343 DAG.setRoot(DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops, 3));
3346 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
3348 void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I,
3349 unsigned Intrinsic) {
3350 bool HasChain = !I.doesNotAccessMemory();
3351 bool OnlyLoad = HasChain && I.onlyReadsMemory();
3353 // Build the operand list.
3354 SmallVector<SDValue, 8> Ops;
3355 if (HasChain) { // If this intrinsic has side-effects, chainify it.
3357 // We don't need to serialize loads against other loads.
3358 Ops.push_back(DAG.getRoot());
3360 Ops.push_back(getRoot());
3364 // Info is set by getTgtMemInstrinsic
3365 TargetLowering::IntrinsicInfo Info;
3366 bool IsTgtIntrinsic = TLI.getTgtMemIntrinsic(Info, I, Intrinsic);
3368 // Add the intrinsic ID as an integer operand if it's not a target intrinsic.
3369 if (!IsTgtIntrinsic || Info.opc == ISD::INTRINSIC_VOID ||
3370 Info.opc == ISD::INTRINSIC_W_CHAIN)
3371 Ops.push_back(DAG.getConstant(Intrinsic, TLI.getPointerTy()));
3373 // Add all operands of the call to the operand list.
3374 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) {
3375 SDValue Op = getValue(I.getArgOperand(i));
3376 assert(TLI.isTypeLegal(Op.getValueType()) &&
3377 "Intrinsic uses a non-legal type?");
3381 SmallVector<EVT, 4> ValueVTs;
3382 ComputeValueVTs(TLI, I.getType(), ValueVTs);
3384 for (unsigned Val = 0, E = ValueVTs.size(); Val != E; ++Val) {
3385 assert(TLI.isTypeLegal(ValueVTs[Val]) &&
3386 "Intrinsic uses a non-legal type?");
3391 ValueVTs.push_back(MVT::Other);
3393 SDVTList VTs = DAG.getVTList(ValueVTs.data(), ValueVTs.size());
3397 if (IsTgtIntrinsic) {
3398 // This is target intrinsic that touches memory
3399 Result = DAG.getMemIntrinsicNode(Info.opc, getCurDebugLoc(),
3400 VTs, &Ops[0], Ops.size(),
3402 MachinePointerInfo(Info.ptrVal, Info.offset),
3403 Info.align, Info.vol,
3404 Info.readMem, Info.writeMem);
3405 } else if (!HasChain) {
3406 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurDebugLoc(),
3407 VTs, &Ops[0], Ops.size());
3408 } else if (!I.getType()->isVoidTy()) {
3409 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurDebugLoc(),
3410 VTs, &Ops[0], Ops.size());
3412 Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurDebugLoc(),
3413 VTs, &Ops[0], Ops.size());
3417 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1);
3419 PendingLoads.push_back(Chain);
3424 if (!I.getType()->isVoidTy()) {
3425 if (VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
3426 EVT VT = TLI.getValueType(PTy);
3427 Result = DAG.getNode(ISD::BITCAST, getCurDebugLoc(), VT, Result);
3430 setValue(&I, Result);
3434 /// GetSignificand - Get the significand and build it into a floating-point
3435 /// number with exponent of 1:
3437 /// Op = (Op & 0x007fffff) | 0x3f800000;
3439 /// where Op is the hexidecimal representation of floating point value.
3441 GetSignificand(SelectionDAG &DAG, SDValue Op, DebugLoc dl) {
3442 SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3443 DAG.getConstant(0x007fffff, MVT::i32));
3444 SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1,
3445 DAG.getConstant(0x3f800000, MVT::i32));
3446 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, t2);
3449 /// GetExponent - Get the exponent:
3451 /// (float)(int)(((Op & 0x7f800000) >> 23) - 127);
3453 /// where Op is the hexidecimal representation of floating point value.
3455 GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI,
3457 SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3458 DAG.getConstant(0x7f800000, MVT::i32));
3459 SDValue t1 = DAG.getNode(ISD::SRL, dl, MVT::i32, t0,
3460 DAG.getConstant(23, TLI.getPointerTy()));
3461 SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1,
3462 DAG.getConstant(127, MVT::i32));
3463 return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2);
3466 /// getF32Constant - Get 32-bit floating point constant.
3468 getF32Constant(SelectionDAG &DAG, unsigned Flt) {
3469 return DAG.getConstantFP(APFloat(APInt(32, Flt)), MVT::f32);
3472 /// Inlined utility function to implement binary input atomic intrinsics for
3473 /// visitIntrinsicCall: I is a call instruction
3474 /// Op is the associated NodeType for I
3476 SelectionDAGBuilder::implVisitBinaryAtomic(const CallInst& I,
3478 SDValue Root = getRoot();
3480 DAG.getAtomic(Op, getCurDebugLoc(),
3481 getValue(I.getArgOperand(1)).getValueType().getSimpleVT(),
3483 getValue(I.getArgOperand(0)),
3484 getValue(I.getArgOperand(1)),
3485 I.getArgOperand(0), 0 /* Alignment */,
3486 Monotonic, CrossThread);
3488 DAG.setRoot(L.getValue(1));
3492 // implVisitAluOverflow - Lower arithmetic overflow instrinsics.
3494 SelectionDAGBuilder::implVisitAluOverflow(const CallInst &I, ISD::NodeType Op) {
3495 SDValue Op1 = getValue(I.getArgOperand(0));
3496 SDValue Op2 = getValue(I.getArgOperand(1));
3498 SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1);
3499 setValue(&I, DAG.getNode(Op, getCurDebugLoc(), VTs, Op1, Op2));
3503 /// visitExp - Lower an exp intrinsic. Handles the special sequences for
3504 /// limited-precision mode.
3506 SelectionDAGBuilder::visitExp(const CallInst &I) {
3508 DebugLoc dl = getCurDebugLoc();
3510 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3511 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3512 SDValue Op = getValue(I.getArgOperand(0));
3514 // Put the exponent in the right bit position for later addition to the
3517 // #define LOG2OFe 1.4426950f
3518 // IntegerPartOfX = ((int32_t)(X * LOG2OFe));
3519 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3520 getF32Constant(DAG, 0x3fb8aa3b));
3521 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3523 // FractionalPartOfX = (X * LOG2OFe) - (float)IntegerPartOfX;
3524 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3525 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3527 // IntegerPartOfX <<= 23;
3528 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3529 DAG.getConstant(23, TLI.getPointerTy()));
3531 if (LimitFloatPrecision <= 6) {
3532 // For floating-point precision of 6:
3534 // TwoToFractionalPartOfX =
3536 // (0.735607626f + 0.252464424f * x) * x;
3538 // error 0.0144103317, which is 6 bits
3539 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3540 getF32Constant(DAG, 0x3e814304));
3541 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3542 getF32Constant(DAG, 0x3f3c50c8));
3543 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3544 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3545 getF32Constant(DAG, 0x3f7f5e7e));
3546 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BITCAST, dl,MVT::i32, t5);
3548 // Add the exponent into the result in integer domain.
3549 SDValue t6 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3550 TwoToFracPartOfX, IntegerPartOfX);
3552 result = DAG.getNode(ISD::BITCAST, dl, MVT::f32, t6);
3553 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3554 // For floating-point precision of 12:
3556 // TwoToFractionalPartOfX =
3559 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3561 // 0.000107046256 error, which is 13 to 14 bits
3562 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3563 getF32Constant(DAG, 0x3da235e3));
3564 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3565 getF32Constant(DAG, 0x3e65b8f3));
3566 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3567 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3568 getF32Constant(DAG, 0x3f324b07));
3569 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3570 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3571 getF32Constant(DAG, 0x3f7ff8fd));
3572 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BITCAST, dl,MVT::i32, t7);
3574 // Add the exponent into the result in integer domain.
3575 SDValue t8 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3576 TwoToFracPartOfX, IntegerPartOfX);
3578 result = DAG.getNode(ISD::BITCAST, dl, MVT::f32, t8);
3579 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3580 // For floating-point precision of 18:
3582 // TwoToFractionalPartOfX =
3586 // (0.554906021e-1f +
3587 // (0.961591928e-2f +
3588 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3590 // error 2.47208000*10^(-7), which is better than 18 bits
3591 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3592 getF32Constant(DAG, 0x3924b03e));
3593 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3594 getF32Constant(DAG, 0x3ab24b87));
3595 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3596 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3597 getF32Constant(DAG, 0x3c1d8c17));
3598 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3599 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3600 getF32Constant(DAG, 0x3d634a1d));
3601 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3602 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3603 getF32Constant(DAG, 0x3e75fe14));
3604 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3605 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3606 getF32Constant(DAG, 0x3f317234));
3607 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3608 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3609 getF32Constant(DAG, 0x3f800000));
3610 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BITCAST, dl,
3613 // Add the exponent into the result in integer domain.
3614 SDValue t14 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3615 TwoToFracPartOfX, IntegerPartOfX);
3617 result = DAG.getNode(ISD::BITCAST, dl, MVT::f32, t14);
3620 // No special expansion.
3621 result = DAG.getNode(ISD::FEXP, dl,
3622 getValue(I.getArgOperand(0)).getValueType(),
3623 getValue(I.getArgOperand(0)));
3626 setValue(&I, result);
3629 /// visitLog - Lower a log intrinsic. Handles the special sequences for
3630 /// limited-precision mode.
3632 SelectionDAGBuilder::visitLog(const CallInst &I) {
3634 DebugLoc dl = getCurDebugLoc();
3636 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3637 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3638 SDValue Op = getValue(I.getArgOperand(0));
3639 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3641 // Scale the exponent by log(2) [0.69314718f].
3642 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3643 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3644 getF32Constant(DAG, 0x3f317218));
3646 // Get the significand and build it into a floating-point number with
3648 SDValue X = GetSignificand(DAG, Op1, dl);
3650 if (LimitFloatPrecision <= 6) {
3651 // For floating-point precision of 6:
3655 // (1.4034025f - 0.23903021f * x) * x;
3657 // error 0.0034276066, which is better than 8 bits
3658 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3659 getF32Constant(DAG, 0xbe74c456));
3660 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3661 getF32Constant(DAG, 0x3fb3a2b1));
3662 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3663 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3664 getF32Constant(DAG, 0x3f949a29));
3666 result = DAG.getNode(ISD::FADD, dl,
3667 MVT::f32, LogOfExponent, LogOfMantissa);
3668 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3669 // For floating-point precision of 12:
3675 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x;
3677 // error 0.000061011436, which is 14 bits
3678 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3679 getF32Constant(DAG, 0xbd67b6d6));
3680 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3681 getF32Constant(DAG, 0x3ee4f4b8));
3682 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3683 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3684 getF32Constant(DAG, 0x3fbc278b));
3685 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3686 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3687 getF32Constant(DAG, 0x40348e95));
3688 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3689 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3690 getF32Constant(DAG, 0x3fdef31a));
3692 result = DAG.getNode(ISD::FADD, dl,
3693 MVT::f32, LogOfExponent, LogOfMantissa);
3694 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3695 // For floating-point precision of 18:
3703 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x;
3705 // error 0.0000023660568, which is better than 18 bits
3706 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3707 getF32Constant(DAG, 0xbc91e5ac));
3708 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3709 getF32Constant(DAG, 0x3e4350aa));
3710 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3711 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3712 getF32Constant(DAG, 0x3f60d3e3));
3713 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3714 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3715 getF32Constant(DAG, 0x4011cdf0));
3716 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3717 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3718 getF32Constant(DAG, 0x406cfd1c));
3719 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3720 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3721 getF32Constant(DAG, 0x408797cb));
3722 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3723 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3724 getF32Constant(DAG, 0x4006dcab));
3726 result = DAG.getNode(ISD::FADD, dl,
3727 MVT::f32, LogOfExponent, LogOfMantissa);
3730 // No special expansion.
3731 result = DAG.getNode(ISD::FLOG, dl,
3732 getValue(I.getArgOperand(0)).getValueType(),
3733 getValue(I.getArgOperand(0)));
3736 setValue(&I, result);
3739 /// visitLog2 - Lower a log2 intrinsic. Handles the special sequences for
3740 /// limited-precision mode.
3742 SelectionDAGBuilder::visitLog2(const CallInst &I) {
3744 DebugLoc dl = getCurDebugLoc();
3746 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3747 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3748 SDValue Op = getValue(I.getArgOperand(0));
3749 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3751 // Get the exponent.
3752 SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl);
3754 // Get the significand and build it into a floating-point number with
3756 SDValue X = GetSignificand(DAG, Op1, dl);
3758 // Different possible minimax approximations of significand in
3759 // floating-point for various degrees of accuracy over [1,2].
3760 if (LimitFloatPrecision <= 6) {
3761 // For floating-point precision of 6:
3763 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x;
3765 // error 0.0049451742, which is more than 7 bits
3766 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3767 getF32Constant(DAG, 0xbeb08fe0));
3768 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3769 getF32Constant(DAG, 0x40019463));
3770 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3771 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3772 getF32Constant(DAG, 0x3fd6633d));
3774 result = DAG.getNode(ISD::FADD, dl,
3775 MVT::f32, LogOfExponent, Log2ofMantissa);
3776 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3777 // For floating-point precision of 12:
3783 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x;
3785 // error 0.0000876136000, which is better than 13 bits
3786 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3787 getF32Constant(DAG, 0xbda7262e));
3788 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3789 getF32Constant(DAG, 0x3f25280b));
3790 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3791 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3792 getF32Constant(DAG, 0x4007b923));
3793 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3794 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3795 getF32Constant(DAG, 0x40823e2f));
3796 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3797 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3798 getF32Constant(DAG, 0x4020d29c));
3800 result = DAG.getNode(ISD::FADD, dl,
3801 MVT::f32, LogOfExponent, Log2ofMantissa);
3802 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3803 // For floating-point precision of 18:
3812 // 0.25691327e-1f * x) * x) * x) * x) * x) * x;
3814 // error 0.0000018516, which is better than 18 bits
3815 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3816 getF32Constant(DAG, 0xbcd2769e));
3817 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3818 getF32Constant(DAG, 0x3e8ce0b9));
3819 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3820 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3821 getF32Constant(DAG, 0x3fa22ae7));
3822 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3823 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3824 getF32Constant(DAG, 0x40525723));
3825 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3826 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3827 getF32Constant(DAG, 0x40aaf200));
3828 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3829 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3830 getF32Constant(DAG, 0x40c39dad));
3831 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3832 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3833 getF32Constant(DAG, 0x4042902c));
3835 result = DAG.getNode(ISD::FADD, dl,
3836 MVT::f32, LogOfExponent, Log2ofMantissa);
3839 // No special expansion.
3840 result = DAG.getNode(ISD::FLOG2, dl,
3841 getValue(I.getArgOperand(0)).getValueType(),
3842 getValue(I.getArgOperand(0)));
3845 setValue(&I, result);
3848 /// visitLog10 - Lower a log10 intrinsic. Handles the special sequences for
3849 /// limited-precision mode.
3851 SelectionDAGBuilder::visitLog10(const CallInst &I) {
3853 DebugLoc dl = getCurDebugLoc();
3855 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3856 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3857 SDValue Op = getValue(I.getArgOperand(0));
3858 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3860 // Scale the exponent by log10(2) [0.30102999f].
3861 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3862 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3863 getF32Constant(DAG, 0x3e9a209a));
3865 // Get the significand and build it into a floating-point number with
3867 SDValue X = GetSignificand(DAG, Op1, dl);
3869 if (LimitFloatPrecision <= 6) {
3870 // For floating-point precision of 6:
3872 // Log10ofMantissa =
3874 // (0.60948995f - 0.10380950f * x) * x;
3876 // error 0.0014886165, which is 6 bits
3877 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3878 getF32Constant(DAG, 0xbdd49a13));
3879 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3880 getF32Constant(DAG, 0x3f1c0789));
3881 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3882 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3883 getF32Constant(DAG, 0x3f011300));
3885 result = DAG.getNode(ISD::FADD, dl,
3886 MVT::f32, LogOfExponent, Log10ofMantissa);
3887 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3888 // For floating-point precision of 12:
3890 // Log10ofMantissa =
3893 // (-0.31664806f + 0.47637168e-1f * x) * x) * x;
3895 // error 0.00019228036, which is better than 12 bits
3896 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3897 getF32Constant(DAG, 0x3d431f31));
3898 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
3899 getF32Constant(DAG, 0x3ea21fb2));
3900 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3901 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3902 getF32Constant(DAG, 0x3f6ae232));
3903 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3904 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
3905 getF32Constant(DAG, 0x3f25f7c3));
3907 result = DAG.getNode(ISD::FADD, dl,
3908 MVT::f32, LogOfExponent, Log10ofMantissa);
3909 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3910 // For floating-point precision of 18:
3912 // Log10ofMantissa =
3917 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x;
3919 // error 0.0000037995730, which is better than 18 bits
3920 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3921 getF32Constant(DAG, 0x3c5d51ce));
3922 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
3923 getF32Constant(DAG, 0x3e00685a));
3924 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3925 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3926 getF32Constant(DAG, 0x3efb6798));
3927 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3928 SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
3929 getF32Constant(DAG, 0x3f88d192));
3930 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3931 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3932 getF32Constant(DAG, 0x3fc4316c));
3933 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3934 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8,
3935 getF32Constant(DAG, 0x3f57ce70));
3937 result = DAG.getNode(ISD::FADD, dl,
3938 MVT::f32, LogOfExponent, Log10ofMantissa);
3941 // No special expansion.
3942 result = DAG.getNode(ISD::FLOG10, dl,
3943 getValue(I.getArgOperand(0)).getValueType(),
3944 getValue(I.getArgOperand(0)));
3947 setValue(&I, result);
3950 /// visitExp2 - Lower an exp2 intrinsic. Handles the special sequences for
3951 /// limited-precision mode.
3953 SelectionDAGBuilder::visitExp2(const CallInst &I) {
3955 DebugLoc dl = getCurDebugLoc();
3957 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3958 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3959 SDValue Op = getValue(I.getArgOperand(0));
3961 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Op);
3963 // FractionalPartOfX = x - (float)IntegerPartOfX;
3964 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3965 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, Op, t1);
3967 // IntegerPartOfX <<= 23;
3968 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3969 DAG.getConstant(23, TLI.getPointerTy()));
3971 if (LimitFloatPrecision <= 6) {
3972 // For floating-point precision of 6:
3974 // TwoToFractionalPartOfX =
3976 // (0.735607626f + 0.252464424f * x) * x;
3978 // error 0.0144103317, which is 6 bits
3979 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3980 getF32Constant(DAG, 0x3e814304));
3981 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3982 getF32Constant(DAG, 0x3f3c50c8));
3983 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3984 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3985 getF32Constant(DAG, 0x3f7f5e7e));
3986 SDValue t6 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t5);
3987 SDValue TwoToFractionalPartOfX =
3988 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
3990 result = DAG.getNode(ISD::BITCAST, dl,
3991 MVT::f32, TwoToFractionalPartOfX);
3992 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3993 // For floating-point precision of 12:
3995 // TwoToFractionalPartOfX =
3998 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
4000 // error 0.000107046256, which is 13 to 14 bits
4001 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4002 getF32Constant(DAG, 0x3da235e3));
4003 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4004 getF32Constant(DAG, 0x3e65b8f3));
4005 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4006 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4007 getF32Constant(DAG, 0x3f324b07));
4008 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4009 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4010 getF32Constant(DAG, 0x3f7ff8fd));
4011 SDValue t8 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t7);
4012 SDValue TwoToFractionalPartOfX =
4013 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
4015 result = DAG.getNode(ISD::BITCAST, dl,
4016 MVT::f32, TwoToFractionalPartOfX);
4017 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
4018 // For floating-point precision of 18:
4020 // TwoToFractionalPartOfX =
4024 // (0.554906021e-1f +
4025 // (0.961591928e-2f +
4026 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
4027 // error 2.47208000*10^(-7), which is better than 18 bits
4028 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4029 getF32Constant(DAG, 0x3924b03e));
4030 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4031 getF32Constant(DAG, 0x3ab24b87));
4032 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4033 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4034 getF32Constant(DAG, 0x3c1d8c17));
4035 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4036 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4037 getF32Constant(DAG, 0x3d634a1d));
4038 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4039 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4040 getF32Constant(DAG, 0x3e75fe14));
4041 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4042 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
4043 getF32Constant(DAG, 0x3f317234));
4044 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
4045 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
4046 getF32Constant(DAG, 0x3f800000));
4047 SDValue t14 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t13);
4048 SDValue TwoToFractionalPartOfX =
4049 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
4051 result = DAG.getNode(ISD::BITCAST, dl,
4052 MVT::f32, TwoToFractionalPartOfX);
4055 // No special expansion.
4056 result = DAG.getNode(ISD::FEXP2, dl,
4057 getValue(I.getArgOperand(0)).getValueType(),
4058 getValue(I.getArgOperand(0)));
4061 setValue(&I, result);
4064 /// visitPow - Lower a pow intrinsic. Handles the special sequences for
4065 /// limited-precision mode with x == 10.0f.
4067 SelectionDAGBuilder::visitPow(const CallInst &I) {
4069 const Value *Val = I.getArgOperand(0);
4070 DebugLoc dl = getCurDebugLoc();
4071 bool IsExp10 = false;
4073 if (getValue(Val).getValueType() == MVT::f32 &&
4074 getValue(I.getArgOperand(1)).getValueType() == MVT::f32 &&
4075 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4076 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(Val))) {
4077 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
4079 IsExp10 = CFP->getValueAPF().bitwiseIsEqual(Ten);
4084 if (IsExp10 && LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4085 SDValue Op = getValue(I.getArgOperand(1));
4087 // Put the exponent in the right bit position for later addition to the
4090 // #define LOG2OF10 3.3219281f
4091 // IntegerPartOfX = (int32_t)(x * LOG2OF10);
4092 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
4093 getF32Constant(DAG, 0x40549a78));
4094 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
4096 // FractionalPartOfX = x - (float)IntegerPartOfX;
4097 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
4098 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
4100 // IntegerPartOfX <<= 23;
4101 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
4102 DAG.getConstant(23, TLI.getPointerTy()));
4104 if (LimitFloatPrecision <= 6) {
4105 // For floating-point precision of 6:
4107 // twoToFractionalPartOfX =
4109 // (0.735607626f + 0.252464424f * x) * x;
4111 // error 0.0144103317, which is 6 bits
4112 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4113 getF32Constant(DAG, 0x3e814304));
4114 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4115 getF32Constant(DAG, 0x3f3c50c8));
4116 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4117 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4118 getF32Constant(DAG, 0x3f7f5e7e));
4119 SDValue t6 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t5);
4120 SDValue TwoToFractionalPartOfX =
4121 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
4123 result = DAG.getNode(ISD::BITCAST, dl,
4124 MVT::f32, TwoToFractionalPartOfX);
4125 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
4126 // For floating-point precision of 12:
4128 // TwoToFractionalPartOfX =
4131 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
4133 // error 0.000107046256, which is 13 to 14 bits
4134 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4135 getF32Constant(DAG, 0x3da235e3));
4136 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4137 getF32Constant(DAG, 0x3e65b8f3));
4138 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4139 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4140 getF32Constant(DAG, 0x3f324b07));
4141 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4142 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4143 getF32Constant(DAG, 0x3f7ff8fd));
4144 SDValue t8 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t7);
4145 SDValue TwoToFractionalPartOfX =
4146 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
4148 result = DAG.getNode(ISD::BITCAST, dl,
4149 MVT::f32, TwoToFractionalPartOfX);
4150 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
4151 // For floating-point precision of 18:
4153 // TwoToFractionalPartOfX =
4157 // (0.554906021e-1f +
4158 // (0.961591928e-2f +
4159 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
4160 // error 2.47208000*10^(-7), which is better than 18 bits
4161 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4162 getF32Constant(DAG, 0x3924b03e));
4163 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4164 getF32Constant(DAG, 0x3ab24b87));
4165 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4166 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4167 getF32Constant(DAG, 0x3c1d8c17));
4168 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4169 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4170 getF32Constant(DAG, 0x3d634a1d));
4171 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4172 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4173 getF32Constant(DAG, 0x3e75fe14));
4174 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4175 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
4176 getF32Constant(DAG, 0x3f317234));
4177 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
4178 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
4179 getF32Constant(DAG, 0x3f800000));
4180 SDValue t14 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t13);
4181 SDValue TwoToFractionalPartOfX =
4182 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
4184 result = DAG.getNode(ISD::BITCAST, dl,
4185 MVT::f32, TwoToFractionalPartOfX);
4188 // No special expansion.
4189 result = DAG.getNode(ISD::FPOW, dl,
4190 getValue(I.getArgOperand(0)).getValueType(),
4191 getValue(I.getArgOperand(0)),
4192 getValue(I.getArgOperand(1)));
4195 setValue(&I, result);
4199 /// ExpandPowI - Expand a llvm.powi intrinsic.
4200 static SDValue ExpandPowI(DebugLoc DL, SDValue LHS, SDValue RHS,
4201 SelectionDAG &DAG) {
4202 // If RHS is a constant, we can expand this out to a multiplication tree,
4203 // otherwise we end up lowering to a call to __powidf2 (for example). When
4204 // optimizing for size, we only want to do this if the expansion would produce
4205 // a small number of multiplies, otherwise we do the full expansion.
4206 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
4207 // Get the exponent as a positive value.
4208 unsigned Val = RHSC->getSExtValue();
4209 if ((int)Val < 0) Val = -Val;
4211 // powi(x, 0) -> 1.0
4213 return DAG.getConstantFP(1.0, LHS.getValueType());
4215 const Function *F = DAG.getMachineFunction().getFunction();
4216 if (!F->hasFnAttr(Attribute::OptimizeForSize) ||
4217 // If optimizing for size, don't insert too many multiplies. This
4218 // inserts up to 5 multiplies.
4219 CountPopulation_32(Val)+Log2_32(Val) < 7) {
4220 // We use the simple binary decomposition method to generate the multiply
4221 // sequence. There are more optimal ways to do this (for example,
4222 // powi(x,15) generates one more multiply than it should), but this has
4223 // the benefit of being both really simple and much better than a libcall.
4224 SDValue Res; // Logically starts equal to 1.0
4225 SDValue CurSquare = LHS;
4229 Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare);
4231 Res = CurSquare; // 1.0*CurSquare.
4234 CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(),
4235 CurSquare, CurSquare);
4239 // If the original was negative, invert the result, producing 1/(x*x*x).
4240 if (RHSC->getSExtValue() < 0)
4241 Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(),
4242 DAG.getConstantFP(1.0, LHS.getValueType()), Res);
4247 // Otherwise, expand to a libcall.
4248 return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS);
4251 // getTruncatedArgReg - Find underlying register used for an truncated
4253 static unsigned getTruncatedArgReg(const SDValue &N) {
4254 if (N.getOpcode() != ISD::TRUNCATE)
4257 const SDValue &Ext = N.getOperand(0);
4258 if (Ext.getOpcode() == ISD::AssertZext || Ext.getOpcode() == ISD::AssertSext){
4259 const SDValue &CFR = Ext.getOperand(0);
4260 if (CFR.getOpcode() == ISD::CopyFromReg)
4261 return cast<RegisterSDNode>(CFR.getOperand(1))->getReg();
4263 if (CFR.getOpcode() == ISD::TRUNCATE)
4264 return getTruncatedArgReg(CFR);
4269 /// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a function
4270 /// argument, create the corresponding DBG_VALUE machine instruction for it now.
4271 /// At the end of instruction selection, they will be inserted to the entry BB.
4273 SelectionDAGBuilder::EmitFuncArgumentDbgValue(const Value *V, MDNode *Variable,
4276 const Argument *Arg = dyn_cast<Argument>(V);
4280 MachineFunction &MF = DAG.getMachineFunction();
4281 const TargetInstrInfo *TII = DAG.getTarget().getInstrInfo();
4282 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
4284 // Ignore inlined function arguments here.
4285 DIVariable DV(Variable);
4286 if (DV.isInlinedFnArgument(MF.getFunction()))
4290 if (Arg->hasByValAttr()) {
4291 // Byval arguments' frame index is recorded during argument lowering.
4292 // Use this info directly.
4293 Reg = TRI->getFrameRegister(MF);
4294 Offset = FuncInfo.getByValArgumentFrameIndex(Arg);
4295 // If byval argument ofset is not recorded then ignore this.
4301 if (N.getOpcode() == ISD::CopyFromReg)
4302 Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg();
4304 Reg = getTruncatedArgReg(N);
4305 if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) {
4306 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4307 unsigned PR = RegInfo.getLiveInPhysReg(Reg);
4314 // Check if ValueMap has reg number.
4315 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
4316 if (VMI != FuncInfo.ValueMap.end())
4320 if (!Reg && N.getNode()) {
4321 // Check if frame index is available.
4322 if (LoadSDNode *LNode = dyn_cast<LoadSDNode>(N.getNode()))
4323 if (FrameIndexSDNode *FINode =
4324 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode())) {
4325 Reg = TRI->getFrameRegister(MF);
4326 Offset = FINode->getIndex();
4333 MachineInstrBuilder MIB = BuildMI(MF, getCurDebugLoc(),
4334 TII->get(TargetOpcode::DBG_VALUE))
4335 .addReg(Reg, RegState::Debug).addImm(Offset).addMetadata(Variable);
4336 FuncInfo.ArgDbgValues.push_back(&*MIB);
4340 // VisualStudio defines setjmp as _setjmp
4341 #if defined(_MSC_VER) && defined(setjmp) && \
4342 !defined(setjmp_undefined_for_msvc)
4343 # pragma push_macro("setjmp")
4345 # define setjmp_undefined_for_msvc
4348 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
4349 /// we want to emit this as a call to a named external function, return the name
4350 /// otherwise lower it and return null.
4352 SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) {
4353 DebugLoc dl = getCurDebugLoc();
4356 switch (Intrinsic) {
4358 // By default, turn this into a target intrinsic node.
4359 visitTargetIntrinsic(I, Intrinsic);
4361 case Intrinsic::vastart: visitVAStart(I); return 0;
4362 case Intrinsic::vaend: visitVAEnd(I); return 0;
4363 case Intrinsic::vacopy: visitVACopy(I); return 0;
4364 case Intrinsic::returnaddress:
4365 setValue(&I, DAG.getNode(ISD::RETURNADDR, dl, TLI.getPointerTy(),
4366 getValue(I.getArgOperand(0))));
4368 case Intrinsic::frameaddress:
4369 setValue(&I, DAG.getNode(ISD::FRAMEADDR, dl, TLI.getPointerTy(),
4370 getValue(I.getArgOperand(0))));
4372 case Intrinsic::setjmp:
4373 return "_setjmp"+!TLI.usesUnderscoreSetJmp();
4374 case Intrinsic::longjmp:
4375 return "_longjmp"+!TLI.usesUnderscoreLongJmp();
4376 case Intrinsic::memcpy: {
4377 // Assert for address < 256 since we support only user defined address
4379 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4381 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4383 "Unknown address space");
4384 SDValue Op1 = getValue(I.getArgOperand(0));
4385 SDValue Op2 = getValue(I.getArgOperand(1));
4386 SDValue Op3 = getValue(I.getArgOperand(2));
4387 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4388 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4389 DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, isVol, false,
4390 MachinePointerInfo(I.getArgOperand(0)),
4391 MachinePointerInfo(I.getArgOperand(1))));
4394 case Intrinsic::memset: {
4395 // Assert for address < 256 since we support only user defined address
4397 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4399 "Unknown address space");
4400 SDValue Op1 = getValue(I.getArgOperand(0));
4401 SDValue Op2 = getValue(I.getArgOperand(1));
4402 SDValue Op3 = getValue(I.getArgOperand(2));
4403 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4404 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4405 DAG.setRoot(DAG.getMemset(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4406 MachinePointerInfo(I.getArgOperand(0))));
4409 case Intrinsic::memmove: {
4410 // Assert for address < 256 since we support only user defined address
4412 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4414 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4416 "Unknown address space");
4417 SDValue Op1 = getValue(I.getArgOperand(0));
4418 SDValue Op2 = getValue(I.getArgOperand(1));
4419 SDValue Op3 = getValue(I.getArgOperand(2));
4420 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4421 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4422 DAG.setRoot(DAG.getMemmove(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4423 MachinePointerInfo(I.getArgOperand(0)),
4424 MachinePointerInfo(I.getArgOperand(1))));
4427 case Intrinsic::dbg_declare: {
4428 const DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
4429 MDNode *Variable = DI.getVariable();
4430 const Value *Address = DI.getAddress();
4431 if (!Address || !DIVariable(DI.getVariable()).Verify())
4434 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder
4435 // but do not always have a corresponding SDNode built. The SDNodeOrder
4436 // absolute, but not relative, values are different depending on whether
4437 // debug info exists.
4440 // Check if address has undef value.
4441 if (isa<UndefValue>(Address) ||
4442 (Address->use_empty() && !isa<Argument>(Address))) {
4443 DEBUG(dbgs() << "Dropping debug info for " << DI);
4447 SDValue &N = NodeMap[Address];
4448 if (!N.getNode() && isa<Argument>(Address))
4449 // Check unused arguments map.
4450 N = UnusedArgNodeMap[Address];
4453 // Parameters are handled specially.
4455 DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable;
4456 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
4457 Address = BCI->getOperand(0);
4458 const AllocaInst *AI = dyn_cast<AllocaInst>(Address);
4460 if (isParameter && !AI) {
4461 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode());
4463 // Byval parameter. We have a frame index at this point.
4464 SDV = DAG.getDbgValue(Variable, FINode->getIndex(),
4465 0, dl, SDNodeOrder);
4467 // Address is an argument, so try to emit its dbg value using
4468 // virtual register info from the FuncInfo.ValueMap.
4469 EmitFuncArgumentDbgValue(Address, Variable, 0, N);
4473 SDV = DAG.getDbgValue(Variable, N.getNode(), N.getResNo(),
4474 0, dl, SDNodeOrder);
4476 // Can't do anything with other non-AI cases yet.
4477 DEBUG(dbgs() << "Dropping debug info for " << DI);
4480 DAG.AddDbgValue(SDV, N.getNode(), isParameter);
4482 // If Address is an argument then try to emit its dbg value using
4483 // virtual register info from the FuncInfo.ValueMap.
4484 if (!EmitFuncArgumentDbgValue(Address, Variable, 0, N)) {
4485 // If variable is pinned by a alloca in dominating bb then
4486 // use StaticAllocaMap.
4487 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Address)) {
4488 if (AI->getParent() != DI.getParent()) {
4489 DenseMap<const AllocaInst*, int>::iterator SI =
4490 FuncInfo.StaticAllocaMap.find(AI);
4491 if (SI != FuncInfo.StaticAllocaMap.end()) {
4492 SDV = DAG.getDbgValue(Variable, SI->second,
4493 0, dl, SDNodeOrder);
4494 DAG.AddDbgValue(SDV, 0, false);
4499 DEBUG(dbgs() << "Dropping debug info for " << DI);
4504 case Intrinsic::dbg_value: {
4505 const DbgValueInst &DI = cast<DbgValueInst>(I);
4506 if (!DIVariable(DI.getVariable()).Verify())
4509 MDNode *Variable = DI.getVariable();
4510 uint64_t Offset = DI.getOffset();
4511 const Value *V = DI.getValue();
4515 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder
4516 // but do not always have a corresponding SDNode built. The SDNodeOrder
4517 // absolute, but not relative, values are different depending on whether
4518 // debug info exists.
4521 if (isa<ConstantInt>(V) || isa<ConstantFP>(V) || isa<UndefValue>(V)) {
4522 SDV = DAG.getDbgValue(Variable, V, Offset, dl, SDNodeOrder);
4523 DAG.AddDbgValue(SDV, 0, false);
4525 // Do not use getValue() in here; we don't want to generate code at
4526 // this point if it hasn't been done yet.
4527 SDValue N = NodeMap[V];
4528 if (!N.getNode() && isa<Argument>(V))
4529 // Check unused arguments map.
4530 N = UnusedArgNodeMap[V];
4532 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, N)) {
4533 SDV = DAG.getDbgValue(Variable, N.getNode(),
4534 N.getResNo(), Offset, dl, SDNodeOrder);
4535 DAG.AddDbgValue(SDV, N.getNode(), false);
4537 } else if (!V->use_empty() ) {
4538 // Do not call getValue(V) yet, as we don't want to generate code.
4539 // Remember it for later.
4540 DanglingDebugInfo DDI(&DI, dl, SDNodeOrder);
4541 DanglingDebugInfoMap[V] = DDI;
4543 // We may expand this to cover more cases. One case where we have no
4544 // data available is an unreferenced parameter.
4545 DEBUG(dbgs() << "Dropping debug info for " << DI);
4549 // Build a debug info table entry.
4550 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
4551 V = BCI->getOperand(0);
4552 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
4553 // Don't handle byval struct arguments or VLAs, for example.
4556 DenseMap<const AllocaInst*, int>::iterator SI =
4557 FuncInfo.StaticAllocaMap.find(AI);
4558 if (SI == FuncInfo.StaticAllocaMap.end())
4560 int FI = SI->second;
4562 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4563 if (!DI.getDebugLoc().isUnknown() && MMI.hasDebugInfo())
4564 MMI.setVariableDbgInfo(Variable, FI, DI.getDebugLoc());
4567 case Intrinsic::eh_exception: {
4568 // Insert the EXCEPTIONADDR instruction.
4569 assert(FuncInfo.MBB->isLandingPad() &&
4570 "Call to eh.exception not in landing pad!");
4571 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
4573 Ops[0] = DAG.getRoot();
4574 SDValue Op = DAG.getNode(ISD::EXCEPTIONADDR, dl, VTs, Ops, 1);
4576 DAG.setRoot(Op.getValue(1));
4580 case Intrinsic::eh_selector: {
4581 MachineBasicBlock *CallMBB = FuncInfo.MBB;
4582 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4583 if (CallMBB->isLandingPad())
4584 AddCatchInfo(I, &MMI, CallMBB);
4587 FuncInfo.CatchInfoLost.insert(&I);
4589 // FIXME: Mark exception selector register as live in. Hack for PR1508.
4590 unsigned Reg = TLI.getExceptionSelectorRegister();
4591 if (Reg) FuncInfo.MBB->addLiveIn(Reg);
4594 // Insert the EHSELECTION instruction.
4595 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
4597 Ops[0] = getValue(I.getArgOperand(0));
4599 SDValue Op = DAG.getNode(ISD::EHSELECTION, dl, VTs, Ops, 2);
4600 DAG.setRoot(Op.getValue(1));
4601 setValue(&I, DAG.getSExtOrTrunc(Op, dl, MVT::i32));
4605 case Intrinsic::eh_typeid_for: {
4606 // Find the type id for the given typeinfo.
4607 GlobalVariable *GV = ExtractTypeInfo(I.getArgOperand(0));
4608 unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(GV);
4609 Res = DAG.getConstant(TypeID, MVT::i32);
4614 case Intrinsic::eh_return_i32:
4615 case Intrinsic::eh_return_i64:
4616 DAG.getMachineFunction().getMMI().setCallsEHReturn(true);
4617 DAG.setRoot(DAG.getNode(ISD::EH_RETURN, dl,
4620 getValue(I.getArgOperand(0)),
4621 getValue(I.getArgOperand(1))));
4623 case Intrinsic::eh_unwind_init:
4624 DAG.getMachineFunction().getMMI().setCallsUnwindInit(true);
4626 case Intrinsic::eh_dwarf_cfa: {
4627 SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getArgOperand(0)), dl,
4628 TLI.getPointerTy());
4629 SDValue Offset = DAG.getNode(ISD::ADD, dl,
4631 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, dl,
4632 TLI.getPointerTy()),
4634 SDValue FA = DAG.getNode(ISD::FRAMEADDR, dl,
4636 DAG.getConstant(0, TLI.getPointerTy()));
4637 setValue(&I, DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
4641 case Intrinsic::eh_sjlj_callsite: {
4642 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4643 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(0));
4644 assert(CI && "Non-constant call site value in eh.sjlj.callsite!");
4645 assert(MMI.getCurrentCallSite() == 0 && "Overlapping call sites!");
4647 MMI.setCurrentCallSite(CI->getZExtValue());
4650 case Intrinsic::eh_sjlj_setjmp: {
4651 setValue(&I, DAG.getNode(ISD::EH_SJLJ_SETJMP, dl, MVT::i32, getRoot(),
4652 getValue(I.getArgOperand(0))));
4655 case Intrinsic::eh_sjlj_longjmp: {
4656 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, dl, MVT::Other,
4657 getRoot(), getValue(I.getArgOperand(0))));
4660 case Intrinsic::eh_sjlj_dispatch_setup: {
4661 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_DISPATCHSETUP, dl, MVT::Other,
4662 getRoot(), getValue(I.getArgOperand(0))));
4666 case Intrinsic::x86_mmx_pslli_w:
4667 case Intrinsic::x86_mmx_pslli_d:
4668 case Intrinsic::x86_mmx_pslli_q:
4669 case Intrinsic::x86_mmx_psrli_w:
4670 case Intrinsic::x86_mmx_psrli_d:
4671 case Intrinsic::x86_mmx_psrli_q:
4672 case Intrinsic::x86_mmx_psrai_w:
4673 case Intrinsic::x86_mmx_psrai_d: {
4674 SDValue ShAmt = getValue(I.getArgOperand(1));
4675 if (isa<ConstantSDNode>(ShAmt)) {
4676 visitTargetIntrinsic(I, Intrinsic);
4679 unsigned NewIntrinsic = 0;
4680 EVT ShAmtVT = MVT::v2i32;
4681 switch (Intrinsic) {
4682 case Intrinsic::x86_mmx_pslli_w:
4683 NewIntrinsic = Intrinsic::x86_mmx_psll_w;
4685 case Intrinsic::x86_mmx_pslli_d:
4686 NewIntrinsic = Intrinsic::x86_mmx_psll_d;
4688 case Intrinsic::x86_mmx_pslli_q:
4689 NewIntrinsic = Intrinsic::x86_mmx_psll_q;
4691 case Intrinsic::x86_mmx_psrli_w:
4692 NewIntrinsic = Intrinsic::x86_mmx_psrl_w;
4694 case Intrinsic::x86_mmx_psrli_d:
4695 NewIntrinsic = Intrinsic::x86_mmx_psrl_d;
4697 case Intrinsic::x86_mmx_psrli_q:
4698 NewIntrinsic = Intrinsic::x86_mmx_psrl_q;
4700 case Intrinsic::x86_mmx_psrai_w:
4701 NewIntrinsic = Intrinsic::x86_mmx_psra_w;
4703 case Intrinsic::x86_mmx_psrai_d:
4704 NewIntrinsic = Intrinsic::x86_mmx_psra_d;
4706 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4709 // The vector shift intrinsics with scalars uses 32b shift amounts but
4710 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
4712 // We must do this early because v2i32 is not a legal type.
4713 DebugLoc dl = getCurDebugLoc();
4716 ShOps[1] = DAG.getConstant(0, MVT::i32);
4717 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
4718 EVT DestVT = TLI.getValueType(I.getType());
4719 ShAmt = DAG.getNode(ISD::BITCAST, dl, DestVT, ShAmt);
4720 Res = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4721 DAG.getConstant(NewIntrinsic, MVT::i32),
4722 getValue(I.getArgOperand(0)), ShAmt);
4726 case Intrinsic::convertff:
4727 case Intrinsic::convertfsi:
4728 case Intrinsic::convertfui:
4729 case Intrinsic::convertsif:
4730 case Intrinsic::convertuif:
4731 case Intrinsic::convertss:
4732 case Intrinsic::convertsu:
4733 case Intrinsic::convertus:
4734 case Intrinsic::convertuu: {
4735 ISD::CvtCode Code = ISD::CVT_INVALID;
4736 switch (Intrinsic) {
4737 case Intrinsic::convertff: Code = ISD::CVT_FF; break;
4738 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break;
4739 case Intrinsic::convertfui: Code = ISD::CVT_FU; break;
4740 case Intrinsic::convertsif: Code = ISD::CVT_SF; break;
4741 case Intrinsic::convertuif: Code = ISD::CVT_UF; break;
4742 case Intrinsic::convertss: Code = ISD::CVT_SS; break;
4743 case Intrinsic::convertsu: Code = ISD::CVT_SU; break;
4744 case Intrinsic::convertus: Code = ISD::CVT_US; break;
4745 case Intrinsic::convertuu: Code = ISD::CVT_UU; break;
4747 EVT DestVT = TLI.getValueType(I.getType());
4748 const Value *Op1 = I.getArgOperand(0);
4749 Res = DAG.getConvertRndSat(DestVT, getCurDebugLoc(), getValue(Op1),
4750 DAG.getValueType(DestVT),
4751 DAG.getValueType(getValue(Op1).getValueType()),
4752 getValue(I.getArgOperand(1)),
4753 getValue(I.getArgOperand(2)),
4758 case Intrinsic::sqrt:
4759 setValue(&I, DAG.getNode(ISD::FSQRT, dl,
4760 getValue(I.getArgOperand(0)).getValueType(),
4761 getValue(I.getArgOperand(0))));
4763 case Intrinsic::powi:
4764 setValue(&I, ExpandPowI(dl, getValue(I.getArgOperand(0)),
4765 getValue(I.getArgOperand(1)), DAG));
4767 case Intrinsic::sin:
4768 setValue(&I, DAG.getNode(ISD::FSIN, dl,
4769 getValue(I.getArgOperand(0)).getValueType(),
4770 getValue(I.getArgOperand(0))));
4772 case Intrinsic::cos:
4773 setValue(&I, DAG.getNode(ISD::FCOS, dl,
4774 getValue(I.getArgOperand(0)).getValueType(),
4775 getValue(I.getArgOperand(0))));
4777 case Intrinsic::log:
4780 case Intrinsic::log2:
4783 case Intrinsic::log10:
4786 case Intrinsic::exp:
4789 case Intrinsic::exp2:
4792 case Intrinsic::pow:
4795 case Intrinsic::fma:
4796 setValue(&I, DAG.getNode(ISD::FMA, dl,
4797 getValue(I.getArgOperand(0)).getValueType(),
4798 getValue(I.getArgOperand(0)),
4799 getValue(I.getArgOperand(1)),
4800 getValue(I.getArgOperand(2))));
4802 case Intrinsic::convert_to_fp16:
4803 setValue(&I, DAG.getNode(ISD::FP32_TO_FP16, dl,
4804 MVT::i16, getValue(I.getArgOperand(0))));
4806 case Intrinsic::convert_from_fp16:
4807 setValue(&I, DAG.getNode(ISD::FP16_TO_FP32, dl,
4808 MVT::f32, getValue(I.getArgOperand(0))));
4810 case Intrinsic::pcmarker: {
4811 SDValue Tmp = getValue(I.getArgOperand(0));
4812 DAG.setRoot(DAG.getNode(ISD::PCMARKER, dl, MVT::Other, getRoot(), Tmp));
4815 case Intrinsic::readcyclecounter: {
4816 SDValue Op = getRoot();
4817 Res = DAG.getNode(ISD::READCYCLECOUNTER, dl,
4818 DAG.getVTList(MVT::i64, MVT::Other),
4821 DAG.setRoot(Res.getValue(1));
4824 case Intrinsic::bswap:
4825 setValue(&I, DAG.getNode(ISD::BSWAP, dl,
4826 getValue(I.getArgOperand(0)).getValueType(),
4827 getValue(I.getArgOperand(0))));
4829 case Intrinsic::cttz: {
4830 SDValue Arg = getValue(I.getArgOperand(0));
4831 EVT Ty = Arg.getValueType();
4832 setValue(&I, DAG.getNode(ISD::CTTZ, dl, Ty, Arg));
4835 case Intrinsic::ctlz: {
4836 SDValue Arg = getValue(I.getArgOperand(0));
4837 EVT Ty = Arg.getValueType();
4838 setValue(&I, DAG.getNode(ISD::CTLZ, dl, Ty, Arg));
4841 case Intrinsic::ctpop: {
4842 SDValue Arg = getValue(I.getArgOperand(0));
4843 EVT Ty = Arg.getValueType();
4844 setValue(&I, DAG.getNode(ISD::CTPOP, dl, Ty, Arg));
4847 case Intrinsic::stacksave: {
4848 SDValue Op = getRoot();
4849 Res = DAG.getNode(ISD::STACKSAVE, dl,
4850 DAG.getVTList(TLI.getPointerTy(), MVT::Other), &Op, 1);
4852 DAG.setRoot(Res.getValue(1));
4855 case Intrinsic::stackrestore: {
4856 Res = getValue(I.getArgOperand(0));
4857 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, dl, MVT::Other, getRoot(), Res));
4860 case Intrinsic::stackprotector: {
4861 // Emit code into the DAG to store the stack guard onto the stack.
4862 MachineFunction &MF = DAG.getMachineFunction();
4863 MachineFrameInfo *MFI = MF.getFrameInfo();
4864 EVT PtrTy = TLI.getPointerTy();
4866 SDValue Src = getValue(I.getArgOperand(0)); // The guard's value.
4867 AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
4869 int FI = FuncInfo.StaticAllocaMap[Slot];
4870 MFI->setStackProtectorIndex(FI);
4872 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
4874 // Store the stack protector onto the stack.
4875 Res = DAG.getStore(getRoot(), getCurDebugLoc(), Src, FIN,
4876 MachinePointerInfo::getFixedStack(FI),
4882 case Intrinsic::objectsize: {
4883 // If we don't know by now, we're never going to know.
4884 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1));
4886 assert(CI && "Non-constant type in __builtin_object_size?");
4888 SDValue Arg = getValue(I.getCalledValue());
4889 EVT Ty = Arg.getValueType();
4892 Res = DAG.getConstant(-1ULL, Ty);
4894 Res = DAG.getConstant(0, Ty);
4899 case Intrinsic::var_annotation:
4900 // Discard annotate attributes
4903 case Intrinsic::init_trampoline: {
4904 const Function *F = cast<Function>(I.getArgOperand(1)->stripPointerCasts());
4908 Ops[1] = getValue(I.getArgOperand(0));
4909 Ops[2] = getValue(I.getArgOperand(1));
4910 Ops[3] = getValue(I.getArgOperand(2));
4911 Ops[4] = DAG.getSrcValue(I.getArgOperand(0));
4912 Ops[5] = DAG.getSrcValue(F);
4914 Res = DAG.getNode(ISD::TRAMPOLINE, dl,
4915 DAG.getVTList(TLI.getPointerTy(), MVT::Other),
4919 DAG.setRoot(Res.getValue(1));
4922 case Intrinsic::gcroot:
4924 const Value *Alloca = I.getArgOperand(0);
4925 const Constant *TypeMap = cast<Constant>(I.getArgOperand(1));
4927 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode());
4928 GFI->addStackRoot(FI->getIndex(), TypeMap);
4931 case Intrinsic::gcread:
4932 case Intrinsic::gcwrite:
4933 llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!");
4935 case Intrinsic::flt_rounds:
4936 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, dl, MVT::i32));
4939 case Intrinsic::expect: {
4940 // Just replace __builtin_expect(exp, c) with EXP.
4941 setValue(&I, getValue(I.getArgOperand(0)));
4945 case Intrinsic::trap: {
4946 StringRef TrapFuncName = getTrapFunctionName();
4947 if (TrapFuncName.empty()) {
4948 DAG.setRoot(DAG.getNode(ISD::TRAP, dl,MVT::Other, getRoot()));
4951 TargetLowering::ArgListTy Args;
4952 std::pair<SDValue, SDValue> Result =
4953 TLI.LowerCallTo(getRoot(), I.getType(),
4954 false, false, false, false, 0, CallingConv::C,
4955 /*isTailCall=*/false, /*isReturnValueUsed=*/true,
4956 DAG.getExternalSymbol(TrapFuncName.data(), TLI.getPointerTy()),
4957 Args, DAG, getCurDebugLoc());
4958 DAG.setRoot(Result.second);
4961 case Intrinsic::uadd_with_overflow:
4962 return implVisitAluOverflow(I, ISD::UADDO);
4963 case Intrinsic::sadd_with_overflow:
4964 return implVisitAluOverflow(I, ISD::SADDO);
4965 case Intrinsic::usub_with_overflow:
4966 return implVisitAluOverflow(I, ISD::USUBO);
4967 case Intrinsic::ssub_with_overflow:
4968 return implVisitAluOverflow(I, ISD::SSUBO);
4969 case Intrinsic::umul_with_overflow:
4970 return implVisitAluOverflow(I, ISD::UMULO);
4971 case Intrinsic::smul_with_overflow:
4972 return implVisitAluOverflow(I, ISD::SMULO);
4974 case Intrinsic::prefetch: {
4976 unsigned rw = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
4978 Ops[1] = getValue(I.getArgOperand(0));
4979 Ops[2] = getValue(I.getArgOperand(1));
4980 Ops[3] = getValue(I.getArgOperand(2));
4981 Ops[4] = getValue(I.getArgOperand(3));
4982 DAG.setRoot(DAG.getMemIntrinsicNode(ISD::PREFETCH, dl,
4983 DAG.getVTList(MVT::Other),
4985 EVT::getIntegerVT(*Context, 8),
4986 MachinePointerInfo(I.getArgOperand(0)),
4988 false, /* volatile */
4990 rw==1)); /* write */
4993 case Intrinsic::memory_barrier: {
4996 for (int x = 1; x < 6; ++x)
4997 Ops[x] = getValue(I.getArgOperand(x - 1));
4999 DAG.setRoot(DAG.getNode(ISD::MEMBARRIER, dl, MVT::Other, &Ops[0], 6));
5002 case Intrinsic::atomic_cmp_swap: {
5003 SDValue Root = getRoot();
5005 DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, getCurDebugLoc(),
5006 getValue(I.getArgOperand(1)).getValueType().getSimpleVT(),
5008 getValue(I.getArgOperand(0)),
5009 getValue(I.getArgOperand(1)),
5010 getValue(I.getArgOperand(2)),
5011 MachinePointerInfo(I.getArgOperand(0)), 0 /* Alignment */,
5012 Monotonic, CrossThread);
5014 DAG.setRoot(L.getValue(1));
5017 case Intrinsic::atomic_load_add:
5018 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_ADD);
5019 case Intrinsic::atomic_load_sub:
5020 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_SUB);
5021 case Intrinsic::atomic_load_or:
5022 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_OR);
5023 case Intrinsic::atomic_load_xor:
5024 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_XOR);
5025 case Intrinsic::atomic_load_and:
5026 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_AND);
5027 case Intrinsic::atomic_load_nand:
5028 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_NAND);
5029 case Intrinsic::atomic_load_max:
5030 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MAX);
5031 case Intrinsic::atomic_load_min:
5032 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MIN);
5033 case Intrinsic::atomic_load_umin:
5034 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMIN);
5035 case Intrinsic::atomic_load_umax:
5036 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMAX);
5037 case Intrinsic::atomic_swap:
5038 return implVisitBinaryAtomic(I, ISD::ATOMIC_SWAP);
5040 case Intrinsic::invariant_start:
5041 case Intrinsic::lifetime_start:
5042 // Discard region information.
5043 setValue(&I, DAG.getUNDEF(TLI.getPointerTy()));
5045 case Intrinsic::invariant_end:
5046 case Intrinsic::lifetime_end:
5047 // Discard region information.
5052 void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee,
5054 MachineBasicBlock *LandingPad) {
5055 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
5056 FunctionType *FTy = cast<FunctionType>(PT->getElementType());
5057 Type *RetTy = FTy->getReturnType();
5058 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5059 MCSymbol *BeginLabel = 0;
5061 TargetLowering::ArgListTy Args;
5062 TargetLowering::ArgListEntry Entry;
5063 Args.reserve(CS.arg_size());
5065 // Check whether the function can return without sret-demotion.
5066 SmallVector<ISD::OutputArg, 4> Outs;
5067 SmallVector<uint64_t, 4> Offsets;
5068 GetReturnInfo(RetTy, CS.getAttributes().getRetAttributes(),
5069 Outs, TLI, &Offsets);
5071 bool CanLowerReturn = TLI.CanLowerReturn(CS.getCallingConv(),
5072 DAG.getMachineFunction(),
5073 FTy->isVarArg(), Outs,
5076 SDValue DemoteStackSlot;
5077 int DemoteStackIdx = -100;
5079 if (!CanLowerReturn) {
5080 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(
5081 FTy->getReturnType());
5082 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(
5083 FTy->getReturnType());
5084 MachineFunction &MF = DAG.getMachineFunction();
5085 DemoteStackIdx = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
5086 Type *StackSlotPtrType = PointerType::getUnqual(FTy->getReturnType());
5088 DemoteStackSlot = DAG.getFrameIndex(DemoteStackIdx, TLI.getPointerTy());
5089 Entry.Node = DemoteStackSlot;
5090 Entry.Ty = StackSlotPtrType;
5091 Entry.isSExt = false;
5092 Entry.isZExt = false;
5093 Entry.isInReg = false;
5094 Entry.isSRet = true;
5095 Entry.isNest = false;
5096 Entry.isByVal = false;
5097 Entry.Alignment = Align;
5098 Args.push_back(Entry);
5099 RetTy = Type::getVoidTy(FTy->getContext());
5102 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
5104 const Value *V = *i;
5107 if (V->getType()->isEmptyTy())
5110 SDValue ArgNode = getValue(V);
5111 Entry.Node = ArgNode; Entry.Ty = V->getType();
5113 unsigned attrInd = i - CS.arg_begin() + 1;
5114 Entry.isSExt = CS.paramHasAttr(attrInd, Attribute::SExt);
5115 Entry.isZExt = CS.paramHasAttr(attrInd, Attribute::ZExt);
5116 Entry.isInReg = CS.paramHasAttr(attrInd, Attribute::InReg);
5117 Entry.isSRet = CS.paramHasAttr(attrInd, Attribute::StructRet);
5118 Entry.isNest = CS.paramHasAttr(attrInd, Attribute::Nest);
5119 Entry.isByVal = CS.paramHasAttr(attrInd, Attribute::ByVal);
5120 Entry.Alignment = CS.getParamAlignment(attrInd);
5121 Args.push_back(Entry);
5125 // Insert a label before the invoke call to mark the try range. This can be
5126 // used to detect deletion of the invoke via the MachineModuleInfo.
5127 BeginLabel = MMI.getContext().CreateTempSymbol();
5129 // For SjLj, keep track of which landing pads go with which invokes
5130 // so as to maintain the ordering of pads in the LSDA.
5131 unsigned CallSiteIndex = MMI.getCurrentCallSite();
5132 if (CallSiteIndex) {
5133 MMI.setCallSiteBeginLabel(BeginLabel, CallSiteIndex);
5134 // Now that the call site is handled, stop tracking it.
5135 MMI.setCurrentCallSite(0);
5138 // Both PendingLoads and PendingExports must be flushed here;
5139 // this call might not return.
5141 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getControlRoot(), BeginLabel));
5144 // Check if target-independent constraints permit a tail call here.
5145 // Target-dependent constraints are checked within TLI.LowerCallTo.
5147 !isInTailCallPosition(CS, CS.getAttributes().getRetAttributes(), TLI))
5150 // If there's a possibility that fast-isel has already selected some amount
5151 // of the current basic block, don't emit a tail call.
5152 if (isTailCall && EnableFastISel)
5155 std::pair<SDValue,SDValue> Result =
5156 TLI.LowerCallTo(getRoot(), RetTy,
5157 CS.paramHasAttr(0, Attribute::SExt),
5158 CS.paramHasAttr(0, Attribute::ZExt), FTy->isVarArg(),
5159 CS.paramHasAttr(0, Attribute::InReg), FTy->getNumParams(),
5160 CS.getCallingConv(),
5162 !CS.getInstruction()->use_empty(),
5163 Callee, Args, DAG, getCurDebugLoc());
5164 assert((isTailCall || Result.second.getNode()) &&
5165 "Non-null chain expected with non-tail call!");
5166 assert((Result.second.getNode() || !Result.first.getNode()) &&
5167 "Null value expected with tail call!");
5168 if (Result.first.getNode()) {
5169 setValue(CS.getInstruction(), Result.first);
5170 } else if (!CanLowerReturn && Result.second.getNode()) {
5171 // The instruction result is the result of loading from the
5172 // hidden sret parameter.
5173 SmallVector<EVT, 1> PVTs;
5174 Type *PtrRetTy = PointerType::getUnqual(FTy->getReturnType());
5176 ComputeValueVTs(TLI, PtrRetTy, PVTs);
5177 assert(PVTs.size() == 1 && "Pointers should fit in one register");
5178 EVT PtrVT = PVTs[0];
5179 unsigned NumValues = Outs.size();
5180 SmallVector<SDValue, 4> Values(NumValues);
5181 SmallVector<SDValue, 4> Chains(NumValues);
5183 for (unsigned i = 0; i < NumValues; ++i) {
5184 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT,
5186 DAG.getConstant(Offsets[i], PtrVT));
5187 SDValue L = DAG.getLoad(Outs[i].VT, getCurDebugLoc(), Result.second,
5189 MachinePointerInfo::getFixedStack(DemoteStackIdx, Offsets[i]),
5192 Chains[i] = L.getValue(1);
5195 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
5196 MVT::Other, &Chains[0], NumValues);
5197 PendingLoads.push_back(Chain);
5199 // Collect the legal value parts into potentially illegal values
5200 // that correspond to the original function's return values.
5201 SmallVector<EVT, 4> RetTys;
5202 RetTy = FTy->getReturnType();
5203 ComputeValueVTs(TLI, RetTy, RetTys);
5204 ISD::NodeType AssertOp = ISD::DELETED_NODE;
5205 SmallVector<SDValue, 4> ReturnValues;
5206 unsigned CurReg = 0;
5207 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
5209 EVT RegisterVT = TLI.getRegisterType(RetTy->getContext(), VT);
5210 unsigned NumRegs = TLI.getNumRegisters(RetTy->getContext(), VT);
5212 SDValue ReturnValue =
5213 getCopyFromParts(DAG, getCurDebugLoc(), &Values[CurReg], NumRegs,
5214 RegisterVT, VT, AssertOp);
5215 ReturnValues.push_back(ReturnValue);
5219 setValue(CS.getInstruction(),
5220 DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
5221 DAG.getVTList(&RetTys[0], RetTys.size()),
5222 &ReturnValues[0], ReturnValues.size()));
5225 // Assign order to nodes here. If the call does not produce a result, it won't
5226 // be mapped to a SDNode and visit() will not assign it an order number.
5227 if (!Result.second.getNode()) {
5228 // As a special case, a null chain means that a tail call has been emitted and
5229 // the DAG root is already updated.
5232 AssignOrderingToNode(DAG.getRoot().getNode());
5234 DAG.setRoot(Result.second);
5236 AssignOrderingToNode(Result.second.getNode());
5240 // Insert a label at the end of the invoke call to mark the try range. This
5241 // can be used to detect deletion of the invoke via the MachineModuleInfo.
5242 MCSymbol *EndLabel = MMI.getContext().CreateTempSymbol();
5243 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getRoot(), EndLabel));
5245 // Inform MachineModuleInfo of range.
5246 MMI.addInvoke(LandingPad, BeginLabel, EndLabel);
5250 /// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the
5251 /// value is equal or not-equal to zero.
5252 static bool IsOnlyUsedInZeroEqualityComparison(const Value *V) {
5253 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end();
5255 if (const ICmpInst *IC = dyn_cast<ICmpInst>(*UI))
5256 if (IC->isEquality())
5257 if (const Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
5258 if (C->isNullValue())
5260 // Unknown instruction.
5266 static SDValue getMemCmpLoad(const Value *PtrVal, MVT LoadVT,
5268 SelectionDAGBuilder &Builder) {
5270 // Check to see if this load can be trivially constant folded, e.g. if the
5271 // input is from a string literal.
5272 if (const Constant *LoadInput = dyn_cast<Constant>(PtrVal)) {
5273 // Cast pointer to the type we really want to load.
5274 LoadInput = ConstantExpr::getBitCast(const_cast<Constant *>(LoadInput),
5275 PointerType::getUnqual(LoadTy));
5277 if (const Constant *LoadCst =
5278 ConstantFoldLoadFromConstPtr(const_cast<Constant *>(LoadInput),
5280 return Builder.getValue(LoadCst);
5283 // Otherwise, we have to emit the load. If the pointer is to unfoldable but
5284 // still constant memory, the input chain can be the entry node.
5286 bool ConstantMemory = false;
5288 // Do not serialize (non-volatile) loads of constant memory with anything.
5289 if (Builder.AA->pointsToConstantMemory(PtrVal)) {
5290 Root = Builder.DAG.getEntryNode();
5291 ConstantMemory = true;
5293 // Do not serialize non-volatile loads against each other.
5294 Root = Builder.DAG.getRoot();
5297 SDValue Ptr = Builder.getValue(PtrVal);
5298 SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurDebugLoc(), Root,
5299 Ptr, MachinePointerInfo(PtrVal),
5301 false /*nontemporal*/, 1 /* align=1 */);
5303 if (!ConstantMemory)
5304 Builder.PendingLoads.push_back(LoadVal.getValue(1));
5309 /// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form.
5310 /// If so, return true and lower it, otherwise return false and it will be
5311 /// lowered like a normal call.
5312 bool SelectionDAGBuilder::visitMemCmpCall(const CallInst &I) {
5313 // Verify that the prototype makes sense. int memcmp(void*,void*,size_t)
5314 if (I.getNumArgOperands() != 3)
5317 const Value *LHS = I.getArgOperand(0), *RHS = I.getArgOperand(1);
5318 if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() ||
5319 !I.getArgOperand(2)->getType()->isIntegerTy() ||
5320 !I.getType()->isIntegerTy())
5323 const ConstantInt *Size = dyn_cast<ConstantInt>(I.getArgOperand(2));
5325 // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0
5326 // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0
5327 if (Size && IsOnlyUsedInZeroEqualityComparison(&I)) {
5328 bool ActuallyDoIt = true;
5331 switch (Size->getZExtValue()) {
5333 LoadVT = MVT::Other;
5335 ActuallyDoIt = false;
5339 LoadTy = Type::getInt16Ty(Size->getContext());
5343 LoadTy = Type::getInt32Ty(Size->getContext());
5347 LoadTy = Type::getInt64Ty(Size->getContext());
5351 LoadVT = MVT::v4i32;
5352 LoadTy = Type::getInt32Ty(Size->getContext());
5353 LoadTy = VectorType::get(LoadTy, 4);
5358 // This turns into unaligned loads. We only do this if the target natively
5359 // supports the MVT we'll be loading or if it is small enough (<= 4) that
5360 // we'll only produce a small number of byte loads.
5362 // Require that we can find a legal MVT, and only do this if the target
5363 // supports unaligned loads of that type. Expanding into byte loads would
5365 if (ActuallyDoIt && Size->getZExtValue() > 4) {
5366 // TODO: Handle 5 byte compare as 4-byte + 1 byte.
5367 // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads.
5368 if (!TLI.isTypeLegal(LoadVT) ||!TLI.allowsUnalignedMemoryAccesses(LoadVT))
5369 ActuallyDoIt = false;
5373 SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this);
5374 SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this);
5376 SDValue Res = DAG.getSetCC(getCurDebugLoc(), MVT::i1, LHSVal, RHSVal,
5378 EVT CallVT = TLI.getValueType(I.getType(), true);
5379 setValue(&I, DAG.getZExtOrTrunc(Res, getCurDebugLoc(), CallVT));
5389 void SelectionDAGBuilder::visitCall(const CallInst &I) {
5390 // Handle inline assembly differently.
5391 if (isa<InlineAsm>(I.getCalledValue())) {
5396 // See if any floating point values are being passed to this function. This is
5397 // used to emit an undefined reference to fltused on Windows.
5399 cast<FunctionType>(I.getCalledValue()->getType()->getContainedType(0));
5400 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5401 if (FT->isVarArg() &&
5402 !MMI.callsExternalVAFunctionWithFloatingPointArguments()) {
5403 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) {
5404 Type* T = I.getArgOperand(i)->getType();
5405 for (po_iterator<Type*> i = po_begin(T), e = po_end(T);
5407 if (!i->isFloatingPointTy()) continue;
5408 MMI.setCallsExternalVAFunctionWithFloatingPointArguments(true);
5414 const char *RenameFn = 0;
5415 if (Function *F = I.getCalledFunction()) {
5416 if (F->isDeclaration()) {
5417 if (const TargetIntrinsicInfo *II = TM.getIntrinsicInfo()) {
5418 if (unsigned IID = II->getIntrinsicID(F)) {
5419 RenameFn = visitIntrinsicCall(I, IID);
5424 if (unsigned IID = F->getIntrinsicID()) {
5425 RenameFn = visitIntrinsicCall(I, IID);
5431 // Check for well-known libc/libm calls. If the function is internal, it
5432 // can't be a library call.
5433 if (!F->hasLocalLinkage() && F->hasName()) {
5434 StringRef Name = F->getName();
5435 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl") {
5436 if (I.getNumArgOperands() == 2 && // Basic sanity checks.
5437 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5438 I.getType() == I.getArgOperand(0)->getType() &&
5439 I.getType() == I.getArgOperand(1)->getType()) {
5440 SDValue LHS = getValue(I.getArgOperand(0));
5441 SDValue RHS = getValue(I.getArgOperand(1));
5442 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurDebugLoc(),
5443 LHS.getValueType(), LHS, RHS));
5446 } else if (Name == "fabs" || Name == "fabsf" || Name == "fabsl") {
5447 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5448 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5449 I.getType() == I.getArgOperand(0)->getType()) {
5450 SDValue Tmp = getValue(I.getArgOperand(0));
5451 setValue(&I, DAG.getNode(ISD::FABS, getCurDebugLoc(),
5452 Tmp.getValueType(), Tmp));
5455 } else if (Name == "sin" || Name == "sinf" || Name == "sinl") {
5456 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5457 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5458 I.getType() == I.getArgOperand(0)->getType() &&
5459 I.onlyReadsMemory()) {
5460 SDValue Tmp = getValue(I.getArgOperand(0));
5461 setValue(&I, DAG.getNode(ISD::FSIN, getCurDebugLoc(),
5462 Tmp.getValueType(), Tmp));
5465 } else if (Name == "cos" || Name == "cosf" || Name == "cosl") {
5466 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5467 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5468 I.getType() == I.getArgOperand(0)->getType() &&
5469 I.onlyReadsMemory()) {
5470 SDValue Tmp = getValue(I.getArgOperand(0));
5471 setValue(&I, DAG.getNode(ISD::FCOS, getCurDebugLoc(),
5472 Tmp.getValueType(), Tmp));
5475 } else if (Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl") {
5476 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5477 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5478 I.getType() == I.getArgOperand(0)->getType() &&
5479 I.onlyReadsMemory()) {
5480 SDValue Tmp = getValue(I.getArgOperand(0));
5481 setValue(&I, DAG.getNode(ISD::FSQRT, getCurDebugLoc(),
5482 Tmp.getValueType(), Tmp));
5485 } else if (Name == "memcmp") {
5486 if (visitMemCmpCall(I))
5494 Callee = getValue(I.getCalledValue());
5496 Callee = DAG.getExternalSymbol(RenameFn, TLI.getPointerTy());
5498 // Check if we can potentially perform a tail call. More detailed checking is
5499 // be done within LowerCallTo, after more information about the call is known.
5500 LowerCallTo(&I, Callee, I.isTailCall());
5505 /// AsmOperandInfo - This contains information for each constraint that we are
5507 class SDISelAsmOperandInfo : public TargetLowering::AsmOperandInfo {
5509 /// CallOperand - If this is the result output operand or a clobber
5510 /// this is null, otherwise it is the incoming operand to the CallInst.
5511 /// This gets modified as the asm is processed.
5512 SDValue CallOperand;
5514 /// AssignedRegs - If this is a register or register class operand, this
5515 /// contains the set of register corresponding to the operand.
5516 RegsForValue AssignedRegs;
5518 explicit SDISelAsmOperandInfo(const TargetLowering::AsmOperandInfo &info)
5519 : TargetLowering::AsmOperandInfo(info), CallOperand(0,0) {
5522 /// MarkAllocatedRegs - Once AssignedRegs is set, mark the assigned registers
5523 /// busy in OutputRegs/InputRegs.
5524 void MarkAllocatedRegs(bool isOutReg, bool isInReg,
5525 std::set<unsigned> &OutputRegs,
5526 std::set<unsigned> &InputRegs,
5527 const TargetRegisterInfo &TRI) const {
5529 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
5530 MarkRegAndAliases(AssignedRegs.Regs[i], OutputRegs, TRI);
5533 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
5534 MarkRegAndAliases(AssignedRegs.Regs[i], InputRegs, TRI);
5538 /// getCallOperandValEVT - Return the EVT of the Value* that this operand
5539 /// corresponds to. If there is no Value* for this operand, it returns
5541 EVT getCallOperandValEVT(LLVMContext &Context,
5542 const TargetLowering &TLI,
5543 const TargetData *TD) const {
5544 if (CallOperandVal == 0) return MVT::Other;
5546 if (isa<BasicBlock>(CallOperandVal))
5547 return TLI.getPointerTy();
5549 llvm::Type *OpTy = CallOperandVal->getType();
5551 // FIXME: code duplicated from TargetLowering::ParseConstraints().
5552 // If this is an indirect operand, the operand is a pointer to the
5555 llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
5557 report_fatal_error("Indirect operand for inline asm not a pointer!");
5558 OpTy = PtrTy->getElementType();
5561 // Look for vector wrapped in a struct. e.g. { <16 x i8> }.
5562 if (StructType *STy = dyn_cast<StructType>(OpTy))
5563 if (STy->getNumElements() == 1)
5564 OpTy = STy->getElementType(0);
5566 // If OpTy is not a single value, it may be a struct/union that we
5567 // can tile with integers.
5568 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
5569 unsigned BitSize = TD->getTypeSizeInBits(OpTy);
5578 OpTy = IntegerType::get(Context, BitSize);
5583 return TLI.getValueType(OpTy, true);
5587 /// MarkRegAndAliases - Mark the specified register and all aliases in the
5589 static void MarkRegAndAliases(unsigned Reg, std::set<unsigned> &Regs,
5590 const TargetRegisterInfo &TRI) {
5591 assert(TargetRegisterInfo::isPhysicalRegister(Reg) && "Isn't a physreg");
5593 if (const unsigned *Aliases = TRI.getAliasSet(Reg))
5594 for (; *Aliases; ++Aliases)
5595 Regs.insert(*Aliases);
5599 typedef SmallVector<SDISelAsmOperandInfo,16> SDISelAsmOperandInfoVector;
5601 } // end anonymous namespace
5603 /// GetRegistersForValue - Assign registers (virtual or physical) for the
5604 /// specified operand. We prefer to assign virtual registers, to allow the
5605 /// register allocator to handle the assignment process. However, if the asm
5606 /// uses features that we can't model on machineinstrs, we have SDISel do the
5607 /// allocation. This produces generally horrible, but correct, code.
5609 /// OpInfo describes the operand.
5610 /// Input and OutputRegs are the set of already allocated physical registers.
5612 static void GetRegistersForValue(SelectionDAG &DAG,
5613 const TargetLowering &TLI,
5615 SDISelAsmOperandInfo &OpInfo,
5616 std::set<unsigned> &OutputRegs,
5617 std::set<unsigned> &InputRegs) {
5618 LLVMContext &Context = *DAG.getContext();
5620 // Compute whether this value requires an input register, an output register,
5622 bool isOutReg = false;
5623 bool isInReg = false;
5624 switch (OpInfo.Type) {
5625 case InlineAsm::isOutput:
5628 // If there is an input constraint that matches this, we need to reserve
5629 // the input register so no other inputs allocate to it.
5630 isInReg = OpInfo.hasMatchingInput();
5632 case InlineAsm::isInput:
5636 case InlineAsm::isClobber:
5643 MachineFunction &MF = DAG.getMachineFunction();
5644 SmallVector<unsigned, 4> Regs;
5646 // If this is a constraint for a single physreg, or a constraint for a
5647 // register class, find it.
5648 std::pair<unsigned, const TargetRegisterClass*> PhysReg =
5649 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
5650 OpInfo.ConstraintVT);
5652 unsigned NumRegs = 1;
5653 if (OpInfo.ConstraintVT != MVT::Other) {
5654 // If this is a FP input in an integer register (or visa versa) insert a bit
5655 // cast of the input value. More generally, handle any case where the input
5656 // value disagrees with the register class we plan to stick this in.
5657 if (OpInfo.Type == InlineAsm::isInput &&
5658 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) {
5659 // Try to convert to the first EVT that the reg class contains. If the
5660 // types are identical size, use a bitcast to convert (e.g. two differing
5662 EVT RegVT = *PhysReg.second->vt_begin();
5663 if (RegVT.getSizeInBits() == OpInfo.ConstraintVT.getSizeInBits()) {
5664 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
5665 RegVT, OpInfo.CallOperand);
5666 OpInfo.ConstraintVT = RegVT;
5667 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) {
5668 // If the input is a FP value and we want it in FP registers, do a
5669 // bitcast to the corresponding integer type. This turns an f64 value
5670 // into i64, which can be passed with two i32 values on a 32-bit
5672 RegVT = EVT::getIntegerVT(Context,
5673 OpInfo.ConstraintVT.getSizeInBits());
5674 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
5675 RegVT, OpInfo.CallOperand);
5676 OpInfo.ConstraintVT = RegVT;
5680 NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT);
5684 EVT ValueVT = OpInfo.ConstraintVT;
5686 // If this is a constraint for a specific physical register, like {r17},
5688 if (unsigned AssignedReg = PhysReg.first) {
5689 const TargetRegisterClass *RC = PhysReg.second;
5690 if (OpInfo.ConstraintVT == MVT::Other)
5691 ValueVT = *RC->vt_begin();
5693 // Get the actual register value type. This is important, because the user
5694 // may have asked for (e.g.) the AX register in i32 type. We need to
5695 // remember that AX is actually i16 to get the right extension.
5696 RegVT = *RC->vt_begin();
5698 // This is a explicit reference to a physical register.
5699 Regs.push_back(AssignedReg);
5701 // If this is an expanded reference, add the rest of the regs to Regs.
5703 TargetRegisterClass::iterator I = RC->begin();
5704 for (; *I != AssignedReg; ++I)
5705 assert(I != RC->end() && "Didn't find reg!");
5707 // Already added the first reg.
5709 for (; NumRegs; --NumRegs, ++I) {
5710 assert(I != RC->end() && "Ran out of registers to allocate!");
5715 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5716 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
5717 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
5721 // Otherwise, if this was a reference to an LLVM register class, create vregs
5722 // for this reference.
5723 if (const TargetRegisterClass *RC = PhysReg.second) {
5724 RegVT = *RC->vt_begin();
5725 if (OpInfo.ConstraintVT == MVT::Other)
5728 // Create the appropriate number of virtual registers.
5729 MachineRegisterInfo &RegInfo = MF.getRegInfo();
5730 for (; NumRegs; --NumRegs)
5731 Regs.push_back(RegInfo.createVirtualRegister(RC));
5733 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5737 // Otherwise, we couldn't allocate enough registers for this.
5740 /// visitInlineAsm - Handle a call to an InlineAsm object.
5742 void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) {
5743 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
5745 /// ConstraintOperands - Information about all of the constraints.
5746 SDISelAsmOperandInfoVector ConstraintOperands;
5748 std::set<unsigned> OutputRegs, InputRegs;
5750 TargetLowering::AsmOperandInfoVector
5751 TargetConstraints = TLI.ParseConstraints(CS);
5753 bool hasMemory = false;
5755 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
5756 unsigned ResNo = 0; // ResNo - The result number of the next output.
5757 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
5758 ConstraintOperands.push_back(SDISelAsmOperandInfo(TargetConstraints[i]));
5759 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
5761 EVT OpVT = MVT::Other;
5763 // Compute the value type for each operand.
5764 switch (OpInfo.Type) {
5765 case InlineAsm::isOutput:
5766 // Indirect outputs just consume an argument.
5767 if (OpInfo.isIndirect) {
5768 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5772 // The return value of the call is this value. As such, there is no
5773 // corresponding argument.
5774 assert(!CS.getType()->isVoidTy() &&
5776 if (StructType *STy = dyn_cast<StructType>(CS.getType())) {
5777 OpVT = TLI.getValueType(STy->getElementType(ResNo));
5779 assert(ResNo == 0 && "Asm only has one result!");
5780 OpVT = TLI.getValueType(CS.getType());
5784 case InlineAsm::isInput:
5785 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5787 case InlineAsm::isClobber:
5792 // If this is an input or an indirect output, process the call argument.
5793 // BasicBlocks are labels, currently appearing only in asm's.
5794 if (OpInfo.CallOperandVal) {
5795 if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) {
5796 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]);
5798 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
5801 OpVT = OpInfo.getCallOperandValEVT(*DAG.getContext(), TLI, TD);
5804 OpInfo.ConstraintVT = OpVT;
5806 // Indirect operand accesses access memory.
5807 if (OpInfo.isIndirect)
5810 for (unsigned j = 0, ee = OpInfo.Codes.size(); j != ee; ++j) {
5811 TargetLowering::ConstraintType
5812 CType = TLI.getConstraintType(OpInfo.Codes[j]);
5813 if (CType == TargetLowering::C_Memory) {
5821 SDValue Chain, Flag;
5823 // We won't need to flush pending loads if this asm doesn't touch
5824 // memory and is nonvolatile.
5825 if (hasMemory || IA->hasSideEffects())
5828 Chain = DAG.getRoot();
5830 // Second pass over the constraints: compute which constraint option to use
5831 // and assign registers to constraints that want a specific physreg.
5832 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5833 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5835 // If this is an output operand with a matching input operand, look up the
5836 // matching input. If their types mismatch, e.g. one is an integer, the
5837 // other is floating point, or their sizes are different, flag it as an
5839 if (OpInfo.hasMatchingInput()) {
5840 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
5842 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
5843 std::pair<unsigned, const TargetRegisterClass*> MatchRC =
5844 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode, OpInfo.ConstraintVT);
5845 std::pair<unsigned, const TargetRegisterClass*> InputRC =
5846 TLI.getRegForInlineAsmConstraint(Input.ConstraintCode, Input.ConstraintVT);
5847 if ((OpInfo.ConstraintVT.isInteger() !=
5848 Input.ConstraintVT.isInteger()) ||
5849 (MatchRC.second != InputRC.second)) {
5850 report_fatal_error("Unsupported asm: input constraint"
5851 " with a matching output constraint of"
5852 " incompatible type!");
5854 Input.ConstraintVT = OpInfo.ConstraintVT;
5858 // Compute the constraint code and ConstraintType to use.
5859 TLI.ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, &DAG);
5861 // If this is a memory input, and if the operand is not indirect, do what we
5862 // need to to provide an address for the memory input.
5863 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5864 !OpInfo.isIndirect) {
5865 assert((OpInfo.isMultipleAlternative ||
5866 (OpInfo.Type == InlineAsm::isInput)) &&
5867 "Can only indirectify direct input operands!");
5869 // Memory operands really want the address of the value. If we don't have
5870 // an indirect input, put it in the constpool if we can, otherwise spill
5871 // it to a stack slot.
5872 // TODO: This isn't quite right. We need to handle these according to
5873 // the addressing mode that the constraint wants. Also, this may take
5874 // an additional register for the computation and we don't want that
5877 // If the operand is a float, integer, or vector constant, spill to a
5878 // constant pool entry to get its address.
5879 const Value *OpVal = OpInfo.CallOperandVal;
5880 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
5881 isa<ConstantVector>(OpVal)) {
5882 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal),
5883 TLI.getPointerTy());
5885 // Otherwise, create a stack slot and emit a store to it before the
5887 Type *Ty = OpVal->getType();
5888 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
5889 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(Ty);
5890 MachineFunction &MF = DAG.getMachineFunction();
5891 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
5892 SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
5893 Chain = DAG.getStore(Chain, getCurDebugLoc(),
5894 OpInfo.CallOperand, StackSlot,
5895 MachinePointerInfo::getFixedStack(SSFI),
5897 OpInfo.CallOperand = StackSlot;
5900 // There is no longer a Value* corresponding to this operand.
5901 OpInfo.CallOperandVal = 0;
5903 // It is now an indirect operand.
5904 OpInfo.isIndirect = true;
5907 // If this constraint is for a specific register, allocate it before
5909 if (OpInfo.ConstraintType == TargetLowering::C_Register)
5910 GetRegistersForValue(DAG, TLI, getCurDebugLoc(), OpInfo, OutputRegs,
5914 // Second pass - Loop over all of the operands, assigning virtual or physregs
5915 // to register class operands.
5916 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5917 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5919 // C_Register operands have already been allocated, Other/Memory don't need
5921 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
5922 GetRegistersForValue(DAG, TLI, getCurDebugLoc(), OpInfo, OutputRegs,
5926 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
5927 std::vector<SDValue> AsmNodeOperands;
5928 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain
5929 AsmNodeOperands.push_back(
5930 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(),
5931 TLI.getPointerTy()));
5933 // If we have a !srcloc metadata node associated with it, we want to attach
5934 // this to the ultimately generated inline asm machineinstr. To do this, we
5935 // pass in the third operand as this (potentially null) inline asm MDNode.
5936 const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc");
5937 AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc));
5939 // Remember the HasSideEffect and AlignStack bits as operand 3.
5940 unsigned ExtraInfo = 0;
5941 if (IA->hasSideEffects())
5942 ExtraInfo |= InlineAsm::Extra_HasSideEffects;
5943 if (IA->isAlignStack())
5944 ExtraInfo |= InlineAsm::Extra_IsAlignStack;
5945 AsmNodeOperands.push_back(DAG.getTargetConstant(ExtraInfo,
5946 TLI.getPointerTy()));
5948 // Loop over all of the inputs, copying the operand values into the
5949 // appropriate registers and processing the output regs.
5950 RegsForValue RetValRegs;
5952 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
5953 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
5955 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5956 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5958 switch (OpInfo.Type) {
5959 case InlineAsm::isOutput: {
5960 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
5961 OpInfo.ConstraintType != TargetLowering::C_Register) {
5962 // Memory output, or 'other' output (e.g. 'X' constraint).
5963 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
5965 // Add information to the INLINEASM node to know about this output.
5966 unsigned OpFlags = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
5967 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlags,
5968 TLI.getPointerTy()));
5969 AsmNodeOperands.push_back(OpInfo.CallOperand);
5973 // Otherwise, this is a register or register class output.
5975 // Copy the output from the appropriate register. Find a register that
5977 if (OpInfo.AssignedRegs.Regs.empty())
5978 report_fatal_error("Couldn't allocate output reg for constraint '" +
5979 Twine(OpInfo.ConstraintCode) + "'!");
5981 // If this is an indirect operand, store through the pointer after the
5983 if (OpInfo.isIndirect) {
5984 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
5985 OpInfo.CallOperandVal));
5987 // This is the result value of the call.
5988 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
5989 // Concatenate this output onto the outputs list.
5990 RetValRegs.append(OpInfo.AssignedRegs);
5993 // Add information to the INLINEASM node to know that this register is
5995 OpInfo.AssignedRegs.AddInlineAsmOperands(OpInfo.isEarlyClobber ?
5996 InlineAsm::Kind_RegDefEarlyClobber :
5997 InlineAsm::Kind_RegDef,
6004 case InlineAsm::isInput: {
6005 SDValue InOperandVal = OpInfo.CallOperand;
6007 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint?
6008 // If this is required to match an output register we have already set,
6009 // just use its register.
6010 unsigned OperandNo = OpInfo.getMatchedOperand();
6012 // Scan until we find the definition we already emitted of this operand.
6013 // When we find it, create a RegsForValue operand.
6014 unsigned CurOp = InlineAsm::Op_FirstOperand;
6015 for (; OperandNo; --OperandNo) {
6016 // Advance to the next operand.
6018 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6019 assert((InlineAsm::isRegDefKind(OpFlag) ||
6020 InlineAsm::isRegDefEarlyClobberKind(OpFlag) ||
6021 InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?");
6022 CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1;
6026 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6027 if (InlineAsm::isRegDefKind(OpFlag) ||
6028 InlineAsm::isRegDefEarlyClobberKind(OpFlag)) {
6029 // Add (OpFlag&0xffff)>>3 registers to MatchedRegs.
6030 if (OpInfo.isIndirect) {
6031 // This happens on gcc/testsuite/gcc.dg/pr8788-1.c
6032 LLVMContext &Ctx = *DAG.getContext();
6033 Ctx.emitError(CS.getInstruction(), "inline asm not supported yet:"
6034 " don't know how to handle tied "
6035 "indirect register inputs");
6038 RegsForValue MatchedRegs;
6039 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
6040 EVT RegVT = AsmNodeOperands[CurOp+1].getValueType();
6041 MatchedRegs.RegVTs.push_back(RegVT);
6042 MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo();
6043 for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag);
6045 MatchedRegs.Regs.push_back
6046 (RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT)));
6048 // Use the produced MatchedRegs object to
6049 MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
6051 MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse,
6052 true, OpInfo.getMatchedOperand(),
6053 DAG, AsmNodeOperands);
6057 assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!");
6058 assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 &&
6059 "Unexpected number of operands");
6060 // Add information to the INLINEASM node to know about this input.
6061 // See InlineAsm.h isUseOperandTiedToDef.
6062 OpFlag = InlineAsm::getFlagWordForMatchingOp(OpFlag,
6063 OpInfo.getMatchedOperand());
6064 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlag,
6065 TLI.getPointerTy()));
6066 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]);
6070 // Treat indirect 'X' constraint as memory.
6071 if (OpInfo.ConstraintType == TargetLowering::C_Other &&
6073 OpInfo.ConstraintType = TargetLowering::C_Memory;
6075 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
6076 std::vector<SDValue> Ops;
6077 TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode,
6080 report_fatal_error("Invalid operand for inline asm constraint '" +
6081 Twine(OpInfo.ConstraintCode) + "'!");
6083 // Add information to the INLINEASM node to know about this input.
6084 unsigned ResOpType =
6085 InlineAsm::getFlagWord(InlineAsm::Kind_Imm, Ops.size());
6086 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
6087 TLI.getPointerTy()));
6088 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
6092 if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
6093 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
6094 assert(InOperandVal.getValueType() == TLI.getPointerTy() &&
6095 "Memory operands expect pointer values");
6097 // Add information to the INLINEASM node to know about this input.
6098 unsigned ResOpType = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
6099 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
6100 TLI.getPointerTy()));
6101 AsmNodeOperands.push_back(InOperandVal);
6105 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
6106 OpInfo.ConstraintType == TargetLowering::C_Register) &&
6107 "Unknown constraint type!");
6108 assert(!OpInfo.isIndirect &&
6109 "Don't know how to handle indirect register inputs yet!");
6111 // Copy the input into the appropriate registers.
6112 if (OpInfo.AssignedRegs.Regs.empty())
6113 report_fatal_error("Couldn't allocate input reg for constraint '" +
6114 Twine(OpInfo.ConstraintCode) + "'!");
6116 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
6119 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, false, 0,
6120 DAG, AsmNodeOperands);
6123 case InlineAsm::isClobber: {
6124 // Add the clobbered value to the operand list, so that the register
6125 // allocator is aware that the physreg got clobbered.
6126 if (!OpInfo.AssignedRegs.Regs.empty())
6127 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_Clobber,
6135 // Finish up input operands. Set the input chain and add the flag last.
6136 AsmNodeOperands[InlineAsm::Op_InputChain] = Chain;
6137 if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
6139 Chain = DAG.getNode(ISD::INLINEASM, getCurDebugLoc(),
6140 DAG.getVTList(MVT::Other, MVT::Glue),
6141 &AsmNodeOperands[0], AsmNodeOperands.size());
6142 Flag = Chain.getValue(1);
6144 // If this asm returns a register value, copy the result from that register
6145 // and set it as the value of the call.
6146 if (!RetValRegs.Regs.empty()) {
6147 SDValue Val = RetValRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(),
6150 // FIXME: Why don't we do this for inline asms with MRVs?
6151 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) {
6152 EVT ResultType = TLI.getValueType(CS.getType());
6154 // If any of the results of the inline asm is a vector, it may have the
6155 // wrong width/num elts. This can happen for register classes that can
6156 // contain multiple different value types. The preg or vreg allocated may
6157 // not have the same VT as was expected. Convert it to the right type
6158 // with bit_convert.
6159 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) {
6160 Val = DAG.getNode(ISD::BITCAST, getCurDebugLoc(),
6163 } else if (ResultType != Val.getValueType() &&
6164 ResultType.isInteger() && Val.getValueType().isInteger()) {
6165 // If a result value was tied to an input value, the computed result may
6166 // have a wider width than the expected result. Extract the relevant
6168 Val = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), ResultType, Val);
6171 assert(ResultType == Val.getValueType() && "Asm result value mismatch!");
6174 setValue(CS.getInstruction(), Val);
6175 // Don't need to use this as a chain in this case.
6176 if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty())
6180 std::vector<std::pair<SDValue, const Value *> > StoresToEmit;
6182 // Process indirect outputs, first output all of the flagged copies out of
6184 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
6185 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
6186 const Value *Ptr = IndirectStoresToEmit[i].second;
6187 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(),
6189 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
6192 // Emit the non-flagged stores from the physregs.
6193 SmallVector<SDValue, 8> OutChains;
6194 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) {
6195 SDValue Val = DAG.getStore(Chain, getCurDebugLoc(),
6196 StoresToEmit[i].first,
6197 getValue(StoresToEmit[i].second),
6198 MachinePointerInfo(StoresToEmit[i].second),
6200 OutChains.push_back(Val);
6203 if (!OutChains.empty())
6204 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
6205 &OutChains[0], OutChains.size());
6210 void SelectionDAGBuilder::visitVAStart(const CallInst &I) {
6211 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurDebugLoc(),
6212 MVT::Other, getRoot(),
6213 getValue(I.getArgOperand(0)),
6214 DAG.getSrcValue(I.getArgOperand(0))));
6217 void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) {
6218 const TargetData &TD = *TLI.getTargetData();
6219 SDValue V = DAG.getVAArg(TLI.getValueType(I.getType()), getCurDebugLoc(),
6220 getRoot(), getValue(I.getOperand(0)),
6221 DAG.getSrcValue(I.getOperand(0)),
6222 TD.getABITypeAlignment(I.getType()));
6224 DAG.setRoot(V.getValue(1));
6227 void SelectionDAGBuilder::visitVAEnd(const CallInst &I) {
6228 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurDebugLoc(),
6229 MVT::Other, getRoot(),
6230 getValue(I.getArgOperand(0)),
6231 DAG.getSrcValue(I.getArgOperand(0))));
6234 void SelectionDAGBuilder::visitVACopy(const CallInst &I) {
6235 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurDebugLoc(),
6236 MVT::Other, getRoot(),
6237 getValue(I.getArgOperand(0)),
6238 getValue(I.getArgOperand(1)),
6239 DAG.getSrcValue(I.getArgOperand(0)),
6240 DAG.getSrcValue(I.getArgOperand(1))));
6243 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
6244 /// implementation, which just calls LowerCall.
6245 /// FIXME: When all targets are
6246 /// migrated to using LowerCall, this hook should be integrated into SDISel.
6247 std::pair<SDValue, SDValue>
6248 TargetLowering::LowerCallTo(SDValue Chain, Type *RetTy,
6249 bool RetSExt, bool RetZExt, bool isVarArg,
6250 bool isInreg, unsigned NumFixedArgs,
6251 CallingConv::ID CallConv, bool isTailCall,
6252 bool isReturnValueUsed,
6254 ArgListTy &Args, SelectionDAG &DAG,
6255 DebugLoc dl) const {
6256 // Handle all of the outgoing arguments.
6257 SmallVector<ISD::OutputArg, 32> Outs;
6258 SmallVector<SDValue, 32> OutVals;
6259 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
6260 SmallVector<EVT, 4> ValueVTs;
6261 ComputeValueVTs(*this, Args[i].Ty, ValueVTs);
6262 for (unsigned Value = 0, NumValues = ValueVTs.size();
6263 Value != NumValues; ++Value) {
6264 EVT VT = ValueVTs[Value];
6265 Type *ArgTy = VT.getTypeForEVT(RetTy->getContext());
6266 SDValue Op = SDValue(Args[i].Node.getNode(),
6267 Args[i].Node.getResNo() + Value);
6268 ISD::ArgFlagsTy Flags;
6269 unsigned OriginalAlignment =
6270 getTargetData()->getABITypeAlignment(ArgTy);
6276 if (Args[i].isInReg)
6280 if (Args[i].isByVal) {
6282 PointerType *Ty = cast<PointerType>(Args[i].Ty);
6283 Type *ElementTy = Ty->getElementType();
6284 Flags.setByValSize(getTargetData()->getTypeAllocSize(ElementTy));
6285 // For ByVal, alignment should come from FE. BE will guess if this
6286 // info is not there but there are cases it cannot get right.
6287 unsigned FrameAlign;
6288 if (Args[i].Alignment)
6289 FrameAlign = Args[i].Alignment;
6291 FrameAlign = getByValTypeAlignment(ElementTy);
6292 Flags.setByValAlign(FrameAlign);
6296 Flags.setOrigAlign(OriginalAlignment);
6298 EVT PartVT = getRegisterType(RetTy->getContext(), VT);
6299 unsigned NumParts = getNumRegisters(RetTy->getContext(), VT);
6300 SmallVector<SDValue, 4> Parts(NumParts);
6301 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
6304 ExtendKind = ISD::SIGN_EXTEND;
6305 else if (Args[i].isZExt)
6306 ExtendKind = ISD::ZERO_EXTEND;
6308 getCopyToParts(DAG, dl, Op, &Parts[0], NumParts,
6309 PartVT, ExtendKind);
6311 for (unsigned j = 0; j != NumParts; ++j) {
6312 // if it isn't first piece, alignment must be 1
6313 ISD::OutputArg MyFlags(Flags, Parts[j].getValueType(),
6315 if (NumParts > 1 && j == 0)
6316 MyFlags.Flags.setSplit();
6318 MyFlags.Flags.setOrigAlign(1);
6320 Outs.push_back(MyFlags);
6321 OutVals.push_back(Parts[j]);
6326 // Handle the incoming return values from the call.
6327 SmallVector<ISD::InputArg, 32> Ins;
6328 SmallVector<EVT, 4> RetTys;
6329 ComputeValueVTs(*this, RetTy, RetTys);
6330 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
6332 EVT RegisterVT = getRegisterType(RetTy->getContext(), VT);
6333 unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT);
6334 for (unsigned i = 0; i != NumRegs; ++i) {
6335 ISD::InputArg MyFlags;
6336 MyFlags.VT = RegisterVT.getSimpleVT();
6337 MyFlags.Used = isReturnValueUsed;
6339 MyFlags.Flags.setSExt();
6341 MyFlags.Flags.setZExt();
6343 MyFlags.Flags.setInReg();
6344 Ins.push_back(MyFlags);
6348 SmallVector<SDValue, 4> InVals;
6349 Chain = LowerCall(Chain, Callee, CallConv, isVarArg, isTailCall,
6350 Outs, OutVals, Ins, dl, DAG, InVals);
6352 // Verify that the target's LowerCall behaved as expected.
6353 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
6354 "LowerCall didn't return a valid chain!");
6355 assert((!isTailCall || InVals.empty()) &&
6356 "LowerCall emitted a return value for a tail call!");
6357 assert((isTailCall || InVals.size() == Ins.size()) &&
6358 "LowerCall didn't emit the correct number of values!");
6360 // For a tail call, the return value is merely live-out and there aren't
6361 // any nodes in the DAG representing it. Return a special value to
6362 // indicate that a tail call has been emitted and no more Instructions
6363 // should be processed in the current block.
6366 return std::make_pair(SDValue(), SDValue());
6369 DEBUG(for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
6370 assert(InVals[i].getNode() &&
6371 "LowerCall emitted a null value!");
6372 assert(EVT(Ins[i].VT) == InVals[i].getValueType() &&
6373 "LowerCall emitted a value with the wrong type!");
6376 // Collect the legal value parts into potentially illegal values
6377 // that correspond to the original function's return values.
6378 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6380 AssertOp = ISD::AssertSext;
6382 AssertOp = ISD::AssertZext;
6383 SmallVector<SDValue, 4> ReturnValues;
6384 unsigned CurReg = 0;
6385 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
6387 EVT RegisterVT = getRegisterType(RetTy->getContext(), VT);
6388 unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT);
6390 ReturnValues.push_back(getCopyFromParts(DAG, dl, &InVals[CurReg],
6391 NumRegs, RegisterVT, VT,
6396 // For a function returning void, there is no return value. We can't create
6397 // such a node, so we just return a null return value in that case. In
6398 // that case, nothing will actually look at the value.
6399 if (ReturnValues.empty())
6400 return std::make_pair(SDValue(), Chain);
6402 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, dl,
6403 DAG.getVTList(&RetTys[0], RetTys.size()),
6404 &ReturnValues[0], ReturnValues.size());
6405 return std::make_pair(Res, Chain);
6408 void TargetLowering::LowerOperationWrapper(SDNode *N,
6409 SmallVectorImpl<SDValue> &Results,
6410 SelectionDAG &DAG) const {
6411 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
6413 Results.push_back(Res);
6416 SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6417 llvm_unreachable("LowerOperation not implemented for this target!");
6422 SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) {
6423 SDValue Op = getNonRegisterValue(V);
6424 assert((Op.getOpcode() != ISD::CopyFromReg ||
6425 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
6426 "Copy from a reg to the same reg!");
6427 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg");
6429 RegsForValue RFV(V->getContext(), TLI, Reg, V->getType());
6430 SDValue Chain = DAG.getEntryNode();
6431 RFV.getCopyToRegs(Op, DAG, getCurDebugLoc(), Chain, 0);
6432 PendingExports.push_back(Chain);
6435 #include "llvm/CodeGen/SelectionDAGISel.h"
6437 /// isOnlyUsedInEntryBlock - If the specified argument is only used in the
6438 /// entry block, return true. This includes arguments used by switches, since
6439 /// the switch may expand into multiple basic blocks.
6440 static bool isOnlyUsedInEntryBlock(const Argument *A) {
6441 // With FastISel active, we may be splitting blocks, so force creation
6442 // of virtual registers for all non-dead arguments.
6444 return A->use_empty();
6446 const BasicBlock *Entry = A->getParent()->begin();
6447 for (Value::const_use_iterator UI = A->use_begin(), E = A->use_end();
6449 const User *U = *UI;
6450 if (cast<Instruction>(U)->getParent() != Entry || isa<SwitchInst>(U))
6451 return false; // Use not in entry block.
6456 void SelectionDAGISel::LowerArguments(const BasicBlock *LLVMBB) {
6457 // If this is the entry block, emit arguments.
6458 const Function &F = *LLVMBB->getParent();
6459 SelectionDAG &DAG = SDB->DAG;
6460 DebugLoc dl = SDB->getCurDebugLoc();
6461 const TargetData *TD = TLI.getTargetData();
6462 SmallVector<ISD::InputArg, 16> Ins;
6464 // Check whether the function can return without sret-demotion.
6465 SmallVector<ISD::OutputArg, 4> Outs;
6466 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
6469 if (!FuncInfo->CanLowerReturn) {
6470 // Put in an sret pointer parameter before all the other parameters.
6471 SmallVector<EVT, 1> ValueVTs;
6472 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
6474 // NOTE: Assuming that a pointer will never break down to more than one VT
6476 ISD::ArgFlagsTy Flags;
6478 EVT RegisterVT = TLI.getRegisterType(*DAG.getContext(), ValueVTs[0]);
6479 ISD::InputArg RetArg(Flags, RegisterVT, true);
6480 Ins.push_back(RetArg);
6483 // Set up the incoming argument description vector.
6485 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end();
6486 I != E; ++I, ++Idx) {
6487 SmallVector<EVT, 4> ValueVTs;
6488 ComputeValueVTs(TLI, I->getType(), ValueVTs);
6489 bool isArgValueUsed = !I->use_empty();
6490 for (unsigned Value = 0, NumValues = ValueVTs.size();
6491 Value != NumValues; ++Value) {
6492 EVT VT = ValueVTs[Value];
6493 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
6494 ISD::ArgFlagsTy Flags;
6495 unsigned OriginalAlignment =
6496 TD->getABITypeAlignment(ArgTy);
6498 if (F.paramHasAttr(Idx, Attribute::ZExt))
6500 if (F.paramHasAttr(Idx, Attribute::SExt))
6502 if (F.paramHasAttr(Idx, Attribute::InReg))
6504 if (F.paramHasAttr(Idx, Attribute::StructRet))
6506 if (F.paramHasAttr(Idx, Attribute::ByVal)) {
6508 PointerType *Ty = cast<PointerType>(I->getType());
6509 Type *ElementTy = Ty->getElementType();
6510 Flags.setByValSize(TD->getTypeAllocSize(ElementTy));
6511 // For ByVal, alignment should be passed from FE. BE will guess if
6512 // this info is not there but there are cases it cannot get right.
6513 unsigned FrameAlign;
6514 if (F.getParamAlignment(Idx))
6515 FrameAlign = F.getParamAlignment(Idx);
6517 FrameAlign = TLI.getByValTypeAlignment(ElementTy);
6518 Flags.setByValAlign(FrameAlign);
6520 if (F.paramHasAttr(Idx, Attribute::Nest))
6522 Flags.setOrigAlign(OriginalAlignment);
6524 EVT RegisterVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6525 unsigned NumRegs = TLI.getNumRegisters(*CurDAG->getContext(), VT);
6526 for (unsigned i = 0; i != NumRegs; ++i) {
6527 ISD::InputArg MyFlags(Flags, RegisterVT, isArgValueUsed);
6528 if (NumRegs > 1 && i == 0)
6529 MyFlags.Flags.setSplit();
6530 // if it isn't first piece, alignment must be 1
6532 MyFlags.Flags.setOrigAlign(1);
6533 Ins.push_back(MyFlags);
6538 // Call the target to set up the argument values.
6539 SmallVector<SDValue, 8> InVals;
6540 SDValue NewRoot = TLI.LowerFormalArguments(DAG.getRoot(), F.getCallingConv(),
6544 // Verify that the target's LowerFormalArguments behaved as expected.
6545 assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other &&
6546 "LowerFormalArguments didn't return a valid chain!");
6547 assert(InVals.size() == Ins.size() &&
6548 "LowerFormalArguments didn't emit the correct number of values!");
6550 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
6551 assert(InVals[i].getNode() &&
6552 "LowerFormalArguments emitted a null value!");
6553 assert(EVT(Ins[i].VT) == InVals[i].getValueType() &&
6554 "LowerFormalArguments emitted a value with the wrong type!");
6558 // Update the DAG with the new chain value resulting from argument lowering.
6559 DAG.setRoot(NewRoot);
6561 // Set up the argument values.
6564 if (!FuncInfo->CanLowerReturn) {
6565 // Create a virtual register for the sret pointer, and put in a copy
6566 // from the sret argument into it.
6567 SmallVector<EVT, 1> ValueVTs;
6568 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
6569 EVT VT = ValueVTs[0];
6570 EVT RegVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6571 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6572 SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1,
6573 RegVT, VT, AssertOp);
6575 MachineFunction& MF = SDB->DAG.getMachineFunction();
6576 MachineRegisterInfo& RegInfo = MF.getRegInfo();
6577 unsigned SRetReg = RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT));
6578 FuncInfo->DemoteRegister = SRetReg;
6579 NewRoot = SDB->DAG.getCopyToReg(NewRoot, SDB->getCurDebugLoc(),
6581 DAG.setRoot(NewRoot);
6583 // i indexes lowered arguments. Bump it past the hidden sret argument.
6584 // Idx indexes LLVM arguments. Don't touch it.
6588 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
6590 SmallVector<SDValue, 4> ArgValues;
6591 SmallVector<EVT, 4> ValueVTs;
6592 ComputeValueVTs(TLI, I->getType(), ValueVTs);
6593 unsigned NumValues = ValueVTs.size();
6595 // If this argument is unused then remember its value. It is used to generate
6596 // debugging information.
6597 if (I->use_empty() && NumValues)
6598 SDB->setUnusedArgValue(I, InVals[i]);
6600 for (unsigned Val = 0; Val != NumValues; ++Val) {
6601 EVT VT = ValueVTs[Val];
6602 EVT PartVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6603 unsigned NumParts = TLI.getNumRegisters(*CurDAG->getContext(), VT);
6605 if (!I->use_empty()) {
6606 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6607 if (F.paramHasAttr(Idx, Attribute::SExt))
6608 AssertOp = ISD::AssertSext;
6609 else if (F.paramHasAttr(Idx, Attribute::ZExt))
6610 AssertOp = ISD::AssertZext;
6612 ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i],
6613 NumParts, PartVT, VT,
6620 // We don't need to do anything else for unused arguments.
6621 if (ArgValues.empty())
6624 // Note down frame index for byval arguments.
6625 if (I->hasByValAttr())
6626 if (FrameIndexSDNode *FI =
6627 dyn_cast<FrameIndexSDNode>(ArgValues[0].getNode()))
6628 FuncInfo->setByValArgumentFrameIndex(I, FI->getIndex());
6630 SDValue Res = DAG.getMergeValues(&ArgValues[0], NumValues,
6631 SDB->getCurDebugLoc());
6632 SDB->setValue(I, Res);
6634 // If this argument is live outside of the entry block, insert a copy from
6635 // wherever we got it to the vreg that other BB's will reference it as.
6636 if (!EnableFastISel && Res.getOpcode() == ISD::CopyFromReg) {
6637 // If we can, though, try to skip creating an unnecessary vreg.
6638 // FIXME: This isn't very clean... it would be nice to make this more
6639 // general. It's also subtly incompatible with the hacks FastISel
6641 unsigned Reg = cast<RegisterSDNode>(Res.getOperand(1))->getReg();
6642 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
6643 FuncInfo->ValueMap[I] = Reg;
6647 if (!isOnlyUsedInEntryBlock(I)) {
6648 FuncInfo->InitializeRegForValue(I);
6649 SDB->CopyToExportRegsIfNeeded(I);
6653 assert(i == InVals.size() && "Argument register count mismatch!");
6655 // Finally, if the target has anything special to do, allow it to do so.
6656 // FIXME: this should insert code into the DAG!
6657 EmitFunctionEntryCode();
6660 /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
6661 /// ensure constants are generated when needed. Remember the virtual registers
6662 /// that need to be added to the Machine PHI nodes as input. We cannot just
6663 /// directly add them, because expansion might result in multiple MBB's for one
6664 /// BB. As such, the start of the BB might correspond to a different MBB than
6668 SelectionDAGBuilder::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
6669 const TerminatorInst *TI = LLVMBB->getTerminator();
6671 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
6673 // Check successor nodes' PHI nodes that expect a constant to be available
6675 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
6676 const BasicBlock *SuccBB = TI->getSuccessor(succ);
6677 if (!isa<PHINode>(SuccBB->begin())) continue;
6678 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
6680 // If this terminator has multiple identical successors (common for
6681 // switches), only handle each succ once.
6682 if (!SuccsHandled.insert(SuccMBB)) continue;
6684 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
6686 // At this point we know that there is a 1-1 correspondence between LLVM PHI
6687 // nodes and Machine PHI nodes, but the incoming operands have not been
6689 for (BasicBlock::const_iterator I = SuccBB->begin();
6690 const PHINode *PN = dyn_cast<PHINode>(I); ++I) {
6691 // Ignore dead phi's.
6692 if (PN->use_empty()) continue;
6695 if (PN->getType()->isEmptyTy())
6699 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
6701 if (const Constant *C = dyn_cast<Constant>(PHIOp)) {
6702 unsigned &RegOut = ConstantsOut[C];
6704 RegOut = FuncInfo.CreateRegs(C->getType());
6705 CopyValueToVirtualRegister(C, RegOut);
6709 DenseMap<const Value *, unsigned>::iterator I =
6710 FuncInfo.ValueMap.find(PHIOp);
6711 if (I != FuncInfo.ValueMap.end())
6714 assert(isa<AllocaInst>(PHIOp) &&
6715 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
6716 "Didn't codegen value into a register!??");
6717 Reg = FuncInfo.CreateRegs(PHIOp->getType());
6718 CopyValueToVirtualRegister(PHIOp, Reg);
6722 // Remember that this register needs to added to the machine PHI node as
6723 // the input for this MBB.
6724 SmallVector<EVT, 4> ValueVTs;
6725 ComputeValueVTs(TLI, PN->getType(), ValueVTs);
6726 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
6727 EVT VT = ValueVTs[vti];
6728 unsigned NumRegisters = TLI.getNumRegisters(*DAG.getContext(), VT);
6729 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
6730 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
6731 Reg += NumRegisters;
6735 ConstantsOut.clear();