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/DebugInfo.h"
25 #include "llvm/DerivedTypes.h"
26 #include "llvm/Function.h"
27 #include "llvm/GlobalVariable.h"
28 #include "llvm/InlineAsm.h"
29 #include "llvm/Instructions.h"
30 #include "llvm/Intrinsics.h"
31 #include "llvm/IntrinsicInst.h"
32 #include "llvm/LLVMContext.h"
33 #include "llvm/Module.h"
34 #include "llvm/CodeGen/Analysis.h"
35 #include "llvm/CodeGen/FastISel.h"
36 #include "llvm/CodeGen/FunctionLoweringInfo.h"
37 #include "llvm/CodeGen/GCStrategy.h"
38 #include "llvm/CodeGen/GCMetadata.h"
39 #include "llvm/CodeGen/MachineFunction.h"
40 #include "llvm/CodeGen/MachineFrameInfo.h"
41 #include "llvm/CodeGen/MachineInstrBuilder.h"
42 #include "llvm/CodeGen/MachineJumpTableInfo.h"
43 #include "llvm/CodeGen/MachineModuleInfo.h"
44 #include "llvm/CodeGen/MachineRegisterInfo.h"
45 #include "llvm/CodeGen/SelectionDAG.h"
46 #include "llvm/Target/TargetData.h"
47 #include "llvm/Target/TargetFrameLowering.h"
48 #include "llvm/Target/TargetInstrInfo.h"
49 #include "llvm/Target/TargetIntrinsicInfo.h"
50 #include "llvm/Target/TargetLibraryInfo.h"
51 #include "llvm/Target/TargetLowering.h"
52 #include "llvm/Target/TargetOptions.h"
53 #include "llvm/Support/CommandLine.h"
54 #include "llvm/Support/IntegersSubsetMapping.h"
55 #include "llvm/Support/Debug.h"
56 #include "llvm/Support/ErrorHandling.h"
57 #include "llvm/Support/MathExtras.h"
58 #include "llvm/Support/raw_ostream.h"
62 /// LimitFloatPrecision - Generate low-precision inline sequences for
63 /// some float libcalls (6, 8 or 12 bits).
64 static unsigned LimitFloatPrecision;
66 static cl::opt<unsigned, true>
67 LimitFPPrecision("limit-float-precision",
68 cl::desc("Generate low-precision inline sequences "
69 "for some float libcalls"),
70 cl::location(LimitFloatPrecision),
73 // Limit the width of DAG chains. This is important in general to prevent
74 // prevent DAG-based analysis from blowing up. For example, alias analysis and
75 // load clustering may not complete in reasonable time. It is difficult to
76 // recognize and avoid this situation within each individual analysis, and
77 // future analyses are likely to have the same behavior. Limiting DAG width is
78 // the safe approach, and will be especially important with global DAGs.
80 // MaxParallelChains default is arbitrarily high to avoid affecting
81 // optimization, but could be lowered to improve compile time. Any ld-ld-st-st
82 // sequence over this should have been converted to llvm.memcpy by the
83 // frontend. It easy to induce this behavior with .ll code such as:
84 // %buffer = alloca [4096 x i8]
85 // %data = load [4096 x i8]* %argPtr
86 // store [4096 x i8] %data, [4096 x i8]* %buffer
87 static const unsigned MaxParallelChains = 64;
89 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, DebugLoc DL,
90 const SDValue *Parts, unsigned NumParts,
91 EVT PartVT, EVT ValueVT);
93 /// getCopyFromParts - Create a value that contains the specified legal parts
94 /// combined into the value they represent. If the parts combine to a type
95 /// larger then ValueVT then AssertOp can be used to specify whether the extra
96 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
97 /// (ISD::AssertSext).
98 static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc DL,
100 unsigned NumParts, EVT PartVT, EVT ValueVT,
101 ISD::NodeType AssertOp = ISD::DELETED_NODE) {
102 if (ValueVT.isVector())
103 return getCopyFromPartsVector(DAG, DL, Parts, NumParts, PartVT, ValueVT);
105 assert(NumParts > 0 && "No parts to assemble!");
106 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
107 SDValue Val = Parts[0];
110 // Assemble the value from multiple parts.
111 if (ValueVT.isInteger()) {
112 unsigned PartBits = PartVT.getSizeInBits();
113 unsigned ValueBits = ValueVT.getSizeInBits();
115 // Assemble the power of 2 part.
116 unsigned RoundParts = NumParts & (NumParts - 1) ?
117 1 << Log2_32(NumParts) : NumParts;
118 unsigned RoundBits = PartBits * RoundParts;
119 EVT RoundVT = RoundBits == ValueBits ?
120 ValueVT : EVT::getIntegerVT(*DAG.getContext(), RoundBits);
123 EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2);
125 if (RoundParts > 2) {
126 Lo = getCopyFromParts(DAG, DL, Parts, RoundParts / 2,
128 Hi = getCopyFromParts(DAG, DL, Parts + RoundParts / 2,
129 RoundParts / 2, PartVT, HalfVT);
131 Lo = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[0]);
132 Hi = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[1]);
135 if (TLI.isBigEndian())
138 Val = DAG.getNode(ISD::BUILD_PAIR, DL, RoundVT, Lo, Hi);
140 if (RoundParts < NumParts) {
141 // Assemble the trailing non-power-of-2 part.
142 unsigned OddParts = NumParts - RoundParts;
143 EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits);
144 Hi = getCopyFromParts(DAG, DL,
145 Parts + RoundParts, OddParts, PartVT, OddVT);
147 // Combine the round and odd parts.
149 if (TLI.isBigEndian())
151 EVT TotalVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
152 Hi = DAG.getNode(ISD::ANY_EXTEND, DL, TotalVT, Hi);
153 Hi = DAG.getNode(ISD::SHL, DL, TotalVT, Hi,
154 DAG.getConstant(Lo.getValueType().getSizeInBits(),
155 TLI.getPointerTy()));
156 Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, TotalVT, Lo);
157 Val = DAG.getNode(ISD::OR, DL, TotalVT, Lo, Hi);
159 } else if (PartVT.isFloatingPoint()) {
160 // FP split into multiple FP parts (for ppcf128)
161 assert(ValueVT == EVT(MVT::ppcf128) && PartVT == EVT(MVT::f64) &&
164 Lo = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[0]);
165 Hi = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[1]);
166 if (TLI.isBigEndian())
168 Val = DAG.getNode(ISD::BUILD_PAIR, DL, ValueVT, Lo, Hi);
170 // FP split into integer parts (soft fp)
171 assert(ValueVT.isFloatingPoint() && PartVT.isInteger() &&
172 !PartVT.isVector() && "Unexpected split");
173 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits());
174 Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT);
178 // There is now one part, held in Val. Correct it to match ValueVT.
179 PartVT = Val.getValueType();
181 if (PartVT == ValueVT)
184 if (PartVT.isInteger() && ValueVT.isInteger()) {
185 if (ValueVT.bitsLT(PartVT)) {
186 // For a truncate, see if we have any information to
187 // indicate whether the truncated bits will always be
188 // zero or sign-extension.
189 if (AssertOp != ISD::DELETED_NODE)
190 Val = DAG.getNode(AssertOp, DL, PartVT, Val,
191 DAG.getValueType(ValueVT));
192 return DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
194 return DAG.getNode(ISD::ANY_EXTEND, DL, ValueVT, Val);
197 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
198 // FP_ROUND's are always exact here.
199 if (ValueVT.bitsLT(Val.getValueType()))
200 return DAG.getNode(ISD::FP_ROUND, DL, ValueVT, Val,
201 DAG.getTargetConstant(1, TLI.getPointerTy()));
203 return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val);
206 if (PartVT.getSizeInBits() == ValueVT.getSizeInBits())
207 return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
209 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() || PartVT == MVT::x86mmx) &&
357 ValueVT.isInteger() &&
358 "Unknown mismatch!");
359 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
360 Val = DAG.getNode(ExtendKind, DL, ValueVT, Val);
361 if (PartVT == MVT::x86mmx)
362 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
364 } else if (PartBits == ValueVT.getSizeInBits()) {
365 // Different types of the same size.
366 assert(NumParts == 1 && PartVT != ValueVT);
367 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
368 } else if (NumParts * PartBits < ValueVT.getSizeInBits()) {
369 // If the parts cover less bits than value has, truncate the value.
370 assert((PartVT.isInteger() || PartVT == MVT::x86mmx) &&
371 ValueVT.isInteger() &&
372 "Unknown mismatch!");
373 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
374 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
375 if (PartVT == MVT::x86mmx)
376 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
379 // The value may have changed - recompute ValueVT.
380 ValueVT = Val.getValueType();
381 assert(NumParts * PartBits == ValueVT.getSizeInBits() &&
382 "Failed to tile the value with PartVT!");
385 assert(PartVT == ValueVT && "Type conversion failed!");
390 // Expand the value into multiple parts.
391 if (NumParts & (NumParts - 1)) {
392 // The number of parts is not a power of 2. Split off and copy the tail.
393 assert(PartVT.isInteger() && ValueVT.isInteger() &&
394 "Do not know what to expand to!");
395 unsigned RoundParts = 1 << Log2_32(NumParts);
396 unsigned RoundBits = RoundParts * PartBits;
397 unsigned OddParts = NumParts - RoundParts;
398 SDValue OddVal = DAG.getNode(ISD::SRL, DL, ValueVT, Val,
399 DAG.getIntPtrConstant(RoundBits));
400 getCopyToParts(DAG, DL, OddVal, Parts + RoundParts, OddParts, PartVT);
402 if (TLI.isBigEndian())
403 // The odd parts were reversed by getCopyToParts - unreverse them.
404 std::reverse(Parts + RoundParts, Parts + NumParts);
406 NumParts = RoundParts;
407 ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
408 Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
411 // The number of parts is a power of 2. Repeatedly bisect the value using
413 Parts[0] = DAG.getNode(ISD::BITCAST, DL,
414 EVT::getIntegerVT(*DAG.getContext(),
415 ValueVT.getSizeInBits()),
418 for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) {
419 for (unsigned i = 0; i < NumParts; i += StepSize) {
420 unsigned ThisBits = StepSize * PartBits / 2;
421 EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits);
422 SDValue &Part0 = Parts[i];
423 SDValue &Part1 = Parts[i+StepSize/2];
425 Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
426 ThisVT, Part0, DAG.getIntPtrConstant(1));
427 Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
428 ThisVT, Part0, DAG.getIntPtrConstant(0));
430 if (ThisBits == PartBits && ThisVT != PartVT) {
431 Part0 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part0);
432 Part1 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part1);
437 if (TLI.isBigEndian())
438 std::reverse(Parts, Parts + OrigNumParts);
442 /// getCopyToPartsVector - Create a series of nodes that contain the specified
443 /// value split into legal parts.
444 static void getCopyToPartsVector(SelectionDAG &DAG, DebugLoc DL,
445 SDValue Val, SDValue *Parts, unsigned NumParts,
447 EVT ValueVT = Val.getValueType();
448 assert(ValueVT.isVector() && "Not a vector");
449 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
452 if (PartVT == ValueVT) {
454 } else if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) {
455 // Bitconvert vector->vector case.
456 Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
457 } else if (PartVT.isVector() &&
458 PartVT.getVectorElementType() == ValueVT.getVectorElementType() &&
459 PartVT.getVectorNumElements() > ValueVT.getVectorNumElements()) {
460 EVT ElementVT = PartVT.getVectorElementType();
461 // Vector widening case, e.g. <2 x float> -> <4 x float>. Shuffle in
463 SmallVector<SDValue, 16> Ops;
464 for (unsigned i = 0, e = ValueVT.getVectorNumElements(); i != e; ++i)
465 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
466 ElementVT, Val, DAG.getIntPtrConstant(i)));
468 for (unsigned i = ValueVT.getVectorNumElements(),
469 e = PartVT.getVectorNumElements(); i != e; ++i)
470 Ops.push_back(DAG.getUNDEF(ElementVT));
472 Val = DAG.getNode(ISD::BUILD_VECTOR, DL, PartVT, &Ops[0], Ops.size());
474 // FIXME: Use CONCAT for 2x -> 4x.
476 //SDValue UndefElts = DAG.getUNDEF(VectorTy);
477 //Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts);
478 } else if (PartVT.isVector() &&
479 PartVT.getVectorElementType().bitsGE(
480 ValueVT.getVectorElementType()) &&
481 PartVT.getVectorNumElements() == ValueVT.getVectorNumElements()) {
483 // Promoted vector extract
484 bool Smaller = PartVT.bitsLE(ValueVT);
485 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
488 // Vector -> scalar conversion.
489 assert(ValueVT.getVectorNumElements() == 1 &&
490 "Only trivial vector-to-scalar conversions should get here!");
491 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
492 PartVT, Val, DAG.getIntPtrConstant(0));
494 bool Smaller = ValueVT.bitsLE(PartVT);
495 Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
503 // Handle a multi-element vector.
504 EVT IntermediateVT, RegisterVT;
505 unsigned NumIntermediates;
506 unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT,
508 NumIntermediates, RegisterVT);
509 unsigned NumElements = ValueVT.getVectorNumElements();
511 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
512 NumParts = NumRegs; // Silence a compiler warning.
513 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
515 // Split the vector into intermediate operands.
516 SmallVector<SDValue, 8> Ops(NumIntermediates);
517 for (unsigned i = 0; i != NumIntermediates; ++i) {
518 if (IntermediateVT.isVector())
519 Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL,
521 DAG.getIntPtrConstant(i * (NumElements / NumIntermediates)));
523 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
524 IntermediateVT, Val, DAG.getIntPtrConstant(i));
527 // Split the intermediate operands into legal parts.
528 if (NumParts == NumIntermediates) {
529 // If the register was not expanded, promote or copy the value,
531 for (unsigned i = 0; i != NumParts; ++i)
532 getCopyToParts(DAG, DL, Ops[i], &Parts[i], 1, PartVT);
533 } else if (NumParts > 0) {
534 // If the intermediate type was expanded, split each the value into
536 assert(NumParts % NumIntermediates == 0 &&
537 "Must expand into a divisible number of parts!");
538 unsigned Factor = NumParts / NumIntermediates;
539 for (unsigned i = 0; i != NumIntermediates; ++i)
540 getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT);
548 /// RegsForValue - This struct represents the registers (physical or virtual)
549 /// that a particular set of values is assigned, and the type information
550 /// about the value. The most common situation is to represent one value at a
551 /// time, but struct or array values are handled element-wise as multiple
552 /// values. The splitting of aggregates is performed recursively, so that we
553 /// never have aggregate-typed registers. The values at this point do not
554 /// necessarily have legal types, so each value may require one or more
555 /// registers of some legal type.
557 struct RegsForValue {
558 /// ValueVTs - The value types of the values, which may not be legal, and
559 /// may need be promoted or synthesized from one or more registers.
561 SmallVector<EVT, 4> ValueVTs;
563 /// RegVTs - The value types of the registers. This is the same size as
564 /// ValueVTs and it records, for each value, what the type of the assigned
565 /// register or registers are. (Individual values are never synthesized
566 /// from more than one type of register.)
568 /// With virtual registers, the contents of RegVTs is redundant with TLI's
569 /// getRegisterType member function, however when with physical registers
570 /// it is necessary to have a separate record of the types.
572 SmallVector<EVT, 4> RegVTs;
574 /// Regs - This list holds the registers assigned to the values.
575 /// Each legal or promoted value requires one register, and each
576 /// expanded value requires multiple registers.
578 SmallVector<unsigned, 4> Regs;
582 RegsForValue(const SmallVector<unsigned, 4> ®s,
583 EVT regvt, EVT valuevt)
584 : ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {}
586 RegsForValue(LLVMContext &Context, const TargetLowering &tli,
587 unsigned Reg, Type *Ty) {
588 ComputeValueVTs(tli, Ty, ValueVTs);
590 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
591 EVT ValueVT = ValueVTs[Value];
592 unsigned NumRegs = tli.getNumRegisters(Context, ValueVT);
593 EVT RegisterVT = tli.getRegisterType(Context, ValueVT);
594 for (unsigned i = 0; i != NumRegs; ++i)
595 Regs.push_back(Reg + i);
596 RegVTs.push_back(RegisterVT);
601 /// areValueTypesLegal - Return true if types of all the values are legal.
602 bool areValueTypesLegal(const TargetLowering &TLI) {
603 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
604 EVT RegisterVT = RegVTs[Value];
605 if (!TLI.isTypeLegal(RegisterVT))
611 /// append - Add the specified values to this one.
612 void append(const RegsForValue &RHS) {
613 ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end());
614 RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end());
615 Regs.append(RHS.Regs.begin(), RHS.Regs.end());
618 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
619 /// this value and returns the result as a ValueVTs value. This uses
620 /// Chain/Flag as the input and updates them for the output Chain/Flag.
621 /// If the Flag pointer is NULL, no flag is used.
622 SDValue getCopyFromRegs(SelectionDAG &DAG, FunctionLoweringInfo &FuncInfo,
624 SDValue &Chain, SDValue *Flag) const;
626 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
627 /// specified value into the registers specified by this object. This uses
628 /// Chain/Flag as the input and updates them for the output Chain/Flag.
629 /// If the Flag pointer is NULL, no flag is used.
630 void getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
631 SDValue &Chain, SDValue *Flag) const;
633 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
634 /// operand list. This adds the code marker, matching input operand index
635 /// (if applicable), and includes the number of values added into it.
636 void AddInlineAsmOperands(unsigned Kind,
637 bool HasMatching, unsigned MatchingIdx,
639 std::vector<SDValue> &Ops) const;
643 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
644 /// this value and returns the result as a ValueVT value. This uses
645 /// Chain/Flag as the input and updates them for the output Chain/Flag.
646 /// If the Flag pointer is NULL, no flag is used.
647 SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
648 FunctionLoweringInfo &FuncInfo,
650 SDValue &Chain, SDValue *Flag) const {
651 // A Value with type {} or [0 x %t] needs no registers.
652 if (ValueVTs.empty())
655 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
657 // Assemble the legal parts into the final values.
658 SmallVector<SDValue, 4> Values(ValueVTs.size());
659 SmallVector<SDValue, 8> Parts;
660 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
661 // Copy the legal parts from the registers.
662 EVT ValueVT = ValueVTs[Value];
663 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
664 EVT RegisterVT = RegVTs[Value];
666 Parts.resize(NumRegs);
667 for (unsigned i = 0; i != NumRegs; ++i) {
670 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT);
672 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag);
673 *Flag = P.getValue(2);
676 Chain = P.getValue(1);
679 // If the source register was virtual and if we know something about it,
680 // add an assert node.
681 if (!TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) ||
682 !RegisterVT.isInteger() || RegisterVT.isVector())
685 const FunctionLoweringInfo::LiveOutInfo *LOI =
686 FuncInfo.GetLiveOutRegInfo(Regs[Part+i]);
690 unsigned RegSize = RegisterVT.getSizeInBits();
691 unsigned NumSignBits = LOI->NumSignBits;
692 unsigned NumZeroBits = LOI->KnownZero.countLeadingOnes();
694 // FIXME: We capture more information than the dag can represent. For
695 // now, just use the tightest assertzext/assertsext possible.
697 EVT FromVT(MVT::Other);
698 if (NumSignBits == RegSize)
699 isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1
700 else if (NumZeroBits >= RegSize-1)
701 isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1
702 else if (NumSignBits > RegSize-8)
703 isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8
704 else if (NumZeroBits >= RegSize-8)
705 isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8
706 else if (NumSignBits > RegSize-16)
707 isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16
708 else if (NumZeroBits >= RegSize-16)
709 isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16
710 else if (NumSignBits > RegSize-32)
711 isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32
712 else if (NumZeroBits >= RegSize-32)
713 isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32
717 // Add an assertion node.
718 assert(FromVT != MVT::Other);
719 Parts[i] = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl,
720 RegisterVT, P, DAG.getValueType(FromVT));
723 Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(),
724 NumRegs, RegisterVT, ValueVT);
729 return DAG.getNode(ISD::MERGE_VALUES, dl,
730 DAG.getVTList(&ValueVTs[0], ValueVTs.size()),
731 &Values[0], ValueVTs.size());
734 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
735 /// specified value into the registers specified by this object. This uses
736 /// Chain/Flag as the input and updates them for the output Chain/Flag.
737 /// If the Flag pointer is NULL, no flag is used.
738 void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
739 SDValue &Chain, SDValue *Flag) const {
740 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
742 // Get the list of the values's legal parts.
743 unsigned NumRegs = Regs.size();
744 SmallVector<SDValue, 8> Parts(NumRegs);
745 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
746 EVT ValueVT = ValueVTs[Value];
747 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
748 EVT RegisterVT = RegVTs[Value];
750 getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value),
751 &Parts[Part], NumParts, RegisterVT);
755 // Copy the parts into the registers.
756 SmallVector<SDValue, 8> Chains(NumRegs);
757 for (unsigned i = 0; i != NumRegs; ++i) {
760 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]);
762 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag);
763 *Flag = Part.getValue(1);
766 Chains[i] = Part.getValue(0);
769 if (NumRegs == 1 || Flag)
770 // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is
771 // flagged to it. That is the CopyToReg nodes and the user are considered
772 // a single scheduling unit. If we create a TokenFactor and return it as
773 // chain, then the TokenFactor is both a predecessor (operand) of the
774 // user as well as a successor (the TF operands are flagged to the user).
775 // c1, f1 = CopyToReg
776 // c2, f2 = CopyToReg
777 // c3 = TokenFactor c1, c2
780 Chain = Chains[NumRegs-1];
782 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0], NumRegs);
785 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
786 /// operand list. This adds the code marker and includes the number of
787 /// values added into it.
788 void RegsForValue::AddInlineAsmOperands(unsigned Code, bool HasMatching,
789 unsigned MatchingIdx,
791 std::vector<SDValue> &Ops) const {
792 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
794 unsigned Flag = InlineAsm::getFlagWord(Code, Regs.size());
796 Flag = InlineAsm::getFlagWordForMatchingOp(Flag, MatchingIdx);
797 else if (!Regs.empty() &&
798 TargetRegisterInfo::isVirtualRegister(Regs.front())) {
799 // Put the register class of the virtual registers in the flag word. That
800 // way, later passes can recompute register class constraints for inline
801 // assembly as well as normal instructions.
802 // Don't do this for tied operands that can use the regclass information
804 const MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
805 const TargetRegisterClass *RC = MRI.getRegClass(Regs.front());
806 Flag = InlineAsm::getFlagWordForRegClass(Flag, RC->getID());
809 SDValue Res = DAG.getTargetConstant(Flag, MVT::i32);
812 for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) {
813 unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVTs[Value]);
814 EVT RegisterVT = RegVTs[Value];
815 for (unsigned i = 0; i != NumRegs; ++i) {
816 assert(Reg < Regs.size() && "Mismatch in # registers expected");
817 Ops.push_back(DAG.getRegister(Regs[Reg++], RegisterVT));
822 void SelectionDAGBuilder::init(GCFunctionInfo *gfi, AliasAnalysis &aa,
823 const TargetLibraryInfo *li) {
827 TD = DAG.getTarget().getTargetData();
828 LPadToCallSiteMap.clear();
831 /// clear - Clear out the current SelectionDAG and the associated
832 /// state and prepare this SelectionDAGBuilder object to be used
833 /// for a new block. This doesn't clear out information about
834 /// additional blocks that are needed to complete switch lowering
835 /// or PHI node updating; that information is cleared out as it is
837 void SelectionDAGBuilder::clear() {
839 UnusedArgNodeMap.clear();
840 PendingLoads.clear();
841 PendingExports.clear();
842 CurDebugLoc = DebugLoc();
846 /// clearDanglingDebugInfo - Clear the dangling debug information
847 /// map. This function is separated from the clear so that debug
848 /// information that is dangling in a basic block can be properly
849 /// resolved in a different basic block. This allows the
850 /// SelectionDAG to resolve dangling debug information attached
852 void SelectionDAGBuilder::clearDanglingDebugInfo() {
853 DanglingDebugInfoMap.clear();
856 /// getRoot - Return the current virtual root of the Selection DAG,
857 /// flushing any PendingLoad items. This must be done before emitting
858 /// a store or any other node that may need to be ordered after any
859 /// prior load instructions.
861 SDValue SelectionDAGBuilder::getRoot() {
862 if (PendingLoads.empty())
863 return DAG.getRoot();
865 if (PendingLoads.size() == 1) {
866 SDValue Root = PendingLoads[0];
868 PendingLoads.clear();
872 // Otherwise, we have to make a token factor node.
873 SDValue Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
874 &PendingLoads[0], PendingLoads.size());
875 PendingLoads.clear();
880 /// getControlRoot - Similar to getRoot, but instead of flushing all the
881 /// PendingLoad items, flush all the PendingExports items. It is necessary
882 /// to do this before emitting a terminator instruction.
884 SDValue SelectionDAGBuilder::getControlRoot() {
885 SDValue Root = DAG.getRoot();
887 if (PendingExports.empty())
890 // Turn all of the CopyToReg chains into one factored node.
891 if (Root.getOpcode() != ISD::EntryToken) {
892 unsigned i = 0, e = PendingExports.size();
893 for (; i != e; ++i) {
894 assert(PendingExports[i].getNode()->getNumOperands() > 1);
895 if (PendingExports[i].getNode()->getOperand(0) == Root)
896 break; // Don't add the root if we already indirectly depend on it.
900 PendingExports.push_back(Root);
903 Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
905 PendingExports.size());
906 PendingExports.clear();
911 void SelectionDAGBuilder::AssignOrderingToNode(const SDNode *Node) {
912 if (DAG.GetOrdering(Node) != 0) return; // Already has ordering.
913 DAG.AssignOrdering(Node, SDNodeOrder);
915 for (unsigned I = 0, E = Node->getNumOperands(); I != E; ++I)
916 AssignOrderingToNode(Node->getOperand(I).getNode());
919 void SelectionDAGBuilder::visit(const Instruction &I) {
920 // Set up outgoing PHI node register values before emitting the terminator.
921 if (isa<TerminatorInst>(&I))
922 HandlePHINodesInSuccessorBlocks(I.getParent());
924 CurDebugLoc = I.getDebugLoc();
926 visit(I.getOpcode(), I);
928 if (!isa<TerminatorInst>(&I) && !HasTailCall)
929 CopyToExportRegsIfNeeded(&I);
931 CurDebugLoc = DebugLoc();
934 void SelectionDAGBuilder::visitPHI(const PHINode &) {
935 llvm_unreachable("SelectionDAGBuilder shouldn't visit PHI nodes!");
938 void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) {
939 // Note: this doesn't use InstVisitor, because it has to work with
940 // ConstantExpr's in addition to instructions.
942 default: llvm_unreachable("Unknown instruction type encountered!");
943 // Build the switch statement using the Instruction.def file.
944 #define HANDLE_INST(NUM, OPCODE, CLASS) \
945 case Instruction::OPCODE: visit##OPCODE((CLASS&)I); break;
946 #include "llvm/Instruction.def"
949 // Assign the ordering to the freshly created DAG nodes.
950 if (NodeMap.count(&I)) {
952 AssignOrderingToNode(getValue(&I).getNode());
956 // resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V,
957 // generate the debug data structures now that we've seen its definition.
958 void SelectionDAGBuilder::resolveDanglingDebugInfo(const Value *V,
960 DanglingDebugInfo &DDI = DanglingDebugInfoMap[V];
962 const DbgValueInst *DI = DDI.getDI();
963 DebugLoc dl = DDI.getdl();
964 unsigned DbgSDNodeOrder = DDI.getSDNodeOrder();
965 MDNode *Variable = DI->getVariable();
966 uint64_t Offset = DI->getOffset();
969 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, Val)) {
970 SDV = DAG.getDbgValue(Variable, Val.getNode(),
971 Val.getResNo(), Offset, dl, DbgSDNodeOrder);
972 DAG.AddDbgValue(SDV, Val.getNode(), false);
975 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
976 DanglingDebugInfoMap[V] = DanglingDebugInfo();
980 /// getValue - Return an SDValue for the given Value.
981 SDValue SelectionDAGBuilder::getValue(const Value *V) {
982 // If we already have an SDValue for this value, use it. It's important
983 // to do this first, so that we don't create a CopyFromReg if we already
984 // have a regular SDValue.
985 SDValue &N = NodeMap[V];
986 if (N.getNode()) return N;
988 // If there's a virtual register allocated and initialized for this
990 DenseMap<const Value *, unsigned>::iterator It = FuncInfo.ValueMap.find(V);
991 if (It != FuncInfo.ValueMap.end()) {
992 unsigned InReg = It->second;
993 RegsForValue RFV(*DAG.getContext(), TLI, InReg, V->getType());
994 SDValue Chain = DAG.getEntryNode();
995 N = RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain, NULL);
996 resolveDanglingDebugInfo(V, N);
1000 // Otherwise create a new SDValue and remember it.
1001 SDValue Val = getValueImpl(V);
1003 resolveDanglingDebugInfo(V, Val);
1007 /// getNonRegisterValue - Return an SDValue for the given Value, but
1008 /// don't look in FuncInfo.ValueMap for a virtual register.
1009 SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) {
1010 // If we already have an SDValue for this value, use it.
1011 SDValue &N = NodeMap[V];
1012 if (N.getNode()) return N;
1014 // Otherwise create a new SDValue and remember it.
1015 SDValue Val = getValueImpl(V);
1017 resolveDanglingDebugInfo(V, Val);
1021 /// getValueImpl - Helper function for getValue and getNonRegisterValue.
1022 /// Create an SDValue for the given value.
1023 SDValue SelectionDAGBuilder::getValueImpl(const Value *V) {
1024 if (const Constant *C = dyn_cast<Constant>(V)) {
1025 EVT VT = TLI.getValueType(V->getType(), true);
1027 if (const ConstantInt *CI = dyn_cast<ConstantInt>(C))
1028 return DAG.getConstant(*CI, VT);
1030 if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
1031 return DAG.getGlobalAddress(GV, getCurDebugLoc(), VT);
1033 if (isa<ConstantPointerNull>(C))
1034 return DAG.getConstant(0, TLI.getPointerTy());
1036 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
1037 return DAG.getConstantFP(*CFP, VT);
1039 if (isa<UndefValue>(C) && !V->getType()->isAggregateType())
1040 return DAG.getUNDEF(VT);
1042 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
1043 visit(CE->getOpcode(), *CE);
1044 SDValue N1 = NodeMap[V];
1045 assert(N1.getNode() && "visit didn't populate the NodeMap!");
1049 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) {
1050 SmallVector<SDValue, 4> Constants;
1051 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end();
1053 SDNode *Val = getValue(*OI).getNode();
1054 // If the operand is an empty aggregate, there are no values.
1056 // Add each leaf value from the operand to the Constants list
1057 // to form a flattened list of all the values.
1058 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
1059 Constants.push_back(SDValue(Val, i));
1062 return DAG.getMergeValues(&Constants[0], Constants.size(),
1066 if (const ConstantDataSequential *CDS =
1067 dyn_cast<ConstantDataSequential>(C)) {
1068 SmallVector<SDValue, 4> Ops;
1069 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1070 SDNode *Val = getValue(CDS->getElementAsConstant(i)).getNode();
1071 // Add each leaf value from the operand to the Constants list
1072 // to form a flattened list of all the values.
1073 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
1074 Ops.push_back(SDValue(Val, i));
1077 if (isa<ArrayType>(CDS->getType()))
1078 return DAG.getMergeValues(&Ops[0], Ops.size(), getCurDebugLoc());
1079 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
1080 VT, &Ops[0], Ops.size());
1083 if (C->getType()->isStructTy() || C->getType()->isArrayTy()) {
1084 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
1085 "Unknown struct or array constant!");
1087 SmallVector<EVT, 4> ValueVTs;
1088 ComputeValueVTs(TLI, C->getType(), ValueVTs);
1089 unsigned NumElts = ValueVTs.size();
1091 return SDValue(); // empty struct
1092 SmallVector<SDValue, 4> Constants(NumElts);
1093 for (unsigned i = 0; i != NumElts; ++i) {
1094 EVT EltVT = ValueVTs[i];
1095 if (isa<UndefValue>(C))
1096 Constants[i] = DAG.getUNDEF(EltVT);
1097 else if (EltVT.isFloatingPoint())
1098 Constants[i] = DAG.getConstantFP(0, EltVT);
1100 Constants[i] = DAG.getConstant(0, EltVT);
1103 return DAG.getMergeValues(&Constants[0], NumElts,
1107 if (const BlockAddress *BA = dyn_cast<BlockAddress>(C))
1108 return DAG.getBlockAddress(BA, VT);
1110 VectorType *VecTy = cast<VectorType>(V->getType());
1111 unsigned NumElements = VecTy->getNumElements();
1113 // Now that we know the number and type of the elements, get that number of
1114 // elements into the Ops array based on what kind of constant it is.
1115 SmallVector<SDValue, 16> Ops;
1116 if (const ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
1117 for (unsigned i = 0; i != NumElements; ++i)
1118 Ops.push_back(getValue(CV->getOperand(i)));
1120 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!");
1121 EVT EltVT = TLI.getValueType(VecTy->getElementType());
1124 if (EltVT.isFloatingPoint())
1125 Op = DAG.getConstantFP(0, EltVT);
1127 Op = DAG.getConstant(0, EltVT);
1128 Ops.assign(NumElements, Op);
1131 // Create a BUILD_VECTOR node.
1132 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
1133 VT, &Ops[0], Ops.size());
1136 // If this is a static alloca, generate it as the frameindex instead of
1138 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
1139 DenseMap<const AllocaInst*, int>::iterator SI =
1140 FuncInfo.StaticAllocaMap.find(AI);
1141 if (SI != FuncInfo.StaticAllocaMap.end())
1142 return DAG.getFrameIndex(SI->second, TLI.getPointerTy());
1145 // If this is an instruction which fast-isel has deferred, select it now.
1146 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
1147 unsigned InReg = FuncInfo.InitializeRegForValue(Inst);
1148 RegsForValue RFV(*DAG.getContext(), TLI, InReg, Inst->getType());
1149 SDValue Chain = DAG.getEntryNode();
1150 return RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain, NULL);
1153 llvm_unreachable("Can't get register for value!");
1156 void SelectionDAGBuilder::visitRet(const ReturnInst &I) {
1157 SDValue Chain = getControlRoot();
1158 SmallVector<ISD::OutputArg, 8> Outs;
1159 SmallVector<SDValue, 8> OutVals;
1161 if (!FuncInfo.CanLowerReturn) {
1162 unsigned DemoteReg = FuncInfo.DemoteRegister;
1163 const Function *F = I.getParent()->getParent();
1165 // Emit a store of the return value through the virtual register.
1166 // Leave Outs empty so that LowerReturn won't try to load return
1167 // registers the usual way.
1168 SmallVector<EVT, 1> PtrValueVTs;
1169 ComputeValueVTs(TLI, PointerType::getUnqual(F->getReturnType()),
1172 SDValue RetPtr = DAG.getRegister(DemoteReg, PtrValueVTs[0]);
1173 SDValue RetOp = getValue(I.getOperand(0));
1175 SmallVector<EVT, 4> ValueVTs;
1176 SmallVector<uint64_t, 4> Offsets;
1177 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs, &Offsets);
1178 unsigned NumValues = ValueVTs.size();
1180 SmallVector<SDValue, 4> Chains(NumValues);
1181 for (unsigned i = 0; i != NumValues; ++i) {
1182 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(),
1183 RetPtr.getValueType(), RetPtr,
1184 DAG.getIntPtrConstant(Offsets[i]));
1186 DAG.getStore(Chain, getCurDebugLoc(),
1187 SDValue(RetOp.getNode(), RetOp.getResNo() + i),
1188 // FIXME: better loc info would be nice.
1189 Add, MachinePointerInfo(), false, false, 0);
1192 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
1193 MVT::Other, &Chains[0], NumValues);
1194 } else if (I.getNumOperands() != 0) {
1195 SmallVector<EVT, 4> ValueVTs;
1196 ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs);
1197 unsigned NumValues = ValueVTs.size();
1199 SDValue RetOp = getValue(I.getOperand(0));
1200 for (unsigned j = 0, f = NumValues; j != f; ++j) {
1201 EVT VT = ValueVTs[j];
1203 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1205 const Function *F = I.getParent()->getParent();
1206 if (F->paramHasAttr(0, Attribute::SExt))
1207 ExtendKind = ISD::SIGN_EXTEND;
1208 else if (F->paramHasAttr(0, Attribute::ZExt))
1209 ExtendKind = ISD::ZERO_EXTEND;
1211 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger())
1212 VT = TLI.getTypeForExtArgOrReturn(*DAG.getContext(), VT, ExtendKind);
1214 unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), VT);
1215 EVT PartVT = TLI.getRegisterType(*DAG.getContext(), VT);
1216 SmallVector<SDValue, 4> Parts(NumParts);
1217 getCopyToParts(DAG, getCurDebugLoc(),
1218 SDValue(RetOp.getNode(), RetOp.getResNo() + j),
1219 &Parts[0], NumParts, PartVT, ExtendKind);
1221 // 'inreg' on function refers to return value
1222 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1223 if (F->paramHasAttr(0, Attribute::InReg))
1226 // Propagate extension type if any
1227 if (ExtendKind == ISD::SIGN_EXTEND)
1229 else if (ExtendKind == ISD::ZERO_EXTEND)
1232 for (unsigned i = 0; i < NumParts; ++i) {
1233 Outs.push_back(ISD::OutputArg(Flags, Parts[i].getValueType(),
1235 OutVals.push_back(Parts[i]);
1241 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1242 CallingConv::ID CallConv =
1243 DAG.getMachineFunction().getFunction()->getCallingConv();
1244 Chain = TLI.LowerReturn(Chain, CallConv, isVarArg,
1245 Outs, OutVals, getCurDebugLoc(), DAG);
1247 // Verify that the target's LowerReturn behaved as expected.
1248 assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
1249 "LowerReturn didn't return a valid chain!");
1251 // Update the DAG with the new chain value resulting from return lowering.
1255 /// CopyToExportRegsIfNeeded - If the given value has virtual registers
1256 /// created for it, emit nodes to copy the value into the virtual
1258 void SelectionDAGBuilder::CopyToExportRegsIfNeeded(const Value *V) {
1260 if (V->getType()->isEmptyTy())
1263 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
1264 if (VMI != FuncInfo.ValueMap.end()) {
1265 assert(!V->use_empty() && "Unused value assigned virtual registers!");
1266 CopyValueToVirtualRegister(V, VMI->second);
1270 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
1271 /// the current basic block, add it to ValueMap now so that we'll get a
1273 void SelectionDAGBuilder::ExportFromCurrentBlock(const Value *V) {
1274 // No need to export constants.
1275 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
1277 // Already exported?
1278 if (FuncInfo.isExportedInst(V)) return;
1280 unsigned Reg = FuncInfo.InitializeRegForValue(V);
1281 CopyValueToVirtualRegister(V, Reg);
1284 bool SelectionDAGBuilder::isExportableFromCurrentBlock(const Value *V,
1285 const BasicBlock *FromBB) {
1286 // The operands of the setcc have to be in this block. We don't know
1287 // how to export them from some other block.
1288 if (const Instruction *VI = dyn_cast<Instruction>(V)) {
1289 // Can export from current BB.
1290 if (VI->getParent() == FromBB)
1293 // Is already exported, noop.
1294 return FuncInfo.isExportedInst(V);
1297 // If this is an argument, we can export it if the BB is the entry block or
1298 // if it is already exported.
1299 if (isa<Argument>(V)) {
1300 if (FromBB == &FromBB->getParent()->getEntryBlock())
1303 // Otherwise, can only export this if it is already exported.
1304 return FuncInfo.isExportedInst(V);
1307 // Otherwise, constants can always be exported.
1311 /// Return branch probability calculated by BranchProbabilityInfo for IR blocks.
1312 uint32_t SelectionDAGBuilder::getEdgeWeight(const MachineBasicBlock *Src,
1313 const MachineBasicBlock *Dst) const {
1314 BranchProbabilityInfo *BPI = FuncInfo.BPI;
1317 const BasicBlock *SrcBB = Src->getBasicBlock();
1318 const BasicBlock *DstBB = Dst->getBasicBlock();
1319 return BPI->getEdgeWeight(SrcBB, DstBB);
1322 void SelectionDAGBuilder::
1323 addSuccessorWithWeight(MachineBasicBlock *Src, MachineBasicBlock *Dst,
1324 uint32_t Weight /* = 0 */) {
1326 Weight = getEdgeWeight(Src, Dst);
1327 Src->addSuccessor(Dst, Weight);
1331 static bool InBlock(const Value *V, const BasicBlock *BB) {
1332 if (const Instruction *I = dyn_cast<Instruction>(V))
1333 return I->getParent() == BB;
1337 /// EmitBranchForMergedCondition - Helper method for FindMergedConditions.
1338 /// This function emits a branch and is used at the leaves of an OR or an
1339 /// AND operator tree.
1342 SelectionDAGBuilder::EmitBranchForMergedCondition(const Value *Cond,
1343 MachineBasicBlock *TBB,
1344 MachineBasicBlock *FBB,
1345 MachineBasicBlock *CurBB,
1346 MachineBasicBlock *SwitchBB) {
1347 const BasicBlock *BB = CurBB->getBasicBlock();
1349 // If the leaf of the tree is a comparison, merge the condition into
1351 if (const CmpInst *BOp = dyn_cast<CmpInst>(Cond)) {
1352 // The operands of the cmp have to be in this block. We don't know
1353 // how to export them from some other block. If this is the first block
1354 // of the sequence, no exporting is needed.
1355 if (CurBB == SwitchBB ||
1356 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
1357 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) {
1358 ISD::CondCode Condition;
1359 if (const ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
1360 Condition = getICmpCondCode(IC->getPredicate());
1361 } else if (const FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) {
1362 Condition = getFCmpCondCode(FC->getPredicate());
1363 if (TM.Options.NoNaNsFPMath)
1364 Condition = getFCmpCodeWithoutNaN(Condition);
1366 Condition = ISD::SETEQ; // silence warning.
1367 llvm_unreachable("Unknown compare instruction");
1370 CaseBlock CB(Condition, BOp->getOperand(0),
1371 BOp->getOperand(1), NULL, TBB, FBB, CurBB);
1372 SwitchCases.push_back(CB);
1377 // Create a CaseBlock record representing this branch.
1378 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()),
1379 NULL, TBB, FBB, CurBB);
1380 SwitchCases.push_back(CB);
1383 /// FindMergedConditions - If Cond is an expression like
1384 void SelectionDAGBuilder::FindMergedConditions(const Value *Cond,
1385 MachineBasicBlock *TBB,
1386 MachineBasicBlock *FBB,
1387 MachineBasicBlock *CurBB,
1388 MachineBasicBlock *SwitchBB,
1390 // If this node is not part of the or/and tree, emit it as a branch.
1391 const Instruction *BOp = dyn_cast<Instruction>(Cond);
1392 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
1393 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
1394 BOp->getParent() != CurBB->getBasicBlock() ||
1395 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
1396 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
1397 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB);
1401 // Create TmpBB after CurBB.
1402 MachineFunction::iterator BBI = CurBB;
1403 MachineFunction &MF = DAG.getMachineFunction();
1404 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock());
1405 CurBB->getParent()->insert(++BBI, TmpBB);
1407 if (Opc == Instruction::Or) {
1408 // Codegen X | Y as:
1416 // Emit the LHS condition.
1417 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, SwitchBB, Opc);
1419 // Emit the RHS condition into TmpBB.
1420 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
1422 assert(Opc == Instruction::And && "Unknown merge op!");
1423 // Codegen X & Y as:
1430 // This requires creation of TmpBB after CurBB.
1432 // Emit the LHS condition.
1433 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, SwitchBB, Opc);
1435 // Emit the RHS condition into TmpBB.
1436 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
1440 /// If the set of cases should be emitted as a series of branches, return true.
1441 /// If we should emit this as a bunch of and/or'd together conditions, return
1444 SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases){
1445 if (Cases.size() != 2) return true;
1447 // If this is two comparisons of the same values or'd or and'd together, they
1448 // will get folded into a single comparison, so don't emit two blocks.
1449 if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
1450 Cases[0].CmpRHS == Cases[1].CmpRHS) ||
1451 (Cases[0].CmpRHS == Cases[1].CmpLHS &&
1452 Cases[0].CmpLHS == Cases[1].CmpRHS)) {
1456 // Handle: (X != null) | (Y != null) --> (X|Y) != 0
1457 // Handle: (X == null) & (Y == null) --> (X|Y) == 0
1458 if (Cases[0].CmpRHS == Cases[1].CmpRHS &&
1459 Cases[0].CC == Cases[1].CC &&
1460 isa<Constant>(Cases[0].CmpRHS) &&
1461 cast<Constant>(Cases[0].CmpRHS)->isNullValue()) {
1462 if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB)
1464 if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB)
1471 void SelectionDAGBuilder::visitBr(const BranchInst &I) {
1472 MachineBasicBlock *BrMBB = FuncInfo.MBB;
1474 // Update machine-CFG edges.
1475 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
1477 // Figure out which block is immediately after the current one.
1478 MachineBasicBlock *NextBlock = 0;
1479 MachineFunction::iterator BBI = BrMBB;
1480 if (++BBI != FuncInfo.MF->end())
1483 if (I.isUnconditional()) {
1484 // Update machine-CFG edges.
1485 BrMBB->addSuccessor(Succ0MBB);
1487 // If this is not a fall-through branch, emit the branch.
1488 if (Succ0MBB != NextBlock)
1489 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
1490 MVT::Other, getControlRoot(),
1491 DAG.getBasicBlock(Succ0MBB)));
1496 // If this condition is one of the special cases we handle, do special stuff
1498 const Value *CondVal = I.getCondition();
1499 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
1501 // If this is a series of conditions that are or'd or and'd together, emit
1502 // this as a sequence of branches instead of setcc's with and/or operations.
1503 // As long as jumps are not expensive, this should improve performance.
1504 // For example, instead of something like:
1517 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
1518 if (!TLI.isJumpExpensive() &&
1520 (BOp->getOpcode() == Instruction::And ||
1521 BOp->getOpcode() == Instruction::Or)) {
1522 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB,
1524 // If the compares in later blocks need to use values not currently
1525 // exported from this block, export them now. This block should always
1526 // be the first entry.
1527 assert(SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!");
1529 // Allow some cases to be rejected.
1530 if (ShouldEmitAsBranches(SwitchCases)) {
1531 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1532 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1533 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1536 // Emit the branch for this block.
1537 visitSwitchCase(SwitchCases[0], BrMBB);
1538 SwitchCases.erase(SwitchCases.begin());
1542 // Okay, we decided not to do this, remove any inserted MBB's and clear
1544 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
1545 FuncInfo.MF->erase(SwitchCases[i].ThisBB);
1547 SwitchCases.clear();
1551 // Create a CaseBlock record representing this branch.
1552 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()),
1553 NULL, Succ0MBB, Succ1MBB, BrMBB);
1555 // Use visitSwitchCase to actually insert the fast branch sequence for this
1557 visitSwitchCase(CB, BrMBB);
1560 /// visitSwitchCase - Emits the necessary code to represent a single node in
1561 /// the binary search tree resulting from lowering a switch instruction.
1562 void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB,
1563 MachineBasicBlock *SwitchBB) {
1565 SDValue CondLHS = getValue(CB.CmpLHS);
1566 DebugLoc dl = getCurDebugLoc();
1568 // Build the setcc now.
1569 if (CB.CmpMHS == NULL) {
1570 // Fold "(X == true)" to X and "(X == false)" to !X to
1571 // handle common cases produced by branch lowering.
1572 if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) &&
1573 CB.CC == ISD::SETEQ)
1575 else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) &&
1576 CB.CC == ISD::SETEQ) {
1577 SDValue True = DAG.getConstant(1, CondLHS.getValueType());
1578 Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True);
1580 Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1582 assert(CB.CC == ISD::SETCC_INVALID &&
1583 "Condition is undefined for to-the-range belonging check.");
1585 const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue();
1586 const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue();
1588 SDValue CmpOp = getValue(CB.CmpMHS);
1589 EVT VT = CmpOp.getValueType();
1591 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(false)) {
1592 Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, VT),
1595 SDValue SUB = DAG.getNode(ISD::SUB, dl,
1596 VT, CmpOp, DAG.getConstant(Low, VT));
1597 Cond = DAG.getSetCC(dl, MVT::i1, SUB,
1598 DAG.getConstant(High-Low, VT), ISD::SETULE);
1602 // Update successor info
1603 addSuccessorWithWeight(SwitchBB, CB.TrueBB, CB.TrueWeight);
1604 addSuccessorWithWeight(SwitchBB, CB.FalseBB, CB.FalseWeight);
1606 // Set NextBlock to be the MBB immediately after the current one, if any.
1607 // This is used to avoid emitting unnecessary branches to the next block.
1608 MachineBasicBlock *NextBlock = 0;
1609 MachineFunction::iterator BBI = SwitchBB;
1610 if (++BBI != FuncInfo.MF->end())
1613 // If the lhs block is the next block, invert the condition so that we can
1614 // fall through to the lhs instead of the rhs block.
1615 if (CB.TrueBB == NextBlock) {
1616 std::swap(CB.TrueBB, CB.FalseBB);
1617 SDValue True = DAG.getConstant(1, Cond.getValueType());
1618 Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True);
1621 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1622 MVT::Other, getControlRoot(), Cond,
1623 DAG.getBasicBlock(CB.TrueBB));
1625 // Insert the false branch. Do this even if it's a fall through branch,
1626 // this makes it easier to do DAG optimizations which require inverting
1627 // the branch condition.
1628 BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1629 DAG.getBasicBlock(CB.FalseBB));
1631 DAG.setRoot(BrCond);
1634 /// visitJumpTable - Emit JumpTable node in the current MBB
1635 void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) {
1636 // Emit the code for the jump table
1637 assert(JT.Reg != -1U && "Should lower JT Header first!");
1638 EVT PTy = TLI.getPointerTy();
1639 SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(),
1641 SDValue Table = DAG.getJumpTable(JT.JTI, PTy);
1642 SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurDebugLoc(),
1643 MVT::Other, Index.getValue(1),
1645 DAG.setRoot(BrJumpTable);
1648 /// visitJumpTableHeader - This function emits necessary code to produce index
1649 /// in the JumpTable from switch case.
1650 void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT,
1651 JumpTableHeader &JTH,
1652 MachineBasicBlock *SwitchBB) {
1653 // Subtract the lowest switch case value from the value being switched on and
1654 // conditional branch to default mbb if the result is greater than the
1655 // difference between smallest and largest cases.
1656 SDValue SwitchOp = getValue(JTH.SValue);
1657 EVT VT = SwitchOp.getValueType();
1658 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1659 DAG.getConstant(JTH.First, VT));
1661 // The SDNode we just created, which holds the value being switched on minus
1662 // the smallest case value, needs to be copied to a virtual register so it
1663 // can be used as an index into the jump table in a subsequent basic block.
1664 // This value may be smaller or larger than the target's pointer type, and
1665 // therefore require extension or truncating.
1666 SwitchOp = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(), TLI.getPointerTy());
1668 unsigned JumpTableReg = FuncInfo.CreateReg(TLI.getPointerTy());
1669 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1670 JumpTableReg, SwitchOp);
1671 JT.Reg = JumpTableReg;
1673 // Emit the range check for the jump table, and branch to the default block
1674 // for the switch statement if the value being switched on exceeds the largest
1675 // case in the switch.
1676 SDValue CMP = DAG.getSetCC(getCurDebugLoc(),
1677 TLI.getSetCCResultType(Sub.getValueType()), Sub,
1678 DAG.getConstant(JTH.Last-JTH.First,VT),
1681 // Set NextBlock to be the MBB immediately after the current one, if any.
1682 // This is used to avoid emitting unnecessary branches to the next block.
1683 MachineBasicBlock *NextBlock = 0;
1684 MachineFunction::iterator BBI = SwitchBB;
1686 if (++BBI != FuncInfo.MF->end())
1689 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1690 MVT::Other, CopyTo, CMP,
1691 DAG.getBasicBlock(JT.Default));
1693 if (JT.MBB != NextBlock)
1694 BrCond = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrCond,
1695 DAG.getBasicBlock(JT.MBB));
1697 DAG.setRoot(BrCond);
1700 /// visitBitTestHeader - This function emits necessary code to produce value
1701 /// suitable for "bit tests"
1702 void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B,
1703 MachineBasicBlock *SwitchBB) {
1704 // Subtract the minimum value
1705 SDValue SwitchOp = getValue(B.SValue);
1706 EVT VT = SwitchOp.getValueType();
1707 SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1708 DAG.getConstant(B.First, VT));
1711 SDValue RangeCmp = DAG.getSetCC(getCurDebugLoc(),
1712 TLI.getSetCCResultType(Sub.getValueType()),
1713 Sub, DAG.getConstant(B.Range, VT),
1716 // Determine the type of the test operands.
1717 bool UsePtrType = false;
1718 if (!TLI.isTypeLegal(VT))
1721 for (unsigned i = 0, e = B.Cases.size(); i != e; ++i)
1722 if (!isUIntN(VT.getSizeInBits(), B.Cases[i].Mask)) {
1723 // Switch table case range are encoded into series of masks.
1724 // Just use pointer type, it's guaranteed to fit.
1730 VT = TLI.getPointerTy();
1731 Sub = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(), VT);
1735 B.Reg = FuncInfo.CreateReg(VT);
1736 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1739 // Set NextBlock to be the MBB immediately after the current one, if any.
1740 // This is used to avoid emitting unnecessary branches to the next block.
1741 MachineBasicBlock *NextBlock = 0;
1742 MachineFunction::iterator BBI = SwitchBB;
1743 if (++BBI != FuncInfo.MF->end())
1746 MachineBasicBlock* MBB = B.Cases[0].ThisBB;
1748 addSuccessorWithWeight(SwitchBB, B.Default);
1749 addSuccessorWithWeight(SwitchBB, MBB);
1751 SDValue BrRange = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1752 MVT::Other, CopyTo, RangeCmp,
1753 DAG.getBasicBlock(B.Default));
1755 if (MBB != NextBlock)
1756 BrRange = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, CopyTo,
1757 DAG.getBasicBlock(MBB));
1759 DAG.setRoot(BrRange);
1762 /// visitBitTestCase - this function produces one "bit test"
1763 void SelectionDAGBuilder::visitBitTestCase(BitTestBlock &BB,
1764 MachineBasicBlock* NextMBB,
1767 MachineBasicBlock *SwitchBB) {
1769 SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(),
1772 unsigned PopCount = CountPopulation_64(B.Mask);
1773 if (PopCount == 1) {
1774 // Testing for a single bit; just compare the shift count with what it
1775 // would need to be to shift a 1 bit in that position.
1776 Cmp = DAG.getSetCC(getCurDebugLoc(),
1777 TLI.getSetCCResultType(VT),
1779 DAG.getConstant(CountTrailingZeros_64(B.Mask), VT),
1781 } else if (PopCount == BB.Range) {
1782 // There is only one zero bit in the range, test for it directly.
1783 Cmp = DAG.getSetCC(getCurDebugLoc(),
1784 TLI.getSetCCResultType(VT),
1786 DAG.getConstant(CountTrailingOnes_64(B.Mask), VT),
1789 // Make desired shift
1790 SDValue SwitchVal = DAG.getNode(ISD::SHL, getCurDebugLoc(), VT,
1791 DAG.getConstant(1, VT), ShiftOp);
1793 // Emit bit tests and jumps
1794 SDValue AndOp = DAG.getNode(ISD::AND, getCurDebugLoc(),
1795 VT, SwitchVal, DAG.getConstant(B.Mask, VT));
1796 Cmp = DAG.getSetCC(getCurDebugLoc(),
1797 TLI.getSetCCResultType(VT),
1798 AndOp, DAG.getConstant(0, VT),
1802 addSuccessorWithWeight(SwitchBB, B.TargetBB);
1803 addSuccessorWithWeight(SwitchBB, NextMBB);
1805 SDValue BrAnd = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1806 MVT::Other, getControlRoot(),
1807 Cmp, DAG.getBasicBlock(B.TargetBB));
1809 // Set NextBlock to be the MBB immediately after the current one, if any.
1810 // This is used to avoid emitting unnecessary branches to the next block.
1811 MachineBasicBlock *NextBlock = 0;
1812 MachineFunction::iterator BBI = SwitchBB;
1813 if (++BBI != FuncInfo.MF->end())
1816 if (NextMBB != NextBlock)
1817 BrAnd = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrAnd,
1818 DAG.getBasicBlock(NextMBB));
1823 void SelectionDAGBuilder::visitInvoke(const InvokeInst &I) {
1824 MachineBasicBlock *InvokeMBB = FuncInfo.MBB;
1826 // Retrieve successors.
1827 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
1828 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)];
1830 const Value *Callee(I.getCalledValue());
1831 const Function *Fn = dyn_cast<Function>(Callee);
1832 if (isa<InlineAsm>(Callee))
1834 else if (Fn && Fn->isIntrinsic()) {
1835 assert(Fn->getIntrinsicID() == Intrinsic::donothing);
1836 return; // ignore invokes to @llvm.donothing
1838 LowerCallTo(&I, getValue(Callee), false, LandingPad);
1840 // If the value of the invoke is used outside of its defining block, make it
1841 // available as a virtual register.
1842 CopyToExportRegsIfNeeded(&I);
1844 // Update successor info
1845 addSuccessorWithWeight(InvokeMBB, Return);
1846 addSuccessorWithWeight(InvokeMBB, LandingPad);
1848 // Drop into normal successor.
1849 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
1850 MVT::Other, getControlRoot(),
1851 DAG.getBasicBlock(Return)));
1854 void SelectionDAGBuilder::visitResume(const ResumeInst &RI) {
1855 llvm_unreachable("SelectionDAGBuilder shouldn't visit resume instructions!");
1858 void SelectionDAGBuilder::visitLandingPad(const LandingPadInst &LP) {
1859 assert(FuncInfo.MBB->isLandingPad() &&
1860 "Call to landingpad not in landing pad!");
1862 MachineBasicBlock *MBB = FuncInfo.MBB;
1863 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
1864 AddLandingPadInfo(LP, MMI, MBB);
1866 // If there aren't registers to copy the values into (e.g., during SjLj
1867 // exceptions), then don't bother to create these DAG nodes.
1868 if (TLI.getExceptionPointerRegister() == 0 &&
1869 TLI.getExceptionSelectorRegister() == 0)
1872 SmallVector<EVT, 2> ValueVTs;
1873 ComputeValueVTs(TLI, LP.getType(), ValueVTs);
1875 // Insert the EXCEPTIONADDR instruction.
1876 assert(FuncInfo.MBB->isLandingPad() &&
1877 "Call to eh.exception not in landing pad!");
1878 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
1880 Ops[0] = DAG.getRoot();
1881 SDValue Op1 = DAG.getNode(ISD::EXCEPTIONADDR, getCurDebugLoc(), VTs, Ops, 1);
1882 SDValue Chain = Op1.getValue(1);
1884 // Insert the EHSELECTION instruction.
1885 VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
1888 SDValue Op2 = DAG.getNode(ISD::EHSELECTION, getCurDebugLoc(), VTs, Ops, 2);
1889 Chain = Op2.getValue(1);
1890 Op2 = DAG.getSExtOrTrunc(Op2, getCurDebugLoc(), MVT::i32);
1894 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
1895 DAG.getVTList(&ValueVTs[0], ValueVTs.size()),
1898 std::pair<SDValue, SDValue> RetPair = std::make_pair(Res, Chain);
1899 setValue(&LP, RetPair.first);
1900 DAG.setRoot(RetPair.second);
1903 /// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for
1904 /// small case ranges).
1905 bool SelectionDAGBuilder::handleSmallSwitchRange(CaseRec& CR,
1906 CaseRecVector& WorkList,
1908 MachineBasicBlock *Default,
1909 MachineBasicBlock *SwitchBB) {
1910 // Size is the number of Cases represented by this range.
1911 size_t Size = CR.Range.second - CR.Range.first;
1915 // Get the MachineFunction which holds the current MBB. This is used when
1916 // inserting any additional MBBs necessary to represent the switch.
1917 MachineFunction *CurMF = FuncInfo.MF;
1919 // Figure out which block is immediately after the current one.
1920 MachineBasicBlock *NextBlock = 0;
1921 MachineFunction::iterator BBI = CR.CaseBB;
1923 if (++BBI != FuncInfo.MF->end())
1926 // If any two of the cases has the same destination, and if one value
1927 // is the same as the other, but has one bit unset that the other has set,
1928 // use bit manipulation to do two compares at once. For example:
1929 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
1930 // TODO: This could be extended to merge any 2 cases in switches with 3 cases.
1931 // TODO: Handle cases where CR.CaseBB != SwitchBB.
1932 if (Size == 2 && CR.CaseBB == SwitchBB) {
1933 Case &Small = *CR.Range.first;
1934 Case &Big = *(CR.Range.second-1);
1936 if (Small.Low == Small.High && Big.Low == Big.High && Small.BB == Big.BB) {
1937 const APInt& SmallValue = cast<ConstantInt>(Small.Low)->getValue();
1938 const APInt& BigValue = cast<ConstantInt>(Big.Low)->getValue();
1940 // Check that there is only one bit different.
1941 if (BigValue.countPopulation() == SmallValue.countPopulation() + 1 &&
1942 (SmallValue | BigValue) == BigValue) {
1943 // Isolate the common bit.
1944 APInt CommonBit = BigValue & ~SmallValue;
1945 assert((SmallValue | CommonBit) == BigValue &&
1946 CommonBit.countPopulation() == 1 && "Not a common bit?");
1948 SDValue CondLHS = getValue(SV);
1949 EVT VT = CondLHS.getValueType();
1950 DebugLoc DL = getCurDebugLoc();
1952 SDValue Or = DAG.getNode(ISD::OR, DL, VT, CondLHS,
1953 DAG.getConstant(CommonBit, VT));
1954 SDValue Cond = DAG.getSetCC(DL, MVT::i1,
1955 Or, DAG.getConstant(BigValue, VT),
1958 // Update successor info.
1959 addSuccessorWithWeight(SwitchBB, Small.BB);
1960 addSuccessorWithWeight(SwitchBB, Default);
1962 // Insert the true branch.
1963 SDValue BrCond = DAG.getNode(ISD::BRCOND, DL, MVT::Other,
1964 getControlRoot(), Cond,
1965 DAG.getBasicBlock(Small.BB));
1967 // Insert the false branch.
1968 BrCond = DAG.getNode(ISD::BR, DL, MVT::Other, BrCond,
1969 DAG.getBasicBlock(Default));
1971 DAG.setRoot(BrCond);
1977 // Order cases by weight so the most likely case will be checked first.
1978 BranchProbabilityInfo *BPI = FuncInfo.BPI;
1980 for (CaseItr I = CR.Range.first, IE = CR.Range.second; I != IE; ++I) {
1981 uint32_t IWeight = BPI->getEdgeWeight(SwitchBB->getBasicBlock(),
1982 I->BB->getBasicBlock());
1983 for (CaseItr J = CR.Range.first; J < I; ++J) {
1984 uint32_t JWeight = BPI->getEdgeWeight(SwitchBB->getBasicBlock(),
1985 J->BB->getBasicBlock());
1986 if (IWeight > JWeight)
1991 // Rearrange the case blocks so that the last one falls through if possible.
1992 Case &BackCase = *(CR.Range.second-1);
1994 NextBlock && Default != NextBlock && BackCase.BB != NextBlock) {
1995 // The last case block won't fall through into 'NextBlock' if we emit the
1996 // branches in this order. See if rearranging a case value would help.
1997 // We start at the bottom as it's the case with the least weight.
1998 for (Case *I = &*(CR.Range.second-2), *E = &*CR.Range.first-1; I != E; --I){
1999 if (I->BB == NextBlock) {
2000 std::swap(*I, BackCase);
2006 // Create a CaseBlock record representing a conditional branch to
2007 // the Case's target mbb if the value being switched on SV is equal
2009 MachineBasicBlock *CurBlock = CR.CaseBB;
2010 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
2011 MachineBasicBlock *FallThrough;
2013 FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock());
2014 CurMF->insert(BBI, FallThrough);
2016 // Put SV in a virtual register to make it available from the new blocks.
2017 ExportFromCurrentBlock(SV);
2019 // If the last case doesn't match, go to the default block.
2020 FallThrough = Default;
2023 const Value *RHS, *LHS, *MHS;
2025 if (I->High == I->Low) {
2026 // This is just small small case range :) containing exactly 1 case
2028 LHS = SV; RHS = I->High; MHS = NULL;
2030 CC = ISD::SETCC_INVALID;
2031 LHS = I->Low; MHS = SV; RHS = I->High;
2034 uint32_t ExtraWeight = I->ExtraWeight;
2035 CaseBlock CB(CC, LHS, RHS, MHS, /* truebb */ I->BB, /* falsebb */ FallThrough,
2037 /* trueweight */ ExtraWeight / 2, /* falseweight */ ExtraWeight / 2);
2039 // If emitting the first comparison, just call visitSwitchCase to emit the
2040 // code into the current block. Otherwise, push the CaseBlock onto the
2041 // vector to be later processed by SDISel, and insert the node's MBB
2042 // before the next MBB.
2043 if (CurBlock == SwitchBB)
2044 visitSwitchCase(CB, SwitchBB);
2046 SwitchCases.push_back(CB);
2048 CurBlock = FallThrough;
2054 static inline bool areJTsAllowed(const TargetLowering &TLI) {
2055 return TLI.supportJumpTables() &&
2056 (TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
2057 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other));
2060 static APInt ComputeRange(const APInt &First, const APInt &Last) {
2061 uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1;
2062 APInt LastExt = Last.zext(BitWidth), FirstExt = First.zext(BitWidth);
2063 return (LastExt - FirstExt + 1ULL);
2066 /// handleJTSwitchCase - Emit jumptable for current switch case range
2067 bool SelectionDAGBuilder::handleJTSwitchCase(CaseRec &CR,
2068 CaseRecVector &WorkList,
2070 MachineBasicBlock *Default,
2071 MachineBasicBlock *SwitchBB) {
2072 Case& FrontCase = *CR.Range.first;
2073 Case& BackCase = *(CR.Range.second-1);
2075 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
2076 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
2078 APInt TSize(First.getBitWidth(), 0);
2079 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I)
2082 if (!areJTsAllowed(TLI) || TSize.ult(4))
2085 APInt Range = ComputeRange(First, Last);
2086 // The density is TSize / Range. Require at least 40%.
2087 // It should not be possible for IntTSize to saturate for sane code, but make
2088 // sure we handle Range saturation correctly.
2089 uint64_t IntRange = Range.getLimitedValue(UINT64_MAX/10);
2090 uint64_t IntTSize = TSize.getLimitedValue(UINT64_MAX/10);
2091 if (IntTSize * 10 < IntRange * 4)
2094 DEBUG(dbgs() << "Lowering jump table\n"
2095 << "First entry: " << First << ". Last entry: " << Last << '\n'
2096 << "Range: " << Range << ". Size: " << TSize << ".\n\n");
2098 // Get the MachineFunction which holds the current MBB. This is used when
2099 // inserting any additional MBBs necessary to represent the switch.
2100 MachineFunction *CurMF = FuncInfo.MF;
2102 // Figure out which block is immediately after the current one.
2103 MachineFunction::iterator BBI = CR.CaseBB;
2106 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2108 // Create a new basic block to hold the code for loading the address
2109 // of the jump table, and jumping to it. Update successor information;
2110 // we will either branch to the default case for the switch, or the jump
2112 MachineBasicBlock *JumpTableBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2113 CurMF->insert(BBI, JumpTableBB);
2115 addSuccessorWithWeight(CR.CaseBB, Default);
2116 addSuccessorWithWeight(CR.CaseBB, JumpTableBB);
2118 // Build a vector of destination BBs, corresponding to each target
2119 // of the jump table. If the value of the jump table slot corresponds to
2120 // a case statement, push the case's BB onto the vector, otherwise, push
2122 std::vector<MachineBasicBlock*> DestBBs;
2124 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) {
2125 const APInt &Low = cast<ConstantInt>(I->Low)->getValue();
2126 const APInt &High = cast<ConstantInt>(I->High)->getValue();
2128 if (Low.ule(TEI) && TEI.ule(High)) {
2129 DestBBs.push_back(I->BB);
2133 DestBBs.push_back(Default);
2137 // Update successor info. Add one edge to each unique successor.
2138 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs());
2139 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(),
2140 E = DestBBs.end(); I != E; ++I) {
2141 if (!SuccsHandled[(*I)->getNumber()]) {
2142 SuccsHandled[(*I)->getNumber()] = true;
2143 addSuccessorWithWeight(JumpTableBB, *I);
2147 // Create a jump table index for this jump table.
2148 unsigned JTEncoding = TLI.getJumpTableEncoding();
2149 unsigned JTI = CurMF->getOrCreateJumpTableInfo(JTEncoding)
2150 ->createJumpTableIndex(DestBBs);
2152 // Set the jump table information so that we can codegen it as a second
2153 // MachineBasicBlock
2154 JumpTable JT(-1U, JTI, JumpTableBB, Default);
2155 JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == SwitchBB));
2156 if (CR.CaseBB == SwitchBB)
2157 visitJumpTableHeader(JT, JTH, SwitchBB);
2159 JTCases.push_back(JumpTableBlock(JTH, JT));
2163 /// handleBTSplitSwitchCase - emit comparison and split binary search tree into
2165 bool SelectionDAGBuilder::handleBTSplitSwitchCase(CaseRec& CR,
2166 CaseRecVector& WorkList,
2168 MachineBasicBlock *Default,
2169 MachineBasicBlock *SwitchBB) {
2170 // Get the MachineFunction which holds the current MBB. This is used when
2171 // inserting any additional MBBs necessary to represent the switch.
2172 MachineFunction *CurMF = FuncInfo.MF;
2174 // Figure out which block is immediately after the current one.
2175 MachineFunction::iterator BBI = CR.CaseBB;
2178 Case& FrontCase = *CR.Range.first;
2179 Case& BackCase = *(CR.Range.second-1);
2180 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2182 // Size is the number of Cases represented by this range.
2183 unsigned Size = CR.Range.second - CR.Range.first;
2185 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
2186 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
2188 CaseItr Pivot = CR.Range.first + Size/2;
2190 // Select optimal pivot, maximizing sum density of LHS and RHS. This will
2191 // (heuristically) allow us to emit JumpTable's later.
2192 APInt TSize(First.getBitWidth(), 0);
2193 for (CaseItr I = CR.Range.first, E = CR.Range.second;
2197 APInt LSize = FrontCase.size();
2198 APInt RSize = TSize-LSize;
2199 DEBUG(dbgs() << "Selecting best pivot: \n"
2200 << "First: " << First << ", Last: " << Last <<'\n'
2201 << "LSize: " << LSize << ", RSize: " << RSize << '\n');
2202 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second;
2204 const APInt &LEnd = cast<ConstantInt>(I->High)->getValue();
2205 const APInt &RBegin = cast<ConstantInt>(J->Low)->getValue();
2206 APInt Range = ComputeRange(LEnd, RBegin);
2207 assert((Range - 2ULL).isNonNegative() &&
2208 "Invalid case distance");
2209 // Use volatile double here to avoid excess precision issues on some hosts,
2210 // e.g. that use 80-bit X87 registers.
2211 volatile double LDensity =
2212 (double)LSize.roundToDouble() /
2213 (LEnd - First + 1ULL).roundToDouble();
2214 volatile double RDensity =
2215 (double)RSize.roundToDouble() /
2216 (Last - RBegin + 1ULL).roundToDouble();
2217 double Metric = Range.logBase2()*(LDensity+RDensity);
2218 // Should always split in some non-trivial place
2219 DEBUG(dbgs() <<"=>Step\n"
2220 << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n'
2221 << "LDensity: " << LDensity
2222 << ", RDensity: " << RDensity << '\n'
2223 << "Metric: " << Metric << '\n');
2224 if (FMetric < Metric) {
2227 DEBUG(dbgs() << "Current metric set to: " << FMetric << '\n');
2233 if (areJTsAllowed(TLI)) {
2234 // If our case is dense we *really* should handle it earlier!
2235 assert((FMetric > 0) && "Should handle dense range earlier!");
2237 Pivot = CR.Range.first + Size/2;
2240 CaseRange LHSR(CR.Range.first, Pivot);
2241 CaseRange RHSR(Pivot, CR.Range.second);
2242 const Constant *C = Pivot->Low;
2243 MachineBasicBlock *FalseBB = 0, *TrueBB = 0;
2245 // We know that we branch to the LHS if the Value being switched on is
2246 // less than the Pivot value, C. We use this to optimize our binary
2247 // tree a bit, by recognizing that if SV is greater than or equal to the
2248 // LHS's Case Value, and that Case Value is exactly one less than the
2249 // Pivot's Value, then we can branch directly to the LHS's Target,
2250 // rather than creating a leaf node for it.
2251 if ((LHSR.second - LHSR.first) == 1 &&
2252 LHSR.first->High == CR.GE &&
2253 cast<ConstantInt>(C)->getValue() ==
2254 (cast<ConstantInt>(CR.GE)->getValue() + 1LL)) {
2255 TrueBB = LHSR.first->BB;
2257 TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2258 CurMF->insert(BBI, TrueBB);
2259 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR));
2261 // Put SV in a virtual register to make it available from the new blocks.
2262 ExportFromCurrentBlock(SV);
2265 // Similar to the optimization above, if the Value being switched on is
2266 // known to be less than the Constant CR.LT, and the current Case Value
2267 // is CR.LT - 1, then we can branch directly to the target block for
2268 // the current Case Value, rather than emitting a RHS leaf node for it.
2269 if ((RHSR.second - RHSR.first) == 1 && CR.LT &&
2270 cast<ConstantInt>(RHSR.first->Low)->getValue() ==
2271 (cast<ConstantInt>(CR.LT)->getValue() - 1LL)) {
2272 FalseBB = RHSR.first->BB;
2274 FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2275 CurMF->insert(BBI, FalseBB);
2276 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR));
2278 // Put SV in a virtual register to make it available from the new blocks.
2279 ExportFromCurrentBlock(SV);
2282 // Create a CaseBlock record representing a conditional branch to
2283 // the LHS node if the value being switched on SV is less than C.
2284 // Otherwise, branch to LHS.
2285 CaseBlock CB(ISD::SETULT, SV, C, NULL, TrueBB, FalseBB, CR.CaseBB);
2287 if (CR.CaseBB == SwitchBB)
2288 visitSwitchCase(CB, SwitchBB);
2290 SwitchCases.push_back(CB);
2295 /// handleBitTestsSwitchCase - if current case range has few destination and
2296 /// range span less, than machine word bitwidth, encode case range into series
2297 /// of masks and emit bit tests with these masks.
2298 bool SelectionDAGBuilder::handleBitTestsSwitchCase(CaseRec& CR,
2299 CaseRecVector& WorkList,
2301 MachineBasicBlock* Default,
2302 MachineBasicBlock *SwitchBB){
2303 EVT PTy = TLI.getPointerTy();
2304 unsigned IntPtrBits = PTy.getSizeInBits();
2306 Case& FrontCase = *CR.Range.first;
2307 Case& BackCase = *(CR.Range.second-1);
2309 // Get the MachineFunction which holds the current MBB. This is used when
2310 // inserting any additional MBBs necessary to represent the switch.
2311 MachineFunction *CurMF = FuncInfo.MF;
2313 // If target does not have legal shift left, do not emit bit tests at all.
2314 if (!TLI.isOperationLegal(ISD::SHL, TLI.getPointerTy()))
2318 for (CaseItr I = CR.Range.first, E = CR.Range.second;
2320 // Single case counts one, case range - two.
2321 numCmps += (I->Low == I->High ? 1 : 2);
2324 // Count unique destinations
2325 SmallSet<MachineBasicBlock*, 4> Dests;
2326 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2327 Dests.insert(I->BB);
2328 if (Dests.size() > 3)
2329 // Don't bother the code below, if there are too much unique destinations
2332 DEBUG(dbgs() << "Total number of unique destinations: "
2333 << Dests.size() << '\n'
2334 << "Total number of comparisons: " << numCmps << '\n');
2336 // Compute span of values.
2337 const APInt& minValue = cast<ConstantInt>(FrontCase.Low)->getValue();
2338 const APInt& maxValue = cast<ConstantInt>(BackCase.High)->getValue();
2339 APInt cmpRange = maxValue - minValue;
2341 DEBUG(dbgs() << "Compare range: " << cmpRange << '\n'
2342 << "Low bound: " << minValue << '\n'
2343 << "High bound: " << maxValue << '\n');
2345 if (cmpRange.uge(IntPtrBits) ||
2346 (!(Dests.size() == 1 && numCmps >= 3) &&
2347 !(Dests.size() == 2 && numCmps >= 5) &&
2348 !(Dests.size() >= 3 && numCmps >= 6)))
2351 DEBUG(dbgs() << "Emitting bit tests\n");
2352 APInt lowBound = APInt::getNullValue(cmpRange.getBitWidth());
2354 // Optimize the case where all the case values fit in a
2355 // word without having to subtract minValue. In this case,
2356 // we can optimize away the subtraction.
2357 if (maxValue.ult(IntPtrBits)) {
2358 cmpRange = maxValue;
2360 lowBound = minValue;
2363 CaseBitsVector CasesBits;
2364 unsigned i, count = 0;
2366 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2367 MachineBasicBlock* Dest = I->BB;
2368 for (i = 0; i < count; ++i)
2369 if (Dest == CasesBits[i].BB)
2373 assert((count < 3) && "Too much destinations to test!");
2374 CasesBits.push_back(CaseBits(0, Dest, 0));
2378 const APInt& lowValue = cast<ConstantInt>(I->Low)->getValue();
2379 const APInt& highValue = cast<ConstantInt>(I->High)->getValue();
2381 uint64_t lo = (lowValue - lowBound).getZExtValue();
2382 uint64_t hi = (highValue - lowBound).getZExtValue();
2384 for (uint64_t j = lo; j <= hi; j++) {
2385 CasesBits[i].Mask |= 1ULL << j;
2386 CasesBits[i].Bits++;
2390 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp());
2394 // Figure out which block is immediately after the current one.
2395 MachineFunction::iterator BBI = CR.CaseBB;
2398 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2400 DEBUG(dbgs() << "Cases:\n");
2401 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) {
2402 DEBUG(dbgs() << "Mask: " << CasesBits[i].Mask
2403 << ", Bits: " << CasesBits[i].Bits
2404 << ", BB: " << CasesBits[i].BB << '\n');
2406 MachineBasicBlock *CaseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2407 CurMF->insert(BBI, CaseBB);
2408 BTC.push_back(BitTestCase(CasesBits[i].Mask,
2412 // Put SV in a virtual register to make it available from the new blocks.
2413 ExportFromCurrentBlock(SV);
2416 BitTestBlock BTB(lowBound, cmpRange, SV,
2417 -1U, MVT::Other, (CR.CaseBB == SwitchBB),
2418 CR.CaseBB, Default, BTC);
2420 if (CR.CaseBB == SwitchBB)
2421 visitBitTestHeader(BTB, SwitchBB);
2423 BitTestCases.push_back(BTB);
2428 /// Clusterify - Transform simple list of Cases into list of CaseRange's
2429 size_t SelectionDAGBuilder::Clusterify(CaseVector& Cases,
2430 const SwitchInst& SI) {
2432 /// Use a shorter form of declaration, and also
2433 /// show the we want to use CRSBuilder as Clusterifier.
2434 typedef IntegersSubsetMapping<MachineBasicBlock> Clusterifier;
2436 Clusterifier TheClusterifier;
2438 // Start with "simple" cases
2439 for (SwitchInst::ConstCaseIt i = SI.case_begin(), e = SI.case_end();
2441 const BasicBlock *SuccBB = i.getCaseSuccessor();
2442 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SuccBB];
2444 TheClusterifier.add(i.getCaseValueEx(), SMBB);
2447 TheClusterifier.optimize();
2449 BranchProbabilityInfo *BPI = FuncInfo.BPI;
2451 for (Clusterifier::RangeIterator i = TheClusterifier.begin(),
2452 e = TheClusterifier.end(); i != e; ++i, ++numCmps) {
2453 const Clusterifier::RangeEx &R = i->first;
2454 MachineBasicBlock *MBB = i->second;
2457 W = BPI->getEdgeWeight(SI.getParent(), MBB->getBasicBlock());
2461 BPI->setEdgeWeight(SI.getParent(), MBB->getBasicBlock(), W);
2464 // FIXME: Currently work with ConstantInt based numbers.
2465 // Changing it to APInt based is a pretty heavy for this commit.
2466 Cases.push_back(Case(R.getLow().toConstantInt(),
2467 R.getHigh().toConstantInt(), MBB, W));
2469 if (R.getLow() != R.getHigh())
2470 // A range counts double, since it requires two compares.
2477 void SelectionDAGBuilder::UpdateSplitBlock(MachineBasicBlock *First,
2478 MachineBasicBlock *Last) {
2480 for (unsigned i = 0, e = JTCases.size(); i != e; ++i)
2481 if (JTCases[i].first.HeaderBB == First)
2482 JTCases[i].first.HeaderBB = Last;
2484 // Update BitTestCases.
2485 for (unsigned i = 0, e = BitTestCases.size(); i != e; ++i)
2486 if (BitTestCases[i].Parent == First)
2487 BitTestCases[i].Parent = Last;
2490 void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) {
2491 MachineBasicBlock *SwitchMBB = FuncInfo.MBB;
2493 // Figure out which block is immediately after the current one.
2494 MachineBasicBlock *NextBlock = 0;
2495 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()];
2497 // If there is only the default destination, branch to it if it is not the
2498 // next basic block. Otherwise, just fall through.
2499 if (!SI.getNumCases()) {
2500 // Update machine-CFG edges.
2502 // If this is not a fall-through branch, emit the branch.
2503 SwitchMBB->addSuccessor(Default);
2504 if (Default != NextBlock)
2505 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
2506 MVT::Other, getControlRoot(),
2507 DAG.getBasicBlock(Default)));
2512 // If there are any non-default case statements, create a vector of Cases
2513 // representing each one, and sort the vector so that we can efficiently
2514 // create a binary search tree from them.
2516 size_t numCmps = Clusterify(Cases, SI);
2517 DEBUG(dbgs() << "Clusterify finished. Total clusters: " << Cases.size()
2518 << ". Total compares: " << numCmps << '\n');
2521 // Get the Value to be switched on and default basic blocks, which will be
2522 // inserted into CaseBlock records, representing basic blocks in the binary
2524 const Value *SV = SI.getCondition();
2526 // Push the initial CaseRec onto the worklist
2527 CaseRecVector WorkList;
2528 WorkList.push_back(CaseRec(SwitchMBB,0,0,
2529 CaseRange(Cases.begin(),Cases.end())));
2531 while (!WorkList.empty()) {
2532 // Grab a record representing a case range to process off the worklist
2533 CaseRec CR = WorkList.back();
2534 WorkList.pop_back();
2536 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2539 // If the range has few cases (two or less) emit a series of specific
2541 if (handleSmallSwitchRange(CR, WorkList, SV, Default, SwitchMBB))
2544 // If the switch has more than 5 blocks, and at least 40% dense, and the
2545 // target supports indirect branches, then emit a jump table rather than
2546 // lowering the switch to a binary tree of conditional branches.
2547 if (handleJTSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2550 // Emit binary tree. We need to pick a pivot, and push left and right ranges
2551 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call.
2552 handleBTSplitSwitchCase(CR, WorkList, SV, Default, SwitchMBB);
2556 void SelectionDAGBuilder::visitIndirectBr(const IndirectBrInst &I) {
2557 MachineBasicBlock *IndirectBrMBB = FuncInfo.MBB;
2559 // Update machine-CFG edges with unique successors.
2560 SmallVector<BasicBlock*, 32> succs;
2561 succs.reserve(I.getNumSuccessors());
2562 for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i)
2563 succs.push_back(I.getSuccessor(i));
2564 array_pod_sort(succs.begin(), succs.end());
2565 succs.erase(std::unique(succs.begin(), succs.end()), succs.end());
2566 for (unsigned i = 0, e = succs.size(); i != e; ++i) {
2567 MachineBasicBlock *Succ = FuncInfo.MBBMap[succs[i]];
2568 addSuccessorWithWeight(IndirectBrMBB, Succ);
2571 DAG.setRoot(DAG.getNode(ISD::BRIND, getCurDebugLoc(),
2572 MVT::Other, getControlRoot(),
2573 getValue(I.getAddress())));
2576 void SelectionDAGBuilder::visitFSub(const User &I) {
2577 // -0.0 - X --> fneg
2578 Type *Ty = I.getType();
2579 if (isa<Constant>(I.getOperand(0)) &&
2580 I.getOperand(0) == ConstantFP::getZeroValueForNegation(Ty)) {
2581 SDValue Op2 = getValue(I.getOperand(1));
2582 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(),
2583 Op2.getValueType(), Op2));
2587 visitBinary(I, ISD::FSUB);
2590 void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) {
2591 SDValue Op1 = getValue(I.getOperand(0));
2592 SDValue Op2 = getValue(I.getOperand(1));
2593 setValue(&I, DAG.getNode(OpCode, getCurDebugLoc(),
2594 Op1.getValueType(), Op1, Op2));
2597 void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) {
2598 SDValue Op1 = getValue(I.getOperand(0));
2599 SDValue Op2 = getValue(I.getOperand(1));
2601 MVT ShiftTy = TLI.getShiftAmountTy(Op2.getValueType());
2603 // Coerce the shift amount to the right type if we can.
2604 if (!I.getType()->isVectorTy() && Op2.getValueType() != ShiftTy) {
2605 unsigned ShiftSize = ShiftTy.getSizeInBits();
2606 unsigned Op2Size = Op2.getValueType().getSizeInBits();
2607 DebugLoc DL = getCurDebugLoc();
2609 // If the operand is smaller than the shift count type, promote it.
2610 if (ShiftSize > Op2Size)
2611 Op2 = DAG.getNode(ISD::ZERO_EXTEND, DL, ShiftTy, Op2);
2613 // If the operand is larger than the shift count type but the shift
2614 // count type has enough bits to represent any shift value, truncate
2615 // it now. This is a common case and it exposes the truncate to
2616 // optimization early.
2617 else if (ShiftSize >= Log2_32_Ceil(Op2.getValueType().getSizeInBits()))
2618 Op2 = DAG.getNode(ISD::TRUNCATE, DL, ShiftTy, Op2);
2619 // Otherwise we'll need to temporarily settle for some other convenient
2620 // type. Type legalization will make adjustments once the shiftee is split.
2622 Op2 = DAG.getZExtOrTrunc(Op2, DL, MVT::i32);
2625 setValue(&I, DAG.getNode(Opcode, getCurDebugLoc(),
2626 Op1.getValueType(), Op1, Op2));
2629 void SelectionDAGBuilder::visitSDiv(const User &I) {
2630 SDValue Op1 = getValue(I.getOperand(0));
2631 SDValue Op2 = getValue(I.getOperand(1));
2633 // Turn exact SDivs into multiplications.
2634 // FIXME: This should be in DAGCombiner, but it doesn't have access to the
2636 if (isa<BinaryOperator>(&I) && cast<BinaryOperator>(&I)->isExact() &&
2637 !isa<ConstantSDNode>(Op1) &&
2638 isa<ConstantSDNode>(Op2) && !cast<ConstantSDNode>(Op2)->isNullValue())
2639 setValue(&I, TLI.BuildExactSDIV(Op1, Op2, getCurDebugLoc(), DAG));
2641 setValue(&I, DAG.getNode(ISD::SDIV, getCurDebugLoc(), Op1.getValueType(),
2645 void SelectionDAGBuilder::visitICmp(const User &I) {
2646 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2647 if (const ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2648 predicate = IC->getPredicate();
2649 else if (const ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2650 predicate = ICmpInst::Predicate(IC->getPredicate());
2651 SDValue Op1 = getValue(I.getOperand(0));
2652 SDValue Op2 = getValue(I.getOperand(1));
2653 ISD::CondCode Opcode = getICmpCondCode(predicate);
2655 EVT DestVT = TLI.getValueType(I.getType());
2656 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Opcode));
2659 void SelectionDAGBuilder::visitFCmp(const User &I) {
2660 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2661 if (const FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2662 predicate = FC->getPredicate();
2663 else if (const ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2664 predicate = FCmpInst::Predicate(FC->getPredicate());
2665 SDValue Op1 = getValue(I.getOperand(0));
2666 SDValue Op2 = getValue(I.getOperand(1));
2667 ISD::CondCode Condition = getFCmpCondCode(predicate);
2668 if (TM.Options.NoNaNsFPMath)
2669 Condition = getFCmpCodeWithoutNaN(Condition);
2670 EVT DestVT = TLI.getValueType(I.getType());
2671 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Condition));
2674 void SelectionDAGBuilder::visitSelect(const User &I) {
2675 SmallVector<EVT, 4> ValueVTs;
2676 ComputeValueVTs(TLI, I.getType(), ValueVTs);
2677 unsigned NumValues = ValueVTs.size();
2678 if (NumValues == 0) return;
2680 SmallVector<SDValue, 4> Values(NumValues);
2681 SDValue Cond = getValue(I.getOperand(0));
2682 SDValue TrueVal = getValue(I.getOperand(1));
2683 SDValue FalseVal = getValue(I.getOperand(2));
2684 ISD::NodeType OpCode = Cond.getValueType().isVector() ?
2685 ISD::VSELECT : ISD::SELECT;
2687 for (unsigned i = 0; i != NumValues; ++i)
2688 Values[i] = DAG.getNode(OpCode, getCurDebugLoc(),
2689 TrueVal.getNode()->getValueType(TrueVal.getResNo()+i),
2691 SDValue(TrueVal.getNode(),
2692 TrueVal.getResNo() + i),
2693 SDValue(FalseVal.getNode(),
2694 FalseVal.getResNo() + i));
2696 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2697 DAG.getVTList(&ValueVTs[0], NumValues),
2698 &Values[0], NumValues));
2701 void SelectionDAGBuilder::visitTrunc(const User &I) {
2702 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2703 SDValue N = getValue(I.getOperand(0));
2704 EVT DestVT = TLI.getValueType(I.getType());
2705 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), DestVT, N));
2708 void SelectionDAGBuilder::visitZExt(const User &I) {
2709 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2710 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2711 SDValue N = getValue(I.getOperand(0));
2712 EVT DestVT = TLI.getValueType(I.getType());
2713 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(), DestVT, N));
2716 void SelectionDAGBuilder::visitSExt(const User &I) {
2717 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2718 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2719 SDValue N = getValue(I.getOperand(0));
2720 EVT DestVT = TLI.getValueType(I.getType());
2721 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurDebugLoc(), DestVT, N));
2724 void SelectionDAGBuilder::visitFPTrunc(const User &I) {
2725 // FPTrunc is never a no-op cast, no need to check
2726 SDValue N = getValue(I.getOperand(0));
2727 EVT DestVT = TLI.getValueType(I.getType());
2728 setValue(&I, DAG.getNode(ISD::FP_ROUND, getCurDebugLoc(),
2730 DAG.getTargetConstant(0, TLI.getPointerTy())));
2733 void SelectionDAGBuilder::visitFPExt(const User &I){
2734 // FPExt is never a no-op cast, no need to check
2735 SDValue N = getValue(I.getOperand(0));
2736 EVT DestVT = TLI.getValueType(I.getType());
2737 setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurDebugLoc(), DestVT, N));
2740 void SelectionDAGBuilder::visitFPToUI(const User &I) {
2741 // FPToUI is never a no-op cast, no need to check
2742 SDValue N = getValue(I.getOperand(0));
2743 EVT DestVT = TLI.getValueType(I.getType());
2744 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurDebugLoc(), DestVT, N));
2747 void SelectionDAGBuilder::visitFPToSI(const User &I) {
2748 // FPToSI is never a no-op cast, no need to check
2749 SDValue N = getValue(I.getOperand(0));
2750 EVT DestVT = TLI.getValueType(I.getType());
2751 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurDebugLoc(), DestVT, N));
2754 void SelectionDAGBuilder::visitUIToFP(const User &I) {
2755 // UIToFP is never a no-op cast, no need to check
2756 SDValue N = getValue(I.getOperand(0));
2757 EVT DestVT = TLI.getValueType(I.getType());
2758 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurDebugLoc(), DestVT, N));
2761 void SelectionDAGBuilder::visitSIToFP(const User &I){
2762 // SIToFP is never a no-op cast, no need to check
2763 SDValue N = getValue(I.getOperand(0));
2764 EVT DestVT = TLI.getValueType(I.getType());
2765 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurDebugLoc(), DestVT, N));
2768 void SelectionDAGBuilder::visitPtrToInt(const User &I) {
2769 // What to do depends on the size of the integer and the size of the pointer.
2770 // We can either truncate, zero extend, or no-op, accordingly.
2771 SDValue N = getValue(I.getOperand(0));
2772 EVT DestVT = TLI.getValueType(I.getType());
2773 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT));
2776 void SelectionDAGBuilder::visitIntToPtr(const User &I) {
2777 // What to do depends on the size of the integer and the size of the pointer.
2778 // We can either truncate, zero extend, or no-op, accordingly.
2779 SDValue N = getValue(I.getOperand(0));
2780 EVT DestVT = TLI.getValueType(I.getType());
2781 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT));
2784 void SelectionDAGBuilder::visitBitCast(const User &I) {
2785 SDValue N = getValue(I.getOperand(0));
2786 EVT DestVT = TLI.getValueType(I.getType());
2788 // BitCast assures us that source and destination are the same size so this is
2789 // either a BITCAST or a no-op.
2790 if (DestVT != N.getValueType())
2791 setValue(&I, DAG.getNode(ISD::BITCAST, getCurDebugLoc(),
2792 DestVT, N)); // convert types.
2794 setValue(&I, N); // noop cast.
2797 void SelectionDAGBuilder::visitInsertElement(const User &I) {
2798 SDValue InVec = getValue(I.getOperand(0));
2799 SDValue InVal = getValue(I.getOperand(1));
2800 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2802 getValue(I.getOperand(2)));
2803 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurDebugLoc(),
2804 TLI.getValueType(I.getType()),
2805 InVec, InVal, InIdx));
2808 void SelectionDAGBuilder::visitExtractElement(const User &I) {
2809 SDValue InVec = getValue(I.getOperand(0));
2810 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2812 getValue(I.getOperand(1)));
2813 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2814 TLI.getValueType(I.getType()), InVec, InIdx));
2817 // Utility for visitShuffleVector - Return true if every element in Mask,
2818 // beginning from position Pos and ending in Pos+Size, falls within the
2819 // specified sequential range [L, L+Pos). or is undef.
2820 static bool isSequentialInRange(const SmallVectorImpl<int> &Mask,
2821 unsigned Pos, unsigned Size, int Low) {
2822 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
2823 if (Mask[i] >= 0 && Mask[i] != Low)
2828 void SelectionDAGBuilder::visitShuffleVector(const User &I) {
2829 SDValue Src1 = getValue(I.getOperand(0));
2830 SDValue Src2 = getValue(I.getOperand(1));
2832 SmallVector<int, 8> Mask;
2833 ShuffleVectorInst::getShuffleMask(cast<Constant>(I.getOperand(2)), Mask);
2834 unsigned MaskNumElts = Mask.size();
2836 EVT VT = TLI.getValueType(I.getType());
2837 EVT SrcVT = Src1.getValueType();
2838 unsigned SrcNumElts = SrcVT.getVectorNumElements();
2840 if (SrcNumElts == MaskNumElts) {
2841 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2846 // Normalize the shuffle vector since mask and vector length don't match.
2847 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) {
2848 // Mask is longer than the source vectors and is a multiple of the source
2849 // vectors. We can use concatenate vector to make the mask and vectors
2851 if (SrcNumElts*2 == MaskNumElts) {
2852 // First check for Src1 in low and Src2 in high
2853 if (isSequentialInRange(Mask, 0, SrcNumElts, 0) &&
2854 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, SrcNumElts)) {
2855 // The shuffle is concatenating two vectors together.
2856 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurDebugLoc(),
2860 // Then check for Src2 in low and Src1 in high
2861 if (isSequentialInRange(Mask, 0, SrcNumElts, SrcNumElts) &&
2862 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, 0)) {
2863 // The shuffle is concatenating two vectors together.
2864 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurDebugLoc(),
2870 // Pad both vectors with undefs to make them the same length as the mask.
2871 unsigned NumConcat = MaskNumElts / SrcNumElts;
2872 bool Src1U = Src1.getOpcode() == ISD::UNDEF;
2873 bool Src2U = Src2.getOpcode() == ISD::UNDEF;
2874 SDValue UndefVal = DAG.getUNDEF(SrcVT);
2876 SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal);
2877 SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal);
2881 Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2882 getCurDebugLoc(), VT,
2883 &MOps1[0], NumConcat);
2884 Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2885 getCurDebugLoc(), VT,
2886 &MOps2[0], NumConcat);
2888 // Readjust mask for new input vector length.
2889 SmallVector<int, 8> MappedOps;
2890 for (unsigned i = 0; i != MaskNumElts; ++i) {
2892 if (Idx >= (int)SrcNumElts)
2893 Idx -= SrcNumElts - MaskNumElts;
2894 MappedOps.push_back(Idx);
2897 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2902 if (SrcNumElts > MaskNumElts) {
2903 // Analyze the access pattern of the vector to see if we can extract
2904 // two subvectors and do the shuffle. The analysis is done by calculating
2905 // the range of elements the mask access on both vectors.
2906 int MinRange[2] = { static_cast<int>(SrcNumElts),
2907 static_cast<int>(SrcNumElts)};
2908 int MaxRange[2] = {-1, -1};
2910 for (unsigned i = 0; i != MaskNumElts; ++i) {
2916 if (Idx >= (int)SrcNumElts) {
2920 if (Idx > MaxRange[Input])
2921 MaxRange[Input] = Idx;
2922 if (Idx < MinRange[Input])
2923 MinRange[Input] = Idx;
2926 // Check if the access is smaller than the vector size and can we find
2927 // a reasonable extract index.
2928 int RangeUse[2] = { -1, -1 }; // 0 = Unused, 1 = Extract, -1 = Can not
2930 int StartIdx[2]; // StartIdx to extract from
2931 for (unsigned Input = 0; Input < 2; ++Input) {
2932 if (MinRange[Input] >= (int)SrcNumElts && MaxRange[Input] < 0) {
2933 RangeUse[Input] = 0; // Unused
2934 StartIdx[Input] = 0;
2938 // Find a good start index that is a multiple of the mask length. Then
2939 // see if the rest of the elements are in range.
2940 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts;
2941 if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts &&
2942 StartIdx[Input] + MaskNumElts <= SrcNumElts)
2943 RangeUse[Input] = 1; // Extract from a multiple of the mask length.
2946 if (RangeUse[0] == 0 && RangeUse[1] == 0) {
2947 setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used.
2950 if (RangeUse[0] >= 0 && RangeUse[1] >= 0) {
2951 // Extract appropriate subvector and generate a vector shuffle
2952 for (unsigned Input = 0; Input < 2; ++Input) {
2953 SDValue &Src = Input == 0 ? Src1 : Src2;
2954 if (RangeUse[Input] == 0)
2955 Src = DAG.getUNDEF(VT);
2957 Src = DAG.getNode(ISD::EXTRACT_SUBVECTOR, getCurDebugLoc(), VT,
2958 Src, DAG.getIntPtrConstant(StartIdx[Input]));
2961 // Calculate new mask.
2962 SmallVector<int, 8> MappedOps;
2963 for (unsigned i = 0; i != MaskNumElts; ++i) {
2966 if (Idx < (int)SrcNumElts)
2969 Idx -= SrcNumElts + StartIdx[1] - MaskNumElts;
2971 MappedOps.push_back(Idx);
2974 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2980 // We can't use either concat vectors or extract subvectors so fall back to
2981 // replacing the shuffle with extract and build vector.
2982 // to insert and build vector.
2983 EVT EltVT = VT.getVectorElementType();
2984 EVT PtrVT = TLI.getPointerTy();
2985 SmallVector<SDValue,8> Ops;
2986 for (unsigned i = 0; i != MaskNumElts; ++i) {
2991 Res = DAG.getUNDEF(EltVT);
2993 SDValue &Src = Idx < (int)SrcNumElts ? Src1 : Src2;
2994 if (Idx >= (int)SrcNumElts) Idx -= SrcNumElts;
2996 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2997 EltVT, Src, DAG.getConstant(Idx, PtrVT));
3003 setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
3004 VT, &Ops[0], Ops.size()));
3007 void SelectionDAGBuilder::visitInsertValue(const InsertValueInst &I) {
3008 const Value *Op0 = I.getOperand(0);
3009 const Value *Op1 = I.getOperand(1);
3010 Type *AggTy = I.getType();
3011 Type *ValTy = Op1->getType();
3012 bool IntoUndef = isa<UndefValue>(Op0);
3013 bool FromUndef = isa<UndefValue>(Op1);
3015 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
3017 SmallVector<EVT, 4> AggValueVTs;
3018 ComputeValueVTs(TLI, AggTy, AggValueVTs);
3019 SmallVector<EVT, 4> ValValueVTs;
3020 ComputeValueVTs(TLI, ValTy, ValValueVTs);
3022 unsigned NumAggValues = AggValueVTs.size();
3023 unsigned NumValValues = ValValueVTs.size();
3024 SmallVector<SDValue, 4> Values(NumAggValues);
3026 SDValue Agg = getValue(Op0);
3028 // Copy the beginning value(s) from the original aggregate.
3029 for (; i != LinearIndex; ++i)
3030 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
3031 SDValue(Agg.getNode(), Agg.getResNo() + i);
3032 // Copy values from the inserted value(s).
3034 SDValue Val = getValue(Op1);
3035 for (; i != LinearIndex + NumValValues; ++i)
3036 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) :
3037 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex);
3039 // Copy remaining value(s) from the original aggregate.
3040 for (; i != NumAggValues; ++i)
3041 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
3042 SDValue(Agg.getNode(), Agg.getResNo() + i);
3044 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
3045 DAG.getVTList(&AggValueVTs[0], NumAggValues),
3046 &Values[0], NumAggValues));
3049 void SelectionDAGBuilder::visitExtractValue(const ExtractValueInst &I) {
3050 const Value *Op0 = I.getOperand(0);
3051 Type *AggTy = Op0->getType();
3052 Type *ValTy = I.getType();
3053 bool OutOfUndef = isa<UndefValue>(Op0);
3055 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
3057 SmallVector<EVT, 4> ValValueVTs;
3058 ComputeValueVTs(TLI, ValTy, ValValueVTs);
3060 unsigned NumValValues = ValValueVTs.size();
3062 // Ignore a extractvalue that produces an empty object
3063 if (!NumValValues) {
3064 setValue(&I, DAG.getUNDEF(MVT(MVT::Other)));
3068 SmallVector<SDValue, 4> Values(NumValValues);
3070 SDValue Agg = getValue(Op0);
3071 // Copy out the selected value(s).
3072 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i)
3073 Values[i - LinearIndex] =
3075 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) :
3076 SDValue(Agg.getNode(), Agg.getResNo() + i);
3078 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
3079 DAG.getVTList(&ValValueVTs[0], NumValValues),
3080 &Values[0], NumValValues));
3083 void SelectionDAGBuilder::visitGetElementPtr(const User &I) {
3084 SDValue N = getValue(I.getOperand(0));
3085 // Note that the pointer operand may be a vector of pointers. Take the scalar
3086 // element which holds a pointer.
3087 Type *Ty = I.getOperand(0)->getType()->getScalarType();
3089 for (GetElementPtrInst::const_op_iterator OI = I.op_begin()+1, E = I.op_end();
3091 const Value *Idx = *OI;
3092 if (StructType *StTy = dyn_cast<StructType>(Ty)) {
3093 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
3096 uint64_t Offset = TD->getStructLayout(StTy)->getElementOffset(Field);
3097 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
3098 DAG.getIntPtrConstant(Offset));
3101 Ty = StTy->getElementType(Field);
3103 Ty = cast<SequentialType>(Ty)->getElementType();
3105 // If this is a constant subscript, handle it quickly.
3106 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
3107 if (CI->isZero()) continue;
3109 TD->getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
3111 EVT PTy = TLI.getPointerTy();
3112 unsigned PtrBits = PTy.getSizeInBits();
3114 OffsVal = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
3116 DAG.getConstant(Offs, MVT::i64));
3118 OffsVal = DAG.getIntPtrConstant(Offs);
3120 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
3125 // N = N + Idx * ElementSize;
3126 APInt ElementSize = APInt(TLI.getPointerTy().getSizeInBits(),
3127 TD->getTypeAllocSize(Ty));
3128 SDValue IdxN = getValue(Idx);
3130 // If the index is smaller or larger than intptr_t, truncate or extend
3132 IdxN = DAG.getSExtOrTrunc(IdxN, getCurDebugLoc(), N.getValueType());
3134 // If this is a multiply by a power of two, turn it into a shl
3135 // immediately. This is a very common case.
3136 if (ElementSize != 1) {
3137 if (ElementSize.isPowerOf2()) {
3138 unsigned Amt = ElementSize.logBase2();
3139 IdxN = DAG.getNode(ISD::SHL, getCurDebugLoc(),
3140 N.getValueType(), IdxN,
3141 DAG.getConstant(Amt, IdxN.getValueType()));
3143 SDValue Scale = DAG.getConstant(ElementSize, TLI.getPointerTy());
3144 IdxN = DAG.getNode(ISD::MUL, getCurDebugLoc(),
3145 N.getValueType(), IdxN, Scale);
3149 N = DAG.getNode(ISD::ADD, getCurDebugLoc(),
3150 N.getValueType(), N, IdxN);
3157 void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) {
3158 // If this is a fixed sized alloca in the entry block of the function,
3159 // allocate it statically on the stack.
3160 if (FuncInfo.StaticAllocaMap.count(&I))
3161 return; // getValue will auto-populate this.
3163 Type *Ty = I.getAllocatedType();
3164 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
3166 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty),
3169 SDValue AllocSize = getValue(I.getArraySize());
3171 EVT IntPtr = TLI.getPointerTy();
3172 if (AllocSize.getValueType() != IntPtr)
3173 AllocSize = DAG.getZExtOrTrunc(AllocSize, getCurDebugLoc(), IntPtr);
3175 AllocSize = DAG.getNode(ISD::MUL, getCurDebugLoc(), IntPtr,
3177 DAG.getConstant(TySize, IntPtr));
3179 // Handle alignment. If the requested alignment is less than or equal to
3180 // the stack alignment, ignore it. If the size is greater than or equal to
3181 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
3182 unsigned StackAlign = TM.getFrameLowering()->getStackAlignment();
3183 if (Align <= StackAlign)
3186 // Round the size of the allocation up to the stack alignment size
3187 // by add SA-1 to the size.
3188 AllocSize = DAG.getNode(ISD::ADD, getCurDebugLoc(),
3189 AllocSize.getValueType(), AllocSize,
3190 DAG.getIntPtrConstant(StackAlign-1));
3192 // Mask out the low bits for alignment purposes.
3193 AllocSize = DAG.getNode(ISD::AND, getCurDebugLoc(),
3194 AllocSize.getValueType(), AllocSize,
3195 DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1)));
3197 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) };
3198 SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other);
3199 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, getCurDebugLoc(),
3202 DAG.setRoot(DSA.getValue(1));
3204 // Inform the Frame Information that we have just allocated a variable-sized
3206 FuncInfo.MF->getFrameInfo()->CreateVariableSizedObject(Align ? Align : 1);
3209 void SelectionDAGBuilder::visitLoad(const LoadInst &I) {
3211 return visitAtomicLoad(I);
3213 const Value *SV = I.getOperand(0);
3214 SDValue Ptr = getValue(SV);
3216 Type *Ty = I.getType();
3218 bool isVolatile = I.isVolatile();
3219 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
3220 bool isInvariant = I.getMetadata("invariant.load") != 0;
3221 unsigned Alignment = I.getAlignment();
3222 const MDNode *TBAAInfo = I.getMetadata(LLVMContext::MD_tbaa);
3223 const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range);
3225 SmallVector<EVT, 4> ValueVTs;
3226 SmallVector<uint64_t, 4> Offsets;
3227 ComputeValueVTs(TLI, Ty, ValueVTs, &Offsets);
3228 unsigned NumValues = ValueVTs.size();
3233 bool ConstantMemory = false;
3234 if (I.isVolatile() || NumValues > MaxParallelChains)
3235 // Serialize volatile loads with other side effects.
3237 else if (AA->pointsToConstantMemory(
3238 AliasAnalysis::Location(SV, AA->getTypeStoreSize(Ty), TBAAInfo))) {
3239 // Do not serialize (non-volatile) loads of constant memory with anything.
3240 Root = DAG.getEntryNode();
3241 ConstantMemory = true;
3243 // Do not serialize non-volatile loads against each other.
3244 Root = DAG.getRoot();
3247 SmallVector<SDValue, 4> Values(NumValues);
3248 SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains),
3250 EVT PtrVT = Ptr.getValueType();
3251 unsigned ChainI = 0;
3252 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3253 // Serializing loads here may result in excessive register pressure, and
3254 // TokenFactor places arbitrary choke points on the scheduler. SD scheduling
3255 // could recover a bit by hoisting nodes upward in the chain by recognizing
3256 // they are side-effect free or do not alias. The optimizer should really
3257 // avoid this case by converting large object/array copies to llvm.memcpy
3258 // (MaxParallelChains should always remain as failsafe).
3259 if (ChainI == MaxParallelChains) {
3260 assert(PendingLoads.empty() && "PendingLoads must be serialized first");
3261 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3262 MVT::Other, &Chains[0], ChainI);
3266 SDValue A = DAG.getNode(ISD::ADD, getCurDebugLoc(),
3268 DAG.getConstant(Offsets[i], PtrVT));
3269 SDValue L = DAG.getLoad(ValueVTs[i], getCurDebugLoc(), Root,
3270 A, MachinePointerInfo(SV, Offsets[i]), isVolatile,
3271 isNonTemporal, isInvariant, Alignment, TBAAInfo,
3275 Chains[ChainI] = L.getValue(1);
3278 if (!ConstantMemory) {
3279 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3280 MVT::Other, &Chains[0], ChainI);
3284 PendingLoads.push_back(Chain);
3287 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
3288 DAG.getVTList(&ValueVTs[0], NumValues),
3289 &Values[0], NumValues));
3292 void SelectionDAGBuilder::visitStore(const StoreInst &I) {
3294 return visitAtomicStore(I);
3296 const Value *SrcV = I.getOperand(0);
3297 const Value *PtrV = I.getOperand(1);
3299 SmallVector<EVT, 4> ValueVTs;
3300 SmallVector<uint64_t, 4> Offsets;
3301 ComputeValueVTs(TLI, SrcV->getType(), ValueVTs, &Offsets);
3302 unsigned NumValues = ValueVTs.size();
3306 // Get the lowered operands. Note that we do this after
3307 // checking if NumResults is zero, because with zero results
3308 // the operands won't have values in the map.
3309 SDValue Src = getValue(SrcV);
3310 SDValue Ptr = getValue(PtrV);
3312 SDValue Root = getRoot();
3313 SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains),
3315 EVT PtrVT = Ptr.getValueType();
3316 bool isVolatile = I.isVolatile();
3317 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
3318 unsigned Alignment = I.getAlignment();
3319 const MDNode *TBAAInfo = I.getMetadata(LLVMContext::MD_tbaa);
3321 unsigned ChainI = 0;
3322 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3323 // See visitLoad comments.
3324 if (ChainI == MaxParallelChains) {
3325 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3326 MVT::Other, &Chains[0], ChainI);
3330 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT, Ptr,
3331 DAG.getConstant(Offsets[i], PtrVT));
3332 SDValue St = DAG.getStore(Root, getCurDebugLoc(),
3333 SDValue(Src.getNode(), Src.getResNo() + i),
3334 Add, MachinePointerInfo(PtrV, Offsets[i]),
3335 isVolatile, isNonTemporal, Alignment, TBAAInfo);
3336 Chains[ChainI] = St;
3339 SDValue StoreNode = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3340 MVT::Other, &Chains[0], ChainI);
3342 AssignOrderingToNode(StoreNode.getNode());
3343 DAG.setRoot(StoreNode);
3346 static SDValue InsertFenceForAtomic(SDValue Chain, AtomicOrdering Order,
3347 SynchronizationScope Scope,
3348 bool Before, DebugLoc dl,
3350 const TargetLowering &TLI) {
3351 // Fence, if necessary
3353 if (Order == AcquireRelease || Order == SequentiallyConsistent)
3355 else if (Order == Acquire || Order == Monotonic)
3358 if (Order == AcquireRelease)
3360 else if (Order == Release || Order == Monotonic)
3365 Ops[1] = DAG.getConstant(Order, TLI.getPointerTy());
3366 Ops[2] = DAG.getConstant(Scope, TLI.getPointerTy());
3367 return DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops, 3);
3370 void SelectionDAGBuilder::visitAtomicCmpXchg(const AtomicCmpXchgInst &I) {
3371 DebugLoc dl = getCurDebugLoc();
3372 AtomicOrdering Order = I.getOrdering();
3373 SynchronizationScope Scope = I.getSynchScope();
3375 SDValue InChain = getRoot();
3377 if (TLI.getInsertFencesForAtomic())
3378 InChain = InsertFenceForAtomic(InChain, Order, Scope, true, dl,
3382 DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, dl,
3383 getValue(I.getCompareOperand()).getValueType().getSimpleVT(),
3385 getValue(I.getPointerOperand()),
3386 getValue(I.getCompareOperand()),
3387 getValue(I.getNewValOperand()),
3388 MachinePointerInfo(I.getPointerOperand()), 0 /* Alignment */,
3389 TLI.getInsertFencesForAtomic() ? Monotonic : Order,
3392 SDValue OutChain = L.getValue(1);
3394 if (TLI.getInsertFencesForAtomic())
3395 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3399 DAG.setRoot(OutChain);
3402 void SelectionDAGBuilder::visitAtomicRMW(const AtomicRMWInst &I) {
3403 DebugLoc dl = getCurDebugLoc();
3405 switch (I.getOperation()) {
3406 default: llvm_unreachable("Unknown atomicrmw operation");
3407 case AtomicRMWInst::Xchg: NT = ISD::ATOMIC_SWAP; break;
3408 case AtomicRMWInst::Add: NT = ISD::ATOMIC_LOAD_ADD; break;
3409 case AtomicRMWInst::Sub: NT = ISD::ATOMIC_LOAD_SUB; break;
3410 case AtomicRMWInst::And: NT = ISD::ATOMIC_LOAD_AND; break;
3411 case AtomicRMWInst::Nand: NT = ISD::ATOMIC_LOAD_NAND; break;
3412 case AtomicRMWInst::Or: NT = ISD::ATOMIC_LOAD_OR; break;
3413 case AtomicRMWInst::Xor: NT = ISD::ATOMIC_LOAD_XOR; break;
3414 case AtomicRMWInst::Max: NT = ISD::ATOMIC_LOAD_MAX; break;
3415 case AtomicRMWInst::Min: NT = ISD::ATOMIC_LOAD_MIN; break;
3416 case AtomicRMWInst::UMax: NT = ISD::ATOMIC_LOAD_UMAX; break;
3417 case AtomicRMWInst::UMin: NT = ISD::ATOMIC_LOAD_UMIN; break;
3419 AtomicOrdering Order = I.getOrdering();
3420 SynchronizationScope Scope = I.getSynchScope();
3422 SDValue InChain = getRoot();
3424 if (TLI.getInsertFencesForAtomic())
3425 InChain = InsertFenceForAtomic(InChain, Order, Scope, true, dl,
3429 DAG.getAtomic(NT, dl,
3430 getValue(I.getValOperand()).getValueType().getSimpleVT(),
3432 getValue(I.getPointerOperand()),
3433 getValue(I.getValOperand()),
3434 I.getPointerOperand(), 0 /* Alignment */,
3435 TLI.getInsertFencesForAtomic() ? Monotonic : Order,
3438 SDValue OutChain = L.getValue(1);
3440 if (TLI.getInsertFencesForAtomic())
3441 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3445 DAG.setRoot(OutChain);
3448 void SelectionDAGBuilder::visitFence(const FenceInst &I) {
3449 DebugLoc dl = getCurDebugLoc();
3452 Ops[1] = DAG.getConstant(I.getOrdering(), TLI.getPointerTy());
3453 Ops[2] = DAG.getConstant(I.getSynchScope(), TLI.getPointerTy());
3454 DAG.setRoot(DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops, 3));
3457 void SelectionDAGBuilder::visitAtomicLoad(const LoadInst &I) {
3458 DebugLoc dl = getCurDebugLoc();
3459 AtomicOrdering Order = I.getOrdering();
3460 SynchronizationScope Scope = I.getSynchScope();
3462 SDValue InChain = getRoot();
3464 EVT VT = EVT::getEVT(I.getType());
3466 if (I.getAlignment() * 8 < VT.getSizeInBits())
3467 report_fatal_error("Cannot generate unaligned atomic load");
3470 DAG.getAtomic(ISD::ATOMIC_LOAD, dl, VT, VT, InChain,
3471 getValue(I.getPointerOperand()),
3472 I.getPointerOperand(), I.getAlignment(),
3473 TLI.getInsertFencesForAtomic() ? Monotonic : Order,
3476 SDValue OutChain = L.getValue(1);
3478 if (TLI.getInsertFencesForAtomic())
3479 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3483 DAG.setRoot(OutChain);
3486 void SelectionDAGBuilder::visitAtomicStore(const StoreInst &I) {
3487 DebugLoc dl = getCurDebugLoc();
3489 AtomicOrdering Order = I.getOrdering();
3490 SynchronizationScope Scope = I.getSynchScope();
3492 SDValue InChain = getRoot();
3494 EVT VT = EVT::getEVT(I.getValueOperand()->getType());
3496 if (I.getAlignment() * 8 < VT.getSizeInBits())
3497 report_fatal_error("Cannot generate unaligned atomic store");
3499 if (TLI.getInsertFencesForAtomic())
3500 InChain = InsertFenceForAtomic(InChain, Order, Scope, true, dl,
3504 DAG.getAtomic(ISD::ATOMIC_STORE, dl, VT,
3506 getValue(I.getPointerOperand()),
3507 getValue(I.getValueOperand()),
3508 I.getPointerOperand(), I.getAlignment(),
3509 TLI.getInsertFencesForAtomic() ? Monotonic : Order,
3512 if (TLI.getInsertFencesForAtomic())
3513 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3516 DAG.setRoot(OutChain);
3519 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
3521 void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I,
3522 unsigned Intrinsic) {
3523 bool HasChain = !I.doesNotAccessMemory();
3524 bool OnlyLoad = HasChain && I.onlyReadsMemory();
3526 // Build the operand list.
3527 SmallVector<SDValue, 8> Ops;
3528 if (HasChain) { // If this intrinsic has side-effects, chainify it.
3530 // We don't need to serialize loads against other loads.
3531 Ops.push_back(DAG.getRoot());
3533 Ops.push_back(getRoot());
3537 // Info is set by getTgtMemInstrinsic
3538 TargetLowering::IntrinsicInfo Info;
3539 bool IsTgtIntrinsic = TLI.getTgtMemIntrinsic(Info, I, Intrinsic);
3541 // Add the intrinsic ID as an integer operand if it's not a target intrinsic.
3542 if (!IsTgtIntrinsic || Info.opc == ISD::INTRINSIC_VOID ||
3543 Info.opc == ISD::INTRINSIC_W_CHAIN)
3544 Ops.push_back(DAG.getTargetConstant(Intrinsic, TLI.getPointerTy()));
3546 // Add all operands of the call to the operand list.
3547 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) {
3548 SDValue Op = getValue(I.getArgOperand(i));
3552 SmallVector<EVT, 4> ValueVTs;
3553 ComputeValueVTs(TLI, I.getType(), ValueVTs);
3556 ValueVTs.push_back(MVT::Other);
3558 SDVTList VTs = DAG.getVTList(ValueVTs.data(), ValueVTs.size());
3562 if (IsTgtIntrinsic) {
3563 // This is target intrinsic that touches memory
3564 Result = DAG.getMemIntrinsicNode(Info.opc, getCurDebugLoc(),
3565 VTs, &Ops[0], Ops.size(),
3567 MachinePointerInfo(Info.ptrVal, Info.offset),
3568 Info.align, Info.vol,
3569 Info.readMem, Info.writeMem);
3570 } else if (!HasChain) {
3571 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurDebugLoc(),
3572 VTs, &Ops[0], Ops.size());
3573 } else if (!I.getType()->isVoidTy()) {
3574 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurDebugLoc(),
3575 VTs, &Ops[0], Ops.size());
3577 Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurDebugLoc(),
3578 VTs, &Ops[0], Ops.size());
3582 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1);
3584 PendingLoads.push_back(Chain);
3589 if (!I.getType()->isVoidTy()) {
3590 if (VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
3591 EVT VT = TLI.getValueType(PTy);
3592 Result = DAG.getNode(ISD::BITCAST, getCurDebugLoc(), VT, Result);
3595 setValue(&I, Result);
3597 // Assign order to result here. If the intrinsic does not produce a result,
3598 // it won't be mapped to a SDNode and visit() will not assign it an order
3601 AssignOrderingToNode(Result.getNode());
3605 /// GetSignificand - Get the significand and build it into a floating-point
3606 /// number with exponent of 1:
3608 /// Op = (Op & 0x007fffff) | 0x3f800000;
3610 /// where Op is the hexidecimal representation of floating point value.
3612 GetSignificand(SelectionDAG &DAG, SDValue Op, DebugLoc dl) {
3613 SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3614 DAG.getConstant(0x007fffff, MVT::i32));
3615 SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1,
3616 DAG.getConstant(0x3f800000, MVT::i32));
3617 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, t2);
3620 /// GetExponent - Get the exponent:
3622 /// (float)(int)(((Op & 0x7f800000) >> 23) - 127);
3624 /// where Op is the hexidecimal representation of floating point value.
3626 GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI,
3628 SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3629 DAG.getConstant(0x7f800000, MVT::i32));
3630 SDValue t1 = DAG.getNode(ISD::SRL, dl, MVT::i32, t0,
3631 DAG.getConstant(23, TLI.getPointerTy()));
3632 SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1,
3633 DAG.getConstant(127, MVT::i32));
3634 return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2);
3637 /// getF32Constant - Get 32-bit floating point constant.
3639 getF32Constant(SelectionDAG &DAG, unsigned Flt) {
3640 return DAG.getConstantFP(APFloat(APInt(32, Flt)), MVT::f32);
3643 /// visitExp - Lower an exp intrinsic. Handles the special sequences for
3644 /// limited-precision mode.
3646 SelectionDAGBuilder::visitExp(const CallInst &I) {
3648 DebugLoc dl = getCurDebugLoc();
3650 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3651 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3652 SDValue Op = getValue(I.getArgOperand(0));
3654 // Put the exponent in the right bit position for later addition to the
3657 // #define LOG2OFe 1.4426950f
3658 // IntegerPartOfX = ((int32_t)(X * LOG2OFe));
3659 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3660 getF32Constant(DAG, 0x3fb8aa3b));
3661 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3663 // FractionalPartOfX = (X * LOG2OFe) - (float)IntegerPartOfX;
3664 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3665 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3667 // IntegerPartOfX <<= 23;
3668 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3669 DAG.getConstant(23, TLI.getPointerTy()));
3671 if (LimitFloatPrecision <= 6) {
3672 // For floating-point precision of 6:
3674 // TwoToFractionalPartOfX =
3676 // (0.735607626f + 0.252464424f * x) * x;
3678 // error 0.0144103317, which is 6 bits
3679 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3680 getF32Constant(DAG, 0x3e814304));
3681 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3682 getF32Constant(DAG, 0x3f3c50c8));
3683 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3684 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3685 getF32Constant(DAG, 0x3f7f5e7e));
3686 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BITCAST, dl,MVT::i32, t5);
3688 // Add the exponent into the result in integer domain.
3689 SDValue t6 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3690 TwoToFracPartOfX, IntegerPartOfX);
3692 result = DAG.getNode(ISD::BITCAST, dl, MVT::f32, t6);
3693 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3694 // For floating-point precision of 12:
3696 // TwoToFractionalPartOfX =
3699 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3701 // 0.000107046256 error, which is 13 to 14 bits
3702 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3703 getF32Constant(DAG, 0x3da235e3));
3704 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3705 getF32Constant(DAG, 0x3e65b8f3));
3706 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3707 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3708 getF32Constant(DAG, 0x3f324b07));
3709 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3710 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3711 getF32Constant(DAG, 0x3f7ff8fd));
3712 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BITCAST, dl,MVT::i32, t7);
3714 // Add the exponent into the result in integer domain.
3715 SDValue t8 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3716 TwoToFracPartOfX, IntegerPartOfX);
3718 result = DAG.getNode(ISD::BITCAST, dl, MVT::f32, t8);
3719 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3720 // For floating-point precision of 18:
3722 // TwoToFractionalPartOfX =
3726 // (0.554906021e-1f +
3727 // (0.961591928e-2f +
3728 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3730 // error 2.47208000*10^(-7), which is better than 18 bits
3731 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3732 getF32Constant(DAG, 0x3924b03e));
3733 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3734 getF32Constant(DAG, 0x3ab24b87));
3735 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3736 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3737 getF32Constant(DAG, 0x3c1d8c17));
3738 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3739 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3740 getF32Constant(DAG, 0x3d634a1d));
3741 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3742 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3743 getF32Constant(DAG, 0x3e75fe14));
3744 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3745 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3746 getF32Constant(DAG, 0x3f317234));
3747 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3748 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3749 getF32Constant(DAG, 0x3f800000));
3750 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BITCAST, dl,
3753 // Add the exponent into the result in integer domain.
3754 SDValue t14 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3755 TwoToFracPartOfX, IntegerPartOfX);
3757 result = DAG.getNode(ISD::BITCAST, dl, MVT::f32, t14);
3760 // No special expansion.
3761 result = DAG.getNode(ISD::FEXP, dl,
3762 getValue(I.getArgOperand(0)).getValueType(),
3763 getValue(I.getArgOperand(0)));
3766 setValue(&I, result);
3769 /// visitLog - Lower a log intrinsic. Handles the special sequences for
3770 /// limited-precision mode.
3772 SelectionDAGBuilder::visitLog(const CallInst &I) {
3774 DebugLoc dl = getCurDebugLoc();
3776 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3777 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3778 SDValue Op = getValue(I.getArgOperand(0));
3779 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3781 // Scale the exponent by log(2) [0.69314718f].
3782 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3783 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3784 getF32Constant(DAG, 0x3f317218));
3786 // Get the significand and build it into a floating-point number with
3788 SDValue X = GetSignificand(DAG, Op1, dl);
3790 if (LimitFloatPrecision <= 6) {
3791 // For floating-point precision of 6:
3795 // (1.4034025f - 0.23903021f * x) * x;
3797 // error 0.0034276066, which is better than 8 bits
3798 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3799 getF32Constant(DAG, 0xbe74c456));
3800 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3801 getF32Constant(DAG, 0x3fb3a2b1));
3802 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3803 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3804 getF32Constant(DAG, 0x3f949a29));
3806 result = DAG.getNode(ISD::FADD, dl,
3807 MVT::f32, LogOfExponent, LogOfMantissa);
3808 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3809 // For floating-point precision of 12:
3815 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x;
3817 // error 0.000061011436, which is 14 bits
3818 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3819 getF32Constant(DAG, 0xbd67b6d6));
3820 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3821 getF32Constant(DAG, 0x3ee4f4b8));
3822 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3823 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3824 getF32Constant(DAG, 0x3fbc278b));
3825 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3826 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3827 getF32Constant(DAG, 0x40348e95));
3828 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3829 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3830 getF32Constant(DAG, 0x3fdef31a));
3832 result = DAG.getNode(ISD::FADD, dl,
3833 MVT::f32, LogOfExponent, LogOfMantissa);
3834 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3835 // For floating-point precision of 18:
3843 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x;
3845 // error 0.0000023660568, which is better than 18 bits
3846 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3847 getF32Constant(DAG, 0xbc91e5ac));
3848 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3849 getF32Constant(DAG, 0x3e4350aa));
3850 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3851 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3852 getF32Constant(DAG, 0x3f60d3e3));
3853 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3854 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3855 getF32Constant(DAG, 0x4011cdf0));
3856 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3857 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3858 getF32Constant(DAG, 0x406cfd1c));
3859 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3860 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3861 getF32Constant(DAG, 0x408797cb));
3862 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3863 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3864 getF32Constant(DAG, 0x4006dcab));
3866 result = DAG.getNode(ISD::FADD, dl,
3867 MVT::f32, LogOfExponent, LogOfMantissa);
3870 // No special expansion.
3871 result = DAG.getNode(ISD::FLOG, dl,
3872 getValue(I.getArgOperand(0)).getValueType(),
3873 getValue(I.getArgOperand(0)));
3876 setValue(&I, result);
3879 /// visitLog2 - Lower a log2 intrinsic. Handles the special sequences for
3880 /// limited-precision mode.
3882 SelectionDAGBuilder::visitLog2(const CallInst &I) {
3884 DebugLoc dl = getCurDebugLoc();
3886 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3887 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3888 SDValue Op = getValue(I.getArgOperand(0));
3889 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3891 // Get the exponent.
3892 SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl);
3894 // Get the significand and build it into a floating-point number with
3896 SDValue X = GetSignificand(DAG, Op1, dl);
3898 // Different possible minimax approximations of significand in
3899 // floating-point for various degrees of accuracy over [1,2].
3900 if (LimitFloatPrecision <= 6) {
3901 // For floating-point precision of 6:
3903 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x;
3905 // error 0.0049451742, which is more than 7 bits
3906 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3907 getF32Constant(DAG, 0xbeb08fe0));
3908 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3909 getF32Constant(DAG, 0x40019463));
3910 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3911 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3912 getF32Constant(DAG, 0x3fd6633d));
3914 result = DAG.getNode(ISD::FADD, dl,
3915 MVT::f32, LogOfExponent, Log2ofMantissa);
3916 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3917 // For floating-point precision of 12:
3923 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x;
3925 // error 0.0000876136000, which is better than 13 bits
3926 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3927 getF32Constant(DAG, 0xbda7262e));
3928 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3929 getF32Constant(DAG, 0x3f25280b));
3930 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3931 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3932 getF32Constant(DAG, 0x4007b923));
3933 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3934 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3935 getF32Constant(DAG, 0x40823e2f));
3936 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3937 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3938 getF32Constant(DAG, 0x4020d29c));
3940 result = DAG.getNode(ISD::FADD, dl,
3941 MVT::f32, LogOfExponent, Log2ofMantissa);
3942 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3943 // For floating-point precision of 18:
3952 // 0.25691327e-1f * x) * x) * x) * x) * x) * x;
3954 // error 0.0000018516, which is better than 18 bits
3955 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3956 getF32Constant(DAG, 0xbcd2769e));
3957 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3958 getF32Constant(DAG, 0x3e8ce0b9));
3959 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3960 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3961 getF32Constant(DAG, 0x3fa22ae7));
3962 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3963 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3964 getF32Constant(DAG, 0x40525723));
3965 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3966 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3967 getF32Constant(DAG, 0x40aaf200));
3968 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3969 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3970 getF32Constant(DAG, 0x40c39dad));
3971 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3972 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3973 getF32Constant(DAG, 0x4042902c));
3975 result = DAG.getNode(ISD::FADD, dl,
3976 MVT::f32, LogOfExponent, Log2ofMantissa);
3979 // No special expansion.
3980 result = DAG.getNode(ISD::FLOG2, dl,
3981 getValue(I.getArgOperand(0)).getValueType(),
3982 getValue(I.getArgOperand(0)));
3985 setValue(&I, result);
3988 /// visitLog10 - Lower a log10 intrinsic. Handles the special sequences for
3989 /// limited-precision mode.
3991 SelectionDAGBuilder::visitLog10(const CallInst &I) {
3993 DebugLoc dl = getCurDebugLoc();
3995 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3996 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3997 SDValue Op = getValue(I.getArgOperand(0));
3998 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
4000 // Scale the exponent by log10(2) [0.30102999f].
4001 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
4002 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
4003 getF32Constant(DAG, 0x3e9a209a));
4005 // Get the significand and build it into a floating-point number with
4007 SDValue X = GetSignificand(DAG, Op1, dl);
4009 if (LimitFloatPrecision <= 6) {
4010 // For floating-point precision of 6:
4012 // Log10ofMantissa =
4014 // (0.60948995f - 0.10380950f * x) * x;
4016 // error 0.0014886165, which is 6 bits
4017 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4018 getF32Constant(DAG, 0xbdd49a13));
4019 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4020 getF32Constant(DAG, 0x3f1c0789));
4021 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4022 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4023 getF32Constant(DAG, 0x3f011300));
4025 result = DAG.getNode(ISD::FADD, dl,
4026 MVT::f32, LogOfExponent, Log10ofMantissa);
4027 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
4028 // For floating-point precision of 12:
4030 // Log10ofMantissa =
4033 // (-0.31664806f + 0.47637168e-1f * x) * x) * x;
4035 // error 0.00019228036, which is better than 12 bits
4036 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4037 getF32Constant(DAG, 0x3d431f31));
4038 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
4039 getF32Constant(DAG, 0x3ea21fb2));
4040 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4041 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4042 getF32Constant(DAG, 0x3f6ae232));
4043 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4044 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
4045 getF32Constant(DAG, 0x3f25f7c3));
4047 result = DAG.getNode(ISD::FADD, dl,
4048 MVT::f32, LogOfExponent, Log10ofMantissa);
4049 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
4050 // For floating-point precision of 18:
4052 // Log10ofMantissa =
4057 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x;
4059 // error 0.0000037995730, which is better than 18 bits
4060 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4061 getF32Constant(DAG, 0x3c5d51ce));
4062 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
4063 getF32Constant(DAG, 0x3e00685a));
4064 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4065 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4066 getF32Constant(DAG, 0x3efb6798));
4067 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4068 SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
4069 getF32Constant(DAG, 0x3f88d192));
4070 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4071 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4072 getF32Constant(DAG, 0x3fc4316c));
4073 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4074 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8,
4075 getF32Constant(DAG, 0x3f57ce70));
4077 result = DAG.getNode(ISD::FADD, dl,
4078 MVT::f32, LogOfExponent, Log10ofMantissa);
4081 // No special expansion.
4082 result = DAG.getNode(ISD::FLOG10, dl,
4083 getValue(I.getArgOperand(0)).getValueType(),
4084 getValue(I.getArgOperand(0)));
4087 setValue(&I, result);
4090 /// visitExp2 - Lower an exp2 intrinsic. Handles the special sequences for
4091 /// limited-precision mode.
4093 SelectionDAGBuilder::visitExp2(const CallInst &I) {
4095 DebugLoc dl = getCurDebugLoc();
4097 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
4098 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4099 SDValue Op = getValue(I.getArgOperand(0));
4101 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Op);
4103 // FractionalPartOfX = x - (float)IntegerPartOfX;
4104 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
4105 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, Op, t1);
4107 // IntegerPartOfX <<= 23;
4108 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
4109 DAG.getConstant(23, TLI.getPointerTy()));
4111 if (LimitFloatPrecision <= 6) {
4112 // For floating-point precision of 6:
4114 // TwoToFractionalPartOfX =
4116 // (0.735607626f + 0.252464424f * x) * x;
4118 // error 0.0144103317, which is 6 bits
4119 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4120 getF32Constant(DAG, 0x3e814304));
4121 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4122 getF32Constant(DAG, 0x3f3c50c8));
4123 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4124 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4125 getF32Constant(DAG, 0x3f7f5e7e));
4126 SDValue t6 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t5);
4127 SDValue TwoToFractionalPartOfX =
4128 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
4130 result = DAG.getNode(ISD::BITCAST, dl,
4131 MVT::f32, TwoToFractionalPartOfX);
4132 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
4133 // For floating-point precision of 12:
4135 // TwoToFractionalPartOfX =
4138 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
4140 // error 0.000107046256, which is 13 to 14 bits
4141 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4142 getF32Constant(DAG, 0x3da235e3));
4143 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4144 getF32Constant(DAG, 0x3e65b8f3));
4145 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4146 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4147 getF32Constant(DAG, 0x3f324b07));
4148 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4149 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4150 getF32Constant(DAG, 0x3f7ff8fd));
4151 SDValue t8 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t7);
4152 SDValue TwoToFractionalPartOfX =
4153 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
4155 result = DAG.getNode(ISD::BITCAST, dl,
4156 MVT::f32, TwoToFractionalPartOfX);
4157 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
4158 // For floating-point precision of 18:
4160 // TwoToFractionalPartOfX =
4164 // (0.554906021e-1f +
4165 // (0.961591928e-2f +
4166 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
4167 // error 2.47208000*10^(-7), which is better than 18 bits
4168 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4169 getF32Constant(DAG, 0x3924b03e));
4170 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4171 getF32Constant(DAG, 0x3ab24b87));
4172 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4173 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4174 getF32Constant(DAG, 0x3c1d8c17));
4175 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4176 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4177 getF32Constant(DAG, 0x3d634a1d));
4178 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4179 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4180 getF32Constant(DAG, 0x3e75fe14));
4181 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4182 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
4183 getF32Constant(DAG, 0x3f317234));
4184 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
4185 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
4186 getF32Constant(DAG, 0x3f800000));
4187 SDValue t14 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t13);
4188 SDValue TwoToFractionalPartOfX =
4189 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
4191 result = DAG.getNode(ISD::BITCAST, dl,
4192 MVT::f32, TwoToFractionalPartOfX);
4195 // No special expansion.
4196 result = DAG.getNode(ISD::FEXP2, dl,
4197 getValue(I.getArgOperand(0)).getValueType(),
4198 getValue(I.getArgOperand(0)));
4201 setValue(&I, result);
4204 /// visitPow - Lower a pow intrinsic. Handles the special sequences for
4205 /// limited-precision mode with x == 10.0f.
4207 SelectionDAGBuilder::visitPow(const CallInst &I) {
4209 const Value *Val = I.getArgOperand(0);
4210 DebugLoc dl = getCurDebugLoc();
4211 bool IsExp10 = false;
4213 if (getValue(Val).getValueType() == MVT::f32 &&
4214 getValue(I.getArgOperand(1)).getValueType() == MVT::f32 &&
4215 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4216 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(Val))) {
4217 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
4219 IsExp10 = CFP->getValueAPF().bitwiseIsEqual(Ten);
4224 if (IsExp10 && LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4225 SDValue Op = getValue(I.getArgOperand(1));
4227 // Put the exponent in the right bit position for later addition to the
4230 // #define LOG2OF10 3.3219281f
4231 // IntegerPartOfX = (int32_t)(x * LOG2OF10);
4232 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
4233 getF32Constant(DAG, 0x40549a78));
4234 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
4236 // FractionalPartOfX = x - (float)IntegerPartOfX;
4237 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
4238 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
4240 // IntegerPartOfX <<= 23;
4241 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
4242 DAG.getConstant(23, TLI.getPointerTy()));
4244 if (LimitFloatPrecision <= 6) {
4245 // For floating-point precision of 6:
4247 // twoToFractionalPartOfX =
4249 // (0.735607626f + 0.252464424f * x) * x;
4251 // error 0.0144103317, which is 6 bits
4252 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4253 getF32Constant(DAG, 0x3e814304));
4254 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4255 getF32Constant(DAG, 0x3f3c50c8));
4256 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4257 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4258 getF32Constant(DAG, 0x3f7f5e7e));
4259 SDValue t6 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t5);
4260 SDValue TwoToFractionalPartOfX =
4261 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
4263 result = DAG.getNode(ISD::BITCAST, dl,
4264 MVT::f32, TwoToFractionalPartOfX);
4265 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
4266 // For floating-point precision of 12:
4268 // TwoToFractionalPartOfX =
4271 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
4273 // error 0.000107046256, which is 13 to 14 bits
4274 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4275 getF32Constant(DAG, 0x3da235e3));
4276 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4277 getF32Constant(DAG, 0x3e65b8f3));
4278 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4279 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4280 getF32Constant(DAG, 0x3f324b07));
4281 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4282 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4283 getF32Constant(DAG, 0x3f7ff8fd));
4284 SDValue t8 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t7);
4285 SDValue TwoToFractionalPartOfX =
4286 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
4288 result = DAG.getNode(ISD::BITCAST, dl,
4289 MVT::f32, TwoToFractionalPartOfX);
4290 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
4291 // For floating-point precision of 18:
4293 // TwoToFractionalPartOfX =
4297 // (0.554906021e-1f +
4298 // (0.961591928e-2f +
4299 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
4300 // error 2.47208000*10^(-7), which is better than 18 bits
4301 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4302 getF32Constant(DAG, 0x3924b03e));
4303 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4304 getF32Constant(DAG, 0x3ab24b87));
4305 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4306 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4307 getF32Constant(DAG, 0x3c1d8c17));
4308 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4309 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4310 getF32Constant(DAG, 0x3d634a1d));
4311 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4312 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4313 getF32Constant(DAG, 0x3e75fe14));
4314 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4315 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
4316 getF32Constant(DAG, 0x3f317234));
4317 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
4318 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
4319 getF32Constant(DAG, 0x3f800000));
4320 SDValue t14 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t13);
4321 SDValue TwoToFractionalPartOfX =
4322 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
4324 result = DAG.getNode(ISD::BITCAST, dl,
4325 MVT::f32, TwoToFractionalPartOfX);
4328 // No special expansion.
4329 result = DAG.getNode(ISD::FPOW, dl,
4330 getValue(I.getArgOperand(0)).getValueType(),
4331 getValue(I.getArgOperand(0)),
4332 getValue(I.getArgOperand(1)));
4335 setValue(&I, result);
4339 /// ExpandPowI - Expand a llvm.powi intrinsic.
4340 static SDValue ExpandPowI(DebugLoc DL, SDValue LHS, SDValue RHS,
4341 SelectionDAG &DAG) {
4342 // If RHS is a constant, we can expand this out to a multiplication tree,
4343 // otherwise we end up lowering to a call to __powidf2 (for example). When
4344 // optimizing for size, we only want to do this if the expansion would produce
4345 // a small number of multiplies, otherwise we do the full expansion.
4346 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
4347 // Get the exponent as a positive value.
4348 unsigned Val = RHSC->getSExtValue();
4349 if ((int)Val < 0) Val = -Val;
4351 // powi(x, 0) -> 1.0
4353 return DAG.getConstantFP(1.0, LHS.getValueType());
4355 const Function *F = DAG.getMachineFunction().getFunction();
4356 if (!F->hasFnAttr(Attribute::OptimizeForSize) ||
4357 // If optimizing for size, don't insert too many multiplies. This
4358 // inserts up to 5 multiplies.
4359 CountPopulation_32(Val)+Log2_32(Val) < 7) {
4360 // We use the simple binary decomposition method to generate the multiply
4361 // sequence. There are more optimal ways to do this (for example,
4362 // powi(x,15) generates one more multiply than it should), but this has
4363 // the benefit of being both really simple and much better than a libcall.
4364 SDValue Res; // Logically starts equal to 1.0
4365 SDValue CurSquare = LHS;
4369 Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare);
4371 Res = CurSquare; // 1.0*CurSquare.
4374 CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(),
4375 CurSquare, CurSquare);
4379 // If the original was negative, invert the result, producing 1/(x*x*x).
4380 if (RHSC->getSExtValue() < 0)
4381 Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(),
4382 DAG.getConstantFP(1.0, LHS.getValueType()), Res);
4387 // Otherwise, expand to a libcall.
4388 return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS);
4391 // getTruncatedArgReg - Find underlying register used for an truncated
4393 static unsigned getTruncatedArgReg(const SDValue &N) {
4394 if (N.getOpcode() != ISD::TRUNCATE)
4397 const SDValue &Ext = N.getOperand(0);
4398 if (Ext.getOpcode() == ISD::AssertZext || Ext.getOpcode() == ISD::AssertSext){
4399 const SDValue &CFR = Ext.getOperand(0);
4400 if (CFR.getOpcode() == ISD::CopyFromReg)
4401 return cast<RegisterSDNode>(CFR.getOperand(1))->getReg();
4402 if (CFR.getOpcode() == ISD::TRUNCATE)
4403 return getTruncatedArgReg(CFR);
4408 /// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a function
4409 /// argument, create the corresponding DBG_VALUE machine instruction for it now.
4410 /// At the end of instruction selection, they will be inserted to the entry BB.
4412 SelectionDAGBuilder::EmitFuncArgumentDbgValue(const Value *V, MDNode *Variable,
4415 const Argument *Arg = dyn_cast<Argument>(V);
4419 MachineFunction &MF = DAG.getMachineFunction();
4420 const TargetInstrInfo *TII = DAG.getTarget().getInstrInfo();
4421 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
4423 // Ignore inlined function arguments here.
4424 DIVariable DV(Variable);
4425 if (DV.isInlinedFnArgument(MF.getFunction()))
4429 // Some arguments' frame index is recorded during argument lowering.
4430 Offset = FuncInfo.getArgumentFrameIndex(Arg);
4432 Reg = TRI->getFrameRegister(MF);
4434 if (!Reg && N.getNode()) {
4435 if (N.getOpcode() == ISD::CopyFromReg)
4436 Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg();
4438 Reg = getTruncatedArgReg(N);
4439 if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) {
4440 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4441 unsigned PR = RegInfo.getLiveInPhysReg(Reg);
4448 // Check if ValueMap has reg number.
4449 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
4450 if (VMI != FuncInfo.ValueMap.end())
4454 if (!Reg && N.getNode()) {
4455 // Check if frame index is available.
4456 if (LoadSDNode *LNode = dyn_cast<LoadSDNode>(N.getNode()))
4457 if (FrameIndexSDNode *FINode =
4458 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode())) {
4459 Reg = TRI->getFrameRegister(MF);
4460 Offset = FINode->getIndex();
4467 MachineInstrBuilder MIB = BuildMI(MF, getCurDebugLoc(),
4468 TII->get(TargetOpcode::DBG_VALUE))
4469 .addReg(Reg, RegState::Debug).addImm(Offset).addMetadata(Variable);
4470 FuncInfo.ArgDbgValues.push_back(&*MIB);
4474 // VisualStudio defines setjmp as _setjmp
4475 #if defined(_MSC_VER) && defined(setjmp) && \
4476 !defined(setjmp_undefined_for_msvc)
4477 # pragma push_macro("setjmp")
4479 # define setjmp_undefined_for_msvc
4482 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
4483 /// we want to emit this as a call to a named external function, return the name
4484 /// otherwise lower it and return null.
4486 SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) {
4487 DebugLoc dl = getCurDebugLoc();
4490 switch (Intrinsic) {
4492 // By default, turn this into a target intrinsic node.
4493 visitTargetIntrinsic(I, Intrinsic);
4495 case Intrinsic::vastart: visitVAStart(I); return 0;
4496 case Intrinsic::vaend: visitVAEnd(I); return 0;
4497 case Intrinsic::vacopy: visitVACopy(I); return 0;
4498 case Intrinsic::returnaddress:
4499 setValue(&I, DAG.getNode(ISD::RETURNADDR, dl, TLI.getPointerTy(),
4500 getValue(I.getArgOperand(0))));
4502 case Intrinsic::frameaddress:
4503 setValue(&I, DAG.getNode(ISD::FRAMEADDR, dl, TLI.getPointerTy(),
4504 getValue(I.getArgOperand(0))));
4506 case Intrinsic::setjmp:
4507 return &"_setjmp"[!TLI.usesUnderscoreSetJmp()];
4508 case Intrinsic::longjmp:
4509 return &"_longjmp"[!TLI.usesUnderscoreLongJmp()];
4510 case Intrinsic::memcpy: {
4511 // Assert for address < 256 since we support only user defined address
4513 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4515 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4517 "Unknown address space");
4518 SDValue Op1 = getValue(I.getArgOperand(0));
4519 SDValue Op2 = getValue(I.getArgOperand(1));
4520 SDValue Op3 = getValue(I.getArgOperand(2));
4521 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4522 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4523 DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, isVol, false,
4524 MachinePointerInfo(I.getArgOperand(0)),
4525 MachinePointerInfo(I.getArgOperand(1))));
4528 case Intrinsic::memset: {
4529 // Assert for address < 256 since we support only user defined address
4531 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4533 "Unknown address space");
4534 SDValue Op1 = getValue(I.getArgOperand(0));
4535 SDValue Op2 = getValue(I.getArgOperand(1));
4536 SDValue Op3 = getValue(I.getArgOperand(2));
4537 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4538 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4539 DAG.setRoot(DAG.getMemset(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4540 MachinePointerInfo(I.getArgOperand(0))));
4543 case Intrinsic::memmove: {
4544 // Assert for address < 256 since we support only user defined address
4546 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4548 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4550 "Unknown address space");
4551 SDValue Op1 = getValue(I.getArgOperand(0));
4552 SDValue Op2 = getValue(I.getArgOperand(1));
4553 SDValue Op3 = getValue(I.getArgOperand(2));
4554 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4555 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4556 DAG.setRoot(DAG.getMemmove(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4557 MachinePointerInfo(I.getArgOperand(0)),
4558 MachinePointerInfo(I.getArgOperand(1))));
4561 case Intrinsic::dbg_declare: {
4562 const DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
4563 MDNode *Variable = DI.getVariable();
4564 const Value *Address = DI.getAddress();
4565 if (!Address || !DIVariable(Variable).Verify()) {
4566 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4570 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder
4571 // but do not always have a corresponding SDNode built. The SDNodeOrder
4572 // absolute, but not relative, values are different depending on whether
4573 // debug info exists.
4576 // Check if address has undef value.
4577 if (isa<UndefValue>(Address) ||
4578 (Address->use_empty() && !isa<Argument>(Address))) {
4579 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4583 SDValue &N = NodeMap[Address];
4584 if (!N.getNode() && isa<Argument>(Address))
4585 // Check unused arguments map.
4586 N = UnusedArgNodeMap[Address];
4589 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
4590 Address = BCI->getOperand(0);
4591 // Parameters are handled specially.
4593 (DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable ||
4594 isa<Argument>(Address));
4596 const AllocaInst *AI = dyn_cast<AllocaInst>(Address);
4598 if (isParameter && !AI) {
4599 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode());
4601 // Byval parameter. We have a frame index at this point.
4602 SDV = DAG.getDbgValue(Variable, FINode->getIndex(),
4603 0, dl, SDNodeOrder);
4605 // Address is an argument, so try to emit its dbg value using
4606 // virtual register info from the FuncInfo.ValueMap.
4607 EmitFuncArgumentDbgValue(Address, Variable, 0, N);
4611 SDV = DAG.getDbgValue(Variable, N.getNode(), N.getResNo(),
4612 0, dl, SDNodeOrder);
4614 // Can't do anything with other non-AI cases yet.
4615 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4616 DEBUG(dbgs() << "non-AllocaInst issue for Address: \n\t");
4617 DEBUG(Address->dump());
4620 DAG.AddDbgValue(SDV, N.getNode(), isParameter);
4622 // If Address is an argument then try to emit its dbg value using
4623 // virtual register info from the FuncInfo.ValueMap.
4624 if (!EmitFuncArgumentDbgValue(Address, Variable, 0, N)) {
4625 // If variable is pinned by a alloca in dominating bb then
4626 // use StaticAllocaMap.
4627 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Address)) {
4628 if (AI->getParent() != DI.getParent()) {
4629 DenseMap<const AllocaInst*, int>::iterator SI =
4630 FuncInfo.StaticAllocaMap.find(AI);
4631 if (SI != FuncInfo.StaticAllocaMap.end()) {
4632 SDV = DAG.getDbgValue(Variable, SI->second,
4633 0, dl, SDNodeOrder);
4634 DAG.AddDbgValue(SDV, 0, false);
4639 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4644 case Intrinsic::dbg_value: {
4645 const DbgValueInst &DI = cast<DbgValueInst>(I);
4646 if (!DIVariable(DI.getVariable()).Verify())
4649 MDNode *Variable = DI.getVariable();
4650 uint64_t Offset = DI.getOffset();
4651 const Value *V = DI.getValue();
4655 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder
4656 // but do not always have a corresponding SDNode built. The SDNodeOrder
4657 // absolute, but not relative, values are different depending on whether
4658 // debug info exists.
4661 if (isa<ConstantInt>(V) || isa<ConstantFP>(V) || isa<UndefValue>(V)) {
4662 SDV = DAG.getDbgValue(Variable, V, Offset, dl, SDNodeOrder);
4663 DAG.AddDbgValue(SDV, 0, false);
4665 // Do not use getValue() in here; we don't want to generate code at
4666 // this point if it hasn't been done yet.
4667 SDValue N = NodeMap[V];
4668 if (!N.getNode() && isa<Argument>(V))
4669 // Check unused arguments map.
4670 N = UnusedArgNodeMap[V];
4672 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, N)) {
4673 SDV = DAG.getDbgValue(Variable, N.getNode(),
4674 N.getResNo(), Offset, dl, SDNodeOrder);
4675 DAG.AddDbgValue(SDV, N.getNode(), false);
4677 } else if (!V->use_empty() ) {
4678 // Do not call getValue(V) yet, as we don't want to generate code.
4679 // Remember it for later.
4680 DanglingDebugInfo DDI(&DI, dl, SDNodeOrder);
4681 DanglingDebugInfoMap[V] = DDI;
4683 // We may expand this to cover more cases. One case where we have no
4684 // data available is an unreferenced parameter.
4685 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4689 // Build a debug info table entry.
4690 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
4691 V = BCI->getOperand(0);
4692 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
4693 // Don't handle byval struct arguments or VLAs, for example.
4695 DEBUG(dbgs() << "Dropping debug location info for:\n " << DI << "\n");
4696 DEBUG(dbgs() << " Last seen at:\n " << *V << "\n");
4699 DenseMap<const AllocaInst*, int>::iterator SI =
4700 FuncInfo.StaticAllocaMap.find(AI);
4701 if (SI == FuncInfo.StaticAllocaMap.end())
4703 int FI = SI->second;
4705 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4706 if (!DI.getDebugLoc().isUnknown() && MMI.hasDebugInfo())
4707 MMI.setVariableDbgInfo(Variable, FI, DI.getDebugLoc());
4711 case Intrinsic::eh_typeid_for: {
4712 // Find the type id for the given typeinfo.
4713 GlobalVariable *GV = ExtractTypeInfo(I.getArgOperand(0));
4714 unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(GV);
4715 Res = DAG.getConstant(TypeID, MVT::i32);
4720 case Intrinsic::eh_return_i32:
4721 case Intrinsic::eh_return_i64:
4722 DAG.getMachineFunction().getMMI().setCallsEHReturn(true);
4723 DAG.setRoot(DAG.getNode(ISD::EH_RETURN, dl,
4726 getValue(I.getArgOperand(0)),
4727 getValue(I.getArgOperand(1))));
4729 case Intrinsic::eh_unwind_init:
4730 DAG.getMachineFunction().getMMI().setCallsUnwindInit(true);
4732 case Intrinsic::eh_dwarf_cfa: {
4733 SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getArgOperand(0)), dl,
4734 TLI.getPointerTy());
4735 SDValue Offset = DAG.getNode(ISD::ADD, dl,
4737 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, dl,
4738 TLI.getPointerTy()),
4740 SDValue FA = DAG.getNode(ISD::FRAMEADDR, dl,
4742 DAG.getConstant(0, TLI.getPointerTy()));
4743 setValue(&I, DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
4747 case Intrinsic::eh_sjlj_callsite: {
4748 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4749 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(0));
4750 assert(CI && "Non-constant call site value in eh.sjlj.callsite!");
4751 assert(MMI.getCurrentCallSite() == 0 && "Overlapping call sites!");
4753 MMI.setCurrentCallSite(CI->getZExtValue());
4756 case Intrinsic::eh_sjlj_functioncontext: {
4757 // Get and store the index of the function context.
4758 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4760 cast<AllocaInst>(I.getArgOperand(0)->stripPointerCasts());
4761 int FI = FuncInfo.StaticAllocaMap[FnCtx];
4762 MFI->setFunctionContextIndex(FI);
4765 case Intrinsic::eh_sjlj_setjmp: {
4768 Ops[1] = getValue(I.getArgOperand(0));
4769 SDValue Op = DAG.getNode(ISD::EH_SJLJ_SETJMP, dl,
4770 DAG.getVTList(MVT::i32, MVT::Other),
4772 setValue(&I, Op.getValue(0));
4773 DAG.setRoot(Op.getValue(1));
4776 case Intrinsic::eh_sjlj_longjmp: {
4777 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, dl, MVT::Other,
4778 getRoot(), getValue(I.getArgOperand(0))));
4782 case Intrinsic::x86_mmx_pslli_w:
4783 case Intrinsic::x86_mmx_pslli_d:
4784 case Intrinsic::x86_mmx_pslli_q:
4785 case Intrinsic::x86_mmx_psrli_w:
4786 case Intrinsic::x86_mmx_psrli_d:
4787 case Intrinsic::x86_mmx_psrli_q:
4788 case Intrinsic::x86_mmx_psrai_w:
4789 case Intrinsic::x86_mmx_psrai_d: {
4790 SDValue ShAmt = getValue(I.getArgOperand(1));
4791 if (isa<ConstantSDNode>(ShAmt)) {
4792 visitTargetIntrinsic(I, Intrinsic);
4795 unsigned NewIntrinsic = 0;
4796 EVT ShAmtVT = MVT::v2i32;
4797 switch (Intrinsic) {
4798 case Intrinsic::x86_mmx_pslli_w:
4799 NewIntrinsic = Intrinsic::x86_mmx_psll_w;
4801 case Intrinsic::x86_mmx_pslli_d:
4802 NewIntrinsic = Intrinsic::x86_mmx_psll_d;
4804 case Intrinsic::x86_mmx_pslli_q:
4805 NewIntrinsic = Intrinsic::x86_mmx_psll_q;
4807 case Intrinsic::x86_mmx_psrli_w:
4808 NewIntrinsic = Intrinsic::x86_mmx_psrl_w;
4810 case Intrinsic::x86_mmx_psrli_d:
4811 NewIntrinsic = Intrinsic::x86_mmx_psrl_d;
4813 case Intrinsic::x86_mmx_psrli_q:
4814 NewIntrinsic = Intrinsic::x86_mmx_psrl_q;
4816 case Intrinsic::x86_mmx_psrai_w:
4817 NewIntrinsic = Intrinsic::x86_mmx_psra_w;
4819 case Intrinsic::x86_mmx_psrai_d:
4820 NewIntrinsic = Intrinsic::x86_mmx_psra_d;
4822 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4825 // The vector shift intrinsics with scalars uses 32b shift amounts but
4826 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
4828 // We must do this early because v2i32 is not a legal type.
4829 DebugLoc dl = getCurDebugLoc();
4832 ShOps[1] = DAG.getConstant(0, MVT::i32);
4833 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
4834 EVT DestVT = TLI.getValueType(I.getType());
4835 ShAmt = DAG.getNode(ISD::BITCAST, dl, DestVT, ShAmt);
4836 Res = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4837 DAG.getConstant(NewIntrinsic, MVT::i32),
4838 getValue(I.getArgOperand(0)), ShAmt);
4842 case Intrinsic::x86_avx_vinsertf128_pd_256:
4843 case Intrinsic::x86_avx_vinsertf128_ps_256:
4844 case Intrinsic::x86_avx_vinsertf128_si_256:
4845 case Intrinsic::x86_avx2_vinserti128: {
4846 DebugLoc dl = getCurDebugLoc();
4847 EVT DestVT = TLI.getValueType(I.getType());
4848 EVT ElVT = TLI.getValueType(I.getArgOperand(1)->getType());
4849 uint64_t Idx = (cast<ConstantInt>(I.getArgOperand(2))->getZExtValue() & 1) *
4850 ElVT.getVectorNumElements();
4851 Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, DestVT,
4852 getValue(I.getArgOperand(0)),
4853 getValue(I.getArgOperand(1)),
4854 DAG.getConstant(Idx, MVT::i32));
4858 case Intrinsic::convertff:
4859 case Intrinsic::convertfsi:
4860 case Intrinsic::convertfui:
4861 case Intrinsic::convertsif:
4862 case Intrinsic::convertuif:
4863 case Intrinsic::convertss:
4864 case Intrinsic::convertsu:
4865 case Intrinsic::convertus:
4866 case Intrinsic::convertuu: {
4867 ISD::CvtCode Code = ISD::CVT_INVALID;
4868 switch (Intrinsic) {
4869 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4870 case Intrinsic::convertff: Code = ISD::CVT_FF; break;
4871 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break;
4872 case Intrinsic::convertfui: Code = ISD::CVT_FU; break;
4873 case Intrinsic::convertsif: Code = ISD::CVT_SF; break;
4874 case Intrinsic::convertuif: Code = ISD::CVT_UF; break;
4875 case Intrinsic::convertss: Code = ISD::CVT_SS; break;
4876 case Intrinsic::convertsu: Code = ISD::CVT_SU; break;
4877 case Intrinsic::convertus: Code = ISD::CVT_US; break;
4878 case Intrinsic::convertuu: Code = ISD::CVT_UU; break;
4880 EVT DestVT = TLI.getValueType(I.getType());
4881 const Value *Op1 = I.getArgOperand(0);
4882 Res = DAG.getConvertRndSat(DestVT, getCurDebugLoc(), getValue(Op1),
4883 DAG.getValueType(DestVT),
4884 DAG.getValueType(getValue(Op1).getValueType()),
4885 getValue(I.getArgOperand(1)),
4886 getValue(I.getArgOperand(2)),
4891 case Intrinsic::sqrt:
4892 setValue(&I, DAG.getNode(ISD::FSQRT, dl,
4893 getValue(I.getArgOperand(0)).getValueType(),
4894 getValue(I.getArgOperand(0))));
4896 case Intrinsic::powi:
4897 setValue(&I, ExpandPowI(dl, getValue(I.getArgOperand(0)),
4898 getValue(I.getArgOperand(1)), DAG));
4900 case Intrinsic::sin:
4901 setValue(&I, DAG.getNode(ISD::FSIN, dl,
4902 getValue(I.getArgOperand(0)).getValueType(),
4903 getValue(I.getArgOperand(0))));
4905 case Intrinsic::cos:
4906 setValue(&I, DAG.getNode(ISD::FCOS, dl,
4907 getValue(I.getArgOperand(0)).getValueType(),
4908 getValue(I.getArgOperand(0))));
4910 case Intrinsic::log:
4913 case Intrinsic::log2:
4916 case Intrinsic::log10:
4919 case Intrinsic::exp:
4922 case Intrinsic::exp2:
4925 case Intrinsic::pow:
4928 case Intrinsic::fabs:
4929 setValue(&I, DAG.getNode(ISD::FABS, dl,
4930 getValue(I.getArgOperand(0)).getValueType(),
4931 getValue(I.getArgOperand(0))));
4933 case Intrinsic::fma:
4934 setValue(&I, DAG.getNode(ISD::FMA, dl,
4935 getValue(I.getArgOperand(0)).getValueType(),
4936 getValue(I.getArgOperand(0)),
4937 getValue(I.getArgOperand(1)),
4938 getValue(I.getArgOperand(2))));
4940 case Intrinsic::fmuladd: {
4941 EVT VT = TLI.getValueType(I.getType());
4942 if (TM.Options.AllowFPOpFusion != FPOpFusion::Strict &&
4943 TLI.isOperationLegal(ISD::FMA, VT) &&
4944 TLI.isFMAFasterThanMulAndAdd(VT)){
4945 setValue(&I, DAG.getNode(ISD::FMA, dl,
4946 getValue(I.getArgOperand(0)).getValueType(),
4947 getValue(I.getArgOperand(0)),
4948 getValue(I.getArgOperand(1)),
4949 getValue(I.getArgOperand(2))));
4951 SDValue Mul = DAG.getNode(ISD::FMUL, dl,
4952 getValue(I.getArgOperand(0)).getValueType(),
4953 getValue(I.getArgOperand(0)),
4954 getValue(I.getArgOperand(1)));
4955 SDValue Add = DAG.getNode(ISD::FADD, dl,
4956 getValue(I.getArgOperand(0)).getValueType(),
4958 getValue(I.getArgOperand(2)));
4963 case Intrinsic::convert_to_fp16:
4964 setValue(&I, DAG.getNode(ISD::FP32_TO_FP16, dl,
4965 MVT::i16, getValue(I.getArgOperand(0))));
4967 case Intrinsic::convert_from_fp16:
4968 setValue(&I, DAG.getNode(ISD::FP16_TO_FP32, dl,
4969 MVT::f32, getValue(I.getArgOperand(0))));
4971 case Intrinsic::pcmarker: {
4972 SDValue Tmp = getValue(I.getArgOperand(0));
4973 DAG.setRoot(DAG.getNode(ISD::PCMARKER, dl, MVT::Other, getRoot(), Tmp));
4976 case Intrinsic::readcyclecounter: {
4977 SDValue Op = getRoot();
4978 Res = DAG.getNode(ISD::READCYCLECOUNTER, dl,
4979 DAG.getVTList(MVT::i64, MVT::Other),
4982 DAG.setRoot(Res.getValue(1));
4985 case Intrinsic::bswap:
4986 setValue(&I, DAG.getNode(ISD::BSWAP, dl,
4987 getValue(I.getArgOperand(0)).getValueType(),
4988 getValue(I.getArgOperand(0))));
4990 case Intrinsic::cttz: {
4991 SDValue Arg = getValue(I.getArgOperand(0));
4992 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1));
4993 EVT Ty = Arg.getValueType();
4994 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTTZ : ISD::CTTZ_ZERO_UNDEF,
4998 case Intrinsic::ctlz: {
4999 SDValue Arg = getValue(I.getArgOperand(0));
5000 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1));
5001 EVT Ty = Arg.getValueType();
5002 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTLZ : ISD::CTLZ_ZERO_UNDEF,
5006 case Intrinsic::ctpop: {
5007 SDValue Arg = getValue(I.getArgOperand(0));
5008 EVT Ty = Arg.getValueType();
5009 setValue(&I, DAG.getNode(ISD::CTPOP, dl, Ty, Arg));
5012 case Intrinsic::stacksave: {
5013 SDValue Op = getRoot();
5014 Res = DAG.getNode(ISD::STACKSAVE, dl,
5015 DAG.getVTList(TLI.getPointerTy(), MVT::Other), &Op, 1);
5017 DAG.setRoot(Res.getValue(1));
5020 case Intrinsic::stackrestore: {
5021 Res = getValue(I.getArgOperand(0));
5022 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, dl, MVT::Other, getRoot(), Res));
5025 case Intrinsic::stackprotector: {
5026 // Emit code into the DAG to store the stack guard onto the stack.
5027 MachineFunction &MF = DAG.getMachineFunction();
5028 MachineFrameInfo *MFI = MF.getFrameInfo();
5029 EVT PtrTy = TLI.getPointerTy();
5031 SDValue Src = getValue(I.getArgOperand(0)); // The guard's value.
5032 AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
5034 int FI = FuncInfo.StaticAllocaMap[Slot];
5035 MFI->setStackProtectorIndex(FI);
5037 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
5039 // Store the stack protector onto the stack.
5040 Res = DAG.getStore(getRoot(), getCurDebugLoc(), Src, FIN,
5041 MachinePointerInfo::getFixedStack(FI),
5047 case Intrinsic::objectsize: {
5048 // If we don't know by now, we're never going to know.
5049 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1));
5051 assert(CI && "Non-constant type in __builtin_object_size?");
5053 SDValue Arg = getValue(I.getCalledValue());
5054 EVT Ty = Arg.getValueType();
5057 Res = DAG.getConstant(-1ULL, Ty);
5059 Res = DAG.getConstant(0, Ty);
5064 case Intrinsic::var_annotation:
5065 // Discard annotate attributes
5068 case Intrinsic::init_trampoline: {
5069 const Function *F = cast<Function>(I.getArgOperand(1)->stripPointerCasts());
5073 Ops[1] = getValue(I.getArgOperand(0));
5074 Ops[2] = getValue(I.getArgOperand(1));
5075 Ops[3] = getValue(I.getArgOperand(2));
5076 Ops[4] = DAG.getSrcValue(I.getArgOperand(0));
5077 Ops[5] = DAG.getSrcValue(F);
5079 Res = DAG.getNode(ISD::INIT_TRAMPOLINE, dl, MVT::Other, Ops, 6);
5084 case Intrinsic::adjust_trampoline: {
5085 setValue(&I, DAG.getNode(ISD::ADJUST_TRAMPOLINE, dl,
5087 getValue(I.getArgOperand(0))));
5090 case Intrinsic::gcroot:
5092 const Value *Alloca = I.getArgOperand(0)->stripPointerCasts();
5093 const Constant *TypeMap = cast<Constant>(I.getArgOperand(1));
5095 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode());
5096 GFI->addStackRoot(FI->getIndex(), TypeMap);
5099 case Intrinsic::gcread:
5100 case Intrinsic::gcwrite:
5101 llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!");
5102 case Intrinsic::flt_rounds:
5103 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, dl, MVT::i32));
5106 case Intrinsic::expect: {
5107 // Just replace __builtin_expect(exp, c) with EXP.
5108 setValue(&I, getValue(I.getArgOperand(0)));
5112 case Intrinsic::trap: {
5113 StringRef TrapFuncName = TM.Options.getTrapFunctionName();
5114 if (TrapFuncName.empty()) {
5115 DAG.setRoot(DAG.getNode(ISD::TRAP, dl,MVT::Other, getRoot()));
5118 TargetLowering::ArgListTy Args;
5120 CallLoweringInfo CLI(getRoot(), I.getType(),
5121 false, false, false, false, 0, CallingConv::C,
5122 /*isTailCall=*/false,
5123 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
5124 DAG.getExternalSymbol(TrapFuncName.data(), TLI.getPointerTy()),
5125 Args, DAG, getCurDebugLoc());
5126 std::pair<SDValue, SDValue> Result = TLI.LowerCallTo(CLI);
5127 DAG.setRoot(Result.second);
5130 case Intrinsic::debugtrap: {
5131 DAG.setRoot(DAG.getNode(ISD::DEBUGTRAP, dl,MVT::Other, getRoot()));
5134 case Intrinsic::uadd_with_overflow:
5135 case Intrinsic::sadd_with_overflow:
5136 case Intrinsic::usub_with_overflow:
5137 case Intrinsic::ssub_with_overflow:
5138 case Intrinsic::umul_with_overflow:
5139 case Intrinsic::smul_with_overflow: {
5141 switch (Intrinsic) {
5142 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
5143 case Intrinsic::uadd_with_overflow: Op = ISD::UADDO; break;
5144 case Intrinsic::sadd_with_overflow: Op = ISD::SADDO; break;
5145 case Intrinsic::usub_with_overflow: Op = ISD::USUBO; break;
5146 case Intrinsic::ssub_with_overflow: Op = ISD::SSUBO; break;
5147 case Intrinsic::umul_with_overflow: Op = ISD::UMULO; break;
5148 case Intrinsic::smul_with_overflow: Op = ISD::SMULO; break;
5150 SDValue Op1 = getValue(I.getArgOperand(0));
5151 SDValue Op2 = getValue(I.getArgOperand(1));
5153 SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1);
5154 setValue(&I, DAG.getNode(Op, getCurDebugLoc(), VTs, Op1, Op2));
5157 case Intrinsic::prefetch: {
5159 unsigned rw = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
5161 Ops[1] = getValue(I.getArgOperand(0));
5162 Ops[2] = getValue(I.getArgOperand(1));
5163 Ops[3] = getValue(I.getArgOperand(2));
5164 Ops[4] = getValue(I.getArgOperand(3));
5165 DAG.setRoot(DAG.getMemIntrinsicNode(ISD::PREFETCH, dl,
5166 DAG.getVTList(MVT::Other),
5168 EVT::getIntegerVT(*Context, 8),
5169 MachinePointerInfo(I.getArgOperand(0)),
5171 false, /* volatile */
5173 rw==1)); /* write */
5177 case Intrinsic::invariant_start:
5178 case Intrinsic::lifetime_start:
5179 // Discard region information.
5180 setValue(&I, DAG.getUNDEF(TLI.getPointerTy()));
5182 case Intrinsic::invariant_end:
5183 case Intrinsic::lifetime_end:
5184 // Discard region information.
5186 case Intrinsic::donothing:
5192 void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee,
5194 MachineBasicBlock *LandingPad) {
5195 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
5196 FunctionType *FTy = cast<FunctionType>(PT->getElementType());
5197 Type *RetTy = FTy->getReturnType();
5198 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5199 MCSymbol *BeginLabel = 0;
5201 TargetLowering::ArgListTy Args;
5202 TargetLowering::ArgListEntry Entry;
5203 Args.reserve(CS.arg_size());
5205 // Check whether the function can return without sret-demotion.
5206 SmallVector<ISD::OutputArg, 4> Outs;
5207 GetReturnInfo(RetTy, CS.getAttributes().getRetAttributes(),
5210 bool CanLowerReturn = TLI.CanLowerReturn(CS.getCallingConv(),
5211 DAG.getMachineFunction(),
5212 FTy->isVarArg(), Outs,
5215 SDValue DemoteStackSlot;
5216 int DemoteStackIdx = -100;
5218 if (!CanLowerReturn) {
5219 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(
5220 FTy->getReturnType());
5221 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(
5222 FTy->getReturnType());
5223 MachineFunction &MF = DAG.getMachineFunction();
5224 DemoteStackIdx = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
5225 Type *StackSlotPtrType = PointerType::getUnqual(FTy->getReturnType());
5227 DemoteStackSlot = DAG.getFrameIndex(DemoteStackIdx, TLI.getPointerTy());
5228 Entry.Node = DemoteStackSlot;
5229 Entry.Ty = StackSlotPtrType;
5230 Entry.isSExt = false;
5231 Entry.isZExt = false;
5232 Entry.isInReg = false;
5233 Entry.isSRet = true;
5234 Entry.isNest = false;
5235 Entry.isByVal = false;
5236 Entry.Alignment = Align;
5237 Args.push_back(Entry);
5238 RetTy = Type::getVoidTy(FTy->getContext());
5241 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
5243 const Value *V = *i;
5246 if (V->getType()->isEmptyTy())
5249 SDValue ArgNode = getValue(V);
5250 Entry.Node = ArgNode; Entry.Ty = V->getType();
5252 unsigned attrInd = i - CS.arg_begin() + 1;
5253 Entry.isSExt = CS.paramHasAttr(attrInd, Attribute::SExt);
5254 Entry.isZExt = CS.paramHasAttr(attrInd, Attribute::ZExt);
5255 Entry.isInReg = CS.paramHasAttr(attrInd, Attribute::InReg);
5256 Entry.isSRet = CS.paramHasAttr(attrInd, Attribute::StructRet);
5257 Entry.isNest = CS.paramHasAttr(attrInd, Attribute::Nest);
5258 Entry.isByVal = CS.paramHasAttr(attrInd, Attribute::ByVal);
5259 Entry.Alignment = CS.getParamAlignment(attrInd);
5260 Args.push_back(Entry);
5264 // Insert a label before the invoke call to mark the try range. This can be
5265 // used to detect deletion of the invoke via the MachineModuleInfo.
5266 BeginLabel = MMI.getContext().CreateTempSymbol();
5268 // For SjLj, keep track of which landing pads go with which invokes
5269 // so as to maintain the ordering of pads in the LSDA.
5270 unsigned CallSiteIndex = MMI.getCurrentCallSite();
5271 if (CallSiteIndex) {
5272 MMI.setCallSiteBeginLabel(BeginLabel, CallSiteIndex);
5273 LPadToCallSiteMap[LandingPad].push_back(CallSiteIndex);
5275 // Now that the call site is handled, stop tracking it.
5276 MMI.setCurrentCallSite(0);
5279 // Both PendingLoads and PendingExports must be flushed here;
5280 // this call might not return.
5282 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getControlRoot(), BeginLabel));
5285 // Check if target-independent constraints permit a tail call here.
5286 // Target-dependent constraints are checked within TLI.LowerCallTo.
5288 !isInTailCallPosition(CS, CS.getAttributes().getRetAttributes(), TLI))
5291 // If there's a possibility that fast-isel has already selected some amount
5292 // of the current basic block, don't emit a tail call.
5293 if (isTailCall && TM.Options.EnableFastISel)
5297 CallLoweringInfo CLI(getRoot(), RetTy, FTy, isTailCall, Callee, Args, DAG,
5298 getCurDebugLoc(), CS);
5299 std::pair<SDValue,SDValue> Result = TLI.LowerCallTo(CLI);
5300 assert((isTailCall || Result.second.getNode()) &&
5301 "Non-null chain expected with non-tail call!");
5302 assert((Result.second.getNode() || !Result.first.getNode()) &&
5303 "Null value expected with tail call!");
5304 if (Result.first.getNode()) {
5305 setValue(CS.getInstruction(), Result.first);
5306 } else if (!CanLowerReturn && Result.second.getNode()) {
5307 // The instruction result is the result of loading from the
5308 // hidden sret parameter.
5309 SmallVector<EVT, 1> PVTs;
5310 Type *PtrRetTy = PointerType::getUnqual(FTy->getReturnType());
5312 ComputeValueVTs(TLI, PtrRetTy, PVTs);
5313 assert(PVTs.size() == 1 && "Pointers should fit in one register");
5314 EVT PtrVT = PVTs[0];
5316 SmallVector<EVT, 4> RetTys;
5317 SmallVector<uint64_t, 4> Offsets;
5318 RetTy = FTy->getReturnType();
5319 ComputeValueVTs(TLI, RetTy, RetTys, &Offsets);
5321 unsigned NumValues = RetTys.size();
5322 SmallVector<SDValue, 4> Values(NumValues);
5323 SmallVector<SDValue, 4> Chains(NumValues);
5325 for (unsigned i = 0; i < NumValues; ++i) {
5326 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT,
5328 DAG.getConstant(Offsets[i], PtrVT));
5329 SDValue L = DAG.getLoad(RetTys[i], getCurDebugLoc(), Result.second, Add,
5330 MachinePointerInfo::getFixedStack(DemoteStackIdx, Offsets[i]),
5331 false, false, false, 1);
5333 Chains[i] = L.getValue(1);
5336 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
5337 MVT::Other, &Chains[0], NumValues);
5338 PendingLoads.push_back(Chain);
5340 setValue(CS.getInstruction(),
5341 DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
5342 DAG.getVTList(&RetTys[0], RetTys.size()),
5343 &Values[0], Values.size()));
5346 // Assign order to nodes here. If the call does not produce a result, it won't
5347 // be mapped to a SDNode and visit() will not assign it an order number.
5348 if (!Result.second.getNode()) {
5349 // As a special case, a null chain means that a tail call has been emitted and
5350 // the DAG root is already updated.
5353 AssignOrderingToNode(DAG.getRoot().getNode());
5355 DAG.setRoot(Result.second);
5357 AssignOrderingToNode(Result.second.getNode());
5361 // Insert a label at the end of the invoke call to mark the try range. This
5362 // can be used to detect deletion of the invoke via the MachineModuleInfo.
5363 MCSymbol *EndLabel = MMI.getContext().CreateTempSymbol();
5364 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getRoot(), EndLabel));
5366 // Inform MachineModuleInfo of range.
5367 MMI.addInvoke(LandingPad, BeginLabel, EndLabel);
5371 /// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the
5372 /// value is equal or not-equal to zero.
5373 static bool IsOnlyUsedInZeroEqualityComparison(const Value *V) {
5374 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end();
5376 if (const ICmpInst *IC = dyn_cast<ICmpInst>(*UI))
5377 if (IC->isEquality())
5378 if (const Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
5379 if (C->isNullValue())
5381 // Unknown instruction.
5387 static SDValue getMemCmpLoad(const Value *PtrVal, MVT LoadVT,
5389 SelectionDAGBuilder &Builder) {
5391 // Check to see if this load can be trivially constant folded, e.g. if the
5392 // input is from a string literal.
5393 if (const Constant *LoadInput = dyn_cast<Constant>(PtrVal)) {
5394 // Cast pointer to the type we really want to load.
5395 LoadInput = ConstantExpr::getBitCast(const_cast<Constant *>(LoadInput),
5396 PointerType::getUnqual(LoadTy));
5398 if (const Constant *LoadCst =
5399 ConstantFoldLoadFromConstPtr(const_cast<Constant *>(LoadInput),
5401 return Builder.getValue(LoadCst);
5404 // Otherwise, we have to emit the load. If the pointer is to unfoldable but
5405 // still constant memory, the input chain can be the entry node.
5407 bool ConstantMemory = false;
5409 // Do not serialize (non-volatile) loads of constant memory with anything.
5410 if (Builder.AA->pointsToConstantMemory(PtrVal)) {
5411 Root = Builder.DAG.getEntryNode();
5412 ConstantMemory = true;
5414 // Do not serialize non-volatile loads against each other.
5415 Root = Builder.DAG.getRoot();
5418 SDValue Ptr = Builder.getValue(PtrVal);
5419 SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurDebugLoc(), Root,
5420 Ptr, MachinePointerInfo(PtrVal),
5422 false /*nontemporal*/,
5423 false /*isinvariant*/, 1 /* align=1 */);
5425 if (!ConstantMemory)
5426 Builder.PendingLoads.push_back(LoadVal.getValue(1));
5431 /// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form.
5432 /// If so, return true and lower it, otherwise return false and it will be
5433 /// lowered like a normal call.
5434 bool SelectionDAGBuilder::visitMemCmpCall(const CallInst &I) {
5435 // Verify that the prototype makes sense. int memcmp(void*,void*,size_t)
5436 if (I.getNumArgOperands() != 3)
5439 const Value *LHS = I.getArgOperand(0), *RHS = I.getArgOperand(1);
5440 if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() ||
5441 !I.getArgOperand(2)->getType()->isIntegerTy() ||
5442 !I.getType()->isIntegerTy())
5445 const ConstantInt *Size = dyn_cast<ConstantInt>(I.getArgOperand(2));
5447 // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0
5448 // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0
5449 if (Size && IsOnlyUsedInZeroEqualityComparison(&I)) {
5450 bool ActuallyDoIt = true;
5453 switch (Size->getZExtValue()) {
5455 LoadVT = MVT::Other;
5457 ActuallyDoIt = false;
5461 LoadTy = Type::getInt16Ty(Size->getContext());
5465 LoadTy = Type::getInt32Ty(Size->getContext());
5469 LoadTy = Type::getInt64Ty(Size->getContext());
5473 LoadVT = MVT::v4i32;
5474 LoadTy = Type::getInt32Ty(Size->getContext());
5475 LoadTy = VectorType::get(LoadTy, 4);
5480 // This turns into unaligned loads. We only do this if the target natively
5481 // supports the MVT we'll be loading or if it is small enough (<= 4) that
5482 // we'll only produce a small number of byte loads.
5484 // Require that we can find a legal MVT, and only do this if the target
5485 // supports unaligned loads of that type. Expanding into byte loads would
5487 if (ActuallyDoIt && Size->getZExtValue() > 4) {
5488 // TODO: Handle 5 byte compare as 4-byte + 1 byte.
5489 // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads.
5490 if (!TLI.isTypeLegal(LoadVT) ||!TLI.allowsUnalignedMemoryAccesses(LoadVT))
5491 ActuallyDoIt = false;
5495 SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this);
5496 SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this);
5498 SDValue Res = DAG.getSetCC(getCurDebugLoc(), MVT::i1, LHSVal, RHSVal,
5500 EVT CallVT = TLI.getValueType(I.getType(), true);
5501 setValue(&I, DAG.getZExtOrTrunc(Res, getCurDebugLoc(), CallVT));
5511 void SelectionDAGBuilder::visitCall(const CallInst &I) {
5512 // Handle inline assembly differently.
5513 if (isa<InlineAsm>(I.getCalledValue())) {
5518 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5519 ComputeUsesVAFloatArgument(I, &MMI);
5521 const char *RenameFn = 0;
5522 if (Function *F = I.getCalledFunction()) {
5523 if (F->isDeclaration()) {
5524 if (const TargetIntrinsicInfo *II = TM.getIntrinsicInfo()) {
5525 if (unsigned IID = II->getIntrinsicID(F)) {
5526 RenameFn = visitIntrinsicCall(I, IID);
5531 if (unsigned IID = F->getIntrinsicID()) {
5532 RenameFn = visitIntrinsicCall(I, IID);
5538 // Check for well-known libc/libm calls. If the function is internal, it
5539 // can't be a library call.
5540 if (!F->hasLocalLinkage() && F->hasName()) {
5541 StringRef Name = F->getName();
5542 if ((LibInfo->has(LibFunc::copysign) && Name == "copysign") ||
5543 (LibInfo->has(LibFunc::copysignf) && Name == "copysignf") ||
5544 (LibInfo->has(LibFunc::copysignl) && Name == "copysignl")) {
5545 if (I.getNumArgOperands() == 2 && // Basic sanity checks.
5546 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5547 I.getType() == I.getArgOperand(0)->getType() &&
5548 I.getType() == I.getArgOperand(1)->getType()) {
5549 SDValue LHS = getValue(I.getArgOperand(0));
5550 SDValue RHS = getValue(I.getArgOperand(1));
5551 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurDebugLoc(),
5552 LHS.getValueType(), LHS, RHS));
5555 } else if ((LibInfo->has(LibFunc::fabs) && Name == "fabs") ||
5556 (LibInfo->has(LibFunc::fabsf) && Name == "fabsf") ||
5557 (LibInfo->has(LibFunc::fabsl) && Name == "fabsl")) {
5558 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5559 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5560 I.getType() == I.getArgOperand(0)->getType()) {
5561 SDValue Tmp = getValue(I.getArgOperand(0));
5562 setValue(&I, DAG.getNode(ISD::FABS, getCurDebugLoc(),
5563 Tmp.getValueType(), Tmp));
5566 } else if ((LibInfo->has(LibFunc::sin) && Name == "sin") ||
5567 (LibInfo->has(LibFunc::sinf) && Name == "sinf") ||
5568 (LibInfo->has(LibFunc::sinl) && Name == "sinl")) {
5569 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5570 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5571 I.getType() == I.getArgOperand(0)->getType() &&
5572 I.onlyReadsMemory()) {
5573 SDValue Tmp = getValue(I.getArgOperand(0));
5574 setValue(&I, DAG.getNode(ISD::FSIN, getCurDebugLoc(),
5575 Tmp.getValueType(), Tmp));
5578 } else if ((LibInfo->has(LibFunc::cos) && Name == "cos") ||
5579 (LibInfo->has(LibFunc::cosf) && Name == "cosf") ||
5580 (LibInfo->has(LibFunc::cosl) && Name == "cosl")) {
5581 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5582 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5583 I.getType() == I.getArgOperand(0)->getType() &&
5584 I.onlyReadsMemory()) {
5585 SDValue Tmp = getValue(I.getArgOperand(0));
5586 setValue(&I, DAG.getNode(ISD::FCOS, getCurDebugLoc(),
5587 Tmp.getValueType(), Tmp));
5590 } else if ((LibInfo->has(LibFunc::sqrt) && Name == "sqrt") ||
5591 (LibInfo->has(LibFunc::sqrtf) && Name == "sqrtf") ||
5592 (LibInfo->has(LibFunc::sqrtl) && Name == "sqrtl")) {
5593 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5594 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5595 I.getType() == I.getArgOperand(0)->getType() &&
5596 I.onlyReadsMemory()) {
5597 SDValue Tmp = getValue(I.getArgOperand(0));
5598 setValue(&I, DAG.getNode(ISD::FSQRT, getCurDebugLoc(),
5599 Tmp.getValueType(), Tmp));
5602 } else if ((LibInfo->has(LibFunc::floor) && Name == "floor") ||
5603 (LibInfo->has(LibFunc::floorf) && Name == "floorf") ||
5604 (LibInfo->has(LibFunc::floorl) && Name == "floorl")) {
5605 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5606 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5607 I.getType() == I.getArgOperand(0)->getType()) {
5608 SDValue Tmp = getValue(I.getArgOperand(0));
5609 setValue(&I, DAG.getNode(ISD::FFLOOR, getCurDebugLoc(),
5610 Tmp.getValueType(), Tmp));
5613 } else if ((LibInfo->has(LibFunc::nearbyint) && Name == "nearbyint") ||
5614 (LibInfo->has(LibFunc::nearbyintf) && Name == "nearbyintf") ||
5615 (LibInfo->has(LibFunc::nearbyintl) && Name == "nearbyintl")) {
5616 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5617 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5618 I.getType() == I.getArgOperand(0)->getType()) {
5619 SDValue Tmp = getValue(I.getArgOperand(0));
5620 setValue(&I, DAG.getNode(ISD::FNEARBYINT, getCurDebugLoc(),
5621 Tmp.getValueType(), Tmp));
5624 } else if ((LibInfo->has(LibFunc::ceil) && Name == "ceil") ||
5625 (LibInfo->has(LibFunc::ceilf) && Name == "ceilf") ||
5626 (LibInfo->has(LibFunc::ceill) && Name == "ceill")) {
5627 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5628 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5629 I.getType() == I.getArgOperand(0)->getType()) {
5630 SDValue Tmp = getValue(I.getArgOperand(0));
5631 setValue(&I, DAG.getNode(ISD::FCEIL, getCurDebugLoc(),
5632 Tmp.getValueType(), Tmp));
5635 } else if ((LibInfo->has(LibFunc::rint) && Name == "rint") ||
5636 (LibInfo->has(LibFunc::rintf) && Name == "rintf") ||
5637 (LibInfo->has(LibFunc::rintl) && Name == "rintl")) {
5638 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5639 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5640 I.getType() == I.getArgOperand(0)->getType()) {
5641 SDValue Tmp = getValue(I.getArgOperand(0));
5642 setValue(&I, DAG.getNode(ISD::FRINT, getCurDebugLoc(),
5643 Tmp.getValueType(), Tmp));
5646 } else if ((LibInfo->has(LibFunc::trunc) && Name == "trunc") ||
5647 (LibInfo->has(LibFunc::truncf) && Name == "truncf") ||
5648 (LibInfo->has(LibFunc::truncl) && Name == "truncl")) {
5649 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5650 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5651 I.getType() == I.getArgOperand(0)->getType()) {
5652 SDValue Tmp = getValue(I.getArgOperand(0));
5653 setValue(&I, DAG.getNode(ISD::FTRUNC, getCurDebugLoc(),
5654 Tmp.getValueType(), Tmp));
5657 } else if ((LibInfo->has(LibFunc::log2) && Name == "log2") ||
5658 (LibInfo->has(LibFunc::log2f) && Name == "log2f") ||
5659 (LibInfo->has(LibFunc::log2l) && Name == "log2l")) {
5660 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5661 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5662 I.getType() == I.getArgOperand(0)->getType() &&
5663 I.onlyReadsMemory()) {
5664 SDValue Tmp = getValue(I.getArgOperand(0));
5665 setValue(&I, DAG.getNode(ISD::FLOG2, getCurDebugLoc(),
5666 Tmp.getValueType(), Tmp));
5669 } else if ((LibInfo->has(LibFunc::exp2) && Name == "exp2") ||
5670 (LibInfo->has(LibFunc::exp2f) && Name == "exp2f") ||
5671 (LibInfo->has(LibFunc::exp2l) && Name == "exp2l")) {
5672 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5673 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5674 I.getType() == I.getArgOperand(0)->getType() &&
5675 I.onlyReadsMemory()) {
5676 SDValue Tmp = getValue(I.getArgOperand(0));
5677 setValue(&I, DAG.getNode(ISD::FEXP2, getCurDebugLoc(),
5678 Tmp.getValueType(), Tmp));
5681 } else if (Name == "memcmp") {
5682 if (visitMemCmpCall(I))
5690 Callee = getValue(I.getCalledValue());
5692 Callee = DAG.getExternalSymbol(RenameFn, TLI.getPointerTy());
5694 // Check if we can potentially perform a tail call. More detailed checking is
5695 // be done within LowerCallTo, after more information about the call is known.
5696 LowerCallTo(&I, Callee, I.isTailCall());
5701 /// AsmOperandInfo - This contains information for each constraint that we are
5703 class SDISelAsmOperandInfo : public TargetLowering::AsmOperandInfo {
5705 /// CallOperand - If this is the result output operand or a clobber
5706 /// this is null, otherwise it is the incoming operand to the CallInst.
5707 /// This gets modified as the asm is processed.
5708 SDValue CallOperand;
5710 /// AssignedRegs - If this is a register or register class operand, this
5711 /// contains the set of register corresponding to the operand.
5712 RegsForValue AssignedRegs;
5714 explicit SDISelAsmOperandInfo(const TargetLowering::AsmOperandInfo &info)
5715 : TargetLowering::AsmOperandInfo(info), CallOperand(0,0) {
5718 /// getCallOperandValEVT - Return the EVT of the Value* that this operand
5719 /// corresponds to. If there is no Value* for this operand, it returns
5721 EVT getCallOperandValEVT(LLVMContext &Context,
5722 const TargetLowering &TLI,
5723 const TargetData *TD) const {
5724 if (CallOperandVal == 0) return MVT::Other;
5726 if (isa<BasicBlock>(CallOperandVal))
5727 return TLI.getPointerTy();
5729 llvm::Type *OpTy = CallOperandVal->getType();
5731 // FIXME: code duplicated from TargetLowering::ParseConstraints().
5732 // If this is an indirect operand, the operand is a pointer to the
5735 llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
5737 report_fatal_error("Indirect operand for inline asm not a pointer!");
5738 OpTy = PtrTy->getElementType();
5741 // Look for vector wrapped in a struct. e.g. { <16 x i8> }.
5742 if (StructType *STy = dyn_cast<StructType>(OpTy))
5743 if (STy->getNumElements() == 1)
5744 OpTy = STy->getElementType(0);
5746 // If OpTy is not a single value, it may be a struct/union that we
5747 // can tile with integers.
5748 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
5749 unsigned BitSize = TD->getTypeSizeInBits(OpTy);
5758 OpTy = IntegerType::get(Context, BitSize);
5763 return TLI.getValueType(OpTy, true);
5767 typedef SmallVector<SDISelAsmOperandInfo,16> SDISelAsmOperandInfoVector;
5769 } // end anonymous namespace
5771 /// GetRegistersForValue - Assign registers (virtual or physical) for the
5772 /// specified operand. We prefer to assign virtual registers, to allow the
5773 /// register allocator to handle the assignment process. However, if the asm
5774 /// uses features that we can't model on machineinstrs, we have SDISel do the
5775 /// allocation. This produces generally horrible, but correct, code.
5777 /// OpInfo describes the operand.
5779 static void GetRegistersForValue(SelectionDAG &DAG,
5780 const TargetLowering &TLI,
5782 SDISelAsmOperandInfo &OpInfo) {
5783 LLVMContext &Context = *DAG.getContext();
5785 MachineFunction &MF = DAG.getMachineFunction();
5786 SmallVector<unsigned, 4> Regs;
5788 // If this is a constraint for a single physreg, or a constraint for a
5789 // register class, find it.
5790 std::pair<unsigned, const TargetRegisterClass*> PhysReg =
5791 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
5792 OpInfo.ConstraintVT);
5794 unsigned NumRegs = 1;
5795 if (OpInfo.ConstraintVT != MVT::Other) {
5796 // If this is a FP input in an integer register (or visa versa) insert a bit
5797 // cast of the input value. More generally, handle any case where the input
5798 // value disagrees with the register class we plan to stick this in.
5799 if (OpInfo.Type == InlineAsm::isInput &&
5800 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) {
5801 // Try to convert to the first EVT that the reg class contains. If the
5802 // types are identical size, use a bitcast to convert (e.g. two differing
5804 EVT RegVT = *PhysReg.second->vt_begin();
5805 if (RegVT.getSizeInBits() == OpInfo.ConstraintVT.getSizeInBits()) {
5806 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
5807 RegVT, OpInfo.CallOperand);
5808 OpInfo.ConstraintVT = RegVT;
5809 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) {
5810 // If the input is a FP value and we want it in FP registers, do a
5811 // bitcast to the corresponding integer type. This turns an f64 value
5812 // into i64, which can be passed with two i32 values on a 32-bit
5814 RegVT = EVT::getIntegerVT(Context,
5815 OpInfo.ConstraintVT.getSizeInBits());
5816 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
5817 RegVT, OpInfo.CallOperand);
5818 OpInfo.ConstraintVT = RegVT;
5822 NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT);
5826 EVT ValueVT = OpInfo.ConstraintVT;
5828 // If this is a constraint for a specific physical register, like {r17},
5830 if (unsigned AssignedReg = PhysReg.first) {
5831 const TargetRegisterClass *RC = PhysReg.second;
5832 if (OpInfo.ConstraintVT == MVT::Other)
5833 ValueVT = *RC->vt_begin();
5835 // Get the actual register value type. This is important, because the user
5836 // may have asked for (e.g.) the AX register in i32 type. We need to
5837 // remember that AX is actually i16 to get the right extension.
5838 RegVT = *RC->vt_begin();
5840 // This is a explicit reference to a physical register.
5841 Regs.push_back(AssignedReg);
5843 // If this is an expanded reference, add the rest of the regs to Regs.
5845 TargetRegisterClass::iterator I = RC->begin();
5846 for (; *I != AssignedReg; ++I)
5847 assert(I != RC->end() && "Didn't find reg!");
5849 // Already added the first reg.
5851 for (; NumRegs; --NumRegs, ++I) {
5852 assert(I != RC->end() && "Ran out of registers to allocate!");
5857 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5861 // Otherwise, if this was a reference to an LLVM register class, create vregs
5862 // for this reference.
5863 if (const TargetRegisterClass *RC = PhysReg.second) {
5864 RegVT = *RC->vt_begin();
5865 if (OpInfo.ConstraintVT == MVT::Other)
5868 // Create the appropriate number of virtual registers.
5869 MachineRegisterInfo &RegInfo = MF.getRegInfo();
5870 for (; NumRegs; --NumRegs)
5871 Regs.push_back(RegInfo.createVirtualRegister(RC));
5873 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5877 // Otherwise, we couldn't allocate enough registers for this.
5880 /// visitInlineAsm - Handle a call to an InlineAsm object.
5882 void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) {
5883 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
5885 /// ConstraintOperands - Information about all of the constraints.
5886 SDISelAsmOperandInfoVector ConstraintOperands;
5888 TargetLowering::AsmOperandInfoVector
5889 TargetConstraints = TLI.ParseConstraints(CS);
5891 bool hasMemory = false;
5893 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
5894 unsigned ResNo = 0; // ResNo - The result number of the next output.
5895 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
5896 ConstraintOperands.push_back(SDISelAsmOperandInfo(TargetConstraints[i]));
5897 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
5899 EVT OpVT = MVT::Other;
5901 // Compute the value type for each operand.
5902 switch (OpInfo.Type) {
5903 case InlineAsm::isOutput:
5904 // Indirect outputs just consume an argument.
5905 if (OpInfo.isIndirect) {
5906 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5910 // The return value of the call is this value. As such, there is no
5911 // corresponding argument.
5912 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
5913 if (StructType *STy = dyn_cast<StructType>(CS.getType())) {
5914 OpVT = TLI.getValueType(STy->getElementType(ResNo));
5916 assert(ResNo == 0 && "Asm only has one result!");
5917 OpVT = TLI.getValueType(CS.getType());
5921 case InlineAsm::isInput:
5922 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5924 case InlineAsm::isClobber:
5929 // If this is an input or an indirect output, process the call argument.
5930 // BasicBlocks are labels, currently appearing only in asm's.
5931 if (OpInfo.CallOperandVal) {
5932 if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) {
5933 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]);
5935 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
5938 OpVT = OpInfo.getCallOperandValEVT(*DAG.getContext(), TLI, TD);
5941 OpInfo.ConstraintVT = OpVT;
5943 // Indirect operand accesses access memory.
5944 if (OpInfo.isIndirect)
5947 for (unsigned j = 0, ee = OpInfo.Codes.size(); j != ee; ++j) {
5948 TargetLowering::ConstraintType
5949 CType = TLI.getConstraintType(OpInfo.Codes[j]);
5950 if (CType == TargetLowering::C_Memory) {
5958 SDValue Chain, Flag;
5960 // We won't need to flush pending loads if this asm doesn't touch
5961 // memory and is nonvolatile.
5962 if (hasMemory || IA->hasSideEffects())
5965 Chain = DAG.getRoot();
5967 // Second pass over the constraints: compute which constraint option to use
5968 // and assign registers to constraints that want a specific physreg.
5969 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5970 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5972 // If this is an output operand with a matching input operand, look up the
5973 // matching input. If their types mismatch, e.g. one is an integer, the
5974 // other is floating point, or their sizes are different, flag it as an
5976 if (OpInfo.hasMatchingInput()) {
5977 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
5979 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
5980 std::pair<unsigned, const TargetRegisterClass*> MatchRC =
5981 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
5982 OpInfo.ConstraintVT);
5983 std::pair<unsigned, const TargetRegisterClass*> InputRC =
5984 TLI.getRegForInlineAsmConstraint(Input.ConstraintCode,
5985 Input.ConstraintVT);
5986 if ((OpInfo.ConstraintVT.isInteger() !=
5987 Input.ConstraintVT.isInteger()) ||
5988 (MatchRC.second != InputRC.second)) {
5989 report_fatal_error("Unsupported asm: input constraint"
5990 " with a matching output constraint of"
5991 " incompatible type!");
5993 Input.ConstraintVT = OpInfo.ConstraintVT;
5997 // Compute the constraint code and ConstraintType to use.
5998 TLI.ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, &DAG);
6000 // If this is a memory input, and if the operand is not indirect, do what we
6001 // need to to provide an address for the memory input.
6002 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
6003 !OpInfo.isIndirect) {
6004 assert((OpInfo.isMultipleAlternative ||
6005 (OpInfo.Type == InlineAsm::isInput)) &&
6006 "Can only indirectify direct input operands!");
6008 // Memory operands really want the address of the value. If we don't have
6009 // an indirect input, put it in the constpool if we can, otherwise spill
6010 // it to a stack slot.
6011 // TODO: This isn't quite right. We need to handle these according to
6012 // the addressing mode that the constraint wants. Also, this may take
6013 // an additional register for the computation and we don't want that
6016 // If the operand is a float, integer, or vector constant, spill to a
6017 // constant pool entry to get its address.
6018 const Value *OpVal = OpInfo.CallOperandVal;
6019 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
6020 isa<ConstantVector>(OpVal) || isa<ConstantDataVector>(OpVal)) {
6021 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal),
6022 TLI.getPointerTy());
6024 // Otherwise, create a stack slot and emit a store to it before the
6026 Type *Ty = OpVal->getType();
6027 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
6028 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(Ty);
6029 MachineFunction &MF = DAG.getMachineFunction();
6030 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
6031 SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
6032 Chain = DAG.getStore(Chain, getCurDebugLoc(),
6033 OpInfo.CallOperand, StackSlot,
6034 MachinePointerInfo::getFixedStack(SSFI),
6036 OpInfo.CallOperand = StackSlot;
6039 // There is no longer a Value* corresponding to this operand.
6040 OpInfo.CallOperandVal = 0;
6042 // It is now an indirect operand.
6043 OpInfo.isIndirect = true;
6046 // If this constraint is for a specific register, allocate it before
6048 if (OpInfo.ConstraintType == TargetLowering::C_Register)
6049 GetRegistersForValue(DAG, TLI, getCurDebugLoc(), OpInfo);
6052 // Second pass - Loop over all of the operands, assigning virtual or physregs
6053 // to register class operands.
6054 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6055 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6057 // C_Register operands have already been allocated, Other/Memory don't need
6059 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
6060 GetRegistersForValue(DAG, TLI, getCurDebugLoc(), OpInfo);
6063 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
6064 std::vector<SDValue> AsmNodeOperands;
6065 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain
6066 AsmNodeOperands.push_back(
6067 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(),
6068 TLI.getPointerTy()));
6070 // If we have a !srcloc metadata node associated with it, we want to attach
6071 // this to the ultimately generated inline asm machineinstr. To do this, we
6072 // pass in the third operand as this (potentially null) inline asm MDNode.
6073 const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc");
6074 AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc));
6076 // Remember the HasSideEffect and AlignStack bits as operand 3.
6077 unsigned ExtraInfo = 0;
6078 if (IA->hasSideEffects())
6079 ExtraInfo |= InlineAsm::Extra_HasSideEffects;
6080 if (IA->isAlignStack())
6081 ExtraInfo |= InlineAsm::Extra_IsAlignStack;
6082 AsmNodeOperands.push_back(DAG.getTargetConstant(ExtraInfo,
6083 TLI.getPointerTy()));
6085 // Loop over all of the inputs, copying the operand values into the
6086 // appropriate registers and processing the output regs.
6087 RegsForValue RetValRegs;
6089 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
6090 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
6092 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6093 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6095 switch (OpInfo.Type) {
6096 case InlineAsm::isOutput: {
6097 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
6098 OpInfo.ConstraintType != TargetLowering::C_Register) {
6099 // Memory output, or 'other' output (e.g. 'X' constraint).
6100 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
6102 // Add information to the INLINEASM node to know about this output.
6103 unsigned OpFlags = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
6104 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlags,
6105 TLI.getPointerTy()));
6106 AsmNodeOperands.push_back(OpInfo.CallOperand);
6110 // Otherwise, this is a register or register class output.
6112 // Copy the output from the appropriate register. Find a register that
6114 if (OpInfo.AssignedRegs.Regs.empty()) {
6115 LLVMContext &Ctx = *DAG.getContext();
6116 Ctx.emitError(CS.getInstruction(),
6117 "couldn't allocate output register for constraint '" +
6118 Twine(OpInfo.ConstraintCode) + "'");
6122 // If this is an indirect operand, store through the pointer after the
6124 if (OpInfo.isIndirect) {
6125 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
6126 OpInfo.CallOperandVal));
6128 // This is the result value of the call.
6129 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
6130 // Concatenate this output onto the outputs list.
6131 RetValRegs.append(OpInfo.AssignedRegs);
6134 // Add information to the INLINEASM node to know that this register is
6136 OpInfo.AssignedRegs.AddInlineAsmOperands(OpInfo.isEarlyClobber ?
6137 InlineAsm::Kind_RegDefEarlyClobber :
6138 InlineAsm::Kind_RegDef,
6145 case InlineAsm::isInput: {
6146 SDValue InOperandVal = OpInfo.CallOperand;
6148 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint?
6149 // If this is required to match an output register we have already set,
6150 // just use its register.
6151 unsigned OperandNo = OpInfo.getMatchedOperand();
6153 // Scan until we find the definition we already emitted of this operand.
6154 // When we find it, create a RegsForValue operand.
6155 unsigned CurOp = InlineAsm::Op_FirstOperand;
6156 for (; OperandNo; --OperandNo) {
6157 // Advance to the next operand.
6159 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6160 assert((InlineAsm::isRegDefKind(OpFlag) ||
6161 InlineAsm::isRegDefEarlyClobberKind(OpFlag) ||
6162 InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?");
6163 CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1;
6167 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6168 if (InlineAsm::isRegDefKind(OpFlag) ||
6169 InlineAsm::isRegDefEarlyClobberKind(OpFlag)) {
6170 // Add (OpFlag&0xffff)>>3 registers to MatchedRegs.
6171 if (OpInfo.isIndirect) {
6172 // This happens on gcc/testsuite/gcc.dg/pr8788-1.c
6173 LLVMContext &Ctx = *DAG.getContext();
6174 Ctx.emitError(CS.getInstruction(), "inline asm not supported yet:"
6175 " don't know how to handle tied "
6176 "indirect register inputs");
6179 RegsForValue MatchedRegs;
6180 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
6181 EVT RegVT = AsmNodeOperands[CurOp+1].getValueType();
6182 MatchedRegs.RegVTs.push_back(RegVT);
6183 MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo();
6184 for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag);
6186 MatchedRegs.Regs.push_back
6187 (RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT)));
6189 // Use the produced MatchedRegs object to
6190 MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
6192 MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse,
6193 true, OpInfo.getMatchedOperand(),
6194 DAG, AsmNodeOperands);
6198 assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!");
6199 assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 &&
6200 "Unexpected number of operands");
6201 // Add information to the INLINEASM node to know about this input.
6202 // See InlineAsm.h isUseOperandTiedToDef.
6203 OpFlag = InlineAsm::getFlagWordForMatchingOp(OpFlag,
6204 OpInfo.getMatchedOperand());
6205 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlag,
6206 TLI.getPointerTy()));
6207 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]);
6211 // Treat indirect 'X' constraint as memory.
6212 if (OpInfo.ConstraintType == TargetLowering::C_Other &&
6214 OpInfo.ConstraintType = TargetLowering::C_Memory;
6216 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
6217 std::vector<SDValue> Ops;
6218 TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode,
6221 LLVMContext &Ctx = *DAG.getContext();
6222 Ctx.emitError(CS.getInstruction(),
6223 "invalid operand for inline asm constraint '" +
6224 Twine(OpInfo.ConstraintCode) + "'");
6228 // Add information to the INLINEASM node to know about this input.
6229 unsigned ResOpType =
6230 InlineAsm::getFlagWord(InlineAsm::Kind_Imm, Ops.size());
6231 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
6232 TLI.getPointerTy()));
6233 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
6237 if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
6238 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
6239 assert(InOperandVal.getValueType() == TLI.getPointerTy() &&
6240 "Memory operands expect pointer values");
6242 // Add information to the INLINEASM node to know about this input.
6243 unsigned ResOpType = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
6244 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
6245 TLI.getPointerTy()));
6246 AsmNodeOperands.push_back(InOperandVal);
6250 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
6251 OpInfo.ConstraintType == TargetLowering::C_Register) &&
6252 "Unknown constraint type!");
6254 // TODO: Support this.
6255 if (OpInfo.isIndirect) {
6256 LLVMContext &Ctx = *DAG.getContext();
6257 Ctx.emitError(CS.getInstruction(),
6258 "Don't know how to handle indirect register inputs yet "
6259 "for constraint '" + Twine(OpInfo.ConstraintCode) + "'");
6263 // Copy the input into the appropriate registers.
6264 if (OpInfo.AssignedRegs.Regs.empty()) {
6265 LLVMContext &Ctx = *DAG.getContext();
6266 Ctx.emitError(CS.getInstruction(),
6267 "couldn't allocate input reg for constraint '" +
6268 Twine(OpInfo.ConstraintCode) + "'");
6272 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
6275 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, false, 0,
6276 DAG, AsmNodeOperands);
6279 case InlineAsm::isClobber: {
6280 // Add the clobbered value to the operand list, so that the register
6281 // allocator is aware that the physreg got clobbered.
6282 if (!OpInfo.AssignedRegs.Regs.empty())
6283 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_Clobber,
6291 // Finish up input operands. Set the input chain and add the flag last.
6292 AsmNodeOperands[InlineAsm::Op_InputChain] = Chain;
6293 if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
6295 Chain = DAG.getNode(ISD::INLINEASM, getCurDebugLoc(),
6296 DAG.getVTList(MVT::Other, MVT::Glue),
6297 &AsmNodeOperands[0], AsmNodeOperands.size());
6298 Flag = Chain.getValue(1);
6300 // If this asm returns a register value, copy the result from that register
6301 // and set it as the value of the call.
6302 if (!RetValRegs.Regs.empty()) {
6303 SDValue Val = RetValRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(),
6306 // FIXME: Why don't we do this for inline asms with MRVs?
6307 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) {
6308 EVT ResultType = TLI.getValueType(CS.getType());
6310 // If any of the results of the inline asm is a vector, it may have the
6311 // wrong width/num elts. This can happen for register classes that can
6312 // contain multiple different value types. The preg or vreg allocated may
6313 // not have the same VT as was expected. Convert it to the right type
6314 // with bit_convert.
6315 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) {
6316 Val = DAG.getNode(ISD::BITCAST, getCurDebugLoc(),
6319 } else if (ResultType != Val.getValueType() &&
6320 ResultType.isInteger() && Val.getValueType().isInteger()) {
6321 // If a result value was tied to an input value, the computed result may
6322 // have a wider width than the expected result. Extract the relevant
6324 Val = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), ResultType, Val);
6327 assert(ResultType == Val.getValueType() && "Asm result value mismatch!");
6330 setValue(CS.getInstruction(), Val);
6331 // Don't need to use this as a chain in this case.
6332 if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty())
6336 std::vector<std::pair<SDValue, const Value *> > StoresToEmit;
6338 // Process indirect outputs, first output all of the flagged copies out of
6340 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
6341 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
6342 const Value *Ptr = IndirectStoresToEmit[i].second;
6343 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(),
6345 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
6348 // Emit the non-flagged stores from the physregs.
6349 SmallVector<SDValue, 8> OutChains;
6350 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) {
6351 SDValue Val = DAG.getStore(Chain, getCurDebugLoc(),
6352 StoresToEmit[i].first,
6353 getValue(StoresToEmit[i].second),
6354 MachinePointerInfo(StoresToEmit[i].second),
6356 OutChains.push_back(Val);
6359 if (!OutChains.empty())
6360 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
6361 &OutChains[0], OutChains.size());
6366 void SelectionDAGBuilder::visitVAStart(const CallInst &I) {
6367 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurDebugLoc(),
6368 MVT::Other, getRoot(),
6369 getValue(I.getArgOperand(0)),
6370 DAG.getSrcValue(I.getArgOperand(0))));
6373 void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) {
6374 const TargetData &TD = *TLI.getTargetData();
6375 SDValue V = DAG.getVAArg(TLI.getValueType(I.getType()), getCurDebugLoc(),
6376 getRoot(), getValue(I.getOperand(0)),
6377 DAG.getSrcValue(I.getOperand(0)),
6378 TD.getABITypeAlignment(I.getType()));
6380 DAG.setRoot(V.getValue(1));
6383 void SelectionDAGBuilder::visitVAEnd(const CallInst &I) {
6384 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurDebugLoc(),
6385 MVT::Other, getRoot(),
6386 getValue(I.getArgOperand(0)),
6387 DAG.getSrcValue(I.getArgOperand(0))));
6390 void SelectionDAGBuilder::visitVACopy(const CallInst &I) {
6391 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurDebugLoc(),
6392 MVT::Other, getRoot(),
6393 getValue(I.getArgOperand(0)),
6394 getValue(I.getArgOperand(1)),
6395 DAG.getSrcValue(I.getArgOperand(0)),
6396 DAG.getSrcValue(I.getArgOperand(1))));
6399 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
6400 /// implementation, which just calls LowerCall.
6401 /// FIXME: When all targets are
6402 /// migrated to using LowerCall, this hook should be integrated into SDISel.
6403 std::pair<SDValue, SDValue>
6404 TargetLowering::LowerCallTo(TargetLowering::CallLoweringInfo &CLI) const {
6405 // Handle all of the outgoing arguments.
6407 CLI.OutVals.clear();
6408 ArgListTy &Args = CLI.Args;
6409 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
6410 SmallVector<EVT, 4> ValueVTs;
6411 ComputeValueVTs(*this, Args[i].Ty, ValueVTs);
6412 for (unsigned Value = 0, NumValues = ValueVTs.size();
6413 Value != NumValues; ++Value) {
6414 EVT VT = ValueVTs[Value];
6415 Type *ArgTy = VT.getTypeForEVT(CLI.RetTy->getContext());
6416 SDValue Op = SDValue(Args[i].Node.getNode(),
6417 Args[i].Node.getResNo() + Value);
6418 ISD::ArgFlagsTy Flags;
6419 unsigned OriginalAlignment =
6420 getTargetData()->getABITypeAlignment(ArgTy);
6426 if (Args[i].isInReg)
6430 if (Args[i].isByVal) {
6432 PointerType *Ty = cast<PointerType>(Args[i].Ty);
6433 Type *ElementTy = Ty->getElementType();
6434 Flags.setByValSize(getTargetData()->getTypeAllocSize(ElementTy));
6435 // For ByVal, alignment should come from FE. BE will guess if this
6436 // info is not there but there are cases it cannot get right.
6437 unsigned FrameAlign;
6438 if (Args[i].Alignment)
6439 FrameAlign = Args[i].Alignment;
6441 FrameAlign = getByValTypeAlignment(ElementTy);
6442 Flags.setByValAlign(FrameAlign);
6446 Flags.setOrigAlign(OriginalAlignment);
6448 EVT PartVT = getRegisterType(CLI.RetTy->getContext(), VT);
6449 unsigned NumParts = getNumRegisters(CLI.RetTy->getContext(), VT);
6450 SmallVector<SDValue, 4> Parts(NumParts);
6451 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
6454 ExtendKind = ISD::SIGN_EXTEND;
6455 else if (Args[i].isZExt)
6456 ExtendKind = ISD::ZERO_EXTEND;
6458 getCopyToParts(CLI.DAG, CLI.DL, Op, &Parts[0], NumParts,
6459 PartVT, ExtendKind);
6461 for (unsigned j = 0; j != NumParts; ++j) {
6462 // if it isn't first piece, alignment must be 1
6463 ISD::OutputArg MyFlags(Flags, Parts[j].getValueType(),
6464 i < CLI.NumFixedArgs);
6465 if (NumParts > 1 && j == 0)
6466 MyFlags.Flags.setSplit();
6468 MyFlags.Flags.setOrigAlign(1);
6470 CLI.Outs.push_back(MyFlags);
6471 CLI.OutVals.push_back(Parts[j]);
6476 // Handle the incoming return values from the call.
6478 SmallVector<EVT, 4> RetTys;
6479 ComputeValueVTs(*this, CLI.RetTy, RetTys);
6480 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
6482 EVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
6483 unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
6484 for (unsigned i = 0; i != NumRegs; ++i) {
6485 ISD::InputArg MyFlags;
6486 MyFlags.VT = RegisterVT.getSimpleVT();
6487 MyFlags.Used = CLI.IsReturnValueUsed;
6489 MyFlags.Flags.setSExt();
6491 MyFlags.Flags.setZExt();
6493 MyFlags.Flags.setInReg();
6494 CLI.Ins.push_back(MyFlags);
6498 SmallVector<SDValue, 4> InVals;
6499 CLI.Chain = LowerCall(CLI, InVals);
6501 // Verify that the target's LowerCall behaved as expected.
6502 assert(CLI.Chain.getNode() && CLI.Chain.getValueType() == MVT::Other &&
6503 "LowerCall didn't return a valid chain!");
6504 assert((!CLI.IsTailCall || InVals.empty()) &&
6505 "LowerCall emitted a return value for a tail call!");
6506 assert((CLI.IsTailCall || InVals.size() == CLI.Ins.size()) &&
6507 "LowerCall didn't emit the correct number of values!");
6509 // For a tail call, the return value is merely live-out and there aren't
6510 // any nodes in the DAG representing it. Return a special value to
6511 // indicate that a tail call has been emitted and no more Instructions
6512 // should be processed in the current block.
6513 if (CLI.IsTailCall) {
6514 CLI.DAG.setRoot(CLI.Chain);
6515 return std::make_pair(SDValue(), SDValue());
6518 DEBUG(for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++i) {
6519 assert(InVals[i].getNode() &&
6520 "LowerCall emitted a null value!");
6521 assert(EVT(CLI.Ins[i].VT) == InVals[i].getValueType() &&
6522 "LowerCall emitted a value with the wrong type!");
6525 // Collect the legal value parts into potentially illegal values
6526 // that correspond to the original function's return values.
6527 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6529 AssertOp = ISD::AssertSext;
6530 else if (CLI.RetZExt)
6531 AssertOp = ISD::AssertZext;
6532 SmallVector<SDValue, 4> ReturnValues;
6533 unsigned CurReg = 0;
6534 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
6536 EVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
6537 unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
6539 ReturnValues.push_back(getCopyFromParts(CLI.DAG, CLI.DL, &InVals[CurReg],
6540 NumRegs, RegisterVT, VT,
6545 // For a function returning void, there is no return value. We can't create
6546 // such a node, so we just return a null return value in that case. In
6547 // that case, nothing will actually look at the value.
6548 if (ReturnValues.empty())
6549 return std::make_pair(SDValue(), CLI.Chain);
6551 SDValue Res = CLI.DAG.getNode(ISD::MERGE_VALUES, CLI.DL,
6552 CLI.DAG.getVTList(&RetTys[0], RetTys.size()),
6553 &ReturnValues[0], ReturnValues.size());
6554 return std::make_pair(Res, CLI.Chain);
6557 void TargetLowering::LowerOperationWrapper(SDNode *N,
6558 SmallVectorImpl<SDValue> &Results,
6559 SelectionDAG &DAG) const {
6560 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
6562 Results.push_back(Res);
6565 SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6566 llvm_unreachable("LowerOperation not implemented for this target!");
6570 SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) {
6571 SDValue Op = getNonRegisterValue(V);
6572 assert((Op.getOpcode() != ISD::CopyFromReg ||
6573 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
6574 "Copy from a reg to the same reg!");
6575 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg");
6577 RegsForValue RFV(V->getContext(), TLI, Reg, V->getType());
6578 SDValue Chain = DAG.getEntryNode();
6579 RFV.getCopyToRegs(Op, DAG, getCurDebugLoc(), Chain, 0);
6580 PendingExports.push_back(Chain);
6583 #include "llvm/CodeGen/SelectionDAGISel.h"
6585 /// isOnlyUsedInEntryBlock - If the specified argument is only used in the
6586 /// entry block, return true. This includes arguments used by switches, since
6587 /// the switch may expand into multiple basic blocks.
6588 static bool isOnlyUsedInEntryBlock(const Argument *A, bool FastISel) {
6589 // With FastISel active, we may be splitting blocks, so force creation
6590 // of virtual registers for all non-dead arguments.
6592 return A->use_empty();
6594 const BasicBlock *Entry = A->getParent()->begin();
6595 for (Value::const_use_iterator UI = A->use_begin(), E = A->use_end();
6597 const User *U = *UI;
6598 if (cast<Instruction>(U)->getParent() != Entry || isa<SwitchInst>(U))
6599 return false; // Use not in entry block.
6604 void SelectionDAGISel::LowerArguments(const BasicBlock *LLVMBB) {
6605 // If this is the entry block, emit arguments.
6606 const Function &F = *LLVMBB->getParent();
6607 SelectionDAG &DAG = SDB->DAG;
6608 DebugLoc dl = SDB->getCurDebugLoc();
6609 const TargetData *TD = TLI.getTargetData();
6610 SmallVector<ISD::InputArg, 16> Ins;
6612 // Check whether the function can return without sret-demotion.
6613 SmallVector<ISD::OutputArg, 4> Outs;
6614 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
6617 if (!FuncInfo->CanLowerReturn) {
6618 // Put in an sret pointer parameter before all the other parameters.
6619 SmallVector<EVT, 1> ValueVTs;
6620 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
6622 // NOTE: Assuming that a pointer will never break down to more than one VT
6624 ISD::ArgFlagsTy Flags;
6626 EVT RegisterVT = TLI.getRegisterType(*DAG.getContext(), ValueVTs[0]);
6627 ISD::InputArg RetArg(Flags, RegisterVT, true);
6628 Ins.push_back(RetArg);
6631 // Set up the incoming argument description vector.
6633 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end();
6634 I != E; ++I, ++Idx) {
6635 SmallVector<EVT, 4> ValueVTs;
6636 ComputeValueVTs(TLI, I->getType(), ValueVTs);
6637 bool isArgValueUsed = !I->use_empty();
6638 for (unsigned Value = 0, NumValues = ValueVTs.size();
6639 Value != NumValues; ++Value) {
6640 EVT VT = ValueVTs[Value];
6641 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
6642 ISD::ArgFlagsTy Flags;
6643 unsigned OriginalAlignment =
6644 TD->getABITypeAlignment(ArgTy);
6646 if (F.paramHasAttr(Idx, Attribute::ZExt))
6648 if (F.paramHasAttr(Idx, Attribute::SExt))
6650 if (F.paramHasAttr(Idx, Attribute::InReg))
6652 if (F.paramHasAttr(Idx, Attribute::StructRet))
6654 if (F.paramHasAttr(Idx, Attribute::ByVal)) {
6656 PointerType *Ty = cast<PointerType>(I->getType());
6657 Type *ElementTy = Ty->getElementType();
6658 Flags.setByValSize(TD->getTypeAllocSize(ElementTy));
6659 // For ByVal, alignment should be passed from FE. BE will guess if
6660 // this info is not there but there are cases it cannot get right.
6661 unsigned FrameAlign;
6662 if (F.getParamAlignment(Idx))
6663 FrameAlign = F.getParamAlignment(Idx);
6665 FrameAlign = TLI.getByValTypeAlignment(ElementTy);
6666 Flags.setByValAlign(FrameAlign);
6668 if (F.paramHasAttr(Idx, Attribute::Nest))
6670 Flags.setOrigAlign(OriginalAlignment);
6672 EVT RegisterVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6673 unsigned NumRegs = TLI.getNumRegisters(*CurDAG->getContext(), VT);
6674 for (unsigned i = 0; i != NumRegs; ++i) {
6675 ISD::InputArg MyFlags(Flags, RegisterVT, isArgValueUsed);
6676 if (NumRegs > 1 && i == 0)
6677 MyFlags.Flags.setSplit();
6678 // if it isn't first piece, alignment must be 1
6680 MyFlags.Flags.setOrigAlign(1);
6681 Ins.push_back(MyFlags);
6686 // Call the target to set up the argument values.
6687 SmallVector<SDValue, 8> InVals;
6688 SDValue NewRoot = TLI.LowerFormalArguments(DAG.getRoot(), F.getCallingConv(),
6692 // Verify that the target's LowerFormalArguments behaved as expected.
6693 assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other &&
6694 "LowerFormalArguments didn't return a valid chain!");
6695 assert(InVals.size() == Ins.size() &&
6696 "LowerFormalArguments didn't emit the correct number of values!");
6698 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
6699 assert(InVals[i].getNode() &&
6700 "LowerFormalArguments emitted a null value!");
6701 assert(EVT(Ins[i].VT) == InVals[i].getValueType() &&
6702 "LowerFormalArguments emitted a value with the wrong type!");
6706 // Update the DAG with the new chain value resulting from argument lowering.
6707 DAG.setRoot(NewRoot);
6709 // Set up the argument values.
6712 if (!FuncInfo->CanLowerReturn) {
6713 // Create a virtual register for the sret pointer, and put in a copy
6714 // from the sret argument into it.
6715 SmallVector<EVT, 1> ValueVTs;
6716 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
6717 EVT VT = ValueVTs[0];
6718 EVT RegVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6719 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6720 SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1,
6721 RegVT, VT, AssertOp);
6723 MachineFunction& MF = SDB->DAG.getMachineFunction();
6724 MachineRegisterInfo& RegInfo = MF.getRegInfo();
6725 unsigned SRetReg = RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT));
6726 FuncInfo->DemoteRegister = SRetReg;
6727 NewRoot = SDB->DAG.getCopyToReg(NewRoot, SDB->getCurDebugLoc(),
6729 DAG.setRoot(NewRoot);
6731 // i indexes lowered arguments. Bump it past the hidden sret argument.
6732 // Idx indexes LLVM arguments. Don't touch it.
6736 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
6738 SmallVector<SDValue, 4> ArgValues;
6739 SmallVector<EVT, 4> ValueVTs;
6740 ComputeValueVTs(TLI, I->getType(), ValueVTs);
6741 unsigned NumValues = ValueVTs.size();
6743 // If this argument is unused then remember its value. It is used to generate
6744 // debugging information.
6745 if (I->use_empty() && NumValues)
6746 SDB->setUnusedArgValue(I, InVals[i]);
6748 for (unsigned Val = 0; Val != NumValues; ++Val) {
6749 EVT VT = ValueVTs[Val];
6750 EVT PartVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6751 unsigned NumParts = TLI.getNumRegisters(*CurDAG->getContext(), VT);
6753 if (!I->use_empty()) {
6754 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6755 if (F.paramHasAttr(Idx, Attribute::SExt))
6756 AssertOp = ISD::AssertSext;
6757 else if (F.paramHasAttr(Idx, Attribute::ZExt))
6758 AssertOp = ISD::AssertZext;
6760 ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i],
6761 NumParts, PartVT, VT,
6768 // We don't need to do anything else for unused arguments.
6769 if (ArgValues.empty())
6772 // Note down frame index.
6773 if (FrameIndexSDNode *FI =
6774 dyn_cast<FrameIndexSDNode>(ArgValues[0].getNode()))
6775 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
6777 SDValue Res = DAG.getMergeValues(&ArgValues[0], NumValues,
6778 SDB->getCurDebugLoc());
6780 SDB->setValue(I, Res);
6781 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::BUILD_PAIR) {
6782 if (LoadSDNode *LNode =
6783 dyn_cast<LoadSDNode>(Res.getOperand(0).getNode()))
6784 if (FrameIndexSDNode *FI =
6785 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode()))
6786 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
6789 // If this argument is live outside of the entry block, insert a copy from
6790 // wherever we got it to the vreg that other BB's will reference it as.
6791 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::CopyFromReg) {
6792 // If we can, though, try to skip creating an unnecessary vreg.
6793 // FIXME: This isn't very clean... it would be nice to make this more
6794 // general. It's also subtly incompatible with the hacks FastISel
6796 unsigned Reg = cast<RegisterSDNode>(Res.getOperand(1))->getReg();
6797 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
6798 FuncInfo->ValueMap[I] = Reg;
6802 if (!isOnlyUsedInEntryBlock(I, TM.Options.EnableFastISel)) {
6803 FuncInfo->InitializeRegForValue(I);
6804 SDB->CopyToExportRegsIfNeeded(I);
6808 assert(i == InVals.size() && "Argument register count mismatch!");
6810 // Finally, if the target has anything special to do, allow it to do so.
6811 // FIXME: this should insert code into the DAG!
6812 EmitFunctionEntryCode();
6815 /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
6816 /// ensure constants are generated when needed. Remember the virtual registers
6817 /// that need to be added to the Machine PHI nodes as input. We cannot just
6818 /// directly add them, because expansion might result in multiple MBB's for one
6819 /// BB. As such, the start of the BB might correspond to a different MBB than
6823 SelectionDAGBuilder::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
6824 const TerminatorInst *TI = LLVMBB->getTerminator();
6826 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
6828 // Check successor nodes' PHI nodes that expect a constant to be available
6830 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
6831 const BasicBlock *SuccBB = TI->getSuccessor(succ);
6832 if (!isa<PHINode>(SuccBB->begin())) continue;
6833 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
6835 // If this terminator has multiple identical successors (common for
6836 // switches), only handle each succ once.
6837 if (!SuccsHandled.insert(SuccMBB)) continue;
6839 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
6841 // At this point we know that there is a 1-1 correspondence between LLVM PHI
6842 // nodes and Machine PHI nodes, but the incoming operands have not been
6844 for (BasicBlock::const_iterator I = SuccBB->begin();
6845 const PHINode *PN = dyn_cast<PHINode>(I); ++I) {
6846 // Ignore dead phi's.
6847 if (PN->use_empty()) continue;
6850 if (PN->getType()->isEmptyTy())
6854 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
6856 if (const Constant *C = dyn_cast<Constant>(PHIOp)) {
6857 unsigned &RegOut = ConstantsOut[C];
6859 RegOut = FuncInfo.CreateRegs(C->getType());
6860 CopyValueToVirtualRegister(C, RegOut);
6864 DenseMap<const Value *, unsigned>::iterator I =
6865 FuncInfo.ValueMap.find(PHIOp);
6866 if (I != FuncInfo.ValueMap.end())
6869 assert(isa<AllocaInst>(PHIOp) &&
6870 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
6871 "Didn't codegen value into a register!??");
6872 Reg = FuncInfo.CreateRegs(PHIOp->getType());
6873 CopyValueToVirtualRegister(PHIOp, Reg);
6877 // Remember that this register needs to added to the machine PHI node as
6878 // the input for this MBB.
6879 SmallVector<EVT, 4> ValueVTs;
6880 ComputeValueVTs(TLI, PN->getType(), ValueVTs);
6881 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
6882 EVT VT = ValueVTs[vti];
6883 unsigned NumRegisters = TLI.getNumRegisters(*DAG.getContext(), VT);
6884 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
6885 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
6886 Reg += NumRegisters;
6890 ConstantsOut.clear();