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 if (isa<InlineAsm>(Callee))
1834 LowerCallTo(&I, getValue(Callee), false, LandingPad);
1836 // If the value of the invoke is used outside of its defining block, make it
1837 // available as a virtual register.
1838 CopyToExportRegsIfNeeded(&I);
1840 // Update successor info
1841 addSuccessorWithWeight(InvokeMBB, Return);
1842 addSuccessorWithWeight(InvokeMBB, LandingPad);
1844 // Drop into normal successor.
1845 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
1846 MVT::Other, getControlRoot(),
1847 DAG.getBasicBlock(Return)));
1850 void SelectionDAGBuilder::visitResume(const ResumeInst &RI) {
1851 llvm_unreachable("SelectionDAGBuilder shouldn't visit resume instructions!");
1854 void SelectionDAGBuilder::visitLandingPad(const LandingPadInst &LP) {
1855 assert(FuncInfo.MBB->isLandingPad() &&
1856 "Call to landingpad not in landing pad!");
1858 MachineBasicBlock *MBB = FuncInfo.MBB;
1859 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
1860 AddLandingPadInfo(LP, MMI, MBB);
1862 // If there aren't registers to copy the values into (e.g., during SjLj
1863 // exceptions), then don't bother to create these DAG nodes.
1864 if (TLI.getExceptionPointerRegister() == 0 &&
1865 TLI.getExceptionSelectorRegister() == 0)
1868 SmallVector<EVT, 2> ValueVTs;
1869 ComputeValueVTs(TLI, LP.getType(), ValueVTs);
1871 // Insert the EXCEPTIONADDR instruction.
1872 assert(FuncInfo.MBB->isLandingPad() &&
1873 "Call to eh.exception not in landing pad!");
1874 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
1876 Ops[0] = DAG.getRoot();
1877 SDValue Op1 = DAG.getNode(ISD::EXCEPTIONADDR, getCurDebugLoc(), VTs, Ops, 1);
1878 SDValue Chain = Op1.getValue(1);
1880 // Insert the EHSELECTION instruction.
1881 VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
1884 SDValue Op2 = DAG.getNode(ISD::EHSELECTION, getCurDebugLoc(), VTs, Ops, 2);
1885 Chain = Op2.getValue(1);
1886 Op2 = DAG.getSExtOrTrunc(Op2, getCurDebugLoc(), MVT::i32);
1890 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
1891 DAG.getVTList(&ValueVTs[0], ValueVTs.size()),
1894 std::pair<SDValue, SDValue> RetPair = std::make_pair(Res, Chain);
1895 setValue(&LP, RetPair.first);
1896 DAG.setRoot(RetPair.second);
1899 /// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for
1900 /// small case ranges).
1901 bool SelectionDAGBuilder::handleSmallSwitchRange(CaseRec& CR,
1902 CaseRecVector& WorkList,
1904 MachineBasicBlock *Default,
1905 MachineBasicBlock *SwitchBB) {
1906 // Size is the number of Cases represented by this range.
1907 size_t Size = CR.Range.second - CR.Range.first;
1911 // Get the MachineFunction which holds the current MBB. This is used when
1912 // inserting any additional MBBs necessary to represent the switch.
1913 MachineFunction *CurMF = FuncInfo.MF;
1915 // Figure out which block is immediately after the current one.
1916 MachineBasicBlock *NextBlock = 0;
1917 MachineFunction::iterator BBI = CR.CaseBB;
1919 if (++BBI != FuncInfo.MF->end())
1922 // If any two of the cases has the same destination, and if one value
1923 // is the same as the other, but has one bit unset that the other has set,
1924 // use bit manipulation to do two compares at once. For example:
1925 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
1926 // TODO: This could be extended to merge any 2 cases in switches with 3 cases.
1927 // TODO: Handle cases where CR.CaseBB != SwitchBB.
1928 if (Size == 2 && CR.CaseBB == SwitchBB) {
1929 Case &Small = *CR.Range.first;
1930 Case &Big = *(CR.Range.second-1);
1932 if (Small.Low == Small.High && Big.Low == Big.High && Small.BB == Big.BB) {
1933 const APInt& SmallValue = cast<ConstantInt>(Small.Low)->getValue();
1934 const APInt& BigValue = cast<ConstantInt>(Big.Low)->getValue();
1936 // Check that there is only one bit different.
1937 if (BigValue.countPopulation() == SmallValue.countPopulation() + 1 &&
1938 (SmallValue | BigValue) == BigValue) {
1939 // Isolate the common bit.
1940 APInt CommonBit = BigValue & ~SmallValue;
1941 assert((SmallValue | CommonBit) == BigValue &&
1942 CommonBit.countPopulation() == 1 && "Not a common bit?");
1944 SDValue CondLHS = getValue(SV);
1945 EVT VT = CondLHS.getValueType();
1946 DebugLoc DL = getCurDebugLoc();
1948 SDValue Or = DAG.getNode(ISD::OR, DL, VT, CondLHS,
1949 DAG.getConstant(CommonBit, VT));
1950 SDValue Cond = DAG.getSetCC(DL, MVT::i1,
1951 Or, DAG.getConstant(BigValue, VT),
1954 // Update successor info.
1955 addSuccessorWithWeight(SwitchBB, Small.BB);
1956 addSuccessorWithWeight(SwitchBB, Default);
1958 // Insert the true branch.
1959 SDValue BrCond = DAG.getNode(ISD::BRCOND, DL, MVT::Other,
1960 getControlRoot(), Cond,
1961 DAG.getBasicBlock(Small.BB));
1963 // Insert the false branch.
1964 BrCond = DAG.getNode(ISD::BR, DL, MVT::Other, BrCond,
1965 DAG.getBasicBlock(Default));
1967 DAG.setRoot(BrCond);
1973 // Order cases by weight so the most likely case will be checked first.
1974 BranchProbabilityInfo *BPI = FuncInfo.BPI;
1976 for (CaseItr I = CR.Range.first, IE = CR.Range.second; I != IE; ++I) {
1977 uint32_t IWeight = BPI->getEdgeWeight(SwitchBB->getBasicBlock(),
1978 I->BB->getBasicBlock());
1979 for (CaseItr J = CR.Range.first; J < I; ++J) {
1980 uint32_t JWeight = BPI->getEdgeWeight(SwitchBB->getBasicBlock(),
1981 J->BB->getBasicBlock());
1982 if (IWeight > JWeight)
1987 // Rearrange the case blocks so that the last one falls through if possible.
1988 Case &BackCase = *(CR.Range.second-1);
1990 NextBlock && Default != NextBlock && BackCase.BB != NextBlock) {
1991 // The last case block won't fall through into 'NextBlock' if we emit the
1992 // branches in this order. See if rearranging a case value would help.
1993 // We start at the bottom as it's the case with the least weight.
1994 for (Case *I = &*(CR.Range.second-2), *E = &*CR.Range.first-1; I != E; --I){
1995 if (I->BB == NextBlock) {
1996 std::swap(*I, BackCase);
2002 // Create a CaseBlock record representing a conditional branch to
2003 // the Case's target mbb if the value being switched on SV is equal
2005 MachineBasicBlock *CurBlock = CR.CaseBB;
2006 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
2007 MachineBasicBlock *FallThrough;
2009 FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock());
2010 CurMF->insert(BBI, FallThrough);
2012 // Put SV in a virtual register to make it available from the new blocks.
2013 ExportFromCurrentBlock(SV);
2015 // If the last case doesn't match, go to the default block.
2016 FallThrough = Default;
2019 const Value *RHS, *LHS, *MHS;
2021 if (I->High == I->Low) {
2022 // This is just small small case range :) containing exactly 1 case
2024 LHS = SV; RHS = I->High; MHS = NULL;
2026 CC = ISD::SETCC_INVALID;
2027 LHS = I->Low; MHS = SV; RHS = I->High;
2030 uint32_t ExtraWeight = I->ExtraWeight;
2031 CaseBlock CB(CC, LHS, RHS, MHS, /* truebb */ I->BB, /* falsebb */ FallThrough,
2033 /* trueweight */ ExtraWeight / 2, /* falseweight */ ExtraWeight / 2);
2035 // If emitting the first comparison, just call visitSwitchCase to emit the
2036 // code into the current block. Otherwise, push the CaseBlock onto the
2037 // vector to be later processed by SDISel, and insert the node's MBB
2038 // before the next MBB.
2039 if (CurBlock == SwitchBB)
2040 visitSwitchCase(CB, SwitchBB);
2042 SwitchCases.push_back(CB);
2044 CurBlock = FallThrough;
2050 static inline bool areJTsAllowed(const TargetLowering &TLI) {
2051 return !TLI.getTargetMachine().Options.DisableJumpTables &&
2052 (TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
2053 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other));
2056 static APInt ComputeRange(const APInt &First, const APInt &Last) {
2057 uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1;
2058 APInt LastExt = Last.zext(BitWidth), FirstExt = First.zext(BitWidth);
2059 return (LastExt - FirstExt + 1ULL);
2062 /// handleJTSwitchCase - Emit jumptable for current switch case range
2063 bool SelectionDAGBuilder::handleJTSwitchCase(CaseRec &CR,
2064 CaseRecVector &WorkList,
2066 MachineBasicBlock *Default,
2067 MachineBasicBlock *SwitchBB) {
2068 Case& FrontCase = *CR.Range.first;
2069 Case& BackCase = *(CR.Range.second-1);
2071 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
2072 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
2074 APInt TSize(First.getBitWidth(), 0);
2075 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I)
2078 if (!areJTsAllowed(TLI) || TSize.ult(4))
2081 APInt Range = ComputeRange(First, Last);
2082 // The density is TSize / Range. Require at least 40%.
2083 // It should not be possible for IntTSize to saturate for sane code, but make
2084 // sure we handle Range saturation correctly.
2085 uint64_t IntRange = Range.getLimitedValue(UINT64_MAX/10);
2086 uint64_t IntTSize = TSize.getLimitedValue(UINT64_MAX/10);
2087 if (IntTSize * 10 < IntRange * 4)
2090 DEBUG(dbgs() << "Lowering jump table\n"
2091 << "First entry: " << First << ". Last entry: " << Last << '\n'
2092 << "Range: " << Range << ". Size: " << TSize << ".\n\n");
2094 // Get the MachineFunction which holds the current MBB. This is used when
2095 // inserting any additional MBBs necessary to represent the switch.
2096 MachineFunction *CurMF = FuncInfo.MF;
2098 // Figure out which block is immediately after the current one.
2099 MachineFunction::iterator BBI = CR.CaseBB;
2102 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2104 // Create a new basic block to hold the code for loading the address
2105 // of the jump table, and jumping to it. Update successor information;
2106 // we will either branch to the default case for the switch, or the jump
2108 MachineBasicBlock *JumpTableBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2109 CurMF->insert(BBI, JumpTableBB);
2111 addSuccessorWithWeight(CR.CaseBB, Default);
2112 addSuccessorWithWeight(CR.CaseBB, JumpTableBB);
2114 // Build a vector of destination BBs, corresponding to each target
2115 // of the jump table. If the value of the jump table slot corresponds to
2116 // a case statement, push the case's BB onto the vector, otherwise, push
2118 std::vector<MachineBasicBlock*> DestBBs;
2120 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) {
2121 const APInt &Low = cast<ConstantInt>(I->Low)->getValue();
2122 const APInt &High = cast<ConstantInt>(I->High)->getValue();
2124 if (Low.ule(TEI) && TEI.ule(High)) {
2125 DestBBs.push_back(I->BB);
2129 DestBBs.push_back(Default);
2133 // Update successor info. Add one edge to each unique successor.
2134 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs());
2135 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(),
2136 E = DestBBs.end(); I != E; ++I) {
2137 if (!SuccsHandled[(*I)->getNumber()]) {
2138 SuccsHandled[(*I)->getNumber()] = true;
2139 addSuccessorWithWeight(JumpTableBB, *I);
2143 // Create a jump table index for this jump table.
2144 unsigned JTEncoding = TLI.getJumpTableEncoding();
2145 unsigned JTI = CurMF->getOrCreateJumpTableInfo(JTEncoding)
2146 ->createJumpTableIndex(DestBBs);
2148 // Set the jump table information so that we can codegen it as a second
2149 // MachineBasicBlock
2150 JumpTable JT(-1U, JTI, JumpTableBB, Default);
2151 JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == SwitchBB));
2152 if (CR.CaseBB == SwitchBB)
2153 visitJumpTableHeader(JT, JTH, SwitchBB);
2155 JTCases.push_back(JumpTableBlock(JTH, JT));
2159 /// handleBTSplitSwitchCase - emit comparison and split binary search tree into
2161 bool SelectionDAGBuilder::handleBTSplitSwitchCase(CaseRec& CR,
2162 CaseRecVector& WorkList,
2164 MachineBasicBlock *Default,
2165 MachineBasicBlock *SwitchBB) {
2166 // Get the MachineFunction which holds the current MBB. This is used when
2167 // inserting any additional MBBs necessary to represent the switch.
2168 MachineFunction *CurMF = FuncInfo.MF;
2170 // Figure out which block is immediately after the current one.
2171 MachineFunction::iterator BBI = CR.CaseBB;
2174 Case& FrontCase = *CR.Range.first;
2175 Case& BackCase = *(CR.Range.second-1);
2176 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2178 // Size is the number of Cases represented by this range.
2179 unsigned Size = CR.Range.second - CR.Range.first;
2181 const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
2182 const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
2184 CaseItr Pivot = CR.Range.first + Size/2;
2186 // Select optimal pivot, maximizing sum density of LHS and RHS. This will
2187 // (heuristically) allow us to emit JumpTable's later.
2188 APInt TSize(First.getBitWidth(), 0);
2189 for (CaseItr I = CR.Range.first, E = CR.Range.second;
2193 APInt LSize = FrontCase.size();
2194 APInt RSize = TSize-LSize;
2195 DEBUG(dbgs() << "Selecting best pivot: \n"
2196 << "First: " << First << ", Last: " << Last <<'\n'
2197 << "LSize: " << LSize << ", RSize: " << RSize << '\n');
2198 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second;
2200 const APInt &LEnd = cast<ConstantInt>(I->High)->getValue();
2201 const APInt &RBegin = cast<ConstantInt>(J->Low)->getValue();
2202 APInt Range = ComputeRange(LEnd, RBegin);
2203 assert((Range - 2ULL).isNonNegative() &&
2204 "Invalid case distance");
2205 // Use volatile double here to avoid excess precision issues on some hosts,
2206 // e.g. that use 80-bit X87 registers.
2207 volatile double LDensity =
2208 (double)LSize.roundToDouble() /
2209 (LEnd - First + 1ULL).roundToDouble();
2210 volatile double RDensity =
2211 (double)RSize.roundToDouble() /
2212 (Last - RBegin + 1ULL).roundToDouble();
2213 double Metric = Range.logBase2()*(LDensity+RDensity);
2214 // Should always split in some non-trivial place
2215 DEBUG(dbgs() <<"=>Step\n"
2216 << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n'
2217 << "LDensity: " << LDensity
2218 << ", RDensity: " << RDensity << '\n'
2219 << "Metric: " << Metric << '\n');
2220 if (FMetric < Metric) {
2223 DEBUG(dbgs() << "Current metric set to: " << FMetric << '\n');
2229 if (areJTsAllowed(TLI)) {
2230 // If our case is dense we *really* should handle it earlier!
2231 assert((FMetric > 0) && "Should handle dense range earlier!");
2233 Pivot = CR.Range.first + Size/2;
2236 CaseRange LHSR(CR.Range.first, Pivot);
2237 CaseRange RHSR(Pivot, CR.Range.second);
2238 const Constant *C = Pivot->Low;
2239 MachineBasicBlock *FalseBB = 0, *TrueBB = 0;
2241 // We know that we branch to the LHS if the Value being switched on is
2242 // less than the Pivot value, C. We use this to optimize our binary
2243 // tree a bit, by recognizing that if SV is greater than or equal to the
2244 // LHS's Case Value, and that Case Value is exactly one less than the
2245 // Pivot's Value, then we can branch directly to the LHS's Target,
2246 // rather than creating a leaf node for it.
2247 if ((LHSR.second - LHSR.first) == 1 &&
2248 LHSR.first->High == CR.GE &&
2249 cast<ConstantInt>(C)->getValue() ==
2250 (cast<ConstantInt>(CR.GE)->getValue() + 1LL)) {
2251 TrueBB = LHSR.first->BB;
2253 TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2254 CurMF->insert(BBI, TrueBB);
2255 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR));
2257 // Put SV in a virtual register to make it available from the new blocks.
2258 ExportFromCurrentBlock(SV);
2261 // Similar to the optimization above, if the Value being switched on is
2262 // known to be less than the Constant CR.LT, and the current Case Value
2263 // is CR.LT - 1, then we can branch directly to the target block for
2264 // the current Case Value, rather than emitting a RHS leaf node for it.
2265 if ((RHSR.second - RHSR.first) == 1 && CR.LT &&
2266 cast<ConstantInt>(RHSR.first->Low)->getValue() ==
2267 (cast<ConstantInt>(CR.LT)->getValue() - 1LL)) {
2268 FalseBB = RHSR.first->BB;
2270 FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2271 CurMF->insert(BBI, FalseBB);
2272 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR));
2274 // Put SV in a virtual register to make it available from the new blocks.
2275 ExportFromCurrentBlock(SV);
2278 // Create a CaseBlock record representing a conditional branch to
2279 // the LHS node if the value being switched on SV is less than C.
2280 // Otherwise, branch to LHS.
2281 CaseBlock CB(ISD::SETULT, SV, C, NULL, TrueBB, FalseBB, CR.CaseBB);
2283 if (CR.CaseBB == SwitchBB)
2284 visitSwitchCase(CB, SwitchBB);
2286 SwitchCases.push_back(CB);
2291 /// handleBitTestsSwitchCase - if current case range has few destination and
2292 /// range span less, than machine word bitwidth, encode case range into series
2293 /// of masks and emit bit tests with these masks.
2294 bool SelectionDAGBuilder::handleBitTestsSwitchCase(CaseRec& CR,
2295 CaseRecVector& WorkList,
2297 MachineBasicBlock* Default,
2298 MachineBasicBlock *SwitchBB){
2299 EVT PTy = TLI.getPointerTy();
2300 unsigned IntPtrBits = PTy.getSizeInBits();
2302 Case& FrontCase = *CR.Range.first;
2303 Case& BackCase = *(CR.Range.second-1);
2305 // Get the MachineFunction which holds the current MBB. This is used when
2306 // inserting any additional MBBs necessary to represent the switch.
2307 MachineFunction *CurMF = FuncInfo.MF;
2309 // If target does not have legal shift left, do not emit bit tests at all.
2310 if (!TLI.isOperationLegal(ISD::SHL, TLI.getPointerTy()))
2314 for (CaseItr I = CR.Range.first, E = CR.Range.second;
2316 // Single case counts one, case range - two.
2317 numCmps += (I->Low == I->High ? 1 : 2);
2320 // Count unique destinations
2321 SmallSet<MachineBasicBlock*, 4> Dests;
2322 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2323 Dests.insert(I->BB);
2324 if (Dests.size() > 3)
2325 // Don't bother the code below, if there are too much unique destinations
2328 DEBUG(dbgs() << "Total number of unique destinations: "
2329 << Dests.size() << '\n'
2330 << "Total number of comparisons: " << numCmps << '\n');
2332 // Compute span of values.
2333 const APInt& minValue = cast<ConstantInt>(FrontCase.Low)->getValue();
2334 const APInt& maxValue = cast<ConstantInt>(BackCase.High)->getValue();
2335 APInt cmpRange = maxValue - minValue;
2337 DEBUG(dbgs() << "Compare range: " << cmpRange << '\n'
2338 << "Low bound: " << minValue << '\n'
2339 << "High bound: " << maxValue << '\n');
2341 if (cmpRange.uge(IntPtrBits) ||
2342 (!(Dests.size() == 1 && numCmps >= 3) &&
2343 !(Dests.size() == 2 && numCmps >= 5) &&
2344 !(Dests.size() >= 3 && numCmps >= 6)))
2347 DEBUG(dbgs() << "Emitting bit tests\n");
2348 APInt lowBound = APInt::getNullValue(cmpRange.getBitWidth());
2350 // Optimize the case where all the case values fit in a
2351 // word without having to subtract minValue. In this case,
2352 // we can optimize away the subtraction.
2353 if (maxValue.ult(IntPtrBits)) {
2354 cmpRange = maxValue;
2356 lowBound = minValue;
2359 CaseBitsVector CasesBits;
2360 unsigned i, count = 0;
2362 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
2363 MachineBasicBlock* Dest = I->BB;
2364 for (i = 0; i < count; ++i)
2365 if (Dest == CasesBits[i].BB)
2369 assert((count < 3) && "Too much destinations to test!");
2370 CasesBits.push_back(CaseBits(0, Dest, 0));
2374 const APInt& lowValue = cast<ConstantInt>(I->Low)->getValue();
2375 const APInt& highValue = cast<ConstantInt>(I->High)->getValue();
2377 uint64_t lo = (lowValue - lowBound).getZExtValue();
2378 uint64_t hi = (highValue - lowBound).getZExtValue();
2380 for (uint64_t j = lo; j <= hi; j++) {
2381 CasesBits[i].Mask |= 1ULL << j;
2382 CasesBits[i].Bits++;
2386 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp());
2390 // Figure out which block is immediately after the current one.
2391 MachineFunction::iterator BBI = CR.CaseBB;
2394 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2396 DEBUG(dbgs() << "Cases:\n");
2397 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) {
2398 DEBUG(dbgs() << "Mask: " << CasesBits[i].Mask
2399 << ", Bits: " << CasesBits[i].Bits
2400 << ", BB: " << CasesBits[i].BB << '\n');
2402 MachineBasicBlock *CaseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2403 CurMF->insert(BBI, CaseBB);
2404 BTC.push_back(BitTestCase(CasesBits[i].Mask,
2408 // Put SV in a virtual register to make it available from the new blocks.
2409 ExportFromCurrentBlock(SV);
2412 BitTestBlock BTB(lowBound, cmpRange, SV,
2413 -1U, MVT::Other, (CR.CaseBB == SwitchBB),
2414 CR.CaseBB, Default, BTC);
2416 if (CR.CaseBB == SwitchBB)
2417 visitBitTestHeader(BTB, SwitchBB);
2419 BitTestCases.push_back(BTB);
2424 /// Clusterify - Transform simple list of Cases into list of CaseRange's
2425 size_t SelectionDAGBuilder::Clusterify(CaseVector& Cases,
2426 const SwitchInst& SI) {
2428 /// Use a shorter form of declaration, and also
2429 /// show the we want to use CRSBuilder as Clusterifier.
2430 typedef IntegersSubsetMapping<MachineBasicBlock> Clusterifier;
2432 Clusterifier TheClusterifier;
2434 // Start with "simple" cases
2435 for (SwitchInst::ConstCaseIt i = SI.case_begin(), e = SI.case_end();
2437 const BasicBlock *SuccBB = i.getCaseSuccessor();
2438 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SuccBB];
2440 TheClusterifier.add(i.getCaseValueEx(), SMBB);
2443 TheClusterifier.optimize();
2445 BranchProbabilityInfo *BPI = FuncInfo.BPI;
2447 for (Clusterifier::RangeIterator i = TheClusterifier.begin(),
2448 e = TheClusterifier.end(); i != e; ++i, ++numCmps) {
2449 Clusterifier::Cluster &C = *i;
2452 W = BPI->getEdgeWeight(SI.getParent(), C.second->getBasicBlock());
2455 W *= C.first.Weight;
2456 BPI->setEdgeWeight(SI.getParent(), C.second->getBasicBlock(), W);
2459 // FIXME: Currently work with ConstantInt based numbers.
2460 // Changing it to APInt based is a pretty heavy for this commit.
2461 Cases.push_back(Case(C.first.getLow().toConstantInt(),
2462 C.first.getHigh().toConstantInt(), C.second, W));
2464 if (C.first.getLow() != C.first.getHigh())
2465 // A range counts double, since it requires two compares.
2472 void SelectionDAGBuilder::UpdateSplitBlock(MachineBasicBlock *First,
2473 MachineBasicBlock *Last) {
2475 for (unsigned i = 0, e = JTCases.size(); i != e; ++i)
2476 if (JTCases[i].first.HeaderBB == First)
2477 JTCases[i].first.HeaderBB = Last;
2479 // Update BitTestCases.
2480 for (unsigned i = 0, e = BitTestCases.size(); i != e; ++i)
2481 if (BitTestCases[i].Parent == First)
2482 BitTestCases[i].Parent = Last;
2485 void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) {
2486 MachineBasicBlock *SwitchMBB = FuncInfo.MBB;
2488 // Figure out which block is immediately after the current one.
2489 MachineBasicBlock *NextBlock = 0;
2490 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()];
2492 // If there is only the default destination, branch to it if it is not the
2493 // next basic block. Otherwise, just fall through.
2494 if (!SI.getNumCases()) {
2495 // Update machine-CFG edges.
2497 // If this is not a fall-through branch, emit the branch.
2498 SwitchMBB->addSuccessor(Default);
2499 if (Default != NextBlock)
2500 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
2501 MVT::Other, getControlRoot(),
2502 DAG.getBasicBlock(Default)));
2507 // If there are any non-default case statements, create a vector of Cases
2508 // representing each one, and sort the vector so that we can efficiently
2509 // create a binary search tree from them.
2511 size_t numCmps = Clusterify(Cases, SI);
2512 DEBUG(dbgs() << "Clusterify finished. Total clusters: " << Cases.size()
2513 << ". Total compares: " << numCmps << '\n');
2516 // Get the Value to be switched on and default basic blocks, which will be
2517 // inserted into CaseBlock records, representing basic blocks in the binary
2519 const Value *SV = SI.getCondition();
2521 // Push the initial CaseRec onto the worklist
2522 CaseRecVector WorkList;
2523 WorkList.push_back(CaseRec(SwitchMBB,0,0,
2524 CaseRange(Cases.begin(),Cases.end())));
2526 while (!WorkList.empty()) {
2527 // Grab a record representing a case range to process off the worklist
2528 CaseRec CR = WorkList.back();
2529 WorkList.pop_back();
2531 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2534 // If the range has few cases (two or less) emit a series of specific
2536 if (handleSmallSwitchRange(CR, WorkList, SV, Default, SwitchMBB))
2539 // If the switch has more than 5 blocks, and at least 40% dense, and the
2540 // target supports indirect branches, then emit a jump table rather than
2541 // lowering the switch to a binary tree of conditional branches.
2542 if (handleJTSwitchCase(CR, WorkList, SV, Default, SwitchMBB))
2545 // Emit binary tree. We need to pick a pivot, and push left and right ranges
2546 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call.
2547 handleBTSplitSwitchCase(CR, WorkList, SV, Default, SwitchMBB);
2551 void SelectionDAGBuilder::visitIndirectBr(const IndirectBrInst &I) {
2552 MachineBasicBlock *IndirectBrMBB = FuncInfo.MBB;
2554 // Update machine-CFG edges with unique successors.
2555 SmallVector<BasicBlock*, 32> succs;
2556 succs.reserve(I.getNumSuccessors());
2557 for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i)
2558 succs.push_back(I.getSuccessor(i));
2559 array_pod_sort(succs.begin(), succs.end());
2560 succs.erase(std::unique(succs.begin(), succs.end()), succs.end());
2561 for (unsigned i = 0, e = succs.size(); i != e; ++i) {
2562 MachineBasicBlock *Succ = FuncInfo.MBBMap[succs[i]];
2563 addSuccessorWithWeight(IndirectBrMBB, Succ);
2566 DAG.setRoot(DAG.getNode(ISD::BRIND, getCurDebugLoc(),
2567 MVT::Other, getControlRoot(),
2568 getValue(I.getAddress())));
2571 void SelectionDAGBuilder::visitFSub(const User &I) {
2572 // -0.0 - X --> fneg
2573 Type *Ty = I.getType();
2574 if (isa<Constant>(I.getOperand(0)) &&
2575 I.getOperand(0) == ConstantFP::getZeroValueForNegation(Ty)) {
2576 SDValue Op2 = getValue(I.getOperand(1));
2577 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(),
2578 Op2.getValueType(), Op2));
2582 visitBinary(I, ISD::FSUB);
2585 void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) {
2586 SDValue Op1 = getValue(I.getOperand(0));
2587 SDValue Op2 = getValue(I.getOperand(1));
2588 setValue(&I, DAG.getNode(OpCode, getCurDebugLoc(),
2589 Op1.getValueType(), Op1, Op2));
2592 void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) {
2593 SDValue Op1 = getValue(I.getOperand(0));
2594 SDValue Op2 = getValue(I.getOperand(1));
2596 MVT ShiftTy = TLI.getShiftAmountTy(Op2.getValueType());
2598 // Coerce the shift amount to the right type if we can.
2599 if (!I.getType()->isVectorTy() && Op2.getValueType() != ShiftTy) {
2600 unsigned ShiftSize = ShiftTy.getSizeInBits();
2601 unsigned Op2Size = Op2.getValueType().getSizeInBits();
2602 DebugLoc DL = getCurDebugLoc();
2604 // If the operand is smaller than the shift count type, promote it.
2605 if (ShiftSize > Op2Size)
2606 Op2 = DAG.getNode(ISD::ZERO_EXTEND, DL, ShiftTy, Op2);
2608 // If the operand is larger than the shift count type but the shift
2609 // count type has enough bits to represent any shift value, truncate
2610 // it now. This is a common case and it exposes the truncate to
2611 // optimization early.
2612 else if (ShiftSize >= Log2_32_Ceil(Op2.getValueType().getSizeInBits()))
2613 Op2 = DAG.getNode(ISD::TRUNCATE, DL, ShiftTy, Op2);
2614 // Otherwise we'll need to temporarily settle for some other convenient
2615 // type. Type legalization will make adjustments once the shiftee is split.
2617 Op2 = DAG.getZExtOrTrunc(Op2, DL, MVT::i32);
2620 setValue(&I, DAG.getNode(Opcode, getCurDebugLoc(),
2621 Op1.getValueType(), Op1, Op2));
2624 void SelectionDAGBuilder::visitSDiv(const User &I) {
2625 SDValue Op1 = getValue(I.getOperand(0));
2626 SDValue Op2 = getValue(I.getOperand(1));
2628 // Turn exact SDivs into multiplications.
2629 // FIXME: This should be in DAGCombiner, but it doesn't have access to the
2631 if (isa<BinaryOperator>(&I) && cast<BinaryOperator>(&I)->isExact() &&
2632 !isa<ConstantSDNode>(Op1) &&
2633 isa<ConstantSDNode>(Op2) && !cast<ConstantSDNode>(Op2)->isNullValue())
2634 setValue(&I, TLI.BuildExactSDIV(Op1, Op2, getCurDebugLoc(), DAG));
2636 setValue(&I, DAG.getNode(ISD::SDIV, getCurDebugLoc(), Op1.getValueType(),
2640 void SelectionDAGBuilder::visitICmp(const User &I) {
2641 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2642 if (const ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2643 predicate = IC->getPredicate();
2644 else if (const ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2645 predicate = ICmpInst::Predicate(IC->getPredicate());
2646 SDValue Op1 = getValue(I.getOperand(0));
2647 SDValue Op2 = getValue(I.getOperand(1));
2648 ISD::CondCode Opcode = getICmpCondCode(predicate);
2650 EVT DestVT = TLI.getValueType(I.getType());
2651 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Opcode));
2654 void SelectionDAGBuilder::visitFCmp(const User &I) {
2655 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2656 if (const FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2657 predicate = FC->getPredicate();
2658 else if (const ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2659 predicate = FCmpInst::Predicate(FC->getPredicate());
2660 SDValue Op1 = getValue(I.getOperand(0));
2661 SDValue Op2 = getValue(I.getOperand(1));
2662 ISD::CondCode Condition = getFCmpCondCode(predicate);
2663 if (TM.Options.NoNaNsFPMath)
2664 Condition = getFCmpCodeWithoutNaN(Condition);
2665 EVT DestVT = TLI.getValueType(I.getType());
2666 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Condition));
2669 void SelectionDAGBuilder::visitSelect(const User &I) {
2670 SmallVector<EVT, 4> ValueVTs;
2671 ComputeValueVTs(TLI, I.getType(), ValueVTs);
2672 unsigned NumValues = ValueVTs.size();
2673 if (NumValues == 0) return;
2675 SmallVector<SDValue, 4> Values(NumValues);
2676 SDValue Cond = getValue(I.getOperand(0));
2677 SDValue TrueVal = getValue(I.getOperand(1));
2678 SDValue FalseVal = getValue(I.getOperand(2));
2679 ISD::NodeType OpCode = Cond.getValueType().isVector() ?
2680 ISD::VSELECT : ISD::SELECT;
2682 for (unsigned i = 0; i != NumValues; ++i)
2683 Values[i] = DAG.getNode(OpCode, getCurDebugLoc(),
2684 TrueVal.getNode()->getValueType(TrueVal.getResNo()+i),
2686 SDValue(TrueVal.getNode(),
2687 TrueVal.getResNo() + i),
2688 SDValue(FalseVal.getNode(),
2689 FalseVal.getResNo() + i));
2691 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2692 DAG.getVTList(&ValueVTs[0], NumValues),
2693 &Values[0], NumValues));
2696 void SelectionDAGBuilder::visitTrunc(const User &I) {
2697 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2698 SDValue N = getValue(I.getOperand(0));
2699 EVT DestVT = TLI.getValueType(I.getType());
2700 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), DestVT, N));
2703 void SelectionDAGBuilder::visitZExt(const User &I) {
2704 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2705 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2706 SDValue N = getValue(I.getOperand(0));
2707 EVT DestVT = TLI.getValueType(I.getType());
2708 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(), DestVT, N));
2711 void SelectionDAGBuilder::visitSExt(const User &I) {
2712 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2713 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2714 SDValue N = getValue(I.getOperand(0));
2715 EVT DestVT = TLI.getValueType(I.getType());
2716 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurDebugLoc(), DestVT, N));
2719 void SelectionDAGBuilder::visitFPTrunc(const User &I) {
2720 // FPTrunc is never a no-op cast, no need to check
2721 SDValue N = getValue(I.getOperand(0));
2722 EVT DestVT = TLI.getValueType(I.getType());
2723 setValue(&I, DAG.getNode(ISD::FP_ROUND, getCurDebugLoc(),
2725 DAG.getTargetConstant(0, TLI.getPointerTy())));
2728 void SelectionDAGBuilder::visitFPExt(const User &I){
2729 // FPExt is never a no-op cast, no need to check
2730 SDValue N = getValue(I.getOperand(0));
2731 EVT DestVT = TLI.getValueType(I.getType());
2732 setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurDebugLoc(), DestVT, N));
2735 void SelectionDAGBuilder::visitFPToUI(const User &I) {
2736 // FPToUI is never a no-op cast, no need to check
2737 SDValue N = getValue(I.getOperand(0));
2738 EVT DestVT = TLI.getValueType(I.getType());
2739 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurDebugLoc(), DestVT, N));
2742 void SelectionDAGBuilder::visitFPToSI(const User &I) {
2743 // FPToSI is never a no-op cast, no need to check
2744 SDValue N = getValue(I.getOperand(0));
2745 EVT DestVT = TLI.getValueType(I.getType());
2746 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurDebugLoc(), DestVT, N));
2749 void SelectionDAGBuilder::visitUIToFP(const User &I) {
2750 // UIToFP is never a no-op cast, no need to check
2751 SDValue N = getValue(I.getOperand(0));
2752 EVT DestVT = TLI.getValueType(I.getType());
2753 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurDebugLoc(), DestVT, N));
2756 void SelectionDAGBuilder::visitSIToFP(const User &I){
2757 // SIToFP is never a no-op cast, no need to check
2758 SDValue N = getValue(I.getOperand(0));
2759 EVT DestVT = TLI.getValueType(I.getType());
2760 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurDebugLoc(), DestVT, N));
2763 void SelectionDAGBuilder::visitPtrToInt(const User &I) {
2764 // What to do depends on the size of the integer and the size of the pointer.
2765 // We can either truncate, zero extend, or no-op, accordingly.
2766 SDValue N = getValue(I.getOperand(0));
2767 EVT DestVT = TLI.getValueType(I.getType());
2768 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT));
2771 void SelectionDAGBuilder::visitIntToPtr(const User &I) {
2772 // What to do depends on the size of the integer and the size of the pointer.
2773 // We can either truncate, zero extend, or no-op, accordingly.
2774 SDValue N = getValue(I.getOperand(0));
2775 EVT DestVT = TLI.getValueType(I.getType());
2776 setValue(&I, DAG.getZExtOrTrunc(N, getCurDebugLoc(), DestVT));
2779 void SelectionDAGBuilder::visitBitCast(const User &I) {
2780 SDValue N = getValue(I.getOperand(0));
2781 EVT DestVT = TLI.getValueType(I.getType());
2783 // BitCast assures us that source and destination are the same size so this is
2784 // either a BITCAST or a no-op.
2785 if (DestVT != N.getValueType())
2786 setValue(&I, DAG.getNode(ISD::BITCAST, getCurDebugLoc(),
2787 DestVT, N)); // convert types.
2789 setValue(&I, N); // noop cast.
2792 void SelectionDAGBuilder::visitInsertElement(const User &I) {
2793 SDValue InVec = getValue(I.getOperand(0));
2794 SDValue InVal = getValue(I.getOperand(1));
2795 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2797 getValue(I.getOperand(2)));
2798 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurDebugLoc(),
2799 TLI.getValueType(I.getType()),
2800 InVec, InVal, InIdx));
2803 void SelectionDAGBuilder::visitExtractElement(const User &I) {
2804 SDValue InVec = getValue(I.getOperand(0));
2805 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2807 getValue(I.getOperand(1)));
2808 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2809 TLI.getValueType(I.getType()), InVec, InIdx));
2812 // Utility for visitShuffleVector - Return true if every element in Mask,
2813 // beginning from position Pos and ending in Pos+Size, falls within the
2814 // specified sequential range [L, L+Pos). or is undef.
2815 static bool isSequentialInRange(const SmallVectorImpl<int> &Mask,
2816 unsigned Pos, unsigned Size, int Low) {
2817 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
2818 if (Mask[i] >= 0 && Mask[i] != Low)
2823 void SelectionDAGBuilder::visitShuffleVector(const User &I) {
2824 SDValue Src1 = getValue(I.getOperand(0));
2825 SDValue Src2 = getValue(I.getOperand(1));
2827 SmallVector<int, 8> Mask;
2828 ShuffleVectorInst::getShuffleMask(cast<Constant>(I.getOperand(2)), Mask);
2829 unsigned MaskNumElts = Mask.size();
2831 EVT VT = TLI.getValueType(I.getType());
2832 EVT SrcVT = Src1.getValueType();
2833 unsigned SrcNumElts = SrcVT.getVectorNumElements();
2835 if (SrcNumElts == MaskNumElts) {
2836 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2841 // Normalize the shuffle vector since mask and vector length don't match.
2842 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) {
2843 // Mask is longer than the source vectors and is a multiple of the source
2844 // vectors. We can use concatenate vector to make the mask and vectors
2846 if (SrcNumElts*2 == MaskNumElts) {
2847 // First check for Src1 in low and Src2 in high
2848 if (isSequentialInRange(Mask, 0, SrcNumElts, 0) &&
2849 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, SrcNumElts)) {
2850 // The shuffle is concatenating two vectors together.
2851 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurDebugLoc(),
2855 // Then check for Src2 in low and Src1 in high
2856 if (isSequentialInRange(Mask, 0, SrcNumElts, SrcNumElts) &&
2857 isSequentialInRange(Mask, SrcNumElts, SrcNumElts, 0)) {
2858 // The shuffle is concatenating two vectors together.
2859 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurDebugLoc(),
2865 // Pad both vectors with undefs to make them the same length as the mask.
2866 unsigned NumConcat = MaskNumElts / SrcNumElts;
2867 bool Src1U = Src1.getOpcode() == ISD::UNDEF;
2868 bool Src2U = Src2.getOpcode() == ISD::UNDEF;
2869 SDValue UndefVal = DAG.getUNDEF(SrcVT);
2871 SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal);
2872 SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal);
2876 Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2877 getCurDebugLoc(), VT,
2878 &MOps1[0], NumConcat);
2879 Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2880 getCurDebugLoc(), VT,
2881 &MOps2[0], NumConcat);
2883 // Readjust mask for new input vector length.
2884 SmallVector<int, 8> MappedOps;
2885 for (unsigned i = 0; i != MaskNumElts; ++i) {
2887 if (Idx >= (int)SrcNumElts)
2888 Idx -= SrcNumElts - MaskNumElts;
2889 MappedOps.push_back(Idx);
2892 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2897 if (SrcNumElts > MaskNumElts) {
2898 // Analyze the access pattern of the vector to see if we can extract
2899 // two subvectors and do the shuffle. The analysis is done by calculating
2900 // the range of elements the mask access on both vectors.
2901 int MinRange[2] = { static_cast<int>(SrcNumElts),
2902 static_cast<int>(SrcNumElts)};
2903 int MaxRange[2] = {-1, -1};
2905 for (unsigned i = 0; i != MaskNumElts; ++i) {
2911 if (Idx >= (int)SrcNumElts) {
2915 if (Idx > MaxRange[Input])
2916 MaxRange[Input] = Idx;
2917 if (Idx < MinRange[Input])
2918 MinRange[Input] = Idx;
2921 // Check if the access is smaller than the vector size and can we find
2922 // a reasonable extract index.
2923 int RangeUse[2] = { -1, -1 }; // 0 = Unused, 1 = Extract, -1 = Can not
2925 int StartIdx[2]; // StartIdx to extract from
2926 for (unsigned Input = 0; Input < 2; ++Input) {
2927 if (MinRange[Input] >= (int)SrcNumElts && MaxRange[Input] < 0) {
2928 RangeUse[Input] = 0; // Unused
2929 StartIdx[Input] = 0;
2933 // Find a good start index that is a multiple of the mask length. Then
2934 // see if the rest of the elements are in range.
2935 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts;
2936 if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts &&
2937 StartIdx[Input] + MaskNumElts <= SrcNumElts)
2938 RangeUse[Input] = 1; // Extract from a multiple of the mask length.
2941 if (RangeUse[0] == 0 && RangeUse[1] == 0) {
2942 setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used.
2945 if (RangeUse[0] >= 0 && RangeUse[1] >= 0) {
2946 // Extract appropriate subvector and generate a vector shuffle
2947 for (unsigned Input = 0; Input < 2; ++Input) {
2948 SDValue &Src = Input == 0 ? Src1 : Src2;
2949 if (RangeUse[Input] == 0)
2950 Src = DAG.getUNDEF(VT);
2952 Src = DAG.getNode(ISD::EXTRACT_SUBVECTOR, getCurDebugLoc(), VT,
2953 Src, DAG.getIntPtrConstant(StartIdx[Input]));
2956 // Calculate new mask.
2957 SmallVector<int, 8> MappedOps;
2958 for (unsigned i = 0; i != MaskNumElts; ++i) {
2961 if (Idx < (int)SrcNumElts)
2964 Idx -= SrcNumElts + StartIdx[1] - MaskNumElts;
2966 MappedOps.push_back(Idx);
2969 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2975 // We can't use either concat vectors or extract subvectors so fall back to
2976 // replacing the shuffle with extract and build vector.
2977 // to insert and build vector.
2978 EVT EltVT = VT.getVectorElementType();
2979 EVT PtrVT = TLI.getPointerTy();
2980 SmallVector<SDValue,8> Ops;
2981 for (unsigned i = 0; i != MaskNumElts; ++i) {
2986 Res = DAG.getUNDEF(EltVT);
2988 SDValue &Src = Idx < (int)SrcNumElts ? Src1 : Src2;
2989 if (Idx >= (int)SrcNumElts) Idx -= SrcNumElts;
2991 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2992 EltVT, Src, DAG.getConstant(Idx, PtrVT));
2998 setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
2999 VT, &Ops[0], Ops.size()));
3002 void SelectionDAGBuilder::visitInsertValue(const InsertValueInst &I) {
3003 const Value *Op0 = I.getOperand(0);
3004 const Value *Op1 = I.getOperand(1);
3005 Type *AggTy = I.getType();
3006 Type *ValTy = Op1->getType();
3007 bool IntoUndef = isa<UndefValue>(Op0);
3008 bool FromUndef = isa<UndefValue>(Op1);
3010 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
3012 SmallVector<EVT, 4> AggValueVTs;
3013 ComputeValueVTs(TLI, AggTy, AggValueVTs);
3014 SmallVector<EVT, 4> ValValueVTs;
3015 ComputeValueVTs(TLI, ValTy, ValValueVTs);
3017 unsigned NumAggValues = AggValueVTs.size();
3018 unsigned NumValValues = ValValueVTs.size();
3019 SmallVector<SDValue, 4> Values(NumAggValues);
3021 SDValue Agg = getValue(Op0);
3023 // Copy the beginning value(s) from the original aggregate.
3024 for (; i != LinearIndex; ++i)
3025 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
3026 SDValue(Agg.getNode(), Agg.getResNo() + i);
3027 // Copy values from the inserted value(s).
3029 SDValue Val = getValue(Op1);
3030 for (; i != LinearIndex + NumValValues; ++i)
3031 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) :
3032 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex);
3034 // Copy remaining value(s) from the original aggregate.
3035 for (; i != NumAggValues; ++i)
3036 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
3037 SDValue(Agg.getNode(), Agg.getResNo() + i);
3039 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
3040 DAG.getVTList(&AggValueVTs[0], NumAggValues),
3041 &Values[0], NumAggValues));
3044 void SelectionDAGBuilder::visitExtractValue(const ExtractValueInst &I) {
3045 const Value *Op0 = I.getOperand(0);
3046 Type *AggTy = Op0->getType();
3047 Type *ValTy = I.getType();
3048 bool OutOfUndef = isa<UndefValue>(Op0);
3050 unsigned LinearIndex = ComputeLinearIndex(AggTy, I.getIndices());
3052 SmallVector<EVT, 4> ValValueVTs;
3053 ComputeValueVTs(TLI, ValTy, ValValueVTs);
3055 unsigned NumValValues = ValValueVTs.size();
3057 // Ignore a extractvalue that produces an empty object
3058 if (!NumValValues) {
3059 setValue(&I, DAG.getUNDEF(MVT(MVT::Other)));
3063 SmallVector<SDValue, 4> Values(NumValValues);
3065 SDValue Agg = getValue(Op0);
3066 // Copy out the selected value(s).
3067 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i)
3068 Values[i - LinearIndex] =
3070 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) :
3071 SDValue(Agg.getNode(), Agg.getResNo() + i);
3073 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
3074 DAG.getVTList(&ValValueVTs[0], NumValValues),
3075 &Values[0], NumValValues));
3078 void SelectionDAGBuilder::visitGetElementPtr(const User &I) {
3079 SDValue N = getValue(I.getOperand(0));
3080 // Note that the pointer operand may be a vector of pointers. Take the scalar
3081 // element which holds a pointer.
3082 Type *Ty = I.getOperand(0)->getType()->getScalarType();
3084 for (GetElementPtrInst::const_op_iterator OI = I.op_begin()+1, E = I.op_end();
3086 const Value *Idx = *OI;
3087 if (StructType *StTy = dyn_cast<StructType>(Ty)) {
3088 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
3091 uint64_t Offset = TD->getStructLayout(StTy)->getElementOffset(Field);
3092 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
3093 DAG.getIntPtrConstant(Offset));
3096 Ty = StTy->getElementType(Field);
3098 Ty = cast<SequentialType>(Ty)->getElementType();
3100 // If this is a constant subscript, handle it quickly.
3101 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
3102 if (CI->isZero()) continue;
3104 TD->getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
3106 EVT PTy = TLI.getPointerTy();
3107 unsigned PtrBits = PTy.getSizeInBits();
3109 OffsVal = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
3111 DAG.getConstant(Offs, MVT::i64));
3113 OffsVal = DAG.getIntPtrConstant(Offs);
3115 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
3120 // N = N + Idx * ElementSize;
3121 APInt ElementSize = APInt(TLI.getPointerTy().getSizeInBits(),
3122 TD->getTypeAllocSize(Ty));
3123 SDValue IdxN = getValue(Idx);
3125 // If the index is smaller or larger than intptr_t, truncate or extend
3127 IdxN = DAG.getSExtOrTrunc(IdxN, getCurDebugLoc(), N.getValueType());
3129 // If this is a multiply by a power of two, turn it into a shl
3130 // immediately. This is a very common case.
3131 if (ElementSize != 1) {
3132 if (ElementSize.isPowerOf2()) {
3133 unsigned Amt = ElementSize.logBase2();
3134 IdxN = DAG.getNode(ISD::SHL, getCurDebugLoc(),
3135 N.getValueType(), IdxN,
3136 DAG.getConstant(Amt, IdxN.getValueType()));
3138 SDValue Scale = DAG.getConstant(ElementSize, TLI.getPointerTy());
3139 IdxN = DAG.getNode(ISD::MUL, getCurDebugLoc(),
3140 N.getValueType(), IdxN, Scale);
3144 N = DAG.getNode(ISD::ADD, getCurDebugLoc(),
3145 N.getValueType(), N, IdxN);
3152 void SelectionDAGBuilder::visitAlloca(const AllocaInst &I) {
3153 // If this is a fixed sized alloca in the entry block of the function,
3154 // allocate it statically on the stack.
3155 if (FuncInfo.StaticAllocaMap.count(&I))
3156 return; // getValue will auto-populate this.
3158 Type *Ty = I.getAllocatedType();
3159 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
3161 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty),
3164 SDValue AllocSize = getValue(I.getArraySize());
3166 EVT IntPtr = TLI.getPointerTy();
3167 if (AllocSize.getValueType() != IntPtr)
3168 AllocSize = DAG.getZExtOrTrunc(AllocSize, getCurDebugLoc(), IntPtr);
3170 AllocSize = DAG.getNode(ISD::MUL, getCurDebugLoc(), IntPtr,
3172 DAG.getConstant(TySize, IntPtr));
3174 // Handle alignment. If the requested alignment is less than or equal to
3175 // the stack alignment, ignore it. If the size is greater than or equal to
3176 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
3177 unsigned StackAlign = TM.getFrameLowering()->getStackAlignment();
3178 if (Align <= StackAlign)
3181 // Round the size of the allocation up to the stack alignment size
3182 // by add SA-1 to the size.
3183 AllocSize = DAG.getNode(ISD::ADD, getCurDebugLoc(),
3184 AllocSize.getValueType(), AllocSize,
3185 DAG.getIntPtrConstant(StackAlign-1));
3187 // Mask out the low bits for alignment purposes.
3188 AllocSize = DAG.getNode(ISD::AND, getCurDebugLoc(),
3189 AllocSize.getValueType(), AllocSize,
3190 DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1)));
3192 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) };
3193 SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other);
3194 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, getCurDebugLoc(),
3197 DAG.setRoot(DSA.getValue(1));
3199 // Inform the Frame Information that we have just allocated a variable-sized
3201 FuncInfo.MF->getFrameInfo()->CreateVariableSizedObject(Align ? Align : 1);
3204 void SelectionDAGBuilder::visitLoad(const LoadInst &I) {
3206 return visitAtomicLoad(I);
3208 const Value *SV = I.getOperand(0);
3209 SDValue Ptr = getValue(SV);
3211 Type *Ty = I.getType();
3213 bool isVolatile = I.isVolatile();
3214 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
3215 bool isInvariant = I.getMetadata("invariant.load") != 0;
3216 unsigned Alignment = I.getAlignment();
3217 const MDNode *TBAAInfo = I.getMetadata(LLVMContext::MD_tbaa);
3218 const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range);
3220 SmallVector<EVT, 4> ValueVTs;
3221 SmallVector<uint64_t, 4> Offsets;
3222 ComputeValueVTs(TLI, Ty, ValueVTs, &Offsets);
3223 unsigned NumValues = ValueVTs.size();
3228 bool ConstantMemory = false;
3229 if (I.isVolatile() || NumValues > MaxParallelChains)
3230 // Serialize volatile loads with other side effects.
3232 else if (AA->pointsToConstantMemory(
3233 AliasAnalysis::Location(SV, AA->getTypeStoreSize(Ty), TBAAInfo))) {
3234 // Do not serialize (non-volatile) loads of constant memory with anything.
3235 Root = DAG.getEntryNode();
3236 ConstantMemory = true;
3238 // Do not serialize non-volatile loads against each other.
3239 Root = DAG.getRoot();
3242 SmallVector<SDValue, 4> Values(NumValues);
3243 SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains),
3245 EVT PtrVT = Ptr.getValueType();
3246 unsigned ChainI = 0;
3247 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3248 // Serializing loads here may result in excessive register pressure, and
3249 // TokenFactor places arbitrary choke points on the scheduler. SD scheduling
3250 // could recover a bit by hoisting nodes upward in the chain by recognizing
3251 // they are side-effect free or do not alias. The optimizer should really
3252 // avoid this case by converting large object/array copies to llvm.memcpy
3253 // (MaxParallelChains should always remain as failsafe).
3254 if (ChainI == MaxParallelChains) {
3255 assert(PendingLoads.empty() && "PendingLoads must be serialized first");
3256 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3257 MVT::Other, &Chains[0], ChainI);
3261 SDValue A = DAG.getNode(ISD::ADD, getCurDebugLoc(),
3263 DAG.getConstant(Offsets[i], PtrVT));
3264 SDValue L = DAG.getLoad(ValueVTs[i], getCurDebugLoc(), Root,
3265 A, MachinePointerInfo(SV, Offsets[i]), isVolatile,
3266 isNonTemporal, isInvariant, Alignment, TBAAInfo,
3270 Chains[ChainI] = L.getValue(1);
3273 if (!ConstantMemory) {
3274 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3275 MVT::Other, &Chains[0], ChainI);
3279 PendingLoads.push_back(Chain);
3282 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
3283 DAG.getVTList(&ValueVTs[0], NumValues),
3284 &Values[0], NumValues));
3287 void SelectionDAGBuilder::visitStore(const StoreInst &I) {
3289 return visitAtomicStore(I);
3291 const Value *SrcV = I.getOperand(0);
3292 const Value *PtrV = I.getOperand(1);
3294 SmallVector<EVT, 4> ValueVTs;
3295 SmallVector<uint64_t, 4> Offsets;
3296 ComputeValueVTs(TLI, SrcV->getType(), ValueVTs, &Offsets);
3297 unsigned NumValues = ValueVTs.size();
3301 // Get the lowered operands. Note that we do this after
3302 // checking if NumResults is zero, because with zero results
3303 // the operands won't have values in the map.
3304 SDValue Src = getValue(SrcV);
3305 SDValue Ptr = getValue(PtrV);
3307 SDValue Root = getRoot();
3308 SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains),
3310 EVT PtrVT = Ptr.getValueType();
3311 bool isVolatile = I.isVolatile();
3312 bool isNonTemporal = I.getMetadata("nontemporal") != 0;
3313 unsigned Alignment = I.getAlignment();
3314 const MDNode *TBAAInfo = I.getMetadata(LLVMContext::MD_tbaa);
3316 unsigned ChainI = 0;
3317 for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
3318 // See visitLoad comments.
3319 if (ChainI == MaxParallelChains) {
3320 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3321 MVT::Other, &Chains[0], ChainI);
3325 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT, Ptr,
3326 DAG.getConstant(Offsets[i], PtrVT));
3327 SDValue St = DAG.getStore(Root, getCurDebugLoc(),
3328 SDValue(Src.getNode(), Src.getResNo() + i),
3329 Add, MachinePointerInfo(PtrV, Offsets[i]),
3330 isVolatile, isNonTemporal, Alignment, TBAAInfo);
3331 Chains[ChainI] = St;
3334 SDValue StoreNode = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
3335 MVT::Other, &Chains[0], ChainI);
3337 AssignOrderingToNode(StoreNode.getNode());
3338 DAG.setRoot(StoreNode);
3341 static SDValue InsertFenceForAtomic(SDValue Chain, AtomicOrdering Order,
3342 SynchronizationScope Scope,
3343 bool Before, DebugLoc dl,
3345 const TargetLowering &TLI) {
3346 // Fence, if necessary
3348 if (Order == AcquireRelease || Order == SequentiallyConsistent)
3350 else if (Order == Acquire || Order == Monotonic)
3353 if (Order == AcquireRelease)
3355 else if (Order == Release || Order == Monotonic)
3360 Ops[1] = DAG.getConstant(Order, TLI.getPointerTy());
3361 Ops[2] = DAG.getConstant(Scope, TLI.getPointerTy());
3362 return DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops, 3);
3365 void SelectionDAGBuilder::visitAtomicCmpXchg(const AtomicCmpXchgInst &I) {
3366 DebugLoc dl = getCurDebugLoc();
3367 AtomicOrdering Order = I.getOrdering();
3368 SynchronizationScope Scope = I.getSynchScope();
3370 SDValue InChain = getRoot();
3372 if (TLI.getInsertFencesForAtomic())
3373 InChain = InsertFenceForAtomic(InChain, Order, Scope, true, dl,
3377 DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, dl,
3378 getValue(I.getCompareOperand()).getValueType().getSimpleVT(),
3380 getValue(I.getPointerOperand()),
3381 getValue(I.getCompareOperand()),
3382 getValue(I.getNewValOperand()),
3383 MachinePointerInfo(I.getPointerOperand()), 0 /* Alignment */,
3384 TLI.getInsertFencesForAtomic() ? Monotonic : Order,
3387 SDValue OutChain = L.getValue(1);
3389 if (TLI.getInsertFencesForAtomic())
3390 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3394 DAG.setRoot(OutChain);
3397 void SelectionDAGBuilder::visitAtomicRMW(const AtomicRMWInst &I) {
3398 DebugLoc dl = getCurDebugLoc();
3400 switch (I.getOperation()) {
3401 default: llvm_unreachable("Unknown atomicrmw operation");
3402 case AtomicRMWInst::Xchg: NT = ISD::ATOMIC_SWAP; break;
3403 case AtomicRMWInst::Add: NT = ISD::ATOMIC_LOAD_ADD; break;
3404 case AtomicRMWInst::Sub: NT = ISD::ATOMIC_LOAD_SUB; break;
3405 case AtomicRMWInst::And: NT = ISD::ATOMIC_LOAD_AND; break;
3406 case AtomicRMWInst::Nand: NT = ISD::ATOMIC_LOAD_NAND; break;
3407 case AtomicRMWInst::Or: NT = ISD::ATOMIC_LOAD_OR; break;
3408 case AtomicRMWInst::Xor: NT = ISD::ATOMIC_LOAD_XOR; break;
3409 case AtomicRMWInst::Max: NT = ISD::ATOMIC_LOAD_MAX; break;
3410 case AtomicRMWInst::Min: NT = ISD::ATOMIC_LOAD_MIN; break;
3411 case AtomicRMWInst::UMax: NT = ISD::ATOMIC_LOAD_UMAX; break;
3412 case AtomicRMWInst::UMin: NT = ISD::ATOMIC_LOAD_UMIN; break;
3414 AtomicOrdering Order = I.getOrdering();
3415 SynchronizationScope Scope = I.getSynchScope();
3417 SDValue InChain = getRoot();
3419 if (TLI.getInsertFencesForAtomic())
3420 InChain = InsertFenceForAtomic(InChain, Order, Scope, true, dl,
3424 DAG.getAtomic(NT, dl,
3425 getValue(I.getValOperand()).getValueType().getSimpleVT(),
3427 getValue(I.getPointerOperand()),
3428 getValue(I.getValOperand()),
3429 I.getPointerOperand(), 0 /* Alignment */,
3430 TLI.getInsertFencesForAtomic() ? Monotonic : Order,
3433 SDValue OutChain = L.getValue(1);
3435 if (TLI.getInsertFencesForAtomic())
3436 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3440 DAG.setRoot(OutChain);
3443 void SelectionDAGBuilder::visitFence(const FenceInst &I) {
3444 DebugLoc dl = getCurDebugLoc();
3447 Ops[1] = DAG.getConstant(I.getOrdering(), TLI.getPointerTy());
3448 Ops[2] = DAG.getConstant(I.getSynchScope(), TLI.getPointerTy());
3449 DAG.setRoot(DAG.getNode(ISD::ATOMIC_FENCE, dl, MVT::Other, Ops, 3));
3452 void SelectionDAGBuilder::visitAtomicLoad(const LoadInst &I) {
3453 DebugLoc dl = getCurDebugLoc();
3454 AtomicOrdering Order = I.getOrdering();
3455 SynchronizationScope Scope = I.getSynchScope();
3457 SDValue InChain = getRoot();
3459 EVT VT = EVT::getEVT(I.getType());
3461 if (I.getAlignment() * 8 < VT.getSizeInBits())
3462 report_fatal_error("Cannot generate unaligned atomic load");
3465 DAG.getAtomic(ISD::ATOMIC_LOAD, dl, VT, VT, InChain,
3466 getValue(I.getPointerOperand()),
3467 I.getPointerOperand(), I.getAlignment(),
3468 TLI.getInsertFencesForAtomic() ? Monotonic : Order,
3471 SDValue OutChain = L.getValue(1);
3473 if (TLI.getInsertFencesForAtomic())
3474 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3478 DAG.setRoot(OutChain);
3481 void SelectionDAGBuilder::visitAtomicStore(const StoreInst &I) {
3482 DebugLoc dl = getCurDebugLoc();
3484 AtomicOrdering Order = I.getOrdering();
3485 SynchronizationScope Scope = I.getSynchScope();
3487 SDValue InChain = getRoot();
3489 EVT VT = EVT::getEVT(I.getValueOperand()->getType());
3491 if (I.getAlignment() * 8 < VT.getSizeInBits())
3492 report_fatal_error("Cannot generate unaligned atomic store");
3494 if (TLI.getInsertFencesForAtomic())
3495 InChain = InsertFenceForAtomic(InChain, Order, Scope, true, dl,
3499 DAG.getAtomic(ISD::ATOMIC_STORE, dl, VT,
3501 getValue(I.getPointerOperand()),
3502 getValue(I.getValueOperand()),
3503 I.getPointerOperand(), I.getAlignment(),
3504 TLI.getInsertFencesForAtomic() ? Monotonic : Order,
3507 if (TLI.getInsertFencesForAtomic())
3508 OutChain = InsertFenceForAtomic(OutChain, Order, Scope, false, dl,
3511 DAG.setRoot(OutChain);
3514 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
3516 void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I,
3517 unsigned Intrinsic) {
3518 bool HasChain = !I.doesNotAccessMemory();
3519 bool OnlyLoad = HasChain && I.onlyReadsMemory();
3521 // Build the operand list.
3522 SmallVector<SDValue, 8> Ops;
3523 if (HasChain) { // If this intrinsic has side-effects, chainify it.
3525 // We don't need to serialize loads against other loads.
3526 Ops.push_back(DAG.getRoot());
3528 Ops.push_back(getRoot());
3532 // Info is set by getTgtMemInstrinsic
3533 TargetLowering::IntrinsicInfo Info;
3534 bool IsTgtIntrinsic = TLI.getTgtMemIntrinsic(Info, I, Intrinsic);
3536 // Add the intrinsic ID as an integer operand if it's not a target intrinsic.
3537 if (!IsTgtIntrinsic || Info.opc == ISD::INTRINSIC_VOID ||
3538 Info.opc == ISD::INTRINSIC_W_CHAIN)
3539 Ops.push_back(DAG.getTargetConstant(Intrinsic, TLI.getPointerTy()));
3541 // Add all operands of the call to the operand list.
3542 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) {
3543 SDValue Op = getValue(I.getArgOperand(i));
3547 SmallVector<EVT, 4> ValueVTs;
3548 ComputeValueVTs(TLI, I.getType(), ValueVTs);
3551 ValueVTs.push_back(MVT::Other);
3553 SDVTList VTs = DAG.getVTList(ValueVTs.data(), ValueVTs.size());
3557 if (IsTgtIntrinsic) {
3558 // This is target intrinsic that touches memory
3559 Result = DAG.getMemIntrinsicNode(Info.opc, getCurDebugLoc(),
3560 VTs, &Ops[0], Ops.size(),
3562 MachinePointerInfo(Info.ptrVal, Info.offset),
3563 Info.align, Info.vol,
3564 Info.readMem, Info.writeMem);
3565 } else if (!HasChain) {
3566 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurDebugLoc(),
3567 VTs, &Ops[0], Ops.size());
3568 } else if (!I.getType()->isVoidTy()) {
3569 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurDebugLoc(),
3570 VTs, &Ops[0], Ops.size());
3572 Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurDebugLoc(),
3573 VTs, &Ops[0], Ops.size());
3577 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1);
3579 PendingLoads.push_back(Chain);
3584 if (!I.getType()->isVoidTy()) {
3585 if (VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
3586 EVT VT = TLI.getValueType(PTy);
3587 Result = DAG.getNode(ISD::BITCAST, getCurDebugLoc(), VT, Result);
3590 setValue(&I, Result);
3592 // Assign order to result here. If the intrinsic does not produce a result,
3593 // it won't be mapped to a SDNode and visit() will not assign it an order
3596 AssignOrderingToNode(Result.getNode());
3600 /// GetSignificand - Get the significand and build it into a floating-point
3601 /// number with exponent of 1:
3603 /// Op = (Op & 0x007fffff) | 0x3f800000;
3605 /// where Op is the hexidecimal representation of floating point value.
3607 GetSignificand(SelectionDAG &DAG, SDValue Op, DebugLoc dl) {
3608 SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3609 DAG.getConstant(0x007fffff, MVT::i32));
3610 SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1,
3611 DAG.getConstant(0x3f800000, MVT::i32));
3612 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, t2);
3615 /// GetExponent - Get the exponent:
3617 /// (float)(int)(((Op & 0x7f800000) >> 23) - 127);
3619 /// where Op is the hexidecimal representation of floating point value.
3621 GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI,
3623 SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3624 DAG.getConstant(0x7f800000, MVT::i32));
3625 SDValue t1 = DAG.getNode(ISD::SRL, dl, MVT::i32, t0,
3626 DAG.getConstant(23, TLI.getPointerTy()));
3627 SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1,
3628 DAG.getConstant(127, MVT::i32));
3629 return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2);
3632 /// getF32Constant - Get 32-bit floating point constant.
3634 getF32Constant(SelectionDAG &DAG, unsigned Flt) {
3635 return DAG.getConstantFP(APFloat(APInt(32, Flt)), MVT::f32);
3638 /// visitExp - Lower an exp intrinsic. Handles the special sequences for
3639 /// limited-precision mode.
3641 SelectionDAGBuilder::visitExp(const CallInst &I) {
3643 DebugLoc dl = getCurDebugLoc();
3645 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3646 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3647 SDValue Op = getValue(I.getArgOperand(0));
3649 // Put the exponent in the right bit position for later addition to the
3652 // #define LOG2OFe 1.4426950f
3653 // IntegerPartOfX = ((int32_t)(X * LOG2OFe));
3654 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3655 getF32Constant(DAG, 0x3fb8aa3b));
3656 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3658 // FractionalPartOfX = (X * LOG2OFe) - (float)IntegerPartOfX;
3659 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3660 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3662 // IntegerPartOfX <<= 23;
3663 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3664 DAG.getConstant(23, TLI.getPointerTy()));
3666 if (LimitFloatPrecision <= 6) {
3667 // For floating-point precision of 6:
3669 // TwoToFractionalPartOfX =
3671 // (0.735607626f + 0.252464424f * x) * x;
3673 // error 0.0144103317, which is 6 bits
3674 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3675 getF32Constant(DAG, 0x3e814304));
3676 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3677 getF32Constant(DAG, 0x3f3c50c8));
3678 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3679 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3680 getF32Constant(DAG, 0x3f7f5e7e));
3681 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BITCAST, dl,MVT::i32, t5);
3683 // Add the exponent into the result in integer domain.
3684 SDValue t6 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3685 TwoToFracPartOfX, IntegerPartOfX);
3687 result = DAG.getNode(ISD::BITCAST, dl, MVT::f32, t6);
3688 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3689 // For floating-point precision of 12:
3691 // TwoToFractionalPartOfX =
3694 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3696 // 0.000107046256 error, which is 13 to 14 bits
3697 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3698 getF32Constant(DAG, 0x3da235e3));
3699 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3700 getF32Constant(DAG, 0x3e65b8f3));
3701 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3702 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3703 getF32Constant(DAG, 0x3f324b07));
3704 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3705 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3706 getF32Constant(DAG, 0x3f7ff8fd));
3707 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BITCAST, dl,MVT::i32, t7);
3709 // Add the exponent into the result in integer domain.
3710 SDValue t8 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3711 TwoToFracPartOfX, IntegerPartOfX);
3713 result = DAG.getNode(ISD::BITCAST, dl, MVT::f32, t8);
3714 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3715 // For floating-point precision of 18:
3717 // TwoToFractionalPartOfX =
3721 // (0.554906021e-1f +
3722 // (0.961591928e-2f +
3723 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3725 // error 2.47208000*10^(-7), which is better than 18 bits
3726 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3727 getF32Constant(DAG, 0x3924b03e));
3728 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3729 getF32Constant(DAG, 0x3ab24b87));
3730 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3731 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3732 getF32Constant(DAG, 0x3c1d8c17));
3733 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3734 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3735 getF32Constant(DAG, 0x3d634a1d));
3736 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3737 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3738 getF32Constant(DAG, 0x3e75fe14));
3739 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3740 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3741 getF32Constant(DAG, 0x3f317234));
3742 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3743 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3744 getF32Constant(DAG, 0x3f800000));
3745 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BITCAST, dl,
3748 // Add the exponent into the result in integer domain.
3749 SDValue t14 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3750 TwoToFracPartOfX, IntegerPartOfX);
3752 result = DAG.getNode(ISD::BITCAST, dl, MVT::f32, t14);
3755 // No special expansion.
3756 result = DAG.getNode(ISD::FEXP, dl,
3757 getValue(I.getArgOperand(0)).getValueType(),
3758 getValue(I.getArgOperand(0)));
3761 setValue(&I, result);
3764 /// visitLog - Lower a log intrinsic. Handles the special sequences for
3765 /// limited-precision mode.
3767 SelectionDAGBuilder::visitLog(const CallInst &I) {
3769 DebugLoc dl = getCurDebugLoc();
3771 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3772 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3773 SDValue Op = getValue(I.getArgOperand(0));
3774 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3776 // Scale the exponent by log(2) [0.69314718f].
3777 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3778 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3779 getF32Constant(DAG, 0x3f317218));
3781 // Get the significand and build it into a floating-point number with
3783 SDValue X = GetSignificand(DAG, Op1, dl);
3785 if (LimitFloatPrecision <= 6) {
3786 // For floating-point precision of 6:
3790 // (1.4034025f - 0.23903021f * x) * x;
3792 // error 0.0034276066, which is better than 8 bits
3793 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3794 getF32Constant(DAG, 0xbe74c456));
3795 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3796 getF32Constant(DAG, 0x3fb3a2b1));
3797 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3798 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3799 getF32Constant(DAG, 0x3f949a29));
3801 result = DAG.getNode(ISD::FADD, dl,
3802 MVT::f32, LogOfExponent, LogOfMantissa);
3803 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3804 // For floating-point precision of 12:
3810 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x;
3812 // error 0.000061011436, which is 14 bits
3813 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3814 getF32Constant(DAG, 0xbd67b6d6));
3815 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3816 getF32Constant(DAG, 0x3ee4f4b8));
3817 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3818 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3819 getF32Constant(DAG, 0x3fbc278b));
3820 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3821 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3822 getF32Constant(DAG, 0x40348e95));
3823 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3824 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3825 getF32Constant(DAG, 0x3fdef31a));
3827 result = DAG.getNode(ISD::FADD, dl,
3828 MVT::f32, LogOfExponent, LogOfMantissa);
3829 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3830 // For floating-point precision of 18:
3838 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x;
3840 // error 0.0000023660568, which is better than 18 bits
3841 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3842 getF32Constant(DAG, 0xbc91e5ac));
3843 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3844 getF32Constant(DAG, 0x3e4350aa));
3845 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3846 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3847 getF32Constant(DAG, 0x3f60d3e3));
3848 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3849 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3850 getF32Constant(DAG, 0x4011cdf0));
3851 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3852 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3853 getF32Constant(DAG, 0x406cfd1c));
3854 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3855 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3856 getF32Constant(DAG, 0x408797cb));
3857 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3858 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3859 getF32Constant(DAG, 0x4006dcab));
3861 result = DAG.getNode(ISD::FADD, dl,
3862 MVT::f32, LogOfExponent, LogOfMantissa);
3865 // No special expansion.
3866 result = DAG.getNode(ISD::FLOG, dl,
3867 getValue(I.getArgOperand(0)).getValueType(),
3868 getValue(I.getArgOperand(0)));
3871 setValue(&I, result);
3874 /// visitLog2 - Lower a log2 intrinsic. Handles the special sequences for
3875 /// limited-precision mode.
3877 SelectionDAGBuilder::visitLog2(const CallInst &I) {
3879 DebugLoc dl = getCurDebugLoc();
3881 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3882 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3883 SDValue Op = getValue(I.getArgOperand(0));
3884 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3886 // Get the exponent.
3887 SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl);
3889 // Get the significand and build it into a floating-point number with
3891 SDValue X = GetSignificand(DAG, Op1, dl);
3893 // Different possible minimax approximations of significand in
3894 // floating-point for various degrees of accuracy over [1,2].
3895 if (LimitFloatPrecision <= 6) {
3896 // For floating-point precision of 6:
3898 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x;
3900 // error 0.0049451742, which is more than 7 bits
3901 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3902 getF32Constant(DAG, 0xbeb08fe0));
3903 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3904 getF32Constant(DAG, 0x40019463));
3905 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3906 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3907 getF32Constant(DAG, 0x3fd6633d));
3909 result = DAG.getNode(ISD::FADD, dl,
3910 MVT::f32, LogOfExponent, Log2ofMantissa);
3911 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3912 // For floating-point precision of 12:
3918 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x;
3920 // error 0.0000876136000, which is better than 13 bits
3921 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3922 getF32Constant(DAG, 0xbda7262e));
3923 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3924 getF32Constant(DAG, 0x3f25280b));
3925 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3926 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3927 getF32Constant(DAG, 0x4007b923));
3928 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3929 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3930 getF32Constant(DAG, 0x40823e2f));
3931 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3932 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3933 getF32Constant(DAG, 0x4020d29c));
3935 result = DAG.getNode(ISD::FADD, dl,
3936 MVT::f32, LogOfExponent, Log2ofMantissa);
3937 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3938 // For floating-point precision of 18:
3947 // 0.25691327e-1f * x) * x) * x) * x) * x) * x;
3949 // error 0.0000018516, which is better than 18 bits
3950 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3951 getF32Constant(DAG, 0xbcd2769e));
3952 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3953 getF32Constant(DAG, 0x3e8ce0b9));
3954 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3955 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3956 getF32Constant(DAG, 0x3fa22ae7));
3957 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3958 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3959 getF32Constant(DAG, 0x40525723));
3960 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3961 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3962 getF32Constant(DAG, 0x40aaf200));
3963 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3964 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3965 getF32Constant(DAG, 0x40c39dad));
3966 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3967 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3968 getF32Constant(DAG, 0x4042902c));
3970 result = DAG.getNode(ISD::FADD, dl,
3971 MVT::f32, LogOfExponent, Log2ofMantissa);
3974 // No special expansion.
3975 result = DAG.getNode(ISD::FLOG2, dl,
3976 getValue(I.getArgOperand(0)).getValueType(),
3977 getValue(I.getArgOperand(0)));
3980 setValue(&I, result);
3983 /// visitLog10 - Lower a log10 intrinsic. Handles the special sequences for
3984 /// limited-precision mode.
3986 SelectionDAGBuilder::visitLog10(const CallInst &I) {
3988 DebugLoc dl = getCurDebugLoc();
3990 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
3991 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3992 SDValue Op = getValue(I.getArgOperand(0));
3993 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3995 // Scale the exponent by log10(2) [0.30102999f].
3996 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3997 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3998 getF32Constant(DAG, 0x3e9a209a));
4000 // Get the significand and build it into a floating-point number with
4002 SDValue X = GetSignificand(DAG, Op1, dl);
4004 if (LimitFloatPrecision <= 6) {
4005 // For floating-point precision of 6:
4007 // Log10ofMantissa =
4009 // (0.60948995f - 0.10380950f * x) * x;
4011 // error 0.0014886165, which is 6 bits
4012 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4013 getF32Constant(DAG, 0xbdd49a13));
4014 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
4015 getF32Constant(DAG, 0x3f1c0789));
4016 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4017 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
4018 getF32Constant(DAG, 0x3f011300));
4020 result = DAG.getNode(ISD::FADD, dl,
4021 MVT::f32, LogOfExponent, Log10ofMantissa);
4022 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
4023 // For floating-point precision of 12:
4025 // Log10ofMantissa =
4028 // (-0.31664806f + 0.47637168e-1f * x) * x) * x;
4030 // error 0.00019228036, which is better than 12 bits
4031 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4032 getF32Constant(DAG, 0x3d431f31));
4033 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
4034 getF32Constant(DAG, 0x3ea21fb2));
4035 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4036 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4037 getF32Constant(DAG, 0x3f6ae232));
4038 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4039 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
4040 getF32Constant(DAG, 0x3f25f7c3));
4042 result = DAG.getNode(ISD::FADD, dl,
4043 MVT::f32, LogOfExponent, Log10ofMantissa);
4044 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
4045 // For floating-point precision of 18:
4047 // Log10ofMantissa =
4052 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x;
4054 // error 0.0000037995730, which is better than 18 bits
4055 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4056 getF32Constant(DAG, 0x3c5d51ce));
4057 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
4058 getF32Constant(DAG, 0x3e00685a));
4059 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
4060 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4061 getF32Constant(DAG, 0x3efb6798));
4062 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4063 SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
4064 getF32Constant(DAG, 0x3f88d192));
4065 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4066 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4067 getF32Constant(DAG, 0x3fc4316c));
4068 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4069 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8,
4070 getF32Constant(DAG, 0x3f57ce70));
4072 result = DAG.getNode(ISD::FADD, dl,
4073 MVT::f32, LogOfExponent, Log10ofMantissa);
4076 // No special expansion.
4077 result = DAG.getNode(ISD::FLOG10, dl,
4078 getValue(I.getArgOperand(0)).getValueType(),
4079 getValue(I.getArgOperand(0)));
4082 setValue(&I, result);
4085 /// visitExp2 - Lower an exp2 intrinsic. Handles the special sequences for
4086 /// limited-precision mode.
4088 SelectionDAGBuilder::visitExp2(const CallInst &I) {
4090 DebugLoc dl = getCurDebugLoc();
4092 if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
4093 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4094 SDValue Op = getValue(I.getArgOperand(0));
4096 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Op);
4098 // FractionalPartOfX = x - (float)IntegerPartOfX;
4099 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
4100 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, Op, t1);
4102 // IntegerPartOfX <<= 23;
4103 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
4104 DAG.getConstant(23, TLI.getPointerTy()));
4106 if (LimitFloatPrecision <= 6) {
4107 // For floating-point precision of 6:
4109 // TwoToFractionalPartOfX =
4111 // (0.735607626f + 0.252464424f * x) * x;
4113 // error 0.0144103317, which is 6 bits
4114 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4115 getF32Constant(DAG, 0x3e814304));
4116 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4117 getF32Constant(DAG, 0x3f3c50c8));
4118 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4119 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4120 getF32Constant(DAG, 0x3f7f5e7e));
4121 SDValue t6 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t5);
4122 SDValue TwoToFractionalPartOfX =
4123 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
4125 result = DAG.getNode(ISD::BITCAST, dl,
4126 MVT::f32, TwoToFractionalPartOfX);
4127 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
4128 // For floating-point precision of 12:
4130 // TwoToFractionalPartOfX =
4133 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
4135 // error 0.000107046256, which is 13 to 14 bits
4136 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4137 getF32Constant(DAG, 0x3da235e3));
4138 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4139 getF32Constant(DAG, 0x3e65b8f3));
4140 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4141 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4142 getF32Constant(DAG, 0x3f324b07));
4143 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4144 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4145 getF32Constant(DAG, 0x3f7ff8fd));
4146 SDValue t8 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t7);
4147 SDValue TwoToFractionalPartOfX =
4148 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
4150 result = DAG.getNode(ISD::BITCAST, dl,
4151 MVT::f32, TwoToFractionalPartOfX);
4152 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
4153 // For floating-point precision of 18:
4155 // TwoToFractionalPartOfX =
4159 // (0.554906021e-1f +
4160 // (0.961591928e-2f +
4161 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
4162 // error 2.47208000*10^(-7), which is better than 18 bits
4163 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4164 getF32Constant(DAG, 0x3924b03e));
4165 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4166 getF32Constant(DAG, 0x3ab24b87));
4167 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4168 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4169 getF32Constant(DAG, 0x3c1d8c17));
4170 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4171 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4172 getF32Constant(DAG, 0x3d634a1d));
4173 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4174 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4175 getF32Constant(DAG, 0x3e75fe14));
4176 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4177 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
4178 getF32Constant(DAG, 0x3f317234));
4179 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
4180 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
4181 getF32Constant(DAG, 0x3f800000));
4182 SDValue t14 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t13);
4183 SDValue TwoToFractionalPartOfX =
4184 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
4186 result = DAG.getNode(ISD::BITCAST, dl,
4187 MVT::f32, TwoToFractionalPartOfX);
4190 // No special expansion.
4191 result = DAG.getNode(ISD::FEXP2, dl,
4192 getValue(I.getArgOperand(0)).getValueType(),
4193 getValue(I.getArgOperand(0)));
4196 setValue(&I, result);
4199 /// visitPow - Lower a pow intrinsic. Handles the special sequences for
4200 /// limited-precision mode with x == 10.0f.
4202 SelectionDAGBuilder::visitPow(const CallInst &I) {
4204 const Value *Val = I.getArgOperand(0);
4205 DebugLoc dl = getCurDebugLoc();
4206 bool IsExp10 = false;
4208 if (getValue(Val).getValueType() == MVT::f32 &&
4209 getValue(I.getArgOperand(1)).getValueType() == MVT::f32 &&
4210 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4211 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(Val))) {
4212 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
4214 IsExp10 = CFP->getValueAPF().bitwiseIsEqual(Ten);
4219 if (IsExp10 && LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
4220 SDValue Op = getValue(I.getArgOperand(1));
4222 // Put the exponent in the right bit position for later addition to the
4225 // #define LOG2OF10 3.3219281f
4226 // IntegerPartOfX = (int32_t)(x * LOG2OF10);
4227 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
4228 getF32Constant(DAG, 0x40549a78));
4229 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
4231 // FractionalPartOfX = x - (float)IntegerPartOfX;
4232 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
4233 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
4235 // IntegerPartOfX <<= 23;
4236 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
4237 DAG.getConstant(23, TLI.getPointerTy()));
4239 if (LimitFloatPrecision <= 6) {
4240 // For floating-point precision of 6:
4242 // twoToFractionalPartOfX =
4244 // (0.735607626f + 0.252464424f * x) * x;
4246 // error 0.0144103317, which is 6 bits
4247 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4248 getF32Constant(DAG, 0x3e814304));
4249 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4250 getF32Constant(DAG, 0x3f3c50c8));
4251 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4252 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4253 getF32Constant(DAG, 0x3f7f5e7e));
4254 SDValue t6 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t5);
4255 SDValue TwoToFractionalPartOfX =
4256 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
4258 result = DAG.getNode(ISD::BITCAST, dl,
4259 MVT::f32, TwoToFractionalPartOfX);
4260 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
4261 // For floating-point precision of 12:
4263 // TwoToFractionalPartOfX =
4266 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
4268 // error 0.000107046256, which is 13 to 14 bits
4269 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4270 getF32Constant(DAG, 0x3da235e3));
4271 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4272 getF32Constant(DAG, 0x3e65b8f3));
4273 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4274 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4275 getF32Constant(DAG, 0x3f324b07));
4276 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4277 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4278 getF32Constant(DAG, 0x3f7ff8fd));
4279 SDValue t8 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t7);
4280 SDValue TwoToFractionalPartOfX =
4281 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
4283 result = DAG.getNode(ISD::BITCAST, dl,
4284 MVT::f32, TwoToFractionalPartOfX);
4285 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
4286 // For floating-point precision of 18:
4288 // TwoToFractionalPartOfX =
4292 // (0.554906021e-1f +
4293 // (0.961591928e-2f +
4294 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
4295 // error 2.47208000*10^(-7), which is better than 18 bits
4296 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
4297 getF32Constant(DAG, 0x3924b03e));
4298 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
4299 getF32Constant(DAG, 0x3ab24b87));
4300 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
4301 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
4302 getF32Constant(DAG, 0x3c1d8c17));
4303 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
4304 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
4305 getF32Constant(DAG, 0x3d634a1d));
4306 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
4307 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
4308 getF32Constant(DAG, 0x3e75fe14));
4309 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
4310 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
4311 getF32Constant(DAG, 0x3f317234));
4312 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
4313 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
4314 getF32Constant(DAG, 0x3f800000));
4315 SDValue t14 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t13);
4316 SDValue TwoToFractionalPartOfX =
4317 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
4319 result = DAG.getNode(ISD::BITCAST, dl,
4320 MVT::f32, TwoToFractionalPartOfX);
4323 // No special expansion.
4324 result = DAG.getNode(ISD::FPOW, dl,
4325 getValue(I.getArgOperand(0)).getValueType(),
4326 getValue(I.getArgOperand(0)),
4327 getValue(I.getArgOperand(1)));
4330 setValue(&I, result);
4334 /// ExpandPowI - Expand a llvm.powi intrinsic.
4335 static SDValue ExpandPowI(DebugLoc DL, SDValue LHS, SDValue RHS,
4336 SelectionDAG &DAG) {
4337 // If RHS is a constant, we can expand this out to a multiplication tree,
4338 // otherwise we end up lowering to a call to __powidf2 (for example). When
4339 // optimizing for size, we only want to do this if the expansion would produce
4340 // a small number of multiplies, otherwise we do the full expansion.
4341 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
4342 // Get the exponent as a positive value.
4343 unsigned Val = RHSC->getSExtValue();
4344 if ((int)Val < 0) Val = -Val;
4346 // powi(x, 0) -> 1.0
4348 return DAG.getConstantFP(1.0, LHS.getValueType());
4350 const Function *F = DAG.getMachineFunction().getFunction();
4351 if (!F->hasFnAttr(Attribute::OptimizeForSize) ||
4352 // If optimizing for size, don't insert too many multiplies. This
4353 // inserts up to 5 multiplies.
4354 CountPopulation_32(Val)+Log2_32(Val) < 7) {
4355 // We use the simple binary decomposition method to generate the multiply
4356 // sequence. There are more optimal ways to do this (for example,
4357 // powi(x,15) generates one more multiply than it should), but this has
4358 // the benefit of being both really simple and much better than a libcall.
4359 SDValue Res; // Logically starts equal to 1.0
4360 SDValue CurSquare = LHS;
4364 Res = DAG.getNode(ISD::FMUL, DL,Res.getValueType(), Res, CurSquare);
4366 Res = CurSquare; // 1.0*CurSquare.
4369 CurSquare = DAG.getNode(ISD::FMUL, DL, CurSquare.getValueType(),
4370 CurSquare, CurSquare);
4374 // If the original was negative, invert the result, producing 1/(x*x*x).
4375 if (RHSC->getSExtValue() < 0)
4376 Res = DAG.getNode(ISD::FDIV, DL, LHS.getValueType(),
4377 DAG.getConstantFP(1.0, LHS.getValueType()), Res);
4382 // Otherwise, expand to a libcall.
4383 return DAG.getNode(ISD::FPOWI, DL, LHS.getValueType(), LHS, RHS);
4386 // getTruncatedArgReg - Find underlying register used for an truncated
4388 static unsigned getTruncatedArgReg(const SDValue &N) {
4389 if (N.getOpcode() != ISD::TRUNCATE)
4392 const SDValue &Ext = N.getOperand(0);
4393 if (Ext.getOpcode() == ISD::AssertZext || Ext.getOpcode() == ISD::AssertSext){
4394 const SDValue &CFR = Ext.getOperand(0);
4395 if (CFR.getOpcode() == ISD::CopyFromReg)
4396 return cast<RegisterSDNode>(CFR.getOperand(1))->getReg();
4397 if (CFR.getOpcode() == ISD::TRUNCATE)
4398 return getTruncatedArgReg(CFR);
4403 /// EmitFuncArgumentDbgValue - If the DbgValueInst is a dbg_value of a function
4404 /// argument, create the corresponding DBG_VALUE machine instruction for it now.
4405 /// At the end of instruction selection, they will be inserted to the entry BB.
4407 SelectionDAGBuilder::EmitFuncArgumentDbgValue(const Value *V, MDNode *Variable,
4410 const Argument *Arg = dyn_cast<Argument>(V);
4414 MachineFunction &MF = DAG.getMachineFunction();
4415 const TargetInstrInfo *TII = DAG.getTarget().getInstrInfo();
4416 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
4418 // Ignore inlined function arguments here.
4419 DIVariable DV(Variable);
4420 if (DV.isInlinedFnArgument(MF.getFunction()))
4424 // Some arguments' frame index is recorded during argument lowering.
4425 Offset = FuncInfo.getArgumentFrameIndex(Arg);
4427 Reg = TRI->getFrameRegister(MF);
4429 if (!Reg && N.getNode()) {
4430 if (N.getOpcode() == ISD::CopyFromReg)
4431 Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg();
4433 Reg = getTruncatedArgReg(N);
4434 if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) {
4435 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4436 unsigned PR = RegInfo.getLiveInPhysReg(Reg);
4443 // Check if ValueMap has reg number.
4444 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
4445 if (VMI != FuncInfo.ValueMap.end())
4449 if (!Reg && N.getNode()) {
4450 // Check if frame index is available.
4451 if (LoadSDNode *LNode = dyn_cast<LoadSDNode>(N.getNode()))
4452 if (FrameIndexSDNode *FINode =
4453 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode())) {
4454 Reg = TRI->getFrameRegister(MF);
4455 Offset = FINode->getIndex();
4462 MachineInstrBuilder MIB = BuildMI(MF, getCurDebugLoc(),
4463 TII->get(TargetOpcode::DBG_VALUE))
4464 .addReg(Reg, RegState::Debug).addImm(Offset).addMetadata(Variable);
4465 FuncInfo.ArgDbgValues.push_back(&*MIB);
4469 // VisualStudio defines setjmp as _setjmp
4470 #if defined(_MSC_VER) && defined(setjmp) && \
4471 !defined(setjmp_undefined_for_msvc)
4472 # pragma push_macro("setjmp")
4474 # define setjmp_undefined_for_msvc
4477 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
4478 /// we want to emit this as a call to a named external function, return the name
4479 /// otherwise lower it and return null.
4481 SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) {
4482 DebugLoc dl = getCurDebugLoc();
4485 switch (Intrinsic) {
4487 // By default, turn this into a target intrinsic node.
4488 visitTargetIntrinsic(I, Intrinsic);
4490 case Intrinsic::vastart: visitVAStart(I); return 0;
4491 case Intrinsic::vaend: visitVAEnd(I); return 0;
4492 case Intrinsic::vacopy: visitVACopy(I); return 0;
4493 case Intrinsic::returnaddress:
4494 setValue(&I, DAG.getNode(ISD::RETURNADDR, dl, TLI.getPointerTy(),
4495 getValue(I.getArgOperand(0))));
4497 case Intrinsic::frameaddress:
4498 setValue(&I, DAG.getNode(ISD::FRAMEADDR, dl, TLI.getPointerTy(),
4499 getValue(I.getArgOperand(0))));
4501 case Intrinsic::setjmp:
4502 return &"_setjmp"[!TLI.usesUnderscoreSetJmp()];
4503 case Intrinsic::longjmp:
4504 return &"_longjmp"[!TLI.usesUnderscoreLongJmp()];
4505 case Intrinsic::memcpy: {
4506 // Assert for address < 256 since we support only user defined address
4508 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4510 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4512 "Unknown address space");
4513 SDValue Op1 = getValue(I.getArgOperand(0));
4514 SDValue Op2 = getValue(I.getArgOperand(1));
4515 SDValue Op3 = getValue(I.getArgOperand(2));
4516 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4517 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4518 DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, isVol, false,
4519 MachinePointerInfo(I.getArgOperand(0)),
4520 MachinePointerInfo(I.getArgOperand(1))));
4523 case Intrinsic::memset: {
4524 // Assert for address < 256 since we support only user defined address
4526 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4528 "Unknown address space");
4529 SDValue Op1 = getValue(I.getArgOperand(0));
4530 SDValue Op2 = getValue(I.getArgOperand(1));
4531 SDValue Op3 = getValue(I.getArgOperand(2));
4532 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4533 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4534 DAG.setRoot(DAG.getMemset(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4535 MachinePointerInfo(I.getArgOperand(0))));
4538 case Intrinsic::memmove: {
4539 // Assert for address < 256 since we support only user defined address
4541 assert(cast<PointerType>(I.getArgOperand(0)->getType())->getAddressSpace()
4543 cast<PointerType>(I.getArgOperand(1)->getType())->getAddressSpace()
4545 "Unknown address space");
4546 SDValue Op1 = getValue(I.getArgOperand(0));
4547 SDValue Op2 = getValue(I.getArgOperand(1));
4548 SDValue Op3 = getValue(I.getArgOperand(2));
4549 unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
4550 bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
4551 DAG.setRoot(DAG.getMemmove(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
4552 MachinePointerInfo(I.getArgOperand(0)),
4553 MachinePointerInfo(I.getArgOperand(1))));
4556 case Intrinsic::dbg_declare: {
4557 const DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
4558 MDNode *Variable = DI.getVariable();
4559 const Value *Address = DI.getAddress();
4560 if (!Address || !DIVariable(Variable).Verify()) {
4561 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4565 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder
4566 // but do not always have a corresponding SDNode built. The SDNodeOrder
4567 // absolute, but not relative, values are different depending on whether
4568 // debug info exists.
4571 // Check if address has undef value.
4572 if (isa<UndefValue>(Address) ||
4573 (Address->use_empty() && !isa<Argument>(Address))) {
4574 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4578 SDValue &N = NodeMap[Address];
4579 if (!N.getNode() && isa<Argument>(Address))
4580 // Check unused arguments map.
4581 N = UnusedArgNodeMap[Address];
4584 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
4585 Address = BCI->getOperand(0);
4586 // Parameters are handled specially.
4588 (DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable ||
4589 isa<Argument>(Address));
4591 const AllocaInst *AI = dyn_cast<AllocaInst>(Address);
4593 if (isParameter && !AI) {
4594 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode());
4596 // Byval parameter. We have a frame index at this point.
4597 SDV = DAG.getDbgValue(Variable, FINode->getIndex(),
4598 0, dl, SDNodeOrder);
4600 // Address is an argument, so try to emit its dbg value using
4601 // virtual register info from the FuncInfo.ValueMap.
4602 EmitFuncArgumentDbgValue(Address, Variable, 0, N);
4606 SDV = DAG.getDbgValue(Variable, N.getNode(), N.getResNo(),
4607 0, dl, SDNodeOrder);
4609 // Can't do anything with other non-AI cases yet.
4610 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4611 DEBUG(dbgs() << "non-AllocaInst issue for Address: \n\t");
4612 DEBUG(Address->dump());
4615 DAG.AddDbgValue(SDV, N.getNode(), isParameter);
4617 // If Address is an argument then try to emit its dbg value using
4618 // virtual register info from the FuncInfo.ValueMap.
4619 if (!EmitFuncArgumentDbgValue(Address, Variable, 0, N)) {
4620 // If variable is pinned by a alloca in dominating bb then
4621 // use StaticAllocaMap.
4622 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Address)) {
4623 if (AI->getParent() != DI.getParent()) {
4624 DenseMap<const AllocaInst*, int>::iterator SI =
4625 FuncInfo.StaticAllocaMap.find(AI);
4626 if (SI != FuncInfo.StaticAllocaMap.end()) {
4627 SDV = DAG.getDbgValue(Variable, SI->second,
4628 0, dl, SDNodeOrder);
4629 DAG.AddDbgValue(SDV, 0, false);
4634 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4639 case Intrinsic::dbg_value: {
4640 const DbgValueInst &DI = cast<DbgValueInst>(I);
4641 if (!DIVariable(DI.getVariable()).Verify())
4644 MDNode *Variable = DI.getVariable();
4645 uint64_t Offset = DI.getOffset();
4646 const Value *V = DI.getValue();
4650 // Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder
4651 // but do not always have a corresponding SDNode built. The SDNodeOrder
4652 // absolute, but not relative, values are different depending on whether
4653 // debug info exists.
4656 if (isa<ConstantInt>(V) || isa<ConstantFP>(V) || isa<UndefValue>(V)) {
4657 SDV = DAG.getDbgValue(Variable, V, Offset, dl, SDNodeOrder);
4658 DAG.AddDbgValue(SDV, 0, false);
4660 // Do not use getValue() in here; we don't want to generate code at
4661 // this point if it hasn't been done yet.
4662 SDValue N = NodeMap[V];
4663 if (!N.getNode() && isa<Argument>(V))
4664 // Check unused arguments map.
4665 N = UnusedArgNodeMap[V];
4667 if (!EmitFuncArgumentDbgValue(V, Variable, Offset, N)) {
4668 SDV = DAG.getDbgValue(Variable, N.getNode(),
4669 N.getResNo(), Offset, dl, SDNodeOrder);
4670 DAG.AddDbgValue(SDV, N.getNode(), false);
4672 } else if (!V->use_empty() ) {
4673 // Do not call getValue(V) yet, as we don't want to generate code.
4674 // Remember it for later.
4675 DanglingDebugInfo DDI(&DI, dl, SDNodeOrder);
4676 DanglingDebugInfoMap[V] = DDI;
4678 // We may expand this to cover more cases. One case where we have no
4679 // data available is an unreferenced parameter.
4680 DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
4684 // Build a debug info table entry.
4685 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(V))
4686 V = BCI->getOperand(0);
4687 const AllocaInst *AI = dyn_cast<AllocaInst>(V);
4688 // Don't handle byval struct arguments or VLAs, for example.
4690 DEBUG(dbgs() << "Dropping debug location info for:\n " << DI << "\n");
4691 DEBUG(dbgs() << " Last seen at:\n " << *V << "\n");
4694 DenseMap<const AllocaInst*, int>::iterator SI =
4695 FuncInfo.StaticAllocaMap.find(AI);
4696 if (SI == FuncInfo.StaticAllocaMap.end())
4698 int FI = SI->second;
4700 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4701 if (!DI.getDebugLoc().isUnknown() && MMI.hasDebugInfo())
4702 MMI.setVariableDbgInfo(Variable, FI, DI.getDebugLoc());
4706 case Intrinsic::eh_typeid_for: {
4707 // Find the type id for the given typeinfo.
4708 GlobalVariable *GV = ExtractTypeInfo(I.getArgOperand(0));
4709 unsigned TypeID = DAG.getMachineFunction().getMMI().getTypeIDFor(GV);
4710 Res = DAG.getConstant(TypeID, MVT::i32);
4715 case Intrinsic::eh_return_i32:
4716 case Intrinsic::eh_return_i64:
4717 DAG.getMachineFunction().getMMI().setCallsEHReturn(true);
4718 DAG.setRoot(DAG.getNode(ISD::EH_RETURN, dl,
4721 getValue(I.getArgOperand(0)),
4722 getValue(I.getArgOperand(1))));
4724 case Intrinsic::eh_unwind_init:
4725 DAG.getMachineFunction().getMMI().setCallsUnwindInit(true);
4727 case Intrinsic::eh_dwarf_cfa: {
4728 SDValue CfaArg = DAG.getSExtOrTrunc(getValue(I.getArgOperand(0)), dl,
4729 TLI.getPointerTy());
4730 SDValue Offset = DAG.getNode(ISD::ADD, dl,
4732 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, dl,
4733 TLI.getPointerTy()),
4735 SDValue FA = DAG.getNode(ISD::FRAMEADDR, dl,
4737 DAG.getConstant(0, TLI.getPointerTy()));
4738 setValue(&I, DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
4742 case Intrinsic::eh_sjlj_callsite: {
4743 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
4744 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(0));
4745 assert(CI && "Non-constant call site value in eh.sjlj.callsite!");
4746 assert(MMI.getCurrentCallSite() == 0 && "Overlapping call sites!");
4748 MMI.setCurrentCallSite(CI->getZExtValue());
4751 case Intrinsic::eh_sjlj_functioncontext: {
4752 // Get and store the index of the function context.
4753 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4755 cast<AllocaInst>(I.getArgOperand(0)->stripPointerCasts());
4756 int FI = FuncInfo.StaticAllocaMap[FnCtx];
4757 MFI->setFunctionContextIndex(FI);
4760 case Intrinsic::eh_sjlj_setjmp: {
4763 Ops[1] = getValue(I.getArgOperand(0));
4764 SDValue Op = DAG.getNode(ISD::EH_SJLJ_SETJMP, dl,
4765 DAG.getVTList(MVT::i32, MVT::Other),
4767 setValue(&I, Op.getValue(0));
4768 DAG.setRoot(Op.getValue(1));
4771 case Intrinsic::eh_sjlj_longjmp: {
4772 DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, dl, MVT::Other,
4773 getRoot(), getValue(I.getArgOperand(0))));
4777 case Intrinsic::x86_mmx_pslli_w:
4778 case Intrinsic::x86_mmx_pslli_d:
4779 case Intrinsic::x86_mmx_pslli_q:
4780 case Intrinsic::x86_mmx_psrli_w:
4781 case Intrinsic::x86_mmx_psrli_d:
4782 case Intrinsic::x86_mmx_psrli_q:
4783 case Intrinsic::x86_mmx_psrai_w:
4784 case Intrinsic::x86_mmx_psrai_d: {
4785 SDValue ShAmt = getValue(I.getArgOperand(1));
4786 if (isa<ConstantSDNode>(ShAmt)) {
4787 visitTargetIntrinsic(I, Intrinsic);
4790 unsigned NewIntrinsic = 0;
4791 EVT ShAmtVT = MVT::v2i32;
4792 switch (Intrinsic) {
4793 case Intrinsic::x86_mmx_pslli_w:
4794 NewIntrinsic = Intrinsic::x86_mmx_psll_w;
4796 case Intrinsic::x86_mmx_pslli_d:
4797 NewIntrinsic = Intrinsic::x86_mmx_psll_d;
4799 case Intrinsic::x86_mmx_pslli_q:
4800 NewIntrinsic = Intrinsic::x86_mmx_psll_q;
4802 case Intrinsic::x86_mmx_psrli_w:
4803 NewIntrinsic = Intrinsic::x86_mmx_psrl_w;
4805 case Intrinsic::x86_mmx_psrli_d:
4806 NewIntrinsic = Intrinsic::x86_mmx_psrl_d;
4808 case Intrinsic::x86_mmx_psrli_q:
4809 NewIntrinsic = Intrinsic::x86_mmx_psrl_q;
4811 case Intrinsic::x86_mmx_psrai_w:
4812 NewIntrinsic = Intrinsic::x86_mmx_psra_w;
4814 case Intrinsic::x86_mmx_psrai_d:
4815 NewIntrinsic = Intrinsic::x86_mmx_psra_d;
4817 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4820 // The vector shift intrinsics with scalars uses 32b shift amounts but
4821 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
4823 // We must do this early because v2i32 is not a legal type.
4824 DebugLoc dl = getCurDebugLoc();
4827 ShOps[1] = DAG.getConstant(0, MVT::i32);
4828 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
4829 EVT DestVT = TLI.getValueType(I.getType());
4830 ShAmt = DAG.getNode(ISD::BITCAST, dl, DestVT, ShAmt);
4831 Res = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4832 DAG.getConstant(NewIntrinsic, MVT::i32),
4833 getValue(I.getArgOperand(0)), ShAmt);
4837 case Intrinsic::x86_avx_vinsertf128_pd_256:
4838 case Intrinsic::x86_avx_vinsertf128_ps_256:
4839 case Intrinsic::x86_avx_vinsertf128_si_256:
4840 case Intrinsic::x86_avx2_vinserti128: {
4841 DebugLoc dl = getCurDebugLoc();
4842 EVT DestVT = TLI.getValueType(I.getType());
4843 EVT ElVT = TLI.getValueType(I.getArgOperand(1)->getType());
4844 uint64_t Idx = (cast<ConstantInt>(I.getArgOperand(2))->getZExtValue() & 1) *
4845 ElVT.getVectorNumElements();
4846 Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, DestVT,
4847 getValue(I.getArgOperand(0)),
4848 getValue(I.getArgOperand(1)),
4849 DAG.getConstant(Idx, MVT::i32));
4853 case Intrinsic::convertff:
4854 case Intrinsic::convertfsi:
4855 case Intrinsic::convertfui:
4856 case Intrinsic::convertsif:
4857 case Intrinsic::convertuif:
4858 case Intrinsic::convertss:
4859 case Intrinsic::convertsu:
4860 case Intrinsic::convertus:
4861 case Intrinsic::convertuu: {
4862 ISD::CvtCode Code = ISD::CVT_INVALID;
4863 switch (Intrinsic) {
4864 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
4865 case Intrinsic::convertff: Code = ISD::CVT_FF; break;
4866 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break;
4867 case Intrinsic::convertfui: Code = ISD::CVT_FU; break;
4868 case Intrinsic::convertsif: Code = ISD::CVT_SF; break;
4869 case Intrinsic::convertuif: Code = ISD::CVT_UF; break;
4870 case Intrinsic::convertss: Code = ISD::CVT_SS; break;
4871 case Intrinsic::convertsu: Code = ISD::CVT_SU; break;
4872 case Intrinsic::convertus: Code = ISD::CVT_US; break;
4873 case Intrinsic::convertuu: Code = ISD::CVT_UU; break;
4875 EVT DestVT = TLI.getValueType(I.getType());
4876 const Value *Op1 = I.getArgOperand(0);
4877 Res = DAG.getConvertRndSat(DestVT, getCurDebugLoc(), getValue(Op1),
4878 DAG.getValueType(DestVT),
4879 DAG.getValueType(getValue(Op1).getValueType()),
4880 getValue(I.getArgOperand(1)),
4881 getValue(I.getArgOperand(2)),
4886 case Intrinsic::sqrt:
4887 setValue(&I, DAG.getNode(ISD::FSQRT, dl,
4888 getValue(I.getArgOperand(0)).getValueType(),
4889 getValue(I.getArgOperand(0))));
4891 case Intrinsic::powi:
4892 setValue(&I, ExpandPowI(dl, getValue(I.getArgOperand(0)),
4893 getValue(I.getArgOperand(1)), DAG));
4895 case Intrinsic::sin:
4896 setValue(&I, DAG.getNode(ISD::FSIN, dl,
4897 getValue(I.getArgOperand(0)).getValueType(),
4898 getValue(I.getArgOperand(0))));
4900 case Intrinsic::cos:
4901 setValue(&I, DAG.getNode(ISD::FCOS, dl,
4902 getValue(I.getArgOperand(0)).getValueType(),
4903 getValue(I.getArgOperand(0))));
4905 case Intrinsic::log:
4908 case Intrinsic::log2:
4911 case Intrinsic::log10:
4914 case Intrinsic::exp:
4917 case Intrinsic::exp2:
4920 case Intrinsic::pow:
4923 case Intrinsic::fabs:
4924 setValue(&I, DAG.getNode(ISD::FABS, dl,
4925 getValue(I.getArgOperand(0)).getValueType(),
4926 getValue(I.getArgOperand(0))));
4928 case Intrinsic::fma:
4929 setValue(&I, DAG.getNode(ISD::FMA, dl,
4930 getValue(I.getArgOperand(0)).getValueType(),
4931 getValue(I.getArgOperand(0)),
4932 getValue(I.getArgOperand(1)),
4933 getValue(I.getArgOperand(2))));
4935 case Intrinsic::fmuladd: {
4936 EVT VT = TLI.getValueType(I.getType());
4937 if (TM.Options.AllowFPOpFusion != FPOpFusion::Strict &&
4938 TLI.isOperationLegal(ISD::FMA, VT) &&
4939 TLI.isFMAFasterThanMulAndAdd(VT)){
4940 setValue(&I, DAG.getNode(ISD::FMA, dl,
4941 getValue(I.getArgOperand(0)).getValueType(),
4942 getValue(I.getArgOperand(0)),
4943 getValue(I.getArgOperand(1)),
4944 getValue(I.getArgOperand(2))));
4946 SDValue Mul = DAG.getNode(ISD::FMUL, dl,
4947 getValue(I.getArgOperand(0)).getValueType(),
4948 getValue(I.getArgOperand(0)),
4949 getValue(I.getArgOperand(1)));
4950 SDValue Add = DAG.getNode(ISD::FADD, dl,
4951 getValue(I.getArgOperand(0)).getValueType(),
4953 getValue(I.getArgOperand(2)));
4958 case Intrinsic::convert_to_fp16:
4959 setValue(&I, DAG.getNode(ISD::FP32_TO_FP16, dl,
4960 MVT::i16, getValue(I.getArgOperand(0))));
4962 case Intrinsic::convert_from_fp16:
4963 setValue(&I, DAG.getNode(ISD::FP16_TO_FP32, dl,
4964 MVT::f32, getValue(I.getArgOperand(0))));
4966 case Intrinsic::pcmarker: {
4967 SDValue Tmp = getValue(I.getArgOperand(0));
4968 DAG.setRoot(DAG.getNode(ISD::PCMARKER, dl, MVT::Other, getRoot(), Tmp));
4971 case Intrinsic::readcyclecounter: {
4972 SDValue Op = getRoot();
4973 Res = DAG.getNode(ISD::READCYCLECOUNTER, dl,
4974 DAG.getVTList(MVT::i64, MVT::Other),
4977 DAG.setRoot(Res.getValue(1));
4980 case Intrinsic::bswap:
4981 setValue(&I, DAG.getNode(ISD::BSWAP, dl,
4982 getValue(I.getArgOperand(0)).getValueType(),
4983 getValue(I.getArgOperand(0))));
4985 case Intrinsic::cttz: {
4986 SDValue Arg = getValue(I.getArgOperand(0));
4987 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1));
4988 EVT Ty = Arg.getValueType();
4989 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTTZ : ISD::CTTZ_ZERO_UNDEF,
4993 case Intrinsic::ctlz: {
4994 SDValue Arg = getValue(I.getArgOperand(0));
4995 ConstantInt *CI = cast<ConstantInt>(I.getArgOperand(1));
4996 EVT Ty = Arg.getValueType();
4997 setValue(&I, DAG.getNode(CI->isZero() ? ISD::CTLZ : ISD::CTLZ_ZERO_UNDEF,
5001 case Intrinsic::ctpop: {
5002 SDValue Arg = getValue(I.getArgOperand(0));
5003 EVT Ty = Arg.getValueType();
5004 setValue(&I, DAG.getNode(ISD::CTPOP, dl, Ty, Arg));
5007 case Intrinsic::stacksave: {
5008 SDValue Op = getRoot();
5009 Res = DAG.getNode(ISD::STACKSAVE, dl,
5010 DAG.getVTList(TLI.getPointerTy(), MVT::Other), &Op, 1);
5012 DAG.setRoot(Res.getValue(1));
5015 case Intrinsic::stackrestore: {
5016 Res = getValue(I.getArgOperand(0));
5017 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, dl, MVT::Other, getRoot(), Res));
5020 case Intrinsic::stackprotector: {
5021 // Emit code into the DAG to store the stack guard onto the stack.
5022 MachineFunction &MF = DAG.getMachineFunction();
5023 MachineFrameInfo *MFI = MF.getFrameInfo();
5024 EVT PtrTy = TLI.getPointerTy();
5026 SDValue Src = getValue(I.getArgOperand(0)); // The guard's value.
5027 AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
5029 int FI = FuncInfo.StaticAllocaMap[Slot];
5030 MFI->setStackProtectorIndex(FI);
5032 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
5034 // Store the stack protector onto the stack.
5035 Res = DAG.getStore(getRoot(), getCurDebugLoc(), Src, FIN,
5036 MachinePointerInfo::getFixedStack(FI),
5042 case Intrinsic::objectsize: {
5043 // If we don't know by now, we're never going to know.
5044 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1));
5046 assert(CI && "Non-constant type in __builtin_object_size?");
5048 SDValue Arg = getValue(I.getCalledValue());
5049 EVT Ty = Arg.getValueType();
5052 Res = DAG.getConstant(-1ULL, Ty);
5054 Res = DAG.getConstant(0, Ty);
5059 case Intrinsic::var_annotation:
5060 // Discard annotate attributes
5063 case Intrinsic::init_trampoline: {
5064 const Function *F = cast<Function>(I.getArgOperand(1)->stripPointerCasts());
5068 Ops[1] = getValue(I.getArgOperand(0));
5069 Ops[2] = getValue(I.getArgOperand(1));
5070 Ops[3] = getValue(I.getArgOperand(2));
5071 Ops[4] = DAG.getSrcValue(I.getArgOperand(0));
5072 Ops[5] = DAG.getSrcValue(F);
5074 Res = DAG.getNode(ISD::INIT_TRAMPOLINE, dl, MVT::Other, Ops, 6);
5079 case Intrinsic::adjust_trampoline: {
5080 setValue(&I, DAG.getNode(ISD::ADJUST_TRAMPOLINE, dl,
5082 getValue(I.getArgOperand(0))));
5085 case Intrinsic::gcroot:
5087 const Value *Alloca = I.getArgOperand(0)->stripPointerCasts();
5088 const Constant *TypeMap = cast<Constant>(I.getArgOperand(1));
5090 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode());
5091 GFI->addStackRoot(FI->getIndex(), TypeMap);
5094 case Intrinsic::gcread:
5095 case Intrinsic::gcwrite:
5096 llvm_unreachable("GC failed to lower gcread/gcwrite intrinsics!");
5097 case Intrinsic::flt_rounds:
5098 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, dl, MVT::i32));
5101 case Intrinsic::expect: {
5102 // Just replace __builtin_expect(exp, c) with EXP.
5103 setValue(&I, getValue(I.getArgOperand(0)));
5107 case Intrinsic::trap: {
5108 StringRef TrapFuncName = TM.Options.getTrapFunctionName();
5109 if (TrapFuncName.empty()) {
5110 DAG.setRoot(DAG.getNode(ISD::TRAP, dl,MVT::Other, getRoot()));
5113 TargetLowering::ArgListTy Args;
5115 CallLoweringInfo CLI(getRoot(), I.getType(),
5116 false, false, false, false, 0, CallingConv::C,
5117 /*isTailCall=*/false,
5118 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
5119 DAG.getExternalSymbol(TrapFuncName.data(), TLI.getPointerTy()),
5120 Args, DAG, getCurDebugLoc());
5121 std::pair<SDValue, SDValue> Result = TLI.LowerCallTo(CLI);
5122 DAG.setRoot(Result.second);
5125 case Intrinsic::debugtrap: {
5126 DAG.setRoot(DAG.getNode(ISD::DEBUGTRAP, dl,MVT::Other, getRoot()));
5129 case Intrinsic::uadd_with_overflow:
5130 case Intrinsic::sadd_with_overflow:
5131 case Intrinsic::usub_with_overflow:
5132 case Intrinsic::ssub_with_overflow:
5133 case Intrinsic::umul_with_overflow:
5134 case Intrinsic::smul_with_overflow: {
5136 switch (Intrinsic) {
5137 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
5138 case Intrinsic::uadd_with_overflow: Op = ISD::UADDO; break;
5139 case Intrinsic::sadd_with_overflow: Op = ISD::SADDO; break;
5140 case Intrinsic::usub_with_overflow: Op = ISD::USUBO; break;
5141 case Intrinsic::ssub_with_overflow: Op = ISD::SSUBO; break;
5142 case Intrinsic::umul_with_overflow: Op = ISD::UMULO; break;
5143 case Intrinsic::smul_with_overflow: Op = ISD::SMULO; break;
5145 SDValue Op1 = getValue(I.getArgOperand(0));
5146 SDValue Op2 = getValue(I.getArgOperand(1));
5148 SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1);
5149 setValue(&I, DAG.getNode(Op, getCurDebugLoc(), VTs, Op1, Op2));
5152 case Intrinsic::prefetch: {
5154 unsigned rw = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
5156 Ops[1] = getValue(I.getArgOperand(0));
5157 Ops[2] = getValue(I.getArgOperand(1));
5158 Ops[3] = getValue(I.getArgOperand(2));
5159 Ops[4] = getValue(I.getArgOperand(3));
5160 DAG.setRoot(DAG.getMemIntrinsicNode(ISD::PREFETCH, dl,
5161 DAG.getVTList(MVT::Other),
5163 EVT::getIntegerVT(*Context, 8),
5164 MachinePointerInfo(I.getArgOperand(0)),
5166 false, /* volatile */
5168 rw==1)); /* write */
5172 case Intrinsic::invariant_start:
5173 case Intrinsic::lifetime_start:
5174 // Discard region information.
5175 setValue(&I, DAG.getUNDEF(TLI.getPointerTy()));
5177 case Intrinsic::invariant_end:
5178 case Intrinsic::lifetime_end:
5179 // Discard region information.
5184 void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee,
5186 MachineBasicBlock *LandingPad) {
5187 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
5188 FunctionType *FTy = cast<FunctionType>(PT->getElementType());
5189 Type *RetTy = FTy->getReturnType();
5190 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5191 MCSymbol *BeginLabel = 0;
5193 TargetLowering::ArgListTy Args;
5194 TargetLowering::ArgListEntry Entry;
5195 Args.reserve(CS.arg_size());
5197 // Check whether the function can return without sret-demotion.
5198 SmallVector<ISD::OutputArg, 4> Outs;
5199 GetReturnInfo(RetTy, CS.getAttributes().getRetAttributes(),
5202 bool CanLowerReturn = TLI.CanLowerReturn(CS.getCallingConv(),
5203 DAG.getMachineFunction(),
5204 FTy->isVarArg(), Outs,
5207 SDValue DemoteStackSlot;
5208 int DemoteStackIdx = -100;
5210 if (!CanLowerReturn) {
5211 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(
5212 FTy->getReturnType());
5213 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(
5214 FTy->getReturnType());
5215 MachineFunction &MF = DAG.getMachineFunction();
5216 DemoteStackIdx = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
5217 Type *StackSlotPtrType = PointerType::getUnqual(FTy->getReturnType());
5219 DemoteStackSlot = DAG.getFrameIndex(DemoteStackIdx, TLI.getPointerTy());
5220 Entry.Node = DemoteStackSlot;
5221 Entry.Ty = StackSlotPtrType;
5222 Entry.isSExt = false;
5223 Entry.isZExt = false;
5224 Entry.isInReg = false;
5225 Entry.isSRet = true;
5226 Entry.isNest = false;
5227 Entry.isByVal = false;
5228 Entry.Alignment = Align;
5229 Args.push_back(Entry);
5230 RetTy = Type::getVoidTy(FTy->getContext());
5233 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
5235 const Value *V = *i;
5238 if (V->getType()->isEmptyTy())
5241 SDValue ArgNode = getValue(V);
5242 Entry.Node = ArgNode; Entry.Ty = V->getType();
5244 unsigned attrInd = i - CS.arg_begin() + 1;
5245 Entry.isSExt = CS.paramHasAttr(attrInd, Attribute::SExt);
5246 Entry.isZExt = CS.paramHasAttr(attrInd, Attribute::ZExt);
5247 Entry.isInReg = CS.paramHasAttr(attrInd, Attribute::InReg);
5248 Entry.isSRet = CS.paramHasAttr(attrInd, Attribute::StructRet);
5249 Entry.isNest = CS.paramHasAttr(attrInd, Attribute::Nest);
5250 Entry.isByVal = CS.paramHasAttr(attrInd, Attribute::ByVal);
5251 Entry.Alignment = CS.getParamAlignment(attrInd);
5252 Args.push_back(Entry);
5256 // Insert a label before the invoke call to mark the try range. This can be
5257 // used to detect deletion of the invoke via the MachineModuleInfo.
5258 BeginLabel = MMI.getContext().CreateTempSymbol();
5260 // For SjLj, keep track of which landing pads go with which invokes
5261 // so as to maintain the ordering of pads in the LSDA.
5262 unsigned CallSiteIndex = MMI.getCurrentCallSite();
5263 if (CallSiteIndex) {
5264 MMI.setCallSiteBeginLabel(BeginLabel, CallSiteIndex);
5265 LPadToCallSiteMap[LandingPad].push_back(CallSiteIndex);
5267 // Now that the call site is handled, stop tracking it.
5268 MMI.setCurrentCallSite(0);
5271 // Both PendingLoads and PendingExports must be flushed here;
5272 // this call might not return.
5274 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getControlRoot(), BeginLabel));
5277 // Check if target-independent constraints permit a tail call here.
5278 // Target-dependent constraints are checked within TLI.LowerCallTo.
5280 !isInTailCallPosition(CS, CS.getAttributes().getRetAttributes(), TLI))
5283 // If there's a possibility that fast-isel has already selected some amount
5284 // of the current basic block, don't emit a tail call.
5285 if (isTailCall && TM.Options.EnableFastISel)
5289 CallLoweringInfo CLI(getRoot(), RetTy, FTy, isTailCall, Callee, Args, DAG,
5290 getCurDebugLoc(), CS);
5291 std::pair<SDValue,SDValue> Result = TLI.LowerCallTo(CLI);
5292 assert((isTailCall || Result.second.getNode()) &&
5293 "Non-null chain expected with non-tail call!");
5294 assert((Result.second.getNode() || !Result.first.getNode()) &&
5295 "Null value expected with tail call!");
5296 if (Result.first.getNode()) {
5297 setValue(CS.getInstruction(), Result.first);
5298 } else if (!CanLowerReturn && Result.second.getNode()) {
5299 // The instruction result is the result of loading from the
5300 // hidden sret parameter.
5301 SmallVector<EVT, 1> PVTs;
5302 Type *PtrRetTy = PointerType::getUnqual(FTy->getReturnType());
5304 ComputeValueVTs(TLI, PtrRetTy, PVTs);
5305 assert(PVTs.size() == 1 && "Pointers should fit in one register");
5306 EVT PtrVT = PVTs[0];
5308 SmallVector<EVT, 4> RetTys;
5309 SmallVector<uint64_t, 4> Offsets;
5310 RetTy = FTy->getReturnType();
5311 ComputeValueVTs(TLI, RetTy, RetTys, &Offsets);
5313 unsigned NumValues = RetTys.size();
5314 SmallVector<SDValue, 4> Values(NumValues);
5315 SmallVector<SDValue, 4> Chains(NumValues);
5317 for (unsigned i = 0; i < NumValues; ++i) {
5318 SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT,
5320 DAG.getConstant(Offsets[i], PtrVT));
5321 SDValue L = DAG.getLoad(RetTys[i], getCurDebugLoc(), Result.second, Add,
5322 MachinePointerInfo::getFixedStack(DemoteStackIdx, Offsets[i]),
5323 false, false, false, 1);
5325 Chains[i] = L.getValue(1);
5328 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
5329 MVT::Other, &Chains[0], NumValues);
5330 PendingLoads.push_back(Chain);
5332 setValue(CS.getInstruction(),
5333 DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
5334 DAG.getVTList(&RetTys[0], RetTys.size()),
5335 &Values[0], Values.size()));
5338 // Assign order to nodes here. If the call does not produce a result, it won't
5339 // be mapped to a SDNode and visit() will not assign it an order number.
5340 if (!Result.second.getNode()) {
5341 // As a special case, a null chain means that a tail call has been emitted and
5342 // the DAG root is already updated.
5345 AssignOrderingToNode(DAG.getRoot().getNode());
5347 DAG.setRoot(Result.second);
5349 AssignOrderingToNode(Result.second.getNode());
5353 // Insert a label at the end of the invoke call to mark the try range. This
5354 // can be used to detect deletion of the invoke via the MachineModuleInfo.
5355 MCSymbol *EndLabel = MMI.getContext().CreateTempSymbol();
5356 DAG.setRoot(DAG.getEHLabel(getCurDebugLoc(), getRoot(), EndLabel));
5358 // Inform MachineModuleInfo of range.
5359 MMI.addInvoke(LandingPad, BeginLabel, EndLabel);
5363 /// IsOnlyUsedInZeroEqualityComparison - Return true if it only matters that the
5364 /// value is equal or not-equal to zero.
5365 static bool IsOnlyUsedInZeroEqualityComparison(const Value *V) {
5366 for (Value::const_use_iterator UI = V->use_begin(), E = V->use_end();
5368 if (const ICmpInst *IC = dyn_cast<ICmpInst>(*UI))
5369 if (IC->isEquality())
5370 if (const Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
5371 if (C->isNullValue())
5373 // Unknown instruction.
5379 static SDValue getMemCmpLoad(const Value *PtrVal, MVT LoadVT,
5381 SelectionDAGBuilder &Builder) {
5383 // Check to see if this load can be trivially constant folded, e.g. if the
5384 // input is from a string literal.
5385 if (const Constant *LoadInput = dyn_cast<Constant>(PtrVal)) {
5386 // Cast pointer to the type we really want to load.
5387 LoadInput = ConstantExpr::getBitCast(const_cast<Constant *>(LoadInput),
5388 PointerType::getUnqual(LoadTy));
5390 if (const Constant *LoadCst =
5391 ConstantFoldLoadFromConstPtr(const_cast<Constant *>(LoadInput),
5393 return Builder.getValue(LoadCst);
5396 // Otherwise, we have to emit the load. If the pointer is to unfoldable but
5397 // still constant memory, the input chain can be the entry node.
5399 bool ConstantMemory = false;
5401 // Do not serialize (non-volatile) loads of constant memory with anything.
5402 if (Builder.AA->pointsToConstantMemory(PtrVal)) {
5403 Root = Builder.DAG.getEntryNode();
5404 ConstantMemory = true;
5406 // Do not serialize non-volatile loads against each other.
5407 Root = Builder.DAG.getRoot();
5410 SDValue Ptr = Builder.getValue(PtrVal);
5411 SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurDebugLoc(), Root,
5412 Ptr, MachinePointerInfo(PtrVal),
5414 false /*nontemporal*/,
5415 false /*isinvariant*/, 1 /* align=1 */);
5417 if (!ConstantMemory)
5418 Builder.PendingLoads.push_back(LoadVal.getValue(1));
5423 /// visitMemCmpCall - See if we can lower a call to memcmp in an optimized form.
5424 /// If so, return true and lower it, otherwise return false and it will be
5425 /// lowered like a normal call.
5426 bool SelectionDAGBuilder::visitMemCmpCall(const CallInst &I) {
5427 // Verify that the prototype makes sense. int memcmp(void*,void*,size_t)
5428 if (I.getNumArgOperands() != 3)
5431 const Value *LHS = I.getArgOperand(0), *RHS = I.getArgOperand(1);
5432 if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() ||
5433 !I.getArgOperand(2)->getType()->isIntegerTy() ||
5434 !I.getType()->isIntegerTy())
5437 const ConstantInt *Size = dyn_cast<ConstantInt>(I.getArgOperand(2));
5439 // memcmp(S1,S2,2) != 0 -> (*(short*)LHS != *(short*)RHS) != 0
5440 // memcmp(S1,S2,4) != 0 -> (*(int*)LHS != *(int*)RHS) != 0
5441 if (Size && IsOnlyUsedInZeroEqualityComparison(&I)) {
5442 bool ActuallyDoIt = true;
5445 switch (Size->getZExtValue()) {
5447 LoadVT = MVT::Other;
5449 ActuallyDoIt = false;
5453 LoadTy = Type::getInt16Ty(Size->getContext());
5457 LoadTy = Type::getInt32Ty(Size->getContext());
5461 LoadTy = Type::getInt64Ty(Size->getContext());
5465 LoadVT = MVT::v4i32;
5466 LoadTy = Type::getInt32Ty(Size->getContext());
5467 LoadTy = VectorType::get(LoadTy, 4);
5472 // This turns into unaligned loads. We only do this if the target natively
5473 // supports the MVT we'll be loading or if it is small enough (<= 4) that
5474 // we'll only produce a small number of byte loads.
5476 // Require that we can find a legal MVT, and only do this if the target
5477 // supports unaligned loads of that type. Expanding into byte loads would
5479 if (ActuallyDoIt && Size->getZExtValue() > 4) {
5480 // TODO: Handle 5 byte compare as 4-byte + 1 byte.
5481 // TODO: Handle 8 byte compare on x86-32 as two 32-bit loads.
5482 if (!TLI.isTypeLegal(LoadVT) ||!TLI.allowsUnalignedMemoryAccesses(LoadVT))
5483 ActuallyDoIt = false;
5487 SDValue LHSVal = getMemCmpLoad(LHS, LoadVT, LoadTy, *this);
5488 SDValue RHSVal = getMemCmpLoad(RHS, LoadVT, LoadTy, *this);
5490 SDValue Res = DAG.getSetCC(getCurDebugLoc(), MVT::i1, LHSVal, RHSVal,
5492 EVT CallVT = TLI.getValueType(I.getType(), true);
5493 setValue(&I, DAG.getZExtOrTrunc(Res, getCurDebugLoc(), CallVT));
5503 void SelectionDAGBuilder::visitCall(const CallInst &I) {
5504 // Handle inline assembly differently.
5505 if (isa<InlineAsm>(I.getCalledValue())) {
5510 MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
5511 ComputeUsesVAFloatArgument(I, &MMI);
5513 const char *RenameFn = 0;
5514 if (Function *F = I.getCalledFunction()) {
5515 if (F->isDeclaration()) {
5516 if (const TargetIntrinsicInfo *II = TM.getIntrinsicInfo()) {
5517 if (unsigned IID = II->getIntrinsicID(F)) {
5518 RenameFn = visitIntrinsicCall(I, IID);
5523 if (unsigned IID = F->getIntrinsicID()) {
5524 RenameFn = visitIntrinsicCall(I, IID);
5530 // Check for well-known libc/libm calls. If the function is internal, it
5531 // can't be a library call.
5532 if (!F->hasLocalLinkage() && F->hasName()) {
5533 StringRef Name = F->getName();
5534 if ((LibInfo->has(LibFunc::copysign) && Name == "copysign") ||
5535 (LibInfo->has(LibFunc::copysignf) && Name == "copysignf") ||
5536 (LibInfo->has(LibFunc::copysignl) && Name == "copysignl")) {
5537 if (I.getNumArgOperands() == 2 && // Basic sanity checks.
5538 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5539 I.getType() == I.getArgOperand(0)->getType() &&
5540 I.getType() == I.getArgOperand(1)->getType()) {
5541 SDValue LHS = getValue(I.getArgOperand(0));
5542 SDValue RHS = getValue(I.getArgOperand(1));
5543 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurDebugLoc(),
5544 LHS.getValueType(), LHS, RHS));
5547 } else if ((LibInfo->has(LibFunc::fabs) && Name == "fabs") ||
5548 (LibInfo->has(LibFunc::fabsf) && Name == "fabsf") ||
5549 (LibInfo->has(LibFunc::fabsl) && Name == "fabsl")) {
5550 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5551 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5552 I.getType() == I.getArgOperand(0)->getType()) {
5553 SDValue Tmp = getValue(I.getArgOperand(0));
5554 setValue(&I, DAG.getNode(ISD::FABS, getCurDebugLoc(),
5555 Tmp.getValueType(), Tmp));
5558 } else if ((LibInfo->has(LibFunc::sin) && Name == "sin") ||
5559 (LibInfo->has(LibFunc::sinf) && Name == "sinf") ||
5560 (LibInfo->has(LibFunc::sinl) && Name == "sinl")) {
5561 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5562 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5563 I.getType() == I.getArgOperand(0)->getType() &&
5564 I.onlyReadsMemory()) {
5565 SDValue Tmp = getValue(I.getArgOperand(0));
5566 setValue(&I, DAG.getNode(ISD::FSIN, getCurDebugLoc(),
5567 Tmp.getValueType(), Tmp));
5570 } else if ((LibInfo->has(LibFunc::cos) && Name == "cos") ||
5571 (LibInfo->has(LibFunc::cosf) && Name == "cosf") ||
5572 (LibInfo->has(LibFunc::cosl) && Name == "cosl")) {
5573 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5574 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5575 I.getType() == I.getArgOperand(0)->getType() &&
5576 I.onlyReadsMemory()) {
5577 SDValue Tmp = getValue(I.getArgOperand(0));
5578 setValue(&I, DAG.getNode(ISD::FCOS, getCurDebugLoc(),
5579 Tmp.getValueType(), Tmp));
5582 } else if ((LibInfo->has(LibFunc::sqrt) && Name == "sqrt") ||
5583 (LibInfo->has(LibFunc::sqrtf) && Name == "sqrtf") ||
5584 (LibInfo->has(LibFunc::sqrtl) && Name == "sqrtl")) {
5585 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5586 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5587 I.getType() == I.getArgOperand(0)->getType() &&
5588 I.onlyReadsMemory()) {
5589 SDValue Tmp = getValue(I.getArgOperand(0));
5590 setValue(&I, DAG.getNode(ISD::FSQRT, getCurDebugLoc(),
5591 Tmp.getValueType(), Tmp));
5594 } else if ((LibInfo->has(LibFunc::floor) && Name == "floor") ||
5595 (LibInfo->has(LibFunc::floorf) && Name == "floorf") ||
5596 (LibInfo->has(LibFunc::floorl) && Name == "floorl")) {
5597 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5598 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5599 I.getType() == I.getArgOperand(0)->getType()) {
5600 SDValue Tmp = getValue(I.getArgOperand(0));
5601 setValue(&I, DAG.getNode(ISD::FFLOOR, getCurDebugLoc(),
5602 Tmp.getValueType(), Tmp));
5605 } else if ((LibInfo->has(LibFunc::nearbyint) && Name == "nearbyint") ||
5606 (LibInfo->has(LibFunc::nearbyintf) && Name == "nearbyintf") ||
5607 (LibInfo->has(LibFunc::nearbyintl) && Name == "nearbyintl")) {
5608 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5609 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5610 I.getType() == I.getArgOperand(0)->getType()) {
5611 SDValue Tmp = getValue(I.getArgOperand(0));
5612 setValue(&I, DAG.getNode(ISD::FNEARBYINT, getCurDebugLoc(),
5613 Tmp.getValueType(), Tmp));
5616 } else if ((LibInfo->has(LibFunc::ceil) && Name == "ceil") ||
5617 (LibInfo->has(LibFunc::ceilf) && Name == "ceilf") ||
5618 (LibInfo->has(LibFunc::ceill) && Name == "ceill")) {
5619 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5620 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5621 I.getType() == I.getArgOperand(0)->getType()) {
5622 SDValue Tmp = getValue(I.getArgOperand(0));
5623 setValue(&I, DAG.getNode(ISD::FCEIL, getCurDebugLoc(),
5624 Tmp.getValueType(), Tmp));
5627 } else if ((LibInfo->has(LibFunc::rint) && Name == "rint") ||
5628 (LibInfo->has(LibFunc::rintf) && Name == "rintf") ||
5629 (LibInfo->has(LibFunc::rintl) && Name == "rintl")) {
5630 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5631 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5632 I.getType() == I.getArgOperand(0)->getType()) {
5633 SDValue Tmp = getValue(I.getArgOperand(0));
5634 setValue(&I, DAG.getNode(ISD::FRINT, getCurDebugLoc(),
5635 Tmp.getValueType(), Tmp));
5638 } else if ((LibInfo->has(LibFunc::trunc) && Name == "trunc") ||
5639 (LibInfo->has(LibFunc::truncf) && Name == "truncf") ||
5640 (LibInfo->has(LibFunc::truncl) && Name == "truncl")) {
5641 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5642 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5643 I.getType() == I.getArgOperand(0)->getType()) {
5644 SDValue Tmp = getValue(I.getArgOperand(0));
5645 setValue(&I, DAG.getNode(ISD::FTRUNC, getCurDebugLoc(),
5646 Tmp.getValueType(), Tmp));
5649 } else if ((LibInfo->has(LibFunc::log2) && Name == "log2") ||
5650 (LibInfo->has(LibFunc::log2f) && Name == "log2f") ||
5651 (LibInfo->has(LibFunc::log2l) && Name == "log2l")) {
5652 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5653 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5654 I.getType() == I.getArgOperand(0)->getType() &&
5655 I.onlyReadsMemory()) {
5656 SDValue Tmp = getValue(I.getArgOperand(0));
5657 setValue(&I, DAG.getNode(ISD::FLOG2, getCurDebugLoc(),
5658 Tmp.getValueType(), Tmp));
5661 } else if ((LibInfo->has(LibFunc::exp2) && Name == "exp2") ||
5662 (LibInfo->has(LibFunc::exp2f) && Name == "exp2f") ||
5663 (LibInfo->has(LibFunc::exp2l) && Name == "exp2l")) {
5664 if (I.getNumArgOperands() == 1 && // Basic sanity checks.
5665 I.getArgOperand(0)->getType()->isFloatingPointTy() &&
5666 I.getType() == I.getArgOperand(0)->getType() &&
5667 I.onlyReadsMemory()) {
5668 SDValue Tmp = getValue(I.getArgOperand(0));
5669 setValue(&I, DAG.getNode(ISD::FEXP2, getCurDebugLoc(),
5670 Tmp.getValueType(), Tmp));
5673 } else if (Name == "memcmp") {
5674 if (visitMemCmpCall(I))
5682 Callee = getValue(I.getCalledValue());
5684 Callee = DAG.getExternalSymbol(RenameFn, TLI.getPointerTy());
5686 // Check if we can potentially perform a tail call. More detailed checking is
5687 // be done within LowerCallTo, after more information about the call is known.
5688 LowerCallTo(&I, Callee, I.isTailCall());
5693 /// AsmOperandInfo - This contains information for each constraint that we are
5695 class SDISelAsmOperandInfo : public TargetLowering::AsmOperandInfo {
5697 /// CallOperand - If this is the result output operand or a clobber
5698 /// this is null, otherwise it is the incoming operand to the CallInst.
5699 /// This gets modified as the asm is processed.
5700 SDValue CallOperand;
5702 /// AssignedRegs - If this is a register or register class operand, this
5703 /// contains the set of register corresponding to the operand.
5704 RegsForValue AssignedRegs;
5706 explicit SDISelAsmOperandInfo(const TargetLowering::AsmOperandInfo &info)
5707 : TargetLowering::AsmOperandInfo(info), CallOperand(0,0) {
5710 /// getCallOperandValEVT - Return the EVT of the Value* that this operand
5711 /// corresponds to. If there is no Value* for this operand, it returns
5713 EVT getCallOperandValEVT(LLVMContext &Context,
5714 const TargetLowering &TLI,
5715 const TargetData *TD) const {
5716 if (CallOperandVal == 0) return MVT::Other;
5718 if (isa<BasicBlock>(CallOperandVal))
5719 return TLI.getPointerTy();
5721 llvm::Type *OpTy = CallOperandVal->getType();
5723 // FIXME: code duplicated from TargetLowering::ParseConstraints().
5724 // If this is an indirect operand, the operand is a pointer to the
5727 llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
5729 report_fatal_error("Indirect operand for inline asm not a pointer!");
5730 OpTy = PtrTy->getElementType();
5733 // Look for vector wrapped in a struct. e.g. { <16 x i8> }.
5734 if (StructType *STy = dyn_cast<StructType>(OpTy))
5735 if (STy->getNumElements() == 1)
5736 OpTy = STy->getElementType(0);
5738 // If OpTy is not a single value, it may be a struct/union that we
5739 // can tile with integers.
5740 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
5741 unsigned BitSize = TD->getTypeSizeInBits(OpTy);
5750 OpTy = IntegerType::get(Context, BitSize);
5755 return TLI.getValueType(OpTy, true);
5759 typedef SmallVector<SDISelAsmOperandInfo,16> SDISelAsmOperandInfoVector;
5761 } // end anonymous namespace
5763 /// GetRegistersForValue - Assign registers (virtual or physical) for the
5764 /// specified operand. We prefer to assign virtual registers, to allow the
5765 /// register allocator to handle the assignment process. However, if the asm
5766 /// uses features that we can't model on machineinstrs, we have SDISel do the
5767 /// allocation. This produces generally horrible, but correct, code.
5769 /// OpInfo describes the operand.
5771 static void GetRegistersForValue(SelectionDAG &DAG,
5772 const TargetLowering &TLI,
5774 SDISelAsmOperandInfo &OpInfo) {
5775 LLVMContext &Context = *DAG.getContext();
5777 MachineFunction &MF = DAG.getMachineFunction();
5778 SmallVector<unsigned, 4> Regs;
5780 // If this is a constraint for a single physreg, or a constraint for a
5781 // register class, find it.
5782 std::pair<unsigned, const TargetRegisterClass*> PhysReg =
5783 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
5784 OpInfo.ConstraintVT);
5786 unsigned NumRegs = 1;
5787 if (OpInfo.ConstraintVT != MVT::Other) {
5788 // If this is a FP input in an integer register (or visa versa) insert a bit
5789 // cast of the input value. More generally, handle any case where the input
5790 // value disagrees with the register class we plan to stick this in.
5791 if (OpInfo.Type == InlineAsm::isInput &&
5792 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) {
5793 // Try to convert to the first EVT that the reg class contains. If the
5794 // types are identical size, use a bitcast to convert (e.g. two differing
5796 EVT RegVT = *PhysReg.second->vt_begin();
5797 if (RegVT.getSizeInBits() == OpInfo.ConstraintVT.getSizeInBits()) {
5798 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
5799 RegVT, OpInfo.CallOperand);
5800 OpInfo.ConstraintVT = RegVT;
5801 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) {
5802 // If the input is a FP value and we want it in FP registers, do a
5803 // bitcast to the corresponding integer type. This turns an f64 value
5804 // into i64, which can be passed with two i32 values on a 32-bit
5806 RegVT = EVT::getIntegerVT(Context,
5807 OpInfo.ConstraintVT.getSizeInBits());
5808 OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, DL,
5809 RegVT, OpInfo.CallOperand);
5810 OpInfo.ConstraintVT = RegVT;
5814 NumRegs = TLI.getNumRegisters(Context, OpInfo.ConstraintVT);
5818 EVT ValueVT = OpInfo.ConstraintVT;
5820 // If this is a constraint for a specific physical register, like {r17},
5822 if (unsigned AssignedReg = PhysReg.first) {
5823 const TargetRegisterClass *RC = PhysReg.second;
5824 if (OpInfo.ConstraintVT == MVT::Other)
5825 ValueVT = *RC->vt_begin();
5827 // Get the actual register value type. This is important, because the user
5828 // may have asked for (e.g.) the AX register in i32 type. We need to
5829 // remember that AX is actually i16 to get the right extension.
5830 RegVT = *RC->vt_begin();
5832 // This is a explicit reference to a physical register.
5833 Regs.push_back(AssignedReg);
5835 // If this is an expanded reference, add the rest of the regs to Regs.
5837 TargetRegisterClass::iterator I = RC->begin();
5838 for (; *I != AssignedReg; ++I)
5839 assert(I != RC->end() && "Didn't find reg!");
5841 // Already added the first reg.
5843 for (; NumRegs; --NumRegs, ++I) {
5844 assert(I != RC->end() && "Ran out of registers to allocate!");
5849 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5853 // Otherwise, if this was a reference to an LLVM register class, create vregs
5854 // for this reference.
5855 if (const TargetRegisterClass *RC = PhysReg.second) {
5856 RegVT = *RC->vt_begin();
5857 if (OpInfo.ConstraintVT == MVT::Other)
5860 // Create the appropriate number of virtual registers.
5861 MachineRegisterInfo &RegInfo = MF.getRegInfo();
5862 for (; NumRegs; --NumRegs)
5863 Regs.push_back(RegInfo.createVirtualRegister(RC));
5865 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
5869 // Otherwise, we couldn't allocate enough registers for this.
5872 /// visitInlineAsm - Handle a call to an InlineAsm object.
5874 void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) {
5875 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
5877 /// ConstraintOperands - Information about all of the constraints.
5878 SDISelAsmOperandInfoVector ConstraintOperands;
5880 TargetLowering::AsmOperandInfoVector
5881 TargetConstraints = TLI.ParseConstraints(CS);
5883 bool hasMemory = false;
5885 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
5886 unsigned ResNo = 0; // ResNo - The result number of the next output.
5887 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
5888 ConstraintOperands.push_back(SDISelAsmOperandInfo(TargetConstraints[i]));
5889 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
5891 EVT OpVT = MVT::Other;
5893 // Compute the value type for each operand.
5894 switch (OpInfo.Type) {
5895 case InlineAsm::isOutput:
5896 // Indirect outputs just consume an argument.
5897 if (OpInfo.isIndirect) {
5898 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5902 // The return value of the call is this value. As such, there is no
5903 // corresponding argument.
5904 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
5905 if (StructType *STy = dyn_cast<StructType>(CS.getType())) {
5906 OpVT = TLI.getValueType(STy->getElementType(ResNo));
5908 assert(ResNo == 0 && "Asm only has one result!");
5909 OpVT = TLI.getValueType(CS.getType());
5913 case InlineAsm::isInput:
5914 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
5916 case InlineAsm::isClobber:
5921 // If this is an input or an indirect output, process the call argument.
5922 // BasicBlocks are labels, currently appearing only in asm's.
5923 if (OpInfo.CallOperandVal) {
5924 if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) {
5925 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]);
5927 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
5930 OpVT = OpInfo.getCallOperandValEVT(*DAG.getContext(), TLI, TD);
5933 OpInfo.ConstraintVT = OpVT;
5935 // Indirect operand accesses access memory.
5936 if (OpInfo.isIndirect)
5939 for (unsigned j = 0, ee = OpInfo.Codes.size(); j != ee; ++j) {
5940 TargetLowering::ConstraintType
5941 CType = TLI.getConstraintType(OpInfo.Codes[j]);
5942 if (CType == TargetLowering::C_Memory) {
5950 SDValue Chain, Flag;
5952 // We won't need to flush pending loads if this asm doesn't touch
5953 // memory and is nonvolatile.
5954 if (hasMemory || IA->hasSideEffects())
5957 Chain = DAG.getRoot();
5959 // Second pass over the constraints: compute which constraint option to use
5960 // and assign registers to constraints that want a specific physreg.
5961 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5962 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5964 // If this is an output operand with a matching input operand, look up the
5965 // matching input. If their types mismatch, e.g. one is an integer, the
5966 // other is floating point, or their sizes are different, flag it as an
5968 if (OpInfo.hasMatchingInput()) {
5969 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
5971 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
5972 std::pair<unsigned, const TargetRegisterClass*> MatchRC =
5973 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
5974 OpInfo.ConstraintVT);
5975 std::pair<unsigned, const TargetRegisterClass*> InputRC =
5976 TLI.getRegForInlineAsmConstraint(Input.ConstraintCode,
5977 Input.ConstraintVT);
5978 if ((OpInfo.ConstraintVT.isInteger() !=
5979 Input.ConstraintVT.isInteger()) ||
5980 (MatchRC.second != InputRC.second)) {
5981 report_fatal_error("Unsupported asm: input constraint"
5982 " with a matching output constraint of"
5983 " incompatible type!");
5985 Input.ConstraintVT = OpInfo.ConstraintVT;
5989 // Compute the constraint code and ConstraintType to use.
5990 TLI.ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, &DAG);
5992 // If this is a memory input, and if the operand is not indirect, do what we
5993 // need to to provide an address for the memory input.
5994 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5995 !OpInfo.isIndirect) {
5996 assert((OpInfo.isMultipleAlternative ||
5997 (OpInfo.Type == InlineAsm::isInput)) &&
5998 "Can only indirectify direct input operands!");
6000 // Memory operands really want the address of the value. If we don't have
6001 // an indirect input, put it in the constpool if we can, otherwise spill
6002 // it to a stack slot.
6003 // TODO: This isn't quite right. We need to handle these according to
6004 // the addressing mode that the constraint wants. Also, this may take
6005 // an additional register for the computation and we don't want that
6008 // If the operand is a float, integer, or vector constant, spill to a
6009 // constant pool entry to get its address.
6010 const Value *OpVal = OpInfo.CallOperandVal;
6011 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
6012 isa<ConstantVector>(OpVal) || isa<ConstantDataVector>(OpVal)) {
6013 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal),
6014 TLI.getPointerTy());
6016 // Otherwise, create a stack slot and emit a store to it before the
6018 Type *Ty = OpVal->getType();
6019 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
6020 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(Ty);
6021 MachineFunction &MF = DAG.getMachineFunction();
6022 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
6023 SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
6024 Chain = DAG.getStore(Chain, getCurDebugLoc(),
6025 OpInfo.CallOperand, StackSlot,
6026 MachinePointerInfo::getFixedStack(SSFI),
6028 OpInfo.CallOperand = StackSlot;
6031 // There is no longer a Value* corresponding to this operand.
6032 OpInfo.CallOperandVal = 0;
6034 // It is now an indirect operand.
6035 OpInfo.isIndirect = true;
6038 // If this constraint is for a specific register, allocate it before
6040 if (OpInfo.ConstraintType == TargetLowering::C_Register)
6041 GetRegistersForValue(DAG, TLI, getCurDebugLoc(), OpInfo);
6044 // Second pass - Loop over all of the operands, assigning virtual or physregs
6045 // to register class operands.
6046 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6047 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6049 // C_Register operands have already been allocated, Other/Memory don't need
6051 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
6052 GetRegistersForValue(DAG, TLI, getCurDebugLoc(), OpInfo);
6055 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
6056 std::vector<SDValue> AsmNodeOperands;
6057 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain
6058 AsmNodeOperands.push_back(
6059 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(),
6060 TLI.getPointerTy()));
6062 // If we have a !srcloc metadata node associated with it, we want to attach
6063 // this to the ultimately generated inline asm machineinstr. To do this, we
6064 // pass in the third operand as this (potentially null) inline asm MDNode.
6065 const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc");
6066 AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc));
6068 // Remember the HasSideEffect and AlignStack bits as operand 3.
6069 unsigned ExtraInfo = 0;
6070 if (IA->hasSideEffects())
6071 ExtraInfo |= InlineAsm::Extra_HasSideEffects;
6072 if (IA->isAlignStack())
6073 ExtraInfo |= InlineAsm::Extra_IsAlignStack;
6074 AsmNodeOperands.push_back(DAG.getTargetConstant(ExtraInfo,
6075 TLI.getPointerTy()));
6077 // Loop over all of the inputs, copying the operand values into the
6078 // appropriate registers and processing the output regs.
6079 RegsForValue RetValRegs;
6081 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
6082 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
6084 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
6085 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
6087 switch (OpInfo.Type) {
6088 case InlineAsm::isOutput: {
6089 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
6090 OpInfo.ConstraintType != TargetLowering::C_Register) {
6091 // Memory output, or 'other' output (e.g. 'X' constraint).
6092 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
6094 // Add information to the INLINEASM node to know about this output.
6095 unsigned OpFlags = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
6096 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlags,
6097 TLI.getPointerTy()));
6098 AsmNodeOperands.push_back(OpInfo.CallOperand);
6102 // Otherwise, this is a register or register class output.
6104 // Copy the output from the appropriate register. Find a register that
6106 if (OpInfo.AssignedRegs.Regs.empty()) {
6107 LLVMContext &Ctx = *DAG.getContext();
6108 Ctx.emitError(CS.getInstruction(),
6109 "couldn't allocate output register for constraint '" +
6110 Twine(OpInfo.ConstraintCode) + "'");
6114 // If this is an indirect operand, store through the pointer after the
6116 if (OpInfo.isIndirect) {
6117 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
6118 OpInfo.CallOperandVal));
6120 // This is the result value of the call.
6121 assert(!CS.getType()->isVoidTy() && "Bad inline asm!");
6122 // Concatenate this output onto the outputs list.
6123 RetValRegs.append(OpInfo.AssignedRegs);
6126 // Add information to the INLINEASM node to know that this register is
6128 OpInfo.AssignedRegs.AddInlineAsmOperands(OpInfo.isEarlyClobber ?
6129 InlineAsm::Kind_RegDefEarlyClobber :
6130 InlineAsm::Kind_RegDef,
6137 case InlineAsm::isInput: {
6138 SDValue InOperandVal = OpInfo.CallOperand;
6140 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint?
6141 // If this is required to match an output register we have already set,
6142 // just use its register.
6143 unsigned OperandNo = OpInfo.getMatchedOperand();
6145 // Scan until we find the definition we already emitted of this operand.
6146 // When we find it, create a RegsForValue operand.
6147 unsigned CurOp = InlineAsm::Op_FirstOperand;
6148 for (; OperandNo; --OperandNo) {
6149 // Advance to the next operand.
6151 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6152 assert((InlineAsm::isRegDefKind(OpFlag) ||
6153 InlineAsm::isRegDefEarlyClobberKind(OpFlag) ||
6154 InlineAsm::isMemKind(OpFlag)) && "Skipped past definitions?");
6155 CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1;
6159 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
6160 if (InlineAsm::isRegDefKind(OpFlag) ||
6161 InlineAsm::isRegDefEarlyClobberKind(OpFlag)) {
6162 // Add (OpFlag&0xffff)>>3 registers to MatchedRegs.
6163 if (OpInfo.isIndirect) {
6164 // This happens on gcc/testsuite/gcc.dg/pr8788-1.c
6165 LLVMContext &Ctx = *DAG.getContext();
6166 Ctx.emitError(CS.getInstruction(), "inline asm not supported yet:"
6167 " don't know how to handle tied "
6168 "indirect register inputs");
6171 RegsForValue MatchedRegs;
6172 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
6173 EVT RegVT = AsmNodeOperands[CurOp+1].getValueType();
6174 MatchedRegs.RegVTs.push_back(RegVT);
6175 MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo();
6176 for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag);
6178 MatchedRegs.Regs.push_back
6179 (RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT)));
6181 // Use the produced MatchedRegs object to
6182 MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
6184 MatchedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse,
6185 true, OpInfo.getMatchedOperand(),
6186 DAG, AsmNodeOperands);
6190 assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!");
6191 assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 &&
6192 "Unexpected number of operands");
6193 // Add information to the INLINEASM node to know about this input.
6194 // See InlineAsm.h isUseOperandTiedToDef.
6195 OpFlag = InlineAsm::getFlagWordForMatchingOp(OpFlag,
6196 OpInfo.getMatchedOperand());
6197 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlag,
6198 TLI.getPointerTy()));
6199 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]);
6203 // Treat indirect 'X' constraint as memory.
6204 if (OpInfo.ConstraintType == TargetLowering::C_Other &&
6206 OpInfo.ConstraintType = TargetLowering::C_Memory;
6208 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
6209 std::vector<SDValue> Ops;
6210 TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode,
6213 LLVMContext &Ctx = *DAG.getContext();
6214 Ctx.emitError(CS.getInstruction(),
6215 "invalid operand for inline asm constraint '" +
6216 Twine(OpInfo.ConstraintCode) + "'");
6220 // Add information to the INLINEASM node to know about this input.
6221 unsigned ResOpType =
6222 InlineAsm::getFlagWord(InlineAsm::Kind_Imm, Ops.size());
6223 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
6224 TLI.getPointerTy()));
6225 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
6229 if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
6230 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
6231 assert(InOperandVal.getValueType() == TLI.getPointerTy() &&
6232 "Memory operands expect pointer values");
6234 // Add information to the INLINEASM node to know about this input.
6235 unsigned ResOpType = InlineAsm::getFlagWord(InlineAsm::Kind_Mem, 1);
6236 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
6237 TLI.getPointerTy()));
6238 AsmNodeOperands.push_back(InOperandVal);
6242 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
6243 OpInfo.ConstraintType == TargetLowering::C_Register) &&
6244 "Unknown constraint type!");
6245 assert(!OpInfo.isIndirect &&
6246 "Don't know how to handle indirect register inputs yet!");
6248 // Copy the input into the appropriate registers.
6249 if (OpInfo.AssignedRegs.Regs.empty()) {
6250 LLVMContext &Ctx = *DAG.getContext();
6251 Ctx.emitError(CS.getInstruction(),
6252 "couldn't allocate input reg for constraint '" +
6253 Twine(OpInfo.ConstraintCode) + "'");
6257 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
6260 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_RegUse, false, 0,
6261 DAG, AsmNodeOperands);
6264 case InlineAsm::isClobber: {
6265 // Add the clobbered value to the operand list, so that the register
6266 // allocator is aware that the physreg got clobbered.
6267 if (!OpInfo.AssignedRegs.Regs.empty())
6268 OpInfo.AssignedRegs.AddInlineAsmOperands(InlineAsm::Kind_Clobber,
6276 // Finish up input operands. Set the input chain and add the flag last.
6277 AsmNodeOperands[InlineAsm::Op_InputChain] = Chain;
6278 if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
6280 Chain = DAG.getNode(ISD::INLINEASM, getCurDebugLoc(),
6281 DAG.getVTList(MVT::Other, MVT::Glue),
6282 &AsmNodeOperands[0], AsmNodeOperands.size());
6283 Flag = Chain.getValue(1);
6285 // If this asm returns a register value, copy the result from that register
6286 // and set it as the value of the call.
6287 if (!RetValRegs.Regs.empty()) {
6288 SDValue Val = RetValRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(),
6291 // FIXME: Why don't we do this for inline asms with MRVs?
6292 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) {
6293 EVT ResultType = TLI.getValueType(CS.getType());
6295 // If any of the results of the inline asm is a vector, it may have the
6296 // wrong width/num elts. This can happen for register classes that can
6297 // contain multiple different value types. The preg or vreg allocated may
6298 // not have the same VT as was expected. Convert it to the right type
6299 // with bit_convert.
6300 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) {
6301 Val = DAG.getNode(ISD::BITCAST, getCurDebugLoc(),
6304 } else if (ResultType != Val.getValueType() &&
6305 ResultType.isInteger() && Val.getValueType().isInteger()) {
6306 // If a result value was tied to an input value, the computed result may
6307 // have a wider width than the expected result. Extract the relevant
6309 Val = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), ResultType, Val);
6312 assert(ResultType == Val.getValueType() && "Asm result value mismatch!");
6315 setValue(CS.getInstruction(), Val);
6316 // Don't need to use this as a chain in this case.
6317 if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty())
6321 std::vector<std::pair<SDValue, const Value *> > StoresToEmit;
6323 // Process indirect outputs, first output all of the flagged copies out of
6325 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
6326 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
6327 const Value *Ptr = IndirectStoresToEmit[i].second;
6328 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(),
6330 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
6333 // Emit the non-flagged stores from the physregs.
6334 SmallVector<SDValue, 8> OutChains;
6335 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) {
6336 SDValue Val = DAG.getStore(Chain, getCurDebugLoc(),
6337 StoresToEmit[i].first,
6338 getValue(StoresToEmit[i].second),
6339 MachinePointerInfo(StoresToEmit[i].second),
6341 OutChains.push_back(Val);
6344 if (!OutChains.empty())
6345 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
6346 &OutChains[0], OutChains.size());
6351 void SelectionDAGBuilder::visitVAStart(const CallInst &I) {
6352 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurDebugLoc(),
6353 MVT::Other, getRoot(),
6354 getValue(I.getArgOperand(0)),
6355 DAG.getSrcValue(I.getArgOperand(0))));
6358 void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) {
6359 const TargetData &TD = *TLI.getTargetData();
6360 SDValue V = DAG.getVAArg(TLI.getValueType(I.getType()), getCurDebugLoc(),
6361 getRoot(), getValue(I.getOperand(0)),
6362 DAG.getSrcValue(I.getOperand(0)),
6363 TD.getABITypeAlignment(I.getType()));
6365 DAG.setRoot(V.getValue(1));
6368 void SelectionDAGBuilder::visitVAEnd(const CallInst &I) {
6369 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurDebugLoc(),
6370 MVT::Other, getRoot(),
6371 getValue(I.getArgOperand(0)),
6372 DAG.getSrcValue(I.getArgOperand(0))));
6375 void SelectionDAGBuilder::visitVACopy(const CallInst &I) {
6376 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurDebugLoc(),
6377 MVT::Other, getRoot(),
6378 getValue(I.getArgOperand(0)),
6379 getValue(I.getArgOperand(1)),
6380 DAG.getSrcValue(I.getArgOperand(0)),
6381 DAG.getSrcValue(I.getArgOperand(1))));
6384 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
6385 /// implementation, which just calls LowerCall.
6386 /// FIXME: When all targets are
6387 /// migrated to using LowerCall, this hook should be integrated into SDISel.
6388 std::pair<SDValue, SDValue>
6389 TargetLowering::LowerCallTo(TargetLowering::CallLoweringInfo &CLI) const {
6390 // Handle all of the outgoing arguments.
6392 CLI.OutVals.clear();
6393 ArgListTy &Args = CLI.Args;
6394 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
6395 SmallVector<EVT, 4> ValueVTs;
6396 ComputeValueVTs(*this, Args[i].Ty, ValueVTs);
6397 for (unsigned Value = 0, NumValues = ValueVTs.size();
6398 Value != NumValues; ++Value) {
6399 EVT VT = ValueVTs[Value];
6400 Type *ArgTy = VT.getTypeForEVT(CLI.RetTy->getContext());
6401 SDValue Op = SDValue(Args[i].Node.getNode(),
6402 Args[i].Node.getResNo() + Value);
6403 ISD::ArgFlagsTy Flags;
6404 unsigned OriginalAlignment =
6405 getTargetData()->getABITypeAlignment(ArgTy);
6411 if (Args[i].isInReg)
6415 if (Args[i].isByVal) {
6417 PointerType *Ty = cast<PointerType>(Args[i].Ty);
6418 Type *ElementTy = Ty->getElementType();
6419 Flags.setByValSize(getTargetData()->getTypeAllocSize(ElementTy));
6420 // For ByVal, alignment should come from FE. BE will guess if this
6421 // info is not there but there are cases it cannot get right.
6422 unsigned FrameAlign;
6423 if (Args[i].Alignment)
6424 FrameAlign = Args[i].Alignment;
6426 FrameAlign = getByValTypeAlignment(ElementTy);
6427 Flags.setByValAlign(FrameAlign);
6431 Flags.setOrigAlign(OriginalAlignment);
6433 EVT PartVT = getRegisterType(CLI.RetTy->getContext(), VT);
6434 unsigned NumParts = getNumRegisters(CLI.RetTy->getContext(), VT);
6435 SmallVector<SDValue, 4> Parts(NumParts);
6436 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
6439 ExtendKind = ISD::SIGN_EXTEND;
6440 else if (Args[i].isZExt)
6441 ExtendKind = ISD::ZERO_EXTEND;
6443 getCopyToParts(CLI.DAG, CLI.DL, Op, &Parts[0], NumParts,
6444 PartVT, ExtendKind);
6446 for (unsigned j = 0; j != NumParts; ++j) {
6447 // if it isn't first piece, alignment must be 1
6448 ISD::OutputArg MyFlags(Flags, Parts[j].getValueType(),
6449 i < CLI.NumFixedArgs);
6450 if (NumParts > 1 && j == 0)
6451 MyFlags.Flags.setSplit();
6453 MyFlags.Flags.setOrigAlign(1);
6455 CLI.Outs.push_back(MyFlags);
6456 CLI.OutVals.push_back(Parts[j]);
6461 // Handle the incoming return values from the call.
6463 SmallVector<EVT, 4> RetTys;
6464 ComputeValueVTs(*this, CLI.RetTy, RetTys);
6465 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
6467 EVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
6468 unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
6469 for (unsigned i = 0; i != NumRegs; ++i) {
6470 ISD::InputArg MyFlags;
6471 MyFlags.VT = RegisterVT.getSimpleVT();
6472 MyFlags.Used = CLI.IsReturnValueUsed;
6474 MyFlags.Flags.setSExt();
6476 MyFlags.Flags.setZExt();
6478 MyFlags.Flags.setInReg();
6479 CLI.Ins.push_back(MyFlags);
6483 SmallVector<SDValue, 4> InVals;
6484 CLI.Chain = LowerCall(CLI, InVals);
6486 // Verify that the target's LowerCall behaved as expected.
6487 assert(CLI.Chain.getNode() && CLI.Chain.getValueType() == MVT::Other &&
6488 "LowerCall didn't return a valid chain!");
6489 assert((!CLI.IsTailCall || InVals.empty()) &&
6490 "LowerCall emitted a return value for a tail call!");
6491 assert((CLI.IsTailCall || InVals.size() == CLI.Ins.size()) &&
6492 "LowerCall didn't emit the correct number of values!");
6494 // For a tail call, the return value is merely live-out and there aren't
6495 // any nodes in the DAG representing it. Return a special value to
6496 // indicate that a tail call has been emitted and no more Instructions
6497 // should be processed in the current block.
6498 if (CLI.IsTailCall) {
6499 CLI.DAG.setRoot(CLI.Chain);
6500 return std::make_pair(SDValue(), SDValue());
6503 DEBUG(for (unsigned i = 0, e = CLI.Ins.size(); i != e; ++i) {
6504 assert(InVals[i].getNode() &&
6505 "LowerCall emitted a null value!");
6506 assert(EVT(CLI.Ins[i].VT) == InVals[i].getValueType() &&
6507 "LowerCall emitted a value with the wrong type!");
6510 // Collect the legal value parts into potentially illegal values
6511 // that correspond to the original function's return values.
6512 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6514 AssertOp = ISD::AssertSext;
6515 else if (CLI.RetZExt)
6516 AssertOp = ISD::AssertZext;
6517 SmallVector<SDValue, 4> ReturnValues;
6518 unsigned CurReg = 0;
6519 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
6521 EVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
6522 unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
6524 ReturnValues.push_back(getCopyFromParts(CLI.DAG, CLI.DL, &InVals[CurReg],
6525 NumRegs, RegisterVT, VT,
6530 // For a function returning void, there is no return value. We can't create
6531 // such a node, so we just return a null return value in that case. In
6532 // that case, nothing will actually look at the value.
6533 if (ReturnValues.empty())
6534 return std::make_pair(SDValue(), CLI.Chain);
6536 SDValue Res = CLI.DAG.getNode(ISD::MERGE_VALUES, CLI.DL,
6537 CLI.DAG.getVTList(&RetTys[0], RetTys.size()),
6538 &ReturnValues[0], ReturnValues.size());
6539 return std::make_pair(Res, CLI.Chain);
6542 void TargetLowering::LowerOperationWrapper(SDNode *N,
6543 SmallVectorImpl<SDValue> &Results,
6544 SelectionDAG &DAG) const {
6545 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
6547 Results.push_back(Res);
6550 SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6551 llvm_unreachable("LowerOperation not implemented for this target!");
6555 SelectionDAGBuilder::CopyValueToVirtualRegister(const Value *V, unsigned Reg) {
6556 SDValue Op = getNonRegisterValue(V);
6557 assert((Op.getOpcode() != ISD::CopyFromReg ||
6558 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
6559 "Copy from a reg to the same reg!");
6560 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg");
6562 RegsForValue RFV(V->getContext(), TLI, Reg, V->getType());
6563 SDValue Chain = DAG.getEntryNode();
6564 RFV.getCopyToRegs(Op, DAG, getCurDebugLoc(), Chain, 0);
6565 PendingExports.push_back(Chain);
6568 #include "llvm/CodeGen/SelectionDAGISel.h"
6570 /// isOnlyUsedInEntryBlock - If the specified argument is only used in the
6571 /// entry block, return true. This includes arguments used by switches, since
6572 /// the switch may expand into multiple basic blocks.
6573 static bool isOnlyUsedInEntryBlock(const Argument *A, bool FastISel) {
6574 // With FastISel active, we may be splitting blocks, so force creation
6575 // of virtual registers for all non-dead arguments.
6577 return A->use_empty();
6579 const BasicBlock *Entry = A->getParent()->begin();
6580 for (Value::const_use_iterator UI = A->use_begin(), E = A->use_end();
6582 const User *U = *UI;
6583 if (cast<Instruction>(U)->getParent() != Entry || isa<SwitchInst>(U))
6584 return false; // Use not in entry block.
6589 void SelectionDAGISel::LowerArguments(const BasicBlock *LLVMBB) {
6590 // If this is the entry block, emit arguments.
6591 const Function &F = *LLVMBB->getParent();
6592 SelectionDAG &DAG = SDB->DAG;
6593 DebugLoc dl = SDB->getCurDebugLoc();
6594 const TargetData *TD = TLI.getTargetData();
6595 SmallVector<ISD::InputArg, 16> Ins;
6597 // Check whether the function can return without sret-demotion.
6598 SmallVector<ISD::OutputArg, 4> Outs;
6599 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
6602 if (!FuncInfo->CanLowerReturn) {
6603 // Put in an sret pointer parameter before all the other parameters.
6604 SmallVector<EVT, 1> ValueVTs;
6605 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
6607 // NOTE: Assuming that a pointer will never break down to more than one VT
6609 ISD::ArgFlagsTy Flags;
6611 EVT RegisterVT = TLI.getRegisterType(*DAG.getContext(), ValueVTs[0]);
6612 ISD::InputArg RetArg(Flags, RegisterVT, true);
6613 Ins.push_back(RetArg);
6616 // Set up the incoming argument description vector.
6618 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end();
6619 I != E; ++I, ++Idx) {
6620 SmallVector<EVT, 4> ValueVTs;
6621 ComputeValueVTs(TLI, I->getType(), ValueVTs);
6622 bool isArgValueUsed = !I->use_empty();
6623 for (unsigned Value = 0, NumValues = ValueVTs.size();
6624 Value != NumValues; ++Value) {
6625 EVT VT = ValueVTs[Value];
6626 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
6627 ISD::ArgFlagsTy Flags;
6628 unsigned OriginalAlignment =
6629 TD->getABITypeAlignment(ArgTy);
6631 if (F.paramHasAttr(Idx, Attribute::ZExt))
6633 if (F.paramHasAttr(Idx, Attribute::SExt))
6635 if (F.paramHasAttr(Idx, Attribute::InReg))
6637 if (F.paramHasAttr(Idx, Attribute::StructRet))
6639 if (F.paramHasAttr(Idx, Attribute::ByVal)) {
6641 PointerType *Ty = cast<PointerType>(I->getType());
6642 Type *ElementTy = Ty->getElementType();
6643 Flags.setByValSize(TD->getTypeAllocSize(ElementTy));
6644 // For ByVal, alignment should be passed from FE. BE will guess if
6645 // this info is not there but there are cases it cannot get right.
6646 unsigned FrameAlign;
6647 if (F.getParamAlignment(Idx))
6648 FrameAlign = F.getParamAlignment(Idx);
6650 FrameAlign = TLI.getByValTypeAlignment(ElementTy);
6651 Flags.setByValAlign(FrameAlign);
6653 if (F.paramHasAttr(Idx, Attribute::Nest))
6655 Flags.setOrigAlign(OriginalAlignment);
6657 EVT RegisterVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6658 unsigned NumRegs = TLI.getNumRegisters(*CurDAG->getContext(), VT);
6659 for (unsigned i = 0; i != NumRegs; ++i) {
6660 ISD::InputArg MyFlags(Flags, RegisterVT, isArgValueUsed);
6661 if (NumRegs > 1 && i == 0)
6662 MyFlags.Flags.setSplit();
6663 // if it isn't first piece, alignment must be 1
6665 MyFlags.Flags.setOrigAlign(1);
6666 Ins.push_back(MyFlags);
6671 // Call the target to set up the argument values.
6672 SmallVector<SDValue, 8> InVals;
6673 SDValue NewRoot = TLI.LowerFormalArguments(DAG.getRoot(), F.getCallingConv(),
6677 // Verify that the target's LowerFormalArguments behaved as expected.
6678 assert(NewRoot.getNode() && NewRoot.getValueType() == MVT::Other &&
6679 "LowerFormalArguments didn't return a valid chain!");
6680 assert(InVals.size() == Ins.size() &&
6681 "LowerFormalArguments didn't emit the correct number of values!");
6683 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
6684 assert(InVals[i].getNode() &&
6685 "LowerFormalArguments emitted a null value!");
6686 assert(EVT(Ins[i].VT) == InVals[i].getValueType() &&
6687 "LowerFormalArguments emitted a value with the wrong type!");
6691 // Update the DAG with the new chain value resulting from argument lowering.
6692 DAG.setRoot(NewRoot);
6694 // Set up the argument values.
6697 if (!FuncInfo->CanLowerReturn) {
6698 // Create a virtual register for the sret pointer, and put in a copy
6699 // from the sret argument into it.
6700 SmallVector<EVT, 1> ValueVTs;
6701 ComputeValueVTs(TLI, PointerType::getUnqual(F.getReturnType()), ValueVTs);
6702 EVT VT = ValueVTs[0];
6703 EVT RegVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6704 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6705 SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1,
6706 RegVT, VT, AssertOp);
6708 MachineFunction& MF = SDB->DAG.getMachineFunction();
6709 MachineRegisterInfo& RegInfo = MF.getRegInfo();
6710 unsigned SRetReg = RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT));
6711 FuncInfo->DemoteRegister = SRetReg;
6712 NewRoot = SDB->DAG.getCopyToReg(NewRoot, SDB->getCurDebugLoc(),
6714 DAG.setRoot(NewRoot);
6716 // i indexes lowered arguments. Bump it past the hidden sret argument.
6717 // Idx indexes LLVM arguments. Don't touch it.
6721 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
6723 SmallVector<SDValue, 4> ArgValues;
6724 SmallVector<EVT, 4> ValueVTs;
6725 ComputeValueVTs(TLI, I->getType(), ValueVTs);
6726 unsigned NumValues = ValueVTs.size();
6728 // If this argument is unused then remember its value. It is used to generate
6729 // debugging information.
6730 if (I->use_empty() && NumValues)
6731 SDB->setUnusedArgValue(I, InVals[i]);
6733 for (unsigned Val = 0; Val != NumValues; ++Val) {
6734 EVT VT = ValueVTs[Val];
6735 EVT PartVT = TLI.getRegisterType(*CurDAG->getContext(), VT);
6736 unsigned NumParts = TLI.getNumRegisters(*CurDAG->getContext(), VT);
6738 if (!I->use_empty()) {
6739 ISD::NodeType AssertOp = ISD::DELETED_NODE;
6740 if (F.paramHasAttr(Idx, Attribute::SExt))
6741 AssertOp = ISD::AssertSext;
6742 else if (F.paramHasAttr(Idx, Attribute::ZExt))
6743 AssertOp = ISD::AssertZext;
6745 ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i],
6746 NumParts, PartVT, VT,
6753 // We don't need to do anything else for unused arguments.
6754 if (ArgValues.empty())
6757 // Note down frame index.
6758 if (FrameIndexSDNode *FI =
6759 dyn_cast<FrameIndexSDNode>(ArgValues[0].getNode()))
6760 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
6762 SDValue Res = DAG.getMergeValues(&ArgValues[0], NumValues,
6763 SDB->getCurDebugLoc());
6765 SDB->setValue(I, Res);
6766 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::BUILD_PAIR) {
6767 if (LoadSDNode *LNode =
6768 dyn_cast<LoadSDNode>(Res.getOperand(0).getNode()))
6769 if (FrameIndexSDNode *FI =
6770 dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode()))
6771 FuncInfo->setArgumentFrameIndex(I, FI->getIndex());
6774 // If this argument is live outside of the entry block, insert a copy from
6775 // wherever we got it to the vreg that other BB's will reference it as.
6776 if (!TM.Options.EnableFastISel && Res.getOpcode() == ISD::CopyFromReg) {
6777 // If we can, though, try to skip creating an unnecessary vreg.
6778 // FIXME: This isn't very clean... it would be nice to make this more
6779 // general. It's also subtly incompatible with the hacks FastISel
6781 unsigned Reg = cast<RegisterSDNode>(Res.getOperand(1))->getReg();
6782 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
6783 FuncInfo->ValueMap[I] = Reg;
6787 if (!isOnlyUsedInEntryBlock(I, TM.Options.EnableFastISel)) {
6788 FuncInfo->InitializeRegForValue(I);
6789 SDB->CopyToExportRegsIfNeeded(I);
6793 assert(i == InVals.size() && "Argument register count mismatch!");
6795 // Finally, if the target has anything special to do, allow it to do so.
6796 // FIXME: this should insert code into the DAG!
6797 EmitFunctionEntryCode();
6800 /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
6801 /// ensure constants are generated when needed. Remember the virtual registers
6802 /// that need to be added to the Machine PHI nodes as input. We cannot just
6803 /// directly add them, because expansion might result in multiple MBB's for one
6804 /// BB. As such, the start of the BB might correspond to a different MBB than
6808 SelectionDAGBuilder::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
6809 const TerminatorInst *TI = LLVMBB->getTerminator();
6811 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
6813 // Check successor nodes' PHI nodes that expect a constant to be available
6815 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
6816 const BasicBlock *SuccBB = TI->getSuccessor(succ);
6817 if (!isa<PHINode>(SuccBB->begin())) continue;
6818 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
6820 // If this terminator has multiple identical successors (common for
6821 // switches), only handle each succ once.
6822 if (!SuccsHandled.insert(SuccMBB)) continue;
6824 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
6826 // At this point we know that there is a 1-1 correspondence between LLVM PHI
6827 // nodes and Machine PHI nodes, but the incoming operands have not been
6829 for (BasicBlock::const_iterator I = SuccBB->begin();
6830 const PHINode *PN = dyn_cast<PHINode>(I); ++I) {
6831 // Ignore dead phi's.
6832 if (PN->use_empty()) continue;
6835 if (PN->getType()->isEmptyTy())
6839 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
6841 if (const Constant *C = dyn_cast<Constant>(PHIOp)) {
6842 unsigned &RegOut = ConstantsOut[C];
6844 RegOut = FuncInfo.CreateRegs(C->getType());
6845 CopyValueToVirtualRegister(C, RegOut);
6849 DenseMap<const Value *, unsigned>::iterator I =
6850 FuncInfo.ValueMap.find(PHIOp);
6851 if (I != FuncInfo.ValueMap.end())
6854 assert(isa<AllocaInst>(PHIOp) &&
6855 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
6856 "Didn't codegen value into a register!??");
6857 Reg = FuncInfo.CreateRegs(PHIOp->getType());
6858 CopyValueToVirtualRegister(PHIOp, Reg);
6862 // Remember that this register needs to added to the machine PHI node as
6863 // the input for this MBB.
6864 SmallVector<EVT, 4> ValueVTs;
6865 ComputeValueVTs(TLI, PN->getType(), ValueVTs);
6866 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
6867 EVT VT = ValueVTs[vti];
6868 unsigned NumRegisters = TLI.getNumRegisters(*DAG.getContext(), VT);
6869 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
6870 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
6871 Reg += NumRegisters;
6875 ConstantsOut.clear();