1 //===-- LegalizeVectorOps.cpp - Implement SelectionDAG::LegalizeVectors ---===//
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 file implements the SelectionDAG::LegalizeVectors method.
12 // The vector legalizer looks for vector operations which might need to be
13 // scalarized and legalizes them. This is a separate step from Legalize because
14 // scalarizing can introduce illegal types. For example, suppose we have an
15 // ISD::SDIV of type v2i64 on x86-32. The type is legal (for example, addition
16 // on a v2i64 is legal), but ISD::SDIV isn't legal, so we have to unroll the
17 // operation, which introduces nodes with the illegal type i64 which must be
18 // expanded. Similarly, suppose we have an ISD::SRA of type v16i8 on PowerPC;
19 // the operation must be unrolled, which introduces nodes with the illegal
20 // type i8 which must be promoted.
22 // This does not legalize vector manipulations like ISD::BUILD_VECTOR,
23 // or operations that happen to take a vector which are custom-lowered;
24 // the legalization for such operations never produces nodes
25 // with illegal types, so it's okay to put off legalizing them until
26 // SelectionDAG::Legalize runs.
28 //===----------------------------------------------------------------------===//
30 #include "llvm/CodeGen/SelectionDAG.h"
31 #include "llvm/Target/TargetLowering.h"
35 class VectorLegalizer {
37 const TargetLowering &TLI;
38 bool Changed; // Keep track of whether anything changed
40 /// For nodes that are of legal width, and that have more than one use, this
41 /// map indicates what regularized operand to use. This allows us to avoid
42 /// legalizing the same thing more than once.
43 SmallDenseMap<SDValue, SDValue, 64> LegalizedNodes;
45 /// \brief Adds a node to the translation cache.
46 void AddLegalizedOperand(SDValue From, SDValue To) {
47 LegalizedNodes.insert(std::make_pair(From, To));
48 // If someone requests legalization of the new node, return itself.
50 LegalizedNodes.insert(std::make_pair(To, To));
53 /// \brief Legalizes the given node.
54 SDValue LegalizeOp(SDValue Op);
56 /// \brief Assuming the node is legal, "legalize" the results.
57 SDValue TranslateLegalizeResults(SDValue Op, SDValue Result);
59 /// \brief Implements unrolling a VSETCC.
60 SDValue UnrollVSETCC(SDValue Op);
62 /// \brief Implement expand-based legalization of vector operations.
64 /// This is just a high-level routine to dispatch to specific code paths for
65 /// operations to legalize them.
66 SDValue Expand(SDValue Op);
68 /// \brief Implements expansion for FNEG; falls back to UnrollVectorOp if
71 /// Implements expansion for UINT_TO_FLOAT; falls back to UnrollVectorOp if
72 /// SINT_TO_FLOAT and SHR on vectors isn't legal.
73 SDValue ExpandUINT_TO_FLOAT(SDValue Op);
75 /// \brief Implement expansion for SIGN_EXTEND_INREG using SRL and SRA.
76 SDValue ExpandSEXTINREG(SDValue Op);
78 /// \brief Implement expansion for ANY_EXTEND_VECTOR_INREG.
80 /// Shuffles the low lanes of the operand into place and bitcasts to the proper
81 /// type. The contents of the bits in the extended part of each element are
83 SDValue ExpandANY_EXTEND_VECTOR_INREG(SDValue Op);
85 /// \brief Implement expansion for SIGN_EXTEND_VECTOR_INREG.
87 /// Shuffles the low lanes of the operand into place, bitcasts to the proper
88 /// type, then shifts left and arithmetic shifts right to introduce a sign
90 SDValue ExpandSIGN_EXTEND_VECTOR_INREG(SDValue Op);
92 /// \brief Implement expansion for ZERO_EXTEND_VECTOR_INREG.
94 /// Shuffles the low lanes of the operand into place and blends zeros into
95 /// the remaining lanes, finally bitcasting to the proper type.
96 SDValue ExpandZERO_EXTEND_VECTOR_INREG(SDValue Op);
98 /// \brief Expand bswap of vectors into a shuffle if legal.
99 SDValue ExpandBSWAP(SDValue Op);
101 /// \brief Implement vselect in terms of XOR, AND, OR when blend is not
102 /// supported by the target.
103 SDValue ExpandVSELECT(SDValue Op);
104 SDValue ExpandSELECT(SDValue Op);
105 SDValue ExpandLoad(SDValue Op);
106 SDValue ExpandStore(SDValue Op);
107 SDValue ExpandFNEG(SDValue Op);
109 /// \brief Implements vector promotion.
111 /// This is essentially just bitcasting the operands to a different type and
112 /// bitcasting the result back to the original type.
113 SDValue Promote(SDValue Op);
115 /// \brief Implements [SU]INT_TO_FP vector promotion.
117 /// This is a [zs]ext of the input operand to the next size up.
118 SDValue PromoteINT_TO_FP(SDValue Op);
120 /// \brief Implements FP_TO_[SU]INT vector promotion of the result type.
122 /// It is promoted to the next size up integer type. The result is then
123 /// truncated back to the original type.
124 SDValue PromoteFP_TO_INT(SDValue Op, bool isSigned);
127 /// \brief Begin legalizer the vector operations in the DAG.
129 VectorLegalizer(SelectionDAG& dag) :
130 DAG(dag), TLI(dag.getTargetLoweringInfo()), Changed(false) {}
133 bool VectorLegalizer::Run() {
134 // Before we start legalizing vector nodes, check if there are any vectors.
135 bool HasVectors = false;
136 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
137 E = std::prev(DAG.allnodes_end()); I != std::next(E); ++I) {
138 // Check if the values of the nodes contain vectors. We don't need to check
139 // the operands because we are going to check their values at some point.
140 for (SDNode::value_iterator J = I->value_begin(), E = I->value_end();
142 HasVectors |= J->isVector();
144 // If we found a vector node we can start the legalization.
149 // If this basic block has no vectors then no need to legalize vectors.
153 // The legalize process is inherently a bottom-up recursive process (users
154 // legalize their uses before themselves). Given infinite stack space, we
155 // could just start legalizing on the root and traverse the whole graph. In
156 // practice however, this causes us to run out of stack space on large basic
157 // blocks. To avoid this problem, compute an ordering of the nodes where each
158 // node is only legalized after all of its operands are legalized.
159 DAG.AssignTopologicalOrder();
160 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
161 E = std::prev(DAG.allnodes_end()); I != std::next(E); ++I)
162 LegalizeOp(SDValue(I, 0));
164 // Finally, it's possible the root changed. Get the new root.
165 SDValue OldRoot = DAG.getRoot();
166 assert(LegalizedNodes.count(OldRoot) && "Root didn't get legalized?");
167 DAG.setRoot(LegalizedNodes[OldRoot]);
169 LegalizedNodes.clear();
171 // Remove dead nodes now.
172 DAG.RemoveDeadNodes();
177 SDValue VectorLegalizer::TranslateLegalizeResults(SDValue Op, SDValue Result) {
178 // Generic legalization: just pass the operand through.
179 for (unsigned i = 0, e = Op.getNode()->getNumValues(); i != e; ++i)
180 AddLegalizedOperand(Op.getValue(i), Result.getValue(i));
181 return Result.getValue(Op.getResNo());
184 SDValue VectorLegalizer::LegalizeOp(SDValue Op) {
185 // Note that LegalizeOp may be reentered even from single-use nodes, which
186 // means that we always must cache transformed nodes.
187 DenseMap<SDValue, SDValue>::iterator I = LegalizedNodes.find(Op);
188 if (I != LegalizedNodes.end()) return I->second;
190 SDNode* Node = Op.getNode();
192 // Legalize the operands
193 SmallVector<SDValue, 8> Ops;
194 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
195 Ops.push_back(LegalizeOp(Node->getOperand(i)));
197 SDValue Result = SDValue(DAG.UpdateNodeOperands(Op.getNode(), Ops), 0);
199 if (Op.getOpcode() == ISD::LOAD) {
200 LoadSDNode *LD = cast<LoadSDNode>(Op.getNode());
201 ISD::LoadExtType ExtType = LD->getExtensionType();
202 if (LD->getMemoryVT().isVector() && ExtType != ISD::NON_EXTLOAD)
203 switch (TLI.getLoadExtAction(LD->getExtensionType(), LD->getMemoryVT())) {
204 default: llvm_unreachable("This action is not supported yet!");
205 case TargetLowering::Legal:
206 return TranslateLegalizeResults(Op, Result);
207 case TargetLowering::Custom:
208 if (SDValue Lowered = TLI.LowerOperation(Result, DAG)) {
210 if (Lowered->getNumValues() != Op->getNumValues()) {
211 // This expanded to something other than the load. Assume the
212 // lowering code took care of any chain values, and just handle the
214 assert(Result.getValue(1).use_empty() &&
215 "There are still live users of the old chain!");
216 return LegalizeOp(Lowered);
218 return TranslateLegalizeResults(Op, Lowered);
221 case TargetLowering::Expand:
223 return LegalizeOp(ExpandLoad(Op));
225 } else if (Op.getOpcode() == ISD::STORE) {
226 StoreSDNode *ST = cast<StoreSDNode>(Op.getNode());
227 EVT StVT = ST->getMemoryVT();
228 MVT ValVT = ST->getValue().getSimpleValueType();
229 if (StVT.isVector() && ST->isTruncatingStore())
230 switch (TLI.getTruncStoreAction(ValVT, StVT.getSimpleVT())) {
231 default: llvm_unreachable("This action is not supported yet!");
232 case TargetLowering::Legal:
233 return TranslateLegalizeResults(Op, Result);
234 case TargetLowering::Custom:
236 return TranslateLegalizeResults(Op, TLI.LowerOperation(Result, DAG));
237 case TargetLowering::Expand:
239 return LegalizeOp(ExpandStore(Op));
243 bool HasVectorValue = false;
244 for (SDNode::value_iterator J = Node->value_begin(), E = Node->value_end();
247 HasVectorValue |= J->isVector();
249 return TranslateLegalizeResults(Op, Result);
252 switch (Op.getOpcode()) {
254 return TranslateLegalizeResults(Op, Result);
278 case ISD::CTLZ_ZERO_UNDEF:
279 case ISD::CTTZ_ZERO_UNDEF:
285 case ISD::ZERO_EXTEND:
286 case ISD::ANY_EXTEND:
288 case ISD::SIGN_EXTEND:
289 case ISD::FP_TO_SINT:
290 case ISD::FP_TO_UINT:
309 case ISD::FNEARBYINT:
315 case ISD::SIGN_EXTEND_INREG:
316 case ISD::ANY_EXTEND_VECTOR_INREG:
317 case ISD::SIGN_EXTEND_VECTOR_INREG:
318 case ISD::ZERO_EXTEND_VECTOR_INREG:
319 QueryType = Node->getValueType(0);
321 case ISD::FP_ROUND_INREG:
322 QueryType = cast<VTSDNode>(Node->getOperand(1))->getVT();
324 case ISD::SINT_TO_FP:
325 case ISD::UINT_TO_FP:
326 QueryType = Node->getOperand(0).getValueType();
330 switch (TLI.getOperationAction(Node->getOpcode(), QueryType)) {
331 case TargetLowering::Promote:
332 Result = Promote(Op);
335 case TargetLowering::Legal:
337 case TargetLowering::Custom: {
338 SDValue Tmp1 = TLI.LowerOperation(Op, DAG);
339 if (Tmp1.getNode()) {
345 case TargetLowering::Expand:
349 // Make sure that the generated code is itself legal.
351 Result = LegalizeOp(Result);
355 // Note that LegalizeOp may be reentered even from single-use nodes, which
356 // means that we always must cache transformed nodes.
357 AddLegalizedOperand(Op, Result);
361 SDValue VectorLegalizer::Promote(SDValue Op) {
362 // For a few operations there is a specific concept for promotion based on
363 // the operand's type.
364 switch (Op.getOpcode()) {
365 case ISD::SINT_TO_FP:
366 case ISD::UINT_TO_FP:
367 // "Promote" the operation by extending the operand.
368 return PromoteINT_TO_FP(Op);
369 case ISD::FP_TO_UINT:
370 case ISD::FP_TO_SINT:
371 // Promote the operation by extending the operand.
372 return PromoteFP_TO_INT(Op, Op->getOpcode() == ISD::FP_TO_SINT);
375 // There are currently two cases of vector promotion:
376 // 1) Bitcasting a vector of integers to a different type to a vector of the
377 // same overall length. For example, x86 promotes ISD::AND on v2i32 to v1i64.
378 // 2) Extending a vector of floats to a vector of the same number oflarger
379 // floats. For example, AArch64 promotes ISD::FADD on v4f16 to v4f32.
380 MVT VT = Op.getSimpleValueType();
381 assert(Op.getNode()->getNumValues() == 1 &&
382 "Can't promote a vector with multiple results!");
383 MVT NVT = TLI.getTypeToPromoteTo(Op.getOpcode(), VT);
385 SmallVector<SDValue, 4> Operands(Op.getNumOperands());
387 for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
388 if (Op.getOperand(j).getValueType().isVector())
391 .getVectorElementType()
393 Operands[j] = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Op.getOperand(j));
395 Operands[j] = DAG.getNode(ISD::BITCAST, dl, NVT, Op.getOperand(j));
397 Operands[j] = Op.getOperand(j);
400 Op = DAG.getNode(Op.getOpcode(), dl, NVT, Operands);
401 if (VT.isFloatingPoint() ||
402 (VT.isVector() && VT.getVectorElementType().isFloatingPoint()))
403 return DAG.getNode(ISD::FP_ROUND, dl, VT, Op, DAG.getIntPtrConstant(0));
405 return DAG.getNode(ISD::BITCAST, dl, VT, Op);
408 SDValue VectorLegalizer::PromoteINT_TO_FP(SDValue Op) {
409 // INT_TO_FP operations may require the input operand be promoted even
410 // when the type is otherwise legal.
411 EVT VT = Op.getOperand(0).getValueType();
412 assert(Op.getNode()->getNumValues() == 1 &&
413 "Can't promote a vector with multiple results!");
415 // Normal getTypeToPromoteTo() doesn't work here, as that will promote
416 // by widening the vector w/ the same element width and twice the number
417 // of elements. We want the other way around, the same number of elements,
418 // each twice the width.
420 // Increase the bitwidth of the element to the next pow-of-two
421 // (which is greater than 8 bits).
423 EVT NVT = VT.widenIntegerVectorElementType(*DAG.getContext());
424 assert(NVT.isSimple() && "Promoting to a non-simple vector type!");
426 SmallVector<SDValue, 4> Operands(Op.getNumOperands());
428 unsigned Opc = Op.getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND :
430 for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
431 if (Op.getOperand(j).getValueType().isVector())
432 Operands[j] = DAG.getNode(Opc, dl, NVT, Op.getOperand(j));
434 Operands[j] = Op.getOperand(j);
437 return DAG.getNode(Op.getOpcode(), dl, Op.getValueType(), Operands);
440 // For FP_TO_INT we promote the result type to a vector type with wider
441 // elements and then truncate the result. This is different from the default
442 // PromoteVector which uses bitcast to promote thus assumning that the
443 // promoted vector type has the same overall size.
444 SDValue VectorLegalizer::PromoteFP_TO_INT(SDValue Op, bool isSigned) {
445 assert(Op.getNode()->getNumValues() == 1 &&
446 "Can't promote a vector with multiple results!");
447 EVT VT = Op.getValueType();
452 NewVT = VT.widenIntegerVectorElementType(*DAG.getContext());
453 assert(NewVT.isSimple() && "Promoting to a non-simple vector type!");
454 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NewVT)) {
455 NewOpc = ISD::FP_TO_SINT;
458 if (!isSigned && TLI.isOperationLegalOrCustom(ISD::FP_TO_UINT, NewVT)) {
459 NewOpc = ISD::FP_TO_UINT;
465 SDValue promoted = DAG.getNode(NewOpc, SDLoc(Op), NewVT, Op.getOperand(0));
466 return DAG.getNode(ISD::TRUNCATE, SDLoc(Op), VT, promoted);
470 SDValue VectorLegalizer::ExpandLoad(SDValue Op) {
472 LoadSDNode *LD = cast<LoadSDNode>(Op.getNode());
473 SDValue Chain = LD->getChain();
474 SDValue BasePTR = LD->getBasePtr();
475 EVT SrcVT = LD->getMemoryVT();
476 ISD::LoadExtType ExtType = LD->getExtensionType();
478 SmallVector<SDValue, 8> Vals;
479 SmallVector<SDValue, 8> LoadChains;
480 unsigned NumElem = SrcVT.getVectorNumElements();
482 EVT SrcEltVT = SrcVT.getScalarType();
483 EVT DstEltVT = Op.getNode()->getValueType(0).getScalarType();
485 if (SrcVT.getVectorNumElements() > 1 && !SrcEltVT.isByteSized()) {
486 // When elements in a vector is not byte-addressable, we cannot directly
487 // load each element by advancing pointer, which could only address bytes.
488 // Instead, we load all significant words, mask bits off, and concatenate
489 // them to form each element. Finally, they are extended to destination
490 // scalar type to build the destination vector.
491 EVT WideVT = TLI.getPointerTy();
493 assert(WideVT.isRound() &&
494 "Could not handle the sophisticated case when the widest integer is"
496 assert(WideVT.bitsGE(SrcEltVT) &&
497 "Type is not legalized?");
499 unsigned WideBytes = WideVT.getStoreSize();
501 unsigned RemainingBytes = SrcVT.getStoreSize();
502 SmallVector<SDValue, 8> LoadVals;
504 while (RemainingBytes > 0) {
506 unsigned LoadBytes = WideBytes;
508 if (RemainingBytes >= LoadBytes) {
509 ScalarLoad = DAG.getLoad(WideVT, dl, Chain, BasePTR,
510 LD->getPointerInfo().getWithOffset(Offset),
511 LD->isVolatile(), LD->isNonTemporal(),
512 LD->isInvariant(), LD->getAlignment(),
516 while (RemainingBytes < LoadBytes) {
517 LoadBytes >>= 1; // Reduce the load size by half.
518 LoadVT = EVT::getIntegerVT(*DAG.getContext(), LoadBytes << 3);
520 ScalarLoad = DAG.getExtLoad(ISD::EXTLOAD, dl, WideVT, Chain, BasePTR,
521 LD->getPointerInfo().getWithOffset(Offset),
522 LoadVT, LD->isVolatile(),
523 LD->isNonTemporal(), LD->isInvariant(),
524 LD->getAlignment(), LD->getAAInfo());
527 RemainingBytes -= LoadBytes;
529 BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR,
530 DAG.getConstant(LoadBytes, BasePTR.getValueType()));
532 LoadVals.push_back(ScalarLoad.getValue(0));
533 LoadChains.push_back(ScalarLoad.getValue(1));
536 // Extract bits, pack and extend/trunc them into destination type.
537 unsigned SrcEltBits = SrcEltVT.getSizeInBits();
538 SDValue SrcEltBitMask = DAG.getConstant((1U << SrcEltBits) - 1, WideVT);
540 unsigned BitOffset = 0;
541 unsigned WideIdx = 0;
542 unsigned WideBits = WideVT.getSizeInBits();
544 for (unsigned Idx = 0; Idx != NumElem; ++Idx) {
545 SDValue Lo, Hi, ShAmt;
547 if (BitOffset < WideBits) {
548 ShAmt = DAG.getConstant(BitOffset, TLI.getShiftAmountTy(WideVT));
549 Lo = DAG.getNode(ISD::SRL, dl, WideVT, LoadVals[WideIdx], ShAmt);
550 Lo = DAG.getNode(ISD::AND, dl, WideVT, Lo, SrcEltBitMask);
553 BitOffset += SrcEltBits;
554 if (BitOffset >= WideBits) {
558 ShAmt = DAG.getConstant(SrcEltBits - Offset,
559 TLI.getShiftAmountTy(WideVT));
560 Hi = DAG.getNode(ISD::SHL, dl, WideVT, LoadVals[WideIdx], ShAmt);
561 Hi = DAG.getNode(ISD::AND, dl, WideVT, Hi, SrcEltBitMask);
566 Lo = DAG.getNode(ISD::OR, dl, WideVT, Lo, Hi);
569 default: llvm_unreachable("Unknown extended-load op!");
571 Lo = DAG.getAnyExtOrTrunc(Lo, dl, DstEltVT);
574 Lo = DAG.getZExtOrTrunc(Lo, dl, DstEltVT);
577 ShAmt = DAG.getConstant(WideBits - SrcEltBits,
578 TLI.getShiftAmountTy(WideVT));
579 Lo = DAG.getNode(ISD::SHL, dl, WideVT, Lo, ShAmt);
580 Lo = DAG.getNode(ISD::SRA, dl, WideVT, Lo, ShAmt);
581 Lo = DAG.getSExtOrTrunc(Lo, dl, DstEltVT);
587 unsigned Stride = SrcVT.getScalarType().getSizeInBits()/8;
589 for (unsigned Idx=0; Idx<NumElem; Idx++) {
590 SDValue ScalarLoad = DAG.getExtLoad(ExtType, dl,
591 Op.getNode()->getValueType(0).getScalarType(),
592 Chain, BasePTR, LD->getPointerInfo().getWithOffset(Idx * Stride),
593 SrcVT.getScalarType(),
594 LD->isVolatile(), LD->isNonTemporal(), LD->isInvariant(),
595 LD->getAlignment(), LD->getAAInfo());
597 BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR,
598 DAG.getConstant(Stride, BasePTR.getValueType()));
600 Vals.push_back(ScalarLoad.getValue(0));
601 LoadChains.push_back(ScalarLoad.getValue(1));
605 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, LoadChains);
606 SDValue Value = DAG.getNode(ISD::BUILD_VECTOR, dl,
607 Op.getNode()->getValueType(0), Vals);
609 AddLegalizedOperand(Op.getValue(0), Value);
610 AddLegalizedOperand(Op.getValue(1), NewChain);
612 return (Op.getResNo() ? NewChain : Value);
615 SDValue VectorLegalizer::ExpandStore(SDValue Op) {
617 StoreSDNode *ST = cast<StoreSDNode>(Op.getNode());
618 SDValue Chain = ST->getChain();
619 SDValue BasePTR = ST->getBasePtr();
620 SDValue Value = ST->getValue();
621 EVT StVT = ST->getMemoryVT();
623 unsigned Alignment = ST->getAlignment();
624 bool isVolatile = ST->isVolatile();
625 bool isNonTemporal = ST->isNonTemporal();
626 AAMDNodes AAInfo = ST->getAAInfo();
628 unsigned NumElem = StVT.getVectorNumElements();
629 // The type of the data we want to save
630 EVT RegVT = Value.getValueType();
631 EVT RegSclVT = RegVT.getScalarType();
632 // The type of data as saved in memory.
633 EVT MemSclVT = StVT.getScalarType();
635 // Cast floats into integers
636 unsigned ScalarSize = MemSclVT.getSizeInBits();
638 // Round odd types to the next pow of two.
639 if (!isPowerOf2_32(ScalarSize))
640 ScalarSize = NextPowerOf2(ScalarSize);
642 // Store Stride in bytes
643 unsigned Stride = ScalarSize/8;
644 // Extract each of the elements from the original vector
645 // and save them into memory individually.
646 SmallVector<SDValue, 8> Stores;
647 for (unsigned Idx = 0; Idx < NumElem; Idx++) {
648 SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
649 RegSclVT, Value, DAG.getConstant(Idx, TLI.getVectorIdxTy()));
651 // This scalar TruncStore may be illegal, but we legalize it later.
652 SDValue Store = DAG.getTruncStore(Chain, dl, Ex, BasePTR,
653 ST->getPointerInfo().getWithOffset(Idx*Stride), MemSclVT,
654 isVolatile, isNonTemporal, Alignment, AAInfo);
656 BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR,
657 DAG.getConstant(Stride, BasePTR.getValueType()));
659 Stores.push_back(Store);
661 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
662 AddLegalizedOperand(Op, TF);
666 SDValue VectorLegalizer::Expand(SDValue Op) {
667 switch (Op->getOpcode()) {
668 case ISD::SIGN_EXTEND_INREG:
669 return ExpandSEXTINREG(Op);
670 case ISD::ANY_EXTEND_VECTOR_INREG:
671 return ExpandANY_EXTEND_VECTOR_INREG(Op);
672 case ISD::SIGN_EXTEND_VECTOR_INREG:
673 return ExpandSIGN_EXTEND_VECTOR_INREG(Op);
674 case ISD::ZERO_EXTEND_VECTOR_INREG:
675 return ExpandZERO_EXTEND_VECTOR_INREG(Op);
677 return ExpandBSWAP(Op);
679 return ExpandVSELECT(Op);
681 return ExpandSELECT(Op);
682 case ISD::UINT_TO_FP:
683 return ExpandUINT_TO_FLOAT(Op);
685 return ExpandFNEG(Op);
687 return UnrollVSETCC(Op);
689 return DAG.UnrollVectorOp(Op.getNode());
693 SDValue VectorLegalizer::ExpandSELECT(SDValue Op) {
694 // Lower a select instruction where the condition is a scalar and the
695 // operands are vectors. Lower this select to VSELECT and implement it
696 // using XOR AND OR. The selector bit is broadcasted.
697 EVT VT = Op.getValueType();
700 SDValue Mask = Op.getOperand(0);
701 SDValue Op1 = Op.getOperand(1);
702 SDValue Op2 = Op.getOperand(2);
704 assert(VT.isVector() && !Mask.getValueType().isVector()
705 && Op1.getValueType() == Op2.getValueType() && "Invalid type");
707 unsigned NumElem = VT.getVectorNumElements();
709 // If we can't even use the basic vector operations of
710 // AND,OR,XOR, we will have to scalarize the op.
711 // Notice that the operation may be 'promoted' which means that it is
712 // 'bitcasted' to another type which is handled.
713 // Also, we need to be able to construct a splat vector using BUILD_VECTOR.
714 if (TLI.getOperationAction(ISD::AND, VT) == TargetLowering::Expand ||
715 TLI.getOperationAction(ISD::XOR, VT) == TargetLowering::Expand ||
716 TLI.getOperationAction(ISD::OR, VT) == TargetLowering::Expand ||
717 TLI.getOperationAction(ISD::BUILD_VECTOR, VT) == TargetLowering::Expand)
718 return DAG.UnrollVectorOp(Op.getNode());
720 // Generate a mask operand.
721 EVT MaskTy = VT.changeVectorElementTypeToInteger();
723 // What is the size of each element in the vector mask.
724 EVT BitTy = MaskTy.getScalarType();
726 Mask = DAG.getSelect(DL, BitTy, Mask,
727 DAG.getConstant(APInt::getAllOnesValue(BitTy.getSizeInBits()), BitTy),
728 DAG.getConstant(0, BitTy));
730 // Broadcast the mask so that the entire vector is all-one or all zero.
731 SmallVector<SDValue, 8> Ops(NumElem, Mask);
732 Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, MaskTy, Ops);
734 // Bitcast the operands to be the same type as the mask.
735 // This is needed when we select between FP types because
736 // the mask is a vector of integers.
737 Op1 = DAG.getNode(ISD::BITCAST, DL, MaskTy, Op1);
738 Op2 = DAG.getNode(ISD::BITCAST, DL, MaskTy, Op2);
740 SDValue AllOnes = DAG.getConstant(
741 APInt::getAllOnesValue(BitTy.getSizeInBits()), MaskTy);
742 SDValue NotMask = DAG.getNode(ISD::XOR, DL, MaskTy, Mask, AllOnes);
744 Op1 = DAG.getNode(ISD::AND, DL, MaskTy, Op1, Mask);
745 Op2 = DAG.getNode(ISD::AND, DL, MaskTy, Op2, NotMask);
746 SDValue Val = DAG.getNode(ISD::OR, DL, MaskTy, Op1, Op2);
747 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Val);
750 SDValue VectorLegalizer::ExpandSEXTINREG(SDValue Op) {
751 EVT VT = Op.getValueType();
753 // Make sure that the SRA and SHL instructions are available.
754 if (TLI.getOperationAction(ISD::SRA, VT) == TargetLowering::Expand ||
755 TLI.getOperationAction(ISD::SHL, VT) == TargetLowering::Expand)
756 return DAG.UnrollVectorOp(Op.getNode());
759 EVT OrigTy = cast<VTSDNode>(Op->getOperand(1))->getVT();
761 unsigned BW = VT.getScalarType().getSizeInBits();
762 unsigned OrigBW = OrigTy.getScalarType().getSizeInBits();
763 SDValue ShiftSz = DAG.getConstant(BW - OrigBW, VT);
765 Op = Op.getOperand(0);
766 Op = DAG.getNode(ISD::SHL, DL, VT, Op, ShiftSz);
767 return DAG.getNode(ISD::SRA, DL, VT, Op, ShiftSz);
770 // Generically expand a vector anyext in register to a shuffle of the relevant
771 // lanes into the appropriate locations, with other lanes left undef.
772 SDValue VectorLegalizer::ExpandANY_EXTEND_VECTOR_INREG(SDValue Op) {
774 EVT VT = Op.getValueType();
775 int NumElements = VT.getVectorNumElements();
776 SDValue Src = Op.getOperand(0);
777 EVT SrcVT = Src.getValueType();
778 int NumSrcElements = SrcVT.getVectorNumElements();
780 // Build a base mask of undef shuffles.
781 SmallVector<int, 16> ShuffleMask;
782 ShuffleMask.resize(NumSrcElements, -1);
784 // Place the extended lanes into the correct locations.
785 int ExtLaneScale = NumSrcElements / NumElements;
786 int EndianOffset = TLI.isBigEndian() ? ExtLaneScale - 1 : 0;
787 for (int i = 0; i < NumElements; ++i)
788 ShuffleMask[i * ExtLaneScale + EndianOffset] = i;
791 ISD::BITCAST, DL, VT,
792 DAG.getVectorShuffle(SrcVT, DL, Src, DAG.getUNDEF(SrcVT), ShuffleMask));
795 SDValue VectorLegalizer::ExpandSIGN_EXTEND_VECTOR_INREG(SDValue Op) {
797 EVT VT = Op.getValueType();
798 SDValue Src = Op.getOperand(0);
799 EVT SrcVT = Src.getValueType();
801 // First build an any-extend node which can be legalized above when we
802 // recurse through it.
803 Op = DAG.getAnyExtendVectorInReg(Src, DL, VT);
805 // Now we need sign extend. Do this by shifting the elements. Even if these
806 // aren't legal operations, they have a better chance of being legalized
807 // without full scalarization than the sign extension does.
808 unsigned EltWidth = VT.getVectorElementType().getSizeInBits();
809 unsigned SrcEltWidth = SrcVT.getVectorElementType().getSizeInBits();
810 SDValue ShiftAmount = DAG.getConstant(EltWidth - SrcEltWidth, VT);
811 return DAG.getNode(ISD::SRA, DL, VT,
812 DAG.getNode(ISD::SHL, DL, VT, Op, ShiftAmount),
816 // Generically expand a vector zext in register to a shuffle of the relevant
817 // lanes into the appropriate locations, a blend of zero into the high bits,
818 // and a bitcast to the wider element type.
819 SDValue VectorLegalizer::ExpandZERO_EXTEND_VECTOR_INREG(SDValue Op) {
821 EVT VT = Op.getValueType();
822 int NumElements = VT.getVectorNumElements();
823 SDValue Src = Op.getOperand(0);
824 EVT SrcVT = Src.getValueType();
825 int NumSrcElements = SrcVT.getVectorNumElements();
827 // Build up a zero vector to blend into this one.
828 EVT SrcScalarVT = SrcVT.getScalarType();
829 SDValue ScalarZero = DAG.getTargetConstant(0, SrcScalarVT);
830 SmallVector<SDValue, 4> BuildVectorOperands(NumSrcElements, ScalarZero);
831 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, DL, SrcVT, BuildVectorOperands);
833 // Shuffle the incoming lanes into the correct position, and pull all other
834 // lanes from the zero vector.
835 SmallVector<int, 16> ShuffleMask;
836 ShuffleMask.reserve(NumSrcElements);
837 for (int i = 0; i < NumSrcElements; ++i)
838 ShuffleMask.push_back(i);
840 int ExtLaneScale = NumSrcElements / NumElements;
841 int EndianOffset = TLI.isBigEndian() ? ExtLaneScale - 1 : 0;
842 for (int i = 0; i < NumElements; ++i)
843 ShuffleMask[i * ExtLaneScale + EndianOffset] = NumSrcElements + i;
845 return DAG.getNode(ISD::BITCAST, DL, VT,
846 DAG.getVectorShuffle(SrcVT, DL, Zero, Src, ShuffleMask));
849 SDValue VectorLegalizer::ExpandBSWAP(SDValue Op) {
850 EVT VT = Op.getValueType();
852 // Generate a byte wise shuffle mask for the BSWAP.
853 SmallVector<int, 16> ShuffleMask;
854 int ScalarSizeInBytes = VT.getScalarSizeInBits() / 8;
855 for (int I = 0, E = VT.getVectorNumElements(); I != E; ++I)
856 for (int J = ScalarSizeInBytes - 1; J >= 0; --J)
857 ShuffleMask.push_back((I * ScalarSizeInBytes) + J);
859 EVT ByteVT = EVT::getVectorVT(*DAG.getContext(), MVT::i8, ShuffleMask.size());
861 // Only emit a shuffle if the mask is legal.
862 if (!TLI.isShuffleMaskLegal(ShuffleMask, ByteVT))
863 return DAG.UnrollVectorOp(Op.getNode());
866 Op = DAG.getNode(ISD::BITCAST, DL, ByteVT, Op.getOperand(0));
867 Op = DAG.getVectorShuffle(ByteVT, DL, Op, DAG.getUNDEF(ByteVT),
869 return DAG.getNode(ISD::BITCAST, DL, VT, Op);
872 SDValue VectorLegalizer::ExpandVSELECT(SDValue Op) {
873 // Implement VSELECT in terms of XOR, AND, OR
874 // on platforms which do not support blend natively.
877 SDValue Mask = Op.getOperand(0);
878 SDValue Op1 = Op.getOperand(1);
879 SDValue Op2 = Op.getOperand(2);
881 EVT VT = Mask.getValueType();
883 // If we can't even use the basic vector operations of
884 // AND,OR,XOR, we will have to scalarize the op.
885 // Notice that the operation may be 'promoted' which means that it is
886 // 'bitcasted' to another type which is handled.
887 // This operation also isn't safe with AND, OR, XOR when the boolean
888 // type is 0/1 as we need an all ones vector constant to mask with.
889 // FIXME: Sign extend 1 to all ones if thats legal on the target.
890 if (TLI.getOperationAction(ISD::AND, VT) == TargetLowering::Expand ||
891 TLI.getOperationAction(ISD::XOR, VT) == TargetLowering::Expand ||
892 TLI.getOperationAction(ISD::OR, VT) == TargetLowering::Expand ||
893 TLI.getBooleanContents(Op1.getValueType()) !=
894 TargetLowering::ZeroOrNegativeOneBooleanContent)
895 return DAG.UnrollVectorOp(Op.getNode());
897 // If the mask and the type are different sizes, unroll the vector op. This
898 // can occur when getSetCCResultType returns something that is different in
899 // size from the operand types. For example, v4i8 = select v4i32, v4i8, v4i8.
900 if (VT.getSizeInBits() != Op1.getValueType().getSizeInBits())
901 return DAG.UnrollVectorOp(Op.getNode());
903 // Bitcast the operands to be the same type as the mask.
904 // This is needed when we select between FP types because
905 // the mask is a vector of integers.
906 Op1 = DAG.getNode(ISD::BITCAST, DL, VT, Op1);
907 Op2 = DAG.getNode(ISD::BITCAST, DL, VT, Op2);
909 SDValue AllOnes = DAG.getConstant(
910 APInt::getAllOnesValue(VT.getScalarType().getSizeInBits()), VT);
911 SDValue NotMask = DAG.getNode(ISD::XOR, DL, VT, Mask, AllOnes);
913 Op1 = DAG.getNode(ISD::AND, DL, VT, Op1, Mask);
914 Op2 = DAG.getNode(ISD::AND, DL, VT, Op2, NotMask);
915 SDValue Val = DAG.getNode(ISD::OR, DL, VT, Op1, Op2);
916 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Val);
919 SDValue VectorLegalizer::ExpandUINT_TO_FLOAT(SDValue Op) {
920 EVT VT = Op.getOperand(0).getValueType();
923 // Make sure that the SINT_TO_FP and SRL instructions are available.
924 if (TLI.getOperationAction(ISD::SINT_TO_FP, VT) == TargetLowering::Expand ||
925 TLI.getOperationAction(ISD::SRL, VT) == TargetLowering::Expand)
926 return DAG.UnrollVectorOp(Op.getNode());
928 EVT SVT = VT.getScalarType();
929 assert((SVT.getSizeInBits() == 64 || SVT.getSizeInBits() == 32) &&
930 "Elements in vector-UINT_TO_FP must be 32 or 64 bits wide");
932 unsigned BW = SVT.getSizeInBits();
933 SDValue HalfWord = DAG.getConstant(BW/2, VT);
935 // Constants to clear the upper part of the word.
936 // Notice that we can also use SHL+SHR, but using a constant is slightly
938 uint64_t HWMask = (SVT.getSizeInBits()==64)?0x00000000FFFFFFFF:0x0000FFFF;
939 SDValue HalfWordMask = DAG.getConstant(HWMask, VT);
941 // Two to the power of half-word-size.
942 SDValue TWOHW = DAG.getConstantFP((1<<(BW/2)), Op.getValueType());
944 // Clear upper part of LO, lower HI
945 SDValue HI = DAG.getNode(ISD::SRL, DL, VT, Op.getOperand(0), HalfWord);
946 SDValue LO = DAG.getNode(ISD::AND, DL, VT, Op.getOperand(0), HalfWordMask);
948 // Convert hi and lo to floats
949 // Convert the hi part back to the upper values
950 SDValue fHI = DAG.getNode(ISD::SINT_TO_FP, DL, Op.getValueType(), HI);
951 fHI = DAG.getNode(ISD::FMUL, DL, Op.getValueType(), fHI, TWOHW);
952 SDValue fLO = DAG.getNode(ISD::SINT_TO_FP, DL, Op.getValueType(), LO);
954 // Add the two halves
955 return DAG.getNode(ISD::FADD, DL, Op.getValueType(), fHI, fLO);
959 SDValue VectorLegalizer::ExpandFNEG(SDValue Op) {
960 if (TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType())) {
961 SDValue Zero = DAG.getConstantFP(-0.0, Op.getValueType());
962 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
963 Zero, Op.getOperand(0));
965 return DAG.UnrollVectorOp(Op.getNode());
968 SDValue VectorLegalizer::UnrollVSETCC(SDValue Op) {
969 EVT VT = Op.getValueType();
970 unsigned NumElems = VT.getVectorNumElements();
971 EVT EltVT = VT.getVectorElementType();
972 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1), CC = Op.getOperand(2);
973 EVT TmpEltVT = LHS.getValueType().getVectorElementType();
975 SmallVector<SDValue, 8> Ops(NumElems);
976 for (unsigned i = 0; i < NumElems; ++i) {
977 SDValue LHSElem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, LHS,
978 DAG.getConstant(i, TLI.getVectorIdxTy()));
979 SDValue RHSElem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, RHS,
980 DAG.getConstant(i, TLI.getVectorIdxTy()));
981 Ops[i] = DAG.getNode(ISD::SETCC, dl,
982 TLI.getSetCCResultType(*DAG.getContext(), TmpEltVT),
983 LHSElem, RHSElem, CC);
984 Ops[i] = DAG.getSelect(dl, EltVT, Ops[i],
985 DAG.getConstant(APInt::getAllOnesValue
986 (EltVT.getSizeInBits()), EltVT),
987 DAG.getConstant(0, EltVT));
989 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
994 bool SelectionDAG::LegalizeVectors() {
995 return VectorLegalizer(*this).Run();