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 Implements expansion for FNEG; falls back to UnrollVectorOp if
65 /// Implements expansion for UINT_TO_FLOAT; falls back to UnrollVectorOp if
66 /// SINT_TO_FLOAT and SHR on vectors isn't legal.
67 SDValue ExpandUINT_TO_FLOAT(SDValue Op);
69 /// \brief Implement expansion for SIGN_EXTEND_INREG using SRL and SRA.
70 SDValue ExpandSEXTINREG(SDValue Op);
72 /// \brief Expand bswap of vectors into a shuffle if legal.
73 SDValue ExpandBSWAP(SDValue Op);
75 /// \brief Implement vselect in terms of XOR, AND, OR when blend is not
76 /// supported by the target.
77 SDValue ExpandVSELECT(SDValue Op);
78 SDValue ExpandSELECT(SDValue Op);
79 SDValue ExpandLoad(SDValue Op);
80 SDValue ExpandStore(SDValue Op);
81 SDValue ExpandFNEG(SDValue Op);
83 /// \brief Implements vector promotion.
85 /// This is essentially just bitcasting the operands to a different type and
86 /// bitcasting the result back to the original type.
87 SDValue Promote(SDValue Op);
89 /// \brief Implements [SU]INT_TO_FP vector promotion.
91 /// This is a [zs]ext of the input operand to the next size up.
92 SDValue PromoteINT_TO_FP(SDValue Op);
94 /// \brief Implements FP_TO_[SU]INT vector promotion of the result type.
96 /// It is promoted to the next size up integer type. The result is then
97 /// truncated back to the original type.
98 SDValue PromoteFP_TO_INT(SDValue Op, bool isSigned);
101 /// \brief Begin legalizer the vector operations in the DAG.
103 VectorLegalizer(SelectionDAG& dag) :
104 DAG(dag), TLI(dag.getTargetLoweringInfo()), Changed(false) {}
107 bool VectorLegalizer::Run() {
108 // Before we start legalizing vector nodes, check if there are any vectors.
109 bool HasVectors = false;
110 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
111 E = std::prev(DAG.allnodes_end()); I != std::next(E); ++I) {
112 // Check if the values of the nodes contain vectors. We don't need to check
113 // the operands because we are going to check their values at some point.
114 for (SDNode::value_iterator J = I->value_begin(), E = I->value_end();
116 HasVectors |= J->isVector();
118 // If we found a vector node we can start the legalization.
123 // If this basic block has no vectors then no need to legalize vectors.
127 // The legalize process is inherently a bottom-up recursive process (users
128 // legalize their uses before themselves). Given infinite stack space, we
129 // could just start legalizing on the root and traverse the whole graph. In
130 // practice however, this causes us to run out of stack space on large basic
131 // blocks. To avoid this problem, compute an ordering of the nodes where each
132 // node is only legalized after all of its operands are legalized.
133 DAG.AssignTopologicalOrder();
134 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
135 E = std::prev(DAG.allnodes_end()); I != std::next(E); ++I)
136 LegalizeOp(SDValue(I, 0));
138 // Finally, it's possible the root changed. Get the new root.
139 SDValue OldRoot = DAG.getRoot();
140 assert(LegalizedNodes.count(OldRoot) && "Root didn't get legalized?");
141 DAG.setRoot(LegalizedNodes[OldRoot]);
143 LegalizedNodes.clear();
145 // Remove dead nodes now.
146 DAG.RemoveDeadNodes();
151 SDValue VectorLegalizer::TranslateLegalizeResults(SDValue Op, SDValue Result) {
152 // Generic legalization: just pass the operand through.
153 for (unsigned i = 0, e = Op.getNode()->getNumValues(); i != e; ++i)
154 AddLegalizedOperand(Op.getValue(i), Result.getValue(i));
155 return Result.getValue(Op.getResNo());
158 SDValue VectorLegalizer::LegalizeOp(SDValue Op) {
159 // Note that LegalizeOp may be reentered even from single-use nodes, which
160 // means that we always must cache transformed nodes.
161 DenseMap<SDValue, SDValue>::iterator I = LegalizedNodes.find(Op);
162 if (I != LegalizedNodes.end()) return I->second;
164 SDNode* Node = Op.getNode();
166 // Legalize the operands
167 SmallVector<SDValue, 8> Ops;
168 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
169 Ops.push_back(LegalizeOp(Node->getOperand(i)));
171 SDValue Result = SDValue(DAG.UpdateNodeOperands(Op.getNode(), Ops), 0);
173 if (Op.getOpcode() == ISD::LOAD) {
174 LoadSDNode *LD = cast<LoadSDNode>(Op.getNode());
175 ISD::LoadExtType ExtType = LD->getExtensionType();
176 if (LD->getMemoryVT().isVector() && ExtType != ISD::NON_EXTLOAD) {
177 if (TLI.isLoadExtLegal(LD->getExtensionType(), LD->getMemoryVT()))
178 return TranslateLegalizeResults(Op, Result);
180 return LegalizeOp(ExpandLoad(Op));
182 } else if (Op.getOpcode() == ISD::STORE) {
183 StoreSDNode *ST = cast<StoreSDNode>(Op.getNode());
184 EVT StVT = ST->getMemoryVT();
185 MVT ValVT = ST->getValue().getSimpleValueType();
186 if (StVT.isVector() && ST->isTruncatingStore())
187 switch (TLI.getTruncStoreAction(ValVT, StVT.getSimpleVT())) {
188 default: llvm_unreachable("This action is not supported yet!");
189 case TargetLowering::Legal:
190 return TranslateLegalizeResults(Op, Result);
191 case TargetLowering::Custom:
193 return TranslateLegalizeResults(Op, TLI.LowerOperation(Result, DAG));
194 case TargetLowering::Expand:
196 return LegalizeOp(ExpandStore(Op));
200 bool HasVectorValue = false;
201 for (SDNode::value_iterator J = Node->value_begin(), E = Node->value_end();
204 HasVectorValue |= J->isVector();
206 return TranslateLegalizeResults(Op, Result);
209 switch (Op.getOpcode()) {
211 return TranslateLegalizeResults(Op, Result);
235 case ISD::CTLZ_ZERO_UNDEF:
236 case ISD::CTTZ_ZERO_UNDEF:
242 case ISD::ZERO_EXTEND:
243 case ISD::ANY_EXTEND:
245 case ISD::SIGN_EXTEND:
246 case ISD::FP_TO_SINT:
247 case ISD::FP_TO_UINT:
264 case ISD::FNEARBYINT:
270 case ISD::SIGN_EXTEND_INREG:
271 QueryType = Node->getValueType(0);
273 case ISD::FP_ROUND_INREG:
274 QueryType = cast<VTSDNode>(Node->getOperand(1))->getVT();
276 case ISD::SINT_TO_FP:
277 case ISD::UINT_TO_FP:
278 QueryType = Node->getOperand(0).getValueType();
282 switch (TLI.getOperationAction(Node->getOpcode(), QueryType)) {
283 case TargetLowering::Promote:
284 switch (Op.getOpcode()) {
286 // "Promote" the operation by bitcasting
287 Result = Promote(Op);
290 case ISD::SINT_TO_FP:
291 case ISD::UINT_TO_FP:
292 // "Promote" the operation by extending the operand.
293 Result = PromoteINT_TO_FP(Op);
296 case ISD::FP_TO_UINT:
297 case ISD::FP_TO_SINT:
298 // Promote the operation by extending the operand.
299 Result = PromoteFP_TO_INT(Op, Op->getOpcode() == ISD::FP_TO_SINT);
304 case TargetLowering::Legal: break;
305 case TargetLowering::Custom: {
306 SDValue Tmp1 = TLI.LowerOperation(Op, DAG);
307 if (Tmp1.getNode()) {
313 case TargetLowering::Expand:
314 if (Node->getOpcode() == ISD::SIGN_EXTEND_INREG)
315 Result = ExpandSEXTINREG(Op);
316 else if (Node->getOpcode() == ISD::BSWAP)
317 Result = ExpandBSWAP(Op);
318 else if (Node->getOpcode() == ISD::VSELECT)
319 Result = ExpandVSELECT(Op);
320 else if (Node->getOpcode() == ISD::SELECT)
321 Result = ExpandSELECT(Op);
322 else if (Node->getOpcode() == ISD::UINT_TO_FP)
323 Result = ExpandUINT_TO_FLOAT(Op);
324 else if (Node->getOpcode() == ISD::FNEG)
325 Result = ExpandFNEG(Op);
326 else if (Node->getOpcode() == ISD::SETCC)
327 Result = UnrollVSETCC(Op);
329 Result = DAG.UnrollVectorOp(Op.getNode());
333 // Make sure that the generated code is itself legal.
335 Result = LegalizeOp(Result);
339 // Note that LegalizeOp may be reentered even from single-use nodes, which
340 // means that we always must cache transformed nodes.
341 AddLegalizedOperand(Op, Result);
345 SDValue VectorLegalizer::Promote(SDValue Op) {
346 // Vector "promotion" is basically just bitcasting and doing the operation
347 // in a different type. For example, x86 promotes ISD::AND on v2i32 to
349 MVT VT = Op.getSimpleValueType();
350 assert(Op.getNode()->getNumValues() == 1 &&
351 "Can't promote a vector with multiple results!");
352 MVT NVT = TLI.getTypeToPromoteTo(Op.getOpcode(), VT);
354 SmallVector<SDValue, 4> Operands(Op.getNumOperands());
356 for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
357 if (Op.getOperand(j).getValueType().isVector())
358 Operands[j] = DAG.getNode(ISD::BITCAST, dl, NVT, Op.getOperand(j));
360 Operands[j] = Op.getOperand(j);
363 Op = DAG.getNode(Op.getOpcode(), dl, NVT, Operands);
365 return DAG.getNode(ISD::BITCAST, dl, VT, Op);
368 SDValue VectorLegalizer::PromoteINT_TO_FP(SDValue Op) {
369 // INT_TO_FP operations may require the input operand be promoted even
370 // when the type is otherwise legal.
371 EVT VT = Op.getOperand(0).getValueType();
372 assert(Op.getNode()->getNumValues() == 1 &&
373 "Can't promote a vector with multiple results!");
375 // Normal getTypeToPromoteTo() doesn't work here, as that will promote
376 // by widening the vector w/ the same element width and twice the number
377 // of elements. We want the other way around, the same number of elements,
378 // each twice the width.
380 // Increase the bitwidth of the element to the next pow-of-two
381 // (which is greater than 8 bits).
383 EVT NVT = VT.widenIntegerVectorElementType(*DAG.getContext());
384 assert(NVT.isSimple() && "Promoting to a non-simple vector type!");
386 SmallVector<SDValue, 4> Operands(Op.getNumOperands());
388 unsigned Opc = Op.getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND :
390 for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
391 if (Op.getOperand(j).getValueType().isVector())
392 Operands[j] = DAG.getNode(Opc, dl, NVT, Op.getOperand(j));
394 Operands[j] = Op.getOperand(j);
397 return DAG.getNode(Op.getOpcode(), dl, Op.getValueType(), Operands);
400 // For FP_TO_INT we promote the result type to a vector type with wider
401 // elements and then truncate the result. This is different from the default
402 // PromoteVector which uses bitcast to promote thus assumning that the
403 // promoted vector type has the same overall size.
404 SDValue VectorLegalizer::PromoteFP_TO_INT(SDValue Op, bool isSigned) {
405 assert(Op.getNode()->getNumValues() == 1 &&
406 "Can't promote a vector with multiple results!");
407 EVT VT = Op.getValueType();
412 NewVT = VT.widenIntegerVectorElementType(*DAG.getContext());
413 assert(NewVT.isSimple() && "Promoting to a non-simple vector type!");
414 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NewVT)) {
415 NewOpc = ISD::FP_TO_SINT;
418 if (!isSigned && TLI.isOperationLegalOrCustom(ISD::FP_TO_UINT, NewVT)) {
419 NewOpc = ISD::FP_TO_UINT;
425 SDValue promoted = DAG.getNode(NewOpc, SDLoc(Op), NewVT, Op.getOperand(0));
426 return DAG.getNode(ISD::TRUNCATE, SDLoc(Op), VT, promoted);
430 SDValue VectorLegalizer::ExpandLoad(SDValue Op) {
432 LoadSDNode *LD = cast<LoadSDNode>(Op.getNode());
433 SDValue Chain = LD->getChain();
434 SDValue BasePTR = LD->getBasePtr();
435 EVT SrcVT = LD->getMemoryVT();
436 ISD::LoadExtType ExtType = LD->getExtensionType();
438 SmallVector<SDValue, 8> Vals;
439 SmallVector<SDValue, 8> LoadChains;
440 unsigned NumElem = SrcVT.getVectorNumElements();
442 EVT SrcEltVT = SrcVT.getScalarType();
443 EVT DstEltVT = Op.getNode()->getValueType(0).getScalarType();
445 if (SrcVT.getVectorNumElements() > 1 && !SrcEltVT.isByteSized()) {
446 // When elements in a vector is not byte-addressable, we cannot directly
447 // load each element by advancing pointer, which could only address bytes.
448 // Instead, we load all significant words, mask bits off, and concatenate
449 // them to form each element. Finally, they are extended to destination
450 // scalar type to build the destination vector.
451 EVT WideVT = TLI.getPointerTy();
453 assert(WideVT.isRound() &&
454 "Could not handle the sophisticated case when the widest integer is"
456 assert(WideVT.bitsGE(SrcEltVT) &&
457 "Type is not legalized?");
459 unsigned WideBytes = WideVT.getStoreSize();
461 unsigned RemainingBytes = SrcVT.getStoreSize();
462 SmallVector<SDValue, 8> LoadVals;
464 while (RemainingBytes > 0) {
466 unsigned LoadBytes = WideBytes;
468 if (RemainingBytes >= LoadBytes) {
469 ScalarLoad = DAG.getLoad(WideVT, dl, Chain, BasePTR,
470 LD->getPointerInfo().getWithOffset(Offset),
471 LD->isVolatile(), LD->isNonTemporal(),
472 LD->isInvariant(), LD->getAlignment(),
476 while (RemainingBytes < LoadBytes) {
477 LoadBytes >>= 1; // Reduce the load size by half.
478 LoadVT = EVT::getIntegerVT(*DAG.getContext(), LoadBytes << 3);
480 ScalarLoad = DAG.getExtLoad(ISD::EXTLOAD, dl, WideVT, Chain, BasePTR,
481 LD->getPointerInfo().getWithOffset(Offset),
482 LoadVT, LD->isVolatile(),
483 LD->isNonTemporal(), LD->getAlignment(),
487 RemainingBytes -= LoadBytes;
489 BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR,
490 DAG.getConstant(LoadBytes, BasePTR.getValueType()));
492 LoadVals.push_back(ScalarLoad.getValue(0));
493 LoadChains.push_back(ScalarLoad.getValue(1));
496 // Extract bits, pack and extend/trunc them into destination type.
497 unsigned SrcEltBits = SrcEltVT.getSizeInBits();
498 SDValue SrcEltBitMask = DAG.getConstant((1U << SrcEltBits) - 1, WideVT);
500 unsigned BitOffset = 0;
501 unsigned WideIdx = 0;
502 unsigned WideBits = WideVT.getSizeInBits();
504 for (unsigned Idx = 0; Idx != NumElem; ++Idx) {
505 SDValue Lo, Hi, ShAmt;
507 if (BitOffset < WideBits) {
508 ShAmt = DAG.getConstant(BitOffset, TLI.getShiftAmountTy(WideVT));
509 Lo = DAG.getNode(ISD::SRL, dl, WideVT, LoadVals[WideIdx], ShAmt);
510 Lo = DAG.getNode(ISD::AND, dl, WideVT, Lo, SrcEltBitMask);
513 BitOffset += SrcEltBits;
514 if (BitOffset >= WideBits) {
518 ShAmt = DAG.getConstant(SrcEltBits - Offset,
519 TLI.getShiftAmountTy(WideVT));
520 Hi = DAG.getNode(ISD::SHL, dl, WideVT, LoadVals[WideIdx], ShAmt);
521 Hi = DAG.getNode(ISD::AND, dl, WideVT, Hi, SrcEltBitMask);
526 Lo = DAG.getNode(ISD::OR, dl, WideVT, Lo, Hi);
529 default: llvm_unreachable("Unknown extended-load op!");
531 Lo = DAG.getAnyExtOrTrunc(Lo, dl, DstEltVT);
534 Lo = DAG.getZExtOrTrunc(Lo, dl, DstEltVT);
537 ShAmt = DAG.getConstant(WideBits - SrcEltBits,
538 TLI.getShiftAmountTy(WideVT));
539 Lo = DAG.getNode(ISD::SHL, dl, WideVT, Lo, ShAmt);
540 Lo = DAG.getNode(ISD::SRA, dl, WideVT, Lo, ShAmt);
541 Lo = DAG.getSExtOrTrunc(Lo, dl, DstEltVT);
547 unsigned Stride = SrcVT.getScalarType().getSizeInBits()/8;
549 for (unsigned Idx=0; Idx<NumElem; Idx++) {
550 SDValue ScalarLoad = DAG.getExtLoad(ExtType, dl,
551 Op.getNode()->getValueType(0).getScalarType(),
552 Chain, BasePTR, LD->getPointerInfo().getWithOffset(Idx * Stride),
553 SrcVT.getScalarType(),
554 LD->isVolatile(), LD->isNonTemporal(),
555 LD->getAlignment(), LD->getTBAAInfo());
557 BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR,
558 DAG.getConstant(Stride, BasePTR.getValueType()));
560 Vals.push_back(ScalarLoad.getValue(0));
561 LoadChains.push_back(ScalarLoad.getValue(1));
565 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, LoadChains);
566 SDValue Value = DAG.getNode(ISD::BUILD_VECTOR, dl,
567 Op.getNode()->getValueType(0), Vals);
569 AddLegalizedOperand(Op.getValue(0), Value);
570 AddLegalizedOperand(Op.getValue(1), NewChain);
572 return (Op.getResNo() ? NewChain : Value);
575 SDValue VectorLegalizer::ExpandStore(SDValue Op) {
577 StoreSDNode *ST = cast<StoreSDNode>(Op.getNode());
578 SDValue Chain = ST->getChain();
579 SDValue BasePTR = ST->getBasePtr();
580 SDValue Value = ST->getValue();
581 EVT StVT = ST->getMemoryVT();
583 unsigned Alignment = ST->getAlignment();
584 bool isVolatile = ST->isVolatile();
585 bool isNonTemporal = ST->isNonTemporal();
586 const MDNode *TBAAInfo = ST->getTBAAInfo();
588 unsigned NumElem = StVT.getVectorNumElements();
589 // The type of the data we want to save
590 EVT RegVT = Value.getValueType();
591 EVT RegSclVT = RegVT.getScalarType();
592 // The type of data as saved in memory.
593 EVT MemSclVT = StVT.getScalarType();
595 // Cast floats into integers
596 unsigned ScalarSize = MemSclVT.getSizeInBits();
598 // Round odd types to the next pow of two.
599 if (!isPowerOf2_32(ScalarSize))
600 ScalarSize = NextPowerOf2(ScalarSize);
602 // Store Stride in bytes
603 unsigned Stride = ScalarSize/8;
604 // Extract each of the elements from the original vector
605 // and save them into memory individually.
606 SmallVector<SDValue, 8> Stores;
607 for (unsigned Idx = 0; Idx < NumElem; Idx++) {
608 SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
609 RegSclVT, Value, DAG.getConstant(Idx, TLI.getVectorIdxTy()));
611 // This scalar TruncStore may be illegal, but we legalize it later.
612 SDValue Store = DAG.getTruncStore(Chain, dl, Ex, BasePTR,
613 ST->getPointerInfo().getWithOffset(Idx*Stride), MemSclVT,
614 isVolatile, isNonTemporal, Alignment, TBAAInfo);
616 BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR,
617 DAG.getConstant(Stride, BasePTR.getValueType()));
619 Stores.push_back(Store);
621 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
622 AddLegalizedOperand(Op, TF);
626 SDValue VectorLegalizer::ExpandSELECT(SDValue Op) {
627 // Lower a select instruction where the condition is a scalar and the
628 // operands are vectors. Lower this select to VSELECT and implement it
629 // using XOR AND OR. The selector bit is broadcasted.
630 EVT VT = Op.getValueType();
633 SDValue Mask = Op.getOperand(0);
634 SDValue Op1 = Op.getOperand(1);
635 SDValue Op2 = Op.getOperand(2);
637 assert(VT.isVector() && !Mask.getValueType().isVector()
638 && Op1.getValueType() == Op2.getValueType() && "Invalid type");
640 unsigned NumElem = VT.getVectorNumElements();
642 // If we can't even use the basic vector operations of
643 // AND,OR,XOR, we will have to scalarize the op.
644 // Notice that the operation may be 'promoted' which means that it is
645 // 'bitcasted' to another type which is handled.
646 // Also, we need to be able to construct a splat vector using BUILD_VECTOR.
647 if (TLI.getOperationAction(ISD::AND, VT) == TargetLowering::Expand ||
648 TLI.getOperationAction(ISD::XOR, VT) == TargetLowering::Expand ||
649 TLI.getOperationAction(ISD::OR, VT) == TargetLowering::Expand ||
650 TLI.getOperationAction(ISD::BUILD_VECTOR, VT) == TargetLowering::Expand)
651 return DAG.UnrollVectorOp(Op.getNode());
653 // Generate a mask operand.
654 EVT MaskTy = VT.changeVectorElementTypeToInteger();
656 // What is the size of each element in the vector mask.
657 EVT BitTy = MaskTy.getScalarType();
659 Mask = DAG.getSelect(DL, BitTy, Mask,
660 DAG.getConstant(APInt::getAllOnesValue(BitTy.getSizeInBits()), BitTy),
661 DAG.getConstant(0, BitTy));
663 // Broadcast the mask so that the entire vector is all-one or all zero.
664 SmallVector<SDValue, 8> Ops(NumElem, Mask);
665 Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, MaskTy, Ops);
667 // Bitcast the operands to be the same type as the mask.
668 // This is needed when we select between FP types because
669 // the mask is a vector of integers.
670 Op1 = DAG.getNode(ISD::BITCAST, DL, MaskTy, Op1);
671 Op2 = DAG.getNode(ISD::BITCAST, DL, MaskTy, Op2);
673 SDValue AllOnes = DAG.getConstant(
674 APInt::getAllOnesValue(BitTy.getSizeInBits()), MaskTy);
675 SDValue NotMask = DAG.getNode(ISD::XOR, DL, MaskTy, Mask, AllOnes);
677 Op1 = DAG.getNode(ISD::AND, DL, MaskTy, Op1, Mask);
678 Op2 = DAG.getNode(ISD::AND, DL, MaskTy, Op2, NotMask);
679 SDValue Val = DAG.getNode(ISD::OR, DL, MaskTy, Op1, Op2);
680 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Val);
683 SDValue VectorLegalizer::ExpandSEXTINREG(SDValue Op) {
684 EVT VT = Op.getValueType();
686 // Make sure that the SRA and SHL instructions are available.
687 if (TLI.getOperationAction(ISD::SRA, VT) == TargetLowering::Expand ||
688 TLI.getOperationAction(ISD::SHL, VT) == TargetLowering::Expand)
689 return DAG.UnrollVectorOp(Op.getNode());
692 EVT OrigTy = cast<VTSDNode>(Op->getOperand(1))->getVT();
694 unsigned BW = VT.getScalarType().getSizeInBits();
695 unsigned OrigBW = OrigTy.getScalarType().getSizeInBits();
696 SDValue ShiftSz = DAG.getConstant(BW - OrigBW, VT);
698 Op = Op.getOperand(0);
699 Op = DAG.getNode(ISD::SHL, DL, VT, Op, ShiftSz);
700 return DAG.getNode(ISD::SRA, DL, VT, Op, ShiftSz);
703 SDValue VectorLegalizer::ExpandBSWAP(SDValue Op) {
704 EVT VT = Op.getValueType();
706 // Generate a byte wise shuffle mask for the BSWAP.
707 SmallVector<int, 16> ShuffleMask;
708 int ScalarSizeInBytes = VT.getScalarSizeInBits() / 8;
709 for (int I = 0, E = VT.getVectorNumElements(); I != E; ++I)
710 for (int J = ScalarSizeInBytes - 1; J >= 0; --J)
711 ShuffleMask.push_back((I * ScalarSizeInBytes) + J);
713 EVT ByteVT = EVT::getVectorVT(*DAG.getContext(), MVT::i8, ShuffleMask.size());
715 // Only emit a shuffle if the mask is legal.
716 if (!TLI.isShuffleMaskLegal(ShuffleMask, ByteVT))
717 return DAG.UnrollVectorOp(Op.getNode());
720 Op = DAG.getNode(ISD::BITCAST, DL, ByteVT, Op.getOperand(0));
721 Op = DAG.getVectorShuffle(ByteVT, DL, Op, DAG.getUNDEF(ByteVT),
723 return DAG.getNode(ISD::BITCAST, DL, VT, Op);
726 SDValue VectorLegalizer::ExpandVSELECT(SDValue Op) {
727 // Implement VSELECT in terms of XOR, AND, OR
728 // on platforms which do not support blend natively.
731 SDValue Mask = Op.getOperand(0);
732 SDValue Op1 = Op.getOperand(1);
733 SDValue Op2 = Op.getOperand(2);
735 EVT VT = Mask.getValueType();
737 // If we can't even use the basic vector operations of
738 // AND,OR,XOR, we will have to scalarize the op.
739 // Notice that the operation may be 'promoted' which means that it is
740 // 'bitcasted' to another type which is handled.
741 // This operation also isn't safe with AND, OR, XOR when the boolean
742 // type is 0/1 as we need an all ones vector constant to mask with.
743 // FIXME: Sign extend 1 to all ones if thats legal on the target.
744 if (TLI.getOperationAction(ISD::AND, VT) == TargetLowering::Expand ||
745 TLI.getOperationAction(ISD::XOR, VT) == TargetLowering::Expand ||
746 TLI.getOperationAction(ISD::OR, VT) == TargetLowering::Expand ||
747 TLI.getBooleanContents(true) !=
748 TargetLowering::ZeroOrNegativeOneBooleanContent)
749 return DAG.UnrollVectorOp(Op.getNode());
751 // If the mask and the type are different sizes, unroll the vector op. This
752 // can occur when getSetCCResultType returns something that is different in
753 // size from the operand types. For example, v4i8 = select v4i32, v4i8, v4i8.
754 if (VT.getSizeInBits() != Op1.getValueType().getSizeInBits())
755 return DAG.UnrollVectorOp(Op.getNode());
757 // Bitcast the operands to be the same type as the mask.
758 // This is needed when we select between FP types because
759 // the mask is a vector of integers.
760 Op1 = DAG.getNode(ISD::BITCAST, DL, VT, Op1);
761 Op2 = DAG.getNode(ISD::BITCAST, DL, VT, Op2);
763 SDValue AllOnes = DAG.getConstant(
764 APInt::getAllOnesValue(VT.getScalarType().getSizeInBits()), VT);
765 SDValue NotMask = DAG.getNode(ISD::XOR, DL, VT, Mask, AllOnes);
767 Op1 = DAG.getNode(ISD::AND, DL, VT, Op1, Mask);
768 Op2 = DAG.getNode(ISD::AND, DL, VT, Op2, NotMask);
769 SDValue Val = DAG.getNode(ISD::OR, DL, VT, Op1, Op2);
770 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Val);
773 SDValue VectorLegalizer::ExpandUINT_TO_FLOAT(SDValue Op) {
774 EVT VT = Op.getOperand(0).getValueType();
777 // Make sure that the SINT_TO_FP and SRL instructions are available.
778 if (TLI.getOperationAction(ISD::SINT_TO_FP, VT) == TargetLowering::Expand ||
779 TLI.getOperationAction(ISD::SRL, VT) == TargetLowering::Expand)
780 return DAG.UnrollVectorOp(Op.getNode());
782 EVT SVT = VT.getScalarType();
783 assert((SVT.getSizeInBits() == 64 || SVT.getSizeInBits() == 32) &&
784 "Elements in vector-UINT_TO_FP must be 32 or 64 bits wide");
786 unsigned BW = SVT.getSizeInBits();
787 SDValue HalfWord = DAG.getConstant(BW/2, VT);
789 // Constants to clear the upper part of the word.
790 // Notice that we can also use SHL+SHR, but using a constant is slightly
792 uint64_t HWMask = (SVT.getSizeInBits()==64)?0x00000000FFFFFFFF:0x0000FFFF;
793 SDValue HalfWordMask = DAG.getConstant(HWMask, VT);
795 // Two to the power of half-word-size.
796 SDValue TWOHW = DAG.getConstantFP((1<<(BW/2)), Op.getValueType());
798 // Clear upper part of LO, lower HI
799 SDValue HI = DAG.getNode(ISD::SRL, DL, VT, Op.getOperand(0), HalfWord);
800 SDValue LO = DAG.getNode(ISD::AND, DL, VT, Op.getOperand(0), HalfWordMask);
802 // Convert hi and lo to floats
803 // Convert the hi part back to the upper values
804 SDValue fHI = DAG.getNode(ISD::SINT_TO_FP, DL, Op.getValueType(), HI);
805 fHI = DAG.getNode(ISD::FMUL, DL, Op.getValueType(), fHI, TWOHW);
806 SDValue fLO = DAG.getNode(ISD::SINT_TO_FP, DL, Op.getValueType(), LO);
808 // Add the two halves
809 return DAG.getNode(ISD::FADD, DL, Op.getValueType(), fHI, fLO);
813 SDValue VectorLegalizer::ExpandFNEG(SDValue Op) {
814 if (TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType())) {
815 SDValue Zero = DAG.getConstantFP(-0.0, Op.getValueType());
816 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
817 Zero, Op.getOperand(0));
819 return DAG.UnrollVectorOp(Op.getNode());
822 SDValue VectorLegalizer::UnrollVSETCC(SDValue Op) {
823 EVT VT = Op.getValueType();
824 unsigned NumElems = VT.getVectorNumElements();
825 EVT EltVT = VT.getVectorElementType();
826 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1), CC = Op.getOperand(2);
827 EVT TmpEltVT = LHS.getValueType().getVectorElementType();
829 SmallVector<SDValue, 8> Ops(NumElems);
830 for (unsigned i = 0; i < NumElems; ++i) {
831 SDValue LHSElem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, LHS,
832 DAG.getConstant(i, TLI.getVectorIdxTy()));
833 SDValue RHSElem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, RHS,
834 DAG.getConstant(i, TLI.getVectorIdxTy()));
835 Ops[i] = DAG.getNode(ISD::SETCC, dl,
836 TLI.getSetCCResultType(*DAG.getContext(), TmpEltVT),
837 LHSElem, RHSElem, CC);
838 Ops[i] = DAG.getSelect(dl, EltVT, Ops[i],
839 DAG.getConstant(APInt::getAllOnesValue
840 (EltVT.getSizeInBits()), EltVT),
841 DAG.getConstant(0, EltVT));
843 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
848 bool SelectionDAG::LegalizeVectors() {
849 return VectorLegalizer(*this).Run();