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 /// LegalizedNodes - For nodes that are of legal width, and that have more
41 /// than one use, this map indicates what regularized operand to use. This
42 /// allows us to avoid legalizing the same thing more than once.
43 SmallDenseMap<SDValue, SDValue, 64> LegalizedNodes;
45 // 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 // Legalizes the given node
54 SDValue LegalizeOp(SDValue Op);
55 // Assuming the node is legal, "legalize" the results
56 SDValue TranslateLegalizeResults(SDValue Op, SDValue Result);
57 // Implements unrolling a VSETCC.
58 SDValue UnrollVSETCC(SDValue Op);
59 // Implements expansion for FNEG; falls back to UnrollVectorOp if FSUB
61 // Implements expansion for UINT_TO_FLOAT; falls back to UnrollVectorOp if
62 // SINT_TO_FLOAT and SHR on vectors isn't legal.
63 SDValue ExpandUINT_TO_FLOAT(SDValue Op);
64 // Implement expansion for SIGN_EXTEND_INREG using SRL and SRA.
65 SDValue ExpandSEXTINREG(SDValue Op);
66 // Implement vselect in terms of XOR, AND, OR when blend is not supported
68 SDValue ExpandVSELECT(SDValue Op);
69 SDValue ExpandSELECT(SDValue Op);
70 SDValue ExpandLoad(SDValue Op);
71 SDValue ExpandStore(SDValue Op);
72 SDValue ExpandFNEG(SDValue Op);
73 // Implements vector promotion; this is essentially just bitcasting the
74 // operands to a different type and bitcasting the result back to the
76 SDValue PromoteVectorOp(SDValue Op);
77 // Implements [SU]INT_TO_FP vector promotion; this is a [zs]ext of the input
78 // operand to the next size up.
79 SDValue PromoteVectorOpINT_TO_FP(SDValue Op);
80 // Implements FP_TO_[SU]INT vector promotion of the result type; it is
81 // promoted to the next size up integer type. The result is then truncated
82 // back to the original type.
83 SDValue PromoteVectorOpFP_TO_INT(SDValue Op, bool isSigned);
87 VectorLegalizer(SelectionDAG& dag) :
88 DAG(dag), TLI(dag.getTargetLoweringInfo()), Changed(false) {}
91 bool VectorLegalizer::Run() {
92 // Before we start legalizing vector nodes, check if there are any vectors.
93 bool HasVectors = false;
94 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
95 E = std::prev(DAG.allnodes_end()); I != std::next(E); ++I) {
96 // Check if the values of the nodes contain vectors. We don't need to check
97 // the operands because we are going to check their values at some point.
98 for (SDNode::value_iterator J = I->value_begin(), E = I->value_end();
100 HasVectors |= J->isVector();
102 // If we found a vector node we can start the legalization.
107 // If this basic block has no vectors then no need to legalize vectors.
111 // The legalize process is inherently a bottom-up recursive process (users
112 // legalize their uses before themselves). Given infinite stack space, we
113 // could just start legalizing on the root and traverse the whole graph. In
114 // practice however, this causes us to run out of stack space on large basic
115 // blocks. To avoid this problem, compute an ordering of the nodes where each
116 // node is only legalized after all of its operands are legalized.
117 DAG.AssignTopologicalOrder();
118 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
119 E = std::prev(DAG.allnodes_end()); I != std::next(E); ++I)
120 LegalizeOp(SDValue(I, 0));
122 // Finally, it's possible the root changed. Get the new root.
123 SDValue OldRoot = DAG.getRoot();
124 assert(LegalizedNodes.count(OldRoot) && "Root didn't get legalized?");
125 DAG.setRoot(LegalizedNodes[OldRoot]);
127 LegalizedNodes.clear();
129 // Remove dead nodes now.
130 DAG.RemoveDeadNodes();
135 SDValue VectorLegalizer::TranslateLegalizeResults(SDValue Op, SDValue Result) {
136 // Generic legalization: just pass the operand through.
137 for (unsigned i = 0, e = Op.getNode()->getNumValues(); i != e; ++i)
138 AddLegalizedOperand(Op.getValue(i), Result.getValue(i));
139 return Result.getValue(Op.getResNo());
142 SDValue VectorLegalizer::LegalizeOp(SDValue Op) {
143 // Note that LegalizeOp may be reentered even from single-use nodes, which
144 // means that we always must cache transformed nodes.
145 DenseMap<SDValue, SDValue>::iterator I = LegalizedNodes.find(Op);
146 if (I != LegalizedNodes.end()) return I->second;
148 SDNode* Node = Op.getNode();
150 // Legalize the operands
151 SmallVector<SDValue, 8> Ops;
152 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
153 Ops.push_back(LegalizeOp(Node->getOperand(i)));
155 SDValue Result = SDValue(DAG.UpdateNodeOperands(Op.getNode(), Ops), 0);
157 if (Op.getOpcode() == ISD::LOAD) {
158 LoadSDNode *LD = cast<LoadSDNode>(Op.getNode());
159 ISD::LoadExtType ExtType = LD->getExtensionType();
160 if (LD->getMemoryVT().isVector() && ExtType != ISD::NON_EXTLOAD) {
161 if (TLI.isLoadExtLegal(LD->getExtensionType(), LD->getMemoryVT()))
162 return TranslateLegalizeResults(Op, Result);
164 return LegalizeOp(ExpandLoad(Op));
166 } else if (Op.getOpcode() == ISD::STORE) {
167 StoreSDNode *ST = cast<StoreSDNode>(Op.getNode());
168 EVT StVT = ST->getMemoryVT();
169 MVT ValVT = ST->getValue().getSimpleValueType();
170 if (StVT.isVector() && ST->isTruncatingStore())
171 switch (TLI.getTruncStoreAction(ValVT, StVT.getSimpleVT())) {
172 default: llvm_unreachable("This action is not supported yet!");
173 case TargetLowering::Legal:
174 return TranslateLegalizeResults(Op, Result);
175 case TargetLowering::Custom:
177 return TranslateLegalizeResults(Op, TLI.LowerOperation(Result, DAG));
178 case TargetLowering::Expand:
180 return LegalizeOp(ExpandStore(Op));
184 bool HasVectorValue = false;
185 for (SDNode::value_iterator J = Node->value_begin(), E = Node->value_end();
188 HasVectorValue |= J->isVector();
190 return TranslateLegalizeResults(Op, Result);
193 switch (Op.getOpcode()) {
195 return TranslateLegalizeResults(Op, Result);
219 case ISD::CTLZ_ZERO_UNDEF:
220 case ISD::CTTZ_ZERO_UNDEF:
226 case ISD::ZERO_EXTEND:
227 case ISD::ANY_EXTEND:
229 case ISD::SIGN_EXTEND:
230 case ISD::FP_TO_SINT:
231 case ISD::FP_TO_UINT:
248 case ISD::FNEARBYINT:
254 case ISD::SIGN_EXTEND_INREG:
255 QueryType = Node->getValueType(0);
257 case ISD::FP_ROUND_INREG:
258 QueryType = cast<VTSDNode>(Node->getOperand(1))->getVT();
260 case ISD::SINT_TO_FP:
261 case ISD::UINT_TO_FP:
262 QueryType = Node->getOperand(0).getValueType();
266 switch (TLI.getOperationAction(Node->getOpcode(), QueryType)) {
267 case TargetLowering::Promote:
268 switch (Op.getOpcode()) {
270 // "Promote" the operation by bitcasting
271 Result = PromoteVectorOp(Op);
274 case ISD::SINT_TO_FP:
275 case ISD::UINT_TO_FP:
276 // "Promote" the operation by extending the operand.
277 Result = PromoteVectorOpINT_TO_FP(Op);
280 case ISD::FP_TO_UINT:
281 case ISD::FP_TO_SINT:
282 // Promote the operation by extending the operand.
283 Result = PromoteVectorOpFP_TO_INT(Op, Op->getOpcode() == ISD::FP_TO_SINT);
288 case TargetLowering::Legal: break;
289 case TargetLowering::Custom: {
290 SDValue Tmp1 = TLI.LowerOperation(Op, DAG);
291 if (Tmp1.getNode()) {
297 case TargetLowering::Expand:
298 if (Node->getOpcode() == ISD::SIGN_EXTEND_INREG)
299 Result = ExpandSEXTINREG(Op);
300 else if (Node->getOpcode() == ISD::VSELECT)
301 Result = ExpandVSELECT(Op);
302 else if (Node->getOpcode() == ISD::SELECT)
303 Result = ExpandSELECT(Op);
304 else if (Node->getOpcode() == ISD::UINT_TO_FP)
305 Result = ExpandUINT_TO_FLOAT(Op);
306 else if (Node->getOpcode() == ISD::FNEG)
307 Result = ExpandFNEG(Op);
308 else if (Node->getOpcode() == ISD::SETCC)
309 Result = UnrollVSETCC(Op);
311 Result = DAG.UnrollVectorOp(Op.getNode());
315 // Make sure that the generated code is itself legal.
317 Result = LegalizeOp(Result);
321 // Note that LegalizeOp may be reentered even from single-use nodes, which
322 // means that we always must cache transformed nodes.
323 AddLegalizedOperand(Op, Result);
327 SDValue VectorLegalizer::PromoteVectorOp(SDValue Op) {
328 // Vector "promotion" is basically just bitcasting and doing the operation
329 // in a different type. For example, x86 promotes ISD::AND on v2i32 to
331 MVT VT = Op.getSimpleValueType();
332 assert(Op.getNode()->getNumValues() == 1 &&
333 "Can't promote a vector with multiple results!");
334 MVT NVT = TLI.getTypeToPromoteTo(Op.getOpcode(), VT);
336 SmallVector<SDValue, 4> Operands(Op.getNumOperands());
338 for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
339 if (Op.getOperand(j).getValueType().isVector())
340 Operands[j] = DAG.getNode(ISD::BITCAST, dl, NVT, Op.getOperand(j));
342 Operands[j] = Op.getOperand(j);
345 Op = DAG.getNode(Op.getOpcode(), dl, NVT, Operands);
347 return DAG.getNode(ISD::BITCAST, dl, VT, Op);
350 SDValue VectorLegalizer::PromoteVectorOpINT_TO_FP(SDValue Op) {
351 // INT_TO_FP operations may require the input operand be promoted even
352 // when the type is otherwise legal.
353 EVT VT = Op.getOperand(0).getValueType();
354 assert(Op.getNode()->getNumValues() == 1 &&
355 "Can't promote a vector with multiple results!");
357 // Normal getTypeToPromoteTo() doesn't work here, as that will promote
358 // by widening the vector w/ the same element width and twice the number
359 // of elements. We want the other way around, the same number of elements,
360 // each twice the width.
362 // Increase the bitwidth of the element to the next pow-of-two
363 // (which is greater than 8 bits).
365 EVT NVT = VT.widenIntegerVectorElementType(*DAG.getContext());
366 assert(NVT.isSimple() && "Promoting to a non-simple vector type!");
368 SmallVector<SDValue, 4> Operands(Op.getNumOperands());
370 unsigned Opc = Op.getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND :
372 for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
373 if (Op.getOperand(j).getValueType().isVector())
374 Operands[j] = DAG.getNode(Opc, dl, NVT, Op.getOperand(j));
376 Operands[j] = Op.getOperand(j);
379 return DAG.getNode(Op.getOpcode(), dl, Op.getValueType(), Operands);
382 // For FP_TO_INT we promote the result type to a vector type with wider
383 // elements and then truncate the result. This is different from the default
384 // PromoteVector which uses bitcast to promote thus assumning that the
385 // promoted vector type has the same overall size.
386 SDValue VectorLegalizer::PromoteVectorOpFP_TO_INT(SDValue Op, bool isSigned) {
387 assert(Op.getNode()->getNumValues() == 1 &&
388 "Can't promote a vector with multiple results!");
389 EVT VT = Op.getValueType();
394 NewVT = VT.widenIntegerVectorElementType(*DAG.getContext());
395 assert(NewVT.isSimple() && "Promoting to a non-simple vector type!");
396 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NewVT)) {
397 NewOpc = ISD::FP_TO_SINT;
400 if (!isSigned && TLI.isOperationLegalOrCustom(ISD::FP_TO_UINT, NewVT)) {
401 NewOpc = ISD::FP_TO_UINT;
407 SDValue promoted = DAG.getNode(NewOpc, SDLoc(Op), NewVT, Op.getOperand(0));
408 return DAG.getNode(ISD::TRUNCATE, SDLoc(Op), VT, promoted);
412 SDValue VectorLegalizer::ExpandLoad(SDValue Op) {
414 LoadSDNode *LD = cast<LoadSDNode>(Op.getNode());
415 SDValue Chain = LD->getChain();
416 SDValue BasePTR = LD->getBasePtr();
417 EVT SrcVT = LD->getMemoryVT();
418 ISD::LoadExtType ExtType = LD->getExtensionType();
420 SmallVector<SDValue, 8> Vals;
421 SmallVector<SDValue, 8> LoadChains;
422 unsigned NumElem = SrcVT.getVectorNumElements();
424 EVT SrcEltVT = SrcVT.getScalarType();
425 EVT DstEltVT = Op.getNode()->getValueType(0).getScalarType();
427 if (SrcVT.getVectorNumElements() > 1 && !SrcEltVT.isByteSized()) {
428 // When elements in a vector is not byte-addressable, we cannot directly
429 // load each element by advancing pointer, which could only address bytes.
430 // Instead, we load all significant words, mask bits off, and concatenate
431 // them to form each element. Finally, they are extended to destination
432 // scalar type to build the destination vector.
433 EVT WideVT = TLI.getPointerTy();
435 assert(WideVT.isRound() &&
436 "Could not handle the sophisticated case when the widest integer is"
438 assert(WideVT.bitsGE(SrcEltVT) &&
439 "Type is not legalized?");
441 unsigned WideBytes = WideVT.getStoreSize();
443 unsigned RemainingBytes = SrcVT.getStoreSize();
444 SmallVector<SDValue, 8> LoadVals;
446 while (RemainingBytes > 0) {
448 unsigned LoadBytes = WideBytes;
450 if (RemainingBytes >= LoadBytes) {
451 ScalarLoad = DAG.getLoad(WideVT, dl, Chain, BasePTR,
452 LD->getPointerInfo().getWithOffset(Offset),
453 LD->isVolatile(), LD->isNonTemporal(),
454 LD->isInvariant(), LD->getAlignment(),
458 while (RemainingBytes < LoadBytes) {
459 LoadBytes >>= 1; // Reduce the load size by half.
460 LoadVT = EVT::getIntegerVT(*DAG.getContext(), LoadBytes << 3);
462 ScalarLoad = DAG.getExtLoad(ISD::EXTLOAD, dl, WideVT, Chain, BasePTR,
463 LD->getPointerInfo().getWithOffset(Offset),
464 LoadVT, LD->isVolatile(),
465 LD->isNonTemporal(), LD->getAlignment(),
469 RemainingBytes -= LoadBytes;
471 BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR,
472 DAG.getConstant(LoadBytes, BasePTR.getValueType()));
474 LoadVals.push_back(ScalarLoad.getValue(0));
475 LoadChains.push_back(ScalarLoad.getValue(1));
478 // Extract bits, pack and extend/trunc them into destination type.
479 unsigned SrcEltBits = SrcEltVT.getSizeInBits();
480 SDValue SrcEltBitMask = DAG.getConstant((1U << SrcEltBits) - 1, WideVT);
482 unsigned BitOffset = 0;
483 unsigned WideIdx = 0;
484 unsigned WideBits = WideVT.getSizeInBits();
486 for (unsigned Idx = 0; Idx != NumElem; ++Idx) {
487 SDValue Lo, Hi, ShAmt;
489 if (BitOffset < WideBits) {
490 ShAmt = DAG.getConstant(BitOffset, TLI.getShiftAmountTy(WideVT));
491 Lo = DAG.getNode(ISD::SRL, dl, WideVT, LoadVals[WideIdx], ShAmt);
492 Lo = DAG.getNode(ISD::AND, dl, WideVT, Lo, SrcEltBitMask);
495 BitOffset += SrcEltBits;
496 if (BitOffset >= WideBits) {
500 ShAmt = DAG.getConstant(SrcEltBits - Offset,
501 TLI.getShiftAmountTy(WideVT));
502 Hi = DAG.getNode(ISD::SHL, dl, WideVT, LoadVals[WideIdx], ShAmt);
503 Hi = DAG.getNode(ISD::AND, dl, WideVT, Hi, SrcEltBitMask);
508 Lo = DAG.getNode(ISD::OR, dl, WideVT, Lo, Hi);
511 default: llvm_unreachable("Unknown extended-load op!");
513 Lo = DAG.getAnyExtOrTrunc(Lo, dl, DstEltVT);
516 Lo = DAG.getZExtOrTrunc(Lo, dl, DstEltVT);
519 ShAmt = DAG.getConstant(WideBits - SrcEltBits,
520 TLI.getShiftAmountTy(WideVT));
521 Lo = DAG.getNode(ISD::SHL, dl, WideVT, Lo, ShAmt);
522 Lo = DAG.getNode(ISD::SRA, dl, WideVT, Lo, ShAmt);
523 Lo = DAG.getSExtOrTrunc(Lo, dl, DstEltVT);
529 unsigned Stride = SrcVT.getScalarType().getSizeInBits()/8;
531 for (unsigned Idx=0; Idx<NumElem; Idx++) {
532 SDValue ScalarLoad = DAG.getExtLoad(ExtType, dl,
533 Op.getNode()->getValueType(0).getScalarType(),
534 Chain, BasePTR, LD->getPointerInfo().getWithOffset(Idx * Stride),
535 SrcVT.getScalarType(),
536 LD->isVolatile(), LD->isNonTemporal(),
537 LD->getAlignment(), LD->getTBAAInfo());
539 BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR,
540 DAG.getConstant(Stride, BasePTR.getValueType()));
542 Vals.push_back(ScalarLoad.getValue(0));
543 LoadChains.push_back(ScalarLoad.getValue(1));
547 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, LoadChains);
548 SDValue Value = DAG.getNode(ISD::BUILD_VECTOR, dl,
549 Op.getNode()->getValueType(0), Vals);
551 AddLegalizedOperand(Op.getValue(0), Value);
552 AddLegalizedOperand(Op.getValue(1), NewChain);
554 return (Op.getResNo() ? NewChain : Value);
557 SDValue VectorLegalizer::ExpandStore(SDValue Op) {
559 StoreSDNode *ST = cast<StoreSDNode>(Op.getNode());
560 SDValue Chain = ST->getChain();
561 SDValue BasePTR = ST->getBasePtr();
562 SDValue Value = ST->getValue();
563 EVT StVT = ST->getMemoryVT();
565 unsigned Alignment = ST->getAlignment();
566 bool isVolatile = ST->isVolatile();
567 bool isNonTemporal = ST->isNonTemporal();
568 const MDNode *TBAAInfo = ST->getTBAAInfo();
570 unsigned NumElem = StVT.getVectorNumElements();
571 // The type of the data we want to save
572 EVT RegVT = Value.getValueType();
573 EVT RegSclVT = RegVT.getScalarType();
574 // The type of data as saved in memory.
575 EVT MemSclVT = StVT.getScalarType();
577 // Cast floats into integers
578 unsigned ScalarSize = MemSclVT.getSizeInBits();
580 // Round odd types to the next pow of two.
581 if (!isPowerOf2_32(ScalarSize))
582 ScalarSize = NextPowerOf2(ScalarSize);
584 // Store Stride in bytes
585 unsigned Stride = ScalarSize/8;
586 // Extract each of the elements from the original vector
587 // and save them into memory individually.
588 SmallVector<SDValue, 8> Stores;
589 for (unsigned Idx = 0; Idx < NumElem; Idx++) {
590 SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
591 RegSclVT, Value, DAG.getConstant(Idx, TLI.getVectorIdxTy()));
593 // This scalar TruncStore may be illegal, but we legalize it later.
594 SDValue Store = DAG.getTruncStore(Chain, dl, Ex, BasePTR,
595 ST->getPointerInfo().getWithOffset(Idx*Stride), MemSclVT,
596 isVolatile, isNonTemporal, Alignment, TBAAInfo);
598 BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR,
599 DAG.getConstant(Stride, BasePTR.getValueType()));
601 Stores.push_back(Store);
603 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
604 AddLegalizedOperand(Op, TF);
608 SDValue VectorLegalizer::ExpandSELECT(SDValue Op) {
609 // Lower a select instruction where the condition is a scalar and the
610 // operands are vectors. Lower this select to VSELECT and implement it
611 // using XOR AND OR. The selector bit is broadcasted.
612 EVT VT = Op.getValueType();
615 SDValue Mask = Op.getOperand(0);
616 SDValue Op1 = Op.getOperand(1);
617 SDValue Op2 = Op.getOperand(2);
619 assert(VT.isVector() && !Mask.getValueType().isVector()
620 && Op1.getValueType() == Op2.getValueType() && "Invalid type");
622 unsigned NumElem = VT.getVectorNumElements();
624 // If we can't even use the basic vector operations of
625 // AND,OR,XOR, we will have to scalarize the op.
626 // Notice that the operation may be 'promoted' which means that it is
627 // 'bitcasted' to another type which is handled.
628 // Also, we need to be able to construct a splat vector using BUILD_VECTOR.
629 if (TLI.getOperationAction(ISD::AND, VT) == TargetLowering::Expand ||
630 TLI.getOperationAction(ISD::XOR, VT) == TargetLowering::Expand ||
631 TLI.getOperationAction(ISD::OR, VT) == TargetLowering::Expand ||
632 TLI.getOperationAction(ISD::BUILD_VECTOR, VT) == TargetLowering::Expand)
633 return DAG.UnrollVectorOp(Op.getNode());
635 // Generate a mask operand.
636 EVT MaskTy = VT.changeVectorElementTypeToInteger();
638 // What is the size of each element in the vector mask.
639 EVT BitTy = MaskTy.getScalarType();
641 Mask = DAG.getSelect(DL, BitTy, Mask,
642 DAG.getConstant(APInt::getAllOnesValue(BitTy.getSizeInBits()), BitTy),
643 DAG.getConstant(0, BitTy));
645 // Broadcast the mask so that the entire vector is all-one or all zero.
646 SmallVector<SDValue, 8> Ops(NumElem, Mask);
647 Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, MaskTy, Ops);
649 // Bitcast the operands to be the same type as the mask.
650 // This is needed when we select between FP types because
651 // the mask is a vector of integers.
652 Op1 = DAG.getNode(ISD::BITCAST, DL, MaskTy, Op1);
653 Op2 = DAG.getNode(ISD::BITCAST, DL, MaskTy, Op2);
655 SDValue AllOnes = DAG.getConstant(
656 APInt::getAllOnesValue(BitTy.getSizeInBits()), MaskTy);
657 SDValue NotMask = DAG.getNode(ISD::XOR, DL, MaskTy, Mask, AllOnes);
659 Op1 = DAG.getNode(ISD::AND, DL, MaskTy, Op1, Mask);
660 Op2 = DAG.getNode(ISD::AND, DL, MaskTy, Op2, NotMask);
661 SDValue Val = DAG.getNode(ISD::OR, DL, MaskTy, Op1, Op2);
662 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Val);
665 SDValue VectorLegalizer::ExpandSEXTINREG(SDValue Op) {
666 EVT VT = Op.getValueType();
668 // Make sure that the SRA and SHL instructions are available.
669 if (TLI.getOperationAction(ISD::SRA, VT) == TargetLowering::Expand ||
670 TLI.getOperationAction(ISD::SHL, VT) == TargetLowering::Expand)
671 return DAG.UnrollVectorOp(Op.getNode());
674 EVT OrigTy = cast<VTSDNode>(Op->getOperand(1))->getVT();
676 unsigned BW = VT.getScalarType().getSizeInBits();
677 unsigned OrigBW = OrigTy.getScalarType().getSizeInBits();
678 SDValue ShiftSz = DAG.getConstant(BW - OrigBW, VT);
680 Op = Op.getOperand(0);
681 Op = DAG.getNode(ISD::SHL, DL, VT, Op, ShiftSz);
682 return DAG.getNode(ISD::SRA, DL, VT, Op, ShiftSz);
685 SDValue VectorLegalizer::ExpandVSELECT(SDValue Op) {
686 // Implement VSELECT in terms of XOR, AND, OR
687 // on platforms which do not support blend natively.
690 SDValue Mask = Op.getOperand(0);
691 SDValue Op1 = Op.getOperand(1);
692 SDValue Op2 = Op.getOperand(2);
694 EVT VT = Mask.getValueType();
696 // If we can't even use the basic vector operations of
697 // AND,OR,XOR, we will have to scalarize the op.
698 // Notice that the operation may be 'promoted' which means that it is
699 // 'bitcasted' to another type which is handled.
700 // This operation also isn't safe with AND, OR, XOR when the boolean
701 // type is 0/1 as we need an all ones vector constant to mask with.
702 // FIXME: Sign extend 1 to all ones if thats legal on the target.
703 if (TLI.getOperationAction(ISD::AND, VT) == TargetLowering::Expand ||
704 TLI.getOperationAction(ISD::XOR, VT) == TargetLowering::Expand ||
705 TLI.getOperationAction(ISD::OR, VT) == TargetLowering::Expand ||
706 TLI.getBooleanContents(true) !=
707 TargetLowering::ZeroOrNegativeOneBooleanContent)
708 return DAG.UnrollVectorOp(Op.getNode());
710 // If the mask and the type are different sizes, unroll the vector op. This
711 // can occur when getSetCCResultType returns something that is different in
712 // size from the operand types. For example, v4i8 = select v4i32, v4i8, v4i8.
713 if (VT.getSizeInBits() != Op1.getValueType().getSizeInBits())
714 return DAG.UnrollVectorOp(Op.getNode());
716 // Bitcast the operands to be the same type as the mask.
717 // This is needed when we select between FP types because
718 // the mask is a vector of integers.
719 Op1 = DAG.getNode(ISD::BITCAST, DL, VT, Op1);
720 Op2 = DAG.getNode(ISD::BITCAST, DL, VT, Op2);
722 SDValue AllOnes = DAG.getConstant(
723 APInt::getAllOnesValue(VT.getScalarType().getSizeInBits()), VT);
724 SDValue NotMask = DAG.getNode(ISD::XOR, DL, VT, Mask, AllOnes);
726 Op1 = DAG.getNode(ISD::AND, DL, VT, Op1, Mask);
727 Op2 = DAG.getNode(ISD::AND, DL, VT, Op2, NotMask);
728 SDValue Val = DAG.getNode(ISD::OR, DL, VT, Op1, Op2);
729 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Val);
732 SDValue VectorLegalizer::ExpandUINT_TO_FLOAT(SDValue Op) {
733 EVT VT = Op.getOperand(0).getValueType();
736 // Make sure that the SINT_TO_FP and SRL instructions are available.
737 if (TLI.getOperationAction(ISD::SINT_TO_FP, VT) == TargetLowering::Expand ||
738 TLI.getOperationAction(ISD::SRL, VT) == TargetLowering::Expand)
739 return DAG.UnrollVectorOp(Op.getNode());
741 EVT SVT = VT.getScalarType();
742 assert((SVT.getSizeInBits() == 64 || SVT.getSizeInBits() == 32) &&
743 "Elements in vector-UINT_TO_FP must be 32 or 64 bits wide");
745 unsigned BW = SVT.getSizeInBits();
746 SDValue HalfWord = DAG.getConstant(BW/2, VT);
748 // Constants to clear the upper part of the word.
749 // Notice that we can also use SHL+SHR, but using a constant is slightly
751 uint64_t HWMask = (SVT.getSizeInBits()==64)?0x00000000FFFFFFFF:0x0000FFFF;
752 SDValue HalfWordMask = DAG.getConstant(HWMask, VT);
754 // Two to the power of half-word-size.
755 SDValue TWOHW = DAG.getConstantFP((1<<(BW/2)), Op.getValueType());
757 // Clear upper part of LO, lower HI
758 SDValue HI = DAG.getNode(ISD::SRL, DL, VT, Op.getOperand(0), HalfWord);
759 SDValue LO = DAG.getNode(ISD::AND, DL, VT, Op.getOperand(0), HalfWordMask);
761 // Convert hi and lo to floats
762 // Convert the hi part back to the upper values
763 SDValue fHI = DAG.getNode(ISD::SINT_TO_FP, DL, Op.getValueType(), HI);
764 fHI = DAG.getNode(ISD::FMUL, DL, Op.getValueType(), fHI, TWOHW);
765 SDValue fLO = DAG.getNode(ISD::SINT_TO_FP, DL, Op.getValueType(), LO);
767 // Add the two halves
768 return DAG.getNode(ISD::FADD, DL, Op.getValueType(), fHI, fLO);
772 SDValue VectorLegalizer::ExpandFNEG(SDValue Op) {
773 if (TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType())) {
774 SDValue Zero = DAG.getConstantFP(-0.0, Op.getValueType());
775 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
776 Zero, Op.getOperand(0));
778 return DAG.UnrollVectorOp(Op.getNode());
781 SDValue VectorLegalizer::UnrollVSETCC(SDValue Op) {
782 EVT VT = Op.getValueType();
783 unsigned NumElems = VT.getVectorNumElements();
784 EVT EltVT = VT.getVectorElementType();
785 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1), CC = Op.getOperand(2);
786 EVT TmpEltVT = LHS.getValueType().getVectorElementType();
788 SmallVector<SDValue, 8> Ops(NumElems);
789 for (unsigned i = 0; i < NumElems; ++i) {
790 SDValue LHSElem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, LHS,
791 DAG.getConstant(i, TLI.getVectorIdxTy()));
792 SDValue RHSElem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, RHS,
793 DAG.getConstant(i, TLI.getVectorIdxTy()));
794 Ops[i] = DAG.getNode(ISD::SETCC, dl,
795 TLI.getSetCCResultType(*DAG.getContext(), TmpEltVT),
796 LHSElem, RHSElem, CC);
797 Ops[i] = DAG.getSelect(dl, EltVT, Ops[i],
798 DAG.getConstant(APInt::getAllOnesValue
799 (EltVT.getSizeInBits()), EltVT),
800 DAG.getConstant(0, EltVT));
802 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
807 bool SelectionDAG::LegalizeVectors() {
808 return VectorLegalizer(*this).Run();