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 // Expand bswap of vectors into a shuffle if legal.
67 SDValue ExpandBSWAP(SDValue Op);
68 // Implement vselect in terms of XOR, AND, OR when blend is not supported
70 SDValue ExpandVSELECT(SDValue Op);
71 SDValue ExpandSELECT(SDValue Op);
72 SDValue ExpandLoad(SDValue Op);
73 SDValue ExpandStore(SDValue Op);
74 SDValue ExpandFNEG(SDValue Op);
75 // Implements vector promotion; this is essentially just bitcasting the
76 // operands to a different type and bitcasting the result back to the
78 SDValue PromoteVectorOp(SDValue Op);
79 // Implements [SU]INT_TO_FP vector promotion; this is a [zs]ext of the input
80 // operand to the next size up.
81 SDValue PromoteVectorOpINT_TO_FP(SDValue Op);
82 // Implements FP_TO_[SU]INT vector promotion of the result type; it is
83 // promoted to the next size up integer type. The result is then truncated
84 // back to the original type.
85 SDValue PromoteVectorOpFP_TO_INT(SDValue Op, bool isSigned);
89 VectorLegalizer(SelectionDAG& dag) :
90 DAG(dag), TLI(dag.getTargetLoweringInfo()), Changed(false) {}
93 bool VectorLegalizer::Run() {
94 // Before we start legalizing vector nodes, check if there are any vectors.
95 bool HasVectors = false;
96 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
97 E = std::prev(DAG.allnodes_end()); I != std::next(E); ++I) {
98 // Check if the values of the nodes contain vectors. We don't need to check
99 // the operands because we are going to check their values at some point.
100 for (SDNode::value_iterator J = I->value_begin(), E = I->value_end();
102 HasVectors |= J->isVector();
104 // If we found a vector node we can start the legalization.
109 // If this basic block has no vectors then no need to legalize vectors.
113 // The legalize process is inherently a bottom-up recursive process (users
114 // legalize their uses before themselves). Given infinite stack space, we
115 // could just start legalizing on the root and traverse the whole graph. In
116 // practice however, this causes us to run out of stack space on large basic
117 // blocks. To avoid this problem, compute an ordering of the nodes where each
118 // node is only legalized after all of its operands are legalized.
119 DAG.AssignTopologicalOrder();
120 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
121 E = std::prev(DAG.allnodes_end()); I != std::next(E); ++I)
122 LegalizeOp(SDValue(I, 0));
124 // Finally, it's possible the root changed. Get the new root.
125 SDValue OldRoot = DAG.getRoot();
126 assert(LegalizedNodes.count(OldRoot) && "Root didn't get legalized?");
127 DAG.setRoot(LegalizedNodes[OldRoot]);
129 LegalizedNodes.clear();
131 // Remove dead nodes now.
132 DAG.RemoveDeadNodes();
137 SDValue VectorLegalizer::TranslateLegalizeResults(SDValue Op, SDValue Result) {
138 // Generic legalization: just pass the operand through.
139 for (unsigned i = 0, e = Op.getNode()->getNumValues(); i != e; ++i)
140 AddLegalizedOperand(Op.getValue(i), Result.getValue(i));
141 return Result.getValue(Op.getResNo());
144 SDValue VectorLegalizer::LegalizeOp(SDValue Op) {
145 // Note that LegalizeOp may be reentered even from single-use nodes, which
146 // means that we always must cache transformed nodes.
147 DenseMap<SDValue, SDValue>::iterator I = LegalizedNodes.find(Op);
148 if (I != LegalizedNodes.end()) return I->second;
150 SDNode* Node = Op.getNode();
152 // Legalize the operands
153 SmallVector<SDValue, 8> Ops;
154 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
155 Ops.push_back(LegalizeOp(Node->getOperand(i)));
157 SDValue Result = SDValue(DAG.UpdateNodeOperands(Op.getNode(), Ops), 0);
159 if (Op.getOpcode() == ISD::LOAD) {
160 LoadSDNode *LD = cast<LoadSDNode>(Op.getNode());
161 ISD::LoadExtType ExtType = LD->getExtensionType();
162 if (LD->getMemoryVT().isVector() && ExtType != ISD::NON_EXTLOAD) {
163 if (TLI.isLoadExtLegal(LD->getExtensionType(), LD->getMemoryVT()))
164 return TranslateLegalizeResults(Op, Result);
166 return LegalizeOp(ExpandLoad(Op));
168 } else if (Op.getOpcode() == ISD::STORE) {
169 StoreSDNode *ST = cast<StoreSDNode>(Op.getNode());
170 EVT StVT = ST->getMemoryVT();
171 MVT ValVT = ST->getValue().getSimpleValueType();
172 if (StVT.isVector() && ST->isTruncatingStore())
173 switch (TLI.getTruncStoreAction(ValVT, StVT.getSimpleVT())) {
174 default: llvm_unreachable("This action is not supported yet!");
175 case TargetLowering::Legal:
176 return TranslateLegalizeResults(Op, Result);
177 case TargetLowering::Custom:
179 return TranslateLegalizeResults(Op, TLI.LowerOperation(Result, DAG));
180 case TargetLowering::Expand:
182 return LegalizeOp(ExpandStore(Op));
186 bool HasVectorValue = false;
187 for (SDNode::value_iterator J = Node->value_begin(), E = Node->value_end();
190 HasVectorValue |= J->isVector();
192 return TranslateLegalizeResults(Op, Result);
195 switch (Op.getOpcode()) {
197 return TranslateLegalizeResults(Op, Result);
221 case ISD::CTLZ_ZERO_UNDEF:
222 case ISD::CTTZ_ZERO_UNDEF:
228 case ISD::ZERO_EXTEND:
229 case ISD::ANY_EXTEND:
231 case ISD::SIGN_EXTEND:
232 case ISD::FP_TO_SINT:
233 case ISD::FP_TO_UINT:
250 case ISD::FNEARBYINT:
256 case ISD::SIGN_EXTEND_INREG:
257 QueryType = Node->getValueType(0);
259 case ISD::FP_ROUND_INREG:
260 QueryType = cast<VTSDNode>(Node->getOperand(1))->getVT();
262 case ISD::SINT_TO_FP:
263 case ISD::UINT_TO_FP:
264 QueryType = Node->getOperand(0).getValueType();
268 switch (TLI.getOperationAction(Node->getOpcode(), QueryType)) {
269 case TargetLowering::Promote:
270 switch (Op.getOpcode()) {
272 // "Promote" the operation by bitcasting
273 Result = PromoteVectorOp(Op);
276 case ISD::SINT_TO_FP:
277 case ISD::UINT_TO_FP:
278 // "Promote" the operation by extending the operand.
279 Result = PromoteVectorOpINT_TO_FP(Op);
282 case ISD::FP_TO_UINT:
283 case ISD::FP_TO_SINT:
284 // Promote the operation by extending the operand.
285 Result = PromoteVectorOpFP_TO_INT(Op, Op->getOpcode() == ISD::FP_TO_SINT);
290 case TargetLowering::Legal: break;
291 case TargetLowering::Custom: {
292 SDValue Tmp1 = TLI.LowerOperation(Op, DAG);
293 if (Tmp1.getNode()) {
299 case TargetLowering::Expand:
300 if (Node->getOpcode() == ISD::SIGN_EXTEND_INREG)
301 Result = ExpandSEXTINREG(Op);
302 else if (Node->getOpcode() == ISD::BSWAP)
303 Result = ExpandBSWAP(Op);
304 else if (Node->getOpcode() == ISD::VSELECT)
305 Result = ExpandVSELECT(Op);
306 else if (Node->getOpcode() == ISD::SELECT)
307 Result = ExpandSELECT(Op);
308 else if (Node->getOpcode() == ISD::UINT_TO_FP)
309 Result = ExpandUINT_TO_FLOAT(Op);
310 else if (Node->getOpcode() == ISD::FNEG)
311 Result = ExpandFNEG(Op);
312 else if (Node->getOpcode() == ISD::SETCC)
313 Result = UnrollVSETCC(Op);
315 Result = DAG.UnrollVectorOp(Op.getNode());
319 // Make sure that the generated code is itself legal.
321 Result = LegalizeOp(Result);
325 // Note that LegalizeOp may be reentered even from single-use nodes, which
326 // means that we always must cache transformed nodes.
327 AddLegalizedOperand(Op, Result);
331 SDValue VectorLegalizer::PromoteVectorOp(SDValue Op) {
332 // Vector "promotion" is basically just bitcasting and doing the operation
333 // in a different type. For example, x86 promotes ISD::AND on v2i32 to
335 MVT VT = Op.getSimpleValueType();
336 assert(Op.getNode()->getNumValues() == 1 &&
337 "Can't promote a vector with multiple results!");
338 MVT NVT = TLI.getTypeToPromoteTo(Op.getOpcode(), VT);
340 SmallVector<SDValue, 4> Operands(Op.getNumOperands());
342 for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
343 if (Op.getOperand(j).getValueType().isVector())
344 Operands[j] = DAG.getNode(ISD::BITCAST, dl, NVT, Op.getOperand(j));
346 Operands[j] = Op.getOperand(j);
349 Op = DAG.getNode(Op.getOpcode(), dl, NVT, Operands);
351 return DAG.getNode(ISD::BITCAST, dl, VT, Op);
354 SDValue VectorLegalizer::PromoteVectorOpINT_TO_FP(SDValue Op) {
355 // INT_TO_FP operations may require the input operand be promoted even
356 // when the type is otherwise legal.
357 EVT VT = Op.getOperand(0).getValueType();
358 assert(Op.getNode()->getNumValues() == 1 &&
359 "Can't promote a vector with multiple results!");
361 // Normal getTypeToPromoteTo() doesn't work here, as that will promote
362 // by widening the vector w/ the same element width and twice the number
363 // of elements. We want the other way around, the same number of elements,
364 // each twice the width.
366 // Increase the bitwidth of the element to the next pow-of-two
367 // (which is greater than 8 bits).
369 EVT NVT = VT.widenIntegerVectorElementType(*DAG.getContext());
370 assert(NVT.isSimple() && "Promoting to a non-simple vector type!");
372 SmallVector<SDValue, 4> Operands(Op.getNumOperands());
374 unsigned Opc = Op.getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND :
376 for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
377 if (Op.getOperand(j).getValueType().isVector())
378 Operands[j] = DAG.getNode(Opc, dl, NVT, Op.getOperand(j));
380 Operands[j] = Op.getOperand(j);
383 return DAG.getNode(Op.getOpcode(), dl, Op.getValueType(), Operands);
386 // For FP_TO_INT we promote the result type to a vector type with wider
387 // elements and then truncate the result. This is different from the default
388 // PromoteVector which uses bitcast to promote thus assumning that the
389 // promoted vector type has the same overall size.
390 SDValue VectorLegalizer::PromoteVectorOpFP_TO_INT(SDValue Op, bool isSigned) {
391 assert(Op.getNode()->getNumValues() == 1 &&
392 "Can't promote a vector with multiple results!");
393 EVT VT = Op.getValueType();
398 NewVT = VT.widenIntegerVectorElementType(*DAG.getContext());
399 assert(NewVT.isSimple() && "Promoting to a non-simple vector type!");
400 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NewVT)) {
401 NewOpc = ISD::FP_TO_SINT;
404 if (!isSigned && TLI.isOperationLegalOrCustom(ISD::FP_TO_UINT, NewVT)) {
405 NewOpc = ISD::FP_TO_UINT;
411 SDValue promoted = DAG.getNode(NewOpc, SDLoc(Op), NewVT, Op.getOperand(0));
412 return DAG.getNode(ISD::TRUNCATE, SDLoc(Op), VT, promoted);
416 SDValue VectorLegalizer::ExpandLoad(SDValue Op) {
418 LoadSDNode *LD = cast<LoadSDNode>(Op.getNode());
419 SDValue Chain = LD->getChain();
420 SDValue BasePTR = LD->getBasePtr();
421 EVT SrcVT = LD->getMemoryVT();
422 ISD::LoadExtType ExtType = LD->getExtensionType();
424 SmallVector<SDValue, 8> Vals;
425 SmallVector<SDValue, 8> LoadChains;
426 unsigned NumElem = SrcVT.getVectorNumElements();
428 EVT SrcEltVT = SrcVT.getScalarType();
429 EVT DstEltVT = Op.getNode()->getValueType(0).getScalarType();
431 if (SrcVT.getVectorNumElements() > 1 && !SrcEltVT.isByteSized()) {
432 // When elements in a vector is not byte-addressable, we cannot directly
433 // load each element by advancing pointer, which could only address bytes.
434 // Instead, we load all significant words, mask bits off, and concatenate
435 // them to form each element. Finally, they are extended to destination
436 // scalar type to build the destination vector.
437 EVT WideVT = TLI.getPointerTy();
439 assert(WideVT.isRound() &&
440 "Could not handle the sophisticated case when the widest integer is"
442 assert(WideVT.bitsGE(SrcEltVT) &&
443 "Type is not legalized?");
445 unsigned WideBytes = WideVT.getStoreSize();
447 unsigned RemainingBytes = SrcVT.getStoreSize();
448 SmallVector<SDValue, 8> LoadVals;
450 while (RemainingBytes > 0) {
452 unsigned LoadBytes = WideBytes;
454 if (RemainingBytes >= LoadBytes) {
455 ScalarLoad = DAG.getLoad(WideVT, dl, Chain, BasePTR,
456 LD->getPointerInfo().getWithOffset(Offset),
457 LD->isVolatile(), LD->isNonTemporal(),
458 LD->isInvariant(), LD->getAlignment(),
462 while (RemainingBytes < LoadBytes) {
463 LoadBytes >>= 1; // Reduce the load size by half.
464 LoadVT = EVT::getIntegerVT(*DAG.getContext(), LoadBytes << 3);
466 ScalarLoad = DAG.getExtLoad(ISD::EXTLOAD, dl, WideVT, Chain, BasePTR,
467 LD->getPointerInfo().getWithOffset(Offset),
468 LoadVT, LD->isVolatile(),
469 LD->isNonTemporal(), LD->getAlignment(),
473 RemainingBytes -= LoadBytes;
475 BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR,
476 DAG.getConstant(LoadBytes, BasePTR.getValueType()));
478 LoadVals.push_back(ScalarLoad.getValue(0));
479 LoadChains.push_back(ScalarLoad.getValue(1));
482 // Extract bits, pack and extend/trunc them into destination type.
483 unsigned SrcEltBits = SrcEltVT.getSizeInBits();
484 SDValue SrcEltBitMask = DAG.getConstant((1U << SrcEltBits) - 1, WideVT);
486 unsigned BitOffset = 0;
487 unsigned WideIdx = 0;
488 unsigned WideBits = WideVT.getSizeInBits();
490 for (unsigned Idx = 0; Idx != NumElem; ++Idx) {
491 SDValue Lo, Hi, ShAmt;
493 if (BitOffset < WideBits) {
494 ShAmt = DAG.getConstant(BitOffset, TLI.getShiftAmountTy(WideVT));
495 Lo = DAG.getNode(ISD::SRL, dl, WideVT, LoadVals[WideIdx], ShAmt);
496 Lo = DAG.getNode(ISD::AND, dl, WideVT, Lo, SrcEltBitMask);
499 BitOffset += SrcEltBits;
500 if (BitOffset >= WideBits) {
504 ShAmt = DAG.getConstant(SrcEltBits - Offset,
505 TLI.getShiftAmountTy(WideVT));
506 Hi = DAG.getNode(ISD::SHL, dl, WideVT, LoadVals[WideIdx], ShAmt);
507 Hi = DAG.getNode(ISD::AND, dl, WideVT, Hi, SrcEltBitMask);
512 Lo = DAG.getNode(ISD::OR, dl, WideVT, Lo, Hi);
515 default: llvm_unreachable("Unknown extended-load op!");
517 Lo = DAG.getAnyExtOrTrunc(Lo, dl, DstEltVT);
520 Lo = DAG.getZExtOrTrunc(Lo, dl, DstEltVT);
523 ShAmt = DAG.getConstant(WideBits - SrcEltBits,
524 TLI.getShiftAmountTy(WideVT));
525 Lo = DAG.getNode(ISD::SHL, dl, WideVT, Lo, ShAmt);
526 Lo = DAG.getNode(ISD::SRA, dl, WideVT, Lo, ShAmt);
527 Lo = DAG.getSExtOrTrunc(Lo, dl, DstEltVT);
533 unsigned Stride = SrcVT.getScalarType().getSizeInBits()/8;
535 for (unsigned Idx=0; Idx<NumElem; Idx++) {
536 SDValue ScalarLoad = DAG.getExtLoad(ExtType, dl,
537 Op.getNode()->getValueType(0).getScalarType(),
538 Chain, BasePTR, LD->getPointerInfo().getWithOffset(Idx * Stride),
539 SrcVT.getScalarType(),
540 LD->isVolatile(), LD->isNonTemporal(),
541 LD->getAlignment(), LD->getTBAAInfo());
543 BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR,
544 DAG.getConstant(Stride, BasePTR.getValueType()));
546 Vals.push_back(ScalarLoad.getValue(0));
547 LoadChains.push_back(ScalarLoad.getValue(1));
551 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, LoadChains);
552 SDValue Value = DAG.getNode(ISD::BUILD_VECTOR, dl,
553 Op.getNode()->getValueType(0), Vals);
555 AddLegalizedOperand(Op.getValue(0), Value);
556 AddLegalizedOperand(Op.getValue(1), NewChain);
558 return (Op.getResNo() ? NewChain : Value);
561 SDValue VectorLegalizer::ExpandStore(SDValue Op) {
563 StoreSDNode *ST = cast<StoreSDNode>(Op.getNode());
564 SDValue Chain = ST->getChain();
565 SDValue BasePTR = ST->getBasePtr();
566 SDValue Value = ST->getValue();
567 EVT StVT = ST->getMemoryVT();
569 unsigned Alignment = ST->getAlignment();
570 bool isVolatile = ST->isVolatile();
571 bool isNonTemporal = ST->isNonTemporal();
572 const MDNode *TBAAInfo = ST->getTBAAInfo();
574 unsigned NumElem = StVT.getVectorNumElements();
575 // The type of the data we want to save
576 EVT RegVT = Value.getValueType();
577 EVT RegSclVT = RegVT.getScalarType();
578 // The type of data as saved in memory.
579 EVT MemSclVT = StVT.getScalarType();
581 // Cast floats into integers
582 unsigned ScalarSize = MemSclVT.getSizeInBits();
584 // Round odd types to the next pow of two.
585 if (!isPowerOf2_32(ScalarSize))
586 ScalarSize = NextPowerOf2(ScalarSize);
588 // Store Stride in bytes
589 unsigned Stride = ScalarSize/8;
590 // Extract each of the elements from the original vector
591 // and save them into memory individually.
592 SmallVector<SDValue, 8> Stores;
593 for (unsigned Idx = 0; Idx < NumElem; Idx++) {
594 SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
595 RegSclVT, Value, DAG.getConstant(Idx, TLI.getVectorIdxTy()));
597 // This scalar TruncStore may be illegal, but we legalize it later.
598 SDValue Store = DAG.getTruncStore(Chain, dl, Ex, BasePTR,
599 ST->getPointerInfo().getWithOffset(Idx*Stride), MemSclVT,
600 isVolatile, isNonTemporal, Alignment, TBAAInfo);
602 BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR,
603 DAG.getConstant(Stride, BasePTR.getValueType()));
605 Stores.push_back(Store);
607 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
608 AddLegalizedOperand(Op, TF);
612 SDValue VectorLegalizer::ExpandSELECT(SDValue Op) {
613 // Lower a select instruction where the condition is a scalar and the
614 // operands are vectors. Lower this select to VSELECT and implement it
615 // using XOR AND OR. The selector bit is broadcasted.
616 EVT VT = Op.getValueType();
619 SDValue Mask = Op.getOperand(0);
620 SDValue Op1 = Op.getOperand(1);
621 SDValue Op2 = Op.getOperand(2);
623 assert(VT.isVector() && !Mask.getValueType().isVector()
624 && Op1.getValueType() == Op2.getValueType() && "Invalid type");
626 unsigned NumElem = VT.getVectorNumElements();
628 // If we can't even use the basic vector operations of
629 // AND,OR,XOR, we will have to scalarize the op.
630 // Notice that the operation may be 'promoted' which means that it is
631 // 'bitcasted' to another type which is handled.
632 // Also, we need to be able to construct a splat vector using BUILD_VECTOR.
633 if (TLI.getOperationAction(ISD::AND, VT) == TargetLowering::Expand ||
634 TLI.getOperationAction(ISD::XOR, VT) == TargetLowering::Expand ||
635 TLI.getOperationAction(ISD::OR, VT) == TargetLowering::Expand ||
636 TLI.getOperationAction(ISD::BUILD_VECTOR, VT) == TargetLowering::Expand)
637 return DAG.UnrollVectorOp(Op.getNode());
639 // Generate a mask operand.
640 EVT MaskTy = VT.changeVectorElementTypeToInteger();
642 // What is the size of each element in the vector mask.
643 EVT BitTy = MaskTy.getScalarType();
645 Mask = DAG.getSelect(DL, BitTy, Mask,
646 DAG.getConstant(APInt::getAllOnesValue(BitTy.getSizeInBits()), BitTy),
647 DAG.getConstant(0, BitTy));
649 // Broadcast the mask so that the entire vector is all-one or all zero.
650 SmallVector<SDValue, 8> Ops(NumElem, Mask);
651 Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, MaskTy, Ops);
653 // Bitcast the operands to be the same type as the mask.
654 // This is needed when we select between FP types because
655 // the mask is a vector of integers.
656 Op1 = DAG.getNode(ISD::BITCAST, DL, MaskTy, Op1);
657 Op2 = DAG.getNode(ISD::BITCAST, DL, MaskTy, Op2);
659 SDValue AllOnes = DAG.getConstant(
660 APInt::getAllOnesValue(BitTy.getSizeInBits()), MaskTy);
661 SDValue NotMask = DAG.getNode(ISD::XOR, DL, MaskTy, Mask, AllOnes);
663 Op1 = DAG.getNode(ISD::AND, DL, MaskTy, Op1, Mask);
664 Op2 = DAG.getNode(ISD::AND, DL, MaskTy, Op2, NotMask);
665 SDValue Val = DAG.getNode(ISD::OR, DL, MaskTy, Op1, Op2);
666 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Val);
669 SDValue VectorLegalizer::ExpandSEXTINREG(SDValue Op) {
670 EVT VT = Op.getValueType();
672 // Make sure that the SRA and SHL instructions are available.
673 if (TLI.getOperationAction(ISD::SRA, VT) == TargetLowering::Expand ||
674 TLI.getOperationAction(ISD::SHL, VT) == TargetLowering::Expand)
675 return DAG.UnrollVectorOp(Op.getNode());
678 EVT OrigTy = cast<VTSDNode>(Op->getOperand(1))->getVT();
680 unsigned BW = VT.getScalarType().getSizeInBits();
681 unsigned OrigBW = OrigTy.getScalarType().getSizeInBits();
682 SDValue ShiftSz = DAG.getConstant(BW - OrigBW, VT);
684 Op = Op.getOperand(0);
685 Op = DAG.getNode(ISD::SHL, DL, VT, Op, ShiftSz);
686 return DAG.getNode(ISD::SRA, DL, VT, Op, ShiftSz);
689 SDValue VectorLegalizer::ExpandBSWAP(SDValue Op) {
690 EVT VT = Op.getValueType();
692 // Generate a byte wise shuffle mask for the BSWAP.
693 SmallVector<int, 16> ShuffleMask;
694 int ScalarSizeInBytes = VT.getScalarSizeInBits() / 8;
695 for (int I = 0, E = VT.getVectorNumElements(); I != E; ++I)
696 for (int J = ScalarSizeInBytes - 1; J >= 0; --J)
697 ShuffleMask.push_back((I * ScalarSizeInBytes) + J);
699 EVT ByteVT = EVT::getVectorVT(*DAG.getContext(), MVT::i8, ShuffleMask.size());
701 // Only emit a shuffle if the mask is legal.
702 if (!TLI.isShuffleMaskLegal(ShuffleMask, ByteVT))
703 return DAG.UnrollVectorOp(Op.getNode());
706 Op = DAG.getNode(ISD::BITCAST, DL, ByteVT, Op.getOperand(0));
707 Op = DAG.getVectorShuffle(ByteVT, DL, Op, DAG.getUNDEF(ByteVT),
709 return DAG.getNode(ISD::BITCAST, DL, VT, Op);
712 SDValue VectorLegalizer::ExpandVSELECT(SDValue Op) {
713 // Implement VSELECT in terms of XOR, AND, OR
714 // on platforms which do not support blend natively.
717 SDValue Mask = Op.getOperand(0);
718 SDValue Op1 = Op.getOperand(1);
719 SDValue Op2 = Op.getOperand(2);
721 EVT VT = Mask.getValueType();
723 // If we can't even use the basic vector operations of
724 // AND,OR,XOR, we will have to scalarize the op.
725 // Notice that the operation may be 'promoted' which means that it is
726 // 'bitcasted' to another type which is handled.
727 // This operation also isn't safe with AND, OR, XOR when the boolean
728 // type is 0/1 as we need an all ones vector constant to mask with.
729 // FIXME: Sign extend 1 to all ones if thats legal on the target.
730 if (TLI.getOperationAction(ISD::AND, VT) == TargetLowering::Expand ||
731 TLI.getOperationAction(ISD::XOR, VT) == TargetLowering::Expand ||
732 TLI.getOperationAction(ISD::OR, VT) == TargetLowering::Expand ||
733 TLI.getBooleanContents(true) !=
734 TargetLowering::ZeroOrNegativeOneBooleanContent)
735 return DAG.UnrollVectorOp(Op.getNode());
737 // If the mask and the type are different sizes, unroll the vector op. This
738 // can occur when getSetCCResultType returns something that is different in
739 // size from the operand types. For example, v4i8 = select v4i32, v4i8, v4i8.
740 if (VT.getSizeInBits() != Op1.getValueType().getSizeInBits())
741 return DAG.UnrollVectorOp(Op.getNode());
743 // Bitcast the operands to be the same type as the mask.
744 // This is needed when we select between FP types because
745 // the mask is a vector of integers.
746 Op1 = DAG.getNode(ISD::BITCAST, DL, VT, Op1);
747 Op2 = DAG.getNode(ISD::BITCAST, DL, VT, Op2);
749 SDValue AllOnes = DAG.getConstant(
750 APInt::getAllOnesValue(VT.getScalarType().getSizeInBits()), VT);
751 SDValue NotMask = DAG.getNode(ISD::XOR, DL, VT, Mask, AllOnes);
753 Op1 = DAG.getNode(ISD::AND, DL, VT, Op1, Mask);
754 Op2 = DAG.getNode(ISD::AND, DL, VT, Op2, NotMask);
755 SDValue Val = DAG.getNode(ISD::OR, DL, VT, Op1, Op2);
756 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Val);
759 SDValue VectorLegalizer::ExpandUINT_TO_FLOAT(SDValue Op) {
760 EVT VT = Op.getOperand(0).getValueType();
763 // Make sure that the SINT_TO_FP and SRL instructions are available.
764 if (TLI.getOperationAction(ISD::SINT_TO_FP, VT) == TargetLowering::Expand ||
765 TLI.getOperationAction(ISD::SRL, VT) == TargetLowering::Expand)
766 return DAG.UnrollVectorOp(Op.getNode());
768 EVT SVT = VT.getScalarType();
769 assert((SVT.getSizeInBits() == 64 || SVT.getSizeInBits() == 32) &&
770 "Elements in vector-UINT_TO_FP must be 32 or 64 bits wide");
772 unsigned BW = SVT.getSizeInBits();
773 SDValue HalfWord = DAG.getConstant(BW/2, VT);
775 // Constants to clear the upper part of the word.
776 // Notice that we can also use SHL+SHR, but using a constant is slightly
778 uint64_t HWMask = (SVT.getSizeInBits()==64)?0x00000000FFFFFFFF:0x0000FFFF;
779 SDValue HalfWordMask = DAG.getConstant(HWMask, VT);
781 // Two to the power of half-word-size.
782 SDValue TWOHW = DAG.getConstantFP((1<<(BW/2)), Op.getValueType());
784 // Clear upper part of LO, lower HI
785 SDValue HI = DAG.getNode(ISD::SRL, DL, VT, Op.getOperand(0), HalfWord);
786 SDValue LO = DAG.getNode(ISD::AND, DL, VT, Op.getOperand(0), HalfWordMask);
788 // Convert hi and lo to floats
789 // Convert the hi part back to the upper values
790 SDValue fHI = DAG.getNode(ISD::SINT_TO_FP, DL, Op.getValueType(), HI);
791 fHI = DAG.getNode(ISD::FMUL, DL, Op.getValueType(), fHI, TWOHW);
792 SDValue fLO = DAG.getNode(ISD::SINT_TO_FP, DL, Op.getValueType(), LO);
794 // Add the two halves
795 return DAG.getNode(ISD::FADD, DL, Op.getValueType(), fHI, fLO);
799 SDValue VectorLegalizer::ExpandFNEG(SDValue Op) {
800 if (TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType())) {
801 SDValue Zero = DAG.getConstantFP(-0.0, Op.getValueType());
802 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
803 Zero, Op.getOperand(0));
805 return DAG.UnrollVectorOp(Op.getNode());
808 SDValue VectorLegalizer::UnrollVSETCC(SDValue Op) {
809 EVT VT = Op.getValueType();
810 unsigned NumElems = VT.getVectorNumElements();
811 EVT EltVT = VT.getVectorElementType();
812 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1), CC = Op.getOperand(2);
813 EVT TmpEltVT = LHS.getValueType().getVectorElementType();
815 SmallVector<SDValue, 8> Ops(NumElems);
816 for (unsigned i = 0; i < NumElems; ++i) {
817 SDValue LHSElem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, LHS,
818 DAG.getConstant(i, TLI.getVectorIdxTy()));
819 SDValue RHSElem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, RHS,
820 DAG.getConstant(i, TLI.getVectorIdxTy()));
821 Ops[i] = DAG.getNode(ISD::SETCC, dl,
822 TLI.getSetCCResultType(*DAG.getContext(), TmpEltVT),
823 LHSElem, RHSElem, CC);
824 Ops[i] = DAG.getSelect(dl, EltVT, Ops[i],
825 DAG.getConstant(APInt::getAllOnesValue
826 (EltVT.getSizeInBits()), EltVT),
827 DAG.getConstant(0, EltVT));
829 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
834 bool SelectionDAG::LegalizeVectors() {
835 return VectorLegalizer(*this).Run();