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)));
156 SDValue(DAG.UpdateNodeOperands(Op.getNode(), Ops.data(), Ops.size()), 0);
158 if (Op.getOpcode() == ISD::LOAD) {
159 LoadSDNode *LD = cast<LoadSDNode>(Op.getNode());
160 ISD::LoadExtType ExtType = LD->getExtensionType();
161 if (LD->getMemoryVT().isVector() && ExtType != ISD::NON_EXTLOAD) {
162 if (TLI.isLoadExtLegal(LD->getExtensionType(), LD->getMemoryVT()))
163 return TranslateLegalizeResults(Op, Result);
165 return LegalizeOp(ExpandLoad(Op));
167 } else if (Op.getOpcode() == ISD::STORE) {
168 StoreSDNode *ST = cast<StoreSDNode>(Op.getNode());
169 EVT StVT = ST->getMemoryVT();
170 MVT ValVT = ST->getValue().getSimpleValueType();
171 if (StVT.isVector() && ST->isTruncatingStore())
172 switch (TLI.getTruncStoreAction(ValVT, StVT.getSimpleVT())) {
173 default: llvm_unreachable("This action is not supported yet!");
174 case TargetLowering::Legal:
175 return TranslateLegalizeResults(Op, Result);
176 case TargetLowering::Custom:
178 return TranslateLegalizeResults(Op, TLI.LowerOperation(Result, DAG));
179 case TargetLowering::Expand:
181 return LegalizeOp(ExpandStore(Op));
185 bool HasVectorValue = false;
186 for (SDNode::value_iterator J = Node->value_begin(), E = Node->value_end();
189 HasVectorValue |= J->isVector();
191 return TranslateLegalizeResults(Op, Result);
194 switch (Op.getOpcode()) {
196 return TranslateLegalizeResults(Op, Result);
220 case ISD::CTLZ_ZERO_UNDEF:
221 case ISD::CTTZ_ZERO_UNDEF:
227 case ISD::ZERO_EXTEND:
228 case ISD::ANY_EXTEND:
230 case ISD::SIGN_EXTEND:
231 case ISD::FP_TO_SINT:
232 case ISD::FP_TO_UINT:
249 case ISD::FNEARBYINT:
255 case ISD::SIGN_EXTEND_INREG:
256 QueryType = Node->getValueType(0);
258 case ISD::FP_ROUND_INREG:
259 QueryType = cast<VTSDNode>(Node->getOperand(1))->getVT();
261 case ISD::SINT_TO_FP:
262 case ISD::UINT_TO_FP:
263 QueryType = Node->getOperand(0).getValueType();
267 switch (TLI.getOperationAction(Node->getOpcode(), QueryType)) {
268 case TargetLowering::Promote:
269 switch (Op.getOpcode()) {
271 // "Promote" the operation by bitcasting
272 Result = PromoteVectorOp(Op);
275 case ISD::SINT_TO_FP:
276 case ISD::UINT_TO_FP:
277 // "Promote" the operation by extending the operand.
278 Result = PromoteVectorOpINT_TO_FP(Op);
281 case ISD::FP_TO_UINT:
282 case ISD::FP_TO_SINT:
283 // Promote the operation by extending the operand.
284 Result = PromoteVectorOpFP_TO_INT(Op, Op->getOpcode() == ISD::FP_TO_SINT);
289 case TargetLowering::Legal: break;
290 case TargetLowering::Custom: {
291 SDValue Tmp1 = TLI.LowerOperation(Op, DAG);
292 if (Tmp1.getNode()) {
298 case TargetLowering::Expand:
299 if (Node->getOpcode() == ISD::SIGN_EXTEND_INREG)
300 Result = ExpandSEXTINREG(Op);
301 else if (Node->getOpcode() == ISD::VSELECT)
302 Result = ExpandVSELECT(Op);
303 else if (Node->getOpcode() == ISD::SELECT)
304 Result = ExpandSELECT(Op);
305 else if (Node->getOpcode() == ISD::UINT_TO_FP)
306 Result = ExpandUINT_TO_FLOAT(Op);
307 else if (Node->getOpcode() == ISD::FNEG)
308 Result = ExpandFNEG(Op);
309 else if (Node->getOpcode() == ISD::SETCC)
310 Result = UnrollVSETCC(Op);
312 Result = DAG.UnrollVectorOp(Op.getNode());
316 // Make sure that the generated code is itself legal.
318 Result = LegalizeOp(Result);
322 // Note that LegalizeOp may be reentered even from single-use nodes, which
323 // means that we always must cache transformed nodes.
324 AddLegalizedOperand(Op, Result);
328 SDValue VectorLegalizer::PromoteVectorOp(SDValue Op) {
329 // Vector "promotion" is basically just bitcasting and doing the operation
330 // in a different type. For example, x86 promotes ISD::AND on v2i32 to
332 MVT VT = Op.getSimpleValueType();
333 assert(Op.getNode()->getNumValues() == 1 &&
334 "Can't promote a vector with multiple results!");
335 MVT NVT = TLI.getTypeToPromoteTo(Op.getOpcode(), VT);
337 SmallVector<SDValue, 4> Operands(Op.getNumOperands());
339 for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
340 if (Op.getOperand(j).getValueType().isVector())
341 Operands[j] = DAG.getNode(ISD::BITCAST, dl, NVT, Op.getOperand(j));
343 Operands[j] = Op.getOperand(j);
346 Op = DAG.getNode(Op.getOpcode(), dl, NVT, Operands);
348 return DAG.getNode(ISD::BITCAST, dl, VT, Op);
351 SDValue VectorLegalizer::PromoteVectorOpINT_TO_FP(SDValue Op) {
352 // INT_TO_FP operations may require the input operand be promoted even
353 // when the type is otherwise legal.
354 EVT VT = Op.getOperand(0).getValueType();
355 assert(Op.getNode()->getNumValues() == 1 &&
356 "Can't promote a vector with multiple results!");
358 // Normal getTypeToPromoteTo() doesn't work here, as that will promote
359 // by widening the vector w/ the same element width and twice the number
360 // of elements. We want the other way around, the same number of elements,
361 // each twice the width.
363 // Increase the bitwidth of the element to the next pow-of-two
364 // (which is greater than 8 bits).
366 EVT NVT = VT.widenIntegerVectorElementType(*DAG.getContext());
367 assert(NVT.isSimple() && "Promoting to a non-simple vector type!");
369 SmallVector<SDValue, 4> Operands(Op.getNumOperands());
371 unsigned Opc = Op.getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND :
373 for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
374 if (Op.getOperand(j).getValueType().isVector())
375 Operands[j] = DAG.getNode(Opc, dl, NVT, Op.getOperand(j));
377 Operands[j] = Op.getOperand(j);
380 return DAG.getNode(Op.getOpcode(), dl, Op.getValueType(), Operands);
383 // For FP_TO_INT we promote the result type to a vector type with wider
384 // elements and then truncate the result. This is different from the default
385 // PromoteVector which uses bitcast to promote thus assumning that the
386 // promoted vector type has the same overall size.
387 SDValue VectorLegalizer::PromoteVectorOpFP_TO_INT(SDValue Op, bool isSigned) {
388 assert(Op.getNode()->getNumValues() == 1 &&
389 "Can't promote a vector with multiple results!");
390 EVT VT = Op.getValueType();
395 NewVT = VT.widenIntegerVectorElementType(*DAG.getContext());
396 assert(NewVT.isSimple() && "Promoting to a non-simple vector type!");
397 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NewVT)) {
398 NewOpc = ISD::FP_TO_SINT;
401 if (!isSigned && TLI.isOperationLegalOrCustom(ISD::FP_TO_UINT, NewVT)) {
402 NewOpc = ISD::FP_TO_UINT;
408 SDValue promoted = DAG.getNode(NewOpc, SDLoc(Op), NewVT, Op.getOperand(0));
409 return DAG.getNode(ISD::TRUNCATE, SDLoc(Op), VT, promoted);
413 SDValue VectorLegalizer::ExpandLoad(SDValue Op) {
415 LoadSDNode *LD = cast<LoadSDNode>(Op.getNode());
416 SDValue Chain = LD->getChain();
417 SDValue BasePTR = LD->getBasePtr();
418 EVT SrcVT = LD->getMemoryVT();
419 ISD::LoadExtType ExtType = LD->getExtensionType();
421 SmallVector<SDValue, 8> Vals;
422 SmallVector<SDValue, 8> LoadChains;
423 unsigned NumElem = SrcVT.getVectorNumElements();
425 EVT SrcEltVT = SrcVT.getScalarType();
426 EVT DstEltVT = Op.getNode()->getValueType(0).getScalarType();
428 if (SrcVT.getVectorNumElements() > 1 && !SrcEltVT.isByteSized()) {
429 // When elements in a vector is not byte-addressable, we cannot directly
430 // load each element by advancing pointer, which could only address bytes.
431 // Instead, we load all significant words, mask bits off, and concatenate
432 // them to form each element. Finally, they are extended to destination
433 // scalar type to build the destination vector.
434 EVT WideVT = TLI.getPointerTy();
436 assert(WideVT.isRound() &&
437 "Could not handle the sophisticated case when the widest integer is"
439 assert(WideVT.bitsGE(SrcEltVT) &&
440 "Type is not legalized?");
442 unsigned WideBytes = WideVT.getStoreSize();
444 unsigned RemainingBytes = SrcVT.getStoreSize();
445 SmallVector<SDValue, 8> LoadVals;
447 while (RemainingBytes > 0) {
449 unsigned LoadBytes = WideBytes;
451 if (RemainingBytes >= LoadBytes) {
452 ScalarLoad = DAG.getLoad(WideVT, dl, Chain, BasePTR,
453 LD->getPointerInfo().getWithOffset(Offset),
454 LD->isVolatile(), LD->isNonTemporal(),
455 LD->isInvariant(), LD->getAlignment(),
459 while (RemainingBytes < LoadBytes) {
460 LoadBytes >>= 1; // Reduce the load size by half.
461 LoadVT = EVT::getIntegerVT(*DAG.getContext(), LoadBytes << 3);
463 ScalarLoad = DAG.getExtLoad(ISD::EXTLOAD, dl, WideVT, Chain, BasePTR,
464 LD->getPointerInfo().getWithOffset(Offset),
465 LoadVT, LD->isVolatile(),
466 LD->isNonTemporal(), LD->getAlignment(),
470 RemainingBytes -= LoadBytes;
472 BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR,
473 DAG.getConstant(LoadBytes, BasePTR.getValueType()));
475 LoadVals.push_back(ScalarLoad.getValue(0));
476 LoadChains.push_back(ScalarLoad.getValue(1));
479 // Extract bits, pack and extend/trunc them into destination type.
480 unsigned SrcEltBits = SrcEltVT.getSizeInBits();
481 SDValue SrcEltBitMask = DAG.getConstant((1U << SrcEltBits) - 1, WideVT);
483 unsigned BitOffset = 0;
484 unsigned WideIdx = 0;
485 unsigned WideBits = WideVT.getSizeInBits();
487 for (unsigned Idx = 0; Idx != NumElem; ++Idx) {
488 SDValue Lo, Hi, ShAmt;
490 if (BitOffset < WideBits) {
491 ShAmt = DAG.getConstant(BitOffset, TLI.getShiftAmountTy(WideVT));
492 Lo = DAG.getNode(ISD::SRL, dl, WideVT, LoadVals[WideIdx], ShAmt);
493 Lo = DAG.getNode(ISD::AND, dl, WideVT, Lo, SrcEltBitMask);
496 BitOffset += SrcEltBits;
497 if (BitOffset >= WideBits) {
501 ShAmt = DAG.getConstant(SrcEltBits - Offset,
502 TLI.getShiftAmountTy(WideVT));
503 Hi = DAG.getNode(ISD::SHL, dl, WideVT, LoadVals[WideIdx], ShAmt);
504 Hi = DAG.getNode(ISD::AND, dl, WideVT, Hi, SrcEltBitMask);
509 Lo = DAG.getNode(ISD::OR, dl, WideVT, Lo, Hi);
512 default: llvm_unreachable("Unknown extended-load op!");
514 Lo = DAG.getAnyExtOrTrunc(Lo, dl, DstEltVT);
517 Lo = DAG.getZExtOrTrunc(Lo, dl, DstEltVT);
520 ShAmt = DAG.getConstant(WideBits - SrcEltBits,
521 TLI.getShiftAmountTy(WideVT));
522 Lo = DAG.getNode(ISD::SHL, dl, WideVT, Lo, ShAmt);
523 Lo = DAG.getNode(ISD::SRA, dl, WideVT, Lo, ShAmt);
524 Lo = DAG.getSExtOrTrunc(Lo, dl, DstEltVT);
530 unsigned Stride = SrcVT.getScalarType().getSizeInBits()/8;
532 for (unsigned Idx=0; Idx<NumElem; Idx++) {
533 SDValue ScalarLoad = DAG.getExtLoad(ExtType, dl,
534 Op.getNode()->getValueType(0).getScalarType(),
535 Chain, BasePTR, LD->getPointerInfo().getWithOffset(Idx * Stride),
536 SrcVT.getScalarType(),
537 LD->isVolatile(), LD->isNonTemporal(),
538 LD->getAlignment(), LD->getTBAAInfo());
540 BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR,
541 DAG.getConstant(Stride, BasePTR.getValueType()));
543 Vals.push_back(ScalarLoad.getValue(0));
544 LoadChains.push_back(ScalarLoad.getValue(1));
548 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, LoadChains);
549 SDValue Value = DAG.getNode(ISD::BUILD_VECTOR, dl,
550 Op.getNode()->getValueType(0), Vals);
552 AddLegalizedOperand(Op.getValue(0), Value);
553 AddLegalizedOperand(Op.getValue(1), NewChain);
555 return (Op.getResNo() ? NewChain : Value);
558 SDValue VectorLegalizer::ExpandStore(SDValue Op) {
560 StoreSDNode *ST = cast<StoreSDNode>(Op.getNode());
561 SDValue Chain = ST->getChain();
562 SDValue BasePTR = ST->getBasePtr();
563 SDValue Value = ST->getValue();
564 EVT StVT = ST->getMemoryVT();
566 unsigned Alignment = ST->getAlignment();
567 bool isVolatile = ST->isVolatile();
568 bool isNonTemporal = ST->isNonTemporal();
569 const MDNode *TBAAInfo = ST->getTBAAInfo();
571 unsigned NumElem = StVT.getVectorNumElements();
572 // The type of the data we want to save
573 EVT RegVT = Value.getValueType();
574 EVT RegSclVT = RegVT.getScalarType();
575 // The type of data as saved in memory.
576 EVT MemSclVT = StVT.getScalarType();
578 // Cast floats into integers
579 unsigned ScalarSize = MemSclVT.getSizeInBits();
581 // Round odd types to the next pow of two.
582 if (!isPowerOf2_32(ScalarSize))
583 ScalarSize = NextPowerOf2(ScalarSize);
585 // Store Stride in bytes
586 unsigned Stride = ScalarSize/8;
587 // Extract each of the elements from the original vector
588 // and save them into memory individually.
589 SmallVector<SDValue, 8> Stores;
590 for (unsigned Idx = 0; Idx < NumElem; Idx++) {
591 SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
592 RegSclVT, Value, DAG.getConstant(Idx, TLI.getVectorIdxTy()));
594 // This scalar TruncStore may be illegal, but we legalize it later.
595 SDValue Store = DAG.getTruncStore(Chain, dl, Ex, BasePTR,
596 ST->getPointerInfo().getWithOffset(Idx*Stride), MemSclVT,
597 isVolatile, isNonTemporal, Alignment, TBAAInfo);
599 BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR,
600 DAG.getConstant(Stride, BasePTR.getValueType()));
602 Stores.push_back(Store);
604 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
605 AddLegalizedOperand(Op, TF);
609 SDValue VectorLegalizer::ExpandSELECT(SDValue Op) {
610 // Lower a select instruction where the condition is a scalar and the
611 // operands are vectors. Lower this select to VSELECT and implement it
612 // using XOR AND OR. The selector bit is broadcasted.
613 EVT VT = Op.getValueType();
616 SDValue Mask = Op.getOperand(0);
617 SDValue Op1 = Op.getOperand(1);
618 SDValue Op2 = Op.getOperand(2);
620 assert(VT.isVector() && !Mask.getValueType().isVector()
621 && Op1.getValueType() == Op2.getValueType() && "Invalid type");
623 unsigned NumElem = VT.getVectorNumElements();
625 // If we can't even use the basic vector operations of
626 // AND,OR,XOR, we will have to scalarize the op.
627 // Notice that the operation may be 'promoted' which means that it is
628 // 'bitcasted' to another type which is handled.
629 // Also, we need to be able to construct a splat vector using BUILD_VECTOR.
630 if (TLI.getOperationAction(ISD::AND, VT) == TargetLowering::Expand ||
631 TLI.getOperationAction(ISD::XOR, VT) == TargetLowering::Expand ||
632 TLI.getOperationAction(ISD::OR, VT) == TargetLowering::Expand ||
633 TLI.getOperationAction(ISD::BUILD_VECTOR, VT) == TargetLowering::Expand)
634 return DAG.UnrollVectorOp(Op.getNode());
636 // Generate a mask operand.
637 EVT MaskTy = VT.changeVectorElementTypeToInteger();
639 // What is the size of each element in the vector mask.
640 EVT BitTy = MaskTy.getScalarType();
642 Mask = DAG.getSelect(DL, BitTy, Mask,
643 DAG.getConstant(APInt::getAllOnesValue(BitTy.getSizeInBits()), BitTy),
644 DAG.getConstant(0, BitTy));
646 // Broadcast the mask so that the entire vector is all-one or all zero.
647 SmallVector<SDValue, 8> Ops(NumElem, Mask);
648 Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, MaskTy, Ops);
650 // Bitcast the operands to be the same type as the mask.
651 // This is needed when we select between FP types because
652 // the mask is a vector of integers.
653 Op1 = DAG.getNode(ISD::BITCAST, DL, MaskTy, Op1);
654 Op2 = DAG.getNode(ISD::BITCAST, DL, MaskTy, Op2);
656 SDValue AllOnes = DAG.getConstant(
657 APInt::getAllOnesValue(BitTy.getSizeInBits()), MaskTy);
658 SDValue NotMask = DAG.getNode(ISD::XOR, DL, MaskTy, Mask, AllOnes);
660 Op1 = DAG.getNode(ISD::AND, DL, MaskTy, Op1, Mask);
661 Op2 = DAG.getNode(ISD::AND, DL, MaskTy, Op2, NotMask);
662 SDValue Val = DAG.getNode(ISD::OR, DL, MaskTy, Op1, Op2);
663 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Val);
666 SDValue VectorLegalizer::ExpandSEXTINREG(SDValue Op) {
667 EVT VT = Op.getValueType();
669 // Make sure that the SRA and SHL instructions are available.
670 if (TLI.getOperationAction(ISD::SRA, VT) == TargetLowering::Expand ||
671 TLI.getOperationAction(ISD::SHL, VT) == TargetLowering::Expand)
672 return DAG.UnrollVectorOp(Op.getNode());
675 EVT OrigTy = cast<VTSDNode>(Op->getOperand(1))->getVT();
677 unsigned BW = VT.getScalarType().getSizeInBits();
678 unsigned OrigBW = OrigTy.getScalarType().getSizeInBits();
679 SDValue ShiftSz = DAG.getConstant(BW - OrigBW, VT);
681 Op = Op.getOperand(0);
682 Op = DAG.getNode(ISD::SHL, DL, VT, Op, ShiftSz);
683 return DAG.getNode(ISD::SRA, DL, VT, Op, ShiftSz);
686 SDValue VectorLegalizer::ExpandVSELECT(SDValue Op) {
687 // Implement VSELECT in terms of XOR, AND, OR
688 // on platforms which do not support blend natively.
691 SDValue Mask = Op.getOperand(0);
692 SDValue Op1 = Op.getOperand(1);
693 SDValue Op2 = Op.getOperand(2);
695 EVT VT = Mask.getValueType();
697 // If we can't even use the basic vector operations of
698 // AND,OR,XOR, we will have to scalarize the op.
699 // Notice that the operation may be 'promoted' which means that it is
700 // 'bitcasted' to another type which is handled.
701 // This operation also isn't safe with AND, OR, XOR when the boolean
702 // type is 0/1 as we need an all ones vector constant to mask with.
703 // FIXME: Sign extend 1 to all ones if thats legal on the target.
704 if (TLI.getOperationAction(ISD::AND, VT) == TargetLowering::Expand ||
705 TLI.getOperationAction(ISD::XOR, VT) == TargetLowering::Expand ||
706 TLI.getOperationAction(ISD::OR, VT) == TargetLowering::Expand ||
707 TLI.getBooleanContents(true) !=
708 TargetLowering::ZeroOrNegativeOneBooleanContent)
709 return DAG.UnrollVectorOp(Op.getNode());
711 // If the mask and the type are different sizes, unroll the vector op. This
712 // can occur when getSetCCResultType returns something that is different in
713 // size from the operand types. For example, v4i8 = select v4i32, v4i8, v4i8.
714 if (VT.getSizeInBits() != Op1.getValueType().getSizeInBits())
715 return DAG.UnrollVectorOp(Op.getNode());
717 // Bitcast the operands to be the same type as the mask.
718 // This is needed when we select between FP types because
719 // the mask is a vector of integers.
720 Op1 = DAG.getNode(ISD::BITCAST, DL, VT, Op1);
721 Op2 = DAG.getNode(ISD::BITCAST, DL, VT, Op2);
723 SDValue AllOnes = DAG.getConstant(
724 APInt::getAllOnesValue(VT.getScalarType().getSizeInBits()), VT);
725 SDValue NotMask = DAG.getNode(ISD::XOR, DL, VT, Mask, AllOnes);
727 Op1 = DAG.getNode(ISD::AND, DL, VT, Op1, Mask);
728 Op2 = DAG.getNode(ISD::AND, DL, VT, Op2, NotMask);
729 SDValue Val = DAG.getNode(ISD::OR, DL, VT, Op1, Op2);
730 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Val);
733 SDValue VectorLegalizer::ExpandUINT_TO_FLOAT(SDValue Op) {
734 EVT VT = Op.getOperand(0).getValueType();
737 // Make sure that the SINT_TO_FP and SRL instructions are available.
738 if (TLI.getOperationAction(ISD::SINT_TO_FP, VT) == TargetLowering::Expand ||
739 TLI.getOperationAction(ISD::SRL, VT) == TargetLowering::Expand)
740 return DAG.UnrollVectorOp(Op.getNode());
742 EVT SVT = VT.getScalarType();
743 assert((SVT.getSizeInBits() == 64 || SVT.getSizeInBits() == 32) &&
744 "Elements in vector-UINT_TO_FP must be 32 or 64 bits wide");
746 unsigned BW = SVT.getSizeInBits();
747 SDValue HalfWord = DAG.getConstant(BW/2, VT);
749 // Constants to clear the upper part of the word.
750 // Notice that we can also use SHL+SHR, but using a constant is slightly
752 uint64_t HWMask = (SVT.getSizeInBits()==64)?0x00000000FFFFFFFF:0x0000FFFF;
753 SDValue HalfWordMask = DAG.getConstant(HWMask, VT);
755 // Two to the power of half-word-size.
756 SDValue TWOHW = DAG.getConstantFP((1<<(BW/2)), Op.getValueType());
758 // Clear upper part of LO, lower HI
759 SDValue HI = DAG.getNode(ISD::SRL, DL, VT, Op.getOperand(0), HalfWord);
760 SDValue LO = DAG.getNode(ISD::AND, DL, VT, Op.getOperand(0), HalfWordMask);
762 // Convert hi and lo to floats
763 // Convert the hi part back to the upper values
764 SDValue fHI = DAG.getNode(ISD::SINT_TO_FP, DL, Op.getValueType(), HI);
765 fHI = DAG.getNode(ISD::FMUL, DL, Op.getValueType(), fHI, TWOHW);
766 SDValue fLO = DAG.getNode(ISD::SINT_TO_FP, DL, Op.getValueType(), LO);
768 // Add the two halves
769 return DAG.getNode(ISD::FADD, DL, Op.getValueType(), fHI, fLO);
773 SDValue VectorLegalizer::ExpandFNEG(SDValue Op) {
774 if (TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType())) {
775 SDValue Zero = DAG.getConstantFP(-0.0, Op.getValueType());
776 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
777 Zero, Op.getOperand(0));
779 return DAG.UnrollVectorOp(Op.getNode());
782 SDValue VectorLegalizer::UnrollVSETCC(SDValue Op) {
783 EVT VT = Op.getValueType();
784 unsigned NumElems = VT.getVectorNumElements();
785 EVT EltVT = VT.getVectorElementType();
786 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1), CC = Op.getOperand(2);
787 EVT TmpEltVT = LHS.getValueType().getVectorElementType();
789 SmallVector<SDValue, 8> Ops(NumElems);
790 for (unsigned i = 0; i < NumElems; ++i) {
791 SDValue LHSElem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, LHS,
792 DAG.getConstant(i, TLI.getVectorIdxTy()));
793 SDValue RHSElem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, RHS,
794 DAG.getConstant(i, TLI.getVectorIdxTy()));
795 Ops[i] = DAG.getNode(ISD::SETCC, dl,
796 TLI.getSetCCResultType(*DAG.getContext(), TmpEltVT),
797 LHSElem, RHSElem, CC);
798 Ops[i] = DAG.getSelect(dl, EltVT, Ops[i],
799 DAG.getConstant(APInt::getAllOnesValue
800 (EltVT.getSizeInBits()), EltVT),
801 DAG.getConstant(0, EltVT));
803 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
808 bool SelectionDAG::LegalizeVectors() {
809 return VectorLegalizer(*this).Run();