1 //===-- LegalizeVectorOps.cpp - Implement SelectionDAG::LegalizeVectors ---===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the SelectionDAG::LegalizeVectors method.
12 // The vector legalizer looks for vector operations which might need to be
13 // scalarized and legalizes them. This is a separate step from Legalize because
14 // scalarizing can introduce illegal types. For example, suppose we have an
15 // ISD::SDIV of type v2i64 on x86-32. The type is legal (for example, addition
16 // on a v2i64 is legal), but ISD::SDIV isn't legal, so we have to unroll the
17 // operation, which introduces nodes with the illegal type i64 which must be
18 // expanded. Similarly, suppose we have an ISD::SRA of type v16i8 on PowerPC;
19 // the operation must be unrolled, which introduces nodes with the illegal
20 // type i8 which must be promoted.
22 // This does not legalize vector manipulations like ISD::BUILD_VECTOR,
23 // or operations that happen to take a vector which are custom-lowered;
24 // the legalization for such operations never produces nodes
25 // with illegal types, so it's okay to put off legalizing them until
26 // SelectionDAG::Legalize runs.
28 //===----------------------------------------------------------------------===//
30 #include "llvm/CodeGen/SelectionDAG.h"
31 #include "llvm/Target/TargetLowering.h"
35 class VectorLegalizer {
37 const TargetLowering &TLI;
38 bool Changed; // Keep track of whether anything changed
40 /// For nodes that are of legal width, and that have more than one use, this
41 /// map indicates what regularized operand to use. This allows us to avoid
42 /// legalizing the same thing more than once.
43 SmallDenseMap<SDValue, SDValue, 64> LegalizedNodes;
45 /// \brief Adds a node to the translation cache.
46 void AddLegalizedOperand(SDValue From, SDValue To) {
47 LegalizedNodes.insert(std::make_pair(From, To));
48 // If someone requests legalization of the new node, return itself.
50 LegalizedNodes.insert(std::make_pair(To, To));
53 /// \brief Legalizes the given node.
54 SDValue LegalizeOp(SDValue Op);
56 /// \brief Assuming the node is legal, "legalize" the results.
57 SDValue TranslateLegalizeResults(SDValue Op, SDValue Result);
59 /// \brief Implements unrolling a VSETCC.
60 SDValue UnrollVSETCC(SDValue Op);
62 /// \brief Implement expand-based legalization of vector operations.
64 /// This is just a high-level routine to dispatch to specific code paths for
65 /// operations to legalize them.
66 SDValue Expand(SDValue Op);
68 /// \brief Implements expansion for FNEG; falls back to UnrollVectorOp if
71 /// Implements expansion for UINT_TO_FLOAT; falls back to UnrollVectorOp if
72 /// SINT_TO_FLOAT and SHR on vectors isn't legal.
73 SDValue ExpandUINT_TO_FLOAT(SDValue Op);
75 /// \brief Implement expansion for SIGN_EXTEND_INREG using SRL and SRA.
76 SDValue ExpandSEXTINREG(SDValue Op);
78 /// \brief Implement expansion for ANY_EXTEND_VECTOR_INREG.
80 /// Shuffles the low lanes of the operand into place and bitcasts to the proper
81 /// type. The contents of the bits in the extended part of each element are
83 SDValue ExpandANY_EXTEND_VECTOR_INREG(SDValue Op);
85 /// \brief Implement expansion for SIGN_EXTEND_VECTOR_INREG.
87 /// Shuffles the low lanes of the operand into place, bitcasts to the proper
88 /// type, then shifts left and arithmetic shifts right to introduce a sign
90 SDValue ExpandSIGN_EXTEND_VECTOR_INREG(SDValue Op);
92 /// \brief Implement expansion for ZERO_EXTEND_VECTOR_INREG.
94 /// Shuffles the low lanes of the operand into place and blends zeros into
95 /// the remaining lanes, finally bitcasting to the proper type.
96 SDValue ExpandZERO_EXTEND_VECTOR_INREG(SDValue Op);
98 /// \brief Expand bswap of vectors into a shuffle if legal.
99 SDValue ExpandBSWAP(SDValue Op);
101 /// \brief Implement vselect in terms of XOR, AND, OR when blend is not
102 /// supported by the target.
103 SDValue ExpandVSELECT(SDValue Op);
104 SDValue ExpandSELECT(SDValue Op);
105 SDValue ExpandLoad(SDValue Op);
106 SDValue ExpandStore(SDValue Op);
107 SDValue ExpandFNEG(SDValue Op);
108 SDValue ExpandBITREVERSE(SDValue Op);
109 SDValue ExpandCTLZ_CTTZ_ZERO_UNDEF(SDValue Op);
111 /// \brief Implements vector promotion.
113 /// This is essentially just bitcasting the operands to a different type and
114 /// bitcasting the result back to the original type.
115 SDValue Promote(SDValue Op);
117 /// \brief Implements [SU]INT_TO_FP vector promotion.
119 /// This is a [zs]ext of the input operand to the next size up.
120 SDValue PromoteINT_TO_FP(SDValue Op);
122 /// \brief Implements FP_TO_[SU]INT vector promotion of the result type.
124 /// It is promoted to the next size up integer type. The result is then
125 /// truncated back to the original type.
126 SDValue PromoteFP_TO_INT(SDValue Op, bool isSigned);
129 /// \brief Begin legalizer the vector operations in the DAG.
131 VectorLegalizer(SelectionDAG& dag) :
132 DAG(dag), TLI(dag.getTargetLoweringInfo()), Changed(false) {}
135 bool VectorLegalizer::Run() {
136 // Before we start legalizing vector nodes, check if there are any vectors.
137 bool HasVectors = false;
138 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
139 E = std::prev(DAG.allnodes_end()); I != std::next(E); ++I) {
140 // Check if the values of the nodes contain vectors. We don't need to check
141 // the operands because we are going to check their values at some point.
142 for (SDNode::value_iterator J = I->value_begin(), E = I->value_end();
144 HasVectors |= J->isVector();
146 // If we found a vector node we can start the legalization.
151 // If this basic block has no vectors then no need to legalize vectors.
155 // The legalize process is inherently a bottom-up recursive process (users
156 // legalize their uses before themselves). Given infinite stack space, we
157 // could just start legalizing on the root and traverse the whole graph. In
158 // practice however, this causes us to run out of stack space on large basic
159 // blocks. To avoid this problem, compute an ordering of the nodes where each
160 // node is only legalized after all of its operands are legalized.
161 DAG.AssignTopologicalOrder();
162 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
163 E = std::prev(DAG.allnodes_end()); I != std::next(E); ++I)
164 LegalizeOp(SDValue(&*I, 0));
166 // Finally, it's possible the root changed. Get the new root.
167 SDValue OldRoot = DAG.getRoot();
168 assert(LegalizedNodes.count(OldRoot) && "Root didn't get legalized?");
169 DAG.setRoot(LegalizedNodes[OldRoot]);
171 LegalizedNodes.clear();
173 // Remove dead nodes now.
174 DAG.RemoveDeadNodes();
179 SDValue VectorLegalizer::TranslateLegalizeResults(SDValue Op, SDValue Result) {
180 // Generic legalization: just pass the operand through.
181 for (unsigned i = 0, e = Op.getNode()->getNumValues(); i != e; ++i)
182 AddLegalizedOperand(Op.getValue(i), Result.getValue(i));
183 return Result.getValue(Op.getResNo());
186 SDValue VectorLegalizer::LegalizeOp(SDValue Op) {
187 // Note that LegalizeOp may be reentered even from single-use nodes, which
188 // means that we always must cache transformed nodes.
189 DenseMap<SDValue, SDValue>::iterator I = LegalizedNodes.find(Op);
190 if (I != LegalizedNodes.end()) return I->second;
192 SDNode* Node = Op.getNode();
194 // Legalize the operands
195 SmallVector<SDValue, 8> Ops;
196 for (const SDValue &Op : Node->op_values())
197 Ops.push_back(LegalizeOp(Op));
199 SDValue Result = SDValue(DAG.UpdateNodeOperands(Op.getNode(), Ops), 0);
201 bool HasVectorValue = false;
202 if (Op.getOpcode() == ISD::LOAD) {
203 LoadSDNode *LD = cast<LoadSDNode>(Op.getNode());
204 ISD::LoadExtType ExtType = LD->getExtensionType();
205 if (LD->getMemoryVT().isVector() && ExtType != ISD::NON_EXTLOAD)
206 switch (TLI.getLoadExtAction(LD->getExtensionType(), LD->getValueType(0),
207 LD->getMemoryVT())) {
208 default: llvm_unreachable("This action is not supported yet!");
209 case TargetLowering::Legal:
210 return TranslateLegalizeResults(Op, Result);
211 case TargetLowering::Custom:
212 if (SDValue Lowered = TLI.LowerOperation(Result, DAG)) {
213 if (Lowered == Result)
214 return TranslateLegalizeResults(Op, Lowered);
216 if (Lowered->getNumValues() != Op->getNumValues()) {
217 // This expanded to something other than the load. Assume the
218 // lowering code took care of any chain values, and just handle the
220 assert(Result.getValue(1).use_empty() &&
221 "There are still live users of the old chain!");
222 return LegalizeOp(Lowered);
224 return TranslateLegalizeResults(Op, Lowered);
226 case TargetLowering::Expand:
228 return LegalizeOp(ExpandLoad(Op));
230 } else if (Op.getOpcode() == ISD::STORE) {
231 StoreSDNode *ST = cast<StoreSDNode>(Op.getNode());
232 EVT StVT = ST->getMemoryVT();
233 MVT ValVT = ST->getValue().getSimpleValueType();
234 if (StVT.isVector() && ST->isTruncatingStore())
235 switch (TLI.getTruncStoreAction(ValVT, StVT)) {
236 default: llvm_unreachable("This action is not supported yet!");
237 case TargetLowering::Legal:
238 return TranslateLegalizeResults(Op, Result);
239 case TargetLowering::Custom: {
240 SDValue Lowered = TLI.LowerOperation(Result, DAG);
241 Changed = Lowered != Result;
242 return TranslateLegalizeResults(Op, Lowered);
244 case TargetLowering::Expand:
246 return LegalizeOp(ExpandStore(Op));
248 } else if (Op.getOpcode() == ISD::MSCATTER || Op.getOpcode() == ISD::MSTORE)
249 HasVectorValue = true;
251 for (SDNode::value_iterator J = Node->value_begin(), E = Node->value_end();
254 HasVectorValue |= J->isVector();
256 return TranslateLegalizeResults(Op, Result);
259 switch (Op.getOpcode()) {
261 return TranslateLegalizeResults(Op, Result);
285 case ISD::BITREVERSE:
288 case ISD::CTLZ_ZERO_UNDEF:
289 case ISD::CTTZ_ZERO_UNDEF:
295 case ISD::ZERO_EXTEND:
296 case ISD::ANY_EXTEND:
298 case ISD::SIGN_EXTEND:
299 case ISD::FP_TO_SINT:
300 case ISD::FP_TO_UINT:
321 case ISD::FNEARBYINT:
327 case ISD::SIGN_EXTEND_INREG:
328 case ISD::ANY_EXTEND_VECTOR_INREG:
329 case ISD::SIGN_EXTEND_VECTOR_INREG:
330 case ISD::ZERO_EXTEND_VECTOR_INREG:
335 QueryType = Node->getValueType(0);
337 case ISD::FP_ROUND_INREG:
338 QueryType = cast<VTSDNode>(Node->getOperand(1))->getVT();
340 case ISD::SINT_TO_FP:
341 case ISD::UINT_TO_FP:
342 QueryType = Node->getOperand(0).getValueType();
345 QueryType = cast<MaskedScatterSDNode>(Node)->getValue().getValueType();
348 QueryType = cast<MaskedStoreSDNode>(Node)->getValue().getValueType();
352 switch (TLI.getOperationAction(Node->getOpcode(), QueryType)) {
353 default: llvm_unreachable("This action is not supported yet!");
354 case TargetLowering::Promote:
355 Result = Promote(Op);
358 case TargetLowering::Legal:
360 case TargetLowering::Custom: {
361 if (SDValue Tmp1 = TLI.LowerOperation(Op, DAG)) {
367 case TargetLowering::Expand:
371 // Make sure that the generated code is itself legal.
373 Result = LegalizeOp(Result);
377 // Note that LegalizeOp may be reentered even from single-use nodes, which
378 // means that we always must cache transformed nodes.
379 AddLegalizedOperand(Op, Result);
383 SDValue VectorLegalizer::Promote(SDValue Op) {
384 // For a few operations there is a specific concept for promotion based on
385 // the operand's type.
386 switch (Op.getOpcode()) {
387 case ISD::SINT_TO_FP:
388 case ISD::UINT_TO_FP:
389 // "Promote" the operation by extending the operand.
390 return PromoteINT_TO_FP(Op);
391 case ISD::FP_TO_UINT:
392 case ISD::FP_TO_SINT:
393 // Promote the operation by extending the operand.
394 return PromoteFP_TO_INT(Op, Op->getOpcode() == ISD::FP_TO_SINT);
397 // There are currently two cases of vector promotion:
398 // 1) Bitcasting a vector of integers to a different type to a vector of the
399 // same overall length. For example, x86 promotes ISD::AND v2i32 to v1i64.
400 // 2) Extending a vector of floats to a vector of the same number of larger
401 // floats. For example, AArch64 promotes ISD::FADD on v4f16 to v4f32.
402 MVT VT = Op.getSimpleValueType();
403 assert(Op.getNode()->getNumValues() == 1 &&
404 "Can't promote a vector with multiple results!");
405 MVT NVT = TLI.getTypeToPromoteTo(Op.getOpcode(), VT);
407 SmallVector<SDValue, 4> Operands(Op.getNumOperands());
409 for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
410 if (Op.getOperand(j).getValueType().isVector())
413 .getVectorElementType()
414 .isFloatingPoint() &&
415 NVT.isVector() && NVT.getVectorElementType().isFloatingPoint())
416 Operands[j] = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Op.getOperand(j));
418 Operands[j] = DAG.getNode(ISD::BITCAST, dl, NVT, Op.getOperand(j));
420 Operands[j] = Op.getOperand(j);
423 Op = DAG.getNode(Op.getOpcode(), dl, NVT, Operands, Op.getNode()->getFlags());
424 if ((VT.isFloatingPoint() && NVT.isFloatingPoint()) ||
425 (VT.isVector() && VT.getVectorElementType().isFloatingPoint() &&
426 NVT.isVector() && NVT.getVectorElementType().isFloatingPoint()))
427 return DAG.getNode(ISD::FP_ROUND, dl, VT, Op, DAG.getIntPtrConstant(0, dl));
429 return DAG.getNode(ISD::BITCAST, dl, VT, Op);
432 SDValue VectorLegalizer::PromoteINT_TO_FP(SDValue Op) {
433 // INT_TO_FP operations may require the input operand be promoted even
434 // when the type is otherwise legal.
435 EVT VT = Op.getOperand(0).getValueType();
436 assert(Op.getNode()->getNumValues() == 1 &&
437 "Can't promote a vector with multiple results!");
439 // Normal getTypeToPromoteTo() doesn't work here, as that will promote
440 // by widening the vector w/ the same element width and twice the number
441 // of elements. We want the other way around, the same number of elements,
442 // each twice the width.
444 // Increase the bitwidth of the element to the next pow-of-two
445 // (which is greater than 8 bits).
447 EVT NVT = VT.widenIntegerVectorElementType(*DAG.getContext());
448 assert(NVT.isSimple() && "Promoting to a non-simple vector type!");
450 SmallVector<SDValue, 4> Operands(Op.getNumOperands());
452 unsigned Opc = Op.getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND :
454 for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
455 if (Op.getOperand(j).getValueType().isVector())
456 Operands[j] = DAG.getNode(Opc, dl, NVT, Op.getOperand(j));
458 Operands[j] = Op.getOperand(j);
461 return DAG.getNode(Op.getOpcode(), dl, Op.getValueType(), Operands);
464 // For FP_TO_INT we promote the result type to a vector type with wider
465 // elements and then truncate the result. This is different from the default
466 // PromoteVector which uses bitcast to promote thus assumning that the
467 // promoted vector type has the same overall size.
468 SDValue VectorLegalizer::PromoteFP_TO_INT(SDValue Op, bool isSigned) {
469 assert(Op.getNode()->getNumValues() == 1 &&
470 "Can't promote a vector with multiple results!");
471 EVT VT = Op.getValueType();
476 NewVT = VT.widenIntegerVectorElementType(*DAG.getContext());
477 assert(NewVT.isSimple() && "Promoting to a non-simple vector type!");
478 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NewVT)) {
479 NewOpc = ISD::FP_TO_SINT;
482 if (!isSigned && TLI.isOperationLegalOrCustom(ISD::FP_TO_UINT, NewVT)) {
483 NewOpc = ISD::FP_TO_UINT;
489 SDValue promoted = DAG.getNode(NewOpc, SDLoc(Op), NewVT, Op.getOperand(0));
490 return DAG.getNode(ISD::TRUNCATE, SDLoc(Op), VT, promoted);
494 SDValue VectorLegalizer::ExpandLoad(SDValue Op) {
495 LoadSDNode *LD = cast<LoadSDNode>(Op.getNode());
497 EVT SrcVT = LD->getMemoryVT();
498 EVT SrcEltVT = SrcVT.getScalarType();
499 unsigned NumElem = SrcVT.getVectorNumElements();
504 if (SrcVT.getVectorNumElements() > 1 && !SrcEltVT.isByteSized()) {
507 SmallVector<SDValue, 8> Vals;
508 SmallVector<SDValue, 8> LoadChains;
510 EVT DstEltVT = LD->getValueType(0).getScalarType();
511 SDValue Chain = LD->getChain();
512 SDValue BasePTR = LD->getBasePtr();
513 ISD::LoadExtType ExtType = LD->getExtensionType();
515 // When elements in a vector is not byte-addressable, we cannot directly
516 // load each element by advancing pointer, which could only address bytes.
517 // Instead, we load all significant words, mask bits off, and concatenate
518 // them to form each element. Finally, they are extended to destination
519 // scalar type to build the destination vector.
520 EVT WideVT = TLI.getPointerTy(DAG.getDataLayout());
522 assert(WideVT.isRound() &&
523 "Could not handle the sophisticated case when the widest integer is"
525 assert(WideVT.bitsGE(SrcEltVT) &&
526 "Type is not legalized?");
528 unsigned WideBytes = WideVT.getStoreSize();
530 unsigned RemainingBytes = SrcVT.getStoreSize();
531 SmallVector<SDValue, 8> LoadVals;
533 while (RemainingBytes > 0) {
535 unsigned LoadBytes = WideBytes;
537 if (RemainingBytes >= LoadBytes) {
539 DAG.getLoad(WideVT, dl, Chain, BasePTR,
540 LD->getPointerInfo().getWithOffset(Offset),
541 MinAlign(LD->getAlignment(), Offset),
542 LD->getMemOperand()->getFlags(), LD->getAAInfo());
545 while (RemainingBytes < LoadBytes) {
546 LoadBytes >>= 1; // Reduce the load size by half.
547 LoadVT = EVT::getIntegerVT(*DAG.getContext(), LoadBytes << 3);
550 DAG.getExtLoad(ISD::EXTLOAD, dl, WideVT, Chain, BasePTR,
551 LD->getPointerInfo().getWithOffset(Offset), LoadVT,
552 MinAlign(LD->getAlignment(), Offset),
553 LD->getMemOperand()->getFlags(), LD->getAAInfo());
556 RemainingBytes -= LoadBytes;
558 BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR,
559 DAG.getConstant(LoadBytes, dl,
560 BasePTR.getValueType()));
562 LoadVals.push_back(ScalarLoad.getValue(0));
563 LoadChains.push_back(ScalarLoad.getValue(1));
566 // Extract bits, pack and extend/trunc them into destination type.
567 unsigned SrcEltBits = SrcEltVT.getSizeInBits();
568 SDValue SrcEltBitMask = DAG.getConstant((1U << SrcEltBits) - 1, dl, WideVT);
570 unsigned BitOffset = 0;
571 unsigned WideIdx = 0;
572 unsigned WideBits = WideVT.getSizeInBits();
574 for (unsigned Idx = 0; Idx != NumElem; ++Idx) {
575 SDValue Lo, Hi, ShAmt;
577 if (BitOffset < WideBits) {
578 ShAmt = DAG.getConstant(
579 BitOffset, dl, TLI.getShiftAmountTy(WideVT, DAG.getDataLayout()));
580 Lo = DAG.getNode(ISD::SRL, dl, WideVT, LoadVals[WideIdx], ShAmt);
581 Lo = DAG.getNode(ISD::AND, dl, WideVT, Lo, SrcEltBitMask);
584 BitOffset += SrcEltBits;
585 if (BitOffset >= WideBits) {
587 BitOffset -= WideBits;
589 ShAmt = DAG.getConstant(
590 SrcEltBits - BitOffset, dl,
591 TLI.getShiftAmountTy(WideVT, DAG.getDataLayout()));
592 Hi = DAG.getNode(ISD::SHL, dl, WideVT, LoadVals[WideIdx], ShAmt);
593 Hi = DAG.getNode(ISD::AND, dl, WideVT, Hi, SrcEltBitMask);
598 Lo = DAG.getNode(ISD::OR, dl, WideVT, Lo, Hi);
601 default: llvm_unreachable("Unknown extended-load op!");
603 Lo = DAG.getAnyExtOrTrunc(Lo, dl, DstEltVT);
606 Lo = DAG.getZExtOrTrunc(Lo, dl, DstEltVT);
610 DAG.getConstant(WideBits - SrcEltBits, dl,
611 TLI.getShiftAmountTy(WideVT, DAG.getDataLayout()));
612 Lo = DAG.getNode(ISD::SHL, dl, WideVT, Lo, ShAmt);
613 Lo = DAG.getNode(ISD::SRA, dl, WideVT, Lo, ShAmt);
614 Lo = DAG.getSExtOrTrunc(Lo, dl, DstEltVT);
620 NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, LoadChains);
621 Value = DAG.getNode(ISD::BUILD_VECTOR, dl,
622 Op.getNode()->getValueType(0), Vals);
624 SDValue Scalarized = TLI.scalarizeVectorLoad(LD, DAG);
626 NewChain = Scalarized.getValue(1);
627 Value = Scalarized.getValue(0);
630 AddLegalizedOperand(Op.getValue(0), Value);
631 AddLegalizedOperand(Op.getValue(1), NewChain);
633 return (Op.getResNo() ? NewChain : Value);
636 SDValue VectorLegalizer::ExpandStore(SDValue Op) {
637 StoreSDNode *ST = cast<StoreSDNode>(Op.getNode());
639 EVT StVT = ST->getMemoryVT();
640 EVT MemSclVT = StVT.getScalarType();
641 unsigned ScalarSize = MemSclVT.getSizeInBits();
643 // Round odd types to the next pow of two.
644 if (!isPowerOf2_32(ScalarSize)) {
645 // FIXME: This is completely broken and inconsistent with ExpandLoad
648 // For sub-byte element sizes, this ends up with 0 stride between elements,
649 // so the same element just gets re-written to the same location. There seem
650 // to be tests explicitly testing for this broken behavior though. tests
651 // for this broken behavior.
653 LLVMContext &Ctx = *DAG.getContext();
656 = EVT::getVectorVT(Ctx,
657 MemSclVT.getIntegerVT(Ctx, NextPowerOf2(ScalarSize)),
658 StVT.getVectorNumElements());
660 SDValue NewVectorStore = DAG.getTruncStore(
661 ST->getChain(), SDLoc(Op), ST->getValue(), ST->getBasePtr(),
662 ST->getPointerInfo(), NewMemVT, ST->getAlignment(),
663 ST->getMemOperand()->getFlags(), ST->getAAInfo());
664 ST = cast<StoreSDNode>(NewVectorStore.getNode());
667 SDValue TF = TLI.scalarizeVectorStore(ST, DAG);
668 AddLegalizedOperand(Op, TF);
672 SDValue VectorLegalizer::Expand(SDValue Op) {
673 switch (Op->getOpcode()) {
674 case ISD::SIGN_EXTEND_INREG:
675 return ExpandSEXTINREG(Op);
676 case ISD::ANY_EXTEND_VECTOR_INREG:
677 return ExpandANY_EXTEND_VECTOR_INREG(Op);
678 case ISD::SIGN_EXTEND_VECTOR_INREG:
679 return ExpandSIGN_EXTEND_VECTOR_INREG(Op);
680 case ISD::ZERO_EXTEND_VECTOR_INREG:
681 return ExpandZERO_EXTEND_VECTOR_INREG(Op);
683 return ExpandBSWAP(Op);
685 return ExpandVSELECT(Op);
687 return ExpandSELECT(Op);
688 case ISD::UINT_TO_FP:
689 return ExpandUINT_TO_FLOAT(Op);
691 return ExpandFNEG(Op);
693 return UnrollVSETCC(Op);
694 case ISD::BITREVERSE:
695 return ExpandBITREVERSE(Op);
696 case ISD::CTLZ_ZERO_UNDEF:
697 case ISD::CTTZ_ZERO_UNDEF:
698 return ExpandCTLZ_CTTZ_ZERO_UNDEF(Op);
700 return DAG.UnrollVectorOp(Op.getNode());
704 SDValue VectorLegalizer::ExpandSELECT(SDValue Op) {
705 // Lower a select instruction where the condition is a scalar and the
706 // operands are vectors. Lower this select to VSELECT and implement it
707 // using XOR AND OR. The selector bit is broadcasted.
708 EVT VT = Op.getValueType();
711 SDValue Mask = Op.getOperand(0);
712 SDValue Op1 = Op.getOperand(1);
713 SDValue Op2 = Op.getOperand(2);
715 assert(VT.isVector() && !Mask.getValueType().isVector()
716 && Op1.getValueType() == Op2.getValueType() && "Invalid type");
718 unsigned NumElem = VT.getVectorNumElements();
720 // If we can't even use the basic vector operations of
721 // AND,OR,XOR, we will have to scalarize the op.
722 // Notice that the operation may be 'promoted' which means that it is
723 // 'bitcasted' to another type which is handled.
724 // Also, we need to be able to construct a splat vector using BUILD_VECTOR.
725 if (TLI.getOperationAction(ISD::AND, VT) == TargetLowering::Expand ||
726 TLI.getOperationAction(ISD::XOR, VT) == TargetLowering::Expand ||
727 TLI.getOperationAction(ISD::OR, VT) == TargetLowering::Expand ||
728 TLI.getOperationAction(ISD::BUILD_VECTOR, VT) == TargetLowering::Expand)
729 return DAG.UnrollVectorOp(Op.getNode());
731 // Generate a mask operand.
732 EVT MaskTy = VT.changeVectorElementTypeToInteger();
734 // What is the size of each element in the vector mask.
735 EVT BitTy = MaskTy.getScalarType();
737 Mask = DAG.getSelect(DL, BitTy, Mask,
738 DAG.getConstant(APInt::getAllOnesValue(BitTy.getSizeInBits()), DL,
740 DAG.getConstant(0, DL, BitTy));
742 // Broadcast the mask so that the entire vector is all-one or all zero.
743 SmallVector<SDValue, 8> Ops(NumElem, Mask);
744 Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, MaskTy, Ops);
746 // Bitcast the operands to be the same type as the mask.
747 // This is needed when we select between FP types because
748 // the mask is a vector of integers.
749 Op1 = DAG.getNode(ISD::BITCAST, DL, MaskTy, Op1);
750 Op2 = DAG.getNode(ISD::BITCAST, DL, MaskTy, Op2);
752 SDValue AllOnes = DAG.getConstant(
753 APInt::getAllOnesValue(BitTy.getSizeInBits()), DL, MaskTy);
754 SDValue NotMask = DAG.getNode(ISD::XOR, DL, MaskTy, Mask, AllOnes);
756 Op1 = DAG.getNode(ISD::AND, DL, MaskTy, Op1, Mask);
757 Op2 = DAG.getNode(ISD::AND, DL, MaskTy, Op2, NotMask);
758 SDValue Val = DAG.getNode(ISD::OR, DL, MaskTy, Op1, Op2);
759 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Val);
762 SDValue VectorLegalizer::ExpandSEXTINREG(SDValue Op) {
763 EVT VT = Op.getValueType();
765 // Make sure that the SRA and SHL instructions are available.
766 if (TLI.getOperationAction(ISD::SRA, VT) == TargetLowering::Expand ||
767 TLI.getOperationAction(ISD::SHL, VT) == TargetLowering::Expand)
768 return DAG.UnrollVectorOp(Op.getNode());
771 EVT OrigTy = cast<VTSDNode>(Op->getOperand(1))->getVT();
773 unsigned BW = VT.getScalarType().getSizeInBits();
774 unsigned OrigBW = OrigTy.getScalarType().getSizeInBits();
775 SDValue ShiftSz = DAG.getConstant(BW - OrigBW, DL, VT);
777 Op = Op.getOperand(0);
778 Op = DAG.getNode(ISD::SHL, DL, VT, Op, ShiftSz);
779 return DAG.getNode(ISD::SRA, DL, VT, Op, ShiftSz);
782 // Generically expand a vector anyext in register to a shuffle of the relevant
783 // lanes into the appropriate locations, with other lanes left undef.
784 SDValue VectorLegalizer::ExpandANY_EXTEND_VECTOR_INREG(SDValue Op) {
786 EVT VT = Op.getValueType();
787 int NumElements = VT.getVectorNumElements();
788 SDValue Src = Op.getOperand(0);
789 EVT SrcVT = Src.getValueType();
790 int NumSrcElements = SrcVT.getVectorNumElements();
792 // Build a base mask of undef shuffles.
793 SmallVector<int, 16> ShuffleMask;
794 ShuffleMask.resize(NumSrcElements, -1);
796 // Place the extended lanes into the correct locations.
797 int ExtLaneScale = NumSrcElements / NumElements;
798 int EndianOffset = DAG.getDataLayout().isBigEndian() ? ExtLaneScale - 1 : 0;
799 for (int i = 0; i < NumElements; ++i)
800 ShuffleMask[i * ExtLaneScale + EndianOffset] = i;
803 ISD::BITCAST, DL, VT,
804 DAG.getVectorShuffle(SrcVT, DL, Src, DAG.getUNDEF(SrcVT), ShuffleMask));
807 SDValue VectorLegalizer::ExpandSIGN_EXTEND_VECTOR_INREG(SDValue Op) {
809 EVT VT = Op.getValueType();
810 SDValue Src = Op.getOperand(0);
811 EVT SrcVT = Src.getValueType();
813 // First build an any-extend node which can be legalized above when we
814 // recurse through it.
815 Op = DAG.getAnyExtendVectorInReg(Src, DL, VT);
817 // Now we need sign extend. Do this by shifting the elements. Even if these
818 // aren't legal operations, they have a better chance of being legalized
819 // without full scalarization than the sign extension does.
820 unsigned EltWidth = VT.getVectorElementType().getSizeInBits();
821 unsigned SrcEltWidth = SrcVT.getVectorElementType().getSizeInBits();
822 SDValue ShiftAmount = DAG.getConstant(EltWidth - SrcEltWidth, DL, VT);
823 return DAG.getNode(ISD::SRA, DL, VT,
824 DAG.getNode(ISD::SHL, DL, VT, Op, ShiftAmount),
828 // Generically expand a vector zext in register to a shuffle of the relevant
829 // lanes into the appropriate locations, a blend of zero into the high bits,
830 // and a bitcast to the wider element type.
831 SDValue VectorLegalizer::ExpandZERO_EXTEND_VECTOR_INREG(SDValue Op) {
833 EVT VT = Op.getValueType();
834 int NumElements = VT.getVectorNumElements();
835 SDValue Src = Op.getOperand(0);
836 EVT SrcVT = Src.getValueType();
837 int NumSrcElements = SrcVT.getVectorNumElements();
839 // Build up a zero vector to blend into this one.
840 SDValue Zero = DAG.getConstant(0, DL, SrcVT);
842 // Shuffle the incoming lanes into the correct position, and pull all other
843 // lanes from the zero vector.
844 SmallVector<int, 16> ShuffleMask;
845 ShuffleMask.reserve(NumSrcElements);
846 for (int i = 0; i < NumSrcElements; ++i)
847 ShuffleMask.push_back(i);
849 int ExtLaneScale = NumSrcElements / NumElements;
850 int EndianOffset = DAG.getDataLayout().isBigEndian() ? ExtLaneScale - 1 : 0;
851 for (int i = 0; i < NumElements; ++i)
852 ShuffleMask[i * ExtLaneScale + EndianOffset] = NumSrcElements + i;
854 return DAG.getNode(ISD::BITCAST, DL, VT,
855 DAG.getVectorShuffle(SrcVT, DL, Zero, Src, ShuffleMask));
858 static void createBSWAPShuffleMask(EVT VT, SmallVectorImpl<int> &ShuffleMask) {
859 int ScalarSizeInBytes = VT.getScalarSizeInBits() / 8;
860 for (int I = 0, E = VT.getVectorNumElements(); I != E; ++I)
861 for (int J = ScalarSizeInBytes - 1; J >= 0; --J)
862 ShuffleMask.push_back((I * ScalarSizeInBytes) + J);
865 SDValue VectorLegalizer::ExpandBSWAP(SDValue Op) {
866 EVT VT = Op.getValueType();
868 // Generate a byte wise shuffle mask for the BSWAP.
869 SmallVector<int, 16> ShuffleMask;
870 createBSWAPShuffleMask(VT, ShuffleMask);
871 EVT ByteVT = EVT::getVectorVT(*DAG.getContext(), MVT::i8, ShuffleMask.size());
873 // Only emit a shuffle if the mask is legal.
874 if (!TLI.isShuffleMaskLegal(ShuffleMask, ByteVT))
875 return DAG.UnrollVectorOp(Op.getNode());
878 Op = DAG.getNode(ISD::BITCAST, DL, ByteVT, Op.getOperand(0));
879 Op = DAG.getVectorShuffle(ByteVT, DL, Op, DAG.getUNDEF(ByteVT), ShuffleMask);
880 return DAG.getNode(ISD::BITCAST, DL, VT, Op);
883 SDValue VectorLegalizer::ExpandBITREVERSE(SDValue Op) {
884 EVT VT = Op.getValueType();
886 // If we have the scalar operation, it's probably cheaper to unroll it.
887 if (TLI.isOperationLegalOrCustom(ISD::BITREVERSE, VT.getScalarType()))
888 return DAG.UnrollVectorOp(Op.getNode());
890 // If the vector element width is a whole number of bytes, test if its legal
891 // to BSWAP shuffle the bytes and then perform the BITREVERSE on the byte
892 // vector. This greatly reduces the number of bit shifts necessary.
893 unsigned ScalarSizeInBits = VT.getScalarSizeInBits();
894 if (ScalarSizeInBits > 8 && (ScalarSizeInBits % 8) == 0) {
895 SmallVector<int, 16> BSWAPMask;
896 createBSWAPShuffleMask(VT, BSWAPMask);
898 EVT ByteVT = EVT::getVectorVT(*DAG.getContext(), MVT::i8, BSWAPMask.size());
899 if (TLI.isShuffleMaskLegal(BSWAPMask, ByteVT) &&
900 (TLI.isOperationLegalOrCustom(ISD::BITREVERSE, ByteVT) ||
901 (TLI.isOperationLegalOrCustom(ISD::SHL, ByteVT) &&
902 TLI.isOperationLegalOrCustom(ISD::SRL, ByteVT) &&
903 TLI.isOperationLegalOrCustomOrPromote(ISD::AND, ByteVT) &&
904 TLI.isOperationLegalOrCustomOrPromote(ISD::OR, ByteVT)))) {
906 Op = DAG.getNode(ISD::BITCAST, DL, ByteVT, Op.getOperand(0));
907 Op = DAG.getVectorShuffle(ByteVT, DL, Op, DAG.getUNDEF(ByteVT),
909 Op = DAG.getNode(ISD::BITREVERSE, DL, ByteVT, Op);
910 return DAG.getNode(ISD::BITCAST, DL, VT, Op);
914 // If we have the appropriate vector bit operations, it is better to use them
915 // than unrolling and expanding each component.
916 if (!TLI.isOperationLegalOrCustom(ISD::SHL, VT) ||
917 !TLI.isOperationLegalOrCustom(ISD::SRL, VT) ||
918 !TLI.isOperationLegalOrCustomOrPromote(ISD::AND, VT) ||
919 !TLI.isOperationLegalOrCustomOrPromote(ISD::OR, VT))
920 return DAG.UnrollVectorOp(Op.getNode());
922 // Let LegalizeDAG handle this later.
926 SDValue VectorLegalizer::ExpandVSELECT(SDValue Op) {
927 // Implement VSELECT in terms of XOR, AND, OR
928 // on platforms which do not support blend natively.
931 SDValue Mask = Op.getOperand(0);
932 SDValue Op1 = Op.getOperand(1);
933 SDValue Op2 = Op.getOperand(2);
935 EVT VT = Mask.getValueType();
937 // If we can't even use the basic vector operations of
938 // AND,OR,XOR, we will have to scalarize the op.
939 // Notice that the operation may be 'promoted' which means that it is
940 // 'bitcasted' to another type which is handled.
941 // This operation also isn't safe with AND, OR, XOR when the boolean
942 // type is 0/1 as we need an all ones vector constant to mask with.
943 // FIXME: Sign extend 1 to all ones if thats legal on the target.
944 if (TLI.getOperationAction(ISD::AND, VT) == TargetLowering::Expand ||
945 TLI.getOperationAction(ISD::XOR, VT) == TargetLowering::Expand ||
946 TLI.getOperationAction(ISD::OR, VT) == TargetLowering::Expand ||
947 TLI.getBooleanContents(Op1.getValueType()) !=
948 TargetLowering::ZeroOrNegativeOneBooleanContent)
949 return DAG.UnrollVectorOp(Op.getNode());
951 // If the mask and the type are different sizes, unroll the vector op. This
952 // can occur when getSetCCResultType returns something that is different in
953 // size from the operand types. For example, v4i8 = select v4i32, v4i8, v4i8.
954 if (VT.getSizeInBits() != Op1.getValueType().getSizeInBits())
955 return DAG.UnrollVectorOp(Op.getNode());
957 // Bitcast the operands to be the same type as the mask.
958 // This is needed when we select between FP types because
959 // the mask is a vector of integers.
960 Op1 = DAG.getNode(ISD::BITCAST, DL, VT, Op1);
961 Op2 = DAG.getNode(ISD::BITCAST, DL, VT, Op2);
963 SDValue AllOnes = DAG.getConstant(
964 APInt::getAllOnesValue(VT.getScalarType().getSizeInBits()), DL, VT);
965 SDValue NotMask = DAG.getNode(ISD::XOR, DL, VT, Mask, AllOnes);
967 Op1 = DAG.getNode(ISD::AND, DL, VT, Op1, Mask);
968 Op2 = DAG.getNode(ISD::AND, DL, VT, Op2, NotMask);
969 SDValue Val = DAG.getNode(ISD::OR, DL, VT, Op1, Op2);
970 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Val);
973 SDValue VectorLegalizer::ExpandUINT_TO_FLOAT(SDValue Op) {
974 EVT VT = Op.getOperand(0).getValueType();
977 // Make sure that the SINT_TO_FP and SRL instructions are available.
978 if (TLI.getOperationAction(ISD::SINT_TO_FP, VT) == TargetLowering::Expand ||
979 TLI.getOperationAction(ISD::SRL, VT) == TargetLowering::Expand)
980 return DAG.UnrollVectorOp(Op.getNode());
982 EVT SVT = VT.getScalarType();
983 assert((SVT.getSizeInBits() == 64 || SVT.getSizeInBits() == 32) &&
984 "Elements in vector-UINT_TO_FP must be 32 or 64 bits wide");
986 unsigned BW = SVT.getSizeInBits();
987 SDValue HalfWord = DAG.getConstant(BW/2, DL, VT);
989 // Constants to clear the upper part of the word.
990 // Notice that we can also use SHL+SHR, but using a constant is slightly
992 uint64_t HWMask = (SVT.getSizeInBits()==64)?0x00000000FFFFFFFF:0x0000FFFF;
993 SDValue HalfWordMask = DAG.getConstant(HWMask, DL, VT);
995 // Two to the power of half-word-size.
996 SDValue TWOHW = DAG.getConstantFP(1 << (BW/2), DL, Op.getValueType());
998 // Clear upper part of LO, lower HI
999 SDValue HI = DAG.getNode(ISD::SRL, DL, VT, Op.getOperand(0), HalfWord);
1000 SDValue LO = DAG.getNode(ISD::AND, DL, VT, Op.getOperand(0), HalfWordMask);
1002 // Convert hi and lo to floats
1003 // Convert the hi part back to the upper values
1004 // TODO: Can any fast-math-flags be set on these nodes?
1005 SDValue fHI = DAG.getNode(ISD::SINT_TO_FP, DL, Op.getValueType(), HI);
1006 fHI = DAG.getNode(ISD::FMUL, DL, Op.getValueType(), fHI, TWOHW);
1007 SDValue fLO = DAG.getNode(ISD::SINT_TO_FP, DL, Op.getValueType(), LO);
1009 // Add the two halves
1010 return DAG.getNode(ISD::FADD, DL, Op.getValueType(), fHI, fLO);
1014 SDValue VectorLegalizer::ExpandFNEG(SDValue Op) {
1015 if (TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType())) {
1017 SDValue Zero = DAG.getConstantFP(-0.0, DL, Op.getValueType());
1018 // TODO: If FNEG had fast-math-flags, they'd get propagated to this FSUB.
1019 return DAG.getNode(ISD::FSUB, DL, Op.getValueType(),
1020 Zero, Op.getOperand(0));
1022 return DAG.UnrollVectorOp(Op.getNode());
1025 SDValue VectorLegalizer::ExpandCTLZ_CTTZ_ZERO_UNDEF(SDValue Op) {
1026 // If the non-ZERO_UNDEF version is supported we can use that instead.
1027 unsigned Opc = Op.getOpcode() == ISD::CTLZ_ZERO_UNDEF ? ISD::CTLZ : ISD::CTTZ;
1028 if (TLI.isOperationLegalOrCustom(Opc, Op.getValueType())) {
1030 return DAG.getNode(Opc, DL, Op.getValueType(), Op.getOperand(0));
1033 // Otherwise go ahead and unroll.
1034 return DAG.UnrollVectorOp(Op.getNode());
1037 SDValue VectorLegalizer::UnrollVSETCC(SDValue Op) {
1038 EVT VT = Op.getValueType();
1039 unsigned NumElems = VT.getVectorNumElements();
1040 EVT EltVT = VT.getVectorElementType();
1041 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1), CC = Op.getOperand(2);
1042 EVT TmpEltVT = LHS.getValueType().getVectorElementType();
1044 SmallVector<SDValue, 8> Ops(NumElems);
1045 for (unsigned i = 0; i < NumElems; ++i) {
1046 SDValue LHSElem = DAG.getNode(
1047 ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, LHS,
1048 DAG.getConstant(i, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
1049 SDValue RHSElem = DAG.getNode(
1050 ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, RHS,
1051 DAG.getConstant(i, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
1052 Ops[i] = DAG.getNode(ISD::SETCC, dl,
1053 TLI.getSetCCResultType(DAG.getDataLayout(),
1054 *DAG.getContext(), TmpEltVT),
1055 LHSElem, RHSElem, CC);
1056 Ops[i] = DAG.getSelect(dl, EltVT, Ops[i],
1057 DAG.getConstant(APInt::getAllOnesValue
1058 (EltVT.getSizeInBits()), dl, EltVT),
1059 DAG.getConstant(0, dl, EltVT));
1061 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
1066 bool SelectionDAG::LegalizeVectors() {
1067 return VectorLegalizer(*this).Run();