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 ExpandFSUB(SDValue Op);
109 SDValue ExpandBITREVERSE(SDValue Op);
110 SDValue ExpandCTLZ(SDValue Op);
111 SDValue ExpandCTTZ_ZERO_UNDEF(SDValue Op);
113 /// \brief Implements vector promotion.
115 /// This is essentially just bitcasting the operands to a different type and
116 /// bitcasting the result back to the original type.
117 SDValue Promote(SDValue Op);
119 /// \brief Implements [SU]INT_TO_FP vector promotion.
121 /// This is a [zs]ext of the input operand to the next size up.
122 SDValue PromoteINT_TO_FP(SDValue Op);
124 /// \brief Implements FP_TO_[SU]INT vector promotion of the result type.
126 /// It is promoted to the next size up integer type. The result is then
127 /// truncated back to the original type.
128 SDValue PromoteFP_TO_INT(SDValue Op, bool isSigned);
131 /// \brief Begin legalizer the vector operations in the DAG.
133 VectorLegalizer(SelectionDAG& dag) :
134 DAG(dag), TLI(dag.getTargetLoweringInfo()), Changed(false) {}
137 bool VectorLegalizer::Run() {
138 // Before we start legalizing vector nodes, check if there are any vectors.
139 bool HasVectors = false;
140 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
141 E = std::prev(DAG.allnodes_end()); I != std::next(E); ++I) {
142 // Check if the values of the nodes contain vectors. We don't need to check
143 // the operands because we are going to check their values at some point.
144 for (SDNode::value_iterator J = I->value_begin(), E = I->value_end();
146 HasVectors |= J->isVector();
148 // If we found a vector node we can start the legalization.
153 // If this basic block has no vectors then no need to legalize vectors.
157 // The legalize process is inherently a bottom-up recursive process (users
158 // legalize their uses before themselves). Given infinite stack space, we
159 // could just start legalizing on the root and traverse the whole graph. In
160 // practice however, this causes us to run out of stack space on large basic
161 // blocks. To avoid this problem, compute an ordering of the nodes where each
162 // node is only legalized after all of its operands are legalized.
163 DAG.AssignTopologicalOrder();
164 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
165 E = std::prev(DAG.allnodes_end()); I != std::next(E); ++I)
166 LegalizeOp(SDValue(&*I, 0));
168 // Finally, it's possible the root changed. Get the new root.
169 SDValue OldRoot = DAG.getRoot();
170 assert(LegalizedNodes.count(OldRoot) && "Root didn't get legalized?");
171 DAG.setRoot(LegalizedNodes[OldRoot]);
173 LegalizedNodes.clear();
175 // Remove dead nodes now.
176 DAG.RemoveDeadNodes();
181 SDValue VectorLegalizer::TranslateLegalizeResults(SDValue Op, SDValue Result) {
182 // Generic legalization: just pass the operand through.
183 for (unsigned i = 0, e = Op.getNode()->getNumValues(); i != e; ++i)
184 AddLegalizedOperand(Op.getValue(i), Result.getValue(i));
185 return Result.getValue(Op.getResNo());
188 SDValue VectorLegalizer::LegalizeOp(SDValue Op) {
189 // Note that LegalizeOp may be reentered even from single-use nodes, which
190 // means that we always must cache transformed nodes.
191 DenseMap<SDValue, SDValue>::iterator I = LegalizedNodes.find(Op);
192 if (I != LegalizedNodes.end()) return I->second;
194 SDNode* Node = Op.getNode();
196 // Legalize the operands
197 SmallVector<SDValue, 8> Ops;
198 for (const SDValue &Op : Node->op_values())
199 Ops.push_back(LegalizeOp(Op));
201 SDValue Result = SDValue(DAG.UpdateNodeOperands(Op.getNode(), Ops), 0);
203 bool HasVectorValue = false;
204 if (Op.getOpcode() == ISD::LOAD) {
205 LoadSDNode *LD = cast<LoadSDNode>(Op.getNode());
206 ISD::LoadExtType ExtType = LD->getExtensionType();
207 if (LD->getMemoryVT().isVector() && ExtType != ISD::NON_EXTLOAD)
208 switch (TLI.getLoadExtAction(LD->getExtensionType(), LD->getValueType(0),
209 LD->getMemoryVT())) {
210 default: llvm_unreachable("This action is not supported yet!");
211 case TargetLowering::Legal:
212 return TranslateLegalizeResults(Op, Result);
213 case TargetLowering::Custom:
214 if (SDValue Lowered = TLI.LowerOperation(Result, DAG)) {
215 if (Lowered == Result)
216 return TranslateLegalizeResults(Op, Lowered);
218 if (Lowered->getNumValues() != Op->getNumValues()) {
219 // This expanded to something other than the load. Assume the
220 // lowering code took care of any chain values, and just handle the
222 assert(Result.getValue(1).use_empty() &&
223 "There are still live users of the old chain!");
224 return LegalizeOp(Lowered);
226 return TranslateLegalizeResults(Op, Lowered);
228 case TargetLowering::Expand:
230 return LegalizeOp(ExpandLoad(Op));
232 } else if (Op.getOpcode() == ISD::STORE) {
233 StoreSDNode *ST = cast<StoreSDNode>(Op.getNode());
234 EVT StVT = ST->getMemoryVT();
235 MVT ValVT = ST->getValue().getSimpleValueType();
236 if (StVT.isVector() && ST->isTruncatingStore())
237 switch (TLI.getTruncStoreAction(ValVT, StVT)) {
238 default: llvm_unreachable("This action is not supported yet!");
239 case TargetLowering::Legal:
240 return TranslateLegalizeResults(Op, Result);
241 case TargetLowering::Custom: {
242 SDValue Lowered = TLI.LowerOperation(Result, DAG);
243 Changed = Lowered != Result;
244 return TranslateLegalizeResults(Op, Lowered);
246 case TargetLowering::Expand:
248 return LegalizeOp(ExpandStore(Op));
250 } else if (Op.getOpcode() == ISD::MSCATTER || Op.getOpcode() == ISD::MSTORE)
251 HasVectorValue = true;
253 for (SDNode::value_iterator J = Node->value_begin(), E = Node->value_end();
256 HasVectorValue |= J->isVector();
258 return TranslateLegalizeResults(Op, Result);
261 switch (Op.getOpcode()) {
263 return TranslateLegalizeResults(Op, Result);
287 case ISD::BITREVERSE:
290 case ISD::CTLZ_ZERO_UNDEF:
291 case ISD::CTTZ_ZERO_UNDEF:
297 case ISD::ZERO_EXTEND:
298 case ISD::ANY_EXTEND:
300 case ISD::SIGN_EXTEND:
301 case ISD::FP_TO_SINT:
302 case ISD::FP_TO_UINT:
323 case ISD::FNEARBYINT:
329 case ISD::SIGN_EXTEND_INREG:
330 case ISD::ANY_EXTEND_VECTOR_INREG:
331 case ISD::SIGN_EXTEND_VECTOR_INREG:
332 case ISD::ZERO_EXTEND_VECTOR_INREG:
339 QueryType = Node->getValueType(0);
341 case ISD::FP_ROUND_INREG:
342 QueryType = cast<VTSDNode>(Node->getOperand(1))->getVT();
344 case ISD::SINT_TO_FP:
345 case ISD::UINT_TO_FP:
346 QueryType = Node->getOperand(0).getValueType();
349 QueryType = cast<MaskedScatterSDNode>(Node)->getValue().getValueType();
352 QueryType = cast<MaskedStoreSDNode>(Node)->getValue().getValueType();
356 switch (TLI.getOperationAction(Node->getOpcode(), QueryType)) {
357 default: llvm_unreachable("This action is not supported yet!");
358 case TargetLowering::Promote:
359 Result = Promote(Op);
362 case TargetLowering::Legal:
364 case TargetLowering::Custom: {
365 if (SDValue Tmp1 = TLI.LowerOperation(Op, DAG)) {
371 case TargetLowering::Expand:
375 // Make sure that the generated code is itself legal.
377 Result = LegalizeOp(Result);
381 // Note that LegalizeOp may be reentered even from single-use nodes, which
382 // means that we always must cache transformed nodes.
383 AddLegalizedOperand(Op, Result);
387 SDValue VectorLegalizer::Promote(SDValue Op) {
388 // For a few operations there is a specific concept for promotion based on
389 // the operand's type.
390 switch (Op.getOpcode()) {
391 case ISD::SINT_TO_FP:
392 case ISD::UINT_TO_FP:
393 // "Promote" the operation by extending the operand.
394 return PromoteINT_TO_FP(Op);
395 case ISD::FP_TO_UINT:
396 case ISD::FP_TO_SINT:
397 // Promote the operation by extending the operand.
398 return PromoteFP_TO_INT(Op, Op->getOpcode() == ISD::FP_TO_SINT);
401 // There are currently two cases of vector promotion:
402 // 1) Bitcasting a vector of integers to a different type to a vector of the
403 // same overall length. For example, x86 promotes ISD::AND v2i32 to v1i64.
404 // 2) Extending a vector of floats to a vector of the same number of larger
405 // floats. For example, AArch64 promotes ISD::FADD on v4f16 to v4f32.
406 MVT VT = Op.getSimpleValueType();
407 assert(Op.getNode()->getNumValues() == 1 &&
408 "Can't promote a vector with multiple results!");
409 MVT NVT = TLI.getTypeToPromoteTo(Op.getOpcode(), VT);
411 SmallVector<SDValue, 4> Operands(Op.getNumOperands());
413 for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
414 if (Op.getOperand(j).getValueType().isVector())
417 .getVectorElementType()
418 .isFloatingPoint() &&
419 NVT.isVector() && NVT.getVectorElementType().isFloatingPoint())
420 Operands[j] = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Op.getOperand(j));
422 Operands[j] = DAG.getNode(ISD::BITCAST, dl, NVT, Op.getOperand(j));
424 Operands[j] = Op.getOperand(j);
427 Op = DAG.getNode(Op.getOpcode(), dl, NVT, Operands, Op.getNode()->getFlags());
428 if ((VT.isFloatingPoint() && NVT.isFloatingPoint()) ||
429 (VT.isVector() && VT.getVectorElementType().isFloatingPoint() &&
430 NVT.isVector() && NVT.getVectorElementType().isFloatingPoint()))
431 return DAG.getNode(ISD::FP_ROUND, dl, VT, Op, DAG.getIntPtrConstant(0, dl));
433 return DAG.getNode(ISD::BITCAST, dl, VT, Op);
436 SDValue VectorLegalizer::PromoteINT_TO_FP(SDValue Op) {
437 // INT_TO_FP operations may require the input operand be promoted even
438 // when the type is otherwise legal.
439 EVT VT = Op.getOperand(0).getValueType();
440 assert(Op.getNode()->getNumValues() == 1 &&
441 "Can't promote a vector with multiple results!");
443 // Normal getTypeToPromoteTo() doesn't work here, as that will promote
444 // by widening the vector w/ the same element width and twice the number
445 // of elements. We want the other way around, the same number of elements,
446 // each twice the width.
448 // Increase the bitwidth of the element to the next pow-of-two
449 // (which is greater than 8 bits).
451 EVT NVT = VT.widenIntegerVectorElementType(*DAG.getContext());
452 assert(NVT.isSimple() && "Promoting to a non-simple vector type!");
454 SmallVector<SDValue, 4> Operands(Op.getNumOperands());
456 unsigned Opc = Op.getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND :
458 for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
459 if (Op.getOperand(j).getValueType().isVector())
460 Operands[j] = DAG.getNode(Opc, dl, NVT, Op.getOperand(j));
462 Operands[j] = Op.getOperand(j);
465 return DAG.getNode(Op.getOpcode(), dl, Op.getValueType(), Operands);
468 // For FP_TO_INT we promote the result type to a vector type with wider
469 // elements and then truncate the result. This is different from the default
470 // PromoteVector which uses bitcast to promote thus assumning that the
471 // promoted vector type has the same overall size.
472 SDValue VectorLegalizer::PromoteFP_TO_INT(SDValue Op, bool isSigned) {
473 assert(Op.getNode()->getNumValues() == 1 &&
474 "Can't promote a vector with multiple results!");
475 EVT VT = Op.getValueType();
480 NewVT = VT.widenIntegerVectorElementType(*DAG.getContext());
481 assert(NewVT.isSimple() && "Promoting to a non-simple vector type!");
482 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NewVT)) {
483 NewOpc = ISD::FP_TO_SINT;
486 if (!isSigned && TLI.isOperationLegalOrCustom(ISD::FP_TO_UINT, NewVT)) {
487 NewOpc = ISD::FP_TO_UINT;
493 SDValue promoted = DAG.getNode(NewOpc, SDLoc(Op), NewVT, Op.getOperand(0));
494 return DAG.getNode(ISD::TRUNCATE, SDLoc(Op), VT, promoted);
498 SDValue VectorLegalizer::ExpandLoad(SDValue Op) {
499 LoadSDNode *LD = cast<LoadSDNode>(Op.getNode());
501 EVT SrcVT = LD->getMemoryVT();
502 EVT SrcEltVT = SrcVT.getScalarType();
503 unsigned NumElem = SrcVT.getVectorNumElements();
508 if (SrcVT.getVectorNumElements() > 1 && !SrcEltVT.isByteSized()) {
511 SmallVector<SDValue, 8> Vals;
512 SmallVector<SDValue, 8> LoadChains;
514 EVT DstEltVT = LD->getValueType(0).getScalarType();
515 SDValue Chain = LD->getChain();
516 SDValue BasePTR = LD->getBasePtr();
517 ISD::LoadExtType ExtType = LD->getExtensionType();
519 // When elements in a vector is not byte-addressable, we cannot directly
520 // load each element by advancing pointer, which could only address bytes.
521 // Instead, we load all significant words, mask bits off, and concatenate
522 // them to form each element. Finally, they are extended to destination
523 // scalar type to build the destination vector.
524 EVT WideVT = TLI.getPointerTy(DAG.getDataLayout());
526 assert(WideVT.isRound() &&
527 "Could not handle the sophisticated case when the widest integer is"
529 assert(WideVT.bitsGE(SrcEltVT) &&
530 "Type is not legalized?");
532 unsigned WideBytes = WideVT.getStoreSize();
534 unsigned RemainingBytes = SrcVT.getStoreSize();
535 SmallVector<SDValue, 8> LoadVals;
537 while (RemainingBytes > 0) {
539 unsigned LoadBytes = WideBytes;
541 if (RemainingBytes >= LoadBytes) {
543 DAG.getLoad(WideVT, dl, Chain, BasePTR,
544 LD->getPointerInfo().getWithOffset(Offset),
545 MinAlign(LD->getAlignment(), Offset),
546 LD->getMemOperand()->getFlags(), LD->getAAInfo());
549 while (RemainingBytes < LoadBytes) {
550 LoadBytes >>= 1; // Reduce the load size by half.
551 LoadVT = EVT::getIntegerVT(*DAG.getContext(), LoadBytes << 3);
554 DAG.getExtLoad(ISD::EXTLOAD, dl, WideVT, Chain, BasePTR,
555 LD->getPointerInfo().getWithOffset(Offset), LoadVT,
556 MinAlign(LD->getAlignment(), Offset),
557 LD->getMemOperand()->getFlags(), LD->getAAInfo());
560 RemainingBytes -= LoadBytes;
562 BasePTR = DAG.getNode(ISD::ADD, dl, BasePTR.getValueType(), BasePTR,
563 DAG.getConstant(LoadBytes, dl,
564 BasePTR.getValueType()));
566 LoadVals.push_back(ScalarLoad.getValue(0));
567 LoadChains.push_back(ScalarLoad.getValue(1));
570 // Extract bits, pack and extend/trunc them into destination type.
571 unsigned SrcEltBits = SrcEltVT.getSizeInBits();
572 SDValue SrcEltBitMask = DAG.getConstant((1U << SrcEltBits) - 1, dl, WideVT);
574 unsigned BitOffset = 0;
575 unsigned WideIdx = 0;
576 unsigned WideBits = WideVT.getSizeInBits();
578 for (unsigned Idx = 0; Idx != NumElem; ++Idx) {
579 SDValue Lo, Hi, ShAmt;
581 if (BitOffset < WideBits) {
582 ShAmt = DAG.getConstant(
583 BitOffset, dl, TLI.getShiftAmountTy(WideVT, DAG.getDataLayout()));
584 Lo = DAG.getNode(ISD::SRL, dl, WideVT, LoadVals[WideIdx], ShAmt);
585 Lo = DAG.getNode(ISD::AND, dl, WideVT, Lo, SrcEltBitMask);
588 BitOffset += SrcEltBits;
589 if (BitOffset >= WideBits) {
591 BitOffset -= WideBits;
593 ShAmt = DAG.getConstant(
594 SrcEltBits - BitOffset, dl,
595 TLI.getShiftAmountTy(WideVT, DAG.getDataLayout()));
596 Hi = DAG.getNode(ISD::SHL, dl, WideVT, LoadVals[WideIdx], ShAmt);
597 Hi = DAG.getNode(ISD::AND, dl, WideVT, Hi, SrcEltBitMask);
602 Lo = DAG.getNode(ISD::OR, dl, WideVT, Lo, Hi);
605 default: llvm_unreachable("Unknown extended-load op!");
607 Lo = DAG.getAnyExtOrTrunc(Lo, dl, DstEltVT);
610 Lo = DAG.getZExtOrTrunc(Lo, dl, DstEltVT);
614 DAG.getConstant(WideBits - SrcEltBits, dl,
615 TLI.getShiftAmountTy(WideVT, DAG.getDataLayout()));
616 Lo = DAG.getNode(ISD::SHL, dl, WideVT, Lo, ShAmt);
617 Lo = DAG.getNode(ISD::SRA, dl, WideVT, Lo, ShAmt);
618 Lo = DAG.getSExtOrTrunc(Lo, dl, DstEltVT);
624 NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, LoadChains);
625 Value = DAG.getBuildVector(Op.getNode()->getValueType(0), dl, Vals);
627 SDValue Scalarized = TLI.scalarizeVectorLoad(LD, DAG);
629 NewChain = Scalarized.getValue(1);
630 Value = Scalarized.getValue(0);
633 AddLegalizedOperand(Op.getValue(0), Value);
634 AddLegalizedOperand(Op.getValue(1), NewChain);
636 return (Op.getResNo() ? NewChain : Value);
639 SDValue VectorLegalizer::ExpandStore(SDValue Op) {
640 StoreSDNode *ST = cast<StoreSDNode>(Op.getNode());
642 EVT StVT = ST->getMemoryVT();
643 EVT MemSclVT = StVT.getScalarType();
644 unsigned ScalarSize = MemSclVT.getSizeInBits();
646 // Round odd types to the next pow of two.
647 if (!isPowerOf2_32(ScalarSize)) {
648 // FIXME: This is completely broken and inconsistent with ExpandLoad
651 // For sub-byte element sizes, this ends up with 0 stride between elements,
652 // so the same element just gets re-written to the same location. There seem
653 // to be tests explicitly testing for this broken behavior though. tests
654 // for this broken behavior.
656 LLVMContext &Ctx = *DAG.getContext();
659 = EVT::getVectorVT(Ctx,
660 MemSclVT.getIntegerVT(Ctx, NextPowerOf2(ScalarSize)),
661 StVT.getVectorNumElements());
663 SDValue NewVectorStore = DAG.getTruncStore(
664 ST->getChain(), SDLoc(Op), ST->getValue(), ST->getBasePtr(),
665 ST->getPointerInfo(), NewMemVT, ST->getAlignment(),
666 ST->getMemOperand()->getFlags(), ST->getAAInfo());
667 ST = cast<StoreSDNode>(NewVectorStore.getNode());
670 SDValue TF = TLI.scalarizeVectorStore(ST, DAG);
671 AddLegalizedOperand(Op, TF);
675 SDValue VectorLegalizer::Expand(SDValue Op) {
676 switch (Op->getOpcode()) {
677 case ISD::SIGN_EXTEND_INREG:
678 return ExpandSEXTINREG(Op);
679 case ISD::ANY_EXTEND_VECTOR_INREG:
680 return ExpandANY_EXTEND_VECTOR_INREG(Op);
681 case ISD::SIGN_EXTEND_VECTOR_INREG:
682 return ExpandSIGN_EXTEND_VECTOR_INREG(Op);
683 case ISD::ZERO_EXTEND_VECTOR_INREG:
684 return ExpandZERO_EXTEND_VECTOR_INREG(Op);
686 return ExpandBSWAP(Op);
688 return ExpandVSELECT(Op);
690 return ExpandSELECT(Op);
691 case ISD::UINT_TO_FP:
692 return ExpandUINT_TO_FLOAT(Op);
694 return ExpandFNEG(Op);
696 return ExpandFSUB(Op);
698 return UnrollVSETCC(Op);
699 case ISD::BITREVERSE:
700 return ExpandBITREVERSE(Op);
702 case ISD::CTLZ_ZERO_UNDEF:
703 return ExpandCTLZ(Op);
704 case ISD::CTTZ_ZERO_UNDEF:
705 return ExpandCTTZ_ZERO_UNDEF(Op);
707 return DAG.UnrollVectorOp(Op.getNode());
711 SDValue VectorLegalizer::ExpandSELECT(SDValue Op) {
712 // Lower a select instruction where the condition is a scalar and the
713 // operands are vectors. Lower this select to VSELECT and implement it
714 // using XOR AND OR. The selector bit is broadcasted.
715 EVT VT = Op.getValueType();
718 SDValue Mask = Op.getOperand(0);
719 SDValue Op1 = Op.getOperand(1);
720 SDValue Op2 = Op.getOperand(2);
722 assert(VT.isVector() && !Mask.getValueType().isVector()
723 && Op1.getValueType() == Op2.getValueType() && "Invalid type");
725 // If we can't even use the basic vector operations of
726 // AND,OR,XOR, we will have to scalarize the op.
727 // Notice that the operation may be 'promoted' which means that it is
728 // 'bitcasted' to another type which is handled.
729 // Also, we need to be able to construct a splat vector using BUILD_VECTOR.
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.getOperationAction(ISD::BUILD_VECTOR, VT) == TargetLowering::Expand)
734 return DAG.UnrollVectorOp(Op.getNode());
736 // Generate a mask operand.
737 EVT MaskTy = VT.changeVectorElementTypeToInteger();
739 // What is the size of each element in the vector mask.
740 EVT BitTy = MaskTy.getScalarType();
742 Mask = DAG.getSelect(DL, BitTy, Mask,
743 DAG.getConstant(APInt::getAllOnesValue(BitTy.getSizeInBits()), DL,
745 DAG.getConstant(0, DL, BitTy));
747 // Broadcast the mask so that the entire vector is all-one or all zero.
748 Mask = DAG.getSplatBuildVector(MaskTy, DL, Mask);
750 // Bitcast the operands to be the same type as the mask.
751 // This is needed when we select between FP types because
752 // the mask is a vector of integers.
753 Op1 = DAG.getNode(ISD::BITCAST, DL, MaskTy, Op1);
754 Op2 = DAG.getNode(ISD::BITCAST, DL, MaskTy, Op2);
756 SDValue AllOnes = DAG.getConstant(
757 APInt::getAllOnesValue(BitTy.getSizeInBits()), DL, MaskTy);
758 SDValue NotMask = DAG.getNode(ISD::XOR, DL, MaskTy, Mask, AllOnes);
760 Op1 = DAG.getNode(ISD::AND, DL, MaskTy, Op1, Mask);
761 Op2 = DAG.getNode(ISD::AND, DL, MaskTy, Op2, NotMask);
762 SDValue Val = DAG.getNode(ISD::OR, DL, MaskTy, Op1, Op2);
763 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Val);
766 SDValue VectorLegalizer::ExpandSEXTINREG(SDValue Op) {
767 EVT VT = Op.getValueType();
769 // Make sure that the SRA and SHL instructions are available.
770 if (TLI.getOperationAction(ISD::SRA, VT) == TargetLowering::Expand ||
771 TLI.getOperationAction(ISD::SHL, VT) == TargetLowering::Expand)
772 return DAG.UnrollVectorOp(Op.getNode());
775 EVT OrigTy = cast<VTSDNode>(Op->getOperand(1))->getVT();
777 unsigned BW = VT.getScalarSizeInBits();
778 unsigned OrigBW = OrigTy.getScalarSizeInBits();
779 SDValue ShiftSz = DAG.getConstant(BW - OrigBW, DL, VT);
781 Op = Op.getOperand(0);
782 Op = DAG.getNode(ISD::SHL, DL, VT, Op, ShiftSz);
783 return DAG.getNode(ISD::SRA, DL, VT, Op, ShiftSz);
786 // Generically expand a vector anyext in register to a shuffle of the relevant
787 // lanes into the appropriate locations, with other lanes left undef.
788 SDValue VectorLegalizer::ExpandANY_EXTEND_VECTOR_INREG(SDValue Op) {
790 EVT VT = Op.getValueType();
791 int NumElements = VT.getVectorNumElements();
792 SDValue Src = Op.getOperand(0);
793 EVT SrcVT = Src.getValueType();
794 int NumSrcElements = SrcVT.getVectorNumElements();
796 // Build a base mask of undef shuffles.
797 SmallVector<int, 16> ShuffleMask;
798 ShuffleMask.resize(NumSrcElements, -1);
800 // Place the extended lanes into the correct locations.
801 int ExtLaneScale = NumSrcElements / NumElements;
802 int EndianOffset = DAG.getDataLayout().isBigEndian() ? ExtLaneScale - 1 : 0;
803 for (int i = 0; i < NumElements; ++i)
804 ShuffleMask[i * ExtLaneScale + EndianOffset] = i;
807 ISD::BITCAST, DL, VT,
808 DAG.getVectorShuffle(SrcVT, DL, Src, DAG.getUNDEF(SrcVT), ShuffleMask));
811 SDValue VectorLegalizer::ExpandSIGN_EXTEND_VECTOR_INREG(SDValue Op) {
813 EVT VT = Op.getValueType();
814 SDValue Src = Op.getOperand(0);
815 EVT SrcVT = Src.getValueType();
817 // First build an any-extend node which can be legalized above when we
818 // recurse through it.
819 Op = DAG.getAnyExtendVectorInReg(Src, DL, VT);
821 // Now we need sign extend. Do this by shifting the elements. Even if these
822 // aren't legal operations, they have a better chance of being legalized
823 // without full scalarization than the sign extension does.
824 unsigned EltWidth = VT.getScalarSizeInBits();
825 unsigned SrcEltWidth = SrcVT.getScalarSizeInBits();
826 SDValue ShiftAmount = DAG.getConstant(EltWidth - SrcEltWidth, DL, VT);
827 return DAG.getNode(ISD::SRA, DL, VT,
828 DAG.getNode(ISD::SHL, DL, VT, Op, ShiftAmount),
832 // Generically expand a vector zext in register to a shuffle of the relevant
833 // lanes into the appropriate locations, a blend of zero into the high bits,
834 // and a bitcast to the wider element type.
835 SDValue VectorLegalizer::ExpandZERO_EXTEND_VECTOR_INREG(SDValue Op) {
837 EVT VT = Op.getValueType();
838 int NumElements = VT.getVectorNumElements();
839 SDValue Src = Op.getOperand(0);
840 EVT SrcVT = Src.getValueType();
841 int NumSrcElements = SrcVT.getVectorNumElements();
843 // Build up a zero vector to blend into this one.
844 SDValue Zero = DAG.getConstant(0, DL, SrcVT);
846 // Shuffle the incoming lanes into the correct position, and pull all other
847 // lanes from the zero vector.
848 SmallVector<int, 16> ShuffleMask;
849 ShuffleMask.reserve(NumSrcElements);
850 for (int i = 0; i < NumSrcElements; ++i)
851 ShuffleMask.push_back(i);
853 int ExtLaneScale = NumSrcElements / NumElements;
854 int EndianOffset = DAG.getDataLayout().isBigEndian() ? ExtLaneScale - 1 : 0;
855 for (int i = 0; i < NumElements; ++i)
856 ShuffleMask[i * ExtLaneScale + EndianOffset] = NumSrcElements + i;
858 return DAG.getNode(ISD::BITCAST, DL, VT,
859 DAG.getVectorShuffle(SrcVT, DL, Zero, Src, ShuffleMask));
862 static void createBSWAPShuffleMask(EVT VT, SmallVectorImpl<int> &ShuffleMask) {
863 int ScalarSizeInBytes = VT.getScalarSizeInBits() / 8;
864 for (int I = 0, E = VT.getVectorNumElements(); I != E; ++I)
865 for (int J = ScalarSizeInBytes - 1; J >= 0; --J)
866 ShuffleMask.push_back((I * ScalarSizeInBytes) + J);
869 SDValue VectorLegalizer::ExpandBSWAP(SDValue Op) {
870 EVT VT = Op.getValueType();
872 // Generate a byte wise shuffle mask for the BSWAP.
873 SmallVector<int, 16> ShuffleMask;
874 createBSWAPShuffleMask(VT, ShuffleMask);
875 EVT ByteVT = EVT::getVectorVT(*DAG.getContext(), MVT::i8, ShuffleMask.size());
877 // Only emit a shuffle if the mask is legal.
878 if (!TLI.isShuffleMaskLegal(ShuffleMask, ByteVT))
879 return DAG.UnrollVectorOp(Op.getNode());
882 Op = DAG.getNode(ISD::BITCAST, DL, ByteVT, Op.getOperand(0));
883 Op = DAG.getVectorShuffle(ByteVT, DL, Op, DAG.getUNDEF(ByteVT), ShuffleMask);
884 return DAG.getNode(ISD::BITCAST, DL, VT, Op);
887 SDValue VectorLegalizer::ExpandBITREVERSE(SDValue Op) {
888 EVT VT = Op.getValueType();
890 // If we have the scalar operation, it's probably cheaper to unroll it.
891 if (TLI.isOperationLegalOrCustom(ISD::BITREVERSE, VT.getScalarType()))
892 return DAG.UnrollVectorOp(Op.getNode());
894 // If the vector element width is a whole number of bytes, test if its legal
895 // to BSWAP shuffle the bytes and then perform the BITREVERSE on the byte
896 // vector. This greatly reduces the number of bit shifts necessary.
897 unsigned ScalarSizeInBits = VT.getScalarSizeInBits();
898 if (ScalarSizeInBits > 8 && (ScalarSizeInBits % 8) == 0) {
899 SmallVector<int, 16> BSWAPMask;
900 createBSWAPShuffleMask(VT, BSWAPMask);
902 EVT ByteVT = EVT::getVectorVT(*DAG.getContext(), MVT::i8, BSWAPMask.size());
903 if (TLI.isShuffleMaskLegal(BSWAPMask, ByteVT) &&
904 (TLI.isOperationLegalOrCustom(ISD::BITREVERSE, ByteVT) ||
905 (TLI.isOperationLegalOrCustom(ISD::SHL, ByteVT) &&
906 TLI.isOperationLegalOrCustom(ISD::SRL, ByteVT) &&
907 TLI.isOperationLegalOrCustomOrPromote(ISD::AND, ByteVT) &&
908 TLI.isOperationLegalOrCustomOrPromote(ISD::OR, ByteVT)))) {
910 Op = DAG.getNode(ISD::BITCAST, DL, ByteVT, Op.getOperand(0));
911 Op = DAG.getVectorShuffle(ByteVT, DL, Op, DAG.getUNDEF(ByteVT),
913 Op = DAG.getNode(ISD::BITREVERSE, DL, ByteVT, Op);
914 return DAG.getNode(ISD::BITCAST, DL, VT, Op);
918 // If we have the appropriate vector bit operations, it is better to use them
919 // than unrolling and expanding each component.
920 if (!TLI.isOperationLegalOrCustom(ISD::SHL, VT) ||
921 !TLI.isOperationLegalOrCustom(ISD::SRL, VT) ||
922 !TLI.isOperationLegalOrCustomOrPromote(ISD::AND, VT) ||
923 !TLI.isOperationLegalOrCustomOrPromote(ISD::OR, VT))
924 return DAG.UnrollVectorOp(Op.getNode());
926 // Let LegalizeDAG handle this later.
930 SDValue VectorLegalizer::ExpandVSELECT(SDValue Op) {
931 // Implement VSELECT in terms of XOR, AND, OR
932 // on platforms which do not support blend natively.
935 SDValue Mask = Op.getOperand(0);
936 SDValue Op1 = Op.getOperand(1);
937 SDValue Op2 = Op.getOperand(2);
939 EVT VT = Mask.getValueType();
941 // If we can't even use the basic vector operations of
942 // AND,OR,XOR, we will have to scalarize the op.
943 // Notice that the operation may be 'promoted' which means that it is
944 // 'bitcasted' to another type which is handled.
945 // This operation also isn't safe with AND, OR, XOR when the boolean
946 // type is 0/1 as we need an all ones vector constant to mask with.
947 // FIXME: Sign extend 1 to all ones if thats legal on the target.
948 if (TLI.getOperationAction(ISD::AND, VT) == TargetLowering::Expand ||
949 TLI.getOperationAction(ISD::XOR, VT) == TargetLowering::Expand ||
950 TLI.getOperationAction(ISD::OR, VT) == TargetLowering::Expand ||
951 TLI.getBooleanContents(Op1.getValueType()) !=
952 TargetLowering::ZeroOrNegativeOneBooleanContent)
953 return DAG.UnrollVectorOp(Op.getNode());
955 // If the mask and the type are different sizes, unroll the vector op. This
956 // can occur when getSetCCResultType returns something that is different in
957 // size from the operand types. For example, v4i8 = select v4i32, v4i8, v4i8.
958 if (VT.getSizeInBits() != Op1.getValueSizeInBits())
959 return DAG.UnrollVectorOp(Op.getNode());
961 // Bitcast the operands to be the same type as the mask.
962 // This is needed when we select between FP types because
963 // the mask is a vector of integers.
964 Op1 = DAG.getNode(ISD::BITCAST, DL, VT, Op1);
965 Op2 = DAG.getNode(ISD::BITCAST, DL, VT, Op2);
967 SDValue AllOnes = DAG.getConstant(
968 APInt::getAllOnesValue(VT.getScalarSizeInBits()), DL, VT);
969 SDValue NotMask = DAG.getNode(ISD::XOR, DL, VT, Mask, AllOnes);
971 Op1 = DAG.getNode(ISD::AND, DL, VT, Op1, Mask);
972 Op2 = DAG.getNode(ISD::AND, DL, VT, Op2, NotMask);
973 SDValue Val = DAG.getNode(ISD::OR, DL, VT, Op1, Op2);
974 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Val);
977 SDValue VectorLegalizer::ExpandUINT_TO_FLOAT(SDValue Op) {
978 EVT VT = Op.getOperand(0).getValueType();
981 // Make sure that the SINT_TO_FP and SRL instructions are available.
982 if (TLI.getOperationAction(ISD::SINT_TO_FP, VT) == TargetLowering::Expand ||
983 TLI.getOperationAction(ISD::SRL, VT) == TargetLowering::Expand)
984 return DAG.UnrollVectorOp(Op.getNode());
986 unsigned BW = VT.getScalarSizeInBits();
987 assert((BW == 64 || BW == 32) &&
988 "Elements in vector-UINT_TO_FP must be 32 or 64 bits wide");
990 SDValue HalfWord = DAG.getConstant(BW / 2, DL, VT);
992 // Constants to clear the upper part of the word.
993 // Notice that we can also use SHL+SHR, but using a constant is slightly
995 uint64_t HWMask = (BW == 64) ? 0x00000000FFFFFFFF : 0x0000FFFF;
996 SDValue HalfWordMask = DAG.getConstant(HWMask, DL, VT);
998 // Two to the power of half-word-size.
999 SDValue TWOHW = DAG.getConstantFP(1 << (BW / 2), DL, Op.getValueType());
1001 // Clear upper part of LO, lower HI
1002 SDValue HI = DAG.getNode(ISD::SRL, DL, VT, Op.getOperand(0), HalfWord);
1003 SDValue LO = DAG.getNode(ISD::AND, DL, VT, Op.getOperand(0), HalfWordMask);
1005 // Convert hi and lo to floats
1006 // Convert the hi part back to the upper values
1007 // TODO: Can any fast-math-flags be set on these nodes?
1008 SDValue fHI = DAG.getNode(ISD::SINT_TO_FP, DL, Op.getValueType(), HI);
1009 fHI = DAG.getNode(ISD::FMUL, DL, Op.getValueType(), fHI, TWOHW);
1010 SDValue fLO = DAG.getNode(ISD::SINT_TO_FP, DL, Op.getValueType(), LO);
1012 // Add the two halves
1013 return DAG.getNode(ISD::FADD, DL, Op.getValueType(), fHI, fLO);
1016 SDValue VectorLegalizer::ExpandFNEG(SDValue Op) {
1017 if (TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType())) {
1019 SDValue Zero = DAG.getConstantFP(-0.0, DL, Op.getValueType());
1020 // TODO: If FNEG had fast-math-flags, they'd get propagated to this FSUB.
1021 return DAG.getNode(ISD::FSUB, DL, Op.getValueType(),
1022 Zero, Op.getOperand(0));
1024 return DAG.UnrollVectorOp(Op.getNode());
1027 SDValue VectorLegalizer::ExpandFSUB(SDValue Op) {
1028 // For floating-point values, (a-b) is the same as a+(-b). If FNEG is legal,
1029 // we can defer this to operation legalization where it will be lowered as
1031 EVT VT = Op.getValueType();
1032 if (TLI.isOperationLegalOrCustom(ISD::FNEG, VT) &&
1033 TLI.isOperationLegalOrCustom(ISD::FADD, VT))
1034 return Op; // Defer to LegalizeDAG
1036 return DAG.UnrollVectorOp(Op.getNode());
1039 SDValue VectorLegalizer::ExpandCTLZ(SDValue Op) {
1040 EVT VT = Op.getValueType();
1041 unsigned NumBitsPerElt = VT.getScalarSizeInBits();
1043 // If the non-ZERO_UNDEF version is supported we can use that instead.
1044 if (Op.getOpcode() == ISD::CTLZ_ZERO_UNDEF &&
1045 TLI.isOperationLegalOrCustom(ISD::CTLZ, VT)) {
1047 return DAG.getNode(ISD::CTLZ, DL, Op.getValueType(), Op.getOperand(0));
1050 // If CTPOP is available we can lower with a CTPOP based method:
1051 // u16 ctlz(u16 x) {
1056 // return ctpop(~x);
1058 // Ref: "Hacker's Delight" by Henry Warren
1059 if (isPowerOf2_32(NumBitsPerElt) &&
1060 TLI.isOperationLegalOrCustom(ISD::CTPOP, VT) &&
1061 TLI.isOperationLegalOrCustom(ISD::SRL, VT) &&
1062 TLI.isOperationLegalOrCustomOrPromote(ISD::OR, VT) &&
1063 TLI.isOperationLegalOrCustomOrPromote(ISD::XOR, VT)) {
1065 SDValue Res = Op.getOperand(0);
1066 EVT ShiftTy = TLI.getShiftAmountTy(VT, DAG.getDataLayout());
1068 for (unsigned i = 1; i != NumBitsPerElt; i *= 2)
1070 ISD::OR, DL, VT, Res,
1071 DAG.getNode(ISD::SRL, DL, VT, Res, DAG.getConstant(i, DL, ShiftTy)));
1073 Res = DAG.getNOT(DL, Res, VT);
1074 return DAG.getNode(ISD::CTPOP, DL, VT, Res);
1077 // Otherwise go ahead and unroll.
1078 return DAG.UnrollVectorOp(Op.getNode());
1081 SDValue VectorLegalizer::ExpandCTTZ_ZERO_UNDEF(SDValue Op) {
1082 // If the non-ZERO_UNDEF version is supported we can use that instead.
1083 if (TLI.isOperationLegalOrCustom(ISD::CTTZ, Op.getValueType())) {
1085 return DAG.getNode(ISD::CTTZ, DL, Op.getValueType(), Op.getOperand(0));
1088 // Otherwise go ahead and unroll.
1089 return DAG.UnrollVectorOp(Op.getNode());
1092 SDValue VectorLegalizer::UnrollVSETCC(SDValue Op) {
1093 EVT VT = Op.getValueType();
1094 unsigned NumElems = VT.getVectorNumElements();
1095 EVT EltVT = VT.getVectorElementType();
1096 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1), CC = Op.getOperand(2);
1097 EVT TmpEltVT = LHS.getValueType().getVectorElementType();
1099 SmallVector<SDValue, 8> Ops(NumElems);
1100 for (unsigned i = 0; i < NumElems; ++i) {
1101 SDValue LHSElem = DAG.getNode(
1102 ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, LHS,
1103 DAG.getConstant(i, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
1104 SDValue RHSElem = DAG.getNode(
1105 ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, RHS,
1106 DAG.getConstant(i, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
1107 Ops[i] = DAG.getNode(ISD::SETCC, dl,
1108 TLI.getSetCCResultType(DAG.getDataLayout(),
1109 *DAG.getContext(), TmpEltVT),
1110 LHSElem, RHSElem, CC);
1111 Ops[i] = DAG.getSelect(dl, EltVT, Ops[i],
1112 DAG.getConstant(APInt::getAllOnesValue
1113 (EltVT.getSizeInBits()), dl, EltVT),
1114 DAG.getConstant(0, dl, EltVT));
1116 return DAG.getBuildVector(VT, dl, Ops);
1121 bool SelectionDAG::LegalizeVectors() {
1122 return VectorLegalizer(*this).Run();