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/ADT/APInt.h"
31 #include "llvm/ADT/DenseMap.h"
32 #include "llvm/ADT/SmallVector.h"
33 #include "llvm/CodeGen/ISDOpcodes.h"
34 #include "llvm/CodeGen/MachineMemOperand.h"
35 #include "llvm/CodeGen/SelectionDAG.h"
36 #include "llvm/CodeGen/SelectionDAGNodes.h"
37 #include "llvm/CodeGen/TargetLowering.h"
38 #include "llvm/CodeGen/ValueTypes.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/Support/Casting.h"
41 #include "llvm/Support/Compiler.h"
42 #include "llvm/Support/ErrorHandling.h"
43 #include "llvm/Support/MachineValueType.h"
44 #include "llvm/Support/MathExtras.h"
52 #define DEBUG_TYPE "legalizevectorops"
56 class VectorLegalizer {
58 const TargetLowering &TLI;
59 bool Changed = false; // Keep track of whether anything changed
61 /// For nodes that are of legal width, and that have more than one use, this
62 /// map indicates what regularized operand to use. This allows us to avoid
63 /// legalizing the same thing more than once.
64 SmallDenseMap<SDValue, SDValue, 64> LegalizedNodes;
66 /// Adds a node to the translation cache.
67 void AddLegalizedOperand(SDValue From, SDValue To) {
68 LegalizedNodes.insert(std::make_pair(From, To));
69 // If someone requests legalization of the new node, return itself.
71 LegalizedNodes.insert(std::make_pair(To, To));
74 /// Legalizes the given node.
75 SDValue LegalizeOp(SDValue Op);
77 /// Assuming the node is legal, "legalize" the results.
78 SDValue TranslateLegalizeResults(SDValue Op, SDValue Result);
80 /// Implements unrolling a VSETCC.
81 SDValue UnrollVSETCC(SDValue Op);
83 /// Implement expand-based legalization of vector operations.
85 /// This is just a high-level routine to dispatch to specific code paths for
86 /// operations to legalize them.
87 SDValue Expand(SDValue Op);
89 /// Implements expansion for FNEG; falls back to UnrollVectorOp if
92 /// Implements expansion for UINT_TO_FLOAT; falls back to UnrollVectorOp if
93 /// SINT_TO_FLOAT and SHR on vectors isn't legal.
94 SDValue ExpandUINT_TO_FLOAT(SDValue Op);
96 /// Implement expansion for SIGN_EXTEND_INREG using SRL and SRA.
97 SDValue ExpandSEXTINREG(SDValue Op);
99 /// Implement expansion for ANY_EXTEND_VECTOR_INREG.
101 /// Shuffles the low lanes of the operand into place and bitcasts to the proper
102 /// type. The contents of the bits in the extended part of each element are
104 SDValue ExpandANY_EXTEND_VECTOR_INREG(SDValue Op);
106 /// Implement expansion for SIGN_EXTEND_VECTOR_INREG.
108 /// Shuffles the low lanes of the operand into place, bitcasts to the proper
109 /// type, then shifts left and arithmetic shifts right to introduce a sign
111 SDValue ExpandSIGN_EXTEND_VECTOR_INREG(SDValue Op);
113 /// Implement expansion for ZERO_EXTEND_VECTOR_INREG.
115 /// Shuffles the low lanes of the operand into place and blends zeros into
116 /// the remaining lanes, finally bitcasting to the proper type.
117 SDValue ExpandZERO_EXTEND_VECTOR_INREG(SDValue Op);
119 /// Expand bswap of vectors into a shuffle if legal.
120 SDValue ExpandBSWAP(SDValue Op);
122 /// Implement vselect in terms of XOR, AND, OR when blend is not
123 /// supported by the target.
124 SDValue ExpandVSELECT(SDValue Op);
125 SDValue ExpandSELECT(SDValue Op);
126 SDValue ExpandLoad(SDValue Op);
127 SDValue ExpandStore(SDValue Op);
128 SDValue ExpandFNEG(SDValue Op);
129 SDValue ExpandFSUB(SDValue Op);
130 SDValue ExpandBITREVERSE(SDValue Op);
131 SDValue ExpandCTLZ(SDValue Op);
132 SDValue ExpandCTTZ(SDValue Op);
133 SDValue ExpandStrictFPOp(SDValue Op);
135 /// Implements vector promotion.
137 /// This is essentially just bitcasting the operands to a different type and
138 /// bitcasting the result back to the original type.
139 SDValue Promote(SDValue Op);
141 /// Implements [SU]INT_TO_FP vector promotion.
143 /// This is a [zs]ext of the input operand to a larger integer type.
144 SDValue PromoteINT_TO_FP(SDValue Op);
146 /// Implements FP_TO_[SU]INT vector promotion of the result type.
148 /// It is promoted to a larger integer type. The result is then
149 /// truncated back to the original type.
150 SDValue PromoteFP_TO_INT(SDValue Op);
153 VectorLegalizer(SelectionDAG& dag) :
154 DAG(dag), TLI(dag.getTargetLoweringInfo()) {}
156 /// Begin legalizer the vector operations in the DAG.
160 } // end anonymous namespace
162 bool VectorLegalizer::Run() {
163 // Before we start legalizing vector nodes, check if there are any vectors.
164 bool HasVectors = false;
165 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
166 E = std::prev(DAG.allnodes_end()); I != std::next(E); ++I) {
167 // Check if the values of the nodes contain vectors. We don't need to check
168 // the operands because we are going to check their values at some point.
169 for (SDNode::value_iterator J = I->value_begin(), E = I->value_end();
171 HasVectors |= J->isVector();
173 // If we found a vector node we can start the legalization.
178 // If this basic block has no vectors then no need to legalize vectors.
182 // The legalize process is inherently a bottom-up recursive process (users
183 // legalize their uses before themselves). Given infinite stack space, we
184 // could just start legalizing on the root and traverse the whole graph. In
185 // practice however, this causes us to run out of stack space on large basic
186 // blocks. To avoid this problem, compute an ordering of the nodes where each
187 // node is only legalized after all of its operands are legalized.
188 DAG.AssignTopologicalOrder();
189 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
190 E = std::prev(DAG.allnodes_end()); I != std::next(E); ++I)
191 LegalizeOp(SDValue(&*I, 0));
193 // Finally, it's possible the root changed. Get the new root.
194 SDValue OldRoot = DAG.getRoot();
195 assert(LegalizedNodes.count(OldRoot) && "Root didn't get legalized?");
196 DAG.setRoot(LegalizedNodes[OldRoot]);
198 LegalizedNodes.clear();
200 // Remove dead nodes now.
201 DAG.RemoveDeadNodes();
206 SDValue VectorLegalizer::TranslateLegalizeResults(SDValue Op, SDValue Result) {
207 // Generic legalization: just pass the operand through.
208 for (unsigned i = 0, e = Op.getNode()->getNumValues(); i != e; ++i)
209 AddLegalizedOperand(Op.getValue(i), Result.getValue(i));
210 return Result.getValue(Op.getResNo());
213 SDValue VectorLegalizer::LegalizeOp(SDValue Op) {
214 // Note that LegalizeOp may be reentered even from single-use nodes, which
215 // means that we always must cache transformed nodes.
216 DenseMap<SDValue, SDValue>::iterator I = LegalizedNodes.find(Op);
217 if (I != LegalizedNodes.end()) return I->second;
219 SDNode* Node = Op.getNode();
221 // Legalize the operands
222 SmallVector<SDValue, 8> Ops;
223 for (const SDValue &Op : Node->op_values())
224 Ops.push_back(LegalizeOp(Op));
226 SDValue Result = SDValue(DAG.UpdateNodeOperands(Op.getNode(), Ops),
229 if (Op.getOpcode() == ISD::LOAD) {
230 LoadSDNode *LD = cast<LoadSDNode>(Op.getNode());
231 ISD::LoadExtType ExtType = LD->getExtensionType();
232 if (LD->getMemoryVT().isVector() && ExtType != ISD::NON_EXTLOAD) {
233 LLVM_DEBUG(dbgs() << "\nLegalizing extending vector load: ";
235 switch (TLI.getLoadExtAction(LD->getExtensionType(), LD->getValueType(0),
236 LD->getMemoryVT())) {
237 default: llvm_unreachable("This action is not supported yet!");
238 case TargetLowering::Legal:
239 return TranslateLegalizeResults(Op, Result);
240 case TargetLowering::Custom:
241 if (SDValue Lowered = TLI.LowerOperation(Result, DAG)) {
242 assert(Lowered->getNumValues() == Op->getNumValues() &&
243 "Unexpected number of results");
244 Changed = Lowered != Result;
245 return TranslateLegalizeResults(Op, Lowered);
248 case TargetLowering::Expand:
250 return LegalizeOp(ExpandLoad(Op));
253 } else if (Op.getOpcode() == ISD::STORE) {
254 StoreSDNode *ST = cast<StoreSDNode>(Op.getNode());
255 EVT StVT = ST->getMemoryVT();
256 MVT ValVT = ST->getValue().getSimpleValueType();
257 if (StVT.isVector() && ST->isTruncatingStore()) {
258 LLVM_DEBUG(dbgs() << "\nLegalizing truncating vector store: ";
260 switch (TLI.getTruncStoreAction(ValVT, StVT)) {
261 default: llvm_unreachable("This action is not supported yet!");
262 case TargetLowering::Legal:
263 return TranslateLegalizeResults(Op, Result);
264 case TargetLowering::Custom: {
265 SDValue Lowered = TLI.LowerOperation(Result, DAG);
266 Changed = Lowered != Result;
267 return TranslateLegalizeResults(Op, Lowered);
269 case TargetLowering::Expand:
271 return LegalizeOp(ExpandStore(Op));
276 bool HasVectorValue = false;
277 for (SDNode::value_iterator J = Node->value_begin(), E = Node->value_end();
280 HasVectorValue |= J->isVector();
282 return TranslateLegalizeResults(Op, Result);
284 TargetLowering::LegalizeAction Action = TargetLowering::Legal;
285 switch (Op.getOpcode()) {
287 return TranslateLegalizeResults(Op, Result);
288 case ISD::STRICT_FADD:
289 case ISD::STRICT_FSUB:
290 case ISD::STRICT_FMUL:
291 case ISD::STRICT_FDIV:
292 case ISD::STRICT_FREM:
293 case ISD::STRICT_FSQRT:
294 case ISD::STRICT_FMA:
295 case ISD::STRICT_FPOW:
296 case ISD::STRICT_FPOWI:
297 case ISD::STRICT_FSIN:
298 case ISD::STRICT_FCOS:
299 case ISD::STRICT_FEXP:
300 case ISD::STRICT_FEXP2:
301 case ISD::STRICT_FLOG:
302 case ISD::STRICT_FLOG10:
303 case ISD::STRICT_FLOG2:
304 case ISD::STRICT_FRINT:
305 case ISD::STRICT_FNEARBYINT:
306 // These pseudo-ops get legalized as if they were their non-strict
307 // equivalent. For instance, if ISD::FSQRT is legal then ISD::STRICT_FSQRT
308 // is also legal, but if ISD::FSQRT requires expansion then so does
309 // ISD::STRICT_FSQRT.
310 Action = TLI.getStrictFPOperationAction(Node->getOpcode(),
311 Node->getValueType(0));
336 case ISD::BITREVERSE:
339 case ISD::CTLZ_ZERO_UNDEF:
340 case ISD::CTTZ_ZERO_UNDEF:
346 case ISD::ZERO_EXTEND:
347 case ISD::ANY_EXTEND:
349 case ISD::SIGN_EXTEND:
350 case ISD::FP_TO_SINT:
351 case ISD::FP_TO_UINT:
372 case ISD::FNEARBYINT:
378 case ISD::SIGN_EXTEND_INREG:
379 case ISD::ANY_EXTEND_VECTOR_INREG:
380 case ISD::SIGN_EXTEND_VECTOR_INREG:
381 case ISD::ZERO_EXTEND_VECTOR_INREG:
388 case ISD::FCANONICALIZE:
390 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
392 case ISD::FP_ROUND_INREG:
393 Action = TLI.getOperationAction(Node->getOpcode(),
394 cast<VTSDNode>(Node->getOperand(1))->getVT());
396 case ISD::SINT_TO_FP:
397 case ISD::UINT_TO_FP:
398 Action = TLI.getOperationAction(Node->getOpcode(),
399 Node->getOperand(0).getValueType());
403 LLVM_DEBUG(dbgs() << "\nLegalizing vector op: "; Node->dump(&DAG));
406 default: llvm_unreachable("This action is not supported yet!");
407 case TargetLowering::Promote:
408 Result = Promote(Op);
411 case TargetLowering::Legal:
412 LLVM_DEBUG(dbgs() << "Legal node: nothing to do\n");
414 case TargetLowering::Custom: {
415 LLVM_DEBUG(dbgs() << "Trying custom legalization\n");
416 if (SDValue Tmp1 = TLI.LowerOperation(Op, DAG)) {
417 LLVM_DEBUG(dbgs() << "Successfully custom legalized node\n");
421 LLVM_DEBUG(dbgs() << "Could not custom legalize node\n");
424 case TargetLowering::Expand:
428 // Make sure that the generated code is itself legal.
430 Result = LegalizeOp(Result);
434 // Note that LegalizeOp may be reentered even from single-use nodes, which
435 // means that we always must cache transformed nodes.
436 AddLegalizedOperand(Op, Result);
440 SDValue VectorLegalizer::Promote(SDValue Op) {
441 // For a few operations there is a specific concept for promotion based on
442 // the operand's type.
443 switch (Op.getOpcode()) {
444 case ISD::SINT_TO_FP:
445 case ISD::UINT_TO_FP:
446 // "Promote" the operation by extending the operand.
447 return PromoteINT_TO_FP(Op);
448 case ISD::FP_TO_UINT:
449 case ISD::FP_TO_SINT:
450 // Promote the operation by extending the operand.
451 return PromoteFP_TO_INT(Op);
454 // There are currently two cases of vector promotion:
455 // 1) Bitcasting a vector of integers to a different type to a vector of the
456 // same overall length. For example, x86 promotes ISD::AND v2i32 to v1i64.
457 // 2) Extending a vector of floats to a vector of the same number of larger
458 // floats. For example, AArch64 promotes ISD::FADD on v4f16 to v4f32.
459 MVT VT = Op.getSimpleValueType();
460 assert(Op.getNode()->getNumValues() == 1 &&
461 "Can't promote a vector with multiple results!");
462 MVT NVT = TLI.getTypeToPromoteTo(Op.getOpcode(), VT);
464 SmallVector<SDValue, 4> Operands(Op.getNumOperands());
466 for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
467 if (Op.getOperand(j).getValueType().isVector())
470 .getVectorElementType()
471 .isFloatingPoint() &&
472 NVT.isVector() && NVT.getVectorElementType().isFloatingPoint())
473 Operands[j] = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Op.getOperand(j));
475 Operands[j] = DAG.getNode(ISD::BITCAST, dl, NVT, Op.getOperand(j));
477 Operands[j] = Op.getOperand(j);
480 Op = DAG.getNode(Op.getOpcode(), dl, NVT, Operands, Op.getNode()->getFlags());
481 if ((VT.isFloatingPoint() && NVT.isFloatingPoint()) ||
482 (VT.isVector() && VT.getVectorElementType().isFloatingPoint() &&
483 NVT.isVector() && NVT.getVectorElementType().isFloatingPoint()))
484 return DAG.getNode(ISD::FP_ROUND, dl, VT, Op, DAG.getIntPtrConstant(0, dl));
486 return DAG.getNode(ISD::BITCAST, dl, VT, Op);
489 SDValue VectorLegalizer::PromoteINT_TO_FP(SDValue Op) {
490 // INT_TO_FP operations may require the input operand be promoted even
491 // when the type is otherwise legal.
492 MVT VT = Op.getOperand(0).getSimpleValueType();
493 MVT NVT = TLI.getTypeToPromoteTo(Op.getOpcode(), VT);
494 assert(NVT.getVectorNumElements() == VT.getVectorNumElements() &&
495 "Vectors have different number of elements!");
498 SmallVector<SDValue, 4> Operands(Op.getNumOperands());
500 unsigned Opc = Op.getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND :
502 for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
503 if (Op.getOperand(j).getValueType().isVector())
504 Operands[j] = DAG.getNode(Opc, dl, NVT, Op.getOperand(j));
506 Operands[j] = Op.getOperand(j);
509 return DAG.getNode(Op.getOpcode(), dl, Op.getValueType(), Operands);
512 // For FP_TO_INT we promote the result type to a vector type with wider
513 // elements and then truncate the result. This is different from the default
514 // PromoteVector which uses bitcast to promote thus assumning that the
515 // promoted vector type has the same overall size.
516 SDValue VectorLegalizer::PromoteFP_TO_INT(SDValue Op) {
517 MVT VT = Op.getSimpleValueType();
518 MVT NVT = TLI.getTypeToPromoteTo(Op.getOpcode(), VT);
519 assert(NVT.getVectorNumElements() == VT.getVectorNumElements() &&
520 "Vectors have different number of elements!");
522 unsigned NewOpc = Op->getOpcode();
523 // Change FP_TO_UINT to FP_TO_SINT if possible.
524 // TODO: Should we only do this if FP_TO_UINT itself isn't legal?
525 if (NewOpc == ISD::FP_TO_UINT &&
526 TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NVT))
527 NewOpc = ISD::FP_TO_SINT;
530 SDValue Promoted = DAG.getNode(NewOpc, dl, NVT, Op.getOperand(0));
532 // Assert that the converted value fits in the original type. If it doesn't
533 // (eg: because the value being converted is too big), then the result of the
534 // original operation was undefined anyway, so the assert is still correct.
535 Promoted = DAG.getNode(Op->getOpcode() == ISD::FP_TO_UINT ? ISD::AssertZext
538 DAG.getValueType(VT.getScalarType()));
539 return DAG.getNode(ISD::TRUNCATE, dl, VT, Promoted);
542 SDValue VectorLegalizer::ExpandLoad(SDValue Op) {
543 LoadSDNode *LD = cast<LoadSDNode>(Op.getNode());
545 EVT SrcVT = LD->getMemoryVT();
546 EVT SrcEltVT = SrcVT.getScalarType();
547 unsigned NumElem = SrcVT.getVectorNumElements();
551 if (SrcVT.getVectorNumElements() > 1 && !SrcEltVT.isByteSized()) {
554 SmallVector<SDValue, 8> Vals;
555 SmallVector<SDValue, 8> LoadChains;
557 EVT DstEltVT = LD->getValueType(0).getScalarType();
558 SDValue Chain = LD->getChain();
559 SDValue BasePTR = LD->getBasePtr();
560 ISD::LoadExtType ExtType = LD->getExtensionType();
562 // When elements in a vector is not byte-addressable, we cannot directly
563 // load each element by advancing pointer, which could only address bytes.
564 // Instead, we load all significant words, mask bits off, and concatenate
565 // them to form each element. Finally, they are extended to destination
566 // scalar type to build the destination vector.
567 EVT WideVT = TLI.getPointerTy(DAG.getDataLayout());
569 assert(WideVT.isRound() &&
570 "Could not handle the sophisticated case when the widest integer is"
572 assert(WideVT.bitsGE(SrcEltVT) &&
573 "Type is not legalized?");
575 unsigned WideBytes = WideVT.getStoreSize();
577 unsigned RemainingBytes = SrcVT.getStoreSize();
578 SmallVector<SDValue, 8> LoadVals;
579 while (RemainingBytes > 0) {
581 unsigned LoadBytes = WideBytes;
583 if (RemainingBytes >= LoadBytes) {
585 DAG.getLoad(WideVT, dl, Chain, BasePTR,
586 LD->getPointerInfo().getWithOffset(Offset),
587 MinAlign(LD->getAlignment(), Offset),
588 LD->getMemOperand()->getFlags(), LD->getAAInfo());
591 while (RemainingBytes < LoadBytes) {
592 LoadBytes >>= 1; // Reduce the load size by half.
593 LoadVT = EVT::getIntegerVT(*DAG.getContext(), LoadBytes << 3);
596 DAG.getExtLoad(ISD::EXTLOAD, dl, WideVT, Chain, BasePTR,
597 LD->getPointerInfo().getWithOffset(Offset), LoadVT,
598 MinAlign(LD->getAlignment(), Offset),
599 LD->getMemOperand()->getFlags(), LD->getAAInfo());
602 RemainingBytes -= LoadBytes;
605 BasePTR = DAG.getObjectPtrOffset(dl, BasePTR, LoadBytes);
607 LoadVals.push_back(ScalarLoad.getValue(0));
608 LoadChains.push_back(ScalarLoad.getValue(1));
611 // Extract bits, pack and extend/trunc them into destination type.
612 unsigned SrcEltBits = SrcEltVT.getSizeInBits();
613 SDValue SrcEltBitMask = DAG.getConstant((1U << SrcEltBits) - 1, dl, WideVT);
615 unsigned BitOffset = 0;
616 unsigned WideIdx = 0;
617 unsigned WideBits = WideVT.getSizeInBits();
619 for (unsigned Idx = 0; Idx != NumElem; ++Idx) {
620 SDValue Lo, Hi, ShAmt;
622 if (BitOffset < WideBits) {
623 ShAmt = DAG.getConstant(
624 BitOffset, dl, TLI.getShiftAmountTy(WideVT, DAG.getDataLayout()));
625 Lo = DAG.getNode(ISD::SRL, dl, WideVT, LoadVals[WideIdx], ShAmt);
626 Lo = DAG.getNode(ISD::AND, dl, WideVT, Lo, SrcEltBitMask);
629 BitOffset += SrcEltBits;
630 if (BitOffset >= WideBits) {
632 BitOffset -= WideBits;
634 ShAmt = DAG.getConstant(
635 SrcEltBits - BitOffset, dl,
636 TLI.getShiftAmountTy(WideVT, DAG.getDataLayout()));
637 Hi = DAG.getNode(ISD::SHL, dl, WideVT, LoadVals[WideIdx], ShAmt);
638 Hi = DAG.getNode(ISD::AND, dl, WideVT, Hi, SrcEltBitMask);
643 Lo = DAG.getNode(ISD::OR, dl, WideVT, Lo, Hi);
646 default: llvm_unreachable("Unknown extended-load op!");
648 Lo = DAG.getAnyExtOrTrunc(Lo, dl, DstEltVT);
651 Lo = DAG.getZExtOrTrunc(Lo, dl, DstEltVT);
655 DAG.getConstant(WideBits - SrcEltBits, dl,
656 TLI.getShiftAmountTy(WideVT, DAG.getDataLayout()));
657 Lo = DAG.getNode(ISD::SHL, dl, WideVT, Lo, ShAmt);
658 Lo = DAG.getNode(ISD::SRA, dl, WideVT, Lo, ShAmt);
659 Lo = DAG.getSExtOrTrunc(Lo, dl, DstEltVT);
665 NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, LoadChains);
666 Value = DAG.getBuildVector(Op.getNode()->getValueType(0), dl, Vals);
668 SDValue Scalarized = TLI.scalarizeVectorLoad(LD, DAG);
669 // Skip past MERGE_VALUE node if known.
670 if (Scalarized->getOpcode() == ISD::MERGE_VALUES) {
671 NewChain = Scalarized.getOperand(1);
672 Value = Scalarized.getOperand(0);
674 NewChain = Scalarized.getValue(1);
675 Value = Scalarized.getValue(0);
679 AddLegalizedOperand(Op.getValue(0), Value);
680 AddLegalizedOperand(Op.getValue(1), NewChain);
682 return (Op.getResNo() ? NewChain : Value);
685 SDValue VectorLegalizer::ExpandStore(SDValue Op) {
686 StoreSDNode *ST = cast<StoreSDNode>(Op.getNode());
687 SDValue TF = TLI.scalarizeVectorStore(ST, DAG);
688 AddLegalizedOperand(Op, TF);
692 SDValue VectorLegalizer::Expand(SDValue Op) {
693 switch (Op->getOpcode()) {
694 case ISD::SIGN_EXTEND_INREG:
695 return ExpandSEXTINREG(Op);
696 case ISD::ANY_EXTEND_VECTOR_INREG:
697 return ExpandANY_EXTEND_VECTOR_INREG(Op);
698 case ISD::SIGN_EXTEND_VECTOR_INREG:
699 return ExpandSIGN_EXTEND_VECTOR_INREG(Op);
700 case ISD::ZERO_EXTEND_VECTOR_INREG:
701 return ExpandZERO_EXTEND_VECTOR_INREG(Op);
703 return ExpandBSWAP(Op);
705 return ExpandVSELECT(Op);
707 return ExpandSELECT(Op);
708 case ISD::UINT_TO_FP:
709 return ExpandUINT_TO_FLOAT(Op);
711 return ExpandFNEG(Op);
713 return ExpandFSUB(Op);
715 return UnrollVSETCC(Op);
716 case ISD::BITREVERSE:
717 return ExpandBITREVERSE(Op);
719 case ISD::CTLZ_ZERO_UNDEF:
720 return ExpandCTLZ(Op);
722 case ISD::CTTZ_ZERO_UNDEF:
723 return ExpandCTTZ(Op);
724 case ISD::STRICT_FADD:
725 case ISD::STRICT_FSUB:
726 case ISD::STRICT_FMUL:
727 case ISD::STRICT_FDIV:
728 case ISD::STRICT_FREM:
729 case ISD::STRICT_FSQRT:
730 case ISD::STRICT_FMA:
731 case ISD::STRICT_FPOW:
732 case ISD::STRICT_FPOWI:
733 case ISD::STRICT_FSIN:
734 case ISD::STRICT_FCOS:
735 case ISD::STRICT_FEXP:
736 case ISD::STRICT_FEXP2:
737 case ISD::STRICT_FLOG:
738 case ISD::STRICT_FLOG10:
739 case ISD::STRICT_FLOG2:
740 case ISD::STRICT_FRINT:
741 case ISD::STRICT_FNEARBYINT:
742 return ExpandStrictFPOp(Op);
744 return DAG.UnrollVectorOp(Op.getNode());
748 SDValue VectorLegalizer::ExpandSELECT(SDValue Op) {
749 // Lower a select instruction where the condition is a scalar and the
750 // operands are vectors. Lower this select to VSELECT and implement it
751 // using XOR AND OR. The selector bit is broadcasted.
752 EVT VT = Op.getValueType();
755 SDValue Mask = Op.getOperand(0);
756 SDValue Op1 = Op.getOperand(1);
757 SDValue Op2 = Op.getOperand(2);
759 assert(VT.isVector() && !Mask.getValueType().isVector()
760 && Op1.getValueType() == Op2.getValueType() && "Invalid type");
762 // If we can't even use the basic vector operations of
763 // AND,OR,XOR, we will have to scalarize the op.
764 // Notice that the operation may be 'promoted' which means that it is
765 // 'bitcasted' to another type which is handled.
766 // Also, we need to be able to construct a splat vector using BUILD_VECTOR.
767 if (TLI.getOperationAction(ISD::AND, VT) == TargetLowering::Expand ||
768 TLI.getOperationAction(ISD::XOR, VT) == TargetLowering::Expand ||
769 TLI.getOperationAction(ISD::OR, VT) == TargetLowering::Expand ||
770 TLI.getOperationAction(ISD::BUILD_VECTOR, VT) == TargetLowering::Expand)
771 return DAG.UnrollVectorOp(Op.getNode());
773 // Generate a mask operand.
774 EVT MaskTy = VT.changeVectorElementTypeToInteger();
776 // What is the size of each element in the vector mask.
777 EVT BitTy = MaskTy.getScalarType();
779 Mask = DAG.getSelect(DL, BitTy, Mask,
780 DAG.getConstant(APInt::getAllOnesValue(BitTy.getSizeInBits()), DL,
782 DAG.getConstant(0, DL, BitTy));
784 // Broadcast the mask so that the entire vector is all-one or all zero.
785 Mask = DAG.getSplatBuildVector(MaskTy, DL, Mask);
787 // Bitcast the operands to be the same type as the mask.
788 // This is needed when we select between FP types because
789 // the mask is a vector of integers.
790 Op1 = DAG.getNode(ISD::BITCAST, DL, MaskTy, Op1);
791 Op2 = DAG.getNode(ISD::BITCAST, DL, MaskTy, Op2);
793 SDValue AllOnes = DAG.getConstant(
794 APInt::getAllOnesValue(BitTy.getSizeInBits()), DL, MaskTy);
795 SDValue NotMask = DAG.getNode(ISD::XOR, DL, MaskTy, Mask, AllOnes);
797 Op1 = DAG.getNode(ISD::AND, DL, MaskTy, Op1, Mask);
798 Op2 = DAG.getNode(ISD::AND, DL, MaskTy, Op2, NotMask);
799 SDValue Val = DAG.getNode(ISD::OR, DL, MaskTy, Op1, Op2);
800 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Val);
803 SDValue VectorLegalizer::ExpandSEXTINREG(SDValue Op) {
804 EVT VT = Op.getValueType();
806 // Make sure that the SRA and SHL instructions are available.
807 if (TLI.getOperationAction(ISD::SRA, VT) == TargetLowering::Expand ||
808 TLI.getOperationAction(ISD::SHL, VT) == TargetLowering::Expand)
809 return DAG.UnrollVectorOp(Op.getNode());
812 EVT OrigTy = cast<VTSDNode>(Op->getOperand(1))->getVT();
814 unsigned BW = VT.getScalarSizeInBits();
815 unsigned OrigBW = OrigTy.getScalarSizeInBits();
816 SDValue ShiftSz = DAG.getConstant(BW - OrigBW, DL, VT);
818 Op = Op.getOperand(0);
819 Op = DAG.getNode(ISD::SHL, DL, VT, Op, ShiftSz);
820 return DAG.getNode(ISD::SRA, DL, VT, Op, ShiftSz);
823 // Generically expand a vector anyext in register to a shuffle of the relevant
824 // lanes into the appropriate locations, with other lanes left undef.
825 SDValue VectorLegalizer::ExpandANY_EXTEND_VECTOR_INREG(SDValue Op) {
827 EVT VT = Op.getValueType();
828 int NumElements = VT.getVectorNumElements();
829 SDValue Src = Op.getOperand(0);
830 EVT SrcVT = Src.getValueType();
831 int NumSrcElements = SrcVT.getVectorNumElements();
833 // Build a base mask of undef shuffles.
834 SmallVector<int, 16> ShuffleMask;
835 ShuffleMask.resize(NumSrcElements, -1);
837 // Place the extended lanes into the correct locations.
838 int ExtLaneScale = NumSrcElements / NumElements;
839 int EndianOffset = DAG.getDataLayout().isBigEndian() ? ExtLaneScale - 1 : 0;
840 for (int i = 0; i < NumElements; ++i)
841 ShuffleMask[i * ExtLaneScale + EndianOffset] = i;
844 ISD::BITCAST, DL, VT,
845 DAG.getVectorShuffle(SrcVT, DL, Src, DAG.getUNDEF(SrcVT), ShuffleMask));
848 SDValue VectorLegalizer::ExpandSIGN_EXTEND_VECTOR_INREG(SDValue Op) {
850 EVT VT = Op.getValueType();
851 SDValue Src = Op.getOperand(0);
852 EVT SrcVT = Src.getValueType();
854 // First build an any-extend node which can be legalized above when we
855 // recurse through it.
856 Op = DAG.getAnyExtendVectorInReg(Src, DL, VT);
858 // Now we need sign extend. Do this by shifting the elements. Even if these
859 // aren't legal operations, they have a better chance of being legalized
860 // without full scalarization than the sign extension does.
861 unsigned EltWidth = VT.getScalarSizeInBits();
862 unsigned SrcEltWidth = SrcVT.getScalarSizeInBits();
863 SDValue ShiftAmount = DAG.getConstant(EltWidth - SrcEltWidth, DL, VT);
864 return DAG.getNode(ISD::SRA, DL, VT,
865 DAG.getNode(ISD::SHL, DL, VT, Op, ShiftAmount),
869 // Generically expand a vector zext in register to a shuffle of the relevant
870 // lanes into the appropriate locations, a blend of zero into the high bits,
871 // and a bitcast to the wider element type.
872 SDValue VectorLegalizer::ExpandZERO_EXTEND_VECTOR_INREG(SDValue Op) {
874 EVT VT = Op.getValueType();
875 int NumElements = VT.getVectorNumElements();
876 SDValue Src = Op.getOperand(0);
877 EVT SrcVT = Src.getValueType();
878 int NumSrcElements = SrcVT.getVectorNumElements();
880 // Build up a zero vector to blend into this one.
881 SDValue Zero = DAG.getConstant(0, DL, SrcVT);
883 // Shuffle the incoming lanes into the correct position, and pull all other
884 // lanes from the zero vector.
885 SmallVector<int, 16> ShuffleMask;
886 ShuffleMask.reserve(NumSrcElements);
887 for (int i = 0; i < NumSrcElements; ++i)
888 ShuffleMask.push_back(i);
890 int ExtLaneScale = NumSrcElements / NumElements;
891 int EndianOffset = DAG.getDataLayout().isBigEndian() ? ExtLaneScale - 1 : 0;
892 for (int i = 0; i < NumElements; ++i)
893 ShuffleMask[i * ExtLaneScale + EndianOffset] = NumSrcElements + i;
895 return DAG.getNode(ISD::BITCAST, DL, VT,
896 DAG.getVectorShuffle(SrcVT, DL, Zero, Src, ShuffleMask));
899 static void createBSWAPShuffleMask(EVT VT, SmallVectorImpl<int> &ShuffleMask) {
900 int ScalarSizeInBytes = VT.getScalarSizeInBits() / 8;
901 for (int I = 0, E = VT.getVectorNumElements(); I != E; ++I)
902 for (int J = ScalarSizeInBytes - 1; J >= 0; --J)
903 ShuffleMask.push_back((I * ScalarSizeInBytes) + J);
906 SDValue VectorLegalizer::ExpandBSWAP(SDValue Op) {
907 EVT VT = Op.getValueType();
909 // Generate a byte wise shuffle mask for the BSWAP.
910 SmallVector<int, 16> ShuffleMask;
911 createBSWAPShuffleMask(VT, ShuffleMask);
912 EVT ByteVT = EVT::getVectorVT(*DAG.getContext(), MVT::i8, ShuffleMask.size());
914 // Only emit a shuffle if the mask is legal.
915 if (!TLI.isShuffleMaskLegal(ShuffleMask, ByteVT))
916 return DAG.UnrollVectorOp(Op.getNode());
919 Op = DAG.getNode(ISD::BITCAST, DL, ByteVT, Op.getOperand(0));
920 Op = DAG.getVectorShuffle(ByteVT, DL, Op, DAG.getUNDEF(ByteVT), ShuffleMask);
921 return DAG.getNode(ISD::BITCAST, DL, VT, Op);
924 SDValue VectorLegalizer::ExpandBITREVERSE(SDValue Op) {
925 EVT VT = Op.getValueType();
927 // If we have the scalar operation, it's probably cheaper to unroll it.
928 if (TLI.isOperationLegalOrCustom(ISD::BITREVERSE, VT.getScalarType()))
929 return DAG.UnrollVectorOp(Op.getNode());
931 // If the vector element width is a whole number of bytes, test if its legal
932 // to BSWAP shuffle the bytes and then perform the BITREVERSE on the byte
933 // vector. This greatly reduces the number of bit shifts necessary.
934 unsigned ScalarSizeInBits = VT.getScalarSizeInBits();
935 if (ScalarSizeInBits > 8 && (ScalarSizeInBits % 8) == 0) {
936 SmallVector<int, 16> BSWAPMask;
937 createBSWAPShuffleMask(VT, BSWAPMask);
939 EVT ByteVT = EVT::getVectorVT(*DAG.getContext(), MVT::i8, BSWAPMask.size());
940 if (TLI.isShuffleMaskLegal(BSWAPMask, ByteVT) &&
941 (TLI.isOperationLegalOrCustom(ISD::BITREVERSE, ByteVT) ||
942 (TLI.isOperationLegalOrCustom(ISD::SHL, ByteVT) &&
943 TLI.isOperationLegalOrCustom(ISD::SRL, ByteVT) &&
944 TLI.isOperationLegalOrCustomOrPromote(ISD::AND, ByteVT) &&
945 TLI.isOperationLegalOrCustomOrPromote(ISD::OR, ByteVT)))) {
947 Op = DAG.getNode(ISD::BITCAST, DL, ByteVT, Op.getOperand(0));
948 Op = DAG.getVectorShuffle(ByteVT, DL, Op, DAG.getUNDEF(ByteVT),
950 Op = DAG.getNode(ISD::BITREVERSE, DL, ByteVT, Op);
951 return DAG.getNode(ISD::BITCAST, DL, VT, Op);
955 // If we have the appropriate vector bit operations, it is better to use them
956 // than unrolling and expanding each component.
957 if (!TLI.isOperationLegalOrCustom(ISD::SHL, VT) ||
958 !TLI.isOperationLegalOrCustom(ISD::SRL, VT) ||
959 !TLI.isOperationLegalOrCustomOrPromote(ISD::AND, VT) ||
960 !TLI.isOperationLegalOrCustomOrPromote(ISD::OR, VT))
961 return DAG.UnrollVectorOp(Op.getNode());
963 // Let LegalizeDAG handle this later.
967 SDValue VectorLegalizer::ExpandVSELECT(SDValue Op) {
968 // Implement VSELECT in terms of XOR, AND, OR
969 // on platforms which do not support blend natively.
972 SDValue Mask = Op.getOperand(0);
973 SDValue Op1 = Op.getOperand(1);
974 SDValue Op2 = Op.getOperand(2);
976 EVT VT = Mask.getValueType();
978 // If we can't even use the basic vector operations of
979 // AND,OR,XOR, we will have to scalarize the op.
980 // Notice that the operation may be 'promoted' which means that it is
981 // 'bitcasted' to another type which is handled.
982 // This operation also isn't safe with AND, OR, XOR when the boolean
983 // type is 0/1 as we need an all ones vector constant to mask with.
984 // FIXME: Sign extend 1 to all ones if thats legal on the target.
985 if (TLI.getOperationAction(ISD::AND, VT) == TargetLowering::Expand ||
986 TLI.getOperationAction(ISD::XOR, VT) == TargetLowering::Expand ||
987 TLI.getOperationAction(ISD::OR, VT) == TargetLowering::Expand ||
988 TLI.getBooleanContents(Op1.getValueType()) !=
989 TargetLowering::ZeroOrNegativeOneBooleanContent)
990 return DAG.UnrollVectorOp(Op.getNode());
992 // If the mask and the type are different sizes, unroll the vector op. This
993 // can occur when getSetCCResultType returns something that is different in
994 // size from the operand types. For example, v4i8 = select v4i32, v4i8, v4i8.
995 if (VT.getSizeInBits() != Op1.getValueSizeInBits())
996 return DAG.UnrollVectorOp(Op.getNode());
998 // Bitcast the operands to be the same type as the mask.
999 // This is needed when we select between FP types because
1000 // the mask is a vector of integers.
1001 Op1 = DAG.getNode(ISD::BITCAST, DL, VT, Op1);
1002 Op2 = DAG.getNode(ISD::BITCAST, DL, VT, Op2);
1004 SDValue AllOnes = DAG.getConstant(
1005 APInt::getAllOnesValue(VT.getScalarSizeInBits()), DL, VT);
1006 SDValue NotMask = DAG.getNode(ISD::XOR, DL, VT, Mask, AllOnes);
1008 Op1 = DAG.getNode(ISD::AND, DL, VT, Op1, Mask);
1009 Op2 = DAG.getNode(ISD::AND, DL, VT, Op2, NotMask);
1010 SDValue Val = DAG.getNode(ISD::OR, DL, VT, Op1, Op2);
1011 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Val);
1014 SDValue VectorLegalizer::ExpandUINT_TO_FLOAT(SDValue Op) {
1015 EVT VT = Op.getOperand(0).getValueType();
1018 // Make sure that the SINT_TO_FP and SRL instructions are available.
1019 if (TLI.getOperationAction(ISD::SINT_TO_FP, VT) == TargetLowering::Expand ||
1020 TLI.getOperationAction(ISD::SRL, VT) == TargetLowering::Expand)
1021 return DAG.UnrollVectorOp(Op.getNode());
1023 unsigned BW = VT.getScalarSizeInBits();
1024 assert((BW == 64 || BW == 32) &&
1025 "Elements in vector-UINT_TO_FP must be 32 or 64 bits wide");
1027 SDValue HalfWord = DAG.getConstant(BW / 2, DL, VT);
1029 // Constants to clear the upper part of the word.
1030 // Notice that we can also use SHL+SHR, but using a constant is slightly
1032 uint64_t HWMask = (BW == 64) ? 0x00000000FFFFFFFF : 0x0000FFFF;
1033 SDValue HalfWordMask = DAG.getConstant(HWMask, DL, VT);
1035 // Two to the power of half-word-size.
1036 SDValue TWOHW = DAG.getConstantFP(1ULL << (BW / 2), DL, Op.getValueType());
1038 // Clear upper part of LO, lower HI
1039 SDValue HI = DAG.getNode(ISD::SRL, DL, VT, Op.getOperand(0), HalfWord);
1040 SDValue LO = DAG.getNode(ISD::AND, DL, VT, Op.getOperand(0), HalfWordMask);
1042 // Convert hi and lo to floats
1043 // Convert the hi part back to the upper values
1044 // TODO: Can any fast-math-flags be set on these nodes?
1045 SDValue fHI = DAG.getNode(ISD::SINT_TO_FP, DL, Op.getValueType(), HI);
1046 fHI = DAG.getNode(ISD::FMUL, DL, Op.getValueType(), fHI, TWOHW);
1047 SDValue fLO = DAG.getNode(ISD::SINT_TO_FP, DL, Op.getValueType(), LO);
1049 // Add the two halves
1050 return DAG.getNode(ISD::FADD, DL, Op.getValueType(), fHI, fLO);
1053 SDValue VectorLegalizer::ExpandFNEG(SDValue Op) {
1054 if (TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType())) {
1056 SDValue Zero = DAG.getConstantFP(-0.0, DL, Op.getValueType());
1057 // TODO: If FNEG had fast-math-flags, they'd get propagated to this FSUB.
1058 return DAG.getNode(ISD::FSUB, DL, Op.getValueType(),
1059 Zero, Op.getOperand(0));
1061 return DAG.UnrollVectorOp(Op.getNode());
1064 SDValue VectorLegalizer::ExpandFSUB(SDValue Op) {
1065 // For floating-point values, (a-b) is the same as a+(-b). If FNEG is legal,
1066 // we can defer this to operation legalization where it will be lowered as
1068 EVT VT = Op.getValueType();
1069 if (TLI.isOperationLegalOrCustom(ISD::FNEG, VT) &&
1070 TLI.isOperationLegalOrCustom(ISD::FADD, VT))
1071 return Op; // Defer to LegalizeDAG
1073 return DAG.UnrollVectorOp(Op.getNode());
1076 SDValue VectorLegalizer::ExpandCTLZ(SDValue Op) {
1077 EVT VT = Op.getValueType();
1078 unsigned NumBitsPerElt = VT.getScalarSizeInBits();
1080 // If the non-ZERO_UNDEF version is supported we can use that instead.
1081 if (Op.getOpcode() == ISD::CTLZ_ZERO_UNDEF &&
1082 TLI.isOperationLegalOrCustom(ISD::CTLZ, VT)) {
1084 return DAG.getNode(ISD::CTLZ, DL, VT, Op.getOperand(0));
1087 // If we have the appropriate vector bit operations, it is better to use them
1088 // than unrolling and expanding each component.
1089 if (isPowerOf2_32(NumBitsPerElt) &&
1090 TLI.isOperationLegalOrCustom(ISD::CTPOP, VT) &&
1091 TLI.isOperationLegalOrCustom(ISD::SRL, VT) &&
1092 TLI.isOperationLegalOrCustomOrPromote(ISD::OR, VT))
1095 // Otherwise go ahead and unroll.
1096 return DAG.UnrollVectorOp(Op.getNode());
1099 SDValue VectorLegalizer::ExpandCTTZ(SDValue Op) {
1100 EVT VT = Op.getValueType();
1101 unsigned NumBitsPerElt = VT.getScalarSizeInBits();
1103 // If the non-ZERO_UNDEF version is supported we can use that instead.
1104 if (TLI.isOperationLegalOrCustom(ISD::CTTZ, VT)) {
1106 return DAG.getNode(ISD::CTTZ, DL, VT, Op.getOperand(0));
1109 // If we have the appropriate vector bit operations, it is better to use them
1110 // than unrolling and expanding each component.
1111 if (isPowerOf2_32(NumBitsPerElt) &&
1112 (TLI.isOperationLegalOrCustom(ISD::CTPOP, VT) ||
1113 TLI.isOperationLegalOrCustom(ISD::CTLZ, VT)) &&
1114 TLI.isOperationLegalOrCustom(ISD::SUB, VT) &&
1115 TLI.isOperationLegalOrCustomOrPromote(ISD::AND, VT) &&
1116 TLI.isOperationLegalOrCustomOrPromote(ISD::XOR, VT))
1119 // Otherwise go ahead and unroll.
1120 return DAG.UnrollVectorOp(Op.getNode());
1123 SDValue VectorLegalizer::ExpandStrictFPOp(SDValue Op) {
1124 EVT VT = Op.getValueType();
1125 EVT EltVT = VT.getVectorElementType();
1126 unsigned NumElems = VT.getVectorNumElements();
1127 unsigned NumOpers = Op.getNumOperands();
1128 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1129 EVT ValueVTs[] = {EltVT, MVT::Other};
1130 SDValue Chain = Op.getOperand(0);
1133 SmallVector<SDValue, 32> OpValues;
1134 SmallVector<SDValue, 32> OpChains;
1135 for (unsigned i = 0; i < NumElems; ++i) {
1136 SmallVector<SDValue, 4> Opers;
1137 SDValue Idx = DAG.getConstant(i, dl,
1138 TLI.getVectorIdxTy(DAG.getDataLayout()));
1140 // The Chain is the first operand.
1141 Opers.push_back(Chain);
1143 // Now process the remaining operands.
1144 for (unsigned j = 1; j < NumOpers; ++j) {
1145 SDValue Oper = Op.getOperand(j);
1146 EVT OperVT = Oper.getValueType();
1148 if (OperVT.isVector())
1149 Oper = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
1152 Opers.push_back(Oper);
1155 SDValue ScalarOp = DAG.getNode(Op->getOpcode(), dl, ValueVTs, Opers);
1157 OpValues.push_back(ScalarOp.getValue(0));
1158 OpChains.push_back(ScalarOp.getValue(1));
1161 SDValue Result = DAG.getBuildVector(VT, dl, OpValues);
1162 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OpChains);
1164 AddLegalizedOperand(Op.getValue(0), Result);
1165 AddLegalizedOperand(Op.getValue(1), NewChain);
1167 return Op.getResNo() ? NewChain : Result;
1170 SDValue VectorLegalizer::UnrollVSETCC(SDValue Op) {
1171 EVT VT = Op.getValueType();
1172 unsigned NumElems = VT.getVectorNumElements();
1173 EVT EltVT = VT.getVectorElementType();
1174 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1), CC = Op.getOperand(2);
1175 EVT TmpEltVT = LHS.getValueType().getVectorElementType();
1177 SmallVector<SDValue, 8> Ops(NumElems);
1178 for (unsigned i = 0; i < NumElems; ++i) {
1179 SDValue LHSElem = DAG.getNode(
1180 ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, LHS,
1181 DAG.getConstant(i, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
1182 SDValue RHSElem = DAG.getNode(
1183 ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, RHS,
1184 DAG.getConstant(i, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
1185 Ops[i] = DAG.getNode(ISD::SETCC, dl,
1186 TLI.getSetCCResultType(DAG.getDataLayout(),
1187 *DAG.getContext(), TmpEltVT),
1188 LHSElem, RHSElem, CC);
1189 Ops[i] = DAG.getSelect(dl, EltVT, Ops[i],
1190 DAG.getConstant(APInt::getAllOnesValue
1191 (EltVT.getSizeInBits()), dl, EltVT),
1192 DAG.getConstant(0, dl, EltVT));
1194 return DAG.getBuildVector(VT, dl, Ops);
1197 bool SelectionDAG::LegalizeVectors() {
1198 return VectorLegalizer(*this).Run();