1 //===-- TargetLowering.cpp - Implement the TargetLowering class -----------===//
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 implements the TargetLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Target/TargetLowering.h"
15 #include "llvm/ADT/BitVector.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/CodeGen/CallingConvLower.h"
18 #include "llvm/CodeGen/MachineFrameInfo.h"
19 #include "llvm/CodeGen/MachineFunction.h"
20 #include "llvm/CodeGen/MachineJumpTableInfo.h"
21 #include "llvm/CodeGen/MachineRegisterInfo.h"
22 #include "llvm/CodeGen/SelectionDAG.h"
23 #include "llvm/IR/DataLayout.h"
24 #include "llvm/IR/DerivedTypes.h"
25 #include "llvm/IR/GlobalVariable.h"
26 #include "llvm/IR/LLVMContext.h"
27 #include "llvm/MC/MCAsmInfo.h"
28 #include "llvm/MC/MCExpr.h"
29 #include "llvm/Support/ErrorHandling.h"
30 #include "llvm/Support/KnownBits.h"
31 #include "llvm/Support/MathExtras.h"
32 #include "llvm/Target/TargetLoweringObjectFile.h"
33 #include "llvm/Target/TargetMachine.h"
34 #include "llvm/Target/TargetRegisterInfo.h"
35 #include "llvm/Target/TargetSubtargetInfo.h"
39 /// NOTE: The TargetMachine owns TLOF.
40 TargetLowering::TargetLowering(const TargetMachine &tm)
41 : TargetLoweringBase(tm) {}
43 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
47 bool TargetLowering::isPositionIndependent() const {
48 return getTargetMachine().isPositionIndependent();
51 /// Check whether a given call node is in tail position within its function. If
52 /// so, it sets Chain to the input chain of the tail call.
53 bool TargetLowering::isInTailCallPosition(SelectionDAG &DAG, SDNode *Node,
54 SDValue &Chain) const {
55 const Function *F = DAG.getMachineFunction().getFunction();
57 // Conservatively require the attributes of the call to match those of
58 // the return. Ignore noalias because it doesn't affect the call sequence.
59 AttributeList CallerAttrs = F->getAttributes();
60 if (AttrBuilder(CallerAttrs, AttributeList::ReturnIndex)
61 .removeAttribute(Attribute::NoAlias)
65 // It's not safe to eliminate the sign / zero extension of the return value.
66 if (CallerAttrs.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt) ||
67 CallerAttrs.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt))
70 // Check if the only use is a function return node.
71 return isUsedByReturnOnly(Node, Chain);
74 bool TargetLowering::parametersInCSRMatch(const MachineRegisterInfo &MRI,
75 const uint32_t *CallerPreservedMask,
76 const SmallVectorImpl<CCValAssign> &ArgLocs,
77 const SmallVectorImpl<SDValue> &OutVals) const {
78 for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) {
79 const CCValAssign &ArgLoc = ArgLocs[I];
80 if (!ArgLoc.isRegLoc())
82 unsigned Reg = ArgLoc.getLocReg();
83 // Only look at callee saved registers.
84 if (MachineOperand::clobbersPhysReg(CallerPreservedMask, Reg))
86 // Check that we pass the value used for the caller.
87 // (We look for a CopyFromReg reading a virtual register that is used
88 // for the function live-in value of register Reg)
89 SDValue Value = OutVals[I];
90 if (Value->getOpcode() != ISD::CopyFromReg)
92 unsigned ArgReg = cast<RegisterSDNode>(Value->getOperand(1))->getReg();
93 if (MRI.getLiveInPhysReg(ArgReg) != Reg)
99 /// \brief Set CallLoweringInfo attribute flags based on a call instruction
100 /// and called function attributes.
101 void TargetLoweringBase::ArgListEntry::setAttributes(ImmutableCallSite *CS,
103 IsSExt = CS->paramHasAttr(ArgIdx, Attribute::SExt);
104 IsZExt = CS->paramHasAttr(ArgIdx, Attribute::ZExt);
105 IsInReg = CS->paramHasAttr(ArgIdx, Attribute::InReg);
106 IsSRet = CS->paramHasAttr(ArgIdx, Attribute::StructRet);
107 IsNest = CS->paramHasAttr(ArgIdx, Attribute::Nest);
108 IsByVal = CS->paramHasAttr(ArgIdx, Attribute::ByVal);
109 IsInAlloca = CS->paramHasAttr(ArgIdx, Attribute::InAlloca);
110 IsReturned = CS->paramHasAttr(ArgIdx, Attribute::Returned);
111 IsSwiftSelf = CS->paramHasAttr(ArgIdx, Attribute::SwiftSelf);
112 IsSwiftError = CS->paramHasAttr(ArgIdx, Attribute::SwiftError);
113 Alignment = CS->getParamAlignment(ArgIdx);
116 /// Generate a libcall taking the given operands as arguments and returning a
117 /// result of type RetVT.
118 std::pair<SDValue, SDValue>
119 TargetLowering::makeLibCall(SelectionDAG &DAG, RTLIB::Libcall LC, EVT RetVT,
120 ArrayRef<SDValue> Ops, bool isSigned,
121 const SDLoc &dl, bool doesNotReturn,
122 bool isReturnValueUsed) const {
123 TargetLowering::ArgListTy Args;
124 Args.reserve(Ops.size());
126 TargetLowering::ArgListEntry Entry;
127 for (SDValue Op : Ops) {
129 Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext());
130 Entry.IsSExt = shouldSignExtendTypeInLibCall(Op.getValueType(), isSigned);
131 Entry.IsZExt = !shouldSignExtendTypeInLibCall(Op.getValueType(), isSigned);
132 Args.push_back(Entry);
135 if (LC == RTLIB::UNKNOWN_LIBCALL)
136 report_fatal_error("Unsupported library call operation!");
137 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
138 getPointerTy(DAG.getDataLayout()));
140 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
141 TargetLowering::CallLoweringInfo CLI(DAG);
142 bool signExtend = shouldSignExtendTypeInLibCall(RetVT, isSigned);
144 .setChain(DAG.getEntryNode())
145 .setLibCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args))
146 .setNoReturn(doesNotReturn)
147 .setDiscardResult(!isReturnValueUsed)
148 .setSExtResult(signExtend)
149 .setZExtResult(!signExtend);
150 return LowerCallTo(CLI);
153 /// Soften the operands of a comparison. This code is shared among BR_CC,
154 /// SELECT_CC, and SETCC handlers.
155 void TargetLowering::softenSetCCOperands(SelectionDAG &DAG, EVT VT,
156 SDValue &NewLHS, SDValue &NewRHS,
157 ISD::CondCode &CCCode,
158 const SDLoc &dl) const {
159 assert((VT == MVT::f32 || VT == MVT::f64 || VT == MVT::f128 || VT == MVT::ppcf128)
160 && "Unsupported setcc type!");
162 // Expand into one or more soft-fp libcall(s).
163 RTLIB::Libcall LC1 = RTLIB::UNKNOWN_LIBCALL, LC2 = RTLIB::UNKNOWN_LIBCALL;
164 bool ShouldInvertCC = false;
168 LC1 = (VT == MVT::f32) ? RTLIB::OEQ_F32 :
169 (VT == MVT::f64) ? RTLIB::OEQ_F64 :
170 (VT == MVT::f128) ? RTLIB::OEQ_F128 : RTLIB::OEQ_PPCF128;
174 LC1 = (VT == MVT::f32) ? RTLIB::UNE_F32 :
175 (VT == MVT::f64) ? RTLIB::UNE_F64 :
176 (VT == MVT::f128) ? RTLIB::UNE_F128 : RTLIB::UNE_PPCF128;
180 LC1 = (VT == MVT::f32) ? RTLIB::OGE_F32 :
181 (VT == MVT::f64) ? RTLIB::OGE_F64 :
182 (VT == MVT::f128) ? RTLIB::OGE_F128 : RTLIB::OGE_PPCF128;
186 LC1 = (VT == MVT::f32) ? RTLIB::OLT_F32 :
187 (VT == MVT::f64) ? RTLIB::OLT_F64 :
188 (VT == MVT::f128) ? RTLIB::OLT_F128 : RTLIB::OLT_PPCF128;
192 LC1 = (VT == MVT::f32) ? RTLIB::OLE_F32 :
193 (VT == MVT::f64) ? RTLIB::OLE_F64 :
194 (VT == MVT::f128) ? RTLIB::OLE_F128 : RTLIB::OLE_PPCF128;
198 LC1 = (VT == MVT::f32) ? RTLIB::OGT_F32 :
199 (VT == MVT::f64) ? RTLIB::OGT_F64 :
200 (VT == MVT::f128) ? RTLIB::OGT_F128 : RTLIB::OGT_PPCF128;
203 LC1 = (VT == MVT::f32) ? RTLIB::UO_F32 :
204 (VT == MVT::f64) ? RTLIB::UO_F64 :
205 (VT == MVT::f128) ? RTLIB::UO_F128 : RTLIB::UO_PPCF128;
208 LC1 = (VT == MVT::f32) ? RTLIB::O_F32 :
209 (VT == MVT::f64) ? RTLIB::O_F64 :
210 (VT == MVT::f128) ? RTLIB::O_F128 : RTLIB::O_PPCF128;
213 // SETONE = SETOLT | SETOGT
214 LC1 = (VT == MVT::f32) ? RTLIB::OLT_F32 :
215 (VT == MVT::f64) ? RTLIB::OLT_F64 :
216 (VT == MVT::f128) ? RTLIB::OLT_F128 : RTLIB::OLT_PPCF128;
217 LC2 = (VT == MVT::f32) ? RTLIB::OGT_F32 :
218 (VT == MVT::f64) ? RTLIB::OGT_F64 :
219 (VT == MVT::f128) ? RTLIB::OGT_F128 : RTLIB::OGT_PPCF128;
222 LC1 = (VT == MVT::f32) ? RTLIB::UO_F32 :
223 (VT == MVT::f64) ? RTLIB::UO_F64 :
224 (VT == MVT::f128) ? RTLIB::UO_F128 : RTLIB::UO_PPCF128;
225 LC2 = (VT == MVT::f32) ? RTLIB::OEQ_F32 :
226 (VT == MVT::f64) ? RTLIB::OEQ_F64 :
227 (VT == MVT::f128) ? RTLIB::OEQ_F128 : RTLIB::OEQ_PPCF128;
230 // Invert CC for unordered comparisons
231 ShouldInvertCC = true;
234 LC1 = (VT == MVT::f32) ? RTLIB::OGE_F32 :
235 (VT == MVT::f64) ? RTLIB::OGE_F64 :
236 (VT == MVT::f128) ? RTLIB::OGE_F128 : RTLIB::OGE_PPCF128;
239 LC1 = (VT == MVT::f32) ? RTLIB::OGT_F32 :
240 (VT == MVT::f64) ? RTLIB::OGT_F64 :
241 (VT == MVT::f128) ? RTLIB::OGT_F128 : RTLIB::OGT_PPCF128;
244 LC1 = (VT == MVT::f32) ? RTLIB::OLE_F32 :
245 (VT == MVT::f64) ? RTLIB::OLE_F64 :
246 (VT == MVT::f128) ? RTLIB::OLE_F128 : RTLIB::OLE_PPCF128;
249 LC1 = (VT == MVT::f32) ? RTLIB::OLT_F32 :
250 (VT == MVT::f64) ? RTLIB::OLT_F64 :
251 (VT == MVT::f128) ? RTLIB::OLT_F128 : RTLIB::OLT_PPCF128;
253 default: llvm_unreachable("Do not know how to soften this setcc!");
257 // Use the target specific return value for comparions lib calls.
258 EVT RetVT = getCmpLibcallReturnType();
259 SDValue Ops[2] = {NewLHS, NewRHS};
260 NewLHS = makeLibCall(DAG, LC1, RetVT, Ops, false /*sign irrelevant*/,
262 NewRHS = DAG.getConstant(0, dl, RetVT);
264 CCCode = getCmpLibcallCC(LC1);
266 CCCode = getSetCCInverse(CCCode, /*isInteger=*/true);
268 if (LC2 != RTLIB::UNKNOWN_LIBCALL) {
269 SDValue Tmp = DAG.getNode(
271 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), RetVT),
272 NewLHS, NewRHS, DAG.getCondCode(CCCode));
273 NewLHS = makeLibCall(DAG, LC2, RetVT, Ops, false/*sign irrelevant*/,
275 NewLHS = DAG.getNode(
277 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), RetVT),
278 NewLHS, NewRHS, DAG.getCondCode(getCmpLibcallCC(LC2)));
279 NewLHS = DAG.getNode(ISD::OR, dl, Tmp.getValueType(), Tmp, NewLHS);
284 /// Return the entry encoding for a jump table in the current function. The
285 /// returned value is a member of the MachineJumpTableInfo::JTEntryKind enum.
286 unsigned TargetLowering::getJumpTableEncoding() const {
287 // In non-pic modes, just use the address of a block.
288 if (!isPositionIndependent())
289 return MachineJumpTableInfo::EK_BlockAddress;
291 // In PIC mode, if the target supports a GPRel32 directive, use it.
292 if (getTargetMachine().getMCAsmInfo()->getGPRel32Directive() != nullptr)
293 return MachineJumpTableInfo::EK_GPRel32BlockAddress;
295 // Otherwise, use a label difference.
296 return MachineJumpTableInfo::EK_LabelDifference32;
299 SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table,
300 SelectionDAG &DAG) const {
301 // If our PIC model is GP relative, use the global offset table as the base.
302 unsigned JTEncoding = getJumpTableEncoding();
304 if ((JTEncoding == MachineJumpTableInfo::EK_GPRel64BlockAddress) ||
305 (JTEncoding == MachineJumpTableInfo::EK_GPRel32BlockAddress))
306 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy(DAG.getDataLayout()));
311 /// This returns the relocation base for the given PIC jumptable, the same as
312 /// getPICJumpTableRelocBase, but as an MCExpr.
314 TargetLowering::getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
315 unsigned JTI,MCContext &Ctx) const{
316 // The normal PIC reloc base is the label at the start of the jump table.
317 return MCSymbolRefExpr::create(MF->getJTISymbol(JTI, Ctx), Ctx);
321 TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
322 const TargetMachine &TM = getTargetMachine();
323 const GlobalValue *GV = GA->getGlobal();
325 // If the address is not even local to this DSO we will have to load it from
326 // a got and then add the offset.
327 if (!TM.shouldAssumeDSOLocal(*GV->getParent(), GV))
330 // If the code is position independent we will have to add a base register.
331 if (isPositionIndependent())
334 // Otherwise we can do it.
338 //===----------------------------------------------------------------------===//
339 // Optimization Methods
340 //===----------------------------------------------------------------------===//
342 /// If the specified instruction has a constant integer operand and there are
343 /// bits set in that constant that are not demanded, then clear those bits and
345 bool TargetLowering::ShrinkDemandedConstant(SDValue Op, const APInt &Demanded,
346 TargetLoweringOpt &TLO) const {
347 SelectionDAG &DAG = TLO.DAG;
349 unsigned Opcode = Op.getOpcode();
351 // Do target-specific constant optimization.
352 if (targetShrinkDemandedConstant(Op, Demanded, TLO))
353 return TLO.New.getNode();
355 // FIXME: ISD::SELECT, ISD::SELECT_CC
362 auto *Op1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
366 // If this is a 'not' op, don't touch it because that's a canonical form.
367 const APInt &C = Op1C->getAPIntValue();
368 if (Opcode == ISD::XOR && (C | ~Demanded).isAllOnesValue())
371 if (C.intersects(~Demanded)) {
372 EVT VT = Op.getValueType();
373 SDValue NewC = DAG.getConstant(Demanded & C, DL, VT);
374 SDValue NewOp = DAG.getNode(Opcode, DL, VT, Op.getOperand(0), NewC);
375 return TLO.CombineTo(Op, NewOp);
385 /// Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the casts are free.
386 /// This uses isZExtFree and ZERO_EXTEND for the widening cast, but it could be
387 /// generalized for targets with other types of implicit widening casts.
388 bool TargetLowering::ShrinkDemandedOp(SDValue Op, unsigned BitWidth,
389 const APInt &Demanded,
390 TargetLoweringOpt &TLO) const {
391 assert(Op.getNumOperands() == 2 &&
392 "ShrinkDemandedOp only supports binary operators!");
393 assert(Op.getNode()->getNumValues() == 1 &&
394 "ShrinkDemandedOp only supports nodes with one result!");
396 SelectionDAG &DAG = TLO.DAG;
399 // Early return, as this function cannot handle vector types.
400 if (Op.getValueType().isVector())
403 // Don't do this if the node has another user, which may require the
405 if (!Op.getNode()->hasOneUse())
408 // Search for the smallest integer type with free casts to and from
409 // Op's type. For expedience, just check power-of-2 integer types.
410 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
411 unsigned DemandedSize = BitWidth - Demanded.countLeadingZeros();
412 unsigned SmallVTBits = DemandedSize;
413 if (!isPowerOf2_32(SmallVTBits))
414 SmallVTBits = NextPowerOf2(SmallVTBits);
415 for (; SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) {
416 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits);
417 if (TLI.isTruncateFree(Op.getValueType(), SmallVT) &&
418 TLI.isZExtFree(SmallVT, Op.getValueType())) {
419 // We found a type with free casts.
420 SDValue X = DAG.getNode(
421 Op.getOpcode(), dl, SmallVT,
422 DAG.getNode(ISD::TRUNCATE, dl, SmallVT, Op.getOperand(0)),
423 DAG.getNode(ISD::TRUNCATE, dl, SmallVT, Op.getOperand(1)));
424 bool NeedZext = DemandedSize > SmallVTBits;
425 SDValue Z = DAG.getNode(NeedZext ? ISD::ZERO_EXTEND : ISD::ANY_EXTEND,
426 dl, Op.getValueType(), X);
427 return TLO.CombineTo(Op, Z);
434 TargetLowering::SimplifyDemandedBits(SDNode *User, unsigned OpIdx,
435 const APInt &Demanded,
436 DAGCombinerInfo &DCI,
437 TargetLoweringOpt &TLO) const {
438 SDValue Op = User->getOperand(OpIdx);
441 if (!SimplifyDemandedBits(Op, Demanded, Known, TLO, 0, true))
445 // Old will not always be the same as Op. For example:
447 // Demanded = 0xffffff
448 // Op = i64 truncate (i32 and x, 0xffffff)
449 // In this case simplify demand bits will want to replace the 'and' node
450 // with the value 'x', which will give us:
451 // Old = i32 and x, 0xffffff
453 if (TLO.Old.hasOneUse()) {
454 // For the one use case, we just commit the change.
455 DCI.CommitTargetLoweringOpt(TLO);
459 // If Old has more than one use then it must be Op, because the
460 // AssumeSingleUse flag is not propogated to recursive calls of
461 // SimplifyDemanded bits, so the only node with multiple use that
462 // it will attempt to combine will be opt.
463 assert(TLO.Old == Op);
465 SmallVector <SDValue, 4> NewOps;
466 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
468 NewOps.push_back(TLO.New);
471 NewOps.push_back(User->getOperand(i));
473 TLO.DAG.UpdateNodeOperands(User, NewOps);
474 // Op has less users now, so we may be able to perform additional combines
476 DCI.AddToWorklist(Op.getNode());
477 // User's operands have been updated, so we may be able to do new combines
479 DCI.AddToWorklist(User);
483 bool TargetLowering::SimplifyDemandedBits(SDValue Op, APInt &DemandedMask,
484 DAGCombinerInfo &DCI) const {
486 SelectionDAG &DAG = DCI.DAG;
487 TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
488 !DCI.isBeforeLegalizeOps());
491 bool Simplified = SimplifyDemandedBits(Op, DemandedMask, Known, TLO);
493 DCI.CommitTargetLoweringOpt(TLO);
497 /// Look at Op. At this point, we know that only the DemandedMask bits of the
498 /// result of Op are ever used downstream. If we can use this information to
499 /// simplify Op, create a new simplified DAG node and return true, returning the
500 /// original and new nodes in Old and New. Otherwise, analyze the expression and
501 /// return a mask of Known bits for the expression (used to simplify the
502 /// caller). The Known bits may only be accurate for those bits in the
504 bool TargetLowering::SimplifyDemandedBits(SDValue Op,
505 const APInt &DemandedMask,
507 TargetLoweringOpt &TLO,
509 bool AssumeSingleUse) const {
510 unsigned BitWidth = DemandedMask.getBitWidth();
511 assert(Op.getScalarValueSizeInBits() == BitWidth &&
512 "Mask size mismatches value type size!");
513 APInt NewMask = DemandedMask;
515 auto &DL = TLO.DAG.getDataLayout();
517 // Don't know anything.
518 Known = KnownBits(BitWidth);
520 // Other users may use these bits.
521 if (!Op.getNode()->hasOneUse() && !AssumeSingleUse) {
523 // If not at the root, Just compute the Known bits to
524 // simplify things downstream.
525 TLO.DAG.computeKnownBits(Op, Known, Depth);
528 // If this is the root being simplified, allow it to have multiple uses,
529 // just set the NewMask to all bits.
530 NewMask = APInt::getAllOnesValue(BitWidth);
531 } else if (DemandedMask == 0) {
532 // Not demanding any bits from Op.
534 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType()));
536 } else if (Depth == 6) { // Limit search depth.
540 KnownBits Known2, KnownOut;
541 switch (Op.getOpcode()) {
543 // We know all of the bits for a constant!
544 Known.One = cast<ConstantSDNode>(Op)->getAPIntValue();
545 Known.Zero = ~Known.One;
546 return false; // Don't fall through, will infinitely loop.
547 case ISD::BUILD_VECTOR:
548 // Collect the known bits that are shared by every constant vector element.
549 Known.Zero.setAllBits(); Known.One.setAllBits();
550 for (SDValue SrcOp : Op->ops()) {
551 if (!isa<ConstantSDNode>(SrcOp)) {
552 // We can only handle all constant values - bail out with no known bits.
553 Known = KnownBits(BitWidth);
556 Known2.One = cast<ConstantSDNode>(SrcOp)->getAPIntValue();
557 Known2.Zero = ~Known2.One;
559 // BUILD_VECTOR can implicitly truncate sources, we must handle this.
560 if (Known2.One.getBitWidth() != BitWidth) {
561 assert(Known2.getBitWidth() > BitWidth &&
562 "Expected BUILD_VECTOR implicit truncation");
563 Known2 = Known2.trunc(BitWidth);
566 // Known bits are the values that are shared by every element.
567 // TODO: support per-element known bits.
568 Known.One &= Known2.One;
569 Known.Zero &= Known2.Zero;
571 return false; // Don't fall through, will infinitely loop.
573 // If the RHS is a constant, check to see if the LHS would be zero without
574 // using the bits from the RHS. Below, we use knowledge about the RHS to
575 // simplify the LHS, here we're using information from the LHS to simplify
577 if (ConstantSDNode *RHSC = isConstOrConstSplat(Op.getOperand(1))) {
578 SDValue Op0 = Op.getOperand(0);
580 // Do not increment Depth here; that can cause an infinite loop.
581 TLO.DAG.computeKnownBits(Op0, LHSKnown, Depth);
582 // If the LHS already has zeros where RHSC does, this and is dead.
583 if ((LHSKnown.Zero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
584 return TLO.CombineTo(Op, Op0);
586 // If any of the set bits in the RHS are known zero on the LHS, shrink
588 if (ShrinkDemandedConstant(Op, ~LHSKnown.Zero & NewMask, TLO))
591 // Bitwise-not (xor X, -1) is a special case: we don't usually shrink its
592 // constant, but if this 'and' is only clearing bits that were just set by
593 // the xor, then this 'and' can be eliminated by shrinking the mask of
594 // the xor. For example, for a 32-bit X:
595 // and (xor (srl X, 31), -1), 1 --> xor (srl X, 31), 1
596 if (isBitwiseNot(Op0) && Op0.hasOneUse() &&
597 LHSKnown.One == ~RHSC->getAPIntValue()) {
598 SDValue Xor = TLO.DAG.getNode(ISD::XOR, dl, Op.getValueType(),
599 Op0.getOperand(0), Op.getOperand(1));
600 return TLO.CombineTo(Op, Xor);
604 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, Known, TLO, Depth+1))
606 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
607 if (SimplifyDemandedBits(Op.getOperand(0), ~Known.Zero & NewMask,
608 Known2, TLO, Depth+1))
610 assert(!Known2.hasConflict() && "Bits known to be one AND zero?");
612 // If all of the demanded bits are known one on one side, return the other.
613 // These bits cannot contribute to the result of the 'and'.
614 if (NewMask.isSubsetOf(Known2.Zero | Known.One))
615 return TLO.CombineTo(Op, Op.getOperand(0));
616 if (NewMask.isSubsetOf(Known.Zero | Known2.One))
617 return TLO.CombineTo(Op, Op.getOperand(1));
618 // If all of the demanded bits in the inputs are known zeros, return zero.
619 if (NewMask.isSubsetOf(Known.Zero | Known2.Zero))
620 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, dl, Op.getValueType()));
621 // If the RHS is a constant, see if we can simplify it.
622 if (ShrinkDemandedConstant(Op, ~Known2.Zero & NewMask, TLO))
624 // If the operation can be done in a smaller type, do so.
625 if (ShrinkDemandedOp(Op, BitWidth, NewMask, TLO))
628 // Output known-1 bits are only known if set in both the LHS & RHS.
629 Known.One &= Known2.One;
630 // Output known-0 are known to be clear if zero in either the LHS | RHS.
631 Known.Zero |= Known2.Zero;
634 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, Known, TLO, Depth+1))
636 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
637 if (SimplifyDemandedBits(Op.getOperand(0), ~Known.One & NewMask,
638 Known2, TLO, Depth+1))
640 assert(!Known2.hasConflict() && "Bits known to be one AND zero?");
642 // If all of the demanded bits are known zero on one side, return the other.
643 // These bits cannot contribute to the result of the 'or'.
644 if (NewMask.isSubsetOf(Known2.One | Known.Zero))
645 return TLO.CombineTo(Op, Op.getOperand(0));
646 if (NewMask.isSubsetOf(Known.One | Known2.Zero))
647 return TLO.CombineTo(Op, Op.getOperand(1));
648 // If the RHS is a constant, see if we can simplify it.
649 if (ShrinkDemandedConstant(Op, NewMask, TLO))
651 // If the operation can be done in a smaller type, do so.
652 if (ShrinkDemandedOp(Op, BitWidth, NewMask, TLO))
655 // Output known-0 bits are only known if clear in both the LHS & RHS.
656 Known.Zero &= Known2.Zero;
657 // Output known-1 are known to be set if set in either the LHS | RHS.
658 Known.One |= Known2.One;
661 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, Known, TLO, Depth+1))
663 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
664 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, Known2, TLO, Depth+1))
666 assert(!Known2.hasConflict() && "Bits known to be one AND zero?");
668 // If all of the demanded bits are known zero on one side, return the other.
669 // These bits cannot contribute to the result of the 'xor'.
670 if (NewMask.isSubsetOf(Known.Zero))
671 return TLO.CombineTo(Op, Op.getOperand(0));
672 if (NewMask.isSubsetOf(Known2.Zero))
673 return TLO.CombineTo(Op, Op.getOperand(1));
674 // If the operation can be done in a smaller type, do so.
675 if (ShrinkDemandedOp(Op, BitWidth, NewMask, TLO))
678 // If all of the unknown bits are known to be zero on one side or the other
679 // (but not both) turn this into an *inclusive* or.
680 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
681 if ((NewMask & ~Known.Zero & ~Known2.Zero) == 0)
682 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(),
686 // Output known-0 bits are known if clear or set in both the LHS & RHS.
687 KnownOut.Zero = (Known.Zero & Known2.Zero) | (Known.One & Known2.One);
688 // Output known-1 are known to be set if set in only one of the LHS, RHS.
689 KnownOut.One = (Known.Zero & Known2.One) | (Known.One & Known2.Zero);
691 // If all of the demanded bits on one side are known, and all of the set
692 // bits on that side are also known to be set on the other side, turn this
693 // into an AND, as we know the bits will be cleared.
694 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
695 // NB: it is okay if more bits are known than are requested
696 if (NewMask.isSubsetOf(Known.Zero|Known.One)) { // all known on one side
697 if (Known.One == Known2.One) { // set bits are the same on both sides
698 EVT VT = Op.getValueType();
699 SDValue ANDC = TLO.DAG.getConstant(~Known.One & NewMask, dl, VT);
700 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT,
701 Op.getOperand(0), ANDC));
705 // If the RHS is a constant, see if we can change it. Don't alter a -1
706 // constant because that's a 'not' op, and that is better for combining and
708 ConstantSDNode *C = isConstOrConstSplat(Op.getOperand(1));
709 if (C && !C->isAllOnesValue()) {
710 if (NewMask.isSubsetOf(C->getAPIntValue())) {
711 // We're flipping all demanded bits. Flip the undemanded bits too.
712 SDValue New = TLO.DAG.getNOT(dl, Op.getOperand(0), Op.getValueType());
713 return TLO.CombineTo(Op, New);
715 // If we can't turn this into a 'not', try to shrink the constant.
716 if (ShrinkDemandedConstant(Op, NewMask, TLO))
720 Known = std::move(KnownOut);
724 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, Known, TLO, Depth+1))
726 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, Known2, TLO, Depth+1))
728 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
729 assert(!Known2.hasConflict() && "Bits known to be one AND zero?");
731 // If the operands are constants, see if we can simplify them.
732 if (ShrinkDemandedConstant(Op, NewMask, TLO))
735 // Only known if known in both the LHS and RHS.
736 Known.One &= Known2.One;
737 Known.Zero &= Known2.Zero;
740 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, Known, TLO, Depth+1))
742 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, Known2, TLO, Depth+1))
744 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
745 assert(!Known2.hasConflict() && "Bits known to be one AND zero?");
747 // If the operands are constants, see if we can simplify them.
748 if (ShrinkDemandedConstant(Op, NewMask, TLO))
751 // Only known if known in both the LHS and RHS.
752 Known.One &= Known2.One;
753 Known.Zero &= Known2.Zero;
756 SDValue Op0 = Op.getOperand(0);
757 SDValue Op1 = Op.getOperand(1);
758 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
759 // If (1) we only need the sign-bit, (2) the setcc operands are the same
760 // width as the setcc result, and (3) the result of a setcc conforms to 0 or
761 // -1, we may be able to bypass the setcc.
762 if (NewMask.isSignMask() && Op0.getScalarValueSizeInBits() == BitWidth &&
763 getBooleanContents(Op.getValueType()) ==
764 BooleanContent::ZeroOrNegativeOneBooleanContent) {
765 // If we're testing X < 0, then this compare isn't needed - just use X!
766 // FIXME: We're limiting to integer types here, but this should also work
767 // if we don't care about FP signed-zero. The use of SETLT with FP means
768 // that we don't care about NaNs.
769 if (CC == ISD::SETLT && Op1.getValueType().isInteger() &&
770 (isNullConstant(Op1) || ISD::isBuildVectorAllZeros(Op1.getNode())))
771 return TLO.CombineTo(Op, Op0);
773 // TODO: Should we check for other forms of sign-bit comparisons?
774 // Examples: X <= -1, X >= 0
776 if (getBooleanContents(Op0.getValueType()) ==
777 TargetLowering::ZeroOrOneBooleanContent &&
779 Known.Zero.setBitsFrom(1);
783 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
784 unsigned ShAmt = SA->getZExtValue();
785 SDValue InOp = Op.getOperand(0);
787 // If the shift count is an invalid immediate, don't do anything.
788 if (ShAmt >= BitWidth)
791 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
792 // single shift. We can do this if the bottom bits (which are shifted
793 // out) are never demanded.
794 if (InOp.getOpcode() == ISD::SRL &&
795 isa<ConstantSDNode>(InOp.getOperand(1))) {
796 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
797 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
798 unsigned Opc = ISD::SHL;
806 TLO.DAG.getConstant(Diff, dl, Op.getOperand(1).getValueType());
807 EVT VT = Op.getValueType();
808 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
809 InOp.getOperand(0), NewSA));
813 if (SimplifyDemandedBits(InOp, NewMask.lshr(ShAmt), Known, TLO, Depth+1))
816 // Convert (shl (anyext x, c)) to (anyext (shl x, c)) if the high bits
817 // are not demanded. This will likely allow the anyext to be folded away.
818 if (InOp.getNode()->getOpcode() == ISD::ANY_EXTEND) {
819 SDValue InnerOp = InOp.getOperand(0);
820 EVT InnerVT = InnerOp.getValueType();
821 unsigned InnerBits = InnerVT.getSizeInBits();
822 if (ShAmt < InnerBits && NewMask.getActiveBits() <= InnerBits &&
823 isTypeDesirableForOp(ISD::SHL, InnerVT)) {
824 EVT ShTy = getShiftAmountTy(InnerVT, DL);
825 if (!APInt(BitWidth, ShAmt).isIntN(ShTy.getSizeInBits()))
828 TLO.DAG.getNode(ISD::SHL, dl, InnerVT, InnerOp,
829 TLO.DAG.getConstant(ShAmt, dl, ShTy));
832 TLO.DAG.getNode(ISD::ANY_EXTEND, dl, Op.getValueType(),
835 // Repeat the SHL optimization above in cases where an extension
836 // intervenes: (shl (anyext (shr x, c1)), c2) to
837 // (shl (anyext x), c2-c1). This requires that the bottom c1 bits
838 // aren't demanded (as above) and that the shifted upper c1 bits of
839 // x aren't demanded.
840 if (InOp.hasOneUse() &&
841 InnerOp.getOpcode() == ISD::SRL &&
842 InnerOp.hasOneUse() &&
843 isa<ConstantSDNode>(InnerOp.getOperand(1))) {
844 unsigned InnerShAmt = cast<ConstantSDNode>(InnerOp.getOperand(1))
846 if (InnerShAmt < ShAmt &&
847 InnerShAmt < InnerBits &&
848 NewMask.getActiveBits() <= (InnerBits - InnerShAmt + ShAmt) &&
849 NewMask.countTrailingZeros() >= ShAmt) {
851 TLO.DAG.getConstant(ShAmt - InnerShAmt, dl,
852 Op.getOperand(1).getValueType());
853 EVT VT = Op.getValueType();
854 SDValue NewExt = TLO.DAG.getNode(ISD::ANY_EXTEND, dl, VT,
855 InnerOp.getOperand(0));
856 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl, VT,
862 Known.Zero <<= SA->getZExtValue();
863 Known.One <<= SA->getZExtValue();
864 // low bits known zero.
865 Known.Zero.setLowBits(SA->getZExtValue());
869 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
870 EVT VT = Op.getValueType();
871 unsigned ShAmt = SA->getZExtValue();
872 unsigned VTSize = VT.getSizeInBits();
873 SDValue InOp = Op.getOperand(0);
875 // If the shift count is an invalid immediate, don't do anything.
876 if (ShAmt >= BitWidth)
879 APInt InDemandedMask = (NewMask << ShAmt);
881 // If the shift is exact, then it does demand the low bits (and knows that
883 if (Op->getFlags().hasExact())
884 InDemandedMask.setLowBits(ShAmt);
886 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
887 // single shift. We can do this if the top bits (which are shifted out)
888 // are never demanded.
889 if (InOp.getOpcode() == ISD::SHL &&
890 isa<ConstantSDNode>(InOp.getOperand(1))) {
891 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
892 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
893 unsigned Opc = ISD::SRL;
901 TLO.DAG.getConstant(Diff, dl, Op.getOperand(1).getValueType());
902 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
903 InOp.getOperand(0), NewSA));
907 // Compute the new bits that are at the top now.
908 if (SimplifyDemandedBits(InOp, InDemandedMask, Known, TLO, Depth+1))
910 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
911 Known.Zero.lshrInPlace(ShAmt);
912 Known.One.lshrInPlace(ShAmt);
914 Known.Zero.setHighBits(ShAmt); // High bits known zero.
918 // If this is an arithmetic shift right and only the low-bit is set, we can
919 // always convert this into a logical shr, even if the shift amount is
920 // variable. The low bit of the shift cannot be an input sign bit unless
921 // the shift amount is >= the size of the datatype, which is undefined.
923 return TLO.CombineTo(Op,
924 TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(),
925 Op.getOperand(0), Op.getOperand(1)));
927 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
928 EVT VT = Op.getValueType();
929 unsigned ShAmt = SA->getZExtValue();
931 // If the shift count is an invalid immediate, don't do anything.
932 if (ShAmt >= BitWidth)
935 APInt InDemandedMask = (NewMask << ShAmt);
937 // If the shift is exact, then it does demand the low bits (and knows that
939 if (Op->getFlags().hasExact())
940 InDemandedMask.setLowBits(ShAmt);
942 // If any of the demanded bits are produced by the sign extension, we also
943 // demand the input sign bit.
944 if (NewMask.countLeadingZeros() < ShAmt)
945 InDemandedMask.setSignBit();
947 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask, Known, TLO,
950 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
951 Known.Zero.lshrInPlace(ShAmt);
952 Known.One.lshrInPlace(ShAmt);
954 // If the input sign bit is known to be zero, or if none of the top bits
955 // are demanded, turn this into an unsigned shift right.
956 if (Known.Zero[BitWidth - ShAmt - 1] ||
957 NewMask.countLeadingZeros() >= ShAmt) {
959 Flags.setExact(Op->getFlags().hasExact());
960 return TLO.CombineTo(Op,
961 TLO.DAG.getNode(ISD::SRL, dl, VT, Op.getOperand(0),
962 Op.getOperand(1), Flags));
965 int Log2 = NewMask.exactLogBase2();
967 // The bit must come from the sign.
969 TLO.DAG.getConstant(BitWidth - 1 - Log2, dl,
970 Op.getOperand(1).getValueType());
971 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT,
972 Op.getOperand(0), NewSA));
975 if (Known.One[BitWidth - ShAmt - 1])
976 // New bits are known one.
977 Known.One.setHighBits(ShAmt);
980 case ISD::SIGN_EXTEND_INREG: {
981 EVT ExVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
983 APInt MsbMask = APInt::getHighBitsSet(BitWidth, 1);
984 // If we only care about the highest bit, don't bother shifting right.
985 if (MsbMask == NewMask) {
986 unsigned ShAmt = ExVT.getScalarSizeInBits();
987 SDValue InOp = Op.getOperand(0);
988 unsigned VTBits = Op->getValueType(0).getScalarSizeInBits();
989 bool AlreadySignExtended =
990 TLO.DAG.ComputeNumSignBits(InOp) >= VTBits-ShAmt+1;
991 // However if the input is already sign extended we expect the sign
992 // extension to be dropped altogether later and do not simplify.
993 if (!AlreadySignExtended) {
994 // Compute the correct shift amount type, which must be getShiftAmountTy
995 // for scalar types after legalization.
996 EVT ShiftAmtTy = Op.getValueType();
997 if (TLO.LegalTypes() && !ShiftAmtTy.isVector())
998 ShiftAmtTy = getShiftAmountTy(ShiftAmtTy, DL);
1000 SDValue ShiftAmt = TLO.DAG.getConstant(BitWidth - ShAmt, dl,
1002 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl,
1003 Op.getValueType(), InOp,
1008 // Sign extension. Compute the demanded bits in the result that are not
1009 // present in the input.
1011 APInt::getHighBitsSet(BitWidth,
1012 BitWidth - ExVT.getScalarSizeInBits());
1014 // If none of the extended bits are demanded, eliminate the sextinreg.
1015 if ((NewBits & NewMask) == 0)
1016 return TLO.CombineTo(Op, Op.getOperand(0));
1019 APInt::getSignMask(ExVT.getScalarSizeInBits()).zext(BitWidth);
1020 APInt InputDemandedBits =
1021 APInt::getLowBitsSet(BitWidth,
1022 ExVT.getScalarSizeInBits()) &
1025 // Since the sign extended bits are demanded, we know that the sign
1027 InputDemandedBits |= InSignBit;
1029 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
1030 Known, TLO, Depth+1))
1032 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
1034 // If the sign bit of the input is known set or clear, then we know the
1035 // top bits of the result.
1037 // If the input sign bit is known zero, convert this into a zero extension.
1038 if (Known.Zero.intersects(InSignBit))
1039 return TLO.CombineTo(Op, TLO.DAG.getZeroExtendInReg(
1040 Op.getOperand(0), dl, ExVT.getScalarType()));
1042 if (Known.One.intersects(InSignBit)) { // Input sign bit known set
1043 Known.One |= NewBits;
1044 Known.Zero &= ~NewBits;
1045 } else { // Input sign bit unknown
1046 Known.Zero &= ~NewBits;
1047 Known.One &= ~NewBits;
1051 case ISD::BUILD_PAIR: {
1052 EVT HalfVT = Op.getOperand(0).getValueType();
1053 unsigned HalfBitWidth = HalfVT.getScalarSizeInBits();
1055 APInt MaskLo = NewMask.getLoBits(HalfBitWidth).trunc(HalfBitWidth);
1056 APInt MaskHi = NewMask.getHiBits(HalfBitWidth).trunc(HalfBitWidth);
1058 KnownBits KnownLo, KnownHi;
1060 if (SimplifyDemandedBits(Op.getOperand(0), MaskLo, KnownLo, TLO, Depth + 1))
1063 if (SimplifyDemandedBits(Op.getOperand(1), MaskHi, KnownHi, TLO, Depth + 1))
1066 Known.Zero = KnownLo.Zero.zext(BitWidth) |
1067 KnownHi.Zero.zext(BitWidth).shl(HalfBitWidth);
1069 Known.One = KnownLo.One.zext(BitWidth) |
1070 KnownHi.One.zext(BitWidth).shl(HalfBitWidth);
1073 case ISD::ZERO_EXTEND: {
1074 unsigned OperandBitWidth = Op.getOperand(0).getScalarValueSizeInBits();
1075 APInt InMask = NewMask.trunc(OperandBitWidth);
1077 // If none of the top bits are demanded, convert this into an any_extend.
1079 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
1080 if (!NewBits.intersects(NewMask))
1081 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1085 if (SimplifyDemandedBits(Op.getOperand(0), InMask, Known, TLO, Depth+1))
1087 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
1088 Known = Known.zext(BitWidth);
1089 Known.Zero |= NewBits;
1092 case ISD::SIGN_EXTEND: {
1093 EVT InVT = Op.getOperand(0).getValueType();
1094 unsigned InBits = InVT.getScalarSizeInBits();
1095 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
1096 APInt InSignBit = APInt::getOneBitSet(BitWidth, InBits - 1);
1097 APInt NewBits = ~InMask & NewMask;
1099 // If none of the top bits are demanded, convert this into an any_extend.
1101 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
1105 // Since some of the sign extended bits are demanded, we know that the sign
1107 APInt InDemandedBits = InMask & NewMask;
1108 InDemandedBits |= InSignBit;
1109 InDemandedBits = InDemandedBits.trunc(InBits);
1111 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, Known, TLO,
1114 Known = Known.zext(BitWidth);
1116 // If the sign bit is known zero, convert this to a zero extend.
1117 if (Known.Zero.intersects(InSignBit))
1118 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl,
1122 // If the sign bit is known one, the top bits match.
1123 if (Known.One.intersects(InSignBit)) {
1124 Known.One |= NewBits;
1125 assert((Known.Zero & NewBits) == 0);
1126 } else { // Otherwise, top bits aren't known.
1127 assert((Known.One & NewBits) == 0);
1128 assert((Known.Zero & NewBits) == 0);
1132 case ISD::ANY_EXTEND: {
1133 unsigned OperandBitWidth = Op.getOperand(0).getScalarValueSizeInBits();
1134 APInt InMask = NewMask.trunc(OperandBitWidth);
1135 if (SimplifyDemandedBits(Op.getOperand(0), InMask, Known, TLO, Depth+1))
1137 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
1138 Known = Known.zext(BitWidth);
1141 case ISD::TRUNCATE: {
1142 // Simplify the input, using demanded bit information, and compute the known
1143 // zero/one bits live out.
1144 unsigned OperandBitWidth = Op.getOperand(0).getScalarValueSizeInBits();
1145 APInt TruncMask = NewMask.zext(OperandBitWidth);
1146 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask, Known, TLO, Depth+1))
1148 Known = Known.trunc(BitWidth);
1150 // If the input is only used by this truncate, see if we can shrink it based
1151 // on the known demanded bits.
1152 if (Op.getOperand(0).getNode()->hasOneUse()) {
1153 SDValue In = Op.getOperand(0);
1154 switch (In.getOpcode()) {
1157 // Shrink SRL by a constant if none of the high bits shifted in are
1159 if (TLO.LegalTypes() &&
1160 !isTypeDesirableForOp(ISD::SRL, Op.getValueType()))
1161 // Do not turn (vt1 truncate (vt2 srl)) into (vt1 srl) if vt1 is
1164 ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1));
1167 SDValue Shift = In.getOperand(1);
1168 if (TLO.LegalTypes()) {
1169 uint64_t ShVal = ShAmt->getZExtValue();
1170 Shift = TLO.DAG.getConstant(ShVal, dl,
1171 getShiftAmountTy(Op.getValueType(), DL));
1174 if (ShAmt->getZExtValue() < BitWidth) {
1175 APInt HighBits = APInt::getHighBitsSet(OperandBitWidth,
1176 OperandBitWidth - BitWidth);
1177 HighBits.lshrInPlace(ShAmt->getZExtValue());
1178 HighBits = HighBits.trunc(BitWidth);
1180 if (!(HighBits & NewMask)) {
1181 // None of the shifted in bits are needed. Add a truncate of the
1182 // shift input, then shift it.
1183 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl,
1186 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl,
1196 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
1199 case ISD::AssertZext: {
1200 // AssertZext demands all of the high bits, plus any of the low bits
1201 // demanded by its users.
1202 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1203 APInt InMask = APInt::getLowBitsSet(BitWidth,
1204 VT.getSizeInBits());
1205 if (SimplifyDemandedBits(Op.getOperand(0), ~InMask | NewMask,
1206 Known, TLO, Depth+1))
1208 assert(!Known.hasConflict() && "Bits known to be one AND zero?");
1210 Known.Zero |= ~InMask;
1214 // If this is an FP->Int bitcast and if the sign bit is the only
1215 // thing demanded, turn this into a FGETSIGN.
1216 if (!TLO.LegalOperations() &&
1217 !Op.getValueType().isVector() &&
1218 !Op.getOperand(0).getValueType().isVector() &&
1219 NewMask == APInt::getSignMask(Op.getValueSizeInBits()) &&
1220 Op.getOperand(0).getValueType().isFloatingPoint()) {
1221 bool OpVTLegal = isOperationLegalOrCustom(ISD::FGETSIGN, Op.getValueType());
1222 bool i32Legal = isOperationLegalOrCustom(ISD::FGETSIGN, MVT::i32);
1223 if ((OpVTLegal || i32Legal) && Op.getValueType().isSimple() &&
1224 Op.getOperand(0).getValueType() != MVT::f128) {
1225 // Cannot eliminate/lower SHL for f128 yet.
1226 EVT Ty = OpVTLegal ? Op.getValueType() : MVT::i32;
1227 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
1228 // place. We expect the SHL to be eliminated by other optimizations.
1229 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, dl, Ty, Op.getOperand(0));
1230 unsigned OpVTSizeInBits = Op.getValueSizeInBits();
1231 if (!OpVTLegal && OpVTSizeInBits > 32)
1232 Sign = TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), Sign);
1233 unsigned ShVal = Op.getValueSizeInBits() - 1;
1234 SDValue ShAmt = TLO.DAG.getConstant(ShVal, dl, Op.getValueType());
1235 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl,
1244 // Add, Sub, and Mul don't demand any bits in positions beyond that
1245 // of the highest bit demanded of them.
1246 APInt LoMask = APInt::getLowBitsSet(BitWidth,
1247 BitWidth - NewMask.countLeadingZeros());
1248 if (SimplifyDemandedBits(Op.getOperand(0), LoMask, Known2, TLO, Depth+1) ||
1249 SimplifyDemandedBits(Op.getOperand(1), LoMask, Known2, TLO, Depth+1) ||
1250 // See if the operation should be performed at a smaller bit width.
1251 ShrinkDemandedOp(Op, BitWidth, NewMask, TLO)) {
1252 SDNodeFlags Flags = Op.getNode()->getFlags();
1253 if (Flags.hasNoSignedWrap() || Flags.hasNoUnsignedWrap()) {
1254 // Disable the nsw and nuw flags. We can no longer guarantee that we
1255 // won't wrap after simplification.
1256 Flags.setNoSignedWrap(false);
1257 Flags.setNoUnsignedWrap(false);
1258 SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), dl, Op.getValueType(),
1259 Op.getOperand(0), Op.getOperand(1),
1261 return TLO.CombineTo(Op, NewOp);
1268 // Just use computeKnownBits to compute output bits.
1269 TLO.DAG.computeKnownBits(Op, Known, Depth);
1273 // If we know the value of all of the demanded bits, return this as a
1275 if (NewMask.isSubsetOf(Known.Zero|Known.One)) {
1276 // Avoid folding to a constant if any OpaqueConstant is involved.
1277 const SDNode *N = Op.getNode();
1278 for (SDNodeIterator I = SDNodeIterator::begin(N),
1279 E = SDNodeIterator::end(N); I != E; ++I) {
1281 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
1285 return TLO.CombineTo(Op,
1286 TLO.DAG.getConstant(Known.One, dl, Op.getValueType()));
1292 /// Determine which of the bits specified in Mask are known to be either zero or
1293 /// one and return them in the Known.
1294 void TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
1296 const APInt &DemandedElts,
1297 const SelectionDAG &DAG,
1298 unsigned Depth) const {
1299 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1300 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1301 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1302 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1303 "Should use MaskedValueIsZero if you don't know whether Op"
1304 " is a target node!");
1308 /// This method can be implemented by targets that want to expose additional
1309 /// information about sign bits to the DAG Combiner.
1310 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
1312 const SelectionDAG &,
1313 unsigned Depth) const {
1314 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1315 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1316 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1317 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1318 "Should use ComputeNumSignBits if you don't know whether Op"
1319 " is a target node!");
1323 // FIXME: Ideally, this would use ISD::isConstantSplatVector(), but that must
1324 // work with truncating build vectors and vectors with elements of less than
1326 bool TargetLowering::isConstTrueVal(const SDNode *N) const {
1331 if (auto *CN = dyn_cast<ConstantSDNode>(N)) {
1332 CVal = CN->getAPIntValue();
1333 } else if (auto *BV = dyn_cast<BuildVectorSDNode>(N)) {
1334 auto *CN = BV->getConstantSplatNode();
1338 // If this is a truncating build vector, truncate the splat value.
1339 // Otherwise, we may fail to match the expected values below.
1340 unsigned BVEltWidth = BV->getValueType(0).getScalarSizeInBits();
1341 CVal = CN->getAPIntValue();
1342 if (BVEltWidth < CVal.getBitWidth())
1343 CVal = CVal.trunc(BVEltWidth);
1348 switch (getBooleanContents(N->getValueType(0))) {
1349 case UndefinedBooleanContent:
1351 case ZeroOrOneBooleanContent:
1353 case ZeroOrNegativeOneBooleanContent:
1354 return CVal.isAllOnesValue();
1357 llvm_unreachable("Invalid boolean contents");
1360 SDValue TargetLowering::getConstTrueVal(SelectionDAG &DAG, EVT VT,
1361 const SDLoc &DL) const {
1362 unsigned ElementWidth = VT.getScalarSizeInBits();
1364 getBooleanContents(VT) == TargetLowering::ZeroOrOneBooleanContent
1365 ? APInt(ElementWidth, 1)
1366 : APInt::getAllOnesValue(ElementWidth);
1367 return DAG.getConstant(TrueInt, DL, VT);
1370 bool TargetLowering::isConstFalseVal(const SDNode *N) const {
1374 const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N);
1376 const BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N);
1380 // Only interested in constant splats, we don't care about undef
1381 // elements in identifying boolean constants and getConstantSplatNode
1382 // returns NULL if all ops are undef;
1383 CN = BV->getConstantSplatNode();
1388 if (getBooleanContents(N->getValueType(0)) == UndefinedBooleanContent)
1389 return !CN->getAPIntValue()[0];
1391 return CN->isNullValue();
1394 bool TargetLowering::isExtendedTrueVal(const ConstantSDNode *N, EVT VT,
1399 TargetLowering::BooleanContent Cnt = getBooleanContents(VT);
1401 case TargetLowering::ZeroOrOneBooleanContent:
1402 // An extended value of 1 is always true, unless its original type is i1,
1403 // in which case it will be sign extended to -1.
1404 return (N->isOne() && !SExt) || (SExt && (N->getValueType(0) != MVT::i1));
1405 case TargetLowering::UndefinedBooleanContent:
1406 case TargetLowering::ZeroOrNegativeOneBooleanContent:
1407 return N->isAllOnesValue() && SExt;
1409 llvm_unreachable("Unexpected enumeration.");
1412 /// This helper function of SimplifySetCC tries to optimize the comparison when
1413 /// either operand of the SetCC node is a bitwise-and instruction.
1414 SDValue TargetLowering::simplifySetCCWithAnd(EVT VT, SDValue N0, SDValue N1,
1416 DAGCombinerInfo &DCI,
1417 const SDLoc &DL) const {
1418 // Match these patterns in any of their permutations:
1421 if (N1.getOpcode() == ISD::AND && N0.getOpcode() != ISD::AND)
1424 EVT OpVT = N0.getValueType();
1425 if (N0.getOpcode() != ISD::AND || !OpVT.isInteger() ||
1426 (Cond != ISD::SETEQ && Cond != ISD::SETNE))
1430 if (N0.getOperand(0) == N1) {
1431 X = N0.getOperand(1);
1432 Y = N0.getOperand(0);
1433 } else if (N0.getOperand(1) == N1) {
1434 X = N0.getOperand(0);
1435 Y = N0.getOperand(1);
1440 SelectionDAG &DAG = DCI.DAG;
1441 SDValue Zero = DAG.getConstant(0, DL, OpVT);
1442 if (DAG.isKnownToBeAPowerOfTwo(Y)) {
1443 // Simplify X & Y == Y to X & Y != 0 if Y has exactly one bit set.
1444 // Note that where Y is variable and is known to have at most one bit set
1445 // (for example, if it is Z & 1) we cannot do this; the expressions are not
1446 // equivalent when Y == 0.
1447 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
1448 if (DCI.isBeforeLegalizeOps() ||
1449 isCondCodeLegal(Cond, N0.getSimpleValueType()))
1450 return DAG.getSetCC(DL, VT, N0, Zero, Cond);
1451 } else if (N0.hasOneUse() && hasAndNotCompare(Y)) {
1452 // If the target supports an 'and-not' or 'and-complement' logic operation,
1453 // try to use that to make a comparison operation more efficient.
1454 // But don't do this transform if the mask is a single bit because there are
1455 // more efficient ways to deal with that case (for example, 'bt' on x86 or
1456 // 'rlwinm' on PPC).
1458 // Bail out if the compare operand that we want to turn into a zero is
1459 // already a zero (otherwise, infinite loop).
1460 auto *YConst = dyn_cast<ConstantSDNode>(Y);
1461 if (YConst && YConst->isNullValue())
1464 // Transform this into: ~X & Y == 0.
1465 SDValue NotX = DAG.getNOT(SDLoc(X), X, OpVT);
1466 SDValue NewAnd = DAG.getNode(ISD::AND, SDLoc(N0), OpVT, NotX, Y);
1467 return DAG.getSetCC(DL, VT, NewAnd, Zero, Cond);
1473 /// Try to simplify a setcc built with the specified operands and cc. If it is
1474 /// unable to simplify it, return a null SDValue.
1475 SDValue TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
1476 ISD::CondCode Cond, bool foldBooleans,
1477 DAGCombinerInfo &DCI,
1478 const SDLoc &dl) const {
1479 SelectionDAG &DAG = DCI.DAG;
1481 // These setcc operations always fold.
1485 case ISD::SETFALSE2: return DAG.getConstant(0, dl, VT);
1487 case ISD::SETTRUE2: {
1488 TargetLowering::BooleanContent Cnt =
1489 getBooleanContents(N0->getValueType(0));
1490 return DAG.getConstant(
1491 Cnt == TargetLowering::ZeroOrNegativeOneBooleanContent ? -1ULL : 1, dl,
1496 // Ensure that the constant occurs on the RHS, and fold constant
1498 ISD::CondCode SwappedCC = ISD::getSetCCSwappedOperands(Cond);
1499 if (isa<ConstantSDNode>(N0.getNode()) &&
1500 (DCI.isBeforeLegalizeOps() ||
1501 isCondCodeLegal(SwappedCC, N0.getSimpleValueType())))
1502 return DAG.getSetCC(dl, VT, N1, N0, SwappedCC);
1504 if (auto *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1505 const APInt &C1 = N1C->getAPIntValue();
1507 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1508 // equality comparison, then we're just comparing whether X itself is
1510 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1511 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1512 N0.getOperand(1).getOpcode() == ISD::Constant) {
1514 = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
1515 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1516 ShAmt == Log2_32(N0.getValueSizeInBits())) {
1517 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1518 // (srl (ctlz x), 5) == 0 -> X != 0
1519 // (srl (ctlz x), 5) != 1 -> X != 0
1522 // (srl (ctlz x), 5) != 0 -> X == 0
1523 // (srl (ctlz x), 5) == 1 -> X == 0
1526 SDValue Zero = DAG.getConstant(0, dl, N0.getValueType());
1527 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0),
1533 // Look through truncs that don't change the value of a ctpop.
1534 if (N0.hasOneUse() && N0.getOpcode() == ISD::TRUNCATE)
1535 CTPOP = N0.getOperand(0);
1537 if (CTPOP.hasOneUse() && CTPOP.getOpcode() == ISD::CTPOP &&
1539 N0.getValueSizeInBits() > Log2_32_Ceil(CTPOP.getValueSizeInBits()))) {
1540 EVT CTVT = CTPOP.getValueType();
1541 SDValue CTOp = CTPOP.getOperand(0);
1543 // (ctpop x) u< 2 -> (x & x-1) == 0
1544 // (ctpop x) u> 1 -> (x & x-1) != 0
1545 if ((Cond == ISD::SETULT && C1 == 2) || (Cond == ISD::SETUGT && C1 == 1)){
1546 SDValue Sub = DAG.getNode(ISD::SUB, dl, CTVT, CTOp,
1547 DAG.getConstant(1, dl, CTVT));
1548 SDValue And = DAG.getNode(ISD::AND, dl, CTVT, CTOp, Sub);
1549 ISD::CondCode CC = Cond == ISD::SETULT ? ISD::SETEQ : ISD::SETNE;
1550 return DAG.getSetCC(dl, VT, And, DAG.getConstant(0, dl, CTVT), CC);
1553 // TODO: (ctpop x) == 1 -> x && (x & x-1) == 0 iff ctpop is illegal.
1556 // (zext x) == C --> x == (trunc C)
1557 // (sext x) == C --> x == (trunc C)
1558 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1559 DCI.isBeforeLegalize() && N0->hasOneUse()) {
1560 unsigned MinBits = N0.getValueSizeInBits();
1562 bool Signed = false;
1563 if (N0->getOpcode() == ISD::ZERO_EXTEND) {
1565 MinBits = N0->getOperand(0).getValueSizeInBits();
1566 PreExt = N0->getOperand(0);
1567 } else if (N0->getOpcode() == ISD::AND) {
1568 // DAGCombine turns costly ZExts into ANDs
1569 if (auto *C = dyn_cast<ConstantSDNode>(N0->getOperand(1)))
1570 if ((C->getAPIntValue()+1).isPowerOf2()) {
1571 MinBits = C->getAPIntValue().countTrailingOnes();
1572 PreExt = N0->getOperand(0);
1574 } else if (N0->getOpcode() == ISD::SIGN_EXTEND) {
1576 MinBits = N0->getOperand(0).getValueSizeInBits();
1577 PreExt = N0->getOperand(0);
1579 } else if (auto *LN0 = dyn_cast<LoadSDNode>(N0)) {
1580 // ZEXTLOAD / SEXTLOAD
1581 if (LN0->getExtensionType() == ISD::ZEXTLOAD) {
1582 MinBits = LN0->getMemoryVT().getSizeInBits();
1584 } else if (LN0->getExtensionType() == ISD::SEXTLOAD) {
1586 MinBits = LN0->getMemoryVT().getSizeInBits();
1591 // Figure out how many bits we need to preserve this constant.
1592 unsigned ReqdBits = Signed ?
1593 C1.getBitWidth() - C1.getNumSignBits() + 1 :
1596 // Make sure we're not losing bits from the constant.
1598 MinBits < C1.getBitWidth() &&
1599 MinBits >= ReqdBits) {
1600 EVT MinVT = EVT::getIntegerVT(*DAG.getContext(), MinBits);
1601 if (isTypeDesirableForOp(ISD::SETCC, MinVT)) {
1602 // Will get folded away.
1603 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MinVT, PreExt);
1604 if (MinBits == 1 && C1 == 1)
1605 // Invert the condition.
1606 return DAG.getSetCC(dl, VT, Trunc, DAG.getConstant(0, dl, MVT::i1),
1607 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1608 SDValue C = DAG.getConstant(C1.trunc(MinBits), dl, MinVT);
1609 return DAG.getSetCC(dl, VT, Trunc, C, Cond);
1612 // If truncating the setcc operands is not desirable, we can still
1613 // simplify the expression in some cases:
1614 // setcc ([sz]ext (setcc x, y, cc)), 0, setne) -> setcc (x, y, cc)
1615 // setcc ([sz]ext (setcc x, y, cc)), 0, seteq) -> setcc (x, y, inv(cc))
1616 // setcc (zext (setcc x, y, cc)), 1, setne) -> setcc (x, y, inv(cc))
1617 // setcc (zext (setcc x, y, cc)), 1, seteq) -> setcc (x, y, cc)
1618 // setcc (sext (setcc x, y, cc)), -1, setne) -> setcc (x, y, inv(cc))
1619 // setcc (sext (setcc x, y, cc)), -1, seteq) -> setcc (x, y, cc)
1620 SDValue TopSetCC = N0->getOperand(0);
1621 unsigned N0Opc = N0->getOpcode();
1622 bool SExt = (N0Opc == ISD::SIGN_EXTEND);
1623 if (TopSetCC.getValueType() == MVT::i1 && VT == MVT::i1 &&
1624 TopSetCC.getOpcode() == ISD::SETCC &&
1625 (N0Opc == ISD::ZERO_EXTEND || N0Opc == ISD::SIGN_EXTEND) &&
1626 (isConstFalseVal(N1C) ||
1627 isExtendedTrueVal(N1C, N0->getValueType(0), SExt))) {
1629 bool Inverse = (N1C->isNullValue() && Cond == ISD::SETEQ) ||
1630 (!N1C->isNullValue() && Cond == ISD::SETNE);
1635 ISD::CondCode InvCond = ISD::getSetCCInverse(
1636 cast<CondCodeSDNode>(TopSetCC.getOperand(2))->get(),
1637 TopSetCC.getOperand(0).getValueType().isInteger());
1638 return DAG.getSetCC(dl, VT, TopSetCC.getOperand(0),
1639 TopSetCC.getOperand(1),
1646 // If the LHS is '(and load, const)', the RHS is 0,
1647 // the test is for equality or unsigned, and all 1 bits of the const are
1648 // in the same partial word, see if we can shorten the load.
1649 if (DCI.isBeforeLegalize() &&
1650 !ISD::isSignedIntSetCC(Cond) &&
1651 N0.getOpcode() == ISD::AND && C1 == 0 &&
1652 N0.getNode()->hasOneUse() &&
1653 isa<LoadSDNode>(N0.getOperand(0)) &&
1654 N0.getOperand(0).getNode()->hasOneUse() &&
1655 isa<ConstantSDNode>(N0.getOperand(1))) {
1656 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0));
1658 unsigned bestWidth = 0, bestOffset = 0;
1659 if (!Lod->isVolatile() && Lod->isUnindexed()) {
1660 unsigned origWidth = N0.getValueSizeInBits();
1661 unsigned maskWidth = origWidth;
1662 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to
1663 // 8 bits, but have to be careful...
1664 if (Lod->getExtensionType() != ISD::NON_EXTLOAD)
1665 origWidth = Lod->getMemoryVT().getSizeInBits();
1667 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
1668 for (unsigned width = origWidth / 2; width>=8; width /= 2) {
1669 APInt newMask = APInt::getLowBitsSet(maskWidth, width);
1670 for (unsigned offset=0; offset<origWidth/width; offset++) {
1671 if ((newMask & Mask) == Mask) {
1672 if (!DAG.getDataLayout().isLittleEndian())
1673 bestOffset = (origWidth/width - offset - 1) * (width/8);
1675 bestOffset = (uint64_t)offset * (width/8);
1676 bestMask = Mask.lshr(offset * (width/8) * 8);
1685 EVT newVT = EVT::getIntegerVT(*DAG.getContext(), bestWidth);
1686 if (newVT.isRound()) {
1687 EVT PtrType = Lod->getOperand(1).getValueType();
1688 SDValue Ptr = Lod->getBasePtr();
1689 if (bestOffset != 0)
1690 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(),
1691 DAG.getConstant(bestOffset, dl, PtrType));
1692 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset);
1693 SDValue NewLoad = DAG.getLoad(
1694 newVT, dl, Lod->getChain(), Ptr,
1695 Lod->getPointerInfo().getWithOffset(bestOffset), NewAlign);
1696 return DAG.getSetCC(dl, VT,
1697 DAG.getNode(ISD::AND, dl, newVT, NewLoad,
1698 DAG.getConstant(bestMask.trunc(bestWidth),
1700 DAG.getConstant(0LL, dl, newVT), Cond);
1705 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
1706 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
1707 unsigned InSize = N0.getOperand(0).getValueSizeInBits();
1709 // If the comparison constant has bits in the upper part, the
1710 // zero-extended value could never match.
1711 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
1712 C1.getBitWidth() - InSize))) {
1716 case ISD::SETEQ: return DAG.getConstant(0, dl, VT);
1719 case ISD::SETNE: return DAG.getConstant(1, dl, VT);
1722 // True if the sign bit of C1 is set.
1723 return DAG.getConstant(C1.isNegative(), dl, VT);
1726 // True if the sign bit of C1 isn't set.
1727 return DAG.getConstant(C1.isNonNegative(), dl, VT);
1733 // Otherwise, we can perform the comparison with the low bits.
1741 EVT newVT = N0.getOperand(0).getValueType();
1742 if (DCI.isBeforeLegalizeOps() ||
1743 (isOperationLegal(ISD::SETCC, newVT) &&
1744 getCondCodeAction(Cond, newVT.getSimpleVT()) == Legal)) {
1746 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), newVT);
1747 SDValue NewConst = DAG.getConstant(C1.trunc(InSize), dl, newVT);
1749 SDValue NewSetCC = DAG.getSetCC(dl, NewSetCCVT, N0.getOperand(0),
1751 return DAG.getBoolExtOrTrunc(NewSetCC, dl, VT, N0.getValueType());
1756 break; // todo, be more careful with signed comparisons
1758 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1759 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1760 EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
1761 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
1762 EVT ExtDstTy = N0.getValueType();
1763 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
1765 // If the constant doesn't fit into the number of bits for the source of
1766 // the sign extension, it is impossible for both sides to be equal.
1767 if (C1.getMinSignedBits() > ExtSrcTyBits)
1768 return DAG.getConstant(Cond == ISD::SETNE, dl, VT);
1771 EVT Op0Ty = N0.getOperand(0).getValueType();
1772 if (Op0Ty == ExtSrcTy) {
1773 ZextOp = N0.getOperand(0);
1775 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
1776 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0),
1777 DAG.getConstant(Imm, dl, Op0Ty));
1779 if (!DCI.isCalledByLegalizer())
1780 DCI.AddToWorklist(ZextOp.getNode());
1781 // Otherwise, make this a use of a zext.
1782 return DAG.getSetCC(dl, VT, ZextOp,
1783 DAG.getConstant(C1 & APInt::getLowBitsSet(
1788 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) &&
1789 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1790 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
1791 if (N0.getOpcode() == ISD::SETCC &&
1792 isTypeLegal(VT) && VT.bitsLE(N0.getValueType())) {
1793 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getAPIntValue() != 1);
1795 return DAG.getNode(ISD::TRUNCATE, dl, VT, N0);
1796 // Invert the condition.
1797 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
1798 CC = ISD::getSetCCInverse(CC,
1799 N0.getOperand(0).getValueType().isInteger());
1800 if (DCI.isBeforeLegalizeOps() ||
1801 isCondCodeLegal(CC, N0.getOperand(0).getSimpleValueType()))
1802 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC);
1805 if ((N0.getOpcode() == ISD::XOR ||
1806 (N0.getOpcode() == ISD::AND &&
1807 N0.getOperand(0).getOpcode() == ISD::XOR &&
1808 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
1809 isa<ConstantSDNode>(N0.getOperand(1)) &&
1810 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) {
1811 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
1812 // can only do this if the top bits are known zero.
1813 unsigned BitWidth = N0.getValueSizeInBits();
1814 if (DAG.MaskedValueIsZero(N0,
1815 APInt::getHighBitsSet(BitWidth,
1817 // Okay, get the un-inverted input value.
1819 if (N0.getOpcode() == ISD::XOR)
1820 Val = N0.getOperand(0);
1822 assert(N0.getOpcode() == ISD::AND &&
1823 N0.getOperand(0).getOpcode() == ISD::XOR);
1824 // ((X^1)&1)^1 -> X & 1
1825 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(),
1826 N0.getOperand(0).getOperand(0),
1830 return DAG.getSetCC(dl, VT, Val, N1,
1831 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1833 } else if (N1C->getAPIntValue() == 1 &&
1835 getBooleanContents(N0->getValueType(0)) ==
1836 ZeroOrOneBooleanContent)) {
1838 if (Op0.getOpcode() == ISD::TRUNCATE)
1839 Op0 = Op0.getOperand(0);
1841 if ((Op0.getOpcode() == ISD::XOR) &&
1842 Op0.getOperand(0).getOpcode() == ISD::SETCC &&
1843 Op0.getOperand(1).getOpcode() == ISD::SETCC) {
1844 // (xor (setcc), (setcc)) == / != 1 -> (setcc) != / == (setcc)
1845 Cond = (Cond == ISD::SETEQ) ? ISD::SETNE : ISD::SETEQ;
1846 return DAG.getSetCC(dl, VT, Op0.getOperand(0), Op0.getOperand(1),
1849 if (Op0.getOpcode() == ISD::AND &&
1850 isa<ConstantSDNode>(Op0.getOperand(1)) &&
1851 cast<ConstantSDNode>(Op0.getOperand(1))->getAPIntValue() == 1) {
1852 // If this is (X&1) == / != 1, normalize it to (X&1) != / == 0.
1853 if (Op0.getValueType().bitsGT(VT))
1854 Op0 = DAG.getNode(ISD::AND, dl, VT,
1855 DAG.getNode(ISD::TRUNCATE, dl, VT, Op0.getOperand(0)),
1856 DAG.getConstant(1, dl, VT));
1857 else if (Op0.getValueType().bitsLT(VT))
1858 Op0 = DAG.getNode(ISD::AND, dl, VT,
1859 DAG.getNode(ISD::ANY_EXTEND, dl, VT, Op0.getOperand(0)),
1860 DAG.getConstant(1, dl, VT));
1862 return DAG.getSetCC(dl, VT, Op0,
1863 DAG.getConstant(0, dl, Op0.getValueType()),
1864 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1866 if (Op0.getOpcode() == ISD::AssertZext &&
1867 cast<VTSDNode>(Op0.getOperand(1))->getVT() == MVT::i1)
1868 return DAG.getSetCC(dl, VT, Op0,
1869 DAG.getConstant(0, dl, Op0.getValueType()),
1870 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1874 APInt MinVal, MaxVal;
1875 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits();
1876 if (ISD::isSignedIntSetCC(Cond)) {
1877 MinVal = APInt::getSignedMinValue(OperandBitSize);
1878 MaxVal = APInt::getSignedMaxValue(OperandBitSize);
1880 MinVal = APInt::getMinValue(OperandBitSize);
1881 MaxVal = APInt::getMaxValue(OperandBitSize);
1884 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
1885 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
1886 if (C1 == MinVal) return DAG.getConstant(1, dl, VT); // X >= MIN --> true
1887 // X >= C0 --> X > (C0 - 1)
1889 ISD::CondCode NewCC = (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT;
1890 if ((DCI.isBeforeLegalizeOps() ||
1891 isCondCodeLegal(NewCC, VT.getSimpleVT())) &&
1892 (!N1C->isOpaque() || (N1C->isOpaque() && C.getBitWidth() <= 64 &&
1893 isLegalICmpImmediate(C.getSExtValue())))) {
1894 return DAG.getSetCC(dl, VT, N0,
1895 DAG.getConstant(C, dl, N1.getValueType()),
1900 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
1901 if (C1 == MaxVal) return DAG.getConstant(1, dl, VT); // X <= MAX --> true
1902 // X <= C0 --> X < (C0 + 1)
1904 ISD::CondCode NewCC = (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT;
1905 if ((DCI.isBeforeLegalizeOps() ||
1906 isCondCodeLegal(NewCC, VT.getSimpleVT())) &&
1907 (!N1C->isOpaque() || (N1C->isOpaque() && C.getBitWidth() <= 64 &&
1908 isLegalICmpImmediate(C.getSExtValue())))) {
1909 return DAG.getSetCC(dl, VT, N0,
1910 DAG.getConstant(C, dl, N1.getValueType()),
1915 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
1916 return DAG.getConstant(0, dl, VT); // X < MIN --> false
1917 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
1918 return DAG.getConstant(1, dl, VT); // X >= MIN --> true
1919 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
1920 return DAG.getConstant(0, dl, VT); // X > MAX --> false
1921 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
1922 return DAG.getConstant(1, dl, VT); // X <= MAX --> true
1924 // Canonicalize setgt X, Min --> setne X, Min
1925 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
1926 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
1927 // Canonicalize setlt X, Max --> setne X, Max
1928 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
1929 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
1931 // If we have setult X, 1, turn it into seteq X, 0
1932 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
1933 return DAG.getSetCC(dl, VT, N0,
1934 DAG.getConstant(MinVal, dl, N0.getValueType()),
1936 // If we have setugt X, Max-1, turn it into seteq X, Max
1937 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
1938 return DAG.getSetCC(dl, VT, N0,
1939 DAG.getConstant(MaxVal, dl, N0.getValueType()),
1942 // If we have "setcc X, C0", check to see if we can shrink the immediate
1945 // SETUGT X, SINTMAX -> SETLT X, 0
1946 if (Cond == ISD::SETUGT &&
1947 C1 == APInt::getSignedMaxValue(OperandBitSize))
1948 return DAG.getSetCC(dl, VT, N0,
1949 DAG.getConstant(0, dl, N1.getValueType()),
1952 // SETULT X, SINTMIN -> SETGT X, -1
1953 if (Cond == ISD::SETULT &&
1954 C1 == APInt::getSignedMinValue(OperandBitSize)) {
1955 SDValue ConstMinusOne =
1956 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize), dl,
1958 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT);
1961 // Fold bit comparisons when we can.
1962 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1963 (VT == N0.getValueType() ||
1964 (isTypeLegal(VT) && VT.bitsLE(N0.getValueType()))) &&
1965 N0.getOpcode() == ISD::AND) {
1966 auto &DL = DAG.getDataLayout();
1967 if (auto *AndRHS = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1968 EVT ShiftTy = DCI.isBeforeLegalize()
1970 : getShiftAmountTy(N0.getValueType(), DL);
1971 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
1972 // Perform the xform if the AND RHS is a single bit.
1973 if (AndRHS->getAPIntValue().isPowerOf2()) {
1974 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1975 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0,
1976 DAG.getConstant(AndRHS->getAPIntValue().logBase2(), dl,
1979 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getAPIntValue()) {
1980 // (X & 8) == 8 --> (X & 8) >> 3
1981 // Perform the xform if C1 is a single bit.
1982 if (C1.isPowerOf2()) {
1983 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1984 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0,
1985 DAG.getConstant(C1.logBase2(), dl,
1992 if (C1.getMinSignedBits() <= 64 &&
1993 !isLegalICmpImmediate(C1.getSExtValue())) {
1994 // (X & -256) == 256 -> (X >> 8) == 1
1995 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1996 N0.getOpcode() == ISD::AND && N0.hasOneUse()) {
1997 if (auto *AndRHS = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1998 const APInt &AndRHSC = AndRHS->getAPIntValue();
1999 if ((-AndRHSC).isPowerOf2() && (AndRHSC & C1) == C1) {
2000 unsigned ShiftBits = AndRHSC.countTrailingZeros();
2001 auto &DL = DAG.getDataLayout();
2002 EVT ShiftTy = DCI.isBeforeLegalize()
2004 : getShiftAmountTy(N0.getValueType(), DL);
2005 EVT CmpTy = N0.getValueType();
2006 SDValue Shift = DAG.getNode(ISD::SRL, dl, CmpTy, N0.getOperand(0),
2007 DAG.getConstant(ShiftBits, dl,
2009 SDValue CmpRHS = DAG.getConstant(C1.lshr(ShiftBits), dl, CmpTy);
2010 return DAG.getSetCC(dl, VT, Shift, CmpRHS, Cond);
2013 } else if (Cond == ISD::SETULT || Cond == ISD::SETUGE ||
2014 Cond == ISD::SETULE || Cond == ISD::SETUGT) {
2015 bool AdjOne = (Cond == ISD::SETULE || Cond == ISD::SETUGT);
2016 // X < 0x100000000 -> (X >> 32) < 1
2017 // X >= 0x100000000 -> (X >> 32) >= 1
2018 // X <= 0x0ffffffff -> (X >> 32) < 1
2019 // X > 0x0ffffffff -> (X >> 32) >= 1
2022 ISD::CondCode NewCond = Cond;
2024 ShiftBits = C1.countTrailingOnes();
2026 NewCond = (Cond == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2028 ShiftBits = C1.countTrailingZeros();
2030 NewC.lshrInPlace(ShiftBits);
2031 if (ShiftBits && NewC.getMinSignedBits() <= 64 &&
2032 isLegalICmpImmediate(NewC.getSExtValue())) {
2033 auto &DL = DAG.getDataLayout();
2034 EVT ShiftTy = DCI.isBeforeLegalize()
2036 : getShiftAmountTy(N0.getValueType(), DL);
2037 EVT CmpTy = N0.getValueType();
2038 SDValue Shift = DAG.getNode(ISD::SRL, dl, CmpTy, N0,
2039 DAG.getConstant(ShiftBits, dl, ShiftTy));
2040 SDValue CmpRHS = DAG.getConstant(NewC, dl, CmpTy);
2041 return DAG.getSetCC(dl, VT, Shift, CmpRHS, NewCond);
2047 if (isa<ConstantFPSDNode>(N0.getNode())) {
2048 // Constant fold or commute setcc.
2049 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl);
2050 if (O.getNode()) return O;
2051 } else if (auto *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
2052 // If the RHS of an FP comparison is a constant, simplify it away in
2054 if (CFP->getValueAPF().isNaN()) {
2055 // If an operand is known to be a nan, we can fold it.
2056 switch (ISD::getUnorderedFlavor(Cond)) {
2057 default: llvm_unreachable("Unknown flavor!");
2058 case 0: // Known false.
2059 return DAG.getConstant(0, dl, VT);
2060 case 1: // Known true.
2061 return DAG.getConstant(1, dl, VT);
2062 case 2: // Undefined.
2063 return DAG.getUNDEF(VT);
2067 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
2068 // constant if knowing that the operand is non-nan is enough. We prefer to
2069 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
2071 if (Cond == ISD::SETO || Cond == ISD::SETUO)
2072 return DAG.getSetCC(dl, VT, N0, N0, Cond);
2074 // setcc (fneg x), C -> setcc swap(pred) x, -C
2075 if (N0.getOpcode() == ISD::FNEG) {
2076 ISD::CondCode SwapCond = ISD::getSetCCSwappedOperands(Cond);
2077 if (DCI.isBeforeLegalizeOps() ||
2078 isCondCodeLegal(SwapCond, N0.getSimpleValueType())) {
2079 SDValue NegN1 = DAG.getNode(ISD::FNEG, dl, N0.getValueType(), N1);
2080 return DAG.getSetCC(dl, VT, N0.getOperand(0), NegN1, SwapCond);
2084 // If the condition is not legal, see if we can find an equivalent one
2086 if (!isCondCodeLegal(Cond, N0.getSimpleValueType())) {
2087 // If the comparison was an awkward floating-point == or != and one of
2088 // the comparison operands is infinity or negative infinity, convert the
2089 // condition to a less-awkward <= or >=.
2090 if (CFP->getValueAPF().isInfinity()) {
2091 if (CFP->getValueAPF().isNegative()) {
2092 if (Cond == ISD::SETOEQ &&
2093 isCondCodeLegal(ISD::SETOLE, N0.getSimpleValueType()))
2094 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLE);
2095 if (Cond == ISD::SETUEQ &&
2096 isCondCodeLegal(ISD::SETOLE, N0.getSimpleValueType()))
2097 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULE);
2098 if (Cond == ISD::SETUNE &&
2099 isCondCodeLegal(ISD::SETUGT, N0.getSimpleValueType()))
2100 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGT);
2101 if (Cond == ISD::SETONE &&
2102 isCondCodeLegal(ISD::SETUGT, N0.getSimpleValueType()))
2103 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGT);
2105 if (Cond == ISD::SETOEQ &&
2106 isCondCodeLegal(ISD::SETOGE, N0.getSimpleValueType()))
2107 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGE);
2108 if (Cond == ISD::SETUEQ &&
2109 isCondCodeLegal(ISD::SETOGE, N0.getSimpleValueType()))
2110 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGE);
2111 if (Cond == ISD::SETUNE &&
2112 isCondCodeLegal(ISD::SETULT, N0.getSimpleValueType()))
2113 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULT);
2114 if (Cond == ISD::SETONE &&
2115 isCondCodeLegal(ISD::SETULT, N0.getSimpleValueType()))
2116 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLT);
2123 // The sext(setcc()) => setcc() optimization relies on the appropriate
2124 // constant being emitted.
2126 switch (getBooleanContents(N0.getValueType())) {
2127 case UndefinedBooleanContent:
2128 case ZeroOrOneBooleanContent:
2129 EqVal = ISD::isTrueWhenEqual(Cond);
2131 case ZeroOrNegativeOneBooleanContent:
2132 EqVal = ISD::isTrueWhenEqual(Cond) ? -1 : 0;
2136 // We can always fold X == X for integer setcc's.
2137 if (N0.getValueType().isInteger()) {
2138 return DAG.getConstant(EqVal, dl, VT);
2140 unsigned UOF = ISD::getUnorderedFlavor(Cond);
2141 if (UOF == 2) // FP operators that are undefined on NaNs.
2142 return DAG.getConstant(EqVal, dl, VT);
2143 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
2144 return DAG.getConstant(EqVal, dl, VT);
2145 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
2146 // if it is not already.
2147 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
2148 if (NewCond != Cond && (DCI.isBeforeLegalizeOps() ||
2149 getCondCodeAction(NewCond, N0.getSimpleValueType()) == Legal))
2150 return DAG.getSetCC(dl, VT, N0, N1, NewCond);
2153 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
2154 N0.getValueType().isInteger()) {
2155 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
2156 N0.getOpcode() == ISD::XOR) {
2157 // Simplify (X+Y) == (X+Z) --> Y == Z
2158 if (N0.getOpcode() == N1.getOpcode()) {
2159 if (N0.getOperand(0) == N1.getOperand(0))
2160 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond);
2161 if (N0.getOperand(1) == N1.getOperand(1))
2162 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond);
2163 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
2164 // If X op Y == Y op X, try other combinations.
2165 if (N0.getOperand(0) == N1.getOperand(1))
2166 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0),
2168 if (N0.getOperand(1) == N1.getOperand(0))
2169 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1),
2174 // If RHS is a legal immediate value for a compare instruction, we need
2175 // to be careful about increasing register pressure needlessly.
2176 bool LegalRHSImm = false;
2178 if (auto *RHSC = dyn_cast<ConstantSDNode>(N1)) {
2179 if (auto *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2180 // Turn (X+C1) == C2 --> X == C2-C1
2181 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) {
2182 return DAG.getSetCC(dl, VT, N0.getOperand(0),
2183 DAG.getConstant(RHSC->getAPIntValue()-
2184 LHSR->getAPIntValue(),
2185 dl, N0.getValueType()), Cond);
2188 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
2189 if (N0.getOpcode() == ISD::XOR)
2190 // If we know that all of the inverted bits are zero, don't bother
2191 // performing the inversion.
2192 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
2194 DAG.getSetCC(dl, VT, N0.getOperand(0),
2195 DAG.getConstant(LHSR->getAPIntValue() ^
2196 RHSC->getAPIntValue(),
2197 dl, N0.getValueType()),
2201 // Turn (C1-X) == C2 --> X == C1-C2
2202 if (auto *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
2203 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) {
2205 DAG.getSetCC(dl, VT, N0.getOperand(1),
2206 DAG.getConstant(SUBC->getAPIntValue() -
2207 RHSC->getAPIntValue(),
2208 dl, N0.getValueType()),
2213 // Could RHSC fold directly into a compare?
2214 if (RHSC->getValueType(0).getSizeInBits() <= 64)
2215 LegalRHSImm = isLegalICmpImmediate(RHSC->getSExtValue());
2218 // Simplify (X+Z) == X --> Z == 0
2219 // Don't do this if X is an immediate that can fold into a cmp
2220 // instruction and X+Z has other uses. It could be an induction variable
2221 // chain, and the transform would increase register pressure.
2222 if (!LegalRHSImm || N0.getNode()->hasOneUse()) {
2223 if (N0.getOperand(0) == N1)
2224 return DAG.getSetCC(dl, VT, N0.getOperand(1),
2225 DAG.getConstant(0, dl, N0.getValueType()), Cond);
2226 if (N0.getOperand(1) == N1) {
2227 if (DAG.isCommutativeBinOp(N0.getOpcode()))
2228 return DAG.getSetCC(dl, VT, N0.getOperand(0),
2229 DAG.getConstant(0, dl, N0.getValueType()),
2231 if (N0.getNode()->hasOneUse()) {
2232 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
2233 auto &DL = DAG.getDataLayout();
2234 // (Z-X) == X --> Z == X<<1
2235 SDValue SH = DAG.getNode(
2236 ISD::SHL, dl, N1.getValueType(), N1,
2237 DAG.getConstant(1, dl,
2238 getShiftAmountTy(N1.getValueType(), DL)));
2239 if (!DCI.isCalledByLegalizer())
2240 DCI.AddToWorklist(SH.getNode());
2241 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond);
2247 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
2248 N1.getOpcode() == ISD::XOR) {
2249 // Simplify X == (X+Z) --> Z == 0
2250 if (N1.getOperand(0) == N0)
2251 return DAG.getSetCC(dl, VT, N1.getOperand(1),
2252 DAG.getConstant(0, dl, N1.getValueType()), Cond);
2253 if (N1.getOperand(1) == N0) {
2254 if (DAG.isCommutativeBinOp(N1.getOpcode()))
2255 return DAG.getSetCC(dl, VT, N1.getOperand(0),
2256 DAG.getConstant(0, dl, N1.getValueType()), Cond);
2257 if (N1.getNode()->hasOneUse()) {
2258 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
2259 auto &DL = DAG.getDataLayout();
2260 // X == (Z-X) --> X<<1 == Z
2261 SDValue SH = DAG.getNode(
2262 ISD::SHL, dl, N1.getValueType(), N0,
2263 DAG.getConstant(1, dl, getShiftAmountTy(N0.getValueType(), DL)));
2264 if (!DCI.isCalledByLegalizer())
2265 DCI.AddToWorklist(SH.getNode());
2266 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond);
2271 if (SDValue V = simplifySetCCWithAnd(VT, N0, N1, Cond, DCI, dl))
2275 // Fold away ALL boolean setcc's.
2277 if (N0.getValueType() == MVT::i1 && foldBooleans) {
2279 default: llvm_unreachable("Unknown integer setcc!");
2280 case ISD::SETEQ: // X == Y -> ~(X^Y)
2281 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2282 N0 = DAG.getNOT(dl, Temp, MVT::i1);
2283 if (!DCI.isCalledByLegalizer())
2284 DCI.AddToWorklist(Temp.getNode());
2286 case ISD::SETNE: // X != Y --> (X^Y)
2287 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2289 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y
2290 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y
2291 Temp = DAG.getNOT(dl, N0, MVT::i1);
2292 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp);
2293 if (!DCI.isCalledByLegalizer())
2294 DCI.AddToWorklist(Temp.getNode());
2296 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X
2297 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X
2298 Temp = DAG.getNOT(dl, N1, MVT::i1);
2299 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp);
2300 if (!DCI.isCalledByLegalizer())
2301 DCI.AddToWorklist(Temp.getNode());
2303 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y
2304 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y
2305 Temp = DAG.getNOT(dl, N0, MVT::i1);
2306 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp);
2307 if (!DCI.isCalledByLegalizer())
2308 DCI.AddToWorklist(Temp.getNode());
2310 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X
2311 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X
2312 Temp = DAG.getNOT(dl, N1, MVT::i1);
2313 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp);
2316 if (VT != MVT::i1) {
2317 if (!DCI.isCalledByLegalizer())
2318 DCI.AddToWorklist(N0.getNode());
2319 // FIXME: If running after legalize, we probably can't do this.
2320 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0);
2325 // Could not fold it.
2329 /// Returns true (and the GlobalValue and the offset) if the node is a
2330 /// GlobalAddress + offset.
2331 bool TargetLowering::isGAPlusOffset(SDNode *N, const GlobalValue *&GA,
2332 int64_t &Offset) const {
2333 if (auto *GASD = dyn_cast<GlobalAddressSDNode>(N)) {
2334 GA = GASD->getGlobal();
2335 Offset += GASD->getOffset();
2339 if (N->getOpcode() == ISD::ADD) {
2340 SDValue N1 = N->getOperand(0);
2341 SDValue N2 = N->getOperand(1);
2342 if (isGAPlusOffset(N1.getNode(), GA, Offset)) {
2343 if (auto *V = dyn_cast<ConstantSDNode>(N2)) {
2344 Offset += V->getSExtValue();
2347 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) {
2348 if (auto *V = dyn_cast<ConstantSDNode>(N1)) {
2349 Offset += V->getSExtValue();
2358 SDValue TargetLowering::PerformDAGCombine(SDNode *N,
2359 DAGCombinerInfo &DCI) const {
2360 // Default implementation: no optimization.
2364 //===----------------------------------------------------------------------===//
2365 // Inline Assembler Implementation Methods
2366 //===----------------------------------------------------------------------===//
2368 TargetLowering::ConstraintType
2369 TargetLowering::getConstraintType(StringRef Constraint) const {
2370 unsigned S = Constraint.size();
2373 switch (Constraint[0]) {
2375 case 'r': return C_RegisterClass;
2377 case 'o': // offsetable
2378 case 'V': // not offsetable
2380 case 'i': // Simple Integer or Relocatable Constant
2381 case 'n': // Simple Integer
2382 case 'E': // Floating Point Constant
2383 case 'F': // Floating Point Constant
2384 case 's': // Relocatable Constant
2385 case 'p': // Address.
2386 case 'X': // Allow ANY value.
2387 case 'I': // Target registers.
2401 if (S > 1 && Constraint[0] == '{' && Constraint[S-1] == '}') {
2402 if (S == 8 && Constraint.substr(1, 6) == "memory") // "{memory}"
2409 /// Try to replace an X constraint, which matches anything, with another that
2410 /// has more specific requirements based on the type of the corresponding
2412 const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const{
2413 if (ConstraintVT.isInteger())
2415 if (ConstraintVT.isFloatingPoint())
2416 return "f"; // works for many targets
2420 /// Lower the specified operand into the Ops vector.
2421 /// If it is invalid, don't add anything to Ops.
2422 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
2423 std::string &Constraint,
2424 std::vector<SDValue> &Ops,
2425 SelectionDAG &DAG) const {
2427 if (Constraint.length() > 1) return;
2429 char ConstraintLetter = Constraint[0];
2430 switch (ConstraintLetter) {
2432 case 'X': // Allows any operand; labels (basic block) use this.
2433 if (Op.getOpcode() == ISD::BasicBlock) {
2438 case 'i': // Simple Integer or Relocatable Constant
2439 case 'n': // Simple Integer
2440 case 's': { // Relocatable Constant
2441 // These operands are interested in values of the form (GV+C), where C may
2442 // be folded in as an offset of GV, or it may be explicitly added. Also, it
2443 // is possible and fine if either GV or C are missing.
2444 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2445 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
2447 // If we have "(add GV, C)", pull out GV/C
2448 if (Op.getOpcode() == ISD::ADD) {
2449 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2450 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
2452 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
2453 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
2461 // If we find a valid operand, map to the TargetXXX version so that the
2462 // value itself doesn't get selected.
2463 if (GA) { // Either &GV or &GV+C
2464 if (ConstraintLetter != 'n') {
2465 int64_t Offs = GA->getOffset();
2466 if (C) Offs += C->getZExtValue();
2467 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
2468 C ? SDLoc(C) : SDLoc(),
2469 Op.getValueType(), Offs));
2473 if (C) { // just C, no GV.
2474 // Simple constants are not allowed for 's'.
2475 if (ConstraintLetter != 's') {
2476 // gcc prints these as sign extended. Sign extend value to 64 bits
2477 // now; without this it would get ZExt'd later in
2478 // ScheduleDAGSDNodes::EmitNode, which is very generic.
2479 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(),
2480 SDLoc(C), MVT::i64));
2489 std::pair<unsigned, const TargetRegisterClass *>
2490 TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *RI,
2491 StringRef Constraint,
2493 if (Constraint.empty() || Constraint[0] != '{')
2494 return std::make_pair(0u, static_cast<TargetRegisterClass*>(nullptr));
2495 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
2497 // Remove the braces from around the name.
2498 StringRef RegName(Constraint.data()+1, Constraint.size()-2);
2500 std::pair<unsigned, const TargetRegisterClass*> R =
2501 std::make_pair(0u, static_cast<const TargetRegisterClass*>(nullptr));
2503 // Figure out which register class contains this reg.
2504 for (const TargetRegisterClass *RC : RI->regclasses()) {
2505 // If none of the value types for this register class are valid, we
2506 // can't use it. For example, 64-bit reg classes on 32-bit targets.
2507 if (!isLegalRC(*RI, *RC))
2510 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
2512 if (RegName.equals_lower(RI->getRegAsmName(*I))) {
2513 std::pair<unsigned, const TargetRegisterClass*> S =
2514 std::make_pair(*I, RC);
2516 // If this register class has the requested value type, return it,
2517 // otherwise keep searching and return the first class found
2518 // if no other is found which explicitly has the requested type.
2519 if (RI->isTypeLegalForClass(*RC, VT))
2530 //===----------------------------------------------------------------------===//
2531 // Constraint Selection.
2533 /// Return true of this is an input operand that is a matching constraint like
2535 bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const {
2536 assert(!ConstraintCode.empty() && "No known constraint!");
2537 return isdigit(static_cast<unsigned char>(ConstraintCode[0]));
2540 /// If this is an input matching constraint, this method returns the output
2541 /// operand it matches.
2542 unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const {
2543 assert(!ConstraintCode.empty() && "No known constraint!");
2544 return atoi(ConstraintCode.c_str());
2547 /// Split up the constraint string from the inline assembly value into the
2548 /// specific constraints and their prefixes, and also tie in the associated
2550 /// If this returns an empty vector, and if the constraint string itself
2551 /// isn't empty, there was an error parsing.
2552 TargetLowering::AsmOperandInfoVector
2553 TargetLowering::ParseConstraints(const DataLayout &DL,
2554 const TargetRegisterInfo *TRI,
2555 ImmutableCallSite CS) const {
2556 /// Information about all of the constraints.
2557 AsmOperandInfoVector ConstraintOperands;
2558 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
2559 unsigned maCount = 0; // Largest number of multiple alternative constraints.
2561 // Do a prepass over the constraints, canonicalizing them, and building up the
2562 // ConstraintOperands list.
2563 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
2564 unsigned ResNo = 0; // ResNo - The result number of the next output.
2566 for (InlineAsm::ConstraintInfo &CI : IA->ParseConstraints()) {
2567 ConstraintOperands.emplace_back(std::move(CI));
2568 AsmOperandInfo &OpInfo = ConstraintOperands.back();
2570 // Update multiple alternative constraint count.
2571 if (OpInfo.multipleAlternatives.size() > maCount)
2572 maCount = OpInfo.multipleAlternatives.size();
2574 OpInfo.ConstraintVT = MVT::Other;
2576 // Compute the value type for each operand.
2577 switch (OpInfo.Type) {
2578 case InlineAsm::isOutput:
2579 // Indirect outputs just consume an argument.
2580 if (OpInfo.isIndirect) {
2581 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
2585 // The return value of the call is this value. As such, there is no
2586 // corresponding argument.
2587 assert(!CS.getType()->isVoidTy() &&
2589 if (StructType *STy = dyn_cast<StructType>(CS.getType())) {
2590 OpInfo.ConstraintVT =
2591 getSimpleValueType(DL, STy->getElementType(ResNo));
2593 assert(ResNo == 0 && "Asm only has one result!");
2594 OpInfo.ConstraintVT = getSimpleValueType(DL, CS.getType());
2598 case InlineAsm::isInput:
2599 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
2601 case InlineAsm::isClobber:
2606 if (OpInfo.CallOperandVal) {
2607 llvm::Type *OpTy = OpInfo.CallOperandVal->getType();
2608 if (OpInfo.isIndirect) {
2609 llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
2611 report_fatal_error("Indirect operand for inline asm not a pointer!");
2612 OpTy = PtrTy->getElementType();
2615 // Look for vector wrapped in a struct. e.g. { <16 x i8> }.
2616 if (StructType *STy = dyn_cast<StructType>(OpTy))
2617 if (STy->getNumElements() == 1)
2618 OpTy = STy->getElementType(0);
2620 // If OpTy is not a single value, it may be a struct/union that we
2621 // can tile with integers.
2622 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
2623 unsigned BitSize = DL.getTypeSizeInBits(OpTy);
2632 OpInfo.ConstraintVT =
2633 MVT::getVT(IntegerType::get(OpTy->getContext(), BitSize), true);
2636 } else if (PointerType *PT = dyn_cast<PointerType>(OpTy)) {
2637 unsigned PtrSize = DL.getPointerSizeInBits(PT->getAddressSpace());
2638 OpInfo.ConstraintVT = MVT::getIntegerVT(PtrSize);
2640 OpInfo.ConstraintVT = MVT::getVT(OpTy, true);
2645 // If we have multiple alternative constraints, select the best alternative.
2646 if (!ConstraintOperands.empty()) {
2648 unsigned bestMAIndex = 0;
2649 int bestWeight = -1;
2650 // weight: -1 = invalid match, and 0 = so-so match to 5 = good match.
2653 // Compute the sums of the weights for each alternative, keeping track
2654 // of the best (highest weight) one so far.
2655 for (maIndex = 0; maIndex < maCount; ++maIndex) {
2657 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2658 cIndex != eIndex; ++cIndex) {
2659 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex];
2660 if (OpInfo.Type == InlineAsm::isClobber)
2663 // If this is an output operand with a matching input operand,
2664 // look up the matching input. If their types mismatch, e.g. one
2665 // is an integer, the other is floating point, or their sizes are
2666 // different, flag it as an maCantMatch.
2667 if (OpInfo.hasMatchingInput()) {
2668 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
2669 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
2670 if ((OpInfo.ConstraintVT.isInteger() !=
2671 Input.ConstraintVT.isInteger()) ||
2672 (OpInfo.ConstraintVT.getSizeInBits() !=
2673 Input.ConstraintVT.getSizeInBits())) {
2674 weightSum = -1; // Can't match.
2679 weight = getMultipleConstraintMatchWeight(OpInfo, maIndex);
2684 weightSum += weight;
2687 if (weightSum > bestWeight) {
2688 bestWeight = weightSum;
2689 bestMAIndex = maIndex;
2693 // Now select chosen alternative in each constraint.
2694 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2695 cIndex != eIndex; ++cIndex) {
2696 AsmOperandInfo& cInfo = ConstraintOperands[cIndex];
2697 if (cInfo.Type == InlineAsm::isClobber)
2699 cInfo.selectAlternative(bestMAIndex);
2704 // Check and hook up tied operands, choose constraint code to use.
2705 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2706 cIndex != eIndex; ++cIndex) {
2707 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex];
2709 // If this is an output operand with a matching input operand, look up the
2710 // matching input. If their types mismatch, e.g. one is an integer, the
2711 // other is floating point, or their sizes are different, flag it as an
2713 if (OpInfo.hasMatchingInput()) {
2714 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
2716 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
2717 std::pair<unsigned, const TargetRegisterClass *> MatchRC =
2718 getRegForInlineAsmConstraint(TRI, OpInfo.ConstraintCode,
2719 OpInfo.ConstraintVT);
2720 std::pair<unsigned, const TargetRegisterClass *> InputRC =
2721 getRegForInlineAsmConstraint(TRI, Input.ConstraintCode,
2722 Input.ConstraintVT);
2723 if ((OpInfo.ConstraintVT.isInteger() !=
2724 Input.ConstraintVT.isInteger()) ||
2725 (MatchRC.second != InputRC.second)) {
2726 report_fatal_error("Unsupported asm: input constraint"
2727 " with a matching output constraint of"
2728 " incompatible type!");
2734 return ConstraintOperands;
2737 /// Return an integer indicating how general CT is.
2738 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
2740 case TargetLowering::C_Other:
2741 case TargetLowering::C_Unknown:
2743 case TargetLowering::C_Register:
2745 case TargetLowering::C_RegisterClass:
2747 case TargetLowering::C_Memory:
2750 llvm_unreachable("Invalid constraint type");
2753 /// Examine constraint type and operand type and determine a weight value.
2754 /// This object must already have been set up with the operand type
2755 /// and the current alternative constraint selected.
2756 TargetLowering::ConstraintWeight
2757 TargetLowering::getMultipleConstraintMatchWeight(
2758 AsmOperandInfo &info, int maIndex) const {
2759 InlineAsm::ConstraintCodeVector *rCodes;
2760 if (maIndex >= (int)info.multipleAlternatives.size())
2761 rCodes = &info.Codes;
2763 rCodes = &info.multipleAlternatives[maIndex].Codes;
2764 ConstraintWeight BestWeight = CW_Invalid;
2766 // Loop over the options, keeping track of the most general one.
2767 for (unsigned i = 0, e = rCodes->size(); i != e; ++i) {
2768 ConstraintWeight weight =
2769 getSingleConstraintMatchWeight(info, (*rCodes)[i].c_str());
2770 if (weight > BestWeight)
2771 BestWeight = weight;
2777 /// Examine constraint type and operand type and determine a weight value.
2778 /// This object must already have been set up with the operand type
2779 /// and the current alternative constraint selected.
2780 TargetLowering::ConstraintWeight
2781 TargetLowering::getSingleConstraintMatchWeight(
2782 AsmOperandInfo &info, const char *constraint) const {
2783 ConstraintWeight weight = CW_Invalid;
2784 Value *CallOperandVal = info.CallOperandVal;
2785 // If we don't have a value, we can't do a match,
2786 // but allow it at the lowest weight.
2787 if (!CallOperandVal)
2789 // Look at the constraint type.
2790 switch (*constraint) {
2791 case 'i': // immediate integer.
2792 case 'n': // immediate integer with a known value.
2793 if (isa<ConstantInt>(CallOperandVal))
2794 weight = CW_Constant;
2796 case 's': // non-explicit intregal immediate.
2797 if (isa<GlobalValue>(CallOperandVal))
2798 weight = CW_Constant;
2800 case 'E': // immediate float if host format.
2801 case 'F': // immediate float.
2802 if (isa<ConstantFP>(CallOperandVal))
2803 weight = CW_Constant;
2805 case '<': // memory operand with autodecrement.
2806 case '>': // memory operand with autoincrement.
2807 case 'm': // memory operand.
2808 case 'o': // offsettable memory operand
2809 case 'V': // non-offsettable memory operand
2812 case 'r': // general register.
2813 case 'g': // general register, memory operand or immediate integer.
2814 // note: Clang converts "g" to "imr".
2815 if (CallOperandVal->getType()->isIntegerTy())
2816 weight = CW_Register;
2818 case 'X': // any operand.
2820 weight = CW_Default;
2826 /// If there are multiple different constraints that we could pick for this
2827 /// operand (e.g. "imr") try to pick the 'best' one.
2828 /// This is somewhat tricky: constraints fall into four classes:
2829 /// Other -> immediates and magic values
2830 /// Register -> one specific register
2831 /// RegisterClass -> a group of regs
2832 /// Memory -> memory
2833 /// Ideally, we would pick the most specific constraint possible: if we have
2834 /// something that fits into a register, we would pick it. The problem here
2835 /// is that if we have something that could either be in a register or in
2836 /// memory that use of the register could cause selection of *other*
2837 /// operands to fail: they might only succeed if we pick memory. Because of
2838 /// this the heuristic we use is:
2840 /// 1) If there is an 'other' constraint, and if the operand is valid for
2841 /// that constraint, use it. This makes us take advantage of 'i'
2842 /// constraints when available.
2843 /// 2) Otherwise, pick the most general constraint present. This prefers
2844 /// 'm' over 'r', for example.
2846 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo,
2847 const TargetLowering &TLI,
2848 SDValue Op, SelectionDAG *DAG) {
2849 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options");
2850 unsigned BestIdx = 0;
2851 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown;
2852 int BestGenerality = -1;
2854 // Loop over the options, keeping track of the most general one.
2855 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) {
2856 TargetLowering::ConstraintType CType =
2857 TLI.getConstraintType(OpInfo.Codes[i]);
2859 // If this is an 'other' constraint, see if the operand is valid for it.
2860 // For example, on X86 we might have an 'rI' constraint. If the operand
2861 // is an integer in the range [0..31] we want to use I (saving a load
2862 // of a register), otherwise we must use 'r'.
2863 if (CType == TargetLowering::C_Other && Op.getNode()) {
2864 assert(OpInfo.Codes[i].size() == 1 &&
2865 "Unhandled multi-letter 'other' constraint");
2866 std::vector<SDValue> ResultOps;
2867 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i],
2869 if (!ResultOps.empty()) {
2876 // Things with matching constraints can only be registers, per gcc
2877 // documentation. This mainly affects "g" constraints.
2878 if (CType == TargetLowering::C_Memory && OpInfo.hasMatchingInput())
2881 // This constraint letter is more general than the previous one, use it.
2882 int Generality = getConstraintGenerality(CType);
2883 if (Generality > BestGenerality) {
2886 BestGenerality = Generality;
2890 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx];
2891 OpInfo.ConstraintType = BestType;
2894 /// Determines the constraint code and constraint type to use for the specific
2895 /// AsmOperandInfo, setting OpInfo.ConstraintCode and OpInfo.ConstraintType.
2896 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
2898 SelectionDAG *DAG) const {
2899 assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
2901 // Single-letter constraints ('r') are very common.
2902 if (OpInfo.Codes.size() == 1) {
2903 OpInfo.ConstraintCode = OpInfo.Codes[0];
2904 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2906 ChooseConstraint(OpInfo, *this, Op, DAG);
2909 // 'X' matches anything.
2910 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
2911 // Labels and constants are handled elsewhere ('X' is the only thing
2912 // that matches labels). For Functions, the type here is the type of
2913 // the result, which is not what we want to look at; leave them alone.
2914 Value *v = OpInfo.CallOperandVal;
2915 if (isa<BasicBlock>(v) || isa<ConstantInt>(v) || isa<Function>(v)) {
2916 OpInfo.CallOperandVal = v;
2920 // Otherwise, try to resolve it to something we know about by looking at
2921 // the actual operand type.
2922 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
2923 OpInfo.ConstraintCode = Repl;
2924 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2929 /// \brief Given an exact SDIV by a constant, create a multiplication
2930 /// with the multiplicative inverse of the constant.
2931 static SDValue BuildExactSDIV(const TargetLowering &TLI, SDValue Op1, APInt d,
2932 const SDLoc &dl, SelectionDAG &DAG,
2933 std::vector<SDNode *> &Created) {
2934 assert(d != 0 && "Division by zero!");
2936 // Shift the value upfront if it is even, so the LSB is one.
2937 unsigned ShAmt = d.countTrailingZeros();
2939 // TODO: For UDIV use SRL instead of SRA.
2941 DAG.getConstant(ShAmt, dl, TLI.getShiftAmountTy(Op1.getValueType(),
2942 DAG.getDataLayout()));
2944 Flags.setExact(true);
2945 Op1 = DAG.getNode(ISD::SRA, dl, Op1.getValueType(), Op1, Amt, Flags);
2946 Created.push_back(Op1.getNode());
2947 d.ashrInPlace(ShAmt);
2950 // Calculate the multiplicative inverse, using Newton's method.
2952 while ((t = d*xn) != 1)
2953 xn *= APInt(d.getBitWidth(), 2) - t;
2955 SDValue Op2 = DAG.getConstant(xn, dl, Op1.getValueType());
2956 SDValue Mul = DAG.getNode(ISD::MUL, dl, Op1.getValueType(), Op1, Op2);
2957 Created.push_back(Mul.getNode());
2961 SDValue TargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor,
2963 std::vector<SDNode *> *Created) const {
2964 AttributeList Attr = DAG.getMachineFunction().getFunction()->getAttributes();
2965 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2966 if (TLI.isIntDivCheap(N->getValueType(0), Attr))
2967 return SDValue(N,0); // Lower SDIV as SDIV
2971 /// \brief Given an ISD::SDIV node expressing a divide by constant,
2972 /// return a DAG expression to select that will generate the same value by
2973 /// multiplying by a magic number.
2974 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
2975 SDValue TargetLowering::BuildSDIV(SDNode *N, const APInt &Divisor,
2976 SelectionDAG &DAG, bool IsAfterLegalization,
2977 std::vector<SDNode *> *Created) const {
2978 assert(Created && "No vector to hold sdiv ops.");
2980 EVT VT = N->getValueType(0);
2983 // Check to see if we can do this.
2984 // FIXME: We should be more aggressive here.
2985 if (!isTypeLegal(VT))
2988 // If the sdiv has an 'exact' bit we can use a simpler lowering.
2989 if (N->getFlags().hasExact())
2990 return BuildExactSDIV(*this, N->getOperand(0), Divisor, dl, DAG, *Created);
2992 APInt::ms magics = Divisor.magic();
2994 // Multiply the numerator (operand 0) by the magic value
2995 // FIXME: We should support doing a MUL in a wider type
2997 if (IsAfterLegalization ? isOperationLegal(ISD::MULHS, VT) :
2998 isOperationLegalOrCustom(ISD::MULHS, VT))
2999 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0),
3000 DAG.getConstant(magics.m, dl, VT));
3001 else if (IsAfterLegalization ? isOperationLegal(ISD::SMUL_LOHI, VT) :
3002 isOperationLegalOrCustom(ISD::SMUL_LOHI, VT))
3003 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT),
3005 DAG.getConstant(magics.m, dl, VT)).getNode(), 1);
3007 return SDValue(); // No mulhs or equvialent
3008 // If d > 0 and m < 0, add the numerator
3009 if (Divisor.isStrictlyPositive() && magics.m.isNegative()) {
3010 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0));
3011 Created->push_back(Q.getNode());
3013 // If d < 0 and m > 0, subtract the numerator.
3014 if (Divisor.isNegative() && magics.m.isStrictlyPositive()) {
3015 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0));
3016 Created->push_back(Q.getNode());
3018 auto &DL = DAG.getDataLayout();
3019 // Shift right algebraic if shift value is nonzero
3022 ISD::SRA, dl, VT, Q,
3023 DAG.getConstant(magics.s, dl, getShiftAmountTy(Q.getValueType(), DL)));
3024 Created->push_back(Q.getNode());
3026 // Extract the sign bit and add it to the quotient
3028 DAG.getNode(ISD::SRL, dl, VT, Q,
3029 DAG.getConstant(VT.getScalarSizeInBits() - 1, dl,
3030 getShiftAmountTy(Q.getValueType(), DL)));
3031 Created->push_back(T.getNode());
3032 return DAG.getNode(ISD::ADD, dl, VT, Q, T);
3035 /// \brief Given an ISD::UDIV node expressing a divide by constant,
3036 /// return a DAG expression to select that will generate the same value by
3037 /// multiplying by a magic number.
3038 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
3039 SDValue TargetLowering::BuildUDIV(SDNode *N, const APInt &Divisor,
3040 SelectionDAG &DAG, bool IsAfterLegalization,
3041 std::vector<SDNode *> *Created) const {
3042 assert(Created && "No vector to hold udiv ops.");
3044 EVT VT = N->getValueType(0);
3046 auto &DL = DAG.getDataLayout();
3048 // Check to see if we can do this.
3049 // FIXME: We should be more aggressive here.
3050 if (!isTypeLegal(VT))
3053 // FIXME: We should use a narrower constant when the upper
3054 // bits are known to be zero.
3055 APInt::mu magics = Divisor.magicu();
3057 SDValue Q = N->getOperand(0);
3059 // If the divisor is even, we can avoid using the expensive fixup by shifting
3060 // the divided value upfront.
3061 if (magics.a != 0 && !Divisor[0]) {
3062 unsigned Shift = Divisor.countTrailingZeros();
3064 ISD::SRL, dl, VT, Q,
3065 DAG.getConstant(Shift, dl, getShiftAmountTy(Q.getValueType(), DL)));
3066 Created->push_back(Q.getNode());
3068 // Get magic number for the shifted divisor.
3069 magics = Divisor.lshr(Shift).magicu(Shift);
3070 assert(magics.a == 0 && "Should use cheap fixup now");
3073 // Multiply the numerator (operand 0) by the magic value
3074 // FIXME: We should support doing a MUL in a wider type
3075 if (IsAfterLegalization ? isOperationLegal(ISD::MULHU, VT) :
3076 isOperationLegalOrCustom(ISD::MULHU, VT))
3077 Q = DAG.getNode(ISD::MULHU, dl, VT, Q, DAG.getConstant(magics.m, dl, VT));
3078 else if (IsAfterLegalization ? isOperationLegal(ISD::UMUL_LOHI, VT) :
3079 isOperationLegalOrCustom(ISD::UMUL_LOHI, VT))
3080 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT), Q,
3081 DAG.getConstant(magics.m, dl, VT)).getNode(), 1);
3083 return SDValue(); // No mulhu or equivalent
3085 Created->push_back(Q.getNode());
3087 if (magics.a == 0) {
3088 assert(magics.s < Divisor.getBitWidth() &&
3089 "We shouldn't generate an undefined shift!");
3091 ISD::SRL, dl, VT, Q,
3092 DAG.getConstant(magics.s, dl, getShiftAmountTy(Q.getValueType(), DL)));
3094 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q);
3095 Created->push_back(NPQ.getNode());
3097 ISD::SRL, dl, VT, NPQ,
3098 DAG.getConstant(1, dl, getShiftAmountTy(NPQ.getValueType(), DL)));
3099 Created->push_back(NPQ.getNode());
3100 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q);
3101 Created->push_back(NPQ.getNode());
3103 ISD::SRL, dl, VT, NPQ,
3104 DAG.getConstant(magics.s - 1, dl,
3105 getShiftAmountTy(NPQ.getValueType(), DL)));
3109 bool TargetLowering::
3110 verifyReturnAddressArgumentIsConstant(SDValue Op, SelectionDAG &DAG) const {
3111 if (!isa<ConstantSDNode>(Op.getOperand(0))) {
3112 DAG.getContext()->emitError("argument to '__builtin_return_address' must "
3113 "be a constant integer");
3120 //===----------------------------------------------------------------------===//
3121 // Legalization Utilities
3122 //===----------------------------------------------------------------------===//
3124 bool TargetLowering::expandMUL_LOHI(unsigned Opcode, EVT VT, SDLoc dl,
3125 SDValue LHS, SDValue RHS,
3126 SmallVectorImpl<SDValue> &Result,
3127 EVT HiLoVT, SelectionDAG &DAG,
3128 MulExpansionKind Kind, SDValue LL,
3129 SDValue LH, SDValue RL, SDValue RH) const {
3130 assert(Opcode == ISD::MUL || Opcode == ISD::UMUL_LOHI ||
3131 Opcode == ISD::SMUL_LOHI);
3133 bool HasMULHS = (Kind == MulExpansionKind::Always) ||
3134 isOperationLegalOrCustom(ISD::MULHS, HiLoVT);
3135 bool HasMULHU = (Kind == MulExpansionKind::Always) ||
3136 isOperationLegalOrCustom(ISD::MULHU, HiLoVT);
3137 bool HasSMUL_LOHI = (Kind == MulExpansionKind::Always) ||
3138 isOperationLegalOrCustom(ISD::SMUL_LOHI, HiLoVT);
3139 bool HasUMUL_LOHI = (Kind == MulExpansionKind::Always) ||
3140 isOperationLegalOrCustom(ISD::UMUL_LOHI, HiLoVT);
3142 if (!HasMULHU && !HasMULHS && !HasUMUL_LOHI && !HasSMUL_LOHI)
3145 unsigned OuterBitSize = VT.getScalarSizeInBits();
3146 unsigned InnerBitSize = HiLoVT.getScalarSizeInBits();
3147 unsigned LHSSB = DAG.ComputeNumSignBits(LHS);
3148 unsigned RHSSB = DAG.ComputeNumSignBits(RHS);
3150 // LL, LH, RL, and RH must be either all NULL or all set to a value.
3151 assert((LL.getNode() && LH.getNode() && RL.getNode() && RH.getNode()) ||
3152 (!LL.getNode() && !LH.getNode() && !RL.getNode() && !RH.getNode()));
3154 SDVTList VTs = DAG.getVTList(HiLoVT, HiLoVT);
3155 auto MakeMUL_LOHI = [&](SDValue L, SDValue R, SDValue &Lo, SDValue &Hi,
3156 bool Signed) -> bool {
3157 if ((Signed && HasSMUL_LOHI) || (!Signed && HasUMUL_LOHI)) {
3158 Lo = DAG.getNode(Signed ? ISD::SMUL_LOHI : ISD::UMUL_LOHI, dl, VTs, L, R);
3159 Hi = SDValue(Lo.getNode(), 1);
3162 if ((Signed && HasMULHS) || (!Signed && HasMULHU)) {
3163 Lo = DAG.getNode(ISD::MUL, dl, HiLoVT, L, R);
3164 Hi = DAG.getNode(Signed ? ISD::MULHS : ISD::MULHU, dl, HiLoVT, L, R);
3172 if (!LL.getNode() && !RL.getNode() &&
3173 isOperationLegalOrCustom(ISD::TRUNCATE, HiLoVT)) {
3174 LL = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, LHS);
3175 RL = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, RHS);
3181 APInt HighMask = APInt::getHighBitsSet(OuterBitSize, InnerBitSize);
3182 if (DAG.MaskedValueIsZero(LHS, HighMask) &&
3183 DAG.MaskedValueIsZero(RHS, HighMask)) {
3184 // The inputs are both zero-extended.
3185 if (MakeMUL_LOHI(LL, RL, Lo, Hi, false)) {
3186 Result.push_back(Lo);
3187 Result.push_back(Hi);
3188 if (Opcode != ISD::MUL) {
3189 SDValue Zero = DAG.getConstant(0, dl, HiLoVT);
3190 Result.push_back(Zero);
3191 Result.push_back(Zero);
3197 if (!VT.isVector() && Opcode == ISD::MUL && LHSSB > InnerBitSize &&
3198 RHSSB > InnerBitSize) {
3199 // The input values are both sign-extended.
3200 // TODO non-MUL case?
3201 if (MakeMUL_LOHI(LL, RL, Lo, Hi, true)) {
3202 Result.push_back(Lo);
3203 Result.push_back(Hi);
3208 unsigned ShiftAmount = OuterBitSize - InnerBitSize;
3209 EVT ShiftAmountTy = getShiftAmountTy(VT, DAG.getDataLayout());
3210 if (APInt::getMaxValue(ShiftAmountTy.getSizeInBits()).ult(ShiftAmount)) {
3211 // FIXME getShiftAmountTy does not always return a sensible result when VT
3212 // is an illegal type, and so the type may be too small to fit the shift
3213 // amount. Override it with i32. The shift will have to be legalized.
3214 ShiftAmountTy = MVT::i32;
3216 SDValue Shift = DAG.getConstant(ShiftAmount, dl, ShiftAmountTy);
3218 if (!LH.getNode() && !RH.getNode() &&
3219 isOperationLegalOrCustom(ISD::SRL, VT) &&
3220 isOperationLegalOrCustom(ISD::TRUNCATE, HiLoVT)) {
3221 LH = DAG.getNode(ISD::SRL, dl, VT, LHS, Shift);
3222 LH = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, LH);
3223 RH = DAG.getNode(ISD::SRL, dl, VT, RHS, Shift);
3224 RH = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, RH);
3230 if (!MakeMUL_LOHI(LL, RL, Lo, Hi, false))
3233 Result.push_back(Lo);
3235 if (Opcode == ISD::MUL) {
3236 RH = DAG.getNode(ISD::MUL, dl, HiLoVT, LL, RH);
3237 LH = DAG.getNode(ISD::MUL, dl, HiLoVT, LH, RL);
3238 Hi = DAG.getNode(ISD::ADD, dl, HiLoVT, Hi, RH);
3239 Hi = DAG.getNode(ISD::ADD, dl, HiLoVT, Hi, LH);
3240 Result.push_back(Hi);
3244 // Compute the full width result.
3245 auto Merge = [&](SDValue Lo, SDValue Hi) -> SDValue {
3246 Lo = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Lo);
3247 Hi = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Hi);
3248 Hi = DAG.getNode(ISD::SHL, dl, VT, Hi, Shift);
3249 return DAG.getNode(ISD::OR, dl, VT, Lo, Hi);
3252 SDValue Next = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Hi);
3253 if (!MakeMUL_LOHI(LL, RH, Lo, Hi, false))
3256 // This is effectively the add part of a multiply-add of half-sized operands,
3257 // so it cannot overflow.
3258 Next = DAG.getNode(ISD::ADD, dl, VT, Next, Merge(Lo, Hi));
3260 if (!MakeMUL_LOHI(LH, RL, Lo, Hi, false))
3263 Next = DAG.getNode(ISD::ADDC, dl, DAG.getVTList(VT, MVT::Glue), Next,
3266 SDValue Carry = Next.getValue(1);
3267 Result.push_back(DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, Next));
3268 Next = DAG.getNode(ISD::SRL, dl, VT, Next, Shift);
3270 if (!MakeMUL_LOHI(LH, RH, Lo, Hi, Opcode == ISD::SMUL_LOHI))
3273 SDValue Zero = DAG.getConstant(0, dl, HiLoVT);
3274 Hi = DAG.getNode(ISD::ADDE, dl, DAG.getVTList(HiLoVT, MVT::Glue), Hi, Zero,
3276 Next = DAG.getNode(ISD::ADD, dl, VT, Next, Merge(Lo, Hi));
3278 if (Opcode == ISD::SMUL_LOHI) {
3279 SDValue NextSub = DAG.getNode(ISD::SUB, dl, VT, Next,
3280 DAG.getNode(ISD::ZERO_EXTEND, dl, VT, RL));
3281 Next = DAG.getSelectCC(dl, LH, Zero, NextSub, Next, ISD::SETLT);
3283 NextSub = DAG.getNode(ISD::SUB, dl, VT, Next,
3284 DAG.getNode(ISD::ZERO_EXTEND, dl, VT, LL));
3285 Next = DAG.getSelectCC(dl, RH, Zero, NextSub, Next, ISD::SETLT);
3288 Result.push_back(DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, Next));
3289 Next = DAG.getNode(ISD::SRL, dl, VT, Next, Shift);
3290 Result.push_back(DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, Next));
3294 bool TargetLowering::expandMUL(SDNode *N, SDValue &Lo, SDValue &Hi, EVT HiLoVT,
3295 SelectionDAG &DAG, MulExpansionKind Kind,
3296 SDValue LL, SDValue LH, SDValue RL,
3298 SmallVector<SDValue, 2> Result;
3299 bool Ok = expandMUL_LOHI(N->getOpcode(), N->getValueType(0), N,
3300 N->getOperand(0), N->getOperand(1), Result, HiLoVT,
3301 DAG, Kind, LL, LH, RL, RH);
3303 assert(Result.size() == 2);
3310 bool TargetLowering::expandFP_TO_SINT(SDNode *Node, SDValue &Result,
3311 SelectionDAG &DAG) const {
3312 EVT VT = Node->getOperand(0).getValueType();
3313 EVT NVT = Node->getValueType(0);
3314 SDLoc dl(SDValue(Node, 0));
3316 // FIXME: Only f32 to i64 conversions are supported.
3317 if (VT != MVT::f32 || NVT != MVT::i64)
3320 // Expand f32 -> i64 conversion
3321 // This algorithm comes from compiler-rt's implementation of fixsfdi:
3322 // https://github.com/llvm-mirror/compiler-rt/blob/master/lib/builtins/fixsfdi.c
3323 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(),
3324 VT.getSizeInBits());
3325 SDValue ExponentMask = DAG.getConstant(0x7F800000, dl, IntVT);
3326 SDValue ExponentLoBit = DAG.getConstant(23, dl, IntVT);
3327 SDValue Bias = DAG.getConstant(127, dl, IntVT);
3328 SDValue SignMask = DAG.getConstant(APInt::getSignMask(VT.getSizeInBits()), dl,
3330 SDValue SignLowBit = DAG.getConstant(VT.getSizeInBits() - 1, dl, IntVT);
3331 SDValue MantissaMask = DAG.getConstant(0x007FFFFF, dl, IntVT);
3333 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, IntVT, Node->getOperand(0));
3335 auto &DL = DAG.getDataLayout();
3336 SDValue ExponentBits = DAG.getNode(
3337 ISD::SRL, dl, IntVT, DAG.getNode(ISD::AND, dl, IntVT, Bits, ExponentMask),
3338 DAG.getZExtOrTrunc(ExponentLoBit, dl, getShiftAmountTy(IntVT, DL)));
3339 SDValue Exponent = DAG.getNode(ISD::SUB, dl, IntVT, ExponentBits, Bias);
3341 SDValue Sign = DAG.getNode(
3342 ISD::SRA, dl, IntVT, DAG.getNode(ISD::AND, dl, IntVT, Bits, SignMask),
3343 DAG.getZExtOrTrunc(SignLowBit, dl, getShiftAmountTy(IntVT, DL)));
3344 Sign = DAG.getSExtOrTrunc(Sign, dl, NVT);
3346 SDValue R = DAG.getNode(ISD::OR, dl, IntVT,
3347 DAG.getNode(ISD::AND, dl, IntVT, Bits, MantissaMask),
3348 DAG.getConstant(0x00800000, dl, IntVT));
3350 R = DAG.getZExtOrTrunc(R, dl, NVT);
3352 R = DAG.getSelectCC(
3353 dl, Exponent, ExponentLoBit,
3354 DAG.getNode(ISD::SHL, dl, NVT, R,
3356 DAG.getNode(ISD::SUB, dl, IntVT, Exponent, ExponentLoBit),
3357 dl, getShiftAmountTy(IntVT, DL))),
3358 DAG.getNode(ISD::SRL, dl, NVT, R,
3360 DAG.getNode(ISD::SUB, dl, IntVT, ExponentLoBit, Exponent),
3361 dl, getShiftAmountTy(IntVT, DL))),
3364 SDValue Ret = DAG.getNode(ISD::SUB, dl, NVT,
3365 DAG.getNode(ISD::XOR, dl, NVT, R, Sign),
3368 Result = DAG.getSelectCC(dl, Exponent, DAG.getConstant(0, dl, IntVT),
3369 DAG.getConstant(0, dl, NVT), Ret, ISD::SETLT);
3373 SDValue TargetLowering::scalarizeVectorLoad(LoadSDNode *LD,
3374 SelectionDAG &DAG) const {
3376 SDValue Chain = LD->getChain();
3377 SDValue BasePTR = LD->getBasePtr();
3378 EVT SrcVT = LD->getMemoryVT();
3379 ISD::LoadExtType ExtType = LD->getExtensionType();
3381 unsigned NumElem = SrcVT.getVectorNumElements();
3383 EVT SrcEltVT = SrcVT.getScalarType();
3384 EVT DstEltVT = LD->getValueType(0).getScalarType();
3386 unsigned Stride = SrcEltVT.getSizeInBits() / 8;
3387 assert(SrcEltVT.isByteSized());
3389 EVT PtrVT = BasePTR.getValueType();
3391 SmallVector<SDValue, 8> Vals;
3392 SmallVector<SDValue, 8> LoadChains;
3394 for (unsigned Idx = 0; Idx < NumElem; ++Idx) {
3395 SDValue ScalarLoad =
3396 DAG.getExtLoad(ExtType, SL, DstEltVT, Chain, BasePTR,
3397 LD->getPointerInfo().getWithOffset(Idx * Stride),
3398 SrcEltVT, MinAlign(LD->getAlignment(), Idx * Stride),
3399 LD->getMemOperand()->getFlags(), LD->getAAInfo());
3401 BasePTR = DAG.getNode(ISD::ADD, SL, PtrVT, BasePTR,
3402 DAG.getConstant(Stride, SL, PtrVT));
3404 Vals.push_back(ScalarLoad.getValue(0));
3405 LoadChains.push_back(ScalarLoad.getValue(1));
3408 SDValue NewChain = DAG.getNode(ISD::TokenFactor, SL, MVT::Other, LoadChains);
3409 SDValue Value = DAG.getBuildVector(LD->getValueType(0), SL, Vals);
3411 return DAG.getMergeValues({ Value, NewChain }, SL);
3414 // FIXME: This relies on each element having a byte size, otherwise the stride
3415 // is 0 and just overwrites the same location. ExpandStore currently expects
3416 // this broken behavior.
3417 SDValue TargetLowering::scalarizeVectorStore(StoreSDNode *ST,
3418 SelectionDAG &DAG) const {
3421 SDValue Chain = ST->getChain();
3422 SDValue BasePtr = ST->getBasePtr();
3423 SDValue Value = ST->getValue();
3424 EVT StVT = ST->getMemoryVT();
3426 // The type of the data we want to save
3427 EVT RegVT = Value.getValueType();
3428 EVT RegSclVT = RegVT.getScalarType();
3430 // The type of data as saved in memory.
3431 EVT MemSclVT = StVT.getScalarType();
3433 EVT PtrVT = BasePtr.getValueType();
3435 // Store Stride in bytes
3436 unsigned Stride = MemSclVT.getSizeInBits() / 8;
3437 EVT IdxVT = getVectorIdxTy(DAG.getDataLayout());
3438 unsigned NumElem = StVT.getVectorNumElements();
3440 // Extract each of the elements from the original vector and save them into
3441 // memory individually.
3442 SmallVector<SDValue, 8> Stores;
3443 for (unsigned Idx = 0; Idx < NumElem; ++Idx) {
3444 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, RegSclVT, Value,
3445 DAG.getConstant(Idx, SL, IdxVT));
3447 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, BasePtr,
3448 DAG.getConstant(Idx * Stride, SL, PtrVT));
3450 // This scalar TruncStore may be illegal, but we legalize it later.
3451 SDValue Store = DAG.getTruncStore(
3452 Chain, SL, Elt, Ptr, ST->getPointerInfo().getWithOffset(Idx * Stride),
3453 MemSclVT, MinAlign(ST->getAlignment(), Idx * Stride),
3454 ST->getMemOperand()->getFlags(), ST->getAAInfo());
3456 Stores.push_back(Store);
3459 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Stores);
3462 std::pair<SDValue, SDValue>
3463 TargetLowering::expandUnalignedLoad(LoadSDNode *LD, SelectionDAG &DAG) const {
3464 assert(LD->getAddressingMode() == ISD::UNINDEXED &&
3465 "unaligned indexed loads not implemented!");
3466 SDValue Chain = LD->getChain();
3467 SDValue Ptr = LD->getBasePtr();
3468 EVT VT = LD->getValueType(0);
3469 EVT LoadedVT = LD->getMemoryVT();
3471 if (VT.isFloatingPoint() || VT.isVector()) {
3472 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), LoadedVT.getSizeInBits());
3473 if (isTypeLegal(intVT) && isTypeLegal(LoadedVT)) {
3474 if (!isOperationLegalOrCustom(ISD::LOAD, intVT)) {
3475 // Scalarize the load and let the individual components be handled.
3476 SDValue Scalarized = scalarizeVectorLoad(LD, DAG);
3477 return std::make_pair(Scalarized.getValue(0), Scalarized.getValue(1));
3480 // Expand to a (misaligned) integer load of the same size,
3481 // then bitconvert to floating point or vector.
3482 SDValue newLoad = DAG.getLoad(intVT, dl, Chain, Ptr,
3483 LD->getMemOperand());
3484 SDValue Result = DAG.getNode(ISD::BITCAST, dl, LoadedVT, newLoad);
3486 Result = DAG.getNode(VT.isFloatingPoint() ? ISD::FP_EXTEND :
3487 ISD::ANY_EXTEND, dl, VT, Result);
3489 return std::make_pair(Result, newLoad.getValue(1));
3492 // Copy the value to a (aligned) stack slot using (unaligned) integer
3493 // loads and stores, then do a (aligned) load from the stack slot.
3494 MVT RegVT = getRegisterType(*DAG.getContext(), intVT);
3495 unsigned LoadedBytes = LoadedVT.getSizeInBits() / 8;
3496 unsigned RegBytes = RegVT.getSizeInBits() / 8;
3497 unsigned NumRegs = (LoadedBytes + RegBytes - 1) / RegBytes;
3499 // Make sure the stack slot is also aligned for the register type.
3500 SDValue StackBase = DAG.CreateStackTemporary(LoadedVT, RegVT);
3502 SmallVector<SDValue, 8> Stores;
3503 SDValue StackPtr = StackBase;
3504 unsigned Offset = 0;
3506 EVT PtrVT = Ptr.getValueType();
3507 EVT StackPtrVT = StackPtr.getValueType();
3509 SDValue PtrIncrement = DAG.getConstant(RegBytes, dl, PtrVT);
3510 SDValue StackPtrIncrement = DAG.getConstant(RegBytes, dl, StackPtrVT);
3512 // Do all but one copies using the full register width.
3513 for (unsigned i = 1; i < NumRegs; i++) {
3514 // Load one integer register's worth from the original location.
3515 SDValue Load = DAG.getLoad(
3516 RegVT, dl, Chain, Ptr, LD->getPointerInfo().getWithOffset(Offset),
3517 MinAlign(LD->getAlignment(), Offset), LD->getMemOperand()->getFlags(),
3519 // Follow the load with a store to the stack slot. Remember the store.
3520 Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, StackPtr,
3521 MachinePointerInfo()));
3522 // Increment the pointers.
3524 Ptr = DAG.getNode(ISD::ADD, dl, PtrVT, Ptr, PtrIncrement);
3525 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtrVT, StackPtr,
3529 // The last copy may be partial. Do an extending load.
3530 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
3531 8 * (LoadedBytes - Offset));
3533 DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Chain, Ptr,
3534 LD->getPointerInfo().getWithOffset(Offset), MemVT,
3535 MinAlign(LD->getAlignment(), Offset),
3536 LD->getMemOperand()->getFlags(), LD->getAAInfo());
3537 // Follow the load with a store to the stack slot. Remember the store.
3538 // On big-endian machines this requires a truncating store to ensure
3539 // that the bits end up in the right place.
3540 Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, StackPtr,
3541 MachinePointerInfo(), MemVT));
3543 // The order of the stores doesn't matter - say it with a TokenFactor.
3544 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
3546 // Finally, perform the original load only redirected to the stack slot.
3547 Load = DAG.getExtLoad(LD->getExtensionType(), dl, VT, TF, StackBase,
3548 MachinePointerInfo(), LoadedVT);
3550 // Callers expect a MERGE_VALUES node.
3551 return std::make_pair(Load, TF);
3554 assert(LoadedVT.isInteger() && !LoadedVT.isVector() &&
3555 "Unaligned load of unsupported type.");
3557 // Compute the new VT that is half the size of the old one. This is an
3559 unsigned NumBits = LoadedVT.getSizeInBits();
3561 NewLoadedVT = EVT::getIntegerVT(*DAG.getContext(), NumBits/2);
3564 unsigned Alignment = LD->getAlignment();
3565 unsigned IncrementSize = NumBits / 8;
3566 ISD::LoadExtType HiExtType = LD->getExtensionType();
3568 // If the original load is NON_EXTLOAD, the hi part load must be ZEXTLOAD.
3569 if (HiExtType == ISD::NON_EXTLOAD)
3570 HiExtType = ISD::ZEXTLOAD;
3572 // Load the value in two parts
3574 if (DAG.getDataLayout().isLittleEndian()) {
3575 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr, LD->getPointerInfo(),
3576 NewLoadedVT, Alignment, LD->getMemOperand()->getFlags(),
3578 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
3579 DAG.getConstant(IncrementSize, dl, Ptr.getValueType()));
3580 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr,
3581 LD->getPointerInfo().getWithOffset(IncrementSize),
3582 NewLoadedVT, MinAlign(Alignment, IncrementSize),
3583 LD->getMemOperand()->getFlags(), LD->getAAInfo());
3585 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr, LD->getPointerInfo(),
3586 NewLoadedVT, Alignment, LD->getMemOperand()->getFlags(),
3588 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
3589 DAG.getConstant(IncrementSize, dl, Ptr.getValueType()));
3590 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr,
3591 LD->getPointerInfo().getWithOffset(IncrementSize),
3592 NewLoadedVT, MinAlign(Alignment, IncrementSize),
3593 LD->getMemOperand()->getFlags(), LD->getAAInfo());
3596 // aggregate the two parts
3597 SDValue ShiftAmount =
3598 DAG.getConstant(NumBits, dl, getShiftAmountTy(Hi.getValueType(),
3599 DAG.getDataLayout()));
3600 SDValue Result = DAG.getNode(ISD::SHL, dl, VT, Hi, ShiftAmount);
3601 Result = DAG.getNode(ISD::OR, dl, VT, Result, Lo);
3603 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
3606 return std::make_pair(Result, TF);
3609 SDValue TargetLowering::expandUnalignedStore(StoreSDNode *ST,
3610 SelectionDAG &DAG) const {
3611 assert(ST->getAddressingMode() == ISD::UNINDEXED &&
3612 "unaligned indexed stores not implemented!");
3613 SDValue Chain = ST->getChain();
3614 SDValue Ptr = ST->getBasePtr();
3615 SDValue Val = ST->getValue();
3616 EVT VT = Val.getValueType();
3617 int Alignment = ST->getAlignment();
3620 if (ST->getMemoryVT().isFloatingPoint() ||
3621 ST->getMemoryVT().isVector()) {
3622 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
3623 if (isTypeLegal(intVT)) {
3624 if (!isOperationLegalOrCustom(ISD::STORE, intVT)) {
3625 // Scalarize the store and let the individual components be handled.
3626 SDValue Result = scalarizeVectorStore(ST, DAG);
3630 // Expand to a bitconvert of the value to the integer type of the
3631 // same size, then a (misaligned) int store.
3632 // FIXME: Does not handle truncating floating point stores!
3633 SDValue Result = DAG.getNode(ISD::BITCAST, dl, intVT, Val);
3634 Result = DAG.getStore(Chain, dl, Result, Ptr, ST->getPointerInfo(),
3635 Alignment, ST->getMemOperand()->getFlags());
3638 // Do a (aligned) store to a stack slot, then copy from the stack slot
3639 // to the final destination using (unaligned) integer loads and stores.
3640 EVT StoredVT = ST->getMemoryVT();
3642 getRegisterType(*DAG.getContext(),
3643 EVT::getIntegerVT(*DAG.getContext(),
3644 StoredVT.getSizeInBits()));
3645 EVT PtrVT = Ptr.getValueType();
3646 unsigned StoredBytes = StoredVT.getSizeInBits() / 8;
3647 unsigned RegBytes = RegVT.getSizeInBits() / 8;
3648 unsigned NumRegs = (StoredBytes + RegBytes - 1) / RegBytes;
3650 // Make sure the stack slot is also aligned for the register type.
3651 SDValue StackPtr = DAG.CreateStackTemporary(StoredVT, RegVT);
3653 // Perform the original store, only redirected to the stack slot.
3654 SDValue Store = DAG.getTruncStore(Chain, dl, Val, StackPtr,
3655 MachinePointerInfo(), StoredVT);
3657 EVT StackPtrVT = StackPtr.getValueType();
3659 SDValue PtrIncrement = DAG.getConstant(RegBytes, dl, PtrVT);
3660 SDValue StackPtrIncrement = DAG.getConstant(RegBytes, dl, StackPtrVT);
3661 SmallVector<SDValue, 8> Stores;
3662 unsigned Offset = 0;
3664 // Do all but one copies using the full register width.
3665 for (unsigned i = 1; i < NumRegs; i++) {
3666 // Load one integer register's worth from the stack slot.
3668 DAG.getLoad(RegVT, dl, Store, StackPtr, MachinePointerInfo());
3669 // Store it to the final location. Remember the store.
3670 Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, Ptr,
3671 ST->getPointerInfo().getWithOffset(Offset),
3672 MinAlign(ST->getAlignment(), Offset),
3673 ST->getMemOperand()->getFlags()));
3674 // Increment the pointers.
3676 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtrVT,
3677 StackPtr, StackPtrIncrement);
3678 Ptr = DAG.getNode(ISD::ADD, dl, PtrVT, Ptr, PtrIncrement);
3681 // The last store may be partial. Do a truncating store. On big-endian
3682 // machines this requires an extending load from the stack slot to ensure
3683 // that the bits are in the right place.
3684 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
3685 8 * (StoredBytes - Offset));
3687 // Load from the stack slot.
3688 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Store, StackPtr,
3689 MachinePointerInfo(), MemVT);
3692 DAG.getTruncStore(Load.getValue(1), dl, Load, Ptr,
3693 ST->getPointerInfo().getWithOffset(Offset), MemVT,
3694 MinAlign(ST->getAlignment(), Offset),
3695 ST->getMemOperand()->getFlags(), ST->getAAInfo()));
3696 // The order of the stores doesn't matter - say it with a TokenFactor.
3697 SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
3701 assert(ST->getMemoryVT().isInteger() &&
3702 !ST->getMemoryVT().isVector() &&
3703 "Unaligned store of unknown type.");
3704 // Get the half-size VT
3705 EVT NewStoredVT = ST->getMemoryVT().getHalfSizedIntegerVT(*DAG.getContext());
3706 int NumBits = NewStoredVT.getSizeInBits();
3707 int IncrementSize = NumBits / 8;
3709 // Divide the stored value in two parts.
3710 SDValue ShiftAmount =
3711 DAG.getConstant(NumBits, dl, getShiftAmountTy(Val.getValueType(),
3712 DAG.getDataLayout()));
3714 SDValue Hi = DAG.getNode(ISD::SRL, dl, VT, Val, ShiftAmount);
3716 // Store the two parts
3717 SDValue Store1, Store2;
3718 Store1 = DAG.getTruncStore(Chain, dl,
3719 DAG.getDataLayout().isLittleEndian() ? Lo : Hi,
3720 Ptr, ST->getPointerInfo(), NewStoredVT, Alignment,
3721 ST->getMemOperand()->getFlags());
3723 EVT PtrVT = Ptr.getValueType();
3724 Ptr = DAG.getNode(ISD::ADD, dl, PtrVT, Ptr,
3725 DAG.getConstant(IncrementSize, dl, PtrVT));
3726 Alignment = MinAlign(Alignment, IncrementSize);
3727 Store2 = DAG.getTruncStore(
3728 Chain, dl, DAG.getDataLayout().isLittleEndian() ? Hi : Lo, Ptr,
3729 ST->getPointerInfo().getWithOffset(IncrementSize), NewStoredVT, Alignment,
3730 ST->getMemOperand()->getFlags(), ST->getAAInfo());
3733 DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Store1, Store2);
3738 TargetLowering::IncrementMemoryAddress(SDValue Addr, SDValue Mask,
3739 const SDLoc &DL, EVT DataVT,
3741 bool IsCompressedMemory) const {
3743 EVT AddrVT = Addr.getValueType();
3744 EVT MaskVT = Mask.getValueType();
3745 assert(DataVT.getVectorNumElements() == MaskVT.getVectorNumElements() &&
3746 "Incompatible types of Data and Mask");
3747 if (IsCompressedMemory) {
3748 // Incrementing the pointer according to number of '1's in the mask.
3749 EVT MaskIntVT = EVT::getIntegerVT(*DAG.getContext(), MaskVT.getSizeInBits());
3750 SDValue MaskInIntReg = DAG.getBitcast(MaskIntVT, Mask);
3751 if (MaskIntVT.getSizeInBits() < 32) {
3752 MaskInIntReg = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, MaskInIntReg);
3753 MaskIntVT = MVT::i32;
3756 // Count '1's with POPCNT.
3757 Increment = DAG.getNode(ISD::CTPOP, DL, MaskIntVT, MaskInIntReg);
3758 Increment = DAG.getZExtOrTrunc(Increment, DL, AddrVT);
3759 // Scale is an element size in bytes.
3760 SDValue Scale = DAG.getConstant(DataVT.getScalarSizeInBits() / 8, DL,
3762 Increment = DAG.getNode(ISD::MUL, DL, AddrVT, Increment, Scale);
3764 Increment = DAG.getConstant(DataVT.getSizeInBits() / 8, DL, AddrVT);
3766 return DAG.getNode(ISD::ADD, DL, AddrVT, Addr, Increment);
3769 static SDValue clampDynamicVectorIndex(SelectionDAG &DAG,
3773 if (isa<ConstantSDNode>(Idx))
3776 EVT IdxVT = Idx.getValueType();
3777 unsigned NElts = VecVT.getVectorNumElements();
3778 if (isPowerOf2_32(NElts)) {
3779 APInt Imm = APInt::getLowBitsSet(IdxVT.getSizeInBits(),
3781 return DAG.getNode(ISD::AND, dl, IdxVT, Idx,
3782 DAG.getConstant(Imm, dl, IdxVT));
3785 return DAG.getNode(ISD::UMIN, dl, IdxVT, Idx,
3786 DAG.getConstant(NElts - 1, dl, IdxVT));
3789 SDValue TargetLowering::getVectorElementPointer(SelectionDAG &DAG,
3790 SDValue VecPtr, EVT VecVT,
3791 SDValue Index) const {
3793 // Make sure the index type is big enough to compute in.
3794 Index = DAG.getZExtOrTrunc(Index, dl, getPointerTy(DAG.getDataLayout()));
3796 EVT EltVT = VecVT.getVectorElementType();
3798 // Calculate the element offset and add it to the pointer.
3799 unsigned EltSize = EltVT.getSizeInBits() / 8; // FIXME: should be ABI size.
3800 assert(EltSize * 8 == EltVT.getSizeInBits() &&
3801 "Converting bits to bytes lost precision");
3803 Index = clampDynamicVectorIndex(DAG, Index, VecVT, dl);
3805 EVT IdxVT = Index.getValueType();
3807 Index = DAG.getNode(ISD::MUL, dl, IdxVT, Index,
3808 DAG.getConstant(EltSize, dl, IdxVT));
3809 return DAG.getNode(ISD::ADD, dl, IdxVT, Index, VecPtr);
3812 //===----------------------------------------------------------------------===//
3813 // Implementation of Emulated TLS Model
3814 //===----------------------------------------------------------------------===//
3816 SDValue TargetLowering::LowerToTLSEmulatedModel(const GlobalAddressSDNode *GA,
3817 SelectionDAG &DAG) const {
3818 // Access to address of TLS varialbe xyz is lowered to a function call:
3819 // __emutls_get_address( address of global variable named "__emutls_v.xyz" )
3820 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3821 PointerType *VoidPtrType = Type::getInt8PtrTy(*DAG.getContext());
3826 std::string NameString = ("__emutls_v." + GA->getGlobal()->getName()).str();
3827 Module *VariableModule = const_cast<Module*>(GA->getGlobal()->getParent());
3828 StringRef EmuTlsVarName(NameString);
3829 GlobalVariable *EmuTlsVar = VariableModule->getNamedGlobal(EmuTlsVarName);
3830 assert(EmuTlsVar && "Cannot find EmuTlsVar ");
3831 Entry.Node = DAG.getGlobalAddress(EmuTlsVar, dl, PtrVT);
3832 Entry.Ty = VoidPtrType;
3833 Args.push_back(Entry);
3835 SDValue EmuTlsGetAddr = DAG.getExternalSymbol("__emutls_get_address", PtrVT);
3837 TargetLowering::CallLoweringInfo CLI(DAG);
3838 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode());
3839 CLI.setLibCallee(CallingConv::C, VoidPtrType, EmuTlsGetAddr, std::move(Args));
3840 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
3842 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
3843 // At last for X86 targets, maybe good for other targets too?
3844 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
3845 MFI.setAdjustsStack(true); // Is this only for X86 target?
3846 MFI.setHasCalls(true);
3848 assert((GA->getOffset() == 0) &&
3849 "Emulated TLS must have zero offset in GlobalAddressSDNode");
3850 return CallResult.first;
3853 SDValue TargetLowering::lowerCmpEqZeroToCtlzSrl(SDValue Op,
3854 SelectionDAG &DAG) const {
3855 assert((Op->getOpcode() == ISD::SETCC) && "Input has to be a SETCC node.");
3858 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
3860 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
3861 if (C->isNullValue() && CC == ISD::SETEQ) {
3862 EVT VT = Op.getOperand(0).getValueType();
3863 SDValue Zext = Op.getOperand(0);
3864 if (VT.bitsLT(MVT::i32)) {
3866 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
3868 unsigned Log2b = Log2_32(VT.getSizeInBits());
3869 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
3870 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
3871 DAG.getConstant(Log2b, dl, MVT::i32));
3872 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);