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/Analysis.h"
18 #include "llvm/CodeGen/MachineFrameInfo.h"
19 #include "llvm/CodeGen/MachineFunction.h"
20 #include "llvm/CodeGen/MachineJumpTableInfo.h"
21 #include "llvm/CodeGen/SelectionDAG.h"
22 #include "llvm/IR/DataLayout.h"
23 #include "llvm/IR/DerivedTypes.h"
24 #include "llvm/IR/GlobalVariable.h"
25 #include "llvm/IR/LLVMContext.h"
26 #include "llvm/MC/MCAsmInfo.h"
27 #include "llvm/MC/MCExpr.h"
28 #include "llvm/Support/CommandLine.h"
29 #include "llvm/Support/ErrorHandling.h"
30 #include "llvm/Support/MathExtras.h"
31 #include "llvm/Target/TargetLoweringObjectFile.h"
32 #include "llvm/Target/TargetMachine.h"
33 #include "llvm/Target/TargetRegisterInfo.h"
34 #include "llvm/Target/TargetSubtargetInfo.h"
38 /// NOTE: The TargetMachine owns TLOF.
39 TargetLowering::TargetLowering(const TargetMachine &tm)
40 : TargetLoweringBase(tm) {}
42 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
46 /// Check whether a given call node is in tail position within its function. If
47 /// so, it sets Chain to the input chain of the tail call.
48 bool TargetLowering::isInTailCallPosition(SelectionDAG &DAG, SDNode *Node,
49 SDValue &Chain) const {
50 const Function *F = DAG.getMachineFunction().getFunction();
52 // Conservatively require the attributes of the call to match those of
53 // the return. Ignore noalias because it doesn't affect the call sequence.
54 AttributeSet CallerAttrs = F->getAttributes();
55 if (AttrBuilder(CallerAttrs, AttributeSet::ReturnIndex)
56 .removeAttribute(Attribute::NoAlias).hasAttributes())
59 // It's not safe to eliminate the sign / zero extension of the return value.
60 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
61 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
64 // Check if the only use is a function return node.
65 return isUsedByReturnOnly(Node, Chain);
68 /// \brief Set CallLoweringInfo attribute flags based on a call instruction
69 /// and called function attributes.
70 void TargetLowering::ArgListEntry::setAttributes(ImmutableCallSite *CS,
72 isSExt = CS->paramHasAttr(AttrIdx, Attribute::SExt);
73 isZExt = CS->paramHasAttr(AttrIdx, Attribute::ZExt);
74 isInReg = CS->paramHasAttr(AttrIdx, Attribute::InReg);
75 isSRet = CS->paramHasAttr(AttrIdx, Attribute::StructRet);
76 isNest = CS->paramHasAttr(AttrIdx, Attribute::Nest);
77 isByVal = CS->paramHasAttr(AttrIdx, Attribute::ByVal);
78 isInAlloca = CS->paramHasAttr(AttrIdx, Attribute::InAlloca);
79 isReturned = CS->paramHasAttr(AttrIdx, Attribute::Returned);
80 Alignment = CS->getParamAlignment(AttrIdx);
83 /// Generate a libcall taking the given operands as arguments and returning a
84 /// result of type RetVT.
85 std::pair<SDValue, SDValue>
86 TargetLowering::makeLibCall(SelectionDAG &DAG,
87 RTLIB::Libcall LC, EVT RetVT,
88 ArrayRef<SDValue> Ops,
89 bool isSigned, SDLoc dl,
91 bool isReturnValueUsed) const {
92 TargetLowering::ArgListTy Args;
93 Args.reserve(Ops.size());
95 TargetLowering::ArgListEntry Entry;
96 for (SDValue Op : Ops) {
98 Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext());
99 Entry.isSExt = shouldSignExtendTypeInLibCall(Op.getValueType(), isSigned);
100 Entry.isZExt = !shouldSignExtendTypeInLibCall(Op.getValueType(), isSigned);
101 Args.push_back(Entry);
104 if (LC == RTLIB::UNKNOWN_LIBCALL)
105 report_fatal_error("Unsupported library call operation!");
106 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
107 getPointerTy(DAG.getDataLayout()));
109 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
110 TargetLowering::CallLoweringInfo CLI(DAG);
111 bool signExtend = shouldSignExtendTypeInLibCall(RetVT, isSigned);
112 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
113 .setCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
114 .setNoReturn(doesNotReturn).setDiscardResult(!isReturnValueUsed)
115 .setSExtResult(signExtend).setZExtResult(!signExtend);
116 return LowerCallTo(CLI);
119 /// Soften the operands of a comparison. This code is shared among BR_CC,
120 /// SELECT_CC, and SETCC handlers.
121 void TargetLowering::softenSetCCOperands(SelectionDAG &DAG, EVT VT,
122 SDValue &NewLHS, SDValue &NewRHS,
123 ISD::CondCode &CCCode,
125 assert((VT == MVT::f32 || VT == MVT::f64 || VT == MVT::f128)
126 && "Unsupported setcc type!");
128 // Expand into one or more soft-fp libcall(s).
129 RTLIB::Libcall LC1 = RTLIB::UNKNOWN_LIBCALL, LC2 = RTLIB::UNKNOWN_LIBCALL;
130 bool ShouldInvertCC = false;
134 LC1 = (VT == MVT::f32) ? RTLIB::OEQ_F32 :
135 (VT == MVT::f64) ? RTLIB::OEQ_F64 : RTLIB::OEQ_F128;
139 LC1 = (VT == MVT::f32) ? RTLIB::UNE_F32 :
140 (VT == MVT::f64) ? RTLIB::UNE_F64 : RTLIB::UNE_F128;
144 LC1 = (VT == MVT::f32) ? RTLIB::OGE_F32 :
145 (VT == MVT::f64) ? RTLIB::OGE_F64 : RTLIB::OGE_F128;
149 LC1 = (VT == MVT::f32) ? RTLIB::OLT_F32 :
150 (VT == MVT::f64) ? RTLIB::OLT_F64 : RTLIB::OLT_F128;
154 LC1 = (VT == MVT::f32) ? RTLIB::OLE_F32 :
155 (VT == MVT::f64) ? RTLIB::OLE_F64 : RTLIB::OLE_F128;
159 LC1 = (VT == MVT::f32) ? RTLIB::OGT_F32 :
160 (VT == MVT::f64) ? RTLIB::OGT_F64 : RTLIB::OGT_F128;
163 LC1 = (VT == MVT::f32) ? RTLIB::UO_F32 :
164 (VT == MVT::f64) ? RTLIB::UO_F64 : RTLIB::UO_F128;
167 LC1 = (VT == MVT::f32) ? RTLIB::O_F32 :
168 (VT == MVT::f64) ? RTLIB::O_F64 : RTLIB::O_F128;
171 // SETONE = SETOLT | SETOGT
172 LC1 = (VT == MVT::f32) ? RTLIB::OLT_F32 :
173 (VT == MVT::f64) ? RTLIB::OLT_F64 : RTLIB::OLT_F128;
174 LC2 = (VT == MVT::f32) ? RTLIB::OGT_F32 :
175 (VT == MVT::f64) ? RTLIB::OGT_F64 : RTLIB::OGT_F128;
178 LC1 = (VT == MVT::f32) ? RTLIB::UO_F32 :
179 (VT == MVT::f64) ? RTLIB::UO_F64 : RTLIB::UO_F128;
180 LC2 = (VT == MVT::f32) ? RTLIB::OEQ_F32 :
181 (VT == MVT::f64) ? RTLIB::OEQ_F64 : RTLIB::OEQ_F128;
184 // Invert CC for unordered comparisons
185 ShouldInvertCC = true;
188 LC1 = (VT == MVT::f32) ? RTLIB::OGE_F32 :
189 (VT == MVT::f64) ? RTLIB::OGE_F64 : RTLIB::OGE_F128;
192 LC1 = (VT == MVT::f32) ? RTLIB::OGT_F32 :
193 (VT == MVT::f64) ? RTLIB::OGT_F64 : RTLIB::OGT_F128;
196 LC1 = (VT == MVT::f32) ? RTLIB::OLE_F32 :
197 (VT == MVT::f64) ? RTLIB::OLE_F64 : RTLIB::OLE_F128;
200 LC1 = (VT == MVT::f32) ? RTLIB::OLT_F32 :
201 (VT == MVT::f64) ? RTLIB::OLT_F64 : RTLIB::OLT_F128;
203 default: llvm_unreachable("Do not know how to soften this setcc!");
207 // Use the target specific return value for comparions lib calls.
208 EVT RetVT = getCmpLibcallReturnType();
209 SDValue Ops[2] = {NewLHS, NewRHS};
210 NewLHS = makeLibCall(DAG, LC1, RetVT, Ops, false /*sign irrelevant*/,
212 NewRHS = DAG.getConstant(0, dl, RetVT);
214 CCCode = getCmpLibcallCC(LC1);
216 CCCode = getSetCCInverse(CCCode, /*isInteger=*/true);
218 if (LC2 != RTLIB::UNKNOWN_LIBCALL) {
219 SDValue Tmp = DAG.getNode(
221 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), RetVT),
222 NewLHS, NewRHS, DAG.getCondCode(CCCode));
223 NewLHS = makeLibCall(DAG, LC2, RetVT, Ops, false/*sign irrelevant*/,
225 NewLHS = DAG.getNode(
227 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), RetVT),
228 NewLHS, NewRHS, DAG.getCondCode(getCmpLibcallCC(LC2)));
229 NewLHS = DAG.getNode(ISD::OR, dl, Tmp.getValueType(), Tmp, NewLHS);
234 /// Return the entry encoding for a jump table in the current function. The
235 /// returned value is a member of the MachineJumpTableInfo::JTEntryKind enum.
236 unsigned TargetLowering::getJumpTableEncoding() const {
237 // In non-pic modes, just use the address of a block.
238 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
239 return MachineJumpTableInfo::EK_BlockAddress;
241 // In PIC mode, if the target supports a GPRel32 directive, use it.
242 if (getTargetMachine().getMCAsmInfo()->getGPRel32Directive() != nullptr)
243 return MachineJumpTableInfo::EK_GPRel32BlockAddress;
245 // Otherwise, use a label difference.
246 return MachineJumpTableInfo::EK_LabelDifference32;
249 SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table,
250 SelectionDAG &DAG) const {
251 // If our PIC model is GP relative, use the global offset table as the base.
252 unsigned JTEncoding = getJumpTableEncoding();
254 if ((JTEncoding == MachineJumpTableInfo::EK_GPRel64BlockAddress) ||
255 (JTEncoding == MachineJumpTableInfo::EK_GPRel32BlockAddress))
256 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy(DAG.getDataLayout()));
261 /// This returns the relocation base for the given PIC jumptable, the same as
262 /// getPICJumpTableRelocBase, but as an MCExpr.
264 TargetLowering::getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
265 unsigned JTI,MCContext &Ctx) const{
266 // The normal PIC reloc base is the label at the start of the jump table.
267 return MCSymbolRefExpr::create(MF->getJTISymbol(JTI, Ctx), Ctx);
271 TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
272 // Assume that everything is safe in static mode.
273 if (getTargetMachine().getRelocationModel() == Reloc::Static)
276 // In dynamic-no-pic mode, assume that known defined values are safe.
277 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC &&
278 GA && GA->getGlobal()->isStrongDefinitionForLinker())
281 // Otherwise assume nothing is safe.
285 //===----------------------------------------------------------------------===//
286 // Optimization Methods
287 //===----------------------------------------------------------------------===//
289 /// Check to see if the specified operand of the specified instruction is a
290 /// constant integer. If so, check to see if there are any bits set in the
291 /// constant that are not demanded. If so, shrink the constant and return true.
292 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op,
293 const APInt &Demanded) {
296 // FIXME: ISD::SELECT, ISD::SELECT_CC
297 switch (Op.getOpcode()) {
302 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
303 if (!C) return false;
305 if (Op.getOpcode() == ISD::XOR &&
306 (C->getAPIntValue() | (~Demanded)).isAllOnesValue())
309 // if we can expand it to have all bits set, do it
310 if (C->getAPIntValue().intersects(~Demanded)) {
311 EVT VT = Op.getValueType();
312 SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0),
313 DAG.getConstant(Demanded &
316 return CombineTo(Op, New);
326 /// Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the casts are free.
327 /// This uses isZExtFree and ZERO_EXTEND for the widening cast, but it could be
328 /// generalized for targets with other types of implicit widening casts.
330 TargetLowering::TargetLoweringOpt::ShrinkDemandedOp(SDValue Op,
332 const APInt &Demanded,
334 assert(Op.getNumOperands() == 2 &&
335 "ShrinkDemandedOp only supports binary operators!");
336 assert(Op.getNode()->getNumValues() == 1 &&
337 "ShrinkDemandedOp only supports nodes with one result!");
339 // Early return, as this function cannot handle vector types.
340 if (Op.getValueType().isVector())
343 // Don't do this if the node has another user, which may require the
345 if (!Op.getNode()->hasOneUse())
348 // Search for the smallest integer type with free casts to and from
349 // Op's type. For expedience, just check power-of-2 integer types.
350 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
351 unsigned DemandedSize = BitWidth - Demanded.countLeadingZeros();
352 unsigned SmallVTBits = DemandedSize;
353 if (!isPowerOf2_32(SmallVTBits))
354 SmallVTBits = NextPowerOf2(SmallVTBits);
355 for (; SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) {
356 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits);
357 if (TLI.isTruncateFree(Op.getValueType(), SmallVT) &&
358 TLI.isZExtFree(SmallVT, Op.getValueType())) {
359 // We found a type with free casts.
360 SDValue X = DAG.getNode(Op.getOpcode(), dl, SmallVT,
361 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
362 Op.getNode()->getOperand(0)),
363 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
364 Op.getNode()->getOperand(1)));
365 bool NeedZext = DemandedSize > SmallVTBits;
366 SDValue Z = DAG.getNode(NeedZext ? ISD::ZERO_EXTEND : ISD::ANY_EXTEND,
367 dl, Op.getValueType(), X);
368 return CombineTo(Op, Z);
374 /// Look at Op. At this point, we know that only the DemandedMask bits of the
375 /// result of Op are ever used downstream. If we can use this information to
376 /// simplify Op, create a new simplified DAG node and return true, returning the
377 /// original and new nodes in Old and New. Otherwise, analyze the expression and
378 /// return a mask of KnownOne and KnownZero bits for the expression (used to
379 /// simplify the caller). The KnownZero/One bits may only be accurate for those
380 /// bits in the DemandedMask.
381 bool TargetLowering::SimplifyDemandedBits(SDValue Op,
382 const APInt &DemandedMask,
385 TargetLoweringOpt &TLO,
386 unsigned Depth) const {
387 unsigned BitWidth = DemandedMask.getBitWidth();
388 assert(Op.getValueType().getScalarType().getSizeInBits() == BitWidth &&
389 "Mask size mismatches value type size!");
390 APInt NewMask = DemandedMask;
392 auto &DL = TLO.DAG.getDataLayout();
394 // Don't know anything.
395 KnownZero = KnownOne = APInt(BitWidth, 0);
397 // Other users may use these bits.
398 if (!Op.getNode()->hasOneUse()) {
400 // If not at the root, Just compute the KnownZero/KnownOne bits to
401 // simplify things downstream.
402 TLO.DAG.computeKnownBits(Op, KnownZero, KnownOne, Depth);
405 // If this is the root being simplified, allow it to have multiple uses,
406 // just set the NewMask to all bits.
407 NewMask = APInt::getAllOnesValue(BitWidth);
408 } else if (DemandedMask == 0) {
409 // Not demanding any bits from Op.
410 if (Op.getOpcode() != ISD::UNDEF)
411 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType()));
413 } else if (Depth == 6) { // Limit search depth.
417 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
418 switch (Op.getOpcode()) {
420 // We know all of the bits for a constant!
421 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue();
422 KnownZero = ~KnownOne;
423 return false; // Don't fall through, will infinitely loop.
425 // If the RHS is a constant, check to see if the LHS would be zero without
426 // using the bits from the RHS. Below, we use knowledge about the RHS to
427 // simplify the LHS, here we're using information from the LHS to simplify
429 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
430 APInt LHSZero, LHSOne;
431 // Do not increment Depth here; that can cause an infinite loop.
432 TLO.DAG.computeKnownBits(Op.getOperand(0), LHSZero, LHSOne, Depth);
433 // If the LHS already has zeros where RHSC does, this and is dead.
434 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
435 return TLO.CombineTo(Op, Op.getOperand(0));
436 // If any of the set bits in the RHS are known zero on the LHS, shrink
438 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask))
442 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
443 KnownOne, TLO, Depth+1))
445 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
446 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask,
447 KnownZero2, KnownOne2, TLO, Depth+1))
449 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
451 // If all of the demanded bits are known one on one side, return the other.
452 // These bits cannot contribute to the result of the 'and'.
453 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
454 return TLO.CombineTo(Op, Op.getOperand(0));
455 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
456 return TLO.CombineTo(Op, Op.getOperand(1));
457 // If all of the demanded bits in the inputs are known zeros, return zero.
458 if ((NewMask & (KnownZero|KnownZero2)) == NewMask)
459 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, dl, Op.getValueType()));
460 // If the RHS is a constant, see if we can simplify it.
461 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask))
463 // If the operation can be done in a smaller type, do so.
464 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
467 // Output known-1 bits are only known if set in both the LHS & RHS.
468 KnownOne &= KnownOne2;
469 // Output known-0 are known to be clear if zero in either the LHS | RHS.
470 KnownZero |= KnownZero2;
473 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
474 KnownOne, TLO, Depth+1))
476 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
477 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask,
478 KnownZero2, KnownOne2, TLO, Depth+1))
480 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
482 // If all of the demanded bits are known zero on one side, return the other.
483 // These bits cannot contribute to the result of the 'or'.
484 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask))
485 return TLO.CombineTo(Op, Op.getOperand(0));
486 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask))
487 return TLO.CombineTo(Op, Op.getOperand(1));
488 // If all of the potentially set bits on one side are known to be set on
489 // the other side, just use the 'other' side.
490 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
491 return TLO.CombineTo(Op, Op.getOperand(0));
492 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
493 return TLO.CombineTo(Op, Op.getOperand(1));
494 // If the RHS is a constant, see if we can simplify it.
495 if (TLO.ShrinkDemandedConstant(Op, NewMask))
497 // If the operation can be done in a smaller type, do so.
498 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
501 // Output known-0 bits are only known if clear in both the LHS & RHS.
502 KnownZero &= KnownZero2;
503 // Output known-1 are known to be set if set in either the LHS | RHS.
504 KnownOne |= KnownOne2;
507 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
508 KnownOne, TLO, Depth+1))
510 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
511 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2,
512 KnownOne2, TLO, Depth+1))
514 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
516 // If all of the demanded bits are known zero on one side, return the other.
517 // These bits cannot contribute to the result of the 'xor'.
518 if ((KnownZero & NewMask) == NewMask)
519 return TLO.CombineTo(Op, Op.getOperand(0));
520 if ((KnownZero2 & NewMask) == NewMask)
521 return TLO.CombineTo(Op, Op.getOperand(1));
522 // If the operation can be done in a smaller type, do so.
523 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
526 // If all of the unknown bits are known to be zero on one side or the other
527 // (but not both) turn this into an *inclusive* or.
528 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
529 if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
530 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(),
534 // Output known-0 bits are known if clear or set in both the LHS & RHS.
535 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
536 // Output known-1 are known to be set if set in only one of the LHS, RHS.
537 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
539 // If all of the demanded bits on one side are known, and all of the set
540 // bits on that side are also known to be set on the other side, turn this
541 // into an AND, as we know the bits will be cleared.
542 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
543 // NB: it is okay if more bits are known than are requested
544 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known on one side
545 if (KnownOne == KnownOne2) { // set bits are the same on both sides
546 EVT VT = Op.getValueType();
547 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, dl, VT);
548 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT,
549 Op.getOperand(0), ANDC));
553 // If the RHS is a constant, see if we can simplify it.
554 // for XOR, we prefer to force bits to 1 if they will make a -1.
555 // if we can't force bits, try to shrink constant
556 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
557 APInt Expanded = C->getAPIntValue() | (~NewMask);
558 // if we can expand it to have all bits set, do it
559 if (Expanded.isAllOnesValue()) {
560 if (Expanded != C->getAPIntValue()) {
561 EVT VT = Op.getValueType();
562 SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0),
563 TLO.DAG.getConstant(Expanded, dl, VT));
564 return TLO.CombineTo(Op, New);
566 // if it already has all the bits set, nothing to change
567 // but don't shrink either!
568 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) {
573 KnownZero = KnownZeroOut;
574 KnownOne = KnownOneOut;
577 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero,
578 KnownOne, TLO, Depth+1))
580 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2,
581 KnownOne2, TLO, Depth+1))
583 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
584 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
586 // If the operands are constants, see if we can simplify them.
587 if (TLO.ShrinkDemandedConstant(Op, NewMask))
590 // Only known if known in both the LHS and RHS.
591 KnownOne &= KnownOne2;
592 KnownZero &= KnownZero2;
595 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero,
596 KnownOne, TLO, Depth+1))
598 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2,
599 KnownOne2, TLO, Depth+1))
601 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
602 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
604 // If the operands are constants, see if we can simplify them.
605 if (TLO.ShrinkDemandedConstant(Op, NewMask))
608 // Only known if known in both the LHS and RHS.
609 KnownOne &= KnownOne2;
610 KnownZero &= KnownZero2;
613 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
614 unsigned ShAmt = SA->getZExtValue();
615 SDValue InOp = Op.getOperand(0);
617 // If the shift count is an invalid immediate, don't do anything.
618 if (ShAmt >= BitWidth)
621 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
622 // single shift. We can do this if the bottom bits (which are shifted
623 // out) are never demanded.
624 if (InOp.getOpcode() == ISD::SRL &&
625 isa<ConstantSDNode>(InOp.getOperand(1))) {
626 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
627 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
628 unsigned Opc = ISD::SHL;
636 TLO.DAG.getConstant(Diff, dl, Op.getOperand(1).getValueType());
637 EVT VT = Op.getValueType();
638 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
639 InOp.getOperand(0), NewSA));
643 if (SimplifyDemandedBits(InOp, NewMask.lshr(ShAmt),
644 KnownZero, KnownOne, TLO, Depth+1))
647 // Convert (shl (anyext x, c)) to (anyext (shl x, c)) if the high bits
648 // are not demanded. This will likely allow the anyext to be folded away.
649 if (InOp.getNode()->getOpcode() == ISD::ANY_EXTEND) {
650 SDValue InnerOp = InOp.getNode()->getOperand(0);
651 EVT InnerVT = InnerOp.getValueType();
652 unsigned InnerBits = InnerVT.getSizeInBits();
653 if (ShAmt < InnerBits && NewMask.lshr(InnerBits) == 0 &&
654 isTypeDesirableForOp(ISD::SHL, InnerVT)) {
655 EVT ShTy = getShiftAmountTy(InnerVT, DL);
656 if (!APInt(BitWidth, ShAmt).isIntN(ShTy.getSizeInBits()))
659 TLO.DAG.getNode(ISD::SHL, dl, InnerVT, InnerOp,
660 TLO.DAG.getConstant(ShAmt, dl, ShTy));
663 TLO.DAG.getNode(ISD::ANY_EXTEND, dl, Op.getValueType(),
666 // Repeat the SHL optimization above in cases where an extension
667 // intervenes: (shl (anyext (shr x, c1)), c2) to
668 // (shl (anyext x), c2-c1). This requires that the bottom c1 bits
669 // aren't demanded (as above) and that the shifted upper c1 bits of
670 // x aren't demanded.
671 if (InOp.hasOneUse() &&
672 InnerOp.getOpcode() == ISD::SRL &&
673 InnerOp.hasOneUse() &&
674 isa<ConstantSDNode>(InnerOp.getOperand(1))) {
675 uint64_t InnerShAmt = cast<ConstantSDNode>(InnerOp.getOperand(1))
677 if (InnerShAmt < ShAmt &&
678 InnerShAmt < InnerBits &&
679 NewMask.lshr(InnerBits - InnerShAmt + ShAmt) == 0 &&
680 NewMask.trunc(ShAmt) == 0) {
682 TLO.DAG.getConstant(ShAmt - InnerShAmt, dl,
683 Op.getOperand(1).getValueType());
684 EVT VT = Op.getValueType();
685 SDValue NewExt = TLO.DAG.getNode(ISD::ANY_EXTEND, dl, VT,
686 InnerOp.getOperand(0));
687 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl, VT,
693 KnownZero <<= SA->getZExtValue();
694 KnownOne <<= SA->getZExtValue();
695 // low bits known zero.
696 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue());
700 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
701 EVT VT = Op.getValueType();
702 unsigned ShAmt = SA->getZExtValue();
703 unsigned VTSize = VT.getSizeInBits();
704 SDValue InOp = Op.getOperand(0);
706 // If the shift count is an invalid immediate, don't do anything.
707 if (ShAmt >= BitWidth)
710 APInt InDemandedMask = (NewMask << ShAmt);
712 // If the shift is exact, then it does demand the low bits (and knows that
714 if (cast<BinaryWithFlagsSDNode>(Op)->Flags.hasExact())
715 InDemandedMask |= APInt::getLowBitsSet(BitWidth, ShAmt);
717 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
718 // single shift. We can do this if the top bits (which are shifted out)
719 // are never demanded.
720 if (InOp.getOpcode() == ISD::SHL &&
721 isa<ConstantSDNode>(InOp.getOperand(1))) {
722 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
723 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
724 unsigned Opc = ISD::SRL;
732 TLO.DAG.getConstant(Diff, dl, Op.getOperand(1).getValueType());
733 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
734 InOp.getOperand(0), NewSA));
738 // Compute the new bits that are at the top now.
739 if (SimplifyDemandedBits(InOp, InDemandedMask,
740 KnownZero, KnownOne, TLO, Depth+1))
742 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
743 KnownZero = KnownZero.lshr(ShAmt);
744 KnownOne = KnownOne.lshr(ShAmt);
746 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
747 KnownZero |= HighBits; // High bits known zero.
751 // If this is an arithmetic shift right and only the low-bit is set, we can
752 // always convert this into a logical shr, even if the shift amount is
753 // variable. The low bit of the shift cannot be an input sign bit unless
754 // the shift amount is >= the size of the datatype, which is undefined.
756 return TLO.CombineTo(Op,
757 TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(),
758 Op.getOperand(0), Op.getOperand(1)));
760 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
761 EVT VT = Op.getValueType();
762 unsigned ShAmt = SA->getZExtValue();
764 // If the shift count is an invalid immediate, don't do anything.
765 if (ShAmt >= BitWidth)
768 APInt InDemandedMask = (NewMask << ShAmt);
770 // If the shift is exact, then it does demand the low bits (and knows that
772 if (cast<BinaryWithFlagsSDNode>(Op)->Flags.hasExact())
773 InDemandedMask |= APInt::getLowBitsSet(BitWidth, ShAmt);
775 // If any of the demanded bits are produced by the sign extension, we also
776 // demand the input sign bit.
777 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
778 if (HighBits.intersects(NewMask))
779 InDemandedMask |= APInt::getSignBit(VT.getScalarType().getSizeInBits());
781 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
782 KnownZero, KnownOne, TLO, Depth+1))
784 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
785 KnownZero = KnownZero.lshr(ShAmt);
786 KnownOne = KnownOne.lshr(ShAmt);
788 // Handle the sign bit, adjusted to where it is now in the mask.
789 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt);
791 // If the input sign bit is known to be zero, or if none of the top bits
792 // are demanded, turn this into an unsigned shift right.
793 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
795 Flags.setExact(cast<BinaryWithFlagsSDNode>(Op)->Flags.hasExact());
796 return TLO.CombineTo(Op,
797 TLO.DAG.getNode(ISD::SRL, dl, VT, Op.getOperand(0),
798 Op.getOperand(1), &Flags));
801 int Log2 = NewMask.exactLogBase2();
803 // The bit must come from the sign.
805 TLO.DAG.getConstant(BitWidth - 1 - Log2, dl,
806 Op.getOperand(1).getValueType());
807 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT,
808 Op.getOperand(0), NewSA));
811 if (KnownOne.intersects(SignBit))
812 // New bits are known one.
813 KnownOne |= HighBits;
816 case ISD::SIGN_EXTEND_INREG: {
817 EVT ExVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
819 APInt MsbMask = APInt::getHighBitsSet(BitWidth, 1);
820 // If we only care about the highest bit, don't bother shifting right.
821 if (MsbMask == NewMask) {
822 unsigned ShAmt = ExVT.getScalarType().getSizeInBits();
823 SDValue InOp = Op.getOperand(0);
824 unsigned VTBits = Op->getValueType(0).getScalarType().getSizeInBits();
825 bool AlreadySignExtended =
826 TLO.DAG.ComputeNumSignBits(InOp) >= VTBits-ShAmt+1;
827 // However if the input is already sign extended we expect the sign
828 // extension to be dropped altogether later and do not simplify.
829 if (!AlreadySignExtended) {
830 // Compute the correct shift amount type, which must be getShiftAmountTy
831 // for scalar types after legalization.
832 EVT ShiftAmtTy = Op.getValueType();
833 if (TLO.LegalTypes() && !ShiftAmtTy.isVector())
834 ShiftAmtTy = getShiftAmountTy(ShiftAmtTy, DL);
836 SDValue ShiftAmt = TLO.DAG.getConstant(BitWidth - ShAmt, dl,
838 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl,
839 Op.getValueType(), InOp,
844 // Sign extension. Compute the demanded bits in the result that are not
845 // present in the input.
847 APInt::getHighBitsSet(BitWidth,
848 BitWidth - ExVT.getScalarType().getSizeInBits());
850 // If none of the extended bits are demanded, eliminate the sextinreg.
851 if ((NewBits & NewMask) == 0)
852 return TLO.CombineTo(Op, Op.getOperand(0));
855 APInt::getSignBit(ExVT.getScalarType().getSizeInBits()).zext(BitWidth);
856 APInt InputDemandedBits =
857 APInt::getLowBitsSet(BitWidth,
858 ExVT.getScalarType().getSizeInBits()) &
861 // Since the sign extended bits are demanded, we know that the sign
863 InputDemandedBits |= InSignBit;
865 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
866 KnownZero, KnownOne, TLO, Depth+1))
868 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
870 // If the sign bit of the input is known set or clear, then we know the
871 // top bits of the result.
873 // If the input sign bit is known zero, convert this into a zero extension.
874 if (KnownZero.intersects(InSignBit))
875 return TLO.CombineTo(Op,
876 TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,ExVT));
878 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
880 KnownZero &= ~NewBits;
881 } else { // Input sign bit unknown
882 KnownZero &= ~NewBits;
883 KnownOne &= ~NewBits;
887 case ISD::BUILD_PAIR: {
888 EVT HalfVT = Op.getOperand(0).getValueType();
889 unsigned HalfBitWidth = HalfVT.getScalarSizeInBits();
891 APInt MaskLo = NewMask.getLoBits(HalfBitWidth).trunc(HalfBitWidth);
892 APInt MaskHi = NewMask.getHiBits(HalfBitWidth).trunc(HalfBitWidth);
894 APInt KnownZeroLo, KnownOneLo;
895 APInt KnownZeroHi, KnownOneHi;
897 if (SimplifyDemandedBits(Op.getOperand(0), MaskLo, KnownZeroLo,
898 KnownOneLo, TLO, Depth + 1))
901 if (SimplifyDemandedBits(Op.getOperand(1), MaskHi, KnownZeroHi,
902 KnownOneHi, TLO, Depth + 1))
905 KnownZero = KnownZeroLo.zext(BitWidth) |
906 KnownZeroHi.zext(BitWidth).shl(HalfBitWidth);
908 KnownOne = KnownOneLo.zext(BitWidth) |
909 KnownOneHi.zext(BitWidth).shl(HalfBitWidth);
912 case ISD::ZERO_EXTEND: {
913 unsigned OperandBitWidth =
914 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
915 APInt InMask = NewMask.trunc(OperandBitWidth);
917 // If none of the top bits are demanded, convert this into an any_extend.
919 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
920 if (!NewBits.intersects(NewMask))
921 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
925 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
926 KnownZero, KnownOne, TLO, Depth+1))
928 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
929 KnownZero = KnownZero.zext(BitWidth);
930 KnownOne = KnownOne.zext(BitWidth);
931 KnownZero |= NewBits;
934 case ISD::SIGN_EXTEND: {
935 EVT InVT = Op.getOperand(0).getValueType();
936 unsigned InBits = InVT.getScalarType().getSizeInBits();
937 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
938 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits);
939 APInt NewBits = ~InMask & NewMask;
941 // If none of the top bits are demanded, convert this into an any_extend.
943 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
947 // Since some of the sign extended bits are demanded, we know that the sign
949 APInt InDemandedBits = InMask & NewMask;
950 InDemandedBits |= InSignBit;
951 InDemandedBits = InDemandedBits.trunc(InBits);
953 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
954 KnownOne, TLO, Depth+1))
956 KnownZero = KnownZero.zext(BitWidth);
957 KnownOne = KnownOne.zext(BitWidth);
959 // If the sign bit is known zero, convert this to a zero extend.
960 if (KnownZero.intersects(InSignBit))
961 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl,
965 // If the sign bit is known one, the top bits match.
966 if (KnownOne.intersects(InSignBit)) {
968 assert((KnownZero & NewBits) == 0);
969 } else { // Otherwise, top bits aren't known.
970 assert((KnownOne & NewBits) == 0);
971 assert((KnownZero & NewBits) == 0);
975 case ISD::ANY_EXTEND: {
976 unsigned OperandBitWidth =
977 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
978 APInt InMask = NewMask.trunc(OperandBitWidth);
979 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
980 KnownZero, KnownOne, TLO, Depth+1))
982 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
983 KnownZero = KnownZero.zext(BitWidth);
984 KnownOne = KnownOne.zext(BitWidth);
987 case ISD::TRUNCATE: {
988 // Simplify the input, using demanded bit information, and compute the known
989 // zero/one bits live out.
990 unsigned OperandBitWidth =
991 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
992 APInt TruncMask = NewMask.zext(OperandBitWidth);
993 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask,
994 KnownZero, KnownOne, TLO, Depth+1))
996 KnownZero = KnownZero.trunc(BitWidth);
997 KnownOne = KnownOne.trunc(BitWidth);
999 // If the input is only used by this truncate, see if we can shrink it based
1000 // on the known demanded bits.
1001 if (Op.getOperand(0).getNode()->hasOneUse()) {
1002 SDValue In = Op.getOperand(0);
1003 switch (In.getOpcode()) {
1006 // Shrink SRL by a constant if none of the high bits shifted in are
1008 if (TLO.LegalTypes() &&
1009 !isTypeDesirableForOp(ISD::SRL, Op.getValueType()))
1010 // Do not turn (vt1 truncate (vt2 srl)) into (vt1 srl) if vt1 is
1013 ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1));
1016 SDValue Shift = In.getOperand(1);
1017 if (TLO.LegalTypes()) {
1018 uint64_t ShVal = ShAmt->getZExtValue();
1019 Shift = TLO.DAG.getConstant(ShVal, dl,
1020 getShiftAmountTy(Op.getValueType(), DL));
1023 APInt HighBits = APInt::getHighBitsSet(OperandBitWidth,
1024 OperandBitWidth - BitWidth);
1025 HighBits = HighBits.lshr(ShAmt->getZExtValue()).trunc(BitWidth);
1027 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) {
1028 // None of the shifted in bits are needed. Add a truncate of the
1029 // shift input, then shift it.
1030 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl,
1033 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl,
1042 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1045 case ISD::AssertZext: {
1046 // AssertZext demands all of the high bits, plus any of the low bits
1047 // demanded by its users.
1048 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1049 APInt InMask = APInt::getLowBitsSet(BitWidth,
1050 VT.getSizeInBits());
1051 if (SimplifyDemandedBits(Op.getOperand(0), ~InMask | NewMask,
1052 KnownZero, KnownOne, TLO, Depth+1))
1054 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1056 KnownZero |= ~InMask & NewMask;
1060 // If this is an FP->Int bitcast and if the sign bit is the only
1061 // thing demanded, turn this into a FGETSIGN.
1062 if (!TLO.LegalOperations() &&
1063 !Op.getValueType().isVector() &&
1064 !Op.getOperand(0).getValueType().isVector() &&
1065 NewMask == APInt::getSignBit(Op.getValueType().getSizeInBits()) &&
1066 Op.getOperand(0).getValueType().isFloatingPoint()) {
1067 bool OpVTLegal = isOperationLegalOrCustom(ISD::FGETSIGN, Op.getValueType());
1068 bool i32Legal = isOperationLegalOrCustom(ISD::FGETSIGN, MVT::i32);
1069 if ((OpVTLegal || i32Legal) && Op.getValueType().isSimple() &&
1070 Op.getOperand(0).getValueType() != MVT::f128) {
1071 // Cannot eliminate/lower SHL for f128 yet.
1072 EVT Ty = OpVTLegal ? Op.getValueType() : MVT::i32;
1073 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
1074 // place. We expect the SHL to be eliminated by other optimizations.
1075 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, dl, Ty, Op.getOperand(0));
1076 unsigned OpVTSizeInBits = Op.getValueType().getSizeInBits();
1077 if (!OpVTLegal && OpVTSizeInBits > 32)
1078 Sign = TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), Sign);
1079 unsigned ShVal = Op.getValueType().getSizeInBits()-1;
1080 SDValue ShAmt = TLO.DAG.getConstant(ShVal, dl, Op.getValueType());
1081 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl,
1090 // Add, Sub, and Mul don't demand any bits in positions beyond that
1091 // of the highest bit demanded of them.
1092 APInt LoMask = APInt::getLowBitsSet(BitWidth,
1093 BitWidth - NewMask.countLeadingZeros());
1094 if (SimplifyDemandedBits(Op.getOperand(0), LoMask, KnownZero2,
1095 KnownOne2, TLO, Depth+1))
1097 if (SimplifyDemandedBits(Op.getOperand(1), LoMask, KnownZero2,
1098 KnownOne2, TLO, Depth+1))
1100 // See if the operation should be performed at a smaller bit width.
1101 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1106 // Just use computeKnownBits to compute output bits.
1107 TLO.DAG.computeKnownBits(Op, KnownZero, KnownOne, Depth);
1111 // If we know the value of all of the demanded bits, return this as a
1113 if ((NewMask & (KnownZero|KnownOne)) == NewMask) {
1114 // Avoid folding to a constant if any OpaqueConstant is involved.
1115 const SDNode *N = Op.getNode();
1116 for (SDNodeIterator I = SDNodeIterator::begin(N),
1117 E = SDNodeIterator::end(N); I != E; ++I) {
1119 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
1123 return TLO.CombineTo(Op,
1124 TLO.DAG.getConstant(KnownOne, dl, Op.getValueType()));
1130 /// Determine which of the bits specified in Mask are known to be either zero or
1131 /// one and return them in the KnownZero/KnownOne bitsets.
1132 void TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
1135 const SelectionDAG &DAG,
1136 unsigned Depth) const {
1137 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1138 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1139 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1140 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1141 "Should use MaskedValueIsZero if you don't know whether Op"
1142 " is a target node!");
1143 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0);
1146 /// This method can be implemented by targets that want to expose additional
1147 /// information about sign bits to the DAG Combiner.
1148 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
1149 const SelectionDAG &,
1150 unsigned Depth) const {
1151 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1152 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1153 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1154 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1155 "Should use ComputeNumSignBits if you don't know whether Op"
1156 " is a target node!");
1160 /// Test if the given value is known to have exactly one bit set. This differs
1161 /// from computeKnownBits in that it doesn't need to determine which bit is set.
1162 static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) {
1163 // A left-shift of a constant one will have exactly one bit set, because
1164 // shifting the bit off the end is undefined.
1165 if (Val.getOpcode() == ISD::SHL)
1166 if (ConstantSDNode *C =
1167 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1168 if (C->getAPIntValue() == 1)
1171 // Similarly, a right-shift of a constant sign-bit will have exactly
1173 if (Val.getOpcode() == ISD::SRL)
1174 if (ConstantSDNode *C =
1175 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1176 if (C->getAPIntValue().isSignBit())
1179 // More could be done here, though the above checks are enough
1180 // to handle some common cases.
1182 // Fall back to computeKnownBits to catch other known cases.
1183 EVT OpVT = Val.getValueType();
1184 unsigned BitWidth = OpVT.getScalarType().getSizeInBits();
1185 APInt KnownZero, KnownOne;
1186 DAG.computeKnownBits(Val, KnownZero, KnownOne);
1187 return (KnownZero.countPopulation() == BitWidth - 1) &&
1188 (KnownOne.countPopulation() == 1);
1191 bool TargetLowering::isConstTrueVal(const SDNode *N) const {
1195 const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N);
1197 const BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N);
1201 BitVector UndefElements;
1202 CN = BV->getConstantSplatNode(&UndefElements);
1203 // Only interested in constant splats, and we don't try to handle undef
1204 // elements in identifying boolean constants.
1205 if (!CN || UndefElements.none())
1209 switch (getBooleanContents(N->getValueType(0))) {
1210 case UndefinedBooleanContent:
1211 return CN->getAPIntValue()[0];
1212 case ZeroOrOneBooleanContent:
1214 case ZeroOrNegativeOneBooleanContent:
1215 return CN->isAllOnesValue();
1218 llvm_unreachable("Invalid boolean contents");
1221 bool TargetLowering::isConstFalseVal(const SDNode *N) const {
1225 const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N);
1227 const BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N);
1231 BitVector UndefElements;
1232 CN = BV->getConstantSplatNode(&UndefElements);
1233 // Only interested in constant splats, and we don't try to handle undef
1234 // elements in identifying boolean constants.
1235 if (!CN || UndefElements.none())
1239 if (getBooleanContents(N->getValueType(0)) == UndefinedBooleanContent)
1240 return !CN->getAPIntValue()[0];
1242 return CN->isNullValue();
1245 /// Try to simplify a setcc built with the specified operands and cc. If it is
1246 /// unable to simplify it, return a null SDValue.
1248 TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
1249 ISD::CondCode Cond, bool foldBooleans,
1250 DAGCombinerInfo &DCI, SDLoc dl) const {
1251 SelectionDAG &DAG = DCI.DAG;
1253 // These setcc operations always fold.
1257 case ISD::SETFALSE2: return DAG.getConstant(0, dl, VT);
1259 case ISD::SETTRUE2: {
1260 TargetLowering::BooleanContent Cnt =
1261 getBooleanContents(N0->getValueType(0));
1262 return DAG.getConstant(
1263 Cnt == TargetLowering::ZeroOrNegativeOneBooleanContent ? -1ULL : 1, dl,
1268 // Ensure that the constant occurs on the RHS, and fold constant
1270 ISD::CondCode SwappedCC = ISD::getSetCCSwappedOperands(Cond);
1271 if (isa<ConstantSDNode>(N0.getNode()) &&
1272 (DCI.isBeforeLegalizeOps() ||
1273 isCondCodeLegal(SwappedCC, N0.getSimpleValueType())))
1274 return DAG.getSetCC(dl, VT, N1, N0, SwappedCC);
1276 if (auto *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1277 const APInt &C1 = N1C->getAPIntValue();
1279 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1280 // equality comparison, then we're just comparing whether X itself is
1282 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1283 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1284 N0.getOperand(1).getOpcode() == ISD::Constant) {
1286 = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
1287 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1288 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) {
1289 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1290 // (srl (ctlz x), 5) == 0 -> X != 0
1291 // (srl (ctlz x), 5) != 1 -> X != 0
1294 // (srl (ctlz x), 5) != 0 -> X == 0
1295 // (srl (ctlz x), 5) == 1 -> X == 0
1298 SDValue Zero = DAG.getConstant(0, dl, N0.getValueType());
1299 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0),
1305 // Look through truncs that don't change the value of a ctpop.
1306 if (N0.hasOneUse() && N0.getOpcode() == ISD::TRUNCATE)
1307 CTPOP = N0.getOperand(0);
1309 if (CTPOP.hasOneUse() && CTPOP.getOpcode() == ISD::CTPOP &&
1310 (N0 == CTPOP || N0.getValueType().getSizeInBits() >
1311 Log2_32_Ceil(CTPOP.getValueType().getSizeInBits()))) {
1312 EVT CTVT = CTPOP.getValueType();
1313 SDValue CTOp = CTPOP.getOperand(0);
1315 // (ctpop x) u< 2 -> (x & x-1) == 0
1316 // (ctpop x) u> 1 -> (x & x-1) != 0
1317 if ((Cond == ISD::SETULT && C1 == 2) || (Cond == ISD::SETUGT && C1 == 1)){
1318 SDValue Sub = DAG.getNode(ISD::SUB, dl, CTVT, CTOp,
1319 DAG.getConstant(1, dl, CTVT));
1320 SDValue And = DAG.getNode(ISD::AND, dl, CTVT, CTOp, Sub);
1321 ISD::CondCode CC = Cond == ISD::SETULT ? ISD::SETEQ : ISD::SETNE;
1322 return DAG.getSetCC(dl, VT, And, DAG.getConstant(0, dl, CTVT), CC);
1325 // TODO: (ctpop x) == 1 -> x && (x & x-1) == 0 iff ctpop is illegal.
1328 // (zext x) == C --> x == (trunc C)
1329 // (sext x) == C --> x == (trunc C)
1330 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1331 DCI.isBeforeLegalize() && N0->hasOneUse()) {
1332 unsigned MinBits = N0.getValueSizeInBits();
1334 bool Signed = false;
1335 if (N0->getOpcode() == ISD::ZERO_EXTEND) {
1337 MinBits = N0->getOperand(0).getValueSizeInBits();
1338 PreExt = N0->getOperand(0);
1339 } else if (N0->getOpcode() == ISD::AND) {
1340 // DAGCombine turns costly ZExts into ANDs
1341 if (auto *C = dyn_cast<ConstantSDNode>(N0->getOperand(1)))
1342 if ((C->getAPIntValue()+1).isPowerOf2()) {
1343 MinBits = C->getAPIntValue().countTrailingOnes();
1344 PreExt = N0->getOperand(0);
1346 } else if (N0->getOpcode() == ISD::SIGN_EXTEND) {
1348 MinBits = N0->getOperand(0).getValueSizeInBits();
1349 PreExt = N0->getOperand(0);
1351 } else if (auto *LN0 = dyn_cast<LoadSDNode>(N0)) {
1352 // ZEXTLOAD / SEXTLOAD
1353 if (LN0->getExtensionType() == ISD::ZEXTLOAD) {
1354 MinBits = LN0->getMemoryVT().getSizeInBits();
1356 } else if (LN0->getExtensionType() == ISD::SEXTLOAD) {
1358 MinBits = LN0->getMemoryVT().getSizeInBits();
1363 // Figure out how many bits we need to preserve this constant.
1364 unsigned ReqdBits = Signed ?
1365 C1.getBitWidth() - C1.getNumSignBits() + 1 :
1368 // Make sure we're not losing bits from the constant.
1370 MinBits < C1.getBitWidth() &&
1371 MinBits >= ReqdBits) {
1372 EVT MinVT = EVT::getIntegerVT(*DAG.getContext(), MinBits);
1373 if (isTypeDesirableForOp(ISD::SETCC, MinVT)) {
1374 // Will get folded away.
1375 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MinVT, PreExt);
1376 SDValue C = DAG.getConstant(C1.trunc(MinBits), dl, MinVT);
1377 return DAG.getSetCC(dl, VT, Trunc, C, Cond);
1382 // If the LHS is '(and load, const)', the RHS is 0,
1383 // the test is for equality or unsigned, and all 1 bits of the const are
1384 // in the same partial word, see if we can shorten the load.
1385 if (DCI.isBeforeLegalize() &&
1386 !ISD::isSignedIntSetCC(Cond) &&
1387 N0.getOpcode() == ISD::AND && C1 == 0 &&
1388 N0.getNode()->hasOneUse() &&
1389 isa<LoadSDNode>(N0.getOperand(0)) &&
1390 N0.getOperand(0).getNode()->hasOneUse() &&
1391 isa<ConstantSDNode>(N0.getOperand(1))) {
1392 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0));
1394 unsigned bestWidth = 0, bestOffset = 0;
1395 if (!Lod->isVolatile() && Lod->isUnindexed()) {
1396 unsigned origWidth = N0.getValueType().getSizeInBits();
1397 unsigned maskWidth = origWidth;
1398 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to
1399 // 8 bits, but have to be careful...
1400 if (Lod->getExtensionType() != ISD::NON_EXTLOAD)
1401 origWidth = Lod->getMemoryVT().getSizeInBits();
1403 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
1404 for (unsigned width = origWidth / 2; width>=8; width /= 2) {
1405 APInt newMask = APInt::getLowBitsSet(maskWidth, width);
1406 for (unsigned offset=0; offset<origWidth/width; offset++) {
1407 if ((newMask & Mask) == Mask) {
1408 if (!DAG.getDataLayout().isLittleEndian())
1409 bestOffset = (origWidth/width - offset - 1) * (width/8);
1411 bestOffset = (uint64_t)offset * (width/8);
1412 bestMask = Mask.lshr(offset * (width/8) * 8);
1416 newMask = newMask << width;
1421 EVT newVT = EVT::getIntegerVT(*DAG.getContext(), bestWidth);
1422 if (newVT.isRound()) {
1423 EVT PtrType = Lod->getOperand(1).getValueType();
1424 SDValue Ptr = Lod->getBasePtr();
1425 if (bestOffset != 0)
1426 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(),
1427 DAG.getConstant(bestOffset, dl, PtrType));
1428 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset);
1429 SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr,
1430 Lod->getPointerInfo().getWithOffset(bestOffset),
1431 false, false, false, NewAlign);
1432 return DAG.getSetCC(dl, VT,
1433 DAG.getNode(ISD::AND, dl, newVT, NewLoad,
1434 DAG.getConstant(bestMask.trunc(bestWidth),
1436 DAG.getConstant(0LL, dl, newVT), Cond);
1441 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
1442 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
1443 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits();
1445 // If the comparison constant has bits in the upper part, the
1446 // zero-extended value could never match.
1447 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
1448 C1.getBitWidth() - InSize))) {
1452 case ISD::SETEQ: return DAG.getConstant(0, dl, VT);
1455 case ISD::SETNE: return DAG.getConstant(1, dl, VT);
1458 // True if the sign bit of C1 is set.
1459 return DAG.getConstant(C1.isNegative(), dl, VT);
1462 // True if the sign bit of C1 isn't set.
1463 return DAG.getConstant(C1.isNonNegative(), dl, VT);
1469 // Otherwise, we can perform the comparison with the low bits.
1477 EVT newVT = N0.getOperand(0).getValueType();
1478 if (DCI.isBeforeLegalizeOps() ||
1479 (isOperationLegal(ISD::SETCC, newVT) &&
1480 getCondCodeAction(Cond, newVT.getSimpleVT()) == Legal)) {
1482 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), newVT);
1483 SDValue NewConst = DAG.getConstant(C1.trunc(InSize), dl, newVT);
1485 SDValue NewSetCC = DAG.getSetCC(dl, NewSetCCVT, N0.getOperand(0),
1487 return DAG.getBoolExtOrTrunc(NewSetCC, dl, VT, N0.getValueType());
1492 break; // todo, be more careful with signed comparisons
1494 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1495 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1496 EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
1497 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
1498 EVT ExtDstTy = N0.getValueType();
1499 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
1501 // If the constant doesn't fit into the number of bits for the source of
1502 // the sign extension, it is impossible for both sides to be equal.
1503 if (C1.getMinSignedBits() > ExtSrcTyBits)
1504 return DAG.getConstant(Cond == ISD::SETNE, dl, VT);
1507 EVT Op0Ty = N0.getOperand(0).getValueType();
1508 if (Op0Ty == ExtSrcTy) {
1509 ZextOp = N0.getOperand(0);
1511 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
1512 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0),
1513 DAG.getConstant(Imm, dl, Op0Ty));
1515 if (!DCI.isCalledByLegalizer())
1516 DCI.AddToWorklist(ZextOp.getNode());
1517 // Otherwise, make this a use of a zext.
1518 return DAG.getSetCC(dl, VT, ZextOp,
1519 DAG.getConstant(C1 & APInt::getLowBitsSet(
1524 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) &&
1525 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1526 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
1527 if (N0.getOpcode() == ISD::SETCC &&
1528 isTypeLegal(VT) && VT.bitsLE(N0.getValueType())) {
1529 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getAPIntValue() != 1);
1531 return DAG.getNode(ISD::TRUNCATE, dl, VT, N0);
1532 // Invert the condition.
1533 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
1534 CC = ISD::getSetCCInverse(CC,
1535 N0.getOperand(0).getValueType().isInteger());
1536 if (DCI.isBeforeLegalizeOps() ||
1537 isCondCodeLegal(CC, N0.getOperand(0).getSimpleValueType()))
1538 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC);
1541 if ((N0.getOpcode() == ISD::XOR ||
1542 (N0.getOpcode() == ISD::AND &&
1543 N0.getOperand(0).getOpcode() == ISD::XOR &&
1544 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
1545 isa<ConstantSDNode>(N0.getOperand(1)) &&
1546 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) {
1547 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
1548 // can only do this if the top bits are known zero.
1549 unsigned BitWidth = N0.getValueSizeInBits();
1550 if (DAG.MaskedValueIsZero(N0,
1551 APInt::getHighBitsSet(BitWidth,
1553 // Okay, get the un-inverted input value.
1555 if (N0.getOpcode() == ISD::XOR)
1556 Val = N0.getOperand(0);
1558 assert(N0.getOpcode() == ISD::AND &&
1559 N0.getOperand(0).getOpcode() == ISD::XOR);
1560 // ((X^1)&1)^1 -> X & 1
1561 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(),
1562 N0.getOperand(0).getOperand(0),
1566 return DAG.getSetCC(dl, VT, Val, N1,
1567 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1569 } else if (N1C->getAPIntValue() == 1 &&
1571 getBooleanContents(N0->getValueType(0)) ==
1572 ZeroOrOneBooleanContent)) {
1574 if (Op0.getOpcode() == ISD::TRUNCATE)
1575 Op0 = Op0.getOperand(0);
1577 if ((Op0.getOpcode() == ISD::XOR) &&
1578 Op0.getOperand(0).getOpcode() == ISD::SETCC &&
1579 Op0.getOperand(1).getOpcode() == ISD::SETCC) {
1580 // (xor (setcc), (setcc)) == / != 1 -> (setcc) != / == (setcc)
1581 Cond = (Cond == ISD::SETEQ) ? ISD::SETNE : ISD::SETEQ;
1582 return DAG.getSetCC(dl, VT, Op0.getOperand(0), Op0.getOperand(1),
1585 if (Op0.getOpcode() == ISD::AND &&
1586 isa<ConstantSDNode>(Op0.getOperand(1)) &&
1587 cast<ConstantSDNode>(Op0.getOperand(1))->getAPIntValue() == 1) {
1588 // If this is (X&1) == / != 1, normalize it to (X&1) != / == 0.
1589 if (Op0.getValueType().bitsGT(VT))
1590 Op0 = DAG.getNode(ISD::AND, dl, VT,
1591 DAG.getNode(ISD::TRUNCATE, dl, VT, Op0.getOperand(0)),
1592 DAG.getConstant(1, dl, VT));
1593 else if (Op0.getValueType().bitsLT(VT))
1594 Op0 = DAG.getNode(ISD::AND, dl, VT,
1595 DAG.getNode(ISD::ANY_EXTEND, dl, VT, Op0.getOperand(0)),
1596 DAG.getConstant(1, dl, VT));
1598 return DAG.getSetCC(dl, VT, Op0,
1599 DAG.getConstant(0, dl, Op0.getValueType()),
1600 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1602 if (Op0.getOpcode() == ISD::AssertZext &&
1603 cast<VTSDNode>(Op0.getOperand(1))->getVT() == MVT::i1)
1604 return DAG.getSetCC(dl, VT, Op0,
1605 DAG.getConstant(0, dl, Op0.getValueType()),
1606 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1610 APInt MinVal, MaxVal;
1611 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits();
1612 if (ISD::isSignedIntSetCC(Cond)) {
1613 MinVal = APInt::getSignedMinValue(OperandBitSize);
1614 MaxVal = APInt::getSignedMaxValue(OperandBitSize);
1616 MinVal = APInt::getMinValue(OperandBitSize);
1617 MaxVal = APInt::getMaxValue(OperandBitSize);
1620 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
1621 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
1622 if (C1 == MinVal) return DAG.getConstant(1, dl, VT); // X >= MIN --> true
1623 // X >= C0 --> X > (C0 - 1)
1625 ISD::CondCode NewCC = (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT;
1626 if ((DCI.isBeforeLegalizeOps() ||
1627 isCondCodeLegal(NewCC, VT.getSimpleVT())) &&
1628 (!N1C->isOpaque() || (N1C->isOpaque() && C.getBitWidth() <= 64 &&
1629 isLegalICmpImmediate(C.getSExtValue())))) {
1630 return DAG.getSetCC(dl, VT, N0,
1631 DAG.getConstant(C, dl, N1.getValueType()),
1636 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
1637 if (C1 == MaxVal) return DAG.getConstant(1, dl, VT); // X <= MAX --> true
1638 // X <= C0 --> X < (C0 + 1)
1640 ISD::CondCode NewCC = (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT;
1641 if ((DCI.isBeforeLegalizeOps() ||
1642 isCondCodeLegal(NewCC, VT.getSimpleVT())) &&
1643 (!N1C->isOpaque() || (N1C->isOpaque() && C.getBitWidth() <= 64 &&
1644 isLegalICmpImmediate(C.getSExtValue())))) {
1645 return DAG.getSetCC(dl, VT, N0,
1646 DAG.getConstant(C, dl, N1.getValueType()),
1651 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
1652 return DAG.getConstant(0, dl, VT); // X < MIN --> false
1653 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
1654 return DAG.getConstant(1, dl, VT); // X >= MIN --> true
1655 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
1656 return DAG.getConstant(0, dl, VT); // X > MAX --> false
1657 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
1658 return DAG.getConstant(1, dl, VT); // X <= MAX --> true
1660 // Canonicalize setgt X, Min --> setne X, Min
1661 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
1662 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
1663 // Canonicalize setlt X, Max --> setne X, Max
1664 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
1665 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
1667 // If we have setult X, 1, turn it into seteq X, 0
1668 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
1669 return DAG.getSetCC(dl, VT, N0,
1670 DAG.getConstant(MinVal, dl, N0.getValueType()),
1672 // If we have setugt X, Max-1, turn it into seteq X, Max
1673 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
1674 return DAG.getSetCC(dl, VT, N0,
1675 DAG.getConstant(MaxVal, dl, N0.getValueType()),
1678 // If we have "setcc X, C0", check to see if we can shrink the immediate
1681 // SETUGT X, SINTMAX -> SETLT X, 0
1682 if (Cond == ISD::SETUGT &&
1683 C1 == APInt::getSignedMaxValue(OperandBitSize))
1684 return DAG.getSetCC(dl, VT, N0,
1685 DAG.getConstant(0, dl, N1.getValueType()),
1688 // SETULT X, SINTMIN -> SETGT X, -1
1689 if (Cond == ISD::SETULT &&
1690 C1 == APInt::getSignedMinValue(OperandBitSize)) {
1691 SDValue ConstMinusOne =
1692 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize), dl,
1694 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT);
1697 // Fold bit comparisons when we can.
1698 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1699 (VT == N0.getValueType() ||
1700 (isTypeLegal(VT) && VT.bitsLE(N0.getValueType()))) &&
1701 N0.getOpcode() == ISD::AND) {
1702 auto &DL = DAG.getDataLayout();
1703 if (auto *AndRHS = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1704 EVT ShiftTy = DCI.isBeforeLegalize()
1706 : getShiftAmountTy(N0.getValueType(), DL);
1707 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
1708 // Perform the xform if the AND RHS is a single bit.
1709 if (AndRHS->getAPIntValue().isPowerOf2()) {
1710 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1711 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0,
1712 DAG.getConstant(AndRHS->getAPIntValue().logBase2(), dl,
1715 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getAPIntValue()) {
1716 // (X & 8) == 8 --> (X & 8) >> 3
1717 // Perform the xform if C1 is a single bit.
1718 if (C1.isPowerOf2()) {
1719 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1720 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0,
1721 DAG.getConstant(C1.logBase2(), dl,
1728 if (C1.getMinSignedBits() <= 64 &&
1729 !isLegalICmpImmediate(C1.getSExtValue())) {
1730 // (X & -256) == 256 -> (X >> 8) == 1
1731 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1732 N0.getOpcode() == ISD::AND && N0.hasOneUse()) {
1733 if (auto *AndRHS = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1734 const APInt &AndRHSC = AndRHS->getAPIntValue();
1735 if ((-AndRHSC).isPowerOf2() && (AndRHSC & C1) == C1) {
1736 unsigned ShiftBits = AndRHSC.countTrailingZeros();
1737 auto &DL = DAG.getDataLayout();
1738 EVT ShiftTy = DCI.isBeforeLegalize()
1740 : getShiftAmountTy(N0.getValueType(), DL);
1741 EVT CmpTy = N0.getValueType();
1742 SDValue Shift = DAG.getNode(ISD::SRL, dl, CmpTy, N0.getOperand(0),
1743 DAG.getConstant(ShiftBits, dl,
1745 SDValue CmpRHS = DAG.getConstant(C1.lshr(ShiftBits), dl, CmpTy);
1746 return DAG.getSetCC(dl, VT, Shift, CmpRHS, Cond);
1749 } else if (Cond == ISD::SETULT || Cond == ISD::SETUGE ||
1750 Cond == ISD::SETULE || Cond == ISD::SETUGT) {
1751 bool AdjOne = (Cond == ISD::SETULE || Cond == ISD::SETUGT);
1752 // X < 0x100000000 -> (X >> 32) < 1
1753 // X >= 0x100000000 -> (X >> 32) >= 1
1754 // X <= 0x0ffffffff -> (X >> 32) < 1
1755 // X > 0x0ffffffff -> (X >> 32) >= 1
1758 ISD::CondCode NewCond = Cond;
1760 ShiftBits = C1.countTrailingOnes();
1762 NewCond = (Cond == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1764 ShiftBits = C1.countTrailingZeros();
1766 NewC = NewC.lshr(ShiftBits);
1767 if (ShiftBits && NewC.getMinSignedBits() <= 64 &&
1768 isLegalICmpImmediate(NewC.getSExtValue())) {
1769 auto &DL = DAG.getDataLayout();
1770 EVT ShiftTy = DCI.isBeforeLegalize()
1772 : getShiftAmountTy(N0.getValueType(), DL);
1773 EVT CmpTy = N0.getValueType();
1774 SDValue Shift = DAG.getNode(ISD::SRL, dl, CmpTy, N0,
1775 DAG.getConstant(ShiftBits, dl, ShiftTy));
1776 SDValue CmpRHS = DAG.getConstant(NewC, dl, CmpTy);
1777 return DAG.getSetCC(dl, VT, Shift, CmpRHS, NewCond);
1783 if (isa<ConstantFPSDNode>(N0.getNode())) {
1784 // Constant fold or commute setcc.
1785 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl);
1786 if (O.getNode()) return O;
1787 } else if (auto *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
1788 // If the RHS of an FP comparison is a constant, simplify it away in
1790 if (CFP->getValueAPF().isNaN()) {
1791 // If an operand is known to be a nan, we can fold it.
1792 switch (ISD::getUnorderedFlavor(Cond)) {
1793 default: llvm_unreachable("Unknown flavor!");
1794 case 0: // Known false.
1795 return DAG.getConstant(0, dl, VT);
1796 case 1: // Known true.
1797 return DAG.getConstant(1, dl, VT);
1798 case 2: // Undefined.
1799 return DAG.getUNDEF(VT);
1803 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
1804 // constant if knowing that the operand is non-nan is enough. We prefer to
1805 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
1807 if (Cond == ISD::SETO || Cond == ISD::SETUO)
1808 return DAG.getSetCC(dl, VT, N0, N0, Cond);
1810 // If the condition is not legal, see if we can find an equivalent one
1812 if (!isCondCodeLegal(Cond, N0.getSimpleValueType())) {
1813 // If the comparison was an awkward floating-point == or != and one of
1814 // the comparison operands is infinity or negative infinity, convert the
1815 // condition to a less-awkward <= or >=.
1816 if (CFP->getValueAPF().isInfinity()) {
1817 if (CFP->getValueAPF().isNegative()) {
1818 if (Cond == ISD::SETOEQ &&
1819 isCondCodeLegal(ISD::SETOLE, N0.getSimpleValueType()))
1820 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLE);
1821 if (Cond == ISD::SETUEQ &&
1822 isCondCodeLegal(ISD::SETOLE, N0.getSimpleValueType()))
1823 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULE);
1824 if (Cond == ISD::SETUNE &&
1825 isCondCodeLegal(ISD::SETUGT, N0.getSimpleValueType()))
1826 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGT);
1827 if (Cond == ISD::SETONE &&
1828 isCondCodeLegal(ISD::SETUGT, N0.getSimpleValueType()))
1829 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGT);
1831 if (Cond == ISD::SETOEQ &&
1832 isCondCodeLegal(ISD::SETOGE, N0.getSimpleValueType()))
1833 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGE);
1834 if (Cond == ISD::SETUEQ &&
1835 isCondCodeLegal(ISD::SETOGE, N0.getSimpleValueType()))
1836 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGE);
1837 if (Cond == ISD::SETUNE &&
1838 isCondCodeLegal(ISD::SETULT, N0.getSimpleValueType()))
1839 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULT);
1840 if (Cond == ISD::SETONE &&
1841 isCondCodeLegal(ISD::SETULT, N0.getSimpleValueType()))
1842 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLT);
1849 // The sext(setcc()) => setcc() optimization relies on the appropriate
1850 // constant being emitted.
1852 switch (getBooleanContents(N0.getValueType())) {
1853 case UndefinedBooleanContent:
1854 case ZeroOrOneBooleanContent:
1855 EqVal = ISD::isTrueWhenEqual(Cond);
1857 case ZeroOrNegativeOneBooleanContent:
1858 EqVal = ISD::isTrueWhenEqual(Cond) ? -1 : 0;
1862 // We can always fold X == X for integer setcc's.
1863 if (N0.getValueType().isInteger()) {
1864 return DAG.getConstant(EqVal, dl, VT);
1866 unsigned UOF = ISD::getUnorderedFlavor(Cond);
1867 if (UOF == 2) // FP operators that are undefined on NaNs.
1868 return DAG.getConstant(EqVal, dl, VT);
1869 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
1870 return DAG.getConstant(EqVal, dl, VT);
1871 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
1872 // if it is not already.
1873 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
1874 if (NewCond != Cond && (DCI.isBeforeLegalizeOps() ||
1875 getCondCodeAction(NewCond, N0.getSimpleValueType()) == Legal))
1876 return DAG.getSetCC(dl, VT, N0, N1, NewCond);
1879 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1880 N0.getValueType().isInteger()) {
1881 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
1882 N0.getOpcode() == ISD::XOR) {
1883 // Simplify (X+Y) == (X+Z) --> Y == Z
1884 if (N0.getOpcode() == N1.getOpcode()) {
1885 if (N0.getOperand(0) == N1.getOperand(0))
1886 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond);
1887 if (N0.getOperand(1) == N1.getOperand(1))
1888 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond);
1889 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
1890 // If X op Y == Y op X, try other combinations.
1891 if (N0.getOperand(0) == N1.getOperand(1))
1892 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0),
1894 if (N0.getOperand(1) == N1.getOperand(0))
1895 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1),
1900 // If RHS is a legal immediate value for a compare instruction, we need
1901 // to be careful about increasing register pressure needlessly.
1902 bool LegalRHSImm = false;
1904 if (auto *RHSC = dyn_cast<ConstantSDNode>(N1)) {
1905 if (auto *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1906 // Turn (X+C1) == C2 --> X == C2-C1
1907 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) {
1908 return DAG.getSetCC(dl, VT, N0.getOperand(0),
1909 DAG.getConstant(RHSC->getAPIntValue()-
1910 LHSR->getAPIntValue(),
1911 dl, N0.getValueType()), Cond);
1914 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
1915 if (N0.getOpcode() == ISD::XOR)
1916 // If we know that all of the inverted bits are zero, don't bother
1917 // performing the inversion.
1918 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
1920 DAG.getSetCC(dl, VT, N0.getOperand(0),
1921 DAG.getConstant(LHSR->getAPIntValue() ^
1922 RHSC->getAPIntValue(),
1923 dl, N0.getValueType()),
1927 // Turn (C1-X) == C2 --> X == C1-C2
1928 if (auto *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
1929 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) {
1931 DAG.getSetCC(dl, VT, N0.getOperand(1),
1932 DAG.getConstant(SUBC->getAPIntValue() -
1933 RHSC->getAPIntValue(),
1934 dl, N0.getValueType()),
1939 // Could RHSC fold directly into a compare?
1940 if (RHSC->getValueType(0).getSizeInBits() <= 64)
1941 LegalRHSImm = isLegalICmpImmediate(RHSC->getSExtValue());
1944 // Simplify (X+Z) == X --> Z == 0
1945 // Don't do this if X is an immediate that can fold into a cmp
1946 // instruction and X+Z has other uses. It could be an induction variable
1947 // chain, and the transform would increase register pressure.
1948 if (!LegalRHSImm || N0.getNode()->hasOneUse()) {
1949 if (N0.getOperand(0) == N1)
1950 return DAG.getSetCC(dl, VT, N0.getOperand(1),
1951 DAG.getConstant(0, dl, N0.getValueType()), Cond);
1952 if (N0.getOperand(1) == N1) {
1953 if (DAG.isCommutativeBinOp(N0.getOpcode()))
1954 return DAG.getSetCC(dl, VT, N0.getOperand(0),
1955 DAG.getConstant(0, dl, N0.getValueType()),
1957 if (N0.getNode()->hasOneUse()) {
1958 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
1959 auto &DL = DAG.getDataLayout();
1960 // (Z-X) == X --> Z == X<<1
1961 SDValue SH = DAG.getNode(
1962 ISD::SHL, dl, N1.getValueType(), N1,
1963 DAG.getConstant(1, dl,
1964 getShiftAmountTy(N1.getValueType(), DL)));
1965 if (!DCI.isCalledByLegalizer())
1966 DCI.AddToWorklist(SH.getNode());
1967 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond);
1973 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
1974 N1.getOpcode() == ISD::XOR) {
1975 // Simplify X == (X+Z) --> Z == 0
1976 if (N1.getOperand(0) == N0)
1977 return DAG.getSetCC(dl, VT, N1.getOperand(1),
1978 DAG.getConstant(0, dl, N1.getValueType()), Cond);
1979 if (N1.getOperand(1) == N0) {
1980 if (DAG.isCommutativeBinOp(N1.getOpcode()))
1981 return DAG.getSetCC(dl, VT, N1.getOperand(0),
1982 DAG.getConstant(0, dl, N1.getValueType()), Cond);
1983 if (N1.getNode()->hasOneUse()) {
1984 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
1985 auto &DL = DAG.getDataLayout();
1986 // X == (Z-X) --> X<<1 == Z
1987 SDValue SH = DAG.getNode(
1988 ISD::SHL, dl, N1.getValueType(), N0,
1989 DAG.getConstant(1, dl, getShiftAmountTy(N0.getValueType(), DL)));
1990 if (!DCI.isCalledByLegalizer())
1991 DCI.AddToWorklist(SH.getNode());
1992 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond);
1997 // Simplify x&y == y to x&y != 0 if y has exactly one bit set.
1998 // Note that where y is variable and is known to have at most
1999 // one bit set (for example, if it is z&1) we cannot do this;
2000 // the expressions are not equivalent when y==0.
2001 if (N0.getOpcode() == ISD::AND)
2002 if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) {
2003 if (ValueHasExactlyOneBitSet(N1, DAG)) {
2004 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
2005 if (DCI.isBeforeLegalizeOps() ||
2006 isCondCodeLegal(Cond, N0.getSimpleValueType())) {
2007 SDValue Zero = DAG.getConstant(0, dl, N1.getValueType());
2008 return DAG.getSetCC(dl, VT, N0, Zero, Cond);
2012 if (N1.getOpcode() == ISD::AND)
2013 if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) {
2014 if (ValueHasExactlyOneBitSet(N0, DAG)) {
2015 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
2016 if (DCI.isBeforeLegalizeOps() ||
2017 isCondCodeLegal(Cond, N1.getSimpleValueType())) {
2018 SDValue Zero = DAG.getConstant(0, dl, N0.getValueType());
2019 return DAG.getSetCC(dl, VT, N1, Zero, Cond);
2025 // Fold away ALL boolean setcc's.
2027 if (N0.getValueType() == MVT::i1 && foldBooleans) {
2029 default: llvm_unreachable("Unknown integer setcc!");
2030 case ISD::SETEQ: // X == Y -> ~(X^Y)
2031 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2032 N0 = DAG.getNOT(dl, Temp, MVT::i1);
2033 if (!DCI.isCalledByLegalizer())
2034 DCI.AddToWorklist(Temp.getNode());
2036 case ISD::SETNE: // X != Y --> (X^Y)
2037 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2039 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y
2040 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y
2041 Temp = DAG.getNOT(dl, N0, MVT::i1);
2042 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp);
2043 if (!DCI.isCalledByLegalizer())
2044 DCI.AddToWorklist(Temp.getNode());
2046 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X
2047 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X
2048 Temp = DAG.getNOT(dl, N1, MVT::i1);
2049 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp);
2050 if (!DCI.isCalledByLegalizer())
2051 DCI.AddToWorklist(Temp.getNode());
2053 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y
2054 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y
2055 Temp = DAG.getNOT(dl, N0, MVT::i1);
2056 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp);
2057 if (!DCI.isCalledByLegalizer())
2058 DCI.AddToWorklist(Temp.getNode());
2060 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X
2061 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X
2062 Temp = DAG.getNOT(dl, N1, MVT::i1);
2063 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp);
2066 if (VT != MVT::i1) {
2067 if (!DCI.isCalledByLegalizer())
2068 DCI.AddToWorklist(N0.getNode());
2069 // FIXME: If running after legalize, we probably can't do this.
2070 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0);
2075 // Could not fold it.
2079 /// Returns true (and the GlobalValue and the offset) if the node is a
2080 /// GlobalAddress + offset.
2081 bool TargetLowering::isGAPlusOffset(SDNode *N, const GlobalValue *&GA,
2082 int64_t &Offset) const {
2083 if (auto *GASD = dyn_cast<GlobalAddressSDNode>(N)) {
2084 GA = GASD->getGlobal();
2085 Offset += GASD->getOffset();
2089 if (N->getOpcode() == ISD::ADD) {
2090 SDValue N1 = N->getOperand(0);
2091 SDValue N2 = N->getOperand(1);
2092 if (isGAPlusOffset(N1.getNode(), GA, Offset)) {
2093 if (auto *V = dyn_cast<ConstantSDNode>(N2)) {
2094 Offset += V->getSExtValue();
2097 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) {
2098 if (auto *V = dyn_cast<ConstantSDNode>(N1)) {
2099 Offset += V->getSExtValue();
2108 SDValue TargetLowering::PerformDAGCombine(SDNode *N,
2109 DAGCombinerInfo &DCI) const {
2110 // Default implementation: no optimization.
2114 //===----------------------------------------------------------------------===//
2115 // Inline Assembler Implementation Methods
2116 //===----------------------------------------------------------------------===//
2118 TargetLowering::ConstraintType
2119 TargetLowering::getConstraintType(StringRef Constraint) const {
2120 unsigned S = Constraint.size();
2123 switch (Constraint[0]) {
2125 case 'r': return C_RegisterClass;
2127 case 'o': // offsetable
2128 case 'V': // not offsetable
2130 case 'i': // Simple Integer or Relocatable Constant
2131 case 'n': // Simple Integer
2132 case 'E': // Floating Point Constant
2133 case 'F': // Floating Point Constant
2134 case 's': // Relocatable Constant
2135 case 'p': // Address.
2136 case 'X': // Allow ANY value.
2137 case 'I': // Target registers.
2151 if (S > 1 && Constraint[0] == '{' && Constraint[S-1] == '}') {
2152 if (S == 8 && Constraint.substr(1, 6) == "memory") // "{memory}"
2159 /// Try to replace an X constraint, which matches anything, with another that
2160 /// has more specific requirements based on the type of the corresponding
2162 const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const{
2163 if (ConstraintVT.isInteger())
2165 if (ConstraintVT.isFloatingPoint())
2166 return "f"; // works for many targets
2170 /// Lower the specified operand into the Ops vector.
2171 /// If it is invalid, don't add anything to Ops.
2172 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
2173 std::string &Constraint,
2174 std::vector<SDValue> &Ops,
2175 SelectionDAG &DAG) const {
2177 if (Constraint.length() > 1) return;
2179 char ConstraintLetter = Constraint[0];
2180 switch (ConstraintLetter) {
2182 case 'X': // Allows any operand; labels (basic block) use this.
2183 if (Op.getOpcode() == ISD::BasicBlock) {
2188 case 'i': // Simple Integer or Relocatable Constant
2189 case 'n': // Simple Integer
2190 case 's': { // Relocatable Constant
2191 // These operands are interested in values of the form (GV+C), where C may
2192 // be folded in as an offset of GV, or it may be explicitly added. Also, it
2193 // is possible and fine if either GV or C are missing.
2194 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2195 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
2197 // If we have "(add GV, C)", pull out GV/C
2198 if (Op.getOpcode() == ISD::ADD) {
2199 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2200 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
2202 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
2203 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
2206 C = nullptr, GA = nullptr;
2209 // If we find a valid operand, map to the TargetXXX version so that the
2210 // value itself doesn't get selected.
2211 if (GA) { // Either &GV or &GV+C
2212 if (ConstraintLetter != 'n') {
2213 int64_t Offs = GA->getOffset();
2214 if (C) Offs += C->getZExtValue();
2215 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
2216 C ? SDLoc(C) : SDLoc(),
2217 Op.getValueType(), Offs));
2221 if (C) { // just C, no GV.
2222 // Simple constants are not allowed for 's'.
2223 if (ConstraintLetter != 's') {
2224 // gcc prints these as sign extended. Sign extend value to 64 bits
2225 // now; without this it would get ZExt'd later in
2226 // ScheduleDAGSDNodes::EmitNode, which is very generic.
2227 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(),
2228 SDLoc(C), MVT::i64));
2237 std::pair<unsigned, const TargetRegisterClass *>
2238 TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *RI,
2239 StringRef Constraint,
2241 if (Constraint.empty() || Constraint[0] != '{')
2242 return std::make_pair(0u, static_cast<TargetRegisterClass*>(nullptr));
2243 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
2245 // Remove the braces from around the name.
2246 StringRef RegName(Constraint.data()+1, Constraint.size()-2);
2248 std::pair<unsigned, const TargetRegisterClass*> R =
2249 std::make_pair(0u, static_cast<const TargetRegisterClass*>(nullptr));
2251 // Figure out which register class contains this reg.
2252 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
2253 E = RI->regclass_end(); RCI != E; ++RCI) {
2254 const TargetRegisterClass *RC = *RCI;
2256 // If none of the value types for this register class are valid, we
2257 // can't use it. For example, 64-bit reg classes on 32-bit targets.
2261 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
2263 if (RegName.equals_lower(RI->getName(*I))) {
2264 std::pair<unsigned, const TargetRegisterClass*> S =
2265 std::make_pair(*I, RC);
2267 // If this register class has the requested value type, return it,
2268 // otherwise keep searching and return the first class found
2269 // if no other is found which explicitly has the requested type.
2270 if (RC->hasType(VT))
2281 //===----------------------------------------------------------------------===//
2282 // Constraint Selection.
2284 /// Return true of this is an input operand that is a matching constraint like
2286 bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const {
2287 assert(!ConstraintCode.empty() && "No known constraint!");
2288 return isdigit(static_cast<unsigned char>(ConstraintCode[0]));
2291 /// If this is an input matching constraint, this method returns the output
2292 /// operand it matches.
2293 unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const {
2294 assert(!ConstraintCode.empty() && "No known constraint!");
2295 return atoi(ConstraintCode.c_str());
2298 /// Split up the constraint string from the inline assembly value into the
2299 /// specific constraints and their prefixes, and also tie in the associated
2301 /// If this returns an empty vector, and if the constraint string itself
2302 /// isn't empty, there was an error parsing.
2303 TargetLowering::AsmOperandInfoVector
2304 TargetLowering::ParseConstraints(const DataLayout &DL,
2305 const TargetRegisterInfo *TRI,
2306 ImmutableCallSite CS) const {
2307 /// Information about all of the constraints.
2308 AsmOperandInfoVector ConstraintOperands;
2309 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
2310 unsigned maCount = 0; // Largest number of multiple alternative constraints.
2312 // Do a prepass over the constraints, canonicalizing them, and building up the
2313 // ConstraintOperands list.
2314 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
2315 unsigned ResNo = 0; // ResNo - The result number of the next output.
2317 for (InlineAsm::ConstraintInfo &CI : IA->ParseConstraints()) {
2318 ConstraintOperands.emplace_back(std::move(CI));
2319 AsmOperandInfo &OpInfo = ConstraintOperands.back();
2321 // Update multiple alternative constraint count.
2322 if (OpInfo.multipleAlternatives.size() > maCount)
2323 maCount = OpInfo.multipleAlternatives.size();
2325 OpInfo.ConstraintVT = MVT::Other;
2327 // Compute the value type for each operand.
2328 switch (OpInfo.Type) {
2329 case InlineAsm::isOutput:
2330 // Indirect outputs just consume an argument.
2331 if (OpInfo.isIndirect) {
2332 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
2336 // The return value of the call is this value. As such, there is no
2337 // corresponding argument.
2338 assert(!CS.getType()->isVoidTy() &&
2340 if (StructType *STy = dyn_cast<StructType>(CS.getType())) {
2341 OpInfo.ConstraintVT =
2342 getSimpleValueType(DL, STy->getElementType(ResNo));
2344 assert(ResNo == 0 && "Asm only has one result!");
2345 OpInfo.ConstraintVT = getSimpleValueType(DL, CS.getType());
2349 case InlineAsm::isInput:
2350 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
2352 case InlineAsm::isClobber:
2357 if (OpInfo.CallOperandVal) {
2358 llvm::Type *OpTy = OpInfo.CallOperandVal->getType();
2359 if (OpInfo.isIndirect) {
2360 llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
2362 report_fatal_error("Indirect operand for inline asm not a pointer!");
2363 OpTy = PtrTy->getElementType();
2366 // Look for vector wrapped in a struct. e.g. { <16 x i8> }.
2367 if (StructType *STy = dyn_cast<StructType>(OpTy))
2368 if (STy->getNumElements() == 1)
2369 OpTy = STy->getElementType(0);
2371 // If OpTy is not a single value, it may be a struct/union that we
2372 // can tile with integers.
2373 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
2374 unsigned BitSize = DL.getTypeSizeInBits(OpTy);
2383 OpInfo.ConstraintVT =
2384 MVT::getVT(IntegerType::get(OpTy->getContext(), BitSize), true);
2387 } else if (PointerType *PT = dyn_cast<PointerType>(OpTy)) {
2388 unsigned PtrSize = DL.getPointerSizeInBits(PT->getAddressSpace());
2389 OpInfo.ConstraintVT = MVT::getIntegerVT(PtrSize);
2391 OpInfo.ConstraintVT = MVT::getVT(OpTy, true);
2396 // If we have multiple alternative constraints, select the best alternative.
2397 if (!ConstraintOperands.empty()) {
2399 unsigned bestMAIndex = 0;
2400 int bestWeight = -1;
2401 // weight: -1 = invalid match, and 0 = so-so match to 5 = good match.
2404 // Compute the sums of the weights for each alternative, keeping track
2405 // of the best (highest weight) one so far.
2406 for (maIndex = 0; maIndex < maCount; ++maIndex) {
2408 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2409 cIndex != eIndex; ++cIndex) {
2410 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex];
2411 if (OpInfo.Type == InlineAsm::isClobber)
2414 // If this is an output operand with a matching input operand,
2415 // look up the matching input. If their types mismatch, e.g. one
2416 // is an integer, the other is floating point, or their sizes are
2417 // different, flag it as an maCantMatch.
2418 if (OpInfo.hasMatchingInput()) {
2419 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
2420 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
2421 if ((OpInfo.ConstraintVT.isInteger() !=
2422 Input.ConstraintVT.isInteger()) ||
2423 (OpInfo.ConstraintVT.getSizeInBits() !=
2424 Input.ConstraintVT.getSizeInBits())) {
2425 weightSum = -1; // Can't match.
2430 weight = getMultipleConstraintMatchWeight(OpInfo, maIndex);
2435 weightSum += weight;
2438 if (weightSum > bestWeight) {
2439 bestWeight = weightSum;
2440 bestMAIndex = maIndex;
2444 // Now select chosen alternative in each constraint.
2445 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2446 cIndex != eIndex; ++cIndex) {
2447 AsmOperandInfo& cInfo = ConstraintOperands[cIndex];
2448 if (cInfo.Type == InlineAsm::isClobber)
2450 cInfo.selectAlternative(bestMAIndex);
2455 // Check and hook up tied operands, choose constraint code to use.
2456 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2457 cIndex != eIndex; ++cIndex) {
2458 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex];
2460 // If this is an output operand with a matching input operand, look up the
2461 // matching input. If their types mismatch, e.g. one is an integer, the
2462 // other is floating point, or their sizes are different, flag it as an
2464 if (OpInfo.hasMatchingInput()) {
2465 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
2467 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
2468 std::pair<unsigned, const TargetRegisterClass *> MatchRC =
2469 getRegForInlineAsmConstraint(TRI, OpInfo.ConstraintCode,
2470 OpInfo.ConstraintVT);
2471 std::pair<unsigned, const TargetRegisterClass *> InputRC =
2472 getRegForInlineAsmConstraint(TRI, Input.ConstraintCode,
2473 Input.ConstraintVT);
2474 if ((OpInfo.ConstraintVT.isInteger() !=
2475 Input.ConstraintVT.isInteger()) ||
2476 (MatchRC.second != InputRC.second)) {
2477 report_fatal_error("Unsupported asm: input constraint"
2478 " with a matching output constraint of"
2479 " incompatible type!");
2485 return ConstraintOperands;
2488 /// Return an integer indicating how general CT is.
2489 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
2491 case TargetLowering::C_Other:
2492 case TargetLowering::C_Unknown:
2494 case TargetLowering::C_Register:
2496 case TargetLowering::C_RegisterClass:
2498 case TargetLowering::C_Memory:
2501 llvm_unreachable("Invalid constraint type");
2504 /// Examine constraint type and operand type and determine a weight value.
2505 /// This object must already have been set up with the operand type
2506 /// and the current alternative constraint selected.
2507 TargetLowering::ConstraintWeight
2508 TargetLowering::getMultipleConstraintMatchWeight(
2509 AsmOperandInfo &info, int maIndex) const {
2510 InlineAsm::ConstraintCodeVector *rCodes;
2511 if (maIndex >= (int)info.multipleAlternatives.size())
2512 rCodes = &info.Codes;
2514 rCodes = &info.multipleAlternatives[maIndex].Codes;
2515 ConstraintWeight BestWeight = CW_Invalid;
2517 // Loop over the options, keeping track of the most general one.
2518 for (unsigned i = 0, e = rCodes->size(); i != e; ++i) {
2519 ConstraintWeight weight =
2520 getSingleConstraintMatchWeight(info, (*rCodes)[i].c_str());
2521 if (weight > BestWeight)
2522 BestWeight = weight;
2528 /// Examine constraint type and operand type and determine a weight value.
2529 /// This object must already have been set up with the operand type
2530 /// and the current alternative constraint selected.
2531 TargetLowering::ConstraintWeight
2532 TargetLowering::getSingleConstraintMatchWeight(
2533 AsmOperandInfo &info, const char *constraint) const {
2534 ConstraintWeight weight = CW_Invalid;
2535 Value *CallOperandVal = info.CallOperandVal;
2536 // If we don't have a value, we can't do a match,
2537 // but allow it at the lowest weight.
2538 if (!CallOperandVal)
2540 // Look at the constraint type.
2541 switch (*constraint) {
2542 case 'i': // immediate integer.
2543 case 'n': // immediate integer with a known value.
2544 if (isa<ConstantInt>(CallOperandVal))
2545 weight = CW_Constant;
2547 case 's': // non-explicit intregal immediate.
2548 if (isa<GlobalValue>(CallOperandVal))
2549 weight = CW_Constant;
2551 case 'E': // immediate float if host format.
2552 case 'F': // immediate float.
2553 if (isa<ConstantFP>(CallOperandVal))
2554 weight = CW_Constant;
2556 case '<': // memory operand with autodecrement.
2557 case '>': // memory operand with autoincrement.
2558 case 'm': // memory operand.
2559 case 'o': // offsettable memory operand
2560 case 'V': // non-offsettable memory operand
2563 case 'r': // general register.
2564 case 'g': // general register, memory operand or immediate integer.
2565 // note: Clang converts "g" to "imr".
2566 if (CallOperandVal->getType()->isIntegerTy())
2567 weight = CW_Register;
2569 case 'X': // any operand.
2571 weight = CW_Default;
2577 /// If there are multiple different constraints that we could pick for this
2578 /// operand (e.g. "imr") try to pick the 'best' one.
2579 /// This is somewhat tricky: constraints fall into four classes:
2580 /// Other -> immediates and magic values
2581 /// Register -> one specific register
2582 /// RegisterClass -> a group of regs
2583 /// Memory -> memory
2584 /// Ideally, we would pick the most specific constraint possible: if we have
2585 /// something that fits into a register, we would pick it. The problem here
2586 /// is that if we have something that could either be in a register or in
2587 /// memory that use of the register could cause selection of *other*
2588 /// operands to fail: they might only succeed if we pick memory. Because of
2589 /// this the heuristic we use is:
2591 /// 1) If there is an 'other' constraint, and if the operand is valid for
2592 /// that constraint, use it. This makes us take advantage of 'i'
2593 /// constraints when available.
2594 /// 2) Otherwise, pick the most general constraint present. This prefers
2595 /// 'm' over 'r', for example.
2597 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo,
2598 const TargetLowering &TLI,
2599 SDValue Op, SelectionDAG *DAG) {
2600 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options");
2601 unsigned BestIdx = 0;
2602 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown;
2603 int BestGenerality = -1;
2605 // Loop over the options, keeping track of the most general one.
2606 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) {
2607 TargetLowering::ConstraintType CType =
2608 TLI.getConstraintType(OpInfo.Codes[i]);
2610 // If this is an 'other' constraint, see if the operand is valid for it.
2611 // For example, on X86 we might have an 'rI' constraint. If the operand
2612 // is an integer in the range [0..31] we want to use I (saving a load
2613 // of a register), otherwise we must use 'r'.
2614 if (CType == TargetLowering::C_Other && Op.getNode()) {
2615 assert(OpInfo.Codes[i].size() == 1 &&
2616 "Unhandled multi-letter 'other' constraint");
2617 std::vector<SDValue> ResultOps;
2618 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i],
2620 if (!ResultOps.empty()) {
2627 // Things with matching constraints can only be registers, per gcc
2628 // documentation. This mainly affects "g" constraints.
2629 if (CType == TargetLowering::C_Memory && OpInfo.hasMatchingInput())
2632 // This constraint letter is more general than the previous one, use it.
2633 int Generality = getConstraintGenerality(CType);
2634 if (Generality > BestGenerality) {
2637 BestGenerality = Generality;
2641 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx];
2642 OpInfo.ConstraintType = BestType;
2645 /// Determines the constraint code and constraint type to use for the specific
2646 /// AsmOperandInfo, setting OpInfo.ConstraintCode and OpInfo.ConstraintType.
2647 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
2649 SelectionDAG *DAG) const {
2650 assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
2652 // Single-letter constraints ('r') are very common.
2653 if (OpInfo.Codes.size() == 1) {
2654 OpInfo.ConstraintCode = OpInfo.Codes[0];
2655 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2657 ChooseConstraint(OpInfo, *this, Op, DAG);
2660 // 'X' matches anything.
2661 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
2662 // Labels and constants are handled elsewhere ('X' is the only thing
2663 // that matches labels). For Functions, the type here is the type of
2664 // the result, which is not what we want to look at; leave them alone.
2665 Value *v = OpInfo.CallOperandVal;
2666 if (isa<BasicBlock>(v) || isa<ConstantInt>(v) || isa<Function>(v)) {
2667 OpInfo.CallOperandVal = v;
2671 // Otherwise, try to resolve it to something we know about by looking at
2672 // the actual operand type.
2673 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
2674 OpInfo.ConstraintCode = Repl;
2675 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2680 /// \brief Given an exact SDIV by a constant, create a multiplication
2681 /// with the multiplicative inverse of the constant.
2682 static SDValue BuildExactSDIV(const TargetLowering &TLI, SDValue Op1, APInt d,
2683 SDLoc dl, SelectionDAG &DAG,
2684 std::vector<SDNode *> &Created) {
2685 assert(d != 0 && "Division by zero!");
2687 // Shift the value upfront if it is even, so the LSB is one.
2688 unsigned ShAmt = d.countTrailingZeros();
2690 // TODO: For UDIV use SRL instead of SRA.
2692 DAG.getConstant(ShAmt, dl, TLI.getShiftAmountTy(Op1.getValueType(),
2693 DAG.getDataLayout()));
2695 Flags.setExact(true);
2696 Op1 = DAG.getNode(ISD::SRA, dl, Op1.getValueType(), Op1, Amt, &Flags);
2697 Created.push_back(Op1.getNode());
2701 // Calculate the multiplicative inverse, using Newton's method.
2703 while ((t = d*xn) != 1)
2704 xn *= APInt(d.getBitWidth(), 2) - t;
2706 SDValue Op2 = DAG.getConstant(xn, dl, Op1.getValueType());
2707 SDValue Mul = DAG.getNode(ISD::MUL, dl, Op1.getValueType(), Op1, Op2);
2708 Created.push_back(Mul.getNode());
2712 SDValue TargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor,
2714 std::vector<SDNode *> *Created) const {
2715 AttributeSet Attr = DAG.getMachineFunction().getFunction()->getAttributes();
2716 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2717 if (TLI.isIntDivCheap(N->getValueType(0), Attr))
2718 return SDValue(N,0); // Lower SDIV as SDIV
2722 /// \brief Given an ISD::SDIV node expressing a divide by constant,
2723 /// return a DAG expression to select that will generate the same value by
2724 /// multiplying by a magic number.
2725 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
2726 SDValue TargetLowering::BuildSDIV(SDNode *N, const APInt &Divisor,
2727 SelectionDAG &DAG, bool IsAfterLegalization,
2728 std::vector<SDNode *> *Created) const {
2729 assert(Created && "No vector to hold sdiv ops.");
2731 EVT VT = N->getValueType(0);
2734 // Check to see if we can do this.
2735 // FIXME: We should be more aggressive here.
2736 if (!isTypeLegal(VT))
2739 // If the sdiv has an 'exact' bit we can use a simpler lowering.
2740 if (cast<BinaryWithFlagsSDNode>(N)->Flags.hasExact())
2741 return BuildExactSDIV(*this, N->getOperand(0), Divisor, dl, DAG, *Created);
2743 APInt::ms magics = Divisor.magic();
2745 // Multiply the numerator (operand 0) by the magic value
2746 // FIXME: We should support doing a MUL in a wider type
2748 if (IsAfterLegalization ? isOperationLegal(ISD::MULHS, VT) :
2749 isOperationLegalOrCustom(ISD::MULHS, VT))
2750 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0),
2751 DAG.getConstant(magics.m, dl, VT));
2752 else if (IsAfterLegalization ? isOperationLegal(ISD::SMUL_LOHI, VT) :
2753 isOperationLegalOrCustom(ISD::SMUL_LOHI, VT))
2754 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT),
2756 DAG.getConstant(magics.m, dl, VT)).getNode(), 1);
2758 return SDValue(); // No mulhs or equvialent
2759 // If d > 0 and m < 0, add the numerator
2760 if (Divisor.isStrictlyPositive() && magics.m.isNegative()) {
2761 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0));
2762 Created->push_back(Q.getNode());
2764 // If d < 0 and m > 0, subtract the numerator.
2765 if (Divisor.isNegative() && magics.m.isStrictlyPositive()) {
2766 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0));
2767 Created->push_back(Q.getNode());
2769 auto &DL = DAG.getDataLayout();
2770 // Shift right algebraic if shift value is nonzero
2773 ISD::SRA, dl, VT, Q,
2774 DAG.getConstant(magics.s, dl, getShiftAmountTy(Q.getValueType(), DL)));
2775 Created->push_back(Q.getNode());
2777 // Extract the sign bit and add it to the quotient
2779 DAG.getNode(ISD::SRL, dl, VT, Q,
2780 DAG.getConstant(VT.getScalarSizeInBits() - 1, dl,
2781 getShiftAmountTy(Q.getValueType(), DL)));
2782 Created->push_back(T.getNode());
2783 return DAG.getNode(ISD::ADD, dl, VT, Q, T);
2786 /// \brief Given an ISD::UDIV node expressing a divide by constant,
2787 /// return a DAG expression to select that will generate the same value by
2788 /// multiplying by a magic number.
2789 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
2790 SDValue TargetLowering::BuildUDIV(SDNode *N, const APInt &Divisor,
2791 SelectionDAG &DAG, bool IsAfterLegalization,
2792 std::vector<SDNode *> *Created) const {
2793 assert(Created && "No vector to hold udiv ops.");
2795 EVT VT = N->getValueType(0);
2797 auto &DL = DAG.getDataLayout();
2799 // Check to see if we can do this.
2800 // FIXME: We should be more aggressive here.
2801 if (!isTypeLegal(VT))
2804 // FIXME: We should use a narrower constant when the upper
2805 // bits are known to be zero.
2806 APInt::mu magics = Divisor.magicu();
2808 SDValue Q = N->getOperand(0);
2810 // If the divisor is even, we can avoid using the expensive fixup by shifting
2811 // the divided value upfront.
2812 if (magics.a != 0 && !Divisor[0]) {
2813 unsigned Shift = Divisor.countTrailingZeros();
2815 ISD::SRL, dl, VT, Q,
2816 DAG.getConstant(Shift, dl, getShiftAmountTy(Q.getValueType(), DL)));
2817 Created->push_back(Q.getNode());
2819 // Get magic number for the shifted divisor.
2820 magics = Divisor.lshr(Shift).magicu(Shift);
2821 assert(magics.a == 0 && "Should use cheap fixup now");
2824 // Multiply the numerator (operand 0) by the magic value
2825 // FIXME: We should support doing a MUL in a wider type
2826 if (IsAfterLegalization ? isOperationLegal(ISD::MULHU, VT) :
2827 isOperationLegalOrCustom(ISD::MULHU, VT))
2828 Q = DAG.getNode(ISD::MULHU, dl, VT, Q, DAG.getConstant(magics.m, dl, VT));
2829 else if (IsAfterLegalization ? isOperationLegal(ISD::UMUL_LOHI, VT) :
2830 isOperationLegalOrCustom(ISD::UMUL_LOHI, VT))
2831 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT), Q,
2832 DAG.getConstant(magics.m, dl, VT)).getNode(), 1);
2834 return SDValue(); // No mulhu or equvialent
2836 Created->push_back(Q.getNode());
2838 if (magics.a == 0) {
2839 assert(magics.s < Divisor.getBitWidth() &&
2840 "We shouldn't generate an undefined shift!");
2842 ISD::SRL, dl, VT, Q,
2843 DAG.getConstant(magics.s, dl, getShiftAmountTy(Q.getValueType(), DL)));
2845 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q);
2846 Created->push_back(NPQ.getNode());
2848 ISD::SRL, dl, VT, NPQ,
2849 DAG.getConstant(1, dl, getShiftAmountTy(NPQ.getValueType(), DL)));
2850 Created->push_back(NPQ.getNode());
2851 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q);
2852 Created->push_back(NPQ.getNode());
2854 ISD::SRL, dl, VT, NPQ,
2855 DAG.getConstant(magics.s - 1, dl,
2856 getShiftAmountTy(NPQ.getValueType(), DL)));
2860 bool TargetLowering::
2861 verifyReturnAddressArgumentIsConstant(SDValue Op, SelectionDAG &DAG) const {
2862 if (!isa<ConstantSDNode>(Op.getOperand(0))) {
2863 DAG.getContext()->emitError("argument to '__builtin_return_address' must "
2864 "be a constant integer");
2871 //===----------------------------------------------------------------------===//
2872 // Legalization Utilities
2873 //===----------------------------------------------------------------------===//
2875 bool TargetLowering::expandMUL(SDNode *N, SDValue &Lo, SDValue &Hi, EVT HiLoVT,
2876 SelectionDAG &DAG, SDValue LL, SDValue LH,
2877 SDValue RL, SDValue RH) const {
2878 EVT VT = N->getValueType(0);
2881 bool HasMULHS = isOperationLegalOrCustom(ISD::MULHS, HiLoVT);
2882 bool HasMULHU = isOperationLegalOrCustom(ISD::MULHU, HiLoVT);
2883 bool HasSMUL_LOHI = isOperationLegalOrCustom(ISD::SMUL_LOHI, HiLoVT);
2884 bool HasUMUL_LOHI = isOperationLegalOrCustom(ISD::UMUL_LOHI, HiLoVT);
2885 if (HasMULHU || HasMULHS || HasUMUL_LOHI || HasSMUL_LOHI) {
2886 unsigned OuterBitSize = VT.getSizeInBits();
2887 unsigned InnerBitSize = HiLoVT.getSizeInBits();
2888 unsigned LHSSB = DAG.ComputeNumSignBits(N->getOperand(0));
2889 unsigned RHSSB = DAG.ComputeNumSignBits(N->getOperand(1));
2891 // LL, LH, RL, and RH must be either all NULL or all set to a value.
2892 assert((LL.getNode() && LH.getNode() && RL.getNode() && RH.getNode()) ||
2893 (!LL.getNode() && !LH.getNode() && !RL.getNode() && !RH.getNode()));
2895 if (!LL.getNode() && !RL.getNode() &&
2896 isOperationLegalOrCustom(ISD::TRUNCATE, HiLoVT)) {
2897 LL = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, N->getOperand(0));
2898 RL = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, N->getOperand(1));
2904 APInt HighMask = APInt::getHighBitsSet(OuterBitSize, InnerBitSize);
2905 if (DAG.MaskedValueIsZero(N->getOperand(0), HighMask) &&
2906 DAG.MaskedValueIsZero(N->getOperand(1), HighMask)) {
2907 // The inputs are both zero-extended.
2909 // We can emit a umul_lohi.
2910 Lo = DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(HiLoVT, HiLoVT), LL,
2912 Hi = SDValue(Lo.getNode(), 1);
2916 // We can emit a mulhu+mul.
2917 Lo = DAG.getNode(ISD::MUL, dl, HiLoVT, LL, RL);
2918 Hi = DAG.getNode(ISD::MULHU, dl, HiLoVT, LL, RL);
2922 if (LHSSB > InnerBitSize && RHSSB > InnerBitSize) {
2923 // The input values are both sign-extended.
2925 // We can emit a smul_lohi.
2926 Lo = DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(HiLoVT, HiLoVT), LL,
2928 Hi = SDValue(Lo.getNode(), 1);
2932 // We can emit a mulhs+mul.
2933 Lo = DAG.getNode(ISD::MUL, dl, HiLoVT, LL, RL);
2934 Hi = DAG.getNode(ISD::MULHS, dl, HiLoVT, LL, RL);
2939 if (!LH.getNode() && !RH.getNode() &&
2940 isOperationLegalOrCustom(ISD::SRL, VT) &&
2941 isOperationLegalOrCustom(ISD::TRUNCATE, HiLoVT)) {
2942 auto &DL = DAG.getDataLayout();
2943 unsigned ShiftAmt = VT.getSizeInBits() - HiLoVT.getSizeInBits();
2944 SDValue Shift = DAG.getConstant(ShiftAmt, dl, getShiftAmountTy(VT, DL));
2945 LH = DAG.getNode(ISD::SRL, dl, VT, N->getOperand(0), Shift);
2946 LH = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, LH);
2947 RH = DAG.getNode(ISD::SRL, dl, VT, N->getOperand(1), Shift);
2948 RH = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, RH);
2955 // Lo,Hi = umul LHS, RHS.
2956 SDValue UMulLOHI = DAG.getNode(ISD::UMUL_LOHI, dl,
2957 DAG.getVTList(HiLoVT, HiLoVT), LL, RL);
2959 Hi = UMulLOHI.getValue(1);
2960 RH = DAG.getNode(ISD::MUL, dl, HiLoVT, LL, RH);
2961 LH = DAG.getNode(ISD::MUL, dl, HiLoVT, LH, RL);
2962 Hi = DAG.getNode(ISD::ADD, dl, HiLoVT, Hi, RH);
2963 Hi = DAG.getNode(ISD::ADD, dl, HiLoVT, Hi, LH);
2967 Lo = DAG.getNode(ISD::MUL, dl, HiLoVT, LL, RL);
2968 Hi = DAG.getNode(ISD::MULHU, dl, HiLoVT, LL, RL);
2969 RH = DAG.getNode(ISD::MUL, dl, HiLoVT, LL, RH);
2970 LH = DAG.getNode(ISD::MUL, dl, HiLoVT, LH, RL);
2971 Hi = DAG.getNode(ISD::ADD, dl, HiLoVT, Hi, RH);
2972 Hi = DAG.getNode(ISD::ADD, dl, HiLoVT, Hi, LH);
2979 bool TargetLowering::expandFP_TO_SINT(SDNode *Node, SDValue &Result,
2980 SelectionDAG &DAG) const {
2981 EVT VT = Node->getOperand(0).getValueType();
2982 EVT NVT = Node->getValueType(0);
2983 SDLoc dl(SDValue(Node, 0));
2985 // FIXME: Only f32 to i64 conversions are supported.
2986 if (VT != MVT::f32 || NVT != MVT::i64)
2989 // Expand f32 -> i64 conversion
2990 // This algorithm comes from compiler-rt's implementation of fixsfdi:
2991 // https://github.com/llvm-mirror/compiler-rt/blob/master/lib/builtins/fixsfdi.c
2992 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(),
2993 VT.getSizeInBits());
2994 SDValue ExponentMask = DAG.getConstant(0x7F800000, dl, IntVT);
2995 SDValue ExponentLoBit = DAG.getConstant(23, dl, IntVT);
2996 SDValue Bias = DAG.getConstant(127, dl, IntVT);
2997 SDValue SignMask = DAG.getConstant(APInt::getSignBit(VT.getSizeInBits()), dl,
2999 SDValue SignLowBit = DAG.getConstant(VT.getSizeInBits() - 1, dl, IntVT);
3000 SDValue MantissaMask = DAG.getConstant(0x007FFFFF, dl, IntVT);
3002 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, IntVT, Node->getOperand(0));
3004 auto &DL = DAG.getDataLayout();
3005 SDValue ExponentBits = DAG.getNode(
3006 ISD::SRL, dl, IntVT, DAG.getNode(ISD::AND, dl, IntVT, Bits, ExponentMask),
3007 DAG.getZExtOrTrunc(ExponentLoBit, dl, getShiftAmountTy(IntVT, DL)));
3008 SDValue Exponent = DAG.getNode(ISD::SUB, dl, IntVT, ExponentBits, Bias);
3010 SDValue Sign = DAG.getNode(
3011 ISD::SRA, dl, IntVT, DAG.getNode(ISD::AND, dl, IntVT, Bits, SignMask),
3012 DAG.getZExtOrTrunc(SignLowBit, dl, getShiftAmountTy(IntVT, DL)));
3013 Sign = DAG.getSExtOrTrunc(Sign, dl, NVT);
3015 SDValue R = DAG.getNode(ISD::OR, dl, IntVT,
3016 DAG.getNode(ISD::AND, dl, IntVT, Bits, MantissaMask),
3017 DAG.getConstant(0x00800000, dl, IntVT));
3019 R = DAG.getZExtOrTrunc(R, dl, NVT);
3021 R = DAG.getSelectCC(
3022 dl, Exponent, ExponentLoBit,
3023 DAG.getNode(ISD::SHL, dl, NVT, R,
3025 DAG.getNode(ISD::SUB, dl, IntVT, Exponent, ExponentLoBit),
3026 dl, getShiftAmountTy(IntVT, DL))),
3027 DAG.getNode(ISD::SRL, dl, NVT, R,
3029 DAG.getNode(ISD::SUB, dl, IntVT, ExponentLoBit, Exponent),
3030 dl, getShiftAmountTy(IntVT, DL))),
3033 SDValue Ret = DAG.getNode(ISD::SUB, dl, NVT,
3034 DAG.getNode(ISD::XOR, dl, NVT, R, Sign),
3037 Result = DAG.getSelectCC(dl, Exponent, DAG.getConstant(0, dl, IntVT),
3038 DAG.getConstant(0, dl, NVT), Ret, ISD::SETLT);
3042 //===----------------------------------------------------------------------===//
3043 // Implementation of Emulated TLS Model
3044 //===----------------------------------------------------------------------===//
3046 SDValue TargetLowering::LowerToTLSEmulatedModel(const GlobalAddressSDNode *GA,
3047 SelectionDAG &DAG) const {
3048 // Access to address of TLS varialbe xyz is lowered to a function call:
3049 // __emutls_get_address( address of global variable named "__emutls_v.xyz" )
3050 EVT PtrVT = getPointerTy(DAG.getDataLayout());
3051 PointerType *VoidPtrType = Type::getInt8PtrTy(*DAG.getContext());
3056 std::string NameString = ("__emutls_v." + GA->getGlobal()->getName()).str();
3057 Module *VariableModule = const_cast<Module*>(GA->getGlobal()->getParent());
3058 StringRef EmuTlsVarName(NameString);
3059 GlobalVariable *EmuTlsVar = VariableModule->getNamedGlobal(EmuTlsVarName);
3061 EmuTlsVar = dyn_cast_or_null<GlobalVariable>(
3062 VariableModule->getOrInsertGlobal(EmuTlsVarName, VoidPtrType));
3063 Entry.Node = DAG.getGlobalAddress(EmuTlsVar, dl, PtrVT);
3064 Entry.Ty = VoidPtrType;
3065 Args.push_back(Entry);
3067 SDValue EmuTlsGetAddr = DAG.getExternalSymbol("__emutls_get_address", PtrVT);
3069 TargetLowering::CallLoweringInfo CLI(DAG);
3070 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode());
3071 CLI.setCallee(CallingConv::C, VoidPtrType, EmuTlsGetAddr, std::move(Args), 0);
3072 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
3074 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
3075 // At last for X86 targets, maybe good for other targets too?
3076 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3077 MFI->setAdjustsStack(true); // Is this only for X86 target?
3078 MFI->setHasCalls(true);
3080 assert((GA->getOffset() == 0) &&
3081 "Emulated TLS must have zero offset in GlobalAddressSDNode");
3082 return CallResult.first;