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 const SDValue *Ops, unsigned NumOps,
89 bool isSigned, SDLoc dl,
91 bool isReturnValueUsed) const {
92 TargetLowering::ArgListTy Args;
95 TargetLowering::ArgListEntry Entry;
96 for (unsigned i = 0; i != NumOps; ++i) {
98 Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext());
99 Entry.isSExt = shouldSignExtendTypeInLibCall(Ops[i].getValueType(), isSigned);
100 Entry.isZExt = !shouldSignExtendTypeInLibCall(Ops[i].getValueType(), isSigned);
101 Args.push_back(Entry);
103 if (LC == RTLIB::UNKNOWN_LIBCALL)
104 report_fatal_error("Unsupported library call operation!");
105 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC), getPointerTy());
107 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
108 TargetLowering::CallLoweringInfo CLI(DAG);
109 bool signExtend = shouldSignExtendTypeInLibCall(RetVT, isSigned);
110 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
111 .setCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
112 .setNoReturn(doesNotReturn).setDiscardResult(!isReturnValueUsed)
113 .setSExtResult(signExtend).setZExtResult(!signExtend);
114 return LowerCallTo(CLI);
118 /// SoftenSetCCOperands - Soften the operands of a comparison. This code is
119 /// shared among BR_CC, SELECT_CC, and SETCC handlers.
120 void TargetLowering::softenSetCCOperands(SelectionDAG &DAG, EVT VT,
121 SDValue &NewLHS, SDValue &NewRHS,
122 ISD::CondCode &CCCode,
124 assert((VT == MVT::f32 || VT == MVT::f64 || VT == MVT::f128)
125 && "Unsupported setcc type!");
127 // Expand into one or more soft-fp libcall(s).
128 RTLIB::Libcall LC1 = RTLIB::UNKNOWN_LIBCALL, LC2 = RTLIB::UNKNOWN_LIBCALL;
132 LC1 = (VT == MVT::f32) ? RTLIB::OEQ_F32 :
133 (VT == MVT::f64) ? RTLIB::OEQ_F64 : RTLIB::OEQ_F128;
137 LC1 = (VT == MVT::f32) ? RTLIB::UNE_F32 :
138 (VT == MVT::f64) ? RTLIB::UNE_F64 : RTLIB::UNE_F128;
142 LC1 = (VT == MVT::f32) ? RTLIB::OGE_F32 :
143 (VT == MVT::f64) ? RTLIB::OGE_F64 : RTLIB::OGE_F128;
147 LC1 = (VT == MVT::f32) ? RTLIB::OLT_F32 :
148 (VT == MVT::f64) ? RTLIB::OLT_F64 : RTLIB::OLT_F128;
152 LC1 = (VT == MVT::f32) ? RTLIB::OLE_F32 :
153 (VT == MVT::f64) ? RTLIB::OLE_F64 : RTLIB::OLE_F128;
157 LC1 = (VT == MVT::f32) ? RTLIB::OGT_F32 :
158 (VT == MVT::f64) ? RTLIB::OGT_F64 : RTLIB::OGT_F128;
161 LC1 = (VT == MVT::f32) ? RTLIB::UO_F32 :
162 (VT == MVT::f64) ? RTLIB::UO_F64 : RTLIB::UO_F128;
165 LC1 = (VT == MVT::f32) ? RTLIB::O_F32 :
166 (VT == MVT::f64) ? RTLIB::O_F64 : RTLIB::O_F128;
169 LC1 = (VT == MVT::f32) ? RTLIB::UO_F32 :
170 (VT == MVT::f64) ? RTLIB::UO_F64 : RTLIB::UO_F128;
173 // SETONE = SETOLT | SETOGT
174 LC1 = (VT == MVT::f32) ? RTLIB::OLT_F32 :
175 (VT == MVT::f64) ? RTLIB::OLT_F64 : RTLIB::OLT_F128;
178 LC2 = (VT == MVT::f32) ? RTLIB::OGT_F32 :
179 (VT == MVT::f64) ? RTLIB::OGT_F64 : RTLIB::OGT_F128;
182 LC2 = (VT == MVT::f32) ? RTLIB::OGE_F32 :
183 (VT == MVT::f64) ? RTLIB::OGE_F64 : RTLIB::OGE_F128;
186 LC2 = (VT == MVT::f32) ? RTLIB::OLT_F32 :
187 (VT == MVT::f64) ? RTLIB::OLT_F64 : RTLIB::OLT_F128;
190 LC2 = (VT == MVT::f32) ? RTLIB::OLE_F32 :
191 (VT == MVT::f64) ? RTLIB::OLE_F64 : RTLIB::OLE_F128;
194 LC2 = (VT == MVT::f32) ? RTLIB::OEQ_F32 :
195 (VT == MVT::f64) ? RTLIB::OEQ_F64 : RTLIB::OEQ_F128;
197 default: llvm_unreachable("Do not know how to soften this setcc!");
201 // Use the target specific return value for comparions lib calls.
202 EVT RetVT = getCmpLibcallReturnType();
203 SDValue Ops[2] = { NewLHS, NewRHS };
204 NewLHS = makeLibCall(DAG, LC1, RetVT, Ops, 2, false/*sign irrelevant*/,
206 NewRHS = DAG.getConstant(0, dl, RetVT);
207 CCCode = getCmpLibcallCC(LC1);
208 if (LC2 != RTLIB::UNKNOWN_LIBCALL) {
209 SDValue Tmp = DAG.getNode(ISD::SETCC, dl,
210 getSetCCResultType(*DAG.getContext(), RetVT),
211 NewLHS, NewRHS, DAG.getCondCode(CCCode));
212 NewLHS = makeLibCall(DAG, LC2, RetVT, Ops, 2, false/*sign irrelevant*/,
214 NewLHS = DAG.getNode(ISD::SETCC, dl,
215 getSetCCResultType(*DAG.getContext(), RetVT), NewLHS,
216 NewRHS, DAG.getCondCode(getCmpLibcallCC(LC2)));
217 NewLHS = DAG.getNode(ISD::OR, dl, Tmp.getValueType(), Tmp, NewLHS);
222 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
223 /// current function. The returned value is a member of the
224 /// MachineJumpTableInfo::JTEntryKind enum.
225 unsigned TargetLowering::getJumpTableEncoding() const {
226 // In non-pic modes, just use the address of a block.
227 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
228 return MachineJumpTableInfo::EK_BlockAddress;
230 // In PIC mode, if the target supports a GPRel32 directive, use it.
231 if (getTargetMachine().getMCAsmInfo()->getGPRel32Directive() != nullptr)
232 return MachineJumpTableInfo::EK_GPRel32BlockAddress;
234 // Otherwise, use a label difference.
235 return MachineJumpTableInfo::EK_LabelDifference32;
238 SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table,
239 SelectionDAG &DAG) const {
240 // If our PIC model is GP relative, use the global offset table as the base.
241 unsigned JTEncoding = getJumpTableEncoding();
243 if ((JTEncoding == MachineJumpTableInfo::EK_GPRel64BlockAddress) ||
244 (JTEncoding == MachineJumpTableInfo::EK_GPRel32BlockAddress))
245 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy(0));
250 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
251 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
254 TargetLowering::getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
255 unsigned JTI,MCContext &Ctx) const{
256 // The normal PIC reloc base is the label at the start of the jump table.
257 return MCSymbolRefExpr::create(MF->getJTISymbol(JTI, Ctx), Ctx);
261 TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
262 // Assume that everything is safe in static mode.
263 if (getTargetMachine().getRelocationModel() == Reloc::Static)
266 // In dynamic-no-pic mode, assume that known defined values are safe.
267 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC &&
269 !GA->getGlobal()->isDeclaration() &&
270 !GA->getGlobal()->isWeakForLinker())
273 // Otherwise assume nothing is safe.
277 //===----------------------------------------------------------------------===//
278 // Optimization Methods
279 //===----------------------------------------------------------------------===//
281 /// ShrinkDemandedConstant - Check to see if the specified operand of the
282 /// specified instruction is a constant integer. If so, check to see if there
283 /// are any bits set in the constant that are not demanded. If so, shrink the
284 /// constant and return true.
285 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op,
286 const APInt &Demanded) {
289 // FIXME: ISD::SELECT, ISD::SELECT_CC
290 switch (Op.getOpcode()) {
295 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
296 if (!C) return false;
298 if (Op.getOpcode() == ISD::XOR &&
299 (C->getAPIntValue() | (~Demanded)).isAllOnesValue())
302 // if we can expand it to have all bits set, do it
303 if (C->getAPIntValue().intersects(~Demanded)) {
304 EVT VT = Op.getValueType();
305 SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0),
306 DAG.getConstant(Demanded &
309 return CombineTo(Op, New);
319 /// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the
320 /// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening
321 /// cast, but it could be generalized for targets with other types of
322 /// implicit widening casts.
324 TargetLowering::TargetLoweringOpt::ShrinkDemandedOp(SDValue Op,
326 const APInt &Demanded,
328 assert(Op.getNumOperands() == 2 &&
329 "ShrinkDemandedOp only supports binary operators!");
330 assert(Op.getNode()->getNumValues() == 1 &&
331 "ShrinkDemandedOp only supports nodes with one result!");
333 // Early return, as this function cannot handle vector types.
334 if (Op.getValueType().isVector())
337 // Don't do this if the node has another user, which may require the
339 if (!Op.getNode()->hasOneUse())
342 // Search for the smallest integer type with free casts to and from
343 // Op's type. For expedience, just check power-of-2 integer types.
344 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
345 unsigned DemandedSize = BitWidth - Demanded.countLeadingZeros();
346 unsigned SmallVTBits = DemandedSize;
347 if (!isPowerOf2_32(SmallVTBits))
348 SmallVTBits = NextPowerOf2(SmallVTBits);
349 for (; SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) {
350 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits);
351 if (TLI.isTruncateFree(Op.getValueType(), SmallVT) &&
352 TLI.isZExtFree(SmallVT, Op.getValueType())) {
353 // We found a type with free casts.
354 SDValue X = DAG.getNode(Op.getOpcode(), dl, SmallVT,
355 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
356 Op.getNode()->getOperand(0)),
357 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
358 Op.getNode()->getOperand(1)));
359 bool NeedZext = DemandedSize > SmallVTBits;
360 SDValue Z = DAG.getNode(NeedZext ? ISD::ZERO_EXTEND : ISD::ANY_EXTEND,
361 dl, Op.getValueType(), X);
362 return CombineTo(Op, Z);
368 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
369 /// DemandedMask bits of the result of Op are ever used downstream. If we can
370 /// use this information to simplify Op, create a new simplified DAG node and
371 /// return true, returning the original and new nodes in Old and New. Otherwise,
372 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
373 /// the expression (used to simplify the caller). The KnownZero/One bits may
374 /// only be accurate for those bits in the DemandedMask.
375 bool TargetLowering::SimplifyDemandedBits(SDValue Op,
376 const APInt &DemandedMask,
379 TargetLoweringOpt &TLO,
380 unsigned Depth) const {
381 unsigned BitWidth = DemandedMask.getBitWidth();
382 assert(Op.getValueType().getScalarType().getSizeInBits() == BitWidth &&
383 "Mask size mismatches value type size!");
384 APInt NewMask = DemandedMask;
387 // Don't know anything.
388 KnownZero = KnownOne = APInt(BitWidth, 0);
390 // Other users may use these bits.
391 if (!Op.getNode()->hasOneUse()) {
393 // If not at the root, Just compute the KnownZero/KnownOne bits to
394 // simplify things downstream.
395 TLO.DAG.computeKnownBits(Op, KnownZero, KnownOne, Depth);
398 // If this is the root being simplified, allow it to have multiple uses,
399 // just set the NewMask to all bits.
400 NewMask = APInt::getAllOnesValue(BitWidth);
401 } else if (DemandedMask == 0) {
402 // Not demanding any bits from Op.
403 if (Op.getOpcode() != ISD::UNDEF)
404 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType()));
406 } else if (Depth == 6) { // Limit search depth.
410 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
411 switch (Op.getOpcode()) {
413 // We know all of the bits for a constant!
414 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue();
415 KnownZero = ~KnownOne;
416 return false; // Don't fall through, will infinitely loop.
418 // If the RHS is a constant, check to see if the LHS would be zero without
419 // using the bits from the RHS. Below, we use knowledge about the RHS to
420 // simplify the LHS, here we're using information from the LHS to simplify
422 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
423 APInt LHSZero, LHSOne;
424 // Do not increment Depth here; that can cause an infinite loop.
425 TLO.DAG.computeKnownBits(Op.getOperand(0), LHSZero, LHSOne, Depth);
426 // If the LHS already has zeros where RHSC does, this and is dead.
427 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
428 return TLO.CombineTo(Op, Op.getOperand(0));
429 // If any of the set bits in the RHS are known zero on the LHS, shrink
431 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask))
435 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
436 KnownOne, TLO, Depth+1))
438 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
439 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask,
440 KnownZero2, KnownOne2, TLO, Depth+1))
442 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
444 // If all of the demanded bits are known one on one side, return the other.
445 // These bits cannot contribute to the result of the 'and'.
446 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
447 return TLO.CombineTo(Op, Op.getOperand(0));
448 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
449 return TLO.CombineTo(Op, Op.getOperand(1));
450 // If all of the demanded bits in the inputs are known zeros, return zero.
451 if ((NewMask & (KnownZero|KnownZero2)) == NewMask)
452 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, dl, Op.getValueType()));
453 // If the RHS is a constant, see if we can simplify it.
454 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask))
456 // If the operation can be done in a smaller type, do so.
457 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
460 // Output known-1 bits are only known if set in both the LHS & RHS.
461 KnownOne &= KnownOne2;
462 // Output known-0 are known to be clear if zero in either the LHS | RHS.
463 KnownZero |= KnownZero2;
466 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
467 KnownOne, TLO, Depth+1))
469 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
470 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask,
471 KnownZero2, KnownOne2, TLO, Depth+1))
473 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
475 // If all of the demanded bits are known zero on one side, return the other.
476 // These bits cannot contribute to the result of the 'or'.
477 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask))
478 return TLO.CombineTo(Op, Op.getOperand(0));
479 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask))
480 return TLO.CombineTo(Op, Op.getOperand(1));
481 // If all of the potentially set bits on one side are known to be set on
482 // the other side, just use the 'other' side.
483 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
484 return TLO.CombineTo(Op, Op.getOperand(0));
485 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
486 return TLO.CombineTo(Op, Op.getOperand(1));
487 // If the RHS is a constant, see if we can simplify it.
488 if (TLO.ShrinkDemandedConstant(Op, NewMask))
490 // If the operation can be done in a smaller type, do so.
491 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
494 // Output known-0 bits are only known if clear in both the LHS & RHS.
495 KnownZero &= KnownZero2;
496 // Output known-1 are known to be set if set in either the LHS | RHS.
497 KnownOne |= KnownOne2;
500 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
501 KnownOne, TLO, Depth+1))
503 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
504 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2,
505 KnownOne2, TLO, Depth+1))
507 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
509 // If all of the demanded bits are known zero on one side, return the other.
510 // These bits cannot contribute to the result of the 'xor'.
511 if ((KnownZero & NewMask) == NewMask)
512 return TLO.CombineTo(Op, Op.getOperand(0));
513 if ((KnownZero2 & NewMask) == NewMask)
514 return TLO.CombineTo(Op, Op.getOperand(1));
515 // If the operation can be done in a smaller type, do so.
516 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
519 // If all of the unknown bits are known to be zero on one side or the other
520 // (but not both) turn this into an *inclusive* or.
521 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
522 if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
523 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(),
527 // Output known-0 bits are known if clear or set in both the LHS & RHS.
528 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
529 // Output known-1 are known to be set if set in only one of the LHS, RHS.
530 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
532 // If all of the demanded bits on one side are known, and all of the set
533 // bits on that side are also known to be set on the other side, turn this
534 // into an AND, as we know the bits will be cleared.
535 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
536 // NB: it is okay if more bits are known than are requested
537 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known on one side
538 if (KnownOne == KnownOne2) { // set bits are the same on both sides
539 EVT VT = Op.getValueType();
540 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, dl, VT);
541 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT,
542 Op.getOperand(0), ANDC));
546 // If the RHS is a constant, see if we can simplify it.
547 // for XOR, we prefer to force bits to 1 if they will make a -1.
548 // if we can't force bits, try to shrink constant
549 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
550 APInt Expanded = C->getAPIntValue() | (~NewMask);
551 // if we can expand it to have all bits set, do it
552 if (Expanded.isAllOnesValue()) {
553 if (Expanded != C->getAPIntValue()) {
554 EVT VT = Op.getValueType();
555 SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0),
556 TLO.DAG.getConstant(Expanded, dl, VT));
557 return TLO.CombineTo(Op, New);
559 // if it already has all the bits set, nothing to change
560 // but don't shrink either!
561 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) {
566 KnownZero = KnownZeroOut;
567 KnownOne = KnownOneOut;
570 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero,
571 KnownOne, TLO, Depth+1))
573 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2,
574 KnownOne2, TLO, Depth+1))
576 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
577 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
579 // If the operands are constants, see if we can simplify them.
580 if (TLO.ShrinkDemandedConstant(Op, NewMask))
583 // Only known if known in both the LHS and RHS.
584 KnownOne &= KnownOne2;
585 KnownZero &= KnownZero2;
588 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero,
589 KnownOne, TLO, Depth+1))
591 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2,
592 KnownOne2, TLO, Depth+1))
594 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
595 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
597 // If the operands are constants, see if we can simplify them.
598 if (TLO.ShrinkDemandedConstant(Op, NewMask))
601 // Only known if known in both the LHS and RHS.
602 KnownOne &= KnownOne2;
603 KnownZero &= KnownZero2;
606 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
607 unsigned ShAmt = SA->getZExtValue();
608 SDValue InOp = Op.getOperand(0);
610 // If the shift count is an invalid immediate, don't do anything.
611 if (ShAmt >= BitWidth)
614 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
615 // single shift. We can do this if the bottom bits (which are shifted
616 // out) are never demanded.
617 if (InOp.getOpcode() == ISD::SRL &&
618 isa<ConstantSDNode>(InOp.getOperand(1))) {
619 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
620 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
621 unsigned Opc = ISD::SHL;
629 TLO.DAG.getConstant(Diff, dl, Op.getOperand(1).getValueType());
630 EVT VT = Op.getValueType();
631 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
632 InOp.getOperand(0), NewSA));
636 if (SimplifyDemandedBits(InOp, NewMask.lshr(ShAmt),
637 KnownZero, KnownOne, TLO, Depth+1))
640 // Convert (shl (anyext x, c)) to (anyext (shl x, c)) if the high bits
641 // are not demanded. This will likely allow the anyext to be folded away.
642 if (InOp.getNode()->getOpcode() == ISD::ANY_EXTEND) {
643 SDValue InnerOp = InOp.getNode()->getOperand(0);
644 EVT InnerVT = InnerOp.getValueType();
645 unsigned InnerBits = InnerVT.getSizeInBits();
646 if (ShAmt < InnerBits && NewMask.lshr(InnerBits) == 0 &&
647 isTypeDesirableForOp(ISD::SHL, InnerVT)) {
648 EVT ShTy = getShiftAmountTy(InnerVT);
649 if (!APInt(BitWidth, ShAmt).isIntN(ShTy.getSizeInBits()))
652 TLO.DAG.getNode(ISD::SHL, dl, InnerVT, InnerOp,
653 TLO.DAG.getConstant(ShAmt, dl, ShTy));
656 TLO.DAG.getNode(ISD::ANY_EXTEND, dl, Op.getValueType(),
659 // Repeat the SHL optimization above in cases where an extension
660 // intervenes: (shl (anyext (shr x, c1)), c2) to
661 // (shl (anyext x), c2-c1). This requires that the bottom c1 bits
662 // aren't demanded (as above) and that the shifted upper c1 bits of
663 // x aren't demanded.
664 if (InOp.hasOneUse() &&
665 InnerOp.getOpcode() == ISD::SRL &&
666 InnerOp.hasOneUse() &&
667 isa<ConstantSDNode>(InnerOp.getOperand(1))) {
668 uint64_t InnerShAmt = cast<ConstantSDNode>(InnerOp.getOperand(1))
670 if (InnerShAmt < ShAmt &&
671 InnerShAmt < InnerBits &&
672 NewMask.lshr(InnerBits - InnerShAmt + ShAmt) == 0 &&
673 NewMask.trunc(ShAmt) == 0) {
675 TLO.DAG.getConstant(ShAmt - InnerShAmt, dl,
676 Op.getOperand(1).getValueType());
677 EVT VT = Op.getValueType();
678 SDValue NewExt = TLO.DAG.getNode(ISD::ANY_EXTEND, dl, VT,
679 InnerOp.getOperand(0));
680 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl, VT,
686 KnownZero <<= SA->getZExtValue();
687 KnownOne <<= SA->getZExtValue();
688 // low bits known zero.
689 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue());
693 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
694 EVT VT = Op.getValueType();
695 unsigned ShAmt = SA->getZExtValue();
696 unsigned VTSize = VT.getSizeInBits();
697 SDValue InOp = Op.getOperand(0);
699 // If the shift count is an invalid immediate, don't do anything.
700 if (ShAmt >= BitWidth)
703 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
704 // single shift. We can do this if the top bits (which are shifted out)
705 // are never demanded.
706 if (InOp.getOpcode() == ISD::SHL &&
707 isa<ConstantSDNode>(InOp.getOperand(1))) {
708 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
709 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
710 unsigned Opc = ISD::SRL;
718 TLO.DAG.getConstant(Diff, dl, Op.getOperand(1).getValueType());
719 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
720 InOp.getOperand(0), NewSA));
724 // Compute the new bits that are at the top now.
725 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt),
726 KnownZero, KnownOne, TLO, Depth+1))
728 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
729 KnownZero = KnownZero.lshr(ShAmt);
730 KnownOne = KnownOne.lshr(ShAmt);
732 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
733 KnownZero |= HighBits; // High bits known zero.
737 // If this is an arithmetic shift right and only the low-bit is set, we can
738 // always convert this into a logical shr, even if the shift amount is
739 // variable. The low bit of the shift cannot be an input sign bit unless
740 // the shift amount is >= the size of the datatype, which is undefined.
742 return TLO.CombineTo(Op,
743 TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(),
744 Op.getOperand(0), Op.getOperand(1)));
746 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
747 EVT VT = Op.getValueType();
748 unsigned ShAmt = SA->getZExtValue();
750 // If the shift count is an invalid immediate, don't do anything.
751 if (ShAmt >= BitWidth)
754 APInt InDemandedMask = (NewMask << ShAmt);
756 // If any of the demanded bits are produced by the sign extension, we also
757 // demand the input sign bit.
758 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
759 if (HighBits.intersects(NewMask))
760 InDemandedMask |= APInt::getSignBit(VT.getScalarType().getSizeInBits());
762 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
763 KnownZero, KnownOne, TLO, Depth+1))
765 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
766 KnownZero = KnownZero.lshr(ShAmt);
767 KnownOne = KnownOne.lshr(ShAmt);
769 // Handle the sign bit, adjusted to where it is now in the mask.
770 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt);
772 // If the input sign bit is known to be zero, or if none of the top bits
773 // are demanded, turn this into an unsigned shift right.
774 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
776 Flags.setExact(cast<BinaryWithFlagsSDNode>(Op)->Flags.hasExact());
777 return TLO.CombineTo(Op,
778 TLO.DAG.getNode(ISD::SRL, dl, VT, Op.getOperand(0),
779 Op.getOperand(1), &Flags));
782 int Log2 = NewMask.exactLogBase2();
784 // The bit must come from the sign.
786 TLO.DAG.getConstant(BitWidth - 1 - Log2, dl,
787 Op.getOperand(1).getValueType());
788 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT,
789 Op.getOperand(0), NewSA));
792 if (KnownOne.intersects(SignBit))
793 // New bits are known one.
794 KnownOne |= HighBits;
797 case ISD::SIGN_EXTEND_INREG: {
798 EVT ExVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
800 APInt MsbMask = APInt::getHighBitsSet(BitWidth, 1);
801 // If we only care about the highest bit, don't bother shifting right.
802 if (MsbMask == NewMask) {
803 unsigned ShAmt = ExVT.getScalarType().getSizeInBits();
804 SDValue InOp = Op.getOperand(0);
805 unsigned VTBits = Op->getValueType(0).getScalarType().getSizeInBits();
806 bool AlreadySignExtended =
807 TLO.DAG.ComputeNumSignBits(InOp) >= VTBits-ShAmt+1;
808 // However if the input is already sign extended we expect the sign
809 // extension to be dropped altogether later and do not simplify.
810 if (!AlreadySignExtended) {
811 // Compute the correct shift amount type, which must be getShiftAmountTy
812 // for scalar types after legalization.
813 EVT ShiftAmtTy = Op.getValueType();
814 if (TLO.LegalTypes() && !ShiftAmtTy.isVector())
815 ShiftAmtTy = getShiftAmountTy(ShiftAmtTy);
817 SDValue ShiftAmt = TLO.DAG.getConstant(BitWidth - ShAmt, dl,
819 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl,
820 Op.getValueType(), InOp,
825 // Sign extension. Compute the demanded bits in the result that are not
826 // present in the input.
828 APInt::getHighBitsSet(BitWidth,
829 BitWidth - ExVT.getScalarType().getSizeInBits());
831 // If none of the extended bits are demanded, eliminate the sextinreg.
832 if ((NewBits & NewMask) == 0)
833 return TLO.CombineTo(Op, Op.getOperand(0));
836 APInt::getSignBit(ExVT.getScalarType().getSizeInBits()).zext(BitWidth);
837 APInt InputDemandedBits =
838 APInt::getLowBitsSet(BitWidth,
839 ExVT.getScalarType().getSizeInBits()) &
842 // Since the sign extended bits are demanded, we know that the sign
844 InputDemandedBits |= InSignBit;
846 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
847 KnownZero, KnownOne, TLO, Depth+1))
849 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
851 // If the sign bit of the input is known set or clear, then we know the
852 // top bits of the result.
854 // If the input sign bit is known zero, convert this into a zero extension.
855 if (KnownZero.intersects(InSignBit))
856 return TLO.CombineTo(Op,
857 TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,ExVT));
859 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
861 KnownZero &= ~NewBits;
862 } else { // Input sign bit unknown
863 KnownZero &= ~NewBits;
864 KnownOne &= ~NewBits;
868 case ISD::BUILD_PAIR: {
869 EVT HalfVT = Op.getOperand(0).getValueType();
870 unsigned HalfBitWidth = HalfVT.getScalarSizeInBits();
872 APInt MaskLo = NewMask.getLoBits(HalfBitWidth).trunc(HalfBitWidth);
873 APInt MaskHi = NewMask.getHiBits(HalfBitWidth).trunc(HalfBitWidth);
875 APInt KnownZeroLo, KnownOneLo;
876 APInt KnownZeroHi, KnownOneHi;
878 if (SimplifyDemandedBits(Op.getOperand(0), MaskLo, KnownZeroLo,
879 KnownOneLo, TLO, Depth + 1))
882 if (SimplifyDemandedBits(Op.getOperand(1), MaskHi, KnownZeroHi,
883 KnownOneHi, TLO, Depth + 1))
886 KnownZero = KnownZeroLo.zext(BitWidth) |
887 KnownZeroHi.zext(BitWidth).shl(HalfBitWidth);
889 KnownOne = KnownOneLo.zext(BitWidth) |
890 KnownOneHi.zext(BitWidth).shl(HalfBitWidth);
893 case ISD::ZERO_EXTEND: {
894 unsigned OperandBitWidth =
895 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
896 APInt InMask = NewMask.trunc(OperandBitWidth);
898 // If none of the top bits are demanded, convert this into an any_extend.
900 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
901 if (!NewBits.intersects(NewMask))
902 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
906 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
907 KnownZero, KnownOne, TLO, Depth+1))
909 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
910 KnownZero = KnownZero.zext(BitWidth);
911 KnownOne = KnownOne.zext(BitWidth);
912 KnownZero |= NewBits;
915 case ISD::SIGN_EXTEND: {
916 EVT InVT = Op.getOperand(0).getValueType();
917 unsigned InBits = InVT.getScalarType().getSizeInBits();
918 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
919 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits);
920 APInt NewBits = ~InMask & NewMask;
922 // If none of the top bits are demanded, convert this into an any_extend.
924 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
928 // Since some of the sign extended bits are demanded, we know that the sign
930 APInt InDemandedBits = InMask & NewMask;
931 InDemandedBits |= InSignBit;
932 InDemandedBits = InDemandedBits.trunc(InBits);
934 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
935 KnownOne, TLO, Depth+1))
937 KnownZero = KnownZero.zext(BitWidth);
938 KnownOne = KnownOne.zext(BitWidth);
940 // If the sign bit is known zero, convert this to a zero extend.
941 if (KnownZero.intersects(InSignBit))
942 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl,
946 // If the sign bit is known one, the top bits match.
947 if (KnownOne.intersects(InSignBit)) {
949 assert((KnownZero & NewBits) == 0);
950 } else { // Otherwise, top bits aren't known.
951 assert((KnownOne & NewBits) == 0);
952 assert((KnownZero & NewBits) == 0);
956 case ISD::ANY_EXTEND: {
957 unsigned OperandBitWidth =
958 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
959 APInt InMask = NewMask.trunc(OperandBitWidth);
960 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
961 KnownZero, KnownOne, TLO, Depth+1))
963 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
964 KnownZero = KnownZero.zext(BitWidth);
965 KnownOne = KnownOne.zext(BitWidth);
968 case ISD::TRUNCATE: {
969 // Simplify the input, using demanded bit information, and compute the known
970 // zero/one bits live out.
971 unsigned OperandBitWidth =
972 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
973 APInt TruncMask = NewMask.zext(OperandBitWidth);
974 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask,
975 KnownZero, KnownOne, TLO, Depth+1))
977 KnownZero = KnownZero.trunc(BitWidth);
978 KnownOne = KnownOne.trunc(BitWidth);
980 // If the input is only used by this truncate, see if we can shrink it based
981 // on the known demanded bits.
982 if (Op.getOperand(0).getNode()->hasOneUse()) {
983 SDValue In = Op.getOperand(0);
984 switch (In.getOpcode()) {
987 // Shrink SRL by a constant if none of the high bits shifted in are
989 if (TLO.LegalTypes() &&
990 !isTypeDesirableForOp(ISD::SRL, Op.getValueType()))
991 // Do not turn (vt1 truncate (vt2 srl)) into (vt1 srl) if vt1 is
994 ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1));
997 SDValue Shift = In.getOperand(1);
998 if (TLO.LegalTypes()) {
999 uint64_t ShVal = ShAmt->getZExtValue();
1001 TLO.DAG.getConstant(ShVal, dl, getShiftAmountTy(Op.getValueType()));
1004 APInt HighBits = APInt::getHighBitsSet(OperandBitWidth,
1005 OperandBitWidth - BitWidth);
1006 HighBits = HighBits.lshr(ShAmt->getZExtValue()).trunc(BitWidth);
1008 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) {
1009 // None of the shifted in bits are needed. Add a truncate of the
1010 // shift input, then shift it.
1011 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl,
1014 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl,
1023 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1026 case ISD::AssertZext: {
1027 // AssertZext demands all of the high bits, plus any of the low bits
1028 // demanded by its users.
1029 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1030 APInt InMask = APInt::getLowBitsSet(BitWidth,
1031 VT.getSizeInBits());
1032 if (SimplifyDemandedBits(Op.getOperand(0), ~InMask | NewMask,
1033 KnownZero, KnownOne, TLO, Depth+1))
1035 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1037 KnownZero |= ~InMask & NewMask;
1041 // If this is an FP->Int bitcast and if the sign bit is the only
1042 // thing demanded, turn this into a FGETSIGN.
1043 if (!TLO.LegalOperations() &&
1044 !Op.getValueType().isVector() &&
1045 !Op.getOperand(0).getValueType().isVector() &&
1046 NewMask == APInt::getSignBit(Op.getValueType().getSizeInBits()) &&
1047 Op.getOperand(0).getValueType().isFloatingPoint()) {
1048 bool OpVTLegal = isOperationLegalOrCustom(ISD::FGETSIGN, Op.getValueType());
1049 bool i32Legal = isOperationLegalOrCustom(ISD::FGETSIGN, MVT::i32);
1050 if ((OpVTLegal || i32Legal) && Op.getValueType().isSimple()) {
1051 EVT Ty = OpVTLegal ? Op.getValueType() : MVT::i32;
1052 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
1053 // place. We expect the SHL to be eliminated by other optimizations.
1054 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, dl, Ty, Op.getOperand(0));
1055 unsigned OpVTSizeInBits = Op.getValueType().getSizeInBits();
1056 if (!OpVTLegal && OpVTSizeInBits > 32)
1057 Sign = TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), Sign);
1058 unsigned ShVal = Op.getValueType().getSizeInBits()-1;
1059 SDValue ShAmt = TLO.DAG.getConstant(ShVal, dl, Op.getValueType());
1060 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl,
1069 // Add, Sub, and Mul don't demand any bits in positions beyond that
1070 // of the highest bit demanded of them.
1071 APInt LoMask = APInt::getLowBitsSet(BitWidth,
1072 BitWidth - NewMask.countLeadingZeros());
1073 if (SimplifyDemandedBits(Op.getOperand(0), LoMask, KnownZero2,
1074 KnownOne2, TLO, Depth+1))
1076 if (SimplifyDemandedBits(Op.getOperand(1), LoMask, KnownZero2,
1077 KnownOne2, TLO, Depth+1))
1079 // See if the operation should be performed at a smaller bit width.
1080 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
1085 // Just use computeKnownBits to compute output bits.
1086 TLO.DAG.computeKnownBits(Op, KnownZero, KnownOne, Depth);
1090 // If we know the value of all of the demanded bits, return this as a
1092 if ((NewMask & (KnownZero|KnownOne)) == NewMask) {
1093 // Avoid folding to a constant if any OpaqueConstant is involved.
1094 const SDNode *N = Op.getNode();
1095 for (SDNodeIterator I = SDNodeIterator::begin(N),
1096 E = SDNodeIterator::end(N); I != E; ++I) {
1098 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
1102 return TLO.CombineTo(Op,
1103 TLO.DAG.getConstant(KnownOne, dl, Op.getValueType()));
1109 /// computeKnownBitsForTargetNode - Determine which of the bits specified
1110 /// in Mask are known to be either zero or one and return them in the
1111 /// KnownZero/KnownOne bitsets.
1112 void TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
1115 const SelectionDAG &DAG,
1116 unsigned Depth) const {
1117 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1118 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1119 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1120 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1121 "Should use MaskedValueIsZero if you don't know whether Op"
1122 " is a target node!");
1123 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0);
1126 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
1127 /// targets that want to expose additional information about sign bits to the
1129 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
1130 const SelectionDAG &,
1131 unsigned Depth) const {
1132 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1133 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1134 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1135 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1136 "Should use ComputeNumSignBits if you don't know whether Op"
1137 " is a target node!");
1141 /// ValueHasExactlyOneBitSet - Test if the given value is known to have exactly
1142 /// one bit set. This differs from computeKnownBits in that it doesn't need to
1143 /// determine which bit is set.
1145 static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) {
1146 // A left-shift of a constant one will have exactly one bit set, because
1147 // shifting the bit off the end is undefined.
1148 if (Val.getOpcode() == ISD::SHL)
1149 if (ConstantSDNode *C =
1150 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1151 if (C->getAPIntValue() == 1)
1154 // Similarly, a right-shift of a constant sign-bit will have exactly
1156 if (Val.getOpcode() == ISD::SRL)
1157 if (ConstantSDNode *C =
1158 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1159 if (C->getAPIntValue().isSignBit())
1162 // More could be done here, though the above checks are enough
1163 // to handle some common cases.
1165 // Fall back to computeKnownBits to catch other known cases.
1166 EVT OpVT = Val.getValueType();
1167 unsigned BitWidth = OpVT.getScalarType().getSizeInBits();
1168 APInt KnownZero, KnownOne;
1169 DAG.computeKnownBits(Val, KnownZero, KnownOne);
1170 return (KnownZero.countPopulation() == BitWidth - 1) &&
1171 (KnownOne.countPopulation() == 1);
1174 bool TargetLowering::isConstTrueVal(const SDNode *N) const {
1178 const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N);
1180 const BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N);
1184 BitVector UndefElements;
1185 CN = BV->getConstantSplatNode(&UndefElements);
1186 // Only interested in constant splats, and we don't try to handle undef
1187 // elements in identifying boolean constants.
1188 if (!CN || UndefElements.none())
1192 switch (getBooleanContents(N->getValueType(0))) {
1193 case UndefinedBooleanContent:
1194 return CN->getAPIntValue()[0];
1195 case ZeroOrOneBooleanContent:
1197 case ZeroOrNegativeOneBooleanContent:
1198 return CN->isAllOnesValue();
1201 llvm_unreachable("Invalid boolean contents");
1204 bool TargetLowering::isConstFalseVal(const SDNode *N) const {
1208 const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N);
1210 const BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N);
1214 BitVector UndefElements;
1215 CN = BV->getConstantSplatNode(&UndefElements);
1216 // Only interested in constant splats, and we don't try to handle undef
1217 // elements in identifying boolean constants.
1218 if (!CN || UndefElements.none())
1222 if (getBooleanContents(N->getValueType(0)) == UndefinedBooleanContent)
1223 return !CN->getAPIntValue()[0];
1225 return CN->isNullValue();
1228 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
1229 /// and cc. If it is unable to simplify it, return a null SDValue.
1231 TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
1232 ISD::CondCode Cond, bool foldBooleans,
1233 DAGCombinerInfo &DCI, SDLoc dl) const {
1234 SelectionDAG &DAG = DCI.DAG;
1236 // These setcc operations always fold.
1240 case ISD::SETFALSE2: return DAG.getConstant(0, dl, VT);
1242 case ISD::SETTRUE2: {
1243 TargetLowering::BooleanContent Cnt =
1244 getBooleanContents(N0->getValueType(0));
1245 return DAG.getConstant(
1246 Cnt == TargetLowering::ZeroOrNegativeOneBooleanContent ? -1ULL : 1, dl,
1251 // Ensure that the constant occurs on the RHS, and fold constant
1253 ISD::CondCode SwappedCC = ISD::getSetCCSwappedOperands(Cond);
1254 if (isa<ConstantSDNode>(N0.getNode()) &&
1255 (DCI.isBeforeLegalizeOps() ||
1256 isCondCodeLegal(SwappedCC, N0.getSimpleValueType())))
1257 return DAG.getSetCC(dl, VT, N1, N0, SwappedCC);
1259 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1260 const APInt &C1 = N1C->getAPIntValue();
1262 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1263 // equality comparison, then we're just comparing whether X itself is
1265 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1266 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1267 N0.getOperand(1).getOpcode() == ISD::Constant) {
1269 = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
1270 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1271 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) {
1272 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1273 // (srl (ctlz x), 5) == 0 -> X != 0
1274 // (srl (ctlz x), 5) != 1 -> X != 0
1277 // (srl (ctlz x), 5) != 0 -> X == 0
1278 // (srl (ctlz x), 5) == 1 -> X == 0
1281 SDValue Zero = DAG.getConstant(0, dl, N0.getValueType());
1282 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0),
1288 // Look through truncs that don't change the value of a ctpop.
1289 if (N0.hasOneUse() && N0.getOpcode() == ISD::TRUNCATE)
1290 CTPOP = N0.getOperand(0);
1292 if (CTPOP.hasOneUse() && CTPOP.getOpcode() == ISD::CTPOP &&
1293 (N0 == CTPOP || N0.getValueType().getSizeInBits() >
1294 Log2_32_Ceil(CTPOP.getValueType().getSizeInBits()))) {
1295 EVT CTVT = CTPOP.getValueType();
1296 SDValue CTOp = CTPOP.getOperand(0);
1298 // (ctpop x) u< 2 -> (x & x-1) == 0
1299 // (ctpop x) u> 1 -> (x & x-1) != 0
1300 if ((Cond == ISD::SETULT && C1 == 2) || (Cond == ISD::SETUGT && C1 == 1)){
1301 SDValue Sub = DAG.getNode(ISD::SUB, dl, CTVT, CTOp,
1302 DAG.getConstant(1, dl, CTVT));
1303 SDValue And = DAG.getNode(ISD::AND, dl, CTVT, CTOp, Sub);
1304 ISD::CondCode CC = Cond == ISD::SETULT ? ISD::SETEQ : ISD::SETNE;
1305 return DAG.getSetCC(dl, VT, And, DAG.getConstant(0, dl, CTVT), CC);
1308 // TODO: (ctpop x) == 1 -> x && (x & x-1) == 0 iff ctpop is illegal.
1311 // (zext x) == C --> x == (trunc C)
1312 // (sext x) == C --> x == (trunc C)
1313 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1314 DCI.isBeforeLegalize() && N0->hasOneUse()) {
1315 unsigned MinBits = N0.getValueSizeInBits();
1317 bool Signed = false;
1318 if (N0->getOpcode() == ISD::ZERO_EXTEND) {
1320 MinBits = N0->getOperand(0).getValueSizeInBits();
1321 PreExt = N0->getOperand(0);
1322 } else if (N0->getOpcode() == ISD::AND) {
1323 // DAGCombine turns costly ZExts into ANDs
1324 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0->getOperand(1)))
1325 if ((C->getAPIntValue()+1).isPowerOf2()) {
1326 MinBits = C->getAPIntValue().countTrailingOnes();
1327 PreExt = N0->getOperand(0);
1329 } else if (N0->getOpcode() == ISD::SIGN_EXTEND) {
1331 MinBits = N0->getOperand(0).getValueSizeInBits();
1332 PreExt = N0->getOperand(0);
1334 } else if (LoadSDNode *LN0 = dyn_cast<LoadSDNode>(N0)) {
1335 // ZEXTLOAD / SEXTLOAD
1336 if (LN0->getExtensionType() == ISD::ZEXTLOAD) {
1337 MinBits = LN0->getMemoryVT().getSizeInBits();
1339 } else if (LN0->getExtensionType() == ISD::SEXTLOAD) {
1341 MinBits = LN0->getMemoryVT().getSizeInBits();
1346 // Figure out how many bits we need to preserve this constant.
1347 unsigned ReqdBits = Signed ?
1348 C1.getBitWidth() - C1.getNumSignBits() + 1 :
1351 // Make sure we're not losing bits from the constant.
1353 MinBits < C1.getBitWidth() &&
1354 MinBits >= ReqdBits) {
1355 EVT MinVT = EVT::getIntegerVT(*DAG.getContext(), MinBits);
1356 if (isTypeDesirableForOp(ISD::SETCC, MinVT)) {
1357 // Will get folded away.
1358 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MinVT, PreExt);
1359 SDValue C = DAG.getConstant(C1.trunc(MinBits), dl, MinVT);
1360 return DAG.getSetCC(dl, VT, Trunc, C, Cond);
1365 // If the LHS is '(and load, const)', the RHS is 0,
1366 // the test is for equality or unsigned, and all 1 bits of the const are
1367 // in the same partial word, see if we can shorten the load.
1368 if (DCI.isBeforeLegalize() &&
1369 !ISD::isSignedIntSetCC(Cond) &&
1370 N0.getOpcode() == ISD::AND && C1 == 0 &&
1371 N0.getNode()->hasOneUse() &&
1372 isa<LoadSDNode>(N0.getOperand(0)) &&
1373 N0.getOperand(0).getNode()->hasOneUse() &&
1374 isa<ConstantSDNode>(N0.getOperand(1))) {
1375 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0));
1377 unsigned bestWidth = 0, bestOffset = 0;
1378 if (!Lod->isVolatile() && Lod->isUnindexed()) {
1379 unsigned origWidth = N0.getValueType().getSizeInBits();
1380 unsigned maskWidth = origWidth;
1381 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to
1382 // 8 bits, but have to be careful...
1383 if (Lod->getExtensionType() != ISD::NON_EXTLOAD)
1384 origWidth = Lod->getMemoryVT().getSizeInBits();
1386 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
1387 for (unsigned width = origWidth / 2; width>=8; width /= 2) {
1388 APInt newMask = APInt::getLowBitsSet(maskWidth, width);
1389 for (unsigned offset=0; offset<origWidth/width; offset++) {
1390 if ((newMask & Mask) == Mask) {
1391 if (!getDataLayout()->isLittleEndian())
1392 bestOffset = (origWidth/width - offset - 1) * (width/8);
1394 bestOffset = (uint64_t)offset * (width/8);
1395 bestMask = Mask.lshr(offset * (width/8) * 8);
1399 newMask = newMask << width;
1404 EVT newVT = EVT::getIntegerVT(*DAG.getContext(), bestWidth);
1405 if (newVT.isRound()) {
1406 EVT PtrType = Lod->getOperand(1).getValueType();
1407 SDValue Ptr = Lod->getBasePtr();
1408 if (bestOffset != 0)
1409 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(),
1410 DAG.getConstant(bestOffset, dl, PtrType));
1411 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset);
1412 SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr,
1413 Lod->getPointerInfo().getWithOffset(bestOffset),
1414 false, false, false, NewAlign);
1415 return DAG.getSetCC(dl, VT,
1416 DAG.getNode(ISD::AND, dl, newVT, NewLoad,
1417 DAG.getConstant(bestMask.trunc(bestWidth),
1419 DAG.getConstant(0LL, dl, newVT), Cond);
1424 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
1425 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
1426 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits();
1428 // If the comparison constant has bits in the upper part, the
1429 // zero-extended value could never match.
1430 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
1431 C1.getBitWidth() - InSize))) {
1435 case ISD::SETEQ: return DAG.getConstant(0, dl, VT);
1438 case ISD::SETNE: return DAG.getConstant(1, dl, VT);
1441 // True if the sign bit of C1 is set.
1442 return DAG.getConstant(C1.isNegative(), dl, VT);
1445 // True if the sign bit of C1 isn't set.
1446 return DAG.getConstant(C1.isNonNegative(), dl, VT);
1452 // Otherwise, we can perform the comparison with the low bits.
1460 EVT newVT = N0.getOperand(0).getValueType();
1461 if (DCI.isBeforeLegalizeOps() ||
1462 (isOperationLegal(ISD::SETCC, newVT) &&
1463 getCondCodeAction(Cond, newVT.getSimpleVT()) == Legal)) {
1464 EVT NewSetCCVT = getSetCCResultType(*DAG.getContext(), newVT);
1465 SDValue NewConst = DAG.getConstant(C1.trunc(InSize), dl, newVT);
1467 SDValue NewSetCC = DAG.getSetCC(dl, NewSetCCVT, N0.getOperand(0),
1469 return DAG.getBoolExtOrTrunc(NewSetCC, dl, VT, N0.getValueType());
1474 break; // todo, be more careful with signed comparisons
1476 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1477 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1478 EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
1479 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
1480 EVT ExtDstTy = N0.getValueType();
1481 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
1483 // If the constant doesn't fit into the number of bits for the source of
1484 // the sign extension, it is impossible for both sides to be equal.
1485 if (C1.getMinSignedBits() > ExtSrcTyBits)
1486 return DAG.getConstant(Cond == ISD::SETNE, dl, VT);
1489 EVT Op0Ty = N0.getOperand(0).getValueType();
1490 if (Op0Ty == ExtSrcTy) {
1491 ZextOp = N0.getOperand(0);
1493 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
1494 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0),
1495 DAG.getConstant(Imm, dl, Op0Ty));
1497 if (!DCI.isCalledByLegalizer())
1498 DCI.AddToWorklist(ZextOp.getNode());
1499 // Otherwise, make this a use of a zext.
1500 return DAG.getSetCC(dl, VT, ZextOp,
1501 DAG.getConstant(C1 & APInt::getLowBitsSet(
1506 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) &&
1507 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1508 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
1509 if (N0.getOpcode() == ISD::SETCC &&
1510 isTypeLegal(VT) && VT.bitsLE(N0.getValueType())) {
1511 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getAPIntValue() != 1);
1513 return DAG.getNode(ISD::TRUNCATE, dl, VT, N0);
1514 // Invert the condition.
1515 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
1516 CC = ISD::getSetCCInverse(CC,
1517 N0.getOperand(0).getValueType().isInteger());
1518 if (DCI.isBeforeLegalizeOps() ||
1519 isCondCodeLegal(CC, N0.getOperand(0).getSimpleValueType()))
1520 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC);
1523 if ((N0.getOpcode() == ISD::XOR ||
1524 (N0.getOpcode() == ISD::AND &&
1525 N0.getOperand(0).getOpcode() == ISD::XOR &&
1526 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
1527 isa<ConstantSDNode>(N0.getOperand(1)) &&
1528 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) {
1529 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
1530 // can only do this if the top bits are known zero.
1531 unsigned BitWidth = N0.getValueSizeInBits();
1532 if (DAG.MaskedValueIsZero(N0,
1533 APInt::getHighBitsSet(BitWidth,
1535 // Okay, get the un-inverted input value.
1537 if (N0.getOpcode() == ISD::XOR)
1538 Val = N0.getOperand(0);
1540 assert(N0.getOpcode() == ISD::AND &&
1541 N0.getOperand(0).getOpcode() == ISD::XOR);
1542 // ((X^1)&1)^1 -> X & 1
1543 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(),
1544 N0.getOperand(0).getOperand(0),
1548 return DAG.getSetCC(dl, VT, Val, N1,
1549 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1551 } else if (N1C->getAPIntValue() == 1 &&
1553 getBooleanContents(N0->getValueType(0)) ==
1554 ZeroOrOneBooleanContent)) {
1556 if (Op0.getOpcode() == ISD::TRUNCATE)
1557 Op0 = Op0.getOperand(0);
1559 if ((Op0.getOpcode() == ISD::XOR) &&
1560 Op0.getOperand(0).getOpcode() == ISD::SETCC &&
1561 Op0.getOperand(1).getOpcode() == ISD::SETCC) {
1562 // (xor (setcc), (setcc)) == / != 1 -> (setcc) != / == (setcc)
1563 Cond = (Cond == ISD::SETEQ) ? ISD::SETNE : ISD::SETEQ;
1564 return DAG.getSetCC(dl, VT, Op0.getOperand(0), Op0.getOperand(1),
1567 if (Op0.getOpcode() == ISD::AND &&
1568 isa<ConstantSDNode>(Op0.getOperand(1)) &&
1569 cast<ConstantSDNode>(Op0.getOperand(1))->getAPIntValue() == 1) {
1570 // If this is (X&1) == / != 1, normalize it to (X&1) != / == 0.
1571 if (Op0.getValueType().bitsGT(VT))
1572 Op0 = DAG.getNode(ISD::AND, dl, VT,
1573 DAG.getNode(ISD::TRUNCATE, dl, VT, Op0.getOperand(0)),
1574 DAG.getConstant(1, dl, VT));
1575 else if (Op0.getValueType().bitsLT(VT))
1576 Op0 = DAG.getNode(ISD::AND, dl, VT,
1577 DAG.getNode(ISD::ANY_EXTEND, dl, VT, Op0.getOperand(0)),
1578 DAG.getConstant(1, dl, VT));
1580 return DAG.getSetCC(dl, VT, Op0,
1581 DAG.getConstant(0, dl, Op0.getValueType()),
1582 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1584 if (Op0.getOpcode() == ISD::AssertZext &&
1585 cast<VTSDNode>(Op0.getOperand(1))->getVT() == MVT::i1)
1586 return DAG.getSetCC(dl, VT, Op0,
1587 DAG.getConstant(0, dl, Op0.getValueType()),
1588 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1592 APInt MinVal, MaxVal;
1593 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits();
1594 if (ISD::isSignedIntSetCC(Cond)) {
1595 MinVal = APInt::getSignedMinValue(OperandBitSize);
1596 MaxVal = APInt::getSignedMaxValue(OperandBitSize);
1598 MinVal = APInt::getMinValue(OperandBitSize);
1599 MaxVal = APInt::getMaxValue(OperandBitSize);
1602 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
1603 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
1604 if (C1 == MinVal) return DAG.getConstant(1, dl, VT); // X >= MIN --> true
1605 // X >= C0 --> X > (C0 - 1)
1607 ISD::CondCode NewCC = (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT;
1608 if ((DCI.isBeforeLegalizeOps() ||
1609 isCondCodeLegal(NewCC, VT.getSimpleVT())) &&
1610 (!N1C->isOpaque() || (N1C->isOpaque() && C.getBitWidth() <= 64 &&
1611 isLegalICmpImmediate(C.getSExtValue())))) {
1612 return DAG.getSetCC(dl, VT, N0,
1613 DAG.getConstant(C, dl, N1.getValueType()),
1618 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
1619 if (C1 == MaxVal) return DAG.getConstant(1, dl, VT); // X <= MAX --> true
1620 // X <= C0 --> X < (C0 + 1)
1622 ISD::CondCode NewCC = (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT;
1623 if ((DCI.isBeforeLegalizeOps() ||
1624 isCondCodeLegal(NewCC, VT.getSimpleVT())) &&
1625 (!N1C->isOpaque() || (N1C->isOpaque() && C.getBitWidth() <= 64 &&
1626 isLegalICmpImmediate(C.getSExtValue())))) {
1627 return DAG.getSetCC(dl, VT, N0,
1628 DAG.getConstant(C, dl, N1.getValueType()),
1633 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
1634 return DAG.getConstant(0, dl, VT); // X < MIN --> false
1635 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
1636 return DAG.getConstant(1, dl, VT); // X >= MIN --> true
1637 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
1638 return DAG.getConstant(0, dl, VT); // X > MAX --> false
1639 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
1640 return DAG.getConstant(1, dl, VT); // X <= MAX --> true
1642 // Canonicalize setgt X, Min --> setne X, Min
1643 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
1644 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
1645 // Canonicalize setlt X, Max --> setne X, Max
1646 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
1647 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
1649 // If we have setult X, 1, turn it into seteq X, 0
1650 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
1651 return DAG.getSetCC(dl, VT, N0,
1652 DAG.getConstant(MinVal, dl, N0.getValueType()),
1654 // If we have setugt X, Max-1, turn it into seteq X, Max
1655 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
1656 return DAG.getSetCC(dl, VT, N0,
1657 DAG.getConstant(MaxVal, dl, N0.getValueType()),
1660 // If we have "setcc X, C0", check to see if we can shrink the immediate
1663 // SETUGT X, SINTMAX -> SETLT X, 0
1664 if (Cond == ISD::SETUGT &&
1665 C1 == APInt::getSignedMaxValue(OperandBitSize))
1666 return DAG.getSetCC(dl, VT, N0,
1667 DAG.getConstant(0, dl, N1.getValueType()),
1670 // SETULT X, SINTMIN -> SETGT X, -1
1671 if (Cond == ISD::SETULT &&
1672 C1 == APInt::getSignedMinValue(OperandBitSize)) {
1673 SDValue ConstMinusOne =
1674 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize), dl,
1676 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT);
1679 // Fold bit comparisons when we can.
1680 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1681 (VT == N0.getValueType() ||
1682 (isTypeLegal(VT) && VT.bitsLE(N0.getValueType()))) &&
1683 N0.getOpcode() == ISD::AND)
1684 if (ConstantSDNode *AndRHS =
1685 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1686 EVT ShiftTy = DCI.isBeforeLegalize() ?
1687 getPointerTy() : getShiftAmountTy(N0.getValueType());
1688 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
1689 // Perform the xform if the AND RHS is a single bit.
1690 if (AndRHS->getAPIntValue().isPowerOf2()) {
1691 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1692 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0,
1693 DAG.getConstant(AndRHS->getAPIntValue().logBase2(), dl,
1696 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getAPIntValue()) {
1697 // (X & 8) == 8 --> (X & 8) >> 3
1698 // Perform the xform if C1 is a single bit.
1699 if (C1.isPowerOf2()) {
1700 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1701 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0,
1702 DAG.getConstant(C1.logBase2(), dl,
1708 if (C1.getMinSignedBits() <= 64 &&
1709 !isLegalICmpImmediate(C1.getSExtValue())) {
1710 // (X & -256) == 256 -> (X >> 8) == 1
1711 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1712 N0.getOpcode() == ISD::AND && N0.hasOneUse()) {
1713 if (ConstantSDNode *AndRHS =
1714 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1715 const APInt &AndRHSC = AndRHS->getAPIntValue();
1716 if ((-AndRHSC).isPowerOf2() && (AndRHSC & C1) == C1) {
1717 unsigned ShiftBits = AndRHSC.countTrailingZeros();
1718 EVT ShiftTy = DCI.isBeforeLegalize() ?
1719 getPointerTy() : getShiftAmountTy(N0.getValueType());
1720 EVT CmpTy = N0.getValueType();
1721 SDValue Shift = DAG.getNode(ISD::SRL, dl, CmpTy, N0.getOperand(0),
1722 DAG.getConstant(ShiftBits, dl,
1724 SDValue CmpRHS = DAG.getConstant(C1.lshr(ShiftBits), dl, CmpTy);
1725 return DAG.getSetCC(dl, VT, Shift, CmpRHS, Cond);
1728 } else if (Cond == ISD::SETULT || Cond == ISD::SETUGE ||
1729 Cond == ISD::SETULE || Cond == ISD::SETUGT) {
1730 bool AdjOne = (Cond == ISD::SETULE || Cond == ISD::SETUGT);
1731 // X < 0x100000000 -> (X >> 32) < 1
1732 // X >= 0x100000000 -> (X >> 32) >= 1
1733 // X <= 0x0ffffffff -> (X >> 32) < 1
1734 // X > 0x0ffffffff -> (X >> 32) >= 1
1737 ISD::CondCode NewCond = Cond;
1739 ShiftBits = C1.countTrailingOnes();
1741 NewCond = (Cond == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1743 ShiftBits = C1.countTrailingZeros();
1745 NewC = NewC.lshr(ShiftBits);
1746 if (ShiftBits && NewC.getMinSignedBits() <= 64 &&
1747 isLegalICmpImmediate(NewC.getSExtValue())) {
1748 EVT ShiftTy = DCI.isBeforeLegalize() ?
1749 getPointerTy() : getShiftAmountTy(N0.getValueType());
1750 EVT CmpTy = N0.getValueType();
1751 SDValue Shift = DAG.getNode(ISD::SRL, dl, CmpTy, N0,
1752 DAG.getConstant(ShiftBits, dl, ShiftTy));
1753 SDValue CmpRHS = DAG.getConstant(NewC, dl, CmpTy);
1754 return DAG.getSetCC(dl, VT, Shift, CmpRHS, NewCond);
1760 if (isa<ConstantFPSDNode>(N0.getNode())) {
1761 // Constant fold or commute setcc.
1762 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl);
1763 if (O.getNode()) return O;
1764 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
1765 // If the RHS of an FP comparison is a constant, simplify it away in
1767 if (CFP->getValueAPF().isNaN()) {
1768 // If an operand is known to be a nan, we can fold it.
1769 switch (ISD::getUnorderedFlavor(Cond)) {
1770 default: llvm_unreachable("Unknown flavor!");
1771 case 0: // Known false.
1772 return DAG.getConstant(0, dl, VT);
1773 case 1: // Known true.
1774 return DAG.getConstant(1, dl, VT);
1775 case 2: // Undefined.
1776 return DAG.getUNDEF(VT);
1780 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
1781 // constant if knowing that the operand is non-nan is enough. We prefer to
1782 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
1784 if (Cond == ISD::SETO || Cond == ISD::SETUO)
1785 return DAG.getSetCC(dl, VT, N0, N0, Cond);
1787 // If the condition is not legal, see if we can find an equivalent one
1789 if (!isCondCodeLegal(Cond, N0.getSimpleValueType())) {
1790 // If the comparison was an awkward floating-point == or != and one of
1791 // the comparison operands is infinity or negative infinity, convert the
1792 // condition to a less-awkward <= or >=.
1793 if (CFP->getValueAPF().isInfinity()) {
1794 if (CFP->getValueAPF().isNegative()) {
1795 if (Cond == ISD::SETOEQ &&
1796 isCondCodeLegal(ISD::SETOLE, N0.getSimpleValueType()))
1797 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLE);
1798 if (Cond == ISD::SETUEQ &&
1799 isCondCodeLegal(ISD::SETOLE, N0.getSimpleValueType()))
1800 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULE);
1801 if (Cond == ISD::SETUNE &&
1802 isCondCodeLegal(ISD::SETUGT, N0.getSimpleValueType()))
1803 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGT);
1804 if (Cond == ISD::SETONE &&
1805 isCondCodeLegal(ISD::SETUGT, N0.getSimpleValueType()))
1806 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGT);
1808 if (Cond == ISD::SETOEQ &&
1809 isCondCodeLegal(ISD::SETOGE, N0.getSimpleValueType()))
1810 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGE);
1811 if (Cond == ISD::SETUEQ &&
1812 isCondCodeLegal(ISD::SETOGE, N0.getSimpleValueType()))
1813 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGE);
1814 if (Cond == ISD::SETUNE &&
1815 isCondCodeLegal(ISD::SETULT, N0.getSimpleValueType()))
1816 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULT);
1817 if (Cond == ISD::SETONE &&
1818 isCondCodeLegal(ISD::SETULT, N0.getSimpleValueType()))
1819 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLT);
1826 // The sext(setcc()) => setcc() optimization relies on the appropriate
1827 // constant being emitted.
1829 switch (getBooleanContents(N0.getValueType())) {
1830 case UndefinedBooleanContent:
1831 case ZeroOrOneBooleanContent:
1832 EqVal = ISD::isTrueWhenEqual(Cond);
1834 case ZeroOrNegativeOneBooleanContent:
1835 EqVal = ISD::isTrueWhenEqual(Cond) ? -1 : 0;
1839 // We can always fold X == X for integer setcc's.
1840 if (N0.getValueType().isInteger()) {
1841 return DAG.getConstant(EqVal, dl, VT);
1843 unsigned UOF = ISD::getUnorderedFlavor(Cond);
1844 if (UOF == 2) // FP operators that are undefined on NaNs.
1845 return DAG.getConstant(EqVal, dl, VT);
1846 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
1847 return DAG.getConstant(EqVal, dl, VT);
1848 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
1849 // if it is not already.
1850 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
1851 if (NewCond != Cond && (DCI.isBeforeLegalizeOps() ||
1852 getCondCodeAction(NewCond, N0.getSimpleValueType()) == Legal))
1853 return DAG.getSetCC(dl, VT, N0, N1, NewCond);
1856 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1857 N0.getValueType().isInteger()) {
1858 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
1859 N0.getOpcode() == ISD::XOR) {
1860 // Simplify (X+Y) == (X+Z) --> Y == Z
1861 if (N0.getOpcode() == N1.getOpcode()) {
1862 if (N0.getOperand(0) == N1.getOperand(0))
1863 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond);
1864 if (N0.getOperand(1) == N1.getOperand(1))
1865 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond);
1866 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
1867 // If X op Y == Y op X, try other combinations.
1868 if (N0.getOperand(0) == N1.getOperand(1))
1869 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0),
1871 if (N0.getOperand(1) == N1.getOperand(0))
1872 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1),
1877 // If RHS is a legal immediate value for a compare instruction, we need
1878 // to be careful about increasing register pressure needlessly.
1879 bool LegalRHSImm = false;
1881 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
1882 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1883 // Turn (X+C1) == C2 --> X == C2-C1
1884 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) {
1885 return DAG.getSetCC(dl, VT, N0.getOperand(0),
1886 DAG.getConstant(RHSC->getAPIntValue()-
1887 LHSR->getAPIntValue(),
1888 dl, N0.getValueType()), Cond);
1891 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
1892 if (N0.getOpcode() == ISD::XOR)
1893 // If we know that all of the inverted bits are zero, don't bother
1894 // performing the inversion.
1895 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
1897 DAG.getSetCC(dl, VT, N0.getOperand(0),
1898 DAG.getConstant(LHSR->getAPIntValue() ^
1899 RHSC->getAPIntValue(),
1900 dl, N0.getValueType()),
1904 // Turn (C1-X) == C2 --> X == C1-C2
1905 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
1906 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) {
1908 DAG.getSetCC(dl, VT, N0.getOperand(1),
1909 DAG.getConstant(SUBC->getAPIntValue() -
1910 RHSC->getAPIntValue(),
1911 dl, N0.getValueType()),
1916 // Could RHSC fold directly into a compare?
1917 if (RHSC->getValueType(0).getSizeInBits() <= 64)
1918 LegalRHSImm = isLegalICmpImmediate(RHSC->getSExtValue());
1921 // Simplify (X+Z) == X --> Z == 0
1922 // Don't do this if X is an immediate that can fold into a cmp
1923 // instruction and X+Z has other uses. It could be an induction variable
1924 // chain, and the transform would increase register pressure.
1925 if (!LegalRHSImm || N0.getNode()->hasOneUse()) {
1926 if (N0.getOperand(0) == N1)
1927 return DAG.getSetCC(dl, VT, N0.getOperand(1),
1928 DAG.getConstant(0, dl, N0.getValueType()), Cond);
1929 if (N0.getOperand(1) == N1) {
1930 if (DAG.isCommutativeBinOp(N0.getOpcode()))
1931 return DAG.getSetCC(dl, VT, N0.getOperand(0),
1932 DAG.getConstant(0, dl, N0.getValueType()),
1934 if (N0.getNode()->hasOneUse()) {
1935 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
1936 // (Z-X) == X --> Z == X<<1
1937 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N1,
1938 DAG.getConstant(1, dl,
1939 getShiftAmountTy(N1.getValueType())));
1940 if (!DCI.isCalledByLegalizer())
1941 DCI.AddToWorklist(SH.getNode());
1942 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond);
1948 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
1949 N1.getOpcode() == ISD::XOR) {
1950 // Simplify X == (X+Z) --> Z == 0
1951 if (N1.getOperand(0) == N0)
1952 return DAG.getSetCC(dl, VT, N1.getOperand(1),
1953 DAG.getConstant(0, dl, N1.getValueType()), Cond);
1954 if (N1.getOperand(1) == N0) {
1955 if (DAG.isCommutativeBinOp(N1.getOpcode()))
1956 return DAG.getSetCC(dl, VT, N1.getOperand(0),
1957 DAG.getConstant(0, dl, N1.getValueType()), Cond);
1958 if (N1.getNode()->hasOneUse()) {
1959 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
1960 // X == (Z-X) --> X<<1 == Z
1961 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N0,
1962 DAG.getConstant(1, dl,
1963 getShiftAmountTy(N0.getValueType())));
1964 if (!DCI.isCalledByLegalizer())
1965 DCI.AddToWorklist(SH.getNode());
1966 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond);
1971 // Simplify x&y == y to x&y != 0 if y has exactly one bit set.
1972 // Note that where y is variable and is known to have at most
1973 // one bit set (for example, if it is z&1) we cannot do this;
1974 // the expressions are not equivalent when y==0.
1975 if (N0.getOpcode() == ISD::AND)
1976 if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) {
1977 if (ValueHasExactlyOneBitSet(N1, DAG)) {
1978 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
1979 if (DCI.isBeforeLegalizeOps() ||
1980 isCondCodeLegal(Cond, N0.getSimpleValueType())) {
1981 SDValue Zero = DAG.getConstant(0, dl, N1.getValueType());
1982 return DAG.getSetCC(dl, VT, N0, Zero, Cond);
1986 if (N1.getOpcode() == ISD::AND)
1987 if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) {
1988 if (ValueHasExactlyOneBitSet(N0, DAG)) {
1989 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
1990 if (DCI.isBeforeLegalizeOps() ||
1991 isCondCodeLegal(Cond, N1.getSimpleValueType())) {
1992 SDValue Zero = DAG.getConstant(0, dl, N0.getValueType());
1993 return DAG.getSetCC(dl, VT, N1, Zero, Cond);
1999 // Fold away ALL boolean setcc's.
2001 if (N0.getValueType() == MVT::i1 && foldBooleans) {
2003 default: llvm_unreachable("Unknown integer setcc!");
2004 case ISD::SETEQ: // X == Y -> ~(X^Y)
2005 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2006 N0 = DAG.getNOT(dl, Temp, MVT::i1);
2007 if (!DCI.isCalledByLegalizer())
2008 DCI.AddToWorklist(Temp.getNode());
2010 case ISD::SETNE: // X != Y --> (X^Y)
2011 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
2013 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y
2014 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y
2015 Temp = DAG.getNOT(dl, N0, MVT::i1);
2016 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp);
2017 if (!DCI.isCalledByLegalizer())
2018 DCI.AddToWorklist(Temp.getNode());
2020 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X
2021 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X
2022 Temp = DAG.getNOT(dl, N1, MVT::i1);
2023 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp);
2024 if (!DCI.isCalledByLegalizer())
2025 DCI.AddToWorklist(Temp.getNode());
2027 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y
2028 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y
2029 Temp = DAG.getNOT(dl, N0, MVT::i1);
2030 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp);
2031 if (!DCI.isCalledByLegalizer())
2032 DCI.AddToWorklist(Temp.getNode());
2034 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X
2035 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X
2036 Temp = DAG.getNOT(dl, N1, MVT::i1);
2037 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp);
2040 if (VT != MVT::i1) {
2041 if (!DCI.isCalledByLegalizer())
2042 DCI.AddToWorklist(N0.getNode());
2043 // FIXME: If running after legalize, we probably can't do this.
2044 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0);
2049 // Could not fold it.
2053 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
2054 /// node is a GlobalAddress + offset.
2055 bool TargetLowering::isGAPlusOffset(SDNode *N, const GlobalValue *&GA,
2056 int64_t &Offset) const {
2057 if (isa<GlobalAddressSDNode>(N)) {
2058 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N);
2059 GA = GASD->getGlobal();
2060 Offset += GASD->getOffset();
2064 if (N->getOpcode() == ISD::ADD) {
2065 SDValue N1 = N->getOperand(0);
2066 SDValue N2 = N->getOperand(1);
2067 if (isGAPlusOffset(N1.getNode(), GA, Offset)) {
2068 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
2070 Offset += V->getSExtValue();
2073 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) {
2074 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
2076 Offset += V->getSExtValue();
2086 SDValue TargetLowering::
2087 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
2088 // Default implementation: no optimization.
2092 //===----------------------------------------------------------------------===//
2093 // Inline Assembler Implementation Methods
2094 //===----------------------------------------------------------------------===//
2097 TargetLowering::ConstraintType
2098 TargetLowering::getConstraintType(const std::string &Constraint) const {
2099 unsigned S = Constraint.size();
2102 switch (Constraint[0]) {
2104 case 'r': return C_RegisterClass;
2106 case 'o': // offsetable
2107 case 'V': // not offsetable
2109 case 'i': // Simple Integer or Relocatable Constant
2110 case 'n': // Simple Integer
2111 case 'E': // Floating Point Constant
2112 case 'F': // Floating Point Constant
2113 case 's': // Relocatable Constant
2114 case 'p': // Address.
2115 case 'X': // Allow ANY value.
2116 case 'I': // Target registers.
2130 if (S > 1 && Constraint[0] == '{' && Constraint[S-1] == '}') {
2131 if (S == 8 && !Constraint.compare(1, 6, "memory", 6)) // "{memory}"
2138 /// LowerXConstraint - try to replace an X constraint, which matches anything,
2139 /// with another that has more specific requirements based on the type of the
2140 /// corresponding operand.
2141 const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const{
2142 if (ConstraintVT.isInteger())
2144 if (ConstraintVT.isFloatingPoint())
2145 return "f"; // works for many targets
2149 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
2150 /// vector. If it is invalid, don't add anything to Ops.
2151 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
2152 std::string &Constraint,
2153 std::vector<SDValue> &Ops,
2154 SelectionDAG &DAG) const {
2156 if (Constraint.length() > 1) return;
2158 char ConstraintLetter = Constraint[0];
2159 switch (ConstraintLetter) {
2161 case 'X': // Allows any operand; labels (basic block) use this.
2162 if (Op.getOpcode() == ISD::BasicBlock) {
2167 case 'i': // Simple Integer or Relocatable Constant
2168 case 'n': // Simple Integer
2169 case 's': { // Relocatable Constant
2170 // These operands are interested in values of the form (GV+C), where C may
2171 // be folded in as an offset of GV, or it may be explicitly added. Also, it
2172 // is possible and fine if either GV or C are missing.
2173 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2174 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
2176 // If we have "(add GV, C)", pull out GV/C
2177 if (Op.getOpcode() == ISD::ADD) {
2178 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2179 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
2181 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
2182 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
2185 C = nullptr, GA = nullptr;
2188 // If we find a valid operand, map to the TargetXXX version so that the
2189 // value itself doesn't get selected.
2190 if (GA) { // Either &GV or &GV+C
2191 if (ConstraintLetter != 'n') {
2192 int64_t Offs = GA->getOffset();
2193 if (C) Offs += C->getZExtValue();
2194 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
2195 C ? SDLoc(C) : SDLoc(),
2196 Op.getValueType(), Offs));
2200 if (C) { // just C, no GV.
2201 // Simple constants are not allowed for 's'.
2202 if (ConstraintLetter != 's') {
2203 // gcc prints these as sign extended. Sign extend value to 64 bits
2204 // now; without this it would get ZExt'd later in
2205 // ScheduleDAGSDNodes::EmitNode, which is very generic.
2206 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(),
2207 SDLoc(C), MVT::i64));
2216 std::pair<unsigned, const TargetRegisterClass *>
2217 TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *RI,
2218 const std::string &Constraint,
2220 if (Constraint.empty() || Constraint[0] != '{')
2221 return std::make_pair(0u, static_cast<TargetRegisterClass*>(nullptr));
2222 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
2224 // Remove the braces from around the name.
2225 StringRef RegName(Constraint.data()+1, Constraint.size()-2);
2227 std::pair<unsigned, const TargetRegisterClass*> R =
2228 std::make_pair(0u, static_cast<const TargetRegisterClass*>(nullptr));
2230 // Figure out which register class contains this reg.
2231 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
2232 E = RI->regclass_end(); RCI != E; ++RCI) {
2233 const TargetRegisterClass *RC = *RCI;
2235 // If none of the value types for this register class are valid, we
2236 // can't use it. For example, 64-bit reg classes on 32-bit targets.
2240 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
2242 if (RegName.equals_lower(RI->getName(*I))) {
2243 std::pair<unsigned, const TargetRegisterClass*> S =
2244 std::make_pair(*I, RC);
2246 // If this register class has the requested value type, return it,
2247 // otherwise keep searching and return the first class found
2248 // if no other is found which explicitly has the requested type.
2249 if (RC->hasType(VT))
2260 //===----------------------------------------------------------------------===//
2261 // Constraint Selection.
2263 /// isMatchingInputConstraint - Return true of this is an input operand that is
2264 /// a matching constraint like "4".
2265 bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const {
2266 assert(!ConstraintCode.empty() && "No known constraint!");
2267 return isdigit(static_cast<unsigned char>(ConstraintCode[0]));
2270 /// getMatchedOperand - If this is an input matching constraint, this method
2271 /// returns the output operand it matches.
2272 unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const {
2273 assert(!ConstraintCode.empty() && "No known constraint!");
2274 return atoi(ConstraintCode.c_str());
2278 /// ParseConstraints - Split up the constraint string from the inline
2279 /// assembly value into the specific constraints and their prefixes,
2280 /// and also tie in the associated operand values.
2281 /// If this returns an empty vector, and if the constraint string itself
2282 /// isn't empty, there was an error parsing.
2283 TargetLowering::AsmOperandInfoVector
2284 TargetLowering::ParseConstraints(const TargetRegisterInfo *TRI,
2285 ImmutableCallSite CS) const {
2286 /// ConstraintOperands - Information about all of the constraints.
2287 AsmOperandInfoVector ConstraintOperands;
2288 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
2289 unsigned maCount = 0; // Largest number of multiple alternative constraints.
2291 // Do a prepass over the constraints, canonicalizing them, and building up the
2292 // ConstraintOperands list.
2293 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
2294 unsigned ResNo = 0; // ResNo - The result number of the next output.
2296 for (InlineAsm::ConstraintInfo &CI : IA->ParseConstraints()) {
2297 ConstraintOperands.emplace_back(std::move(CI));
2298 AsmOperandInfo &OpInfo = ConstraintOperands.back();
2300 // Update multiple alternative constraint count.
2301 if (OpInfo.multipleAlternatives.size() > maCount)
2302 maCount = OpInfo.multipleAlternatives.size();
2304 OpInfo.ConstraintVT = MVT::Other;
2306 // Compute the value type for each operand.
2307 switch (OpInfo.Type) {
2308 case InlineAsm::isOutput:
2309 // Indirect outputs just consume an argument.
2310 if (OpInfo.isIndirect) {
2311 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
2315 // The return value of the call is this value. As such, there is no
2316 // corresponding argument.
2317 assert(!CS.getType()->isVoidTy() &&
2319 if (StructType *STy = dyn_cast<StructType>(CS.getType())) {
2320 OpInfo.ConstraintVT = getSimpleValueType(STy->getElementType(ResNo));
2322 assert(ResNo == 0 && "Asm only has one result!");
2323 OpInfo.ConstraintVT = getSimpleValueType(CS.getType());
2327 case InlineAsm::isInput:
2328 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
2330 case InlineAsm::isClobber:
2335 if (OpInfo.CallOperandVal) {
2336 llvm::Type *OpTy = OpInfo.CallOperandVal->getType();
2337 if (OpInfo.isIndirect) {
2338 llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
2340 report_fatal_error("Indirect operand for inline asm not a pointer!");
2341 OpTy = PtrTy->getElementType();
2344 // Look for vector wrapped in a struct. e.g. { <16 x i8> }.
2345 if (StructType *STy = dyn_cast<StructType>(OpTy))
2346 if (STy->getNumElements() == 1)
2347 OpTy = STy->getElementType(0);
2349 // If OpTy is not a single value, it may be a struct/union that we
2350 // can tile with integers.
2351 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
2352 unsigned BitSize = getDataLayout()->getTypeSizeInBits(OpTy);
2361 OpInfo.ConstraintVT =
2362 MVT::getVT(IntegerType::get(OpTy->getContext(), BitSize), true);
2365 } else if (PointerType *PT = dyn_cast<PointerType>(OpTy)) {
2367 = getDataLayout()->getPointerSizeInBits(PT->getAddressSpace());
2368 OpInfo.ConstraintVT = MVT::getIntegerVT(PtrSize);
2370 OpInfo.ConstraintVT = MVT::getVT(OpTy, true);
2375 // If we have multiple alternative constraints, select the best alternative.
2376 if (!ConstraintOperands.empty()) {
2378 unsigned bestMAIndex = 0;
2379 int bestWeight = -1;
2380 // weight: -1 = invalid match, and 0 = so-so match to 5 = good match.
2383 // Compute the sums of the weights for each alternative, keeping track
2384 // of the best (highest weight) one so far.
2385 for (maIndex = 0; maIndex < maCount; ++maIndex) {
2387 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2388 cIndex != eIndex; ++cIndex) {
2389 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex];
2390 if (OpInfo.Type == InlineAsm::isClobber)
2393 // If this is an output operand with a matching input operand,
2394 // look up the matching input. If their types mismatch, e.g. one
2395 // is an integer, the other is floating point, or their sizes are
2396 // different, flag it as an maCantMatch.
2397 if (OpInfo.hasMatchingInput()) {
2398 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
2399 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
2400 if ((OpInfo.ConstraintVT.isInteger() !=
2401 Input.ConstraintVT.isInteger()) ||
2402 (OpInfo.ConstraintVT.getSizeInBits() !=
2403 Input.ConstraintVT.getSizeInBits())) {
2404 weightSum = -1; // Can't match.
2409 weight = getMultipleConstraintMatchWeight(OpInfo, maIndex);
2414 weightSum += weight;
2417 if (weightSum > bestWeight) {
2418 bestWeight = weightSum;
2419 bestMAIndex = maIndex;
2423 // Now select chosen alternative in each constraint.
2424 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2425 cIndex != eIndex; ++cIndex) {
2426 AsmOperandInfo& cInfo = ConstraintOperands[cIndex];
2427 if (cInfo.Type == InlineAsm::isClobber)
2429 cInfo.selectAlternative(bestMAIndex);
2434 // Check and hook up tied operands, choose constraint code to use.
2435 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2436 cIndex != eIndex; ++cIndex) {
2437 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex];
2439 // If this is an output operand with a matching input operand, look up the
2440 // matching input. If their types mismatch, e.g. one is an integer, the
2441 // other is floating point, or their sizes are different, flag it as an
2443 if (OpInfo.hasMatchingInput()) {
2444 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
2446 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
2447 std::pair<unsigned, const TargetRegisterClass *> MatchRC =
2448 getRegForInlineAsmConstraint(TRI, OpInfo.ConstraintCode,
2449 OpInfo.ConstraintVT);
2450 std::pair<unsigned, const TargetRegisterClass *> InputRC =
2451 getRegForInlineAsmConstraint(TRI, Input.ConstraintCode,
2452 Input.ConstraintVT);
2453 if ((OpInfo.ConstraintVT.isInteger() !=
2454 Input.ConstraintVT.isInteger()) ||
2455 (MatchRC.second != InputRC.second)) {
2456 report_fatal_error("Unsupported asm: input constraint"
2457 " with a matching output constraint of"
2458 " incompatible type!");
2465 return ConstraintOperands;
2469 /// getConstraintGenerality - Return an integer indicating how general CT
2471 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
2473 case TargetLowering::C_Other:
2474 case TargetLowering::C_Unknown:
2476 case TargetLowering::C_Register:
2478 case TargetLowering::C_RegisterClass:
2480 case TargetLowering::C_Memory:
2483 llvm_unreachable("Invalid constraint type");
2486 /// Examine constraint type and operand type and determine a weight value.
2487 /// This object must already have been set up with the operand type
2488 /// and the current alternative constraint selected.
2489 TargetLowering::ConstraintWeight
2490 TargetLowering::getMultipleConstraintMatchWeight(
2491 AsmOperandInfo &info, int maIndex) const {
2492 InlineAsm::ConstraintCodeVector *rCodes;
2493 if (maIndex >= (int)info.multipleAlternatives.size())
2494 rCodes = &info.Codes;
2496 rCodes = &info.multipleAlternatives[maIndex].Codes;
2497 ConstraintWeight BestWeight = CW_Invalid;
2499 // Loop over the options, keeping track of the most general one.
2500 for (unsigned i = 0, e = rCodes->size(); i != e; ++i) {
2501 ConstraintWeight weight =
2502 getSingleConstraintMatchWeight(info, (*rCodes)[i].c_str());
2503 if (weight > BestWeight)
2504 BestWeight = weight;
2510 /// Examine constraint type and operand type and determine a weight value.
2511 /// This object must already have been set up with the operand type
2512 /// and the current alternative constraint selected.
2513 TargetLowering::ConstraintWeight
2514 TargetLowering::getSingleConstraintMatchWeight(
2515 AsmOperandInfo &info, const char *constraint) const {
2516 ConstraintWeight weight = CW_Invalid;
2517 Value *CallOperandVal = info.CallOperandVal;
2518 // If we don't have a value, we can't do a match,
2519 // but allow it at the lowest weight.
2520 if (!CallOperandVal)
2522 // Look at the constraint type.
2523 switch (*constraint) {
2524 case 'i': // immediate integer.
2525 case 'n': // immediate integer with a known value.
2526 if (isa<ConstantInt>(CallOperandVal))
2527 weight = CW_Constant;
2529 case 's': // non-explicit intregal immediate.
2530 if (isa<GlobalValue>(CallOperandVal))
2531 weight = CW_Constant;
2533 case 'E': // immediate float if host format.
2534 case 'F': // immediate float.
2535 if (isa<ConstantFP>(CallOperandVal))
2536 weight = CW_Constant;
2538 case '<': // memory operand with autodecrement.
2539 case '>': // memory operand with autoincrement.
2540 case 'm': // memory operand.
2541 case 'o': // offsettable memory operand
2542 case 'V': // non-offsettable memory operand
2545 case 'r': // general register.
2546 case 'g': // general register, memory operand or immediate integer.
2547 // note: Clang converts "g" to "imr".
2548 if (CallOperandVal->getType()->isIntegerTy())
2549 weight = CW_Register;
2551 case 'X': // any operand.
2553 weight = CW_Default;
2559 /// ChooseConstraint - If there are multiple different constraints that we
2560 /// could pick for this operand (e.g. "imr") try to pick the 'best' one.
2561 /// This is somewhat tricky: constraints fall into four classes:
2562 /// Other -> immediates and magic values
2563 /// Register -> one specific register
2564 /// RegisterClass -> a group of regs
2565 /// Memory -> memory
2566 /// Ideally, we would pick the most specific constraint possible: if we have
2567 /// something that fits into a register, we would pick it. The problem here
2568 /// is that if we have something that could either be in a register or in
2569 /// memory that use of the register could cause selection of *other*
2570 /// operands to fail: they might only succeed if we pick memory. Because of
2571 /// this the heuristic we use is:
2573 /// 1) If there is an 'other' constraint, and if the operand is valid for
2574 /// that constraint, use it. This makes us take advantage of 'i'
2575 /// constraints when available.
2576 /// 2) Otherwise, pick the most general constraint present. This prefers
2577 /// 'm' over 'r', for example.
2579 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo,
2580 const TargetLowering &TLI,
2581 SDValue Op, SelectionDAG *DAG) {
2582 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options");
2583 unsigned BestIdx = 0;
2584 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown;
2585 int BestGenerality = -1;
2587 // Loop over the options, keeping track of the most general one.
2588 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) {
2589 TargetLowering::ConstraintType CType =
2590 TLI.getConstraintType(OpInfo.Codes[i]);
2592 // If this is an 'other' constraint, see if the operand is valid for it.
2593 // For example, on X86 we might have an 'rI' constraint. If the operand
2594 // is an integer in the range [0..31] we want to use I (saving a load
2595 // of a register), otherwise we must use 'r'.
2596 if (CType == TargetLowering::C_Other && Op.getNode()) {
2597 assert(OpInfo.Codes[i].size() == 1 &&
2598 "Unhandled multi-letter 'other' constraint");
2599 std::vector<SDValue> ResultOps;
2600 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i],
2602 if (!ResultOps.empty()) {
2609 // Things with matching constraints can only be registers, per gcc
2610 // documentation. This mainly affects "g" constraints.
2611 if (CType == TargetLowering::C_Memory && OpInfo.hasMatchingInput())
2614 // This constraint letter is more general than the previous one, use it.
2615 int Generality = getConstraintGenerality(CType);
2616 if (Generality > BestGenerality) {
2619 BestGenerality = Generality;
2623 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx];
2624 OpInfo.ConstraintType = BestType;
2627 /// ComputeConstraintToUse - Determines the constraint code and constraint
2628 /// type to use for the specific AsmOperandInfo, setting
2629 /// OpInfo.ConstraintCode and OpInfo.ConstraintType.
2630 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
2632 SelectionDAG *DAG) const {
2633 assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
2635 // Single-letter constraints ('r') are very common.
2636 if (OpInfo.Codes.size() == 1) {
2637 OpInfo.ConstraintCode = OpInfo.Codes[0];
2638 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2640 ChooseConstraint(OpInfo, *this, Op, DAG);
2643 // 'X' matches anything.
2644 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
2645 // Labels and constants are handled elsewhere ('X' is the only thing
2646 // that matches labels). For Functions, the type here is the type of
2647 // the result, which is not what we want to look at; leave them alone.
2648 Value *v = OpInfo.CallOperandVal;
2649 if (isa<BasicBlock>(v) || isa<ConstantInt>(v) || isa<Function>(v)) {
2650 OpInfo.CallOperandVal = v;
2654 // Otherwise, try to resolve it to something we know about by looking at
2655 // the actual operand type.
2656 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
2657 OpInfo.ConstraintCode = Repl;
2658 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2663 /// \brief Given an exact SDIV by a constant, create a multiplication
2664 /// with the multiplicative inverse of the constant.
2665 SDValue TargetLowering::BuildExactSDIV(SDValue Op1, SDValue Op2, SDLoc dl,
2666 SelectionDAG &DAG) const {
2667 ConstantSDNode *C = cast<ConstantSDNode>(Op2);
2668 APInt d = C->getAPIntValue();
2669 assert(d != 0 && "Division by zero!");
2671 // Shift the value upfront if it is even, so the LSB is one.
2672 unsigned ShAmt = d.countTrailingZeros();
2674 // TODO: For UDIV use SRL instead of SRA.
2676 DAG.getConstant(ShAmt, dl, getShiftAmountTy(Op1.getValueType()));
2678 Flags.setExact(true);
2679 Op1 = DAG.getNode(ISD::SRA, dl, Op1.getValueType(), Op1, Amt, &Flags);
2683 // Calculate the multiplicative inverse, using Newton's method.
2685 while ((t = d*xn) != 1)
2686 xn *= APInt(d.getBitWidth(), 2) - t;
2688 Op2 = DAG.getConstant(xn, dl, Op1.getValueType());
2689 return DAG.getNode(ISD::MUL, dl, Op1.getValueType(), Op1, Op2);
2692 /// \brief Given an ISD::SDIV node expressing a divide by constant,
2693 /// return a DAG expression to select that will generate the same value by
2694 /// multiplying by a magic number.
2695 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
2696 SDValue TargetLowering::BuildSDIV(SDNode *N, const APInt &Divisor,
2697 SelectionDAG &DAG, bool IsAfterLegalization,
2698 std::vector<SDNode *> *Created) const {
2699 assert(Created && "No vector to hold sdiv ops.");
2701 EVT VT = N->getValueType(0);
2704 // Check to see if we can do this.
2705 // FIXME: We should be more aggressive here.
2706 if (!isTypeLegal(VT))
2709 APInt::ms magics = Divisor.magic();
2711 // Multiply the numerator (operand 0) by the magic value
2712 // FIXME: We should support doing a MUL in a wider type
2714 if (IsAfterLegalization ? isOperationLegal(ISD::MULHS, VT) :
2715 isOperationLegalOrCustom(ISD::MULHS, VT))
2716 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0),
2717 DAG.getConstant(magics.m, dl, VT));
2718 else if (IsAfterLegalization ? isOperationLegal(ISD::SMUL_LOHI, VT) :
2719 isOperationLegalOrCustom(ISD::SMUL_LOHI, VT))
2720 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT),
2722 DAG.getConstant(magics.m, dl, VT)).getNode(), 1);
2724 return SDValue(); // No mulhs or equvialent
2725 // If d > 0 and m < 0, add the numerator
2726 if (Divisor.isStrictlyPositive() && magics.m.isNegative()) {
2727 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0));
2728 Created->push_back(Q.getNode());
2730 // If d < 0 and m > 0, subtract the numerator.
2731 if (Divisor.isNegative() && magics.m.isStrictlyPositive()) {
2732 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0));
2733 Created->push_back(Q.getNode());
2735 // Shift right algebraic if shift value is nonzero
2737 Q = DAG.getNode(ISD::SRA, dl, VT, Q,
2738 DAG.getConstant(magics.s, dl,
2739 getShiftAmountTy(Q.getValueType())));
2740 Created->push_back(Q.getNode());
2742 // Extract the sign bit and add it to the quotient
2743 SDValue T = DAG.getNode(ISD::SRL, dl, VT, Q,
2744 DAG.getConstant(VT.getScalarSizeInBits() - 1, dl,
2745 getShiftAmountTy(Q.getValueType())));
2746 Created->push_back(T.getNode());
2747 return DAG.getNode(ISD::ADD, dl, VT, Q, T);
2750 /// \brief Given an ISD::UDIV node expressing a divide by constant,
2751 /// return a DAG expression to select that will generate the same value by
2752 /// multiplying by a magic number.
2753 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
2754 SDValue TargetLowering::BuildUDIV(SDNode *N, const APInt &Divisor,
2755 SelectionDAG &DAG, bool IsAfterLegalization,
2756 std::vector<SDNode *> *Created) const {
2757 assert(Created && "No vector to hold udiv ops.");
2759 EVT VT = N->getValueType(0);
2762 // Check to see if we can do this.
2763 // FIXME: We should be more aggressive here.
2764 if (!isTypeLegal(VT))
2767 // FIXME: We should use a narrower constant when the upper
2768 // bits are known to be zero.
2769 APInt::mu magics = Divisor.magicu();
2771 SDValue Q = N->getOperand(0);
2773 // If the divisor is even, we can avoid using the expensive fixup by shifting
2774 // the divided value upfront.
2775 if (magics.a != 0 && !Divisor[0]) {
2776 unsigned Shift = Divisor.countTrailingZeros();
2777 Q = DAG.getNode(ISD::SRL, dl, VT, Q,
2778 DAG.getConstant(Shift, dl,
2779 getShiftAmountTy(Q.getValueType())));
2780 Created->push_back(Q.getNode());
2782 // Get magic number for the shifted divisor.
2783 magics = Divisor.lshr(Shift).magicu(Shift);
2784 assert(magics.a == 0 && "Should use cheap fixup now");
2787 // Multiply the numerator (operand 0) by the magic value
2788 // FIXME: We should support doing a MUL in a wider type
2789 if (IsAfterLegalization ? isOperationLegal(ISD::MULHU, VT) :
2790 isOperationLegalOrCustom(ISD::MULHU, VT))
2791 Q = DAG.getNode(ISD::MULHU, dl, VT, Q, DAG.getConstant(magics.m, dl, VT));
2792 else if (IsAfterLegalization ? isOperationLegal(ISD::UMUL_LOHI, VT) :
2793 isOperationLegalOrCustom(ISD::UMUL_LOHI, VT))
2794 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT), Q,
2795 DAG.getConstant(magics.m, dl, VT)).getNode(), 1);
2797 return SDValue(); // No mulhu or equvialent
2799 Created->push_back(Q.getNode());
2801 if (magics.a == 0) {
2802 assert(magics.s < Divisor.getBitWidth() &&
2803 "We shouldn't generate an undefined shift!");
2804 return DAG.getNode(ISD::SRL, dl, VT, Q,
2805 DAG.getConstant(magics.s, dl,
2806 getShiftAmountTy(Q.getValueType())));
2808 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q);
2809 Created->push_back(NPQ.getNode());
2810 NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ,
2811 DAG.getConstant(1, dl,
2812 getShiftAmountTy(NPQ.getValueType())));
2813 Created->push_back(NPQ.getNode());
2814 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q);
2815 Created->push_back(NPQ.getNode());
2816 return DAG.getNode(ISD::SRL, dl, VT, NPQ,
2817 DAG.getConstant(magics.s - 1, dl,
2818 getShiftAmountTy(NPQ.getValueType())));
2822 bool TargetLowering::
2823 verifyReturnAddressArgumentIsConstant(SDValue Op, SelectionDAG &DAG) const {
2824 if (!isa<ConstantSDNode>(Op.getOperand(0))) {
2825 DAG.getContext()->emitError("argument to '__builtin_return_address' must "
2826 "be a constant integer");
2833 //===----------------------------------------------------------------------===//
2834 // Legalization Utilities
2835 //===----------------------------------------------------------------------===//
2837 bool TargetLowering::expandMUL(SDNode *N, SDValue &Lo, SDValue &Hi, EVT HiLoVT,
2838 SelectionDAG &DAG, SDValue LL, SDValue LH,
2839 SDValue RL, SDValue RH) const {
2840 EVT VT = N->getValueType(0);
2843 bool HasMULHS = isOperationLegalOrCustom(ISD::MULHS, HiLoVT);
2844 bool HasMULHU = isOperationLegalOrCustom(ISD::MULHU, HiLoVT);
2845 bool HasSMUL_LOHI = isOperationLegalOrCustom(ISD::SMUL_LOHI, HiLoVT);
2846 bool HasUMUL_LOHI = isOperationLegalOrCustom(ISD::UMUL_LOHI, HiLoVT);
2847 if (HasMULHU || HasMULHS || HasUMUL_LOHI || HasSMUL_LOHI) {
2848 unsigned OuterBitSize = VT.getSizeInBits();
2849 unsigned InnerBitSize = HiLoVT.getSizeInBits();
2850 unsigned LHSSB = DAG.ComputeNumSignBits(N->getOperand(0));
2851 unsigned RHSSB = DAG.ComputeNumSignBits(N->getOperand(1));
2853 // LL, LH, RL, and RH must be either all NULL or all set to a value.
2854 assert((LL.getNode() && LH.getNode() && RL.getNode() && RH.getNode()) ||
2855 (!LL.getNode() && !LH.getNode() && !RL.getNode() && !RH.getNode()));
2857 if (!LL.getNode() && !RL.getNode() &&
2858 isOperationLegalOrCustom(ISD::TRUNCATE, HiLoVT)) {
2859 LL = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, N->getOperand(0));
2860 RL = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, N->getOperand(1));
2866 APInt HighMask = APInt::getHighBitsSet(OuterBitSize, InnerBitSize);
2867 if (DAG.MaskedValueIsZero(N->getOperand(0), HighMask) &&
2868 DAG.MaskedValueIsZero(N->getOperand(1), HighMask)) {
2869 // The inputs are both zero-extended.
2871 // We can emit a umul_lohi.
2872 Lo = DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(HiLoVT, HiLoVT), LL,
2874 Hi = SDValue(Lo.getNode(), 1);
2878 // We can emit a mulhu+mul.
2879 Lo = DAG.getNode(ISD::MUL, dl, HiLoVT, LL, RL);
2880 Hi = DAG.getNode(ISD::MULHU, dl, HiLoVT, LL, RL);
2884 if (LHSSB > InnerBitSize && RHSSB > InnerBitSize) {
2885 // The input values are both sign-extended.
2887 // We can emit a smul_lohi.
2888 Lo = DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(HiLoVT, HiLoVT), LL,
2890 Hi = SDValue(Lo.getNode(), 1);
2894 // We can emit a mulhs+mul.
2895 Lo = DAG.getNode(ISD::MUL, dl, HiLoVT, LL, RL);
2896 Hi = DAG.getNode(ISD::MULHS, dl, HiLoVT, LL, RL);
2901 if (!LH.getNode() && !RH.getNode() &&
2902 isOperationLegalOrCustom(ISD::SRL, VT) &&
2903 isOperationLegalOrCustom(ISD::TRUNCATE, HiLoVT)) {
2904 unsigned ShiftAmt = VT.getSizeInBits() - HiLoVT.getSizeInBits();
2905 SDValue Shift = DAG.getConstant(ShiftAmt, dl, getShiftAmountTy(VT));
2906 LH = DAG.getNode(ISD::SRL, dl, VT, N->getOperand(0), Shift);
2907 LH = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, LH);
2908 RH = DAG.getNode(ISD::SRL, dl, VT, N->getOperand(1), Shift);
2909 RH = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, RH);
2916 // Lo,Hi = umul LHS, RHS.
2917 SDValue UMulLOHI = DAG.getNode(ISD::UMUL_LOHI, dl,
2918 DAG.getVTList(HiLoVT, HiLoVT), LL, RL);
2920 Hi = UMulLOHI.getValue(1);
2921 RH = DAG.getNode(ISD::MUL, dl, HiLoVT, LL, RH);
2922 LH = DAG.getNode(ISD::MUL, dl, HiLoVT, LH, RL);
2923 Hi = DAG.getNode(ISD::ADD, dl, HiLoVT, Hi, RH);
2924 Hi = DAG.getNode(ISD::ADD, dl, HiLoVT, Hi, LH);
2928 Lo = DAG.getNode(ISD::MUL, dl, HiLoVT, LL, RL);
2929 Hi = DAG.getNode(ISD::MULHU, dl, HiLoVT, LL, RL);
2930 RH = DAG.getNode(ISD::MUL, dl, HiLoVT, LL, RH);
2931 LH = DAG.getNode(ISD::MUL, dl, HiLoVT, LH, RL);
2932 Hi = DAG.getNode(ISD::ADD, dl, HiLoVT, Hi, RH);
2933 Hi = DAG.getNode(ISD::ADD, dl, HiLoVT, Hi, LH);
2940 bool TargetLowering::expandFP_TO_SINT(SDNode *Node, SDValue &Result,
2941 SelectionDAG &DAG) const {
2942 EVT VT = Node->getOperand(0).getValueType();
2943 EVT NVT = Node->getValueType(0);
2944 SDLoc dl(SDValue(Node, 0));
2946 // FIXME: Only f32 to i64 conversions are supported.
2947 if (VT != MVT::f32 || NVT != MVT::i64)
2950 // Expand f32 -> i64 conversion
2951 // This algorithm comes from compiler-rt's implementation of fixsfdi:
2952 // https://github.com/llvm-mirror/compiler-rt/blob/master/lib/builtins/fixsfdi.c
2953 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(),
2954 VT.getSizeInBits());
2955 SDValue ExponentMask = DAG.getConstant(0x7F800000, dl, IntVT);
2956 SDValue ExponentLoBit = DAG.getConstant(23, dl, IntVT);
2957 SDValue Bias = DAG.getConstant(127, dl, IntVT);
2958 SDValue SignMask = DAG.getConstant(APInt::getSignBit(VT.getSizeInBits()), dl,
2960 SDValue SignLowBit = DAG.getConstant(VT.getSizeInBits() - 1, dl, IntVT);
2961 SDValue MantissaMask = DAG.getConstant(0x007FFFFF, dl, IntVT);
2963 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, IntVT, Node->getOperand(0));
2965 SDValue ExponentBits = DAG.getNode(ISD::SRL, dl, IntVT,
2966 DAG.getNode(ISD::AND, dl, IntVT, Bits, ExponentMask),
2967 DAG.getZExtOrTrunc(ExponentLoBit, dl, getShiftAmountTy(IntVT)));
2968 SDValue Exponent = DAG.getNode(ISD::SUB, dl, IntVT, ExponentBits, Bias);
2970 SDValue Sign = DAG.getNode(ISD::SRA, dl, IntVT,
2971 DAG.getNode(ISD::AND, dl, IntVT, Bits, SignMask),
2972 DAG.getZExtOrTrunc(SignLowBit, dl, getShiftAmountTy(IntVT)));
2973 Sign = DAG.getSExtOrTrunc(Sign, dl, NVT);
2975 SDValue R = DAG.getNode(ISD::OR, dl, IntVT,
2976 DAG.getNode(ISD::AND, dl, IntVT, Bits, MantissaMask),
2977 DAG.getConstant(0x00800000, dl, IntVT));
2979 R = DAG.getZExtOrTrunc(R, dl, NVT);
2982 R = DAG.getSelectCC(dl, Exponent, ExponentLoBit,
2983 DAG.getNode(ISD::SHL, dl, NVT, R,
2985 DAG.getNode(ISD::SUB, dl, IntVT, Exponent, ExponentLoBit),
2986 dl, getShiftAmountTy(IntVT))),
2987 DAG.getNode(ISD::SRL, dl, NVT, R,
2989 DAG.getNode(ISD::SUB, dl, IntVT, ExponentLoBit, Exponent),
2990 dl, getShiftAmountTy(IntVT))),
2993 SDValue Ret = DAG.getNode(ISD::SUB, dl, NVT,
2994 DAG.getNode(ISD::XOR, dl, NVT, R, Sign),
2997 Result = DAG.getSelectCC(dl, Exponent, DAG.getConstant(0, dl, IntVT),
2998 DAG.getConstant(0, dl, NVT), Ret, ISD::SETLT);