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/MC/MCAsmInfo.h"
26 #include "llvm/MC/MCExpr.h"
27 #include "llvm/Support/CommandLine.h"
28 #include "llvm/Support/ErrorHandling.h"
29 #include "llvm/Support/MathExtras.h"
30 #include "llvm/Target/TargetLoweringObjectFile.h"
31 #include "llvm/Target/TargetMachine.h"
32 #include "llvm/Target/TargetRegisterInfo.h"
36 /// NOTE: The constructor takes ownership of TLOF.
37 TargetLowering::TargetLowering(const TargetMachine &tm,
38 const TargetLoweringObjectFile *tlof)
39 : TargetLoweringBase(tm, tlof) {}
41 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const {
45 /// Check whether a given call node is in tail position within its function. If
46 /// so, it sets Chain to the input chain of the tail call.
47 bool TargetLowering::isInTailCallPosition(SelectionDAG &DAG, SDNode *Node,
48 SDValue &Chain) const {
49 const Function *F = DAG.getMachineFunction().getFunction();
51 // Conservatively require the attributes of the call to match those of
52 // the return. Ignore noalias because it doesn't affect the call sequence.
53 AttributeSet CallerAttrs = F->getAttributes();
54 if (AttrBuilder(CallerAttrs, AttributeSet::ReturnIndex)
55 .removeAttribute(Attribute::NoAlias).hasAttributes())
58 // It's not safe to eliminate the sign / zero extension of the return value.
59 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
60 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
63 // Check if the only use is a function return node.
64 return isUsedByReturnOnly(Node, Chain);
68 /// Generate a libcall taking the given operands as arguments and returning a
69 /// result of type RetVT.
70 std::pair<SDValue, SDValue>
71 TargetLowering::makeLibCall(SelectionDAG &DAG,
72 RTLIB::Libcall LC, EVT RetVT,
73 const SDValue *Ops, unsigned NumOps,
74 bool isSigned, SDLoc dl,
76 bool isReturnValueUsed) const {
77 TargetLowering::ArgListTy Args;
80 TargetLowering::ArgListEntry Entry;
81 for (unsigned i = 0; i != NumOps; ++i) {
83 Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext());
84 Entry.isSExt = isSigned;
85 Entry.isZExt = !isSigned;
86 Args.push_back(Entry);
88 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC), getPointerTy());
90 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
92 CallLoweringInfo CLI(DAG.getEntryNode(), RetTy, isSigned, !isSigned, false,
93 false, 0, getLibcallCallingConv(LC),
95 doesNotReturn, isReturnValueUsed, Callee, Args,
97 return LowerCallTo(CLI);
101 /// SoftenSetCCOperands - Soften the operands of a comparison. This code is
102 /// shared among BR_CC, SELECT_CC, and SETCC handlers.
103 void TargetLowering::softenSetCCOperands(SelectionDAG &DAG, EVT VT,
104 SDValue &NewLHS, SDValue &NewRHS,
105 ISD::CondCode &CCCode,
107 assert((VT == MVT::f32 || VT == MVT::f64 || VT == MVT::f128)
108 && "Unsupported setcc type!");
110 // Expand into one or more soft-fp libcall(s).
111 RTLIB::Libcall LC1 = RTLIB::UNKNOWN_LIBCALL, LC2 = RTLIB::UNKNOWN_LIBCALL;
115 LC1 = (VT == MVT::f32) ? RTLIB::OEQ_F32 :
116 (VT == MVT::f64) ? RTLIB::OEQ_F64 : RTLIB::OEQ_F128;
120 LC1 = (VT == MVT::f32) ? RTLIB::UNE_F32 :
121 (VT == MVT::f64) ? RTLIB::UNE_F64 : RTLIB::UNE_F128;
125 LC1 = (VT == MVT::f32) ? RTLIB::OGE_F32 :
126 (VT == MVT::f64) ? RTLIB::OGE_F64 : RTLIB::OGE_F128;
130 LC1 = (VT == MVT::f32) ? RTLIB::OLT_F32 :
131 (VT == MVT::f64) ? RTLIB::OLT_F64 : RTLIB::OLT_F128;
135 LC1 = (VT == MVT::f32) ? RTLIB::OLE_F32 :
136 (VT == MVT::f64) ? RTLIB::OLE_F64 : RTLIB::OLE_F128;
140 LC1 = (VT == MVT::f32) ? RTLIB::OGT_F32 :
141 (VT == MVT::f64) ? RTLIB::OGT_F64 : RTLIB::OGT_F128;
144 LC1 = (VT == MVT::f32) ? RTLIB::UO_F32 :
145 (VT == MVT::f64) ? RTLIB::UO_F64 : RTLIB::UO_F128;
148 LC1 = (VT == MVT::f32) ? RTLIB::O_F32 :
149 (VT == MVT::f64) ? RTLIB::O_F64 : RTLIB::O_F128;
152 LC1 = (VT == MVT::f32) ? RTLIB::UO_F32 :
153 (VT == MVT::f64) ? RTLIB::UO_F64 : RTLIB::UO_F128;
156 // SETONE = SETOLT | SETOGT
157 LC1 = (VT == MVT::f32) ? RTLIB::OLT_F32 :
158 (VT == MVT::f64) ? RTLIB::OLT_F64 : RTLIB::OLT_F128;
161 LC2 = (VT == MVT::f32) ? RTLIB::OGT_F32 :
162 (VT == MVT::f64) ? RTLIB::OGT_F64 : RTLIB::OGT_F128;
165 LC2 = (VT == MVT::f32) ? RTLIB::OGE_F32 :
166 (VT == MVT::f64) ? RTLIB::OGE_F64 : RTLIB::OGE_F128;
169 LC2 = (VT == MVT::f32) ? RTLIB::OLT_F32 :
170 (VT == MVT::f64) ? RTLIB::OLT_F64 : RTLIB::OLT_F128;
173 LC2 = (VT == MVT::f32) ? RTLIB::OLE_F32 :
174 (VT == MVT::f64) ? RTLIB::OLE_F64 : RTLIB::OLE_F128;
177 LC2 = (VT == MVT::f32) ? RTLIB::OEQ_F32 :
178 (VT == MVT::f64) ? RTLIB::OEQ_F64 : RTLIB::OEQ_F128;
180 default: llvm_unreachable("Do not know how to soften this setcc!");
184 // Use the target specific return value for comparions lib calls.
185 EVT RetVT = getCmpLibcallReturnType();
186 SDValue Ops[2] = { NewLHS, NewRHS };
187 NewLHS = makeLibCall(DAG, LC1, RetVT, Ops, 2, false/*sign irrelevant*/,
189 NewRHS = DAG.getConstant(0, RetVT);
190 CCCode = getCmpLibcallCC(LC1);
191 if (LC2 != RTLIB::UNKNOWN_LIBCALL) {
192 SDValue Tmp = DAG.getNode(ISD::SETCC, dl,
193 getSetCCResultType(*DAG.getContext(), RetVT),
194 NewLHS, NewRHS, DAG.getCondCode(CCCode));
195 NewLHS = makeLibCall(DAG, LC2, RetVT, Ops, 2, false/*sign irrelevant*/,
197 NewLHS = DAG.getNode(ISD::SETCC, dl,
198 getSetCCResultType(*DAG.getContext(), RetVT), NewLHS,
199 NewRHS, DAG.getCondCode(getCmpLibcallCC(LC2)));
200 NewLHS = DAG.getNode(ISD::OR, dl, Tmp.getValueType(), Tmp, NewLHS);
205 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
206 /// current function. The returned value is a member of the
207 /// MachineJumpTableInfo::JTEntryKind enum.
208 unsigned TargetLowering::getJumpTableEncoding() const {
209 // In non-pic modes, just use the address of a block.
210 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
211 return MachineJumpTableInfo::EK_BlockAddress;
213 // In PIC mode, if the target supports a GPRel32 directive, use it.
214 if (getTargetMachine().getMCAsmInfo()->getGPRel32Directive() != 0)
215 return MachineJumpTableInfo::EK_GPRel32BlockAddress;
217 // Otherwise, use a label difference.
218 return MachineJumpTableInfo::EK_LabelDifference32;
221 SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table,
222 SelectionDAG &DAG) const {
223 // If our PIC model is GP relative, use the global offset table as the base.
224 unsigned JTEncoding = getJumpTableEncoding();
226 if ((JTEncoding == MachineJumpTableInfo::EK_GPRel64BlockAddress) ||
227 (JTEncoding == MachineJumpTableInfo::EK_GPRel32BlockAddress))
228 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy(0));
233 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
234 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
237 TargetLowering::getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
238 unsigned JTI,MCContext &Ctx) const{
239 // The normal PIC reloc base is the label at the start of the jump table.
240 return MCSymbolRefExpr::Create(MF->getJTISymbol(JTI, Ctx), Ctx);
244 TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
245 // Assume that everything is safe in static mode.
246 if (getTargetMachine().getRelocationModel() == Reloc::Static)
249 // In dynamic-no-pic mode, assume that known defined values are safe.
250 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC &&
252 !GA->getGlobal()->isDeclaration() &&
253 !GA->getGlobal()->isWeakForLinker())
256 // Otherwise assume nothing is safe.
260 //===----------------------------------------------------------------------===//
261 // Optimization Methods
262 //===----------------------------------------------------------------------===//
264 /// ShrinkDemandedConstant - Check to see if the specified operand of the
265 /// specified instruction is a constant integer. If so, check to see if there
266 /// are any bits set in the constant that are not demanded. If so, shrink the
267 /// constant and return true.
268 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op,
269 const APInt &Demanded) {
272 // FIXME: ISD::SELECT, ISD::SELECT_CC
273 switch (Op.getOpcode()) {
278 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
279 if (!C) return false;
281 if (Op.getOpcode() == ISD::XOR &&
282 (C->getAPIntValue() | (~Demanded)).isAllOnesValue())
285 // if we can expand it to have all bits set, do it
286 if (C->getAPIntValue().intersects(~Demanded)) {
287 EVT VT = Op.getValueType();
288 SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0),
289 DAG.getConstant(Demanded &
292 return CombineTo(Op, New);
302 /// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the
303 /// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening
304 /// cast, but it could be generalized for targets with other types of
305 /// implicit widening casts.
307 TargetLowering::TargetLoweringOpt::ShrinkDemandedOp(SDValue Op,
309 const APInt &Demanded,
311 assert(Op.getNumOperands() == 2 &&
312 "ShrinkDemandedOp only supports binary operators!");
313 assert(Op.getNode()->getNumValues() == 1 &&
314 "ShrinkDemandedOp only supports nodes with one result!");
316 // Don't do this if the node has another user, which may require the
318 if (!Op.getNode()->hasOneUse())
321 // Search for the smallest integer type with free casts to and from
322 // Op's type. For expedience, just check power-of-2 integer types.
323 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
324 unsigned DemandedSize = BitWidth - Demanded.countLeadingZeros();
325 unsigned SmallVTBits = DemandedSize;
326 if (!isPowerOf2_32(SmallVTBits))
327 SmallVTBits = NextPowerOf2(SmallVTBits);
328 for (; SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) {
329 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits);
330 if (TLI.isTruncateFree(Op.getValueType(), SmallVT) &&
331 TLI.isZExtFree(SmallVT, Op.getValueType())) {
332 // We found a type with free casts.
333 SDValue X = DAG.getNode(Op.getOpcode(), dl, SmallVT,
334 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
335 Op.getNode()->getOperand(0)),
336 DAG.getNode(ISD::TRUNCATE, dl, SmallVT,
337 Op.getNode()->getOperand(1)));
338 bool NeedZext = DemandedSize > SmallVTBits;
339 SDValue Z = DAG.getNode(NeedZext ? ISD::ZERO_EXTEND : ISD::ANY_EXTEND,
340 dl, Op.getValueType(), X);
341 return CombineTo(Op, Z);
347 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the
348 /// DemandedMask bits of the result of Op are ever used downstream. If we can
349 /// use this information to simplify Op, create a new simplified DAG node and
350 /// return true, returning the original and new nodes in Old and New. Otherwise,
351 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
352 /// the expression (used to simplify the caller). The KnownZero/One bits may
353 /// only be accurate for those bits in the DemandedMask.
354 bool TargetLowering::SimplifyDemandedBits(SDValue Op,
355 const APInt &DemandedMask,
358 TargetLoweringOpt &TLO,
359 unsigned Depth) const {
360 unsigned BitWidth = DemandedMask.getBitWidth();
361 assert(Op.getValueType().getScalarType().getSizeInBits() == BitWidth &&
362 "Mask size mismatches value type size!");
363 APInt NewMask = DemandedMask;
366 // Don't know anything.
367 KnownZero = KnownOne = APInt(BitWidth, 0);
369 // Other users may use these bits.
370 if (!Op.getNode()->hasOneUse()) {
372 // If not at the root, Just compute the KnownZero/KnownOne bits to
373 // simplify things downstream.
374 TLO.DAG.ComputeMaskedBits(Op, KnownZero, KnownOne, Depth);
377 // If this is the root being simplified, allow it to have multiple uses,
378 // just set the NewMask to all bits.
379 NewMask = APInt::getAllOnesValue(BitWidth);
380 } else if (DemandedMask == 0) {
381 // Not demanding any bits from Op.
382 if (Op.getOpcode() != ISD::UNDEF)
383 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType()));
385 } else if (Depth == 6) { // Limit search depth.
389 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut;
390 switch (Op.getOpcode()) {
392 // We know all of the bits for a constant!
393 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue();
394 KnownZero = ~KnownOne;
395 return false; // Don't fall through, will infinitely loop.
397 // If the RHS is a constant, check to see if the LHS would be zero without
398 // using the bits from the RHS. Below, we use knowledge about the RHS to
399 // simplify the LHS, here we're using information from the LHS to simplify
401 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
402 APInt LHSZero, LHSOne;
403 // Do not increment Depth here; that can cause an infinite loop.
404 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), LHSZero, LHSOne, Depth);
405 // If the LHS already has zeros where RHSC does, this and is dead.
406 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask))
407 return TLO.CombineTo(Op, Op.getOperand(0));
408 // If any of the set bits in the RHS are known zero on the LHS, shrink
410 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask))
414 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
415 KnownOne, TLO, Depth+1))
417 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
418 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask,
419 KnownZero2, KnownOne2, TLO, Depth+1))
421 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
423 // If all of the demanded bits are known one on one side, return the other.
424 // These bits cannot contribute to the result of the 'and'.
425 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
426 return TLO.CombineTo(Op, Op.getOperand(0));
427 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
428 return TLO.CombineTo(Op, Op.getOperand(1));
429 // If all of the demanded bits in the inputs are known zeros, return zero.
430 if ((NewMask & (KnownZero|KnownZero2)) == NewMask)
431 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType()));
432 // If the RHS is a constant, see if we can simplify it.
433 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask))
435 // If the operation can be done in a smaller type, do so.
436 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
439 // Output known-1 bits are only known if set in both the LHS & RHS.
440 KnownOne &= KnownOne2;
441 // Output known-0 are known to be clear if zero in either the LHS | RHS.
442 KnownZero |= KnownZero2;
445 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
446 KnownOne, TLO, Depth+1))
448 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
449 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask,
450 KnownZero2, KnownOne2, TLO, Depth+1))
452 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
454 // If all of the demanded bits are known zero on one side, return the other.
455 // These bits cannot contribute to the result of the 'or'.
456 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask))
457 return TLO.CombineTo(Op, Op.getOperand(0));
458 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask))
459 return TLO.CombineTo(Op, Op.getOperand(1));
460 // If all of the potentially set bits on one side are known to be set on
461 // the other side, just use the 'other' side.
462 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask))
463 return TLO.CombineTo(Op, Op.getOperand(0));
464 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask))
465 return TLO.CombineTo(Op, Op.getOperand(1));
466 // If the RHS is a constant, see if we can simplify it.
467 if (TLO.ShrinkDemandedConstant(Op, NewMask))
469 // If the operation can be done in a smaller type, do so.
470 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
473 // Output known-0 bits are only known if clear in both the LHS & RHS.
474 KnownZero &= KnownZero2;
475 // Output known-1 are known to be set if set in either the LHS | RHS.
476 KnownOne |= KnownOne2;
479 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero,
480 KnownOne, TLO, Depth+1))
482 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
483 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2,
484 KnownOne2, TLO, Depth+1))
486 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
488 // If all of the demanded bits are known zero on one side, return the other.
489 // These bits cannot contribute to the result of the 'xor'.
490 if ((KnownZero & NewMask) == NewMask)
491 return TLO.CombineTo(Op, Op.getOperand(0));
492 if ((KnownZero2 & NewMask) == NewMask)
493 return TLO.CombineTo(Op, Op.getOperand(1));
494 // If the operation can be done in a smaller type, do so.
495 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
498 // If all of the unknown bits are known to be zero on one side or the other
499 // (but not both) turn this into an *inclusive* or.
500 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
501 if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
502 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(),
506 // Output known-0 bits are known if clear or set in both the LHS & RHS.
507 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
508 // Output known-1 are known to be set if set in only one of the LHS, RHS.
509 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
511 // If all of the demanded bits on one side are known, and all of the set
512 // bits on that side are also known to be set on the other side, turn this
513 // into an AND, as we know the bits will be cleared.
514 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
515 // NB: it is okay if more bits are known than are requested
516 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known on one side
517 if (KnownOne == KnownOne2) { // set bits are the same on both sides
518 EVT VT = Op.getValueType();
519 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT);
520 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT,
521 Op.getOperand(0), ANDC));
525 // If the RHS is a constant, see if we can simplify it.
526 // for XOR, we prefer to force bits to 1 if they will make a -1.
527 // if we can't force bits, try to shrink constant
528 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
529 APInt Expanded = C->getAPIntValue() | (~NewMask);
530 // if we can expand it to have all bits set, do it
531 if (Expanded.isAllOnesValue()) {
532 if (Expanded != C->getAPIntValue()) {
533 EVT VT = Op.getValueType();
534 SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0),
535 TLO.DAG.getConstant(Expanded, VT));
536 return TLO.CombineTo(Op, New);
538 // if it already has all the bits set, nothing to change
539 // but don't shrink either!
540 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) {
545 KnownZero = KnownZeroOut;
546 KnownOne = KnownOneOut;
549 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero,
550 KnownOne, TLO, Depth+1))
552 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2,
553 KnownOne2, TLO, Depth+1))
555 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
556 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
558 // If the operands are constants, see if we can simplify them.
559 if (TLO.ShrinkDemandedConstant(Op, NewMask))
562 // Only known if known in both the LHS and RHS.
563 KnownOne &= KnownOne2;
564 KnownZero &= KnownZero2;
567 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero,
568 KnownOne, TLO, Depth+1))
570 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2,
571 KnownOne2, TLO, Depth+1))
573 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
574 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
576 // If the operands are constants, see if we can simplify them.
577 if (TLO.ShrinkDemandedConstant(Op, NewMask))
580 // Only known if known in both the LHS and RHS.
581 KnownOne &= KnownOne2;
582 KnownZero &= KnownZero2;
585 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
586 unsigned ShAmt = SA->getZExtValue();
587 SDValue InOp = Op.getOperand(0);
589 // If the shift count is an invalid immediate, don't do anything.
590 if (ShAmt >= BitWidth)
593 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
594 // single shift. We can do this if the bottom bits (which are shifted
595 // out) are never demanded.
596 if (InOp.getOpcode() == ISD::SRL &&
597 isa<ConstantSDNode>(InOp.getOperand(1))) {
598 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) {
599 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
600 unsigned Opc = ISD::SHL;
608 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
609 EVT VT = Op.getValueType();
610 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
611 InOp.getOperand(0), NewSA));
615 if (SimplifyDemandedBits(InOp, NewMask.lshr(ShAmt),
616 KnownZero, KnownOne, TLO, Depth+1))
619 // Convert (shl (anyext x, c)) to (anyext (shl x, c)) if the high bits
620 // are not demanded. This will likely allow the anyext to be folded away.
621 if (InOp.getNode()->getOpcode() == ISD::ANY_EXTEND) {
622 SDValue InnerOp = InOp.getNode()->getOperand(0);
623 EVT InnerVT = InnerOp.getValueType();
624 unsigned InnerBits = InnerVT.getSizeInBits();
625 if (ShAmt < InnerBits && NewMask.lshr(InnerBits) == 0 &&
626 isTypeDesirableForOp(ISD::SHL, InnerVT)) {
627 EVT ShTy = getShiftAmountTy(InnerVT);
628 if (!APInt(BitWidth, ShAmt).isIntN(ShTy.getSizeInBits()))
631 TLO.DAG.getNode(ISD::SHL, dl, InnerVT, InnerOp,
632 TLO.DAG.getConstant(ShAmt, ShTy));
635 TLO.DAG.getNode(ISD::ANY_EXTEND, dl, Op.getValueType(),
640 KnownZero <<= SA->getZExtValue();
641 KnownOne <<= SA->getZExtValue();
642 // low bits known zero.
643 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue());
647 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
648 EVT VT = Op.getValueType();
649 unsigned ShAmt = SA->getZExtValue();
650 unsigned VTSize = VT.getSizeInBits();
651 SDValue InOp = Op.getOperand(0);
653 // If the shift count is an invalid immediate, don't do anything.
654 if (ShAmt >= BitWidth)
657 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
658 // single shift. We can do this if the top bits (which are shifted out)
659 // are never demanded.
660 if (InOp.getOpcode() == ISD::SHL &&
661 isa<ConstantSDNode>(InOp.getOperand(1))) {
662 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) {
663 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue();
664 unsigned Opc = ISD::SRL;
672 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
673 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
674 InOp.getOperand(0), NewSA));
678 // Compute the new bits that are at the top now.
679 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt),
680 KnownZero, KnownOne, TLO, Depth+1))
682 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
683 KnownZero = KnownZero.lshr(ShAmt);
684 KnownOne = KnownOne.lshr(ShAmt);
686 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
687 KnownZero |= HighBits; // High bits known zero.
691 // If this is an arithmetic shift right and only the low-bit is set, we can
692 // always convert this into a logical shr, even if the shift amount is
693 // variable. The low bit of the shift cannot be an input sign bit unless
694 // the shift amount is >= the size of the datatype, which is undefined.
696 return TLO.CombineTo(Op,
697 TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(),
698 Op.getOperand(0), Op.getOperand(1)));
700 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
701 EVT VT = Op.getValueType();
702 unsigned ShAmt = SA->getZExtValue();
704 // If the shift count is an invalid immediate, don't do anything.
705 if (ShAmt >= BitWidth)
708 APInt InDemandedMask = (NewMask << ShAmt);
710 // If any of the demanded bits are produced by the sign extension, we also
711 // demand the input sign bit.
712 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt);
713 if (HighBits.intersects(NewMask))
714 InDemandedMask |= APInt::getSignBit(VT.getScalarType().getSizeInBits());
716 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask,
717 KnownZero, KnownOne, TLO, Depth+1))
719 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
720 KnownZero = KnownZero.lshr(ShAmt);
721 KnownOne = KnownOne.lshr(ShAmt);
723 // Handle the sign bit, adjusted to where it is now in the mask.
724 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt);
726 // If the input sign bit is known to be zero, or if none of the top bits
727 // are demanded, turn this into an unsigned shift right.
728 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
729 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT,
732 } else if (KnownOne.intersects(SignBit)) { // New bits are known one.
733 KnownOne |= HighBits;
737 case ISD::SIGN_EXTEND_INREG: {
738 EVT ExVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
740 APInt MsbMask = APInt::getHighBitsSet(BitWidth, 1);
741 // If we only care about the highest bit, don't bother shifting right.
742 if (MsbMask == DemandedMask) {
743 unsigned ShAmt = ExVT.getScalarType().getSizeInBits();
744 SDValue InOp = Op.getOperand(0);
746 // Compute the correct shift amount type, which must be getShiftAmountTy
747 // for scalar types after legalization.
748 EVT ShiftAmtTy = Op.getValueType();
749 if (TLO.LegalTypes() && !ShiftAmtTy.isVector())
750 ShiftAmtTy = getShiftAmountTy(ShiftAmtTy);
752 SDValue ShiftAmt = TLO.DAG.getConstant(BitWidth - ShAmt, ShiftAmtTy);
753 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl,
754 Op.getValueType(), InOp, ShiftAmt));
757 // Sign extension. Compute the demanded bits in the result that are not
758 // present in the input.
760 APInt::getHighBitsSet(BitWidth,
761 BitWidth - ExVT.getScalarType().getSizeInBits());
763 // If none of the extended bits are demanded, eliminate the sextinreg.
764 if ((NewBits & NewMask) == 0)
765 return TLO.CombineTo(Op, Op.getOperand(0));
768 APInt::getSignBit(ExVT.getScalarType().getSizeInBits()).zext(BitWidth);
769 APInt InputDemandedBits =
770 APInt::getLowBitsSet(BitWidth,
771 ExVT.getScalarType().getSizeInBits()) &
774 // Since the sign extended bits are demanded, we know that the sign
776 InputDemandedBits |= InSignBit;
778 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits,
779 KnownZero, KnownOne, TLO, Depth+1))
781 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
783 // If the sign bit of the input is known set or clear, then we know the
784 // top bits of the result.
786 // If the input sign bit is known zero, convert this into a zero extension.
787 if (KnownZero.intersects(InSignBit))
788 return TLO.CombineTo(Op,
789 TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,ExVT));
791 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
793 KnownZero &= ~NewBits;
794 } else { // Input sign bit unknown
795 KnownZero &= ~NewBits;
796 KnownOne &= ~NewBits;
800 case ISD::ZERO_EXTEND: {
801 unsigned OperandBitWidth =
802 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
803 APInt InMask = NewMask.trunc(OperandBitWidth);
805 // If none of the top bits are demanded, convert this into an any_extend.
807 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
808 if (!NewBits.intersects(NewMask))
809 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
813 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
814 KnownZero, KnownOne, TLO, Depth+1))
816 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
817 KnownZero = KnownZero.zext(BitWidth);
818 KnownOne = KnownOne.zext(BitWidth);
819 KnownZero |= NewBits;
822 case ISD::SIGN_EXTEND: {
823 EVT InVT = Op.getOperand(0).getValueType();
824 unsigned InBits = InVT.getScalarType().getSizeInBits();
825 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits);
826 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits);
827 APInt NewBits = ~InMask & NewMask;
829 // If none of the top bits are demanded, convert this into an any_extend.
831 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
835 // Since some of the sign extended bits are demanded, we know that the sign
837 APInt InDemandedBits = InMask & NewMask;
838 InDemandedBits |= InSignBit;
839 InDemandedBits = InDemandedBits.trunc(InBits);
841 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero,
842 KnownOne, TLO, Depth+1))
844 KnownZero = KnownZero.zext(BitWidth);
845 KnownOne = KnownOne.zext(BitWidth);
847 // If the sign bit is known zero, convert this to a zero extend.
848 if (KnownZero.intersects(InSignBit))
849 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl,
853 // If the sign bit is known one, the top bits match.
854 if (KnownOne.intersects(InSignBit)) {
856 assert((KnownZero & NewBits) == 0);
857 } else { // Otherwise, top bits aren't known.
858 assert((KnownOne & NewBits) == 0);
859 assert((KnownZero & NewBits) == 0);
863 case ISD::ANY_EXTEND: {
864 unsigned OperandBitWidth =
865 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
866 APInt InMask = NewMask.trunc(OperandBitWidth);
867 if (SimplifyDemandedBits(Op.getOperand(0), InMask,
868 KnownZero, KnownOne, TLO, Depth+1))
870 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
871 KnownZero = KnownZero.zext(BitWidth);
872 KnownOne = KnownOne.zext(BitWidth);
875 case ISD::TRUNCATE: {
876 // Simplify the input, using demanded bit information, and compute the known
877 // zero/one bits live out.
878 unsigned OperandBitWidth =
879 Op.getOperand(0).getValueType().getScalarType().getSizeInBits();
880 APInt TruncMask = NewMask.zext(OperandBitWidth);
881 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask,
882 KnownZero, KnownOne, TLO, Depth+1))
884 KnownZero = KnownZero.trunc(BitWidth);
885 KnownOne = KnownOne.trunc(BitWidth);
887 // If the input is only used by this truncate, see if we can shrink it based
888 // on the known demanded bits.
889 if (Op.getOperand(0).getNode()->hasOneUse()) {
890 SDValue In = Op.getOperand(0);
891 switch (In.getOpcode()) {
894 // Shrink SRL by a constant if none of the high bits shifted in are
896 if (TLO.LegalTypes() &&
897 !isTypeDesirableForOp(ISD::SRL, Op.getValueType()))
898 // Do not turn (vt1 truncate (vt2 srl)) into (vt1 srl) if vt1 is
901 ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1));
904 SDValue Shift = In.getOperand(1);
905 if (TLO.LegalTypes()) {
906 uint64_t ShVal = ShAmt->getZExtValue();
908 TLO.DAG.getConstant(ShVal, getShiftAmountTy(Op.getValueType()));
911 APInt HighBits = APInt::getHighBitsSet(OperandBitWidth,
912 OperandBitWidth - BitWidth);
913 HighBits = HighBits.lshr(ShAmt->getZExtValue()).trunc(BitWidth);
915 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) {
916 // None of the shifted in bits are needed. Add a truncate of the
917 // shift input, then shift it.
918 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl,
921 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl,
930 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
933 case ISD::AssertZext: {
934 // AssertZext demands all of the high bits, plus any of the low bits
935 // demanded by its users.
936 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT();
937 APInt InMask = APInt::getLowBitsSet(BitWidth,
939 if (SimplifyDemandedBits(Op.getOperand(0), ~InMask | NewMask,
940 KnownZero, KnownOne, TLO, Depth+1))
942 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
944 KnownZero |= ~InMask & NewMask;
948 // If this is an FP->Int bitcast and if the sign bit is the only
949 // thing demanded, turn this into a FGETSIGN.
950 if (!TLO.LegalOperations() &&
951 !Op.getValueType().isVector() &&
952 !Op.getOperand(0).getValueType().isVector() &&
953 NewMask == APInt::getSignBit(Op.getValueType().getSizeInBits()) &&
954 Op.getOperand(0).getValueType().isFloatingPoint()) {
955 bool OpVTLegal = isOperationLegalOrCustom(ISD::FGETSIGN, Op.getValueType());
956 bool i32Legal = isOperationLegalOrCustom(ISD::FGETSIGN, MVT::i32);
957 if ((OpVTLegal || i32Legal) && Op.getValueType().isSimple()) {
958 EVT Ty = OpVTLegal ? Op.getValueType() : MVT::i32;
959 // Make a FGETSIGN + SHL to move the sign bit into the appropriate
960 // place. We expect the SHL to be eliminated by other optimizations.
961 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, dl, Ty, Op.getOperand(0));
962 unsigned OpVTSizeInBits = Op.getValueType().getSizeInBits();
963 if (!OpVTLegal && OpVTSizeInBits > 32)
964 Sign = TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), Sign);
965 unsigned ShVal = Op.getValueType().getSizeInBits()-1;
966 SDValue ShAmt = TLO.DAG.getConstant(ShVal, Op.getValueType());
967 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl,
976 // Add, Sub, and Mul don't demand any bits in positions beyond that
977 // of the highest bit demanded of them.
978 APInt LoMask = APInt::getLowBitsSet(BitWidth,
979 BitWidth - NewMask.countLeadingZeros());
980 if (SimplifyDemandedBits(Op.getOperand(0), LoMask, KnownZero2,
981 KnownOne2, TLO, Depth+1))
983 if (SimplifyDemandedBits(Op.getOperand(1), LoMask, KnownZero2,
984 KnownOne2, TLO, Depth+1))
986 // See if the operation should be performed at a smaller bit width.
987 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl))
992 // Just use ComputeMaskedBits to compute output bits.
993 TLO.DAG.ComputeMaskedBits(Op, KnownZero, KnownOne, Depth);
997 // If we know the value of all of the demanded bits, return this as a
999 if ((NewMask & (KnownZero|KnownOne)) == NewMask)
1000 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType()));
1005 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
1006 /// in Mask are known to be either zero or one and return them in the
1007 /// KnownZero/KnownOne bitsets.
1008 void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
1011 const SelectionDAG &DAG,
1012 unsigned Depth) const {
1013 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1014 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1015 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1016 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1017 "Should use MaskedValueIsZero if you don't know whether Op"
1018 " is a target node!");
1019 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0);
1022 /// ComputeNumSignBitsForTargetNode - This method can be implemented by
1023 /// targets that want to expose additional information about sign bits to the
1025 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
1026 unsigned Depth) const {
1027 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||
1028 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||
1029 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||
1030 Op.getOpcode() == ISD::INTRINSIC_VOID) &&
1031 "Should use ComputeNumSignBits if you don't know whether Op"
1032 " is a target node!");
1036 /// ValueHasExactlyOneBitSet - Test if the given value is known to have exactly
1037 /// one bit set. This differs from ComputeMaskedBits in that it doesn't need to
1038 /// determine which bit is set.
1040 static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) {
1041 // A left-shift of a constant one will have exactly one bit set, because
1042 // shifting the bit off the end is undefined.
1043 if (Val.getOpcode() == ISD::SHL)
1044 if (ConstantSDNode *C =
1045 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1046 if (C->getAPIntValue() == 1)
1049 // Similarly, a right-shift of a constant sign-bit will have exactly
1051 if (Val.getOpcode() == ISD::SRL)
1052 if (ConstantSDNode *C =
1053 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
1054 if (C->getAPIntValue().isSignBit())
1057 // More could be done here, though the above checks are enough
1058 // to handle some common cases.
1060 // Fall back to ComputeMaskedBits to catch other known cases.
1061 EVT OpVT = Val.getValueType();
1062 unsigned BitWidth = OpVT.getScalarType().getSizeInBits();
1063 APInt KnownZero, KnownOne;
1064 DAG.ComputeMaskedBits(Val, KnownZero, KnownOne);
1065 return (KnownZero.countPopulation() == BitWidth - 1) &&
1066 (KnownOne.countPopulation() == 1);
1069 /// SimplifySetCC - Try to simplify a setcc built with the specified operands
1070 /// and cc. If it is unable to simplify it, return a null SDValue.
1072 TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
1073 ISD::CondCode Cond, bool foldBooleans,
1074 DAGCombinerInfo &DCI, SDLoc dl) const {
1075 SelectionDAG &DAG = DCI.DAG;
1077 // These setcc operations always fold.
1081 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
1083 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
1086 // Ensure that the constant occurs on the RHS, and fold constant
1088 if (isa<ConstantSDNode>(N0.getNode()))
1089 return DAG.getSetCC(dl, VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
1091 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
1092 const APInt &C1 = N1C->getAPIntValue();
1094 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
1095 // equality comparison, then we're just comparing whether X itself is
1097 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
1098 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
1099 N0.getOperand(1).getOpcode() == ISD::Constant) {
1101 = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
1102 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1103 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) {
1104 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
1105 // (srl (ctlz x), 5) == 0 -> X != 0
1106 // (srl (ctlz x), 5) != 1 -> X != 0
1109 // (srl (ctlz x), 5) != 0 -> X == 0
1110 // (srl (ctlz x), 5) == 1 -> X == 0
1113 SDValue Zero = DAG.getConstant(0, N0.getValueType());
1114 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0),
1120 // Look through truncs that don't change the value of a ctpop.
1121 if (N0.hasOneUse() && N0.getOpcode() == ISD::TRUNCATE)
1122 CTPOP = N0.getOperand(0);
1124 if (CTPOP.hasOneUse() && CTPOP.getOpcode() == ISD::CTPOP &&
1125 (N0 == CTPOP || N0.getValueType().getSizeInBits() >
1126 Log2_32_Ceil(CTPOP.getValueType().getSizeInBits()))) {
1127 EVT CTVT = CTPOP.getValueType();
1128 SDValue CTOp = CTPOP.getOperand(0);
1130 // (ctpop x) u< 2 -> (x & x-1) == 0
1131 // (ctpop x) u> 1 -> (x & x-1) != 0
1132 if ((Cond == ISD::SETULT && C1 == 2) || (Cond == ISD::SETUGT && C1 == 1)){
1133 SDValue Sub = DAG.getNode(ISD::SUB, dl, CTVT, CTOp,
1134 DAG.getConstant(1, CTVT));
1135 SDValue And = DAG.getNode(ISD::AND, dl, CTVT, CTOp, Sub);
1136 ISD::CondCode CC = Cond == ISD::SETULT ? ISD::SETEQ : ISD::SETNE;
1137 return DAG.getSetCC(dl, VT, And, DAG.getConstant(0, CTVT), CC);
1140 // TODO: (ctpop x) == 1 -> x && (x & x-1) == 0 iff ctpop is illegal.
1143 // (zext x) == C --> x == (trunc C)
1144 if (DCI.isBeforeLegalize() && N0->hasOneUse() &&
1145 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1146 unsigned MinBits = N0.getValueSizeInBits();
1148 if (N0->getOpcode() == ISD::ZERO_EXTEND) {
1150 MinBits = N0->getOperand(0).getValueSizeInBits();
1151 PreZExt = N0->getOperand(0);
1152 } else if (N0->getOpcode() == ISD::AND) {
1153 // DAGCombine turns costly ZExts into ANDs
1154 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0->getOperand(1)))
1155 if ((C->getAPIntValue()+1).isPowerOf2()) {
1156 MinBits = C->getAPIntValue().countTrailingOnes();
1157 PreZExt = N0->getOperand(0);
1159 } else if (LoadSDNode *LN0 = dyn_cast<LoadSDNode>(N0)) {
1161 if (LN0->getExtensionType() == ISD::ZEXTLOAD) {
1162 MinBits = LN0->getMemoryVT().getSizeInBits();
1167 // Make sure we're not losing bits from the constant.
1169 MinBits < C1.getBitWidth() && MinBits >= C1.getActiveBits()) {
1170 EVT MinVT = EVT::getIntegerVT(*DAG.getContext(), MinBits);
1171 if (isTypeDesirableForOp(ISD::SETCC, MinVT)) {
1172 // Will get folded away.
1173 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MinVT, PreZExt);
1174 SDValue C = DAG.getConstant(C1.trunc(MinBits), MinVT);
1175 return DAG.getSetCC(dl, VT, Trunc, C, Cond);
1180 // If the LHS is '(and load, const)', the RHS is 0,
1181 // the test is for equality or unsigned, and all 1 bits of the const are
1182 // in the same partial word, see if we can shorten the load.
1183 if (DCI.isBeforeLegalize() &&
1184 N0.getOpcode() == ISD::AND && C1 == 0 &&
1185 N0.getNode()->hasOneUse() &&
1186 isa<LoadSDNode>(N0.getOperand(0)) &&
1187 N0.getOperand(0).getNode()->hasOneUse() &&
1188 isa<ConstantSDNode>(N0.getOperand(1))) {
1189 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0));
1191 unsigned bestWidth = 0, bestOffset = 0;
1192 if (!Lod->isVolatile() && Lod->isUnindexed()) {
1193 unsigned origWidth = N0.getValueType().getSizeInBits();
1194 unsigned maskWidth = origWidth;
1195 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to
1196 // 8 bits, but have to be careful...
1197 if (Lod->getExtensionType() != ISD::NON_EXTLOAD)
1198 origWidth = Lod->getMemoryVT().getSizeInBits();
1200 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
1201 for (unsigned width = origWidth / 2; width>=8; width /= 2) {
1202 APInt newMask = APInt::getLowBitsSet(maskWidth, width);
1203 for (unsigned offset=0; offset<origWidth/width; offset++) {
1204 if ((newMask & Mask) == Mask) {
1205 if (!getDataLayout()->isLittleEndian())
1206 bestOffset = (origWidth/width - offset - 1) * (width/8);
1208 bestOffset = (uint64_t)offset * (width/8);
1209 bestMask = Mask.lshr(offset * (width/8) * 8);
1213 newMask = newMask << width;
1218 EVT newVT = EVT::getIntegerVT(*DAG.getContext(), bestWidth);
1219 if (newVT.isRound()) {
1220 EVT PtrType = Lod->getOperand(1).getValueType();
1221 SDValue Ptr = Lod->getBasePtr();
1222 if (bestOffset != 0)
1223 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(),
1224 DAG.getConstant(bestOffset, PtrType));
1225 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset);
1226 SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr,
1227 Lod->getPointerInfo().getWithOffset(bestOffset),
1228 false, false, false, NewAlign);
1229 return DAG.getSetCC(dl, VT,
1230 DAG.getNode(ISD::AND, dl, newVT, NewLoad,
1231 DAG.getConstant(bestMask.trunc(bestWidth),
1233 DAG.getConstant(0LL, newVT), Cond);
1238 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
1239 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
1240 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits();
1242 // If the comparison constant has bits in the upper part, the
1243 // zero-extended value could never match.
1244 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(),
1245 C1.getBitWidth() - InSize))) {
1249 case ISD::SETEQ: return DAG.getConstant(0, VT);
1252 case ISD::SETNE: return DAG.getConstant(1, VT);
1255 // True if the sign bit of C1 is set.
1256 return DAG.getConstant(C1.isNegative(), VT);
1259 // True if the sign bit of C1 isn't set.
1260 return DAG.getConstant(C1.isNonNegative(), VT);
1266 // Otherwise, we can perform the comparison with the low bits.
1274 EVT newVT = N0.getOperand(0).getValueType();
1275 if (DCI.isBeforeLegalizeOps() ||
1276 (isOperationLegal(ISD::SETCC, newVT) &&
1277 getCondCodeAction(Cond, newVT.getSimpleVT())==Legal))
1278 return DAG.getSetCC(dl, VT, N0.getOperand(0),
1279 DAG.getConstant(C1.trunc(InSize), newVT),
1284 break; // todo, be more careful with signed comparisons
1286 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1287 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1288 EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
1289 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits();
1290 EVT ExtDstTy = N0.getValueType();
1291 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits();
1293 // If the constant doesn't fit into the number of bits for the source of
1294 // the sign extension, it is impossible for both sides to be equal.
1295 if (C1.getMinSignedBits() > ExtSrcTyBits)
1296 return DAG.getConstant(Cond == ISD::SETNE, VT);
1299 EVT Op0Ty = N0.getOperand(0).getValueType();
1300 if (Op0Ty == ExtSrcTy) {
1301 ZextOp = N0.getOperand(0);
1303 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
1304 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0),
1305 DAG.getConstant(Imm, Op0Ty));
1307 if (!DCI.isCalledByLegalizer())
1308 DCI.AddToWorklist(ZextOp.getNode());
1309 // Otherwise, make this a use of a zext.
1310 return DAG.getSetCC(dl, VT, ZextOp,
1311 DAG.getConstant(C1 & APInt::getLowBitsSet(
1316 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) &&
1317 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
1318 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
1319 if (N0.getOpcode() == ISD::SETCC &&
1320 isTypeLegal(VT) && VT.bitsLE(N0.getValueType())) {
1321 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getAPIntValue() != 1);
1323 return DAG.getNode(ISD::TRUNCATE, dl, VT, N0);
1324 // Invert the condition.
1325 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
1326 CC = ISD::getSetCCInverse(CC,
1327 N0.getOperand(0).getValueType().isInteger());
1328 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC);
1331 if ((N0.getOpcode() == ISD::XOR ||
1332 (N0.getOpcode() == ISD::AND &&
1333 N0.getOperand(0).getOpcode() == ISD::XOR &&
1334 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
1335 isa<ConstantSDNode>(N0.getOperand(1)) &&
1336 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) {
1337 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
1338 // can only do this if the top bits are known zero.
1339 unsigned BitWidth = N0.getValueSizeInBits();
1340 if (DAG.MaskedValueIsZero(N0,
1341 APInt::getHighBitsSet(BitWidth,
1343 // Okay, get the un-inverted input value.
1345 if (N0.getOpcode() == ISD::XOR)
1346 Val = N0.getOperand(0);
1348 assert(N0.getOpcode() == ISD::AND &&
1349 N0.getOperand(0).getOpcode() == ISD::XOR);
1350 // ((X^1)&1)^1 -> X & 1
1351 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(),
1352 N0.getOperand(0).getOperand(0),
1356 return DAG.getSetCC(dl, VT, Val, N1,
1357 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1359 } else if (N1C->getAPIntValue() == 1 &&
1361 getBooleanContents(false) == ZeroOrOneBooleanContent)) {
1363 if (Op0.getOpcode() == ISD::TRUNCATE)
1364 Op0 = Op0.getOperand(0);
1366 if ((Op0.getOpcode() == ISD::XOR) &&
1367 Op0.getOperand(0).getOpcode() == ISD::SETCC &&
1368 Op0.getOperand(1).getOpcode() == ISD::SETCC) {
1369 // (xor (setcc), (setcc)) == / != 1 -> (setcc) != / == (setcc)
1370 Cond = (Cond == ISD::SETEQ) ? ISD::SETNE : ISD::SETEQ;
1371 return DAG.getSetCC(dl, VT, Op0.getOperand(0), Op0.getOperand(1),
1374 if (Op0.getOpcode() == ISD::AND &&
1375 isa<ConstantSDNode>(Op0.getOperand(1)) &&
1376 cast<ConstantSDNode>(Op0.getOperand(1))->getAPIntValue() == 1) {
1377 // If this is (X&1) == / != 1, normalize it to (X&1) != / == 0.
1378 if (Op0.getValueType().bitsGT(VT))
1379 Op0 = DAG.getNode(ISD::AND, dl, VT,
1380 DAG.getNode(ISD::TRUNCATE, dl, VT, Op0.getOperand(0)),
1381 DAG.getConstant(1, VT));
1382 else if (Op0.getValueType().bitsLT(VT))
1383 Op0 = DAG.getNode(ISD::AND, dl, VT,
1384 DAG.getNode(ISD::ANY_EXTEND, dl, VT, Op0.getOperand(0)),
1385 DAG.getConstant(1, VT));
1387 return DAG.getSetCC(dl, VT, Op0,
1388 DAG.getConstant(0, Op0.getValueType()),
1389 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1391 if (Op0.getOpcode() == ISD::AssertZext &&
1392 cast<VTSDNode>(Op0.getOperand(1))->getVT() == MVT::i1)
1393 return DAG.getSetCC(dl, VT, Op0,
1394 DAG.getConstant(0, Op0.getValueType()),
1395 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
1399 APInt MinVal, MaxVal;
1400 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits();
1401 if (ISD::isSignedIntSetCC(Cond)) {
1402 MinVal = APInt::getSignedMinValue(OperandBitSize);
1403 MaxVal = APInt::getSignedMaxValue(OperandBitSize);
1405 MinVal = APInt::getMinValue(OperandBitSize);
1406 MaxVal = APInt::getMaxValue(OperandBitSize);
1409 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
1410 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
1411 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
1412 // X >= C0 --> X > (C0-1)
1413 return DAG.getSetCC(dl, VT, N0,
1414 DAG.getConstant(C1-1, N1.getValueType()),
1415 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
1418 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
1419 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
1420 // X <= C0 --> X < (C0+1)
1421 return DAG.getSetCC(dl, VT, N0,
1422 DAG.getConstant(C1+1, N1.getValueType()),
1423 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
1426 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
1427 return DAG.getConstant(0, VT); // X < MIN --> false
1428 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal)
1429 return DAG.getConstant(1, VT); // X >= MIN --> true
1430 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal)
1431 return DAG.getConstant(0, VT); // X > MAX --> false
1432 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal)
1433 return DAG.getConstant(1, VT); // X <= MAX --> true
1435 // Canonicalize setgt X, Min --> setne X, Min
1436 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
1437 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
1438 // Canonicalize setlt X, Max --> setne X, Max
1439 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
1440 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
1442 // If we have setult X, 1, turn it into seteq X, 0
1443 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
1444 return DAG.getSetCC(dl, VT, N0,
1445 DAG.getConstant(MinVal, N0.getValueType()),
1447 // If we have setugt X, Max-1, turn it into seteq X, Max
1448 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
1449 return DAG.getSetCC(dl, VT, N0,
1450 DAG.getConstant(MaxVal, N0.getValueType()),
1453 // If we have "setcc X, C0", check to see if we can shrink the immediate
1456 // SETUGT X, SINTMAX -> SETLT X, 0
1457 if (Cond == ISD::SETUGT &&
1458 C1 == APInt::getSignedMaxValue(OperandBitSize))
1459 return DAG.getSetCC(dl, VT, N0,
1460 DAG.getConstant(0, N1.getValueType()),
1463 // SETULT X, SINTMIN -> SETGT X, -1
1464 if (Cond == ISD::SETULT &&
1465 C1 == APInt::getSignedMinValue(OperandBitSize)) {
1466 SDValue ConstMinusOne =
1467 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize),
1469 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT);
1472 // Fold bit comparisons when we can.
1473 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1474 (VT == N0.getValueType() ||
1475 (isTypeLegal(VT) && VT.bitsLE(N0.getValueType()))) &&
1476 N0.getOpcode() == ISD::AND)
1477 if (ConstantSDNode *AndRHS =
1478 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1479 EVT ShiftTy = DCI.isBeforeLegalizeOps() ?
1480 getPointerTy() : getShiftAmountTy(N0.getValueType());
1481 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
1482 // Perform the xform if the AND RHS is a single bit.
1483 if (AndRHS->getAPIntValue().isPowerOf2()) {
1484 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1485 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0,
1486 DAG.getConstant(AndRHS->getAPIntValue().logBase2(), ShiftTy)));
1488 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getAPIntValue()) {
1489 // (X & 8) == 8 --> (X & 8) >> 3
1490 // Perform the xform if C1 is a single bit.
1491 if (C1.isPowerOf2()) {
1492 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1493 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0,
1494 DAG.getConstant(C1.logBase2(), ShiftTy)));
1499 if (C1.getMinSignedBits() <= 64 &&
1500 !isLegalICmpImmediate(C1.getSExtValue())) {
1501 // (X & -256) == 256 -> (X >> 8) == 1
1502 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1503 N0.getOpcode() == ISD::AND && N0.hasOneUse()) {
1504 if (ConstantSDNode *AndRHS =
1505 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1506 const APInt &AndRHSC = AndRHS->getAPIntValue();
1507 if ((-AndRHSC).isPowerOf2() && (AndRHSC & C1) == C1) {
1508 unsigned ShiftBits = AndRHSC.countTrailingZeros();
1509 EVT ShiftTy = DCI.isBeforeLegalizeOps() ?
1510 getPointerTy() : getShiftAmountTy(N0.getValueType());
1511 EVT CmpTy = N0.getValueType();
1512 SDValue Shift = DAG.getNode(ISD::SRL, dl, CmpTy, N0.getOperand(0),
1513 DAG.getConstant(ShiftBits, ShiftTy));
1514 SDValue CmpRHS = DAG.getConstant(C1.lshr(ShiftBits), CmpTy);
1515 return DAG.getSetCC(dl, VT, Shift, CmpRHS, Cond);
1518 } else if (Cond == ISD::SETULT || Cond == ISD::SETUGE ||
1519 Cond == ISD::SETULE || Cond == ISD::SETUGT) {
1520 bool AdjOne = (Cond == ISD::SETULE || Cond == ISD::SETUGT);
1521 // X < 0x100000000 -> (X >> 32) < 1
1522 // X >= 0x100000000 -> (X >> 32) >= 1
1523 // X <= 0x0ffffffff -> (X >> 32) < 1
1524 // X > 0x0ffffffff -> (X >> 32) >= 1
1527 ISD::CondCode NewCond = Cond;
1529 ShiftBits = C1.countTrailingOnes();
1531 NewCond = (Cond == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1533 ShiftBits = C1.countTrailingZeros();
1535 NewC = NewC.lshr(ShiftBits);
1536 if (ShiftBits && isLegalICmpImmediate(NewC.getSExtValue())) {
1537 EVT ShiftTy = DCI.isBeforeLegalizeOps() ?
1538 getPointerTy() : getShiftAmountTy(N0.getValueType());
1539 EVT CmpTy = N0.getValueType();
1540 SDValue Shift = DAG.getNode(ISD::SRL, dl, CmpTy, N0,
1541 DAG.getConstant(ShiftBits, ShiftTy));
1542 SDValue CmpRHS = DAG.getConstant(NewC, CmpTy);
1543 return DAG.getSetCC(dl, VT, Shift, CmpRHS, NewCond);
1549 if (isa<ConstantFPSDNode>(N0.getNode())) {
1550 // Constant fold or commute setcc.
1551 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl);
1552 if (O.getNode()) return O;
1553 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
1554 // If the RHS of an FP comparison is a constant, simplify it away in
1556 if (CFP->getValueAPF().isNaN()) {
1557 // If an operand is known to be a nan, we can fold it.
1558 switch (ISD::getUnorderedFlavor(Cond)) {
1559 default: llvm_unreachable("Unknown flavor!");
1560 case 0: // Known false.
1561 return DAG.getConstant(0, VT);
1562 case 1: // Known true.
1563 return DAG.getConstant(1, VT);
1564 case 2: // Undefined.
1565 return DAG.getUNDEF(VT);
1569 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the
1570 // constant if knowing that the operand is non-nan is enough. We prefer to
1571 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
1573 if (Cond == ISD::SETO || Cond == ISD::SETUO)
1574 return DAG.getSetCC(dl, VT, N0, N0, Cond);
1576 // If the condition is not legal, see if we can find an equivalent one
1578 if (!isCondCodeLegal(Cond, N0.getSimpleValueType())) {
1579 // If the comparison was an awkward floating-point == or != and one of
1580 // the comparison operands is infinity or negative infinity, convert the
1581 // condition to a less-awkward <= or >=.
1582 if (CFP->getValueAPF().isInfinity()) {
1583 if (CFP->getValueAPF().isNegative()) {
1584 if (Cond == ISD::SETOEQ &&
1585 isCondCodeLegal(ISD::SETOLE, N0.getSimpleValueType()))
1586 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLE);
1587 if (Cond == ISD::SETUEQ &&
1588 isCondCodeLegal(ISD::SETOLE, N0.getSimpleValueType()))
1589 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULE);
1590 if (Cond == ISD::SETUNE &&
1591 isCondCodeLegal(ISD::SETUGT, N0.getSimpleValueType()))
1592 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGT);
1593 if (Cond == ISD::SETONE &&
1594 isCondCodeLegal(ISD::SETUGT, N0.getSimpleValueType()))
1595 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGT);
1597 if (Cond == ISD::SETOEQ &&
1598 isCondCodeLegal(ISD::SETOGE, N0.getSimpleValueType()))
1599 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGE);
1600 if (Cond == ISD::SETUEQ &&
1601 isCondCodeLegal(ISD::SETOGE, N0.getSimpleValueType()))
1602 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGE);
1603 if (Cond == ISD::SETUNE &&
1604 isCondCodeLegal(ISD::SETULT, N0.getSimpleValueType()))
1605 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULT);
1606 if (Cond == ISD::SETONE &&
1607 isCondCodeLegal(ISD::SETULT, N0.getSimpleValueType()))
1608 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLT);
1615 // The sext(setcc()) => setcc() optimization relies on the appropriate
1616 // constant being emitted.
1618 switch (getBooleanContents(N0.getValueType().isVector())) {
1619 case UndefinedBooleanContent:
1620 case ZeroOrOneBooleanContent:
1621 EqVal = ISD::isTrueWhenEqual(Cond);
1623 case ZeroOrNegativeOneBooleanContent:
1624 EqVal = ISD::isTrueWhenEqual(Cond) ? -1 : 0;
1628 // We can always fold X == X for integer setcc's.
1629 if (N0.getValueType().isInteger()) {
1630 return DAG.getConstant(EqVal, VT);
1632 unsigned UOF = ISD::getUnorderedFlavor(Cond);
1633 if (UOF == 2) // FP operators that are undefined on NaNs.
1634 return DAG.getConstant(EqVal, VT);
1635 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
1636 return DAG.getConstant(EqVal, VT);
1637 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
1638 // if it is not already.
1639 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
1640 if (NewCond != Cond && (DCI.isBeforeLegalizeOps() ||
1641 getCondCodeAction(NewCond, N0.getSimpleValueType()) == Legal))
1642 return DAG.getSetCC(dl, VT, N0, N1, NewCond);
1645 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
1646 N0.getValueType().isInteger()) {
1647 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
1648 N0.getOpcode() == ISD::XOR) {
1649 // Simplify (X+Y) == (X+Z) --> Y == Z
1650 if (N0.getOpcode() == N1.getOpcode()) {
1651 if (N0.getOperand(0) == N1.getOperand(0))
1652 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond);
1653 if (N0.getOperand(1) == N1.getOperand(1))
1654 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond);
1655 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
1656 // If X op Y == Y op X, try other combinations.
1657 if (N0.getOperand(0) == N1.getOperand(1))
1658 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0),
1660 if (N0.getOperand(1) == N1.getOperand(0))
1661 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1),
1666 // If RHS is a legal immediate value for a compare instruction, we need
1667 // to be careful about increasing register pressure needlessly.
1668 bool LegalRHSImm = false;
1670 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
1671 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1672 // Turn (X+C1) == C2 --> X == C2-C1
1673 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) {
1674 return DAG.getSetCC(dl, VT, N0.getOperand(0),
1675 DAG.getConstant(RHSC->getAPIntValue()-
1676 LHSR->getAPIntValue(),
1677 N0.getValueType()), Cond);
1680 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
1681 if (N0.getOpcode() == ISD::XOR)
1682 // If we know that all of the inverted bits are zero, don't bother
1683 // performing the inversion.
1684 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
1686 DAG.getSetCC(dl, VT, N0.getOperand(0),
1687 DAG.getConstant(LHSR->getAPIntValue() ^
1688 RHSC->getAPIntValue(),
1693 // Turn (C1-X) == C2 --> X == C1-C2
1694 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
1695 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) {
1697 DAG.getSetCC(dl, VT, N0.getOperand(1),
1698 DAG.getConstant(SUBC->getAPIntValue() -
1699 RHSC->getAPIntValue(),
1705 // Could RHSC fold directly into a compare?
1706 if (RHSC->getValueType(0).getSizeInBits() <= 64)
1707 LegalRHSImm = isLegalICmpImmediate(RHSC->getSExtValue());
1710 // Simplify (X+Z) == X --> Z == 0
1711 // Don't do this if X is an immediate that can fold into a cmp
1712 // instruction and X+Z has other uses. It could be an induction variable
1713 // chain, and the transform would increase register pressure.
1714 if (!LegalRHSImm || N0.getNode()->hasOneUse()) {
1715 if (N0.getOperand(0) == N1)
1716 return DAG.getSetCC(dl, VT, N0.getOperand(1),
1717 DAG.getConstant(0, N0.getValueType()), Cond);
1718 if (N0.getOperand(1) == N1) {
1719 if (DAG.isCommutativeBinOp(N0.getOpcode()))
1720 return DAG.getSetCC(dl, VT, N0.getOperand(0),
1721 DAG.getConstant(0, N0.getValueType()), Cond);
1722 if (N0.getNode()->hasOneUse()) {
1723 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
1724 // (Z-X) == X --> Z == X<<1
1725 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N1,
1726 DAG.getConstant(1, getShiftAmountTy(N1.getValueType())));
1727 if (!DCI.isCalledByLegalizer())
1728 DCI.AddToWorklist(SH.getNode());
1729 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond);
1735 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
1736 N1.getOpcode() == ISD::XOR) {
1737 // Simplify X == (X+Z) --> Z == 0
1738 if (N1.getOperand(0) == N0)
1739 return DAG.getSetCC(dl, VT, N1.getOperand(1),
1740 DAG.getConstant(0, N1.getValueType()), Cond);
1741 if (N1.getOperand(1) == N0) {
1742 if (DAG.isCommutativeBinOp(N1.getOpcode()))
1743 return DAG.getSetCC(dl, VT, N1.getOperand(0),
1744 DAG.getConstant(0, N1.getValueType()), Cond);
1745 if (N1.getNode()->hasOneUse()) {
1746 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
1747 // X == (Z-X) --> X<<1 == Z
1748 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N0,
1749 DAG.getConstant(1, getShiftAmountTy(N0.getValueType())));
1750 if (!DCI.isCalledByLegalizer())
1751 DCI.AddToWorklist(SH.getNode());
1752 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond);
1757 // Simplify x&y == y to x&y != 0 if y has exactly one bit set.
1758 // Note that where y is variable and is known to have at most
1759 // one bit set (for example, if it is z&1) we cannot do this;
1760 // the expressions are not equivalent when y==0.
1761 if (N0.getOpcode() == ISD::AND)
1762 if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) {
1763 if (ValueHasExactlyOneBitSet(N1, DAG)) {
1764 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
1765 SDValue Zero = DAG.getConstant(0, N1.getValueType());
1766 return DAG.getSetCC(dl, VT, N0, Zero, Cond);
1769 if (N1.getOpcode() == ISD::AND)
1770 if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) {
1771 if (ValueHasExactlyOneBitSet(N0, DAG)) {
1772 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
1773 SDValue Zero = DAG.getConstant(0, N0.getValueType());
1774 return DAG.getSetCC(dl, VT, N1, Zero, Cond);
1779 // Fold away ALL boolean setcc's.
1781 if (N0.getValueType() == MVT::i1 && foldBooleans) {
1783 default: llvm_unreachable("Unknown integer setcc!");
1784 case ISD::SETEQ: // X == Y -> ~(X^Y)
1785 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
1786 N0 = DAG.getNOT(dl, Temp, MVT::i1);
1787 if (!DCI.isCalledByLegalizer())
1788 DCI.AddToWorklist(Temp.getNode());
1790 case ISD::SETNE: // X != Y --> (X^Y)
1791 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
1793 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y
1794 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y
1795 Temp = DAG.getNOT(dl, N0, MVT::i1);
1796 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp);
1797 if (!DCI.isCalledByLegalizer())
1798 DCI.AddToWorklist(Temp.getNode());
1800 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X
1801 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X
1802 Temp = DAG.getNOT(dl, N1, MVT::i1);
1803 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp);
1804 if (!DCI.isCalledByLegalizer())
1805 DCI.AddToWorklist(Temp.getNode());
1807 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y
1808 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y
1809 Temp = DAG.getNOT(dl, N0, MVT::i1);
1810 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp);
1811 if (!DCI.isCalledByLegalizer())
1812 DCI.AddToWorklist(Temp.getNode());
1814 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X
1815 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X
1816 Temp = DAG.getNOT(dl, N1, MVT::i1);
1817 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp);
1820 if (VT != MVT::i1) {
1821 if (!DCI.isCalledByLegalizer())
1822 DCI.AddToWorklist(N0.getNode());
1823 // FIXME: If running after legalize, we probably can't do this.
1824 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0);
1829 // Could not fold it.
1833 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
1834 /// node is a GlobalAddress + offset.
1835 bool TargetLowering::isGAPlusOffset(SDNode *N, const GlobalValue *&GA,
1836 int64_t &Offset) const {
1837 if (isa<GlobalAddressSDNode>(N)) {
1838 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N);
1839 GA = GASD->getGlobal();
1840 Offset += GASD->getOffset();
1844 if (N->getOpcode() == ISD::ADD) {
1845 SDValue N1 = N->getOperand(0);
1846 SDValue N2 = N->getOperand(1);
1847 if (isGAPlusOffset(N1.getNode(), GA, Offset)) {
1848 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2);
1850 Offset += V->getSExtValue();
1853 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) {
1854 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1);
1856 Offset += V->getSExtValue();
1866 SDValue TargetLowering::
1867 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
1868 // Default implementation: no optimization.
1872 //===----------------------------------------------------------------------===//
1873 // Inline Assembler Implementation Methods
1874 //===----------------------------------------------------------------------===//
1877 TargetLowering::ConstraintType
1878 TargetLowering::getConstraintType(const std::string &Constraint) const {
1879 unsigned S = Constraint.size();
1882 switch (Constraint[0]) {
1884 case 'r': return C_RegisterClass;
1886 case 'o': // offsetable
1887 case 'V': // not offsetable
1889 case 'i': // Simple Integer or Relocatable Constant
1890 case 'n': // Simple Integer
1891 case 'E': // Floating Point Constant
1892 case 'F': // Floating Point Constant
1893 case 's': // Relocatable Constant
1894 case 'p': // Address.
1895 case 'X': // Allow ANY value.
1896 case 'I': // Target registers.
1910 if (S > 1 && Constraint[0] == '{' && Constraint[S-1] == '}') {
1911 if (S == 8 && !Constraint.compare(1, 6, "memory", 6)) // "{memory}"
1918 /// LowerXConstraint - try to replace an X constraint, which matches anything,
1919 /// with another that has more specific requirements based on the type of the
1920 /// corresponding operand.
1921 const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const{
1922 if (ConstraintVT.isInteger())
1924 if (ConstraintVT.isFloatingPoint())
1925 return "f"; // works for many targets
1929 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
1930 /// vector. If it is invalid, don't add anything to Ops.
1931 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
1932 std::string &Constraint,
1933 std::vector<SDValue> &Ops,
1934 SelectionDAG &DAG) const {
1936 if (Constraint.length() > 1) return;
1938 char ConstraintLetter = Constraint[0];
1939 switch (ConstraintLetter) {
1941 case 'X': // Allows any operand; labels (basic block) use this.
1942 if (Op.getOpcode() == ISD::BasicBlock) {
1947 case 'i': // Simple Integer or Relocatable Constant
1948 case 'n': // Simple Integer
1949 case 's': { // Relocatable Constant
1950 // These operands are interested in values of the form (GV+C), where C may
1951 // be folded in as an offset of GV, or it may be explicitly added. Also, it
1952 // is possible and fine if either GV or C are missing.
1953 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1954 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op);
1956 // If we have "(add GV, C)", pull out GV/C
1957 if (Op.getOpcode() == ISD::ADD) {
1958 C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1959 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0));
1960 if (C == 0 || GA == 0) {
1961 C = dyn_cast<ConstantSDNode>(Op.getOperand(0));
1962 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1));
1964 if (C == 0 || GA == 0)
1968 // If we find a valid operand, map to the TargetXXX version so that the
1969 // value itself doesn't get selected.
1970 if (GA) { // Either &GV or &GV+C
1971 if (ConstraintLetter != 'n') {
1972 int64_t Offs = GA->getOffset();
1973 if (C) Offs += C->getZExtValue();
1974 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(),
1975 C ? SDLoc(C) : SDLoc(),
1976 Op.getValueType(), Offs));
1980 if (C) { // just C, no GV.
1981 // Simple constants are not allowed for 's'.
1982 if (ConstraintLetter != 's') {
1983 // gcc prints these as sign extended. Sign extend value to 64 bits
1984 // now; without this it would get ZExt'd later in
1985 // ScheduleDAGSDNodes::EmitNode, which is very generic.
1986 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(),
1996 std::pair<unsigned, const TargetRegisterClass*> TargetLowering::
1997 getRegForInlineAsmConstraint(const std::string &Constraint,
1999 if (Constraint[0] != '{')
2000 return std::make_pair(0u, static_cast<TargetRegisterClass*>(0));
2001 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?");
2003 // Remove the braces from around the name.
2004 StringRef RegName(Constraint.data()+1, Constraint.size()-2);
2006 std::pair<unsigned, const TargetRegisterClass*> R =
2007 std::make_pair(0u, static_cast<const TargetRegisterClass*>(0));
2009 // Figure out which register class contains this reg.
2010 const TargetRegisterInfo *RI = getTargetMachine().getRegisterInfo();
2011 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(),
2012 E = RI->regclass_end(); RCI != E; ++RCI) {
2013 const TargetRegisterClass *RC = *RCI;
2015 // If none of the value types for this register class are valid, we
2016 // can't use it. For example, 64-bit reg classes on 32-bit targets.
2020 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end();
2022 if (RegName.equals_lower(RI->getName(*I))) {
2023 std::pair<unsigned, const TargetRegisterClass*> S =
2024 std::make_pair(*I, RC);
2026 // If this register class has the requested value type, return it,
2027 // otherwise keep searching and return the first class found
2028 // if no other is found which explicitly has the requested type.
2029 if (RC->hasType(VT))
2040 //===----------------------------------------------------------------------===//
2041 // Constraint Selection.
2043 /// isMatchingInputConstraint - Return true of this is an input operand that is
2044 /// a matching constraint like "4".
2045 bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const {
2046 assert(!ConstraintCode.empty() && "No known constraint!");
2047 return isdigit(static_cast<unsigned char>(ConstraintCode[0]));
2050 /// getMatchedOperand - If this is an input matching constraint, this method
2051 /// returns the output operand it matches.
2052 unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const {
2053 assert(!ConstraintCode.empty() && "No known constraint!");
2054 return atoi(ConstraintCode.c_str());
2058 /// ParseConstraints - Split up the constraint string from the inline
2059 /// assembly value into the specific constraints and their prefixes,
2060 /// and also tie in the associated operand values.
2061 /// If this returns an empty vector, and if the constraint string itself
2062 /// isn't empty, there was an error parsing.
2063 TargetLowering::AsmOperandInfoVector TargetLowering::ParseConstraints(
2064 ImmutableCallSite CS) const {
2065 /// ConstraintOperands - Information about all of the constraints.
2066 AsmOperandInfoVector ConstraintOperands;
2067 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
2068 unsigned maCount = 0; // Largest number of multiple alternative constraints.
2070 // Do a prepass over the constraints, canonicalizing them, and building up the
2071 // ConstraintOperands list.
2072 InlineAsm::ConstraintInfoVector
2073 ConstraintInfos = IA->ParseConstraints();
2075 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
2076 unsigned ResNo = 0; // ResNo - The result number of the next output.
2078 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
2079 ConstraintOperands.push_back(AsmOperandInfo(ConstraintInfos[i]));
2080 AsmOperandInfo &OpInfo = ConstraintOperands.back();
2082 // Update multiple alternative constraint count.
2083 if (OpInfo.multipleAlternatives.size() > maCount)
2084 maCount = OpInfo.multipleAlternatives.size();
2086 OpInfo.ConstraintVT = MVT::Other;
2088 // Compute the value type for each operand.
2089 switch (OpInfo.Type) {
2090 case InlineAsm::isOutput:
2091 // Indirect outputs just consume an argument.
2092 if (OpInfo.isIndirect) {
2093 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
2097 // The return value of the call is this value. As such, there is no
2098 // corresponding argument.
2099 assert(!CS.getType()->isVoidTy() &&
2101 if (StructType *STy = dyn_cast<StructType>(CS.getType())) {
2102 OpInfo.ConstraintVT = getSimpleValueType(STy->getElementType(ResNo));
2104 assert(ResNo == 0 && "Asm only has one result!");
2105 OpInfo.ConstraintVT = getSimpleValueType(CS.getType());
2109 case InlineAsm::isInput:
2110 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++));
2112 case InlineAsm::isClobber:
2117 if (OpInfo.CallOperandVal) {
2118 llvm::Type *OpTy = OpInfo.CallOperandVal->getType();
2119 if (OpInfo.isIndirect) {
2120 llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy);
2122 report_fatal_error("Indirect operand for inline asm not a pointer!");
2123 OpTy = PtrTy->getElementType();
2126 // Look for vector wrapped in a struct. e.g. { <16 x i8> }.
2127 if (StructType *STy = dyn_cast<StructType>(OpTy))
2128 if (STy->getNumElements() == 1)
2129 OpTy = STy->getElementType(0);
2131 // If OpTy is not a single value, it may be a struct/union that we
2132 // can tile with integers.
2133 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
2134 unsigned BitSize = getDataLayout()->getTypeSizeInBits(OpTy);
2143 OpInfo.ConstraintVT =
2144 MVT::getVT(IntegerType::get(OpTy->getContext(), BitSize), true);
2147 } else if (PointerType *PT = dyn_cast<PointerType>(OpTy)) {
2148 OpInfo.ConstraintVT = MVT::getIntegerVT(
2149 8*getDataLayout()->getPointerSize(PT->getAddressSpace()));
2151 OpInfo.ConstraintVT = MVT::getVT(OpTy, true);
2156 // If we have multiple alternative constraints, select the best alternative.
2157 if (ConstraintInfos.size()) {
2159 unsigned bestMAIndex = 0;
2160 int bestWeight = -1;
2161 // weight: -1 = invalid match, and 0 = so-so match to 5 = good match.
2164 // Compute the sums of the weights for each alternative, keeping track
2165 // of the best (highest weight) one so far.
2166 for (maIndex = 0; maIndex < maCount; ++maIndex) {
2168 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2169 cIndex != eIndex; ++cIndex) {
2170 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex];
2171 if (OpInfo.Type == InlineAsm::isClobber)
2174 // If this is an output operand with a matching input operand,
2175 // look up the matching input. If their types mismatch, e.g. one
2176 // is an integer, the other is floating point, or their sizes are
2177 // different, flag it as an maCantMatch.
2178 if (OpInfo.hasMatchingInput()) {
2179 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
2180 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
2181 if ((OpInfo.ConstraintVT.isInteger() !=
2182 Input.ConstraintVT.isInteger()) ||
2183 (OpInfo.ConstraintVT.getSizeInBits() !=
2184 Input.ConstraintVT.getSizeInBits())) {
2185 weightSum = -1; // Can't match.
2190 weight = getMultipleConstraintMatchWeight(OpInfo, maIndex);
2195 weightSum += weight;
2198 if (weightSum > bestWeight) {
2199 bestWeight = weightSum;
2200 bestMAIndex = maIndex;
2204 // Now select chosen alternative in each constraint.
2205 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2206 cIndex != eIndex; ++cIndex) {
2207 AsmOperandInfo& cInfo = ConstraintOperands[cIndex];
2208 if (cInfo.Type == InlineAsm::isClobber)
2210 cInfo.selectAlternative(bestMAIndex);
2215 // Check and hook up tied operands, choose constraint code to use.
2216 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size();
2217 cIndex != eIndex; ++cIndex) {
2218 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex];
2220 // If this is an output operand with a matching input operand, look up the
2221 // matching input. If their types mismatch, e.g. one is an integer, the
2222 // other is floating point, or their sizes are different, flag it as an
2224 if (OpInfo.hasMatchingInput()) {
2225 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
2227 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
2228 std::pair<unsigned, const TargetRegisterClass*> MatchRC =
2229 getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
2230 OpInfo.ConstraintVT);
2231 std::pair<unsigned, const TargetRegisterClass*> InputRC =
2232 getRegForInlineAsmConstraint(Input.ConstraintCode,
2233 Input.ConstraintVT);
2234 if ((OpInfo.ConstraintVT.isInteger() !=
2235 Input.ConstraintVT.isInteger()) ||
2236 (MatchRC.second != InputRC.second)) {
2237 report_fatal_error("Unsupported asm: input constraint"
2238 " with a matching output constraint of"
2239 " incompatible type!");
2246 return ConstraintOperands;
2250 /// getConstraintGenerality - Return an integer indicating how general CT
2252 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
2254 case TargetLowering::C_Other:
2255 case TargetLowering::C_Unknown:
2257 case TargetLowering::C_Register:
2259 case TargetLowering::C_RegisterClass:
2261 case TargetLowering::C_Memory:
2264 llvm_unreachable("Invalid constraint type");
2267 /// Examine constraint type and operand type and determine a weight value.
2268 /// This object must already have been set up with the operand type
2269 /// and the current alternative constraint selected.
2270 TargetLowering::ConstraintWeight
2271 TargetLowering::getMultipleConstraintMatchWeight(
2272 AsmOperandInfo &info, int maIndex) const {
2273 InlineAsm::ConstraintCodeVector *rCodes;
2274 if (maIndex >= (int)info.multipleAlternatives.size())
2275 rCodes = &info.Codes;
2277 rCodes = &info.multipleAlternatives[maIndex].Codes;
2278 ConstraintWeight BestWeight = CW_Invalid;
2280 // Loop over the options, keeping track of the most general one.
2281 for (unsigned i = 0, e = rCodes->size(); i != e; ++i) {
2282 ConstraintWeight weight =
2283 getSingleConstraintMatchWeight(info, (*rCodes)[i].c_str());
2284 if (weight > BestWeight)
2285 BestWeight = weight;
2291 /// Examine constraint type and operand type and determine a weight value.
2292 /// This object must already have been set up with the operand type
2293 /// and the current alternative constraint selected.
2294 TargetLowering::ConstraintWeight
2295 TargetLowering::getSingleConstraintMatchWeight(
2296 AsmOperandInfo &info, const char *constraint) const {
2297 ConstraintWeight weight = CW_Invalid;
2298 Value *CallOperandVal = info.CallOperandVal;
2299 // If we don't have a value, we can't do a match,
2300 // but allow it at the lowest weight.
2301 if (CallOperandVal == NULL)
2303 // Look at the constraint type.
2304 switch (*constraint) {
2305 case 'i': // immediate integer.
2306 case 'n': // immediate integer with a known value.
2307 if (isa<ConstantInt>(CallOperandVal))
2308 weight = CW_Constant;
2310 case 's': // non-explicit intregal immediate.
2311 if (isa<GlobalValue>(CallOperandVal))
2312 weight = CW_Constant;
2314 case 'E': // immediate float if host format.
2315 case 'F': // immediate float.
2316 if (isa<ConstantFP>(CallOperandVal))
2317 weight = CW_Constant;
2319 case '<': // memory operand with autodecrement.
2320 case '>': // memory operand with autoincrement.
2321 case 'm': // memory operand.
2322 case 'o': // offsettable memory operand
2323 case 'V': // non-offsettable memory operand
2326 case 'r': // general register.
2327 case 'g': // general register, memory operand or immediate integer.
2328 // note: Clang converts "g" to "imr".
2329 if (CallOperandVal->getType()->isIntegerTy())
2330 weight = CW_Register;
2332 case 'X': // any operand.
2334 weight = CW_Default;
2340 /// ChooseConstraint - If there are multiple different constraints that we
2341 /// could pick for this operand (e.g. "imr") try to pick the 'best' one.
2342 /// This is somewhat tricky: constraints fall into four classes:
2343 /// Other -> immediates and magic values
2344 /// Register -> one specific register
2345 /// RegisterClass -> a group of regs
2346 /// Memory -> memory
2347 /// Ideally, we would pick the most specific constraint possible: if we have
2348 /// something that fits into a register, we would pick it. The problem here
2349 /// is that if we have something that could either be in a register or in
2350 /// memory that use of the register could cause selection of *other*
2351 /// operands to fail: they might only succeed if we pick memory. Because of
2352 /// this the heuristic we use is:
2354 /// 1) If there is an 'other' constraint, and if the operand is valid for
2355 /// that constraint, use it. This makes us take advantage of 'i'
2356 /// constraints when available.
2357 /// 2) Otherwise, pick the most general constraint present. This prefers
2358 /// 'm' over 'r', for example.
2360 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo,
2361 const TargetLowering &TLI,
2362 SDValue Op, SelectionDAG *DAG) {
2363 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options");
2364 unsigned BestIdx = 0;
2365 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown;
2366 int BestGenerality = -1;
2368 // Loop over the options, keeping track of the most general one.
2369 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) {
2370 TargetLowering::ConstraintType CType =
2371 TLI.getConstraintType(OpInfo.Codes[i]);
2373 // If this is an 'other' constraint, see if the operand is valid for it.
2374 // For example, on X86 we might have an 'rI' constraint. If the operand
2375 // is an integer in the range [0..31] we want to use I (saving a load
2376 // of a register), otherwise we must use 'r'.
2377 if (CType == TargetLowering::C_Other && Op.getNode()) {
2378 assert(OpInfo.Codes[i].size() == 1 &&
2379 "Unhandled multi-letter 'other' constraint");
2380 std::vector<SDValue> ResultOps;
2381 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i],
2383 if (!ResultOps.empty()) {
2390 // Things with matching constraints can only be registers, per gcc
2391 // documentation. This mainly affects "g" constraints.
2392 if (CType == TargetLowering::C_Memory && OpInfo.hasMatchingInput())
2395 // This constraint letter is more general than the previous one, use it.
2396 int Generality = getConstraintGenerality(CType);
2397 if (Generality > BestGenerality) {
2400 BestGenerality = Generality;
2404 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx];
2405 OpInfo.ConstraintType = BestType;
2408 /// ComputeConstraintToUse - Determines the constraint code and constraint
2409 /// type to use for the specific AsmOperandInfo, setting
2410 /// OpInfo.ConstraintCode and OpInfo.ConstraintType.
2411 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
2413 SelectionDAG *DAG) const {
2414 assert(!OpInfo.Codes.empty() && "Must have at least one constraint");
2416 // Single-letter constraints ('r') are very common.
2417 if (OpInfo.Codes.size() == 1) {
2418 OpInfo.ConstraintCode = OpInfo.Codes[0];
2419 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2421 ChooseConstraint(OpInfo, *this, Op, DAG);
2424 // 'X' matches anything.
2425 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) {
2426 // Labels and constants are handled elsewhere ('X' is the only thing
2427 // that matches labels). For Functions, the type here is the type of
2428 // the result, which is not what we want to look at; leave them alone.
2429 Value *v = OpInfo.CallOperandVal;
2430 if (isa<BasicBlock>(v) || isa<ConstantInt>(v) || isa<Function>(v)) {
2431 OpInfo.CallOperandVal = v;
2435 // Otherwise, try to resolve it to something we know about by looking at
2436 // the actual operand type.
2437 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) {
2438 OpInfo.ConstraintCode = Repl;
2439 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode);
2444 /// \brief Given an exact SDIV by a constant, create a multiplication
2445 /// with the multiplicative inverse of the constant.
2446 SDValue TargetLowering::BuildExactSDIV(SDValue Op1, SDValue Op2, SDLoc dl,
2447 SelectionDAG &DAG) const {
2448 ConstantSDNode *C = cast<ConstantSDNode>(Op2);
2449 APInt d = C->getAPIntValue();
2450 assert(d != 0 && "Division by zero!");
2452 // Shift the value upfront if it is even, so the LSB is one.
2453 unsigned ShAmt = d.countTrailingZeros();
2455 // TODO: For UDIV use SRL instead of SRA.
2456 SDValue Amt = DAG.getConstant(ShAmt, getShiftAmountTy(Op1.getValueType()));
2457 Op1 = DAG.getNode(ISD::SRA, dl, Op1.getValueType(), Op1, Amt);
2461 // Calculate the multiplicative inverse, using Newton's method.
2463 while ((t = d*xn) != 1)
2464 xn *= APInt(d.getBitWidth(), 2) - t;
2466 Op2 = DAG.getConstant(xn, Op1.getValueType());
2467 return DAG.getNode(ISD::MUL, dl, Op1.getValueType(), Op1, Op2);
2470 /// \brief Given an ISD::SDIV node expressing a divide by constant,
2471 /// return a DAG expression to select that will generate the same value by
2472 /// multiplying by a magic number. See:
2473 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2474 SDValue TargetLowering::
2475 BuildSDIV(SDNode *N, SelectionDAG &DAG, bool IsAfterLegalization,
2476 std::vector<SDNode*> *Created) const {
2477 EVT VT = N->getValueType(0);
2480 // Check to see if we can do this.
2481 // FIXME: We should be more aggressive here.
2482 if (!isTypeLegal(VT))
2485 APInt d = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue();
2486 APInt::ms magics = d.magic();
2488 // Multiply the numerator (operand 0) by the magic value
2489 // FIXME: We should support doing a MUL in a wider type
2491 if (IsAfterLegalization ? isOperationLegal(ISD::MULHS, VT) :
2492 isOperationLegalOrCustom(ISD::MULHS, VT))
2493 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0),
2494 DAG.getConstant(magics.m, VT));
2495 else if (IsAfterLegalization ? isOperationLegal(ISD::SMUL_LOHI, VT) :
2496 isOperationLegalOrCustom(ISD::SMUL_LOHI, VT))
2497 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT),
2499 DAG.getConstant(magics.m, VT)).getNode(), 1);
2501 return SDValue(); // No mulhs or equvialent
2502 // If d > 0 and m < 0, add the numerator
2503 if (d.isStrictlyPositive() && magics.m.isNegative()) {
2504 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0));
2506 Created->push_back(Q.getNode());
2508 // If d < 0 and m > 0, subtract the numerator.
2509 if (d.isNegative() && magics.m.isStrictlyPositive()) {
2510 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0));
2512 Created->push_back(Q.getNode());
2514 // Shift right algebraic if shift value is nonzero
2516 Q = DAG.getNode(ISD::SRA, dl, VT, Q,
2517 DAG.getConstant(magics.s, getShiftAmountTy(Q.getValueType())));
2519 Created->push_back(Q.getNode());
2521 // Extract the sign bit and add it to the quotient
2523 DAG.getNode(ISD::SRL, dl, VT, Q, DAG.getConstant(VT.getSizeInBits()-1,
2524 getShiftAmountTy(Q.getValueType())));
2526 Created->push_back(T.getNode());
2527 return DAG.getNode(ISD::ADD, dl, VT, Q, T);
2530 /// \brief Given an ISD::UDIV node expressing a divide by constant,
2531 /// return a DAG expression to select that will generate the same value by
2532 /// multiplying by a magic number. See:
2533 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
2534 SDValue TargetLowering::
2535 BuildUDIV(SDNode *N, SelectionDAG &DAG, bool IsAfterLegalization,
2536 std::vector<SDNode*> *Created) const {
2537 EVT VT = N->getValueType(0);
2540 // Check to see if we can do this.
2541 // FIXME: We should be more aggressive here.
2542 if (!isTypeLegal(VT))
2545 // FIXME: We should use a narrower constant when the upper
2546 // bits are known to be zero.
2547 const APInt &N1C = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue();
2548 APInt::mu magics = N1C.magicu();
2550 SDValue Q = N->getOperand(0);
2552 // If the divisor is even, we can avoid using the expensive fixup by shifting
2553 // the divided value upfront.
2554 if (magics.a != 0 && !N1C[0]) {
2555 unsigned Shift = N1C.countTrailingZeros();
2556 Q = DAG.getNode(ISD::SRL, dl, VT, Q,
2557 DAG.getConstant(Shift, getShiftAmountTy(Q.getValueType())));
2559 Created->push_back(Q.getNode());
2561 // Get magic number for the shifted divisor.
2562 magics = N1C.lshr(Shift).magicu(Shift);
2563 assert(magics.a == 0 && "Should use cheap fixup now");
2566 // Multiply the numerator (operand 0) by the magic value
2567 // FIXME: We should support doing a MUL in a wider type
2568 if (IsAfterLegalization ? isOperationLegal(ISD::MULHU, VT) :
2569 isOperationLegalOrCustom(ISD::MULHU, VT))
2570 Q = DAG.getNode(ISD::MULHU, dl, VT, Q, DAG.getConstant(magics.m, VT));
2571 else if (IsAfterLegalization ? isOperationLegal(ISD::UMUL_LOHI, VT) :
2572 isOperationLegalOrCustom(ISD::UMUL_LOHI, VT))
2573 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT), Q,
2574 DAG.getConstant(magics.m, VT)).getNode(), 1);
2576 return SDValue(); // No mulhu or equvialent
2578 Created->push_back(Q.getNode());
2580 if (magics.a == 0) {
2581 assert(magics.s < N1C.getBitWidth() &&
2582 "We shouldn't generate an undefined shift!");
2583 return DAG.getNode(ISD::SRL, dl, VT, Q,
2584 DAG.getConstant(magics.s, getShiftAmountTy(Q.getValueType())));
2586 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q);
2588 Created->push_back(NPQ.getNode());
2589 NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ,
2590 DAG.getConstant(1, getShiftAmountTy(NPQ.getValueType())));
2592 Created->push_back(NPQ.getNode());
2593 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q);
2595 Created->push_back(NPQ.getNode());
2596 return DAG.getNode(ISD::SRL, dl, VT, NPQ,
2597 DAG.getConstant(magics.s-1, getShiftAmountTy(NPQ.getValueType())));