void IndVarSimplify::EliminateIVRemainders() {
// Look for SRem and URem users.
for (IVUsers::iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
IVStrideUse &UI = *I;
BinaryOperator *Rem = dyn_cast<BinaryOperator>(UI.getUser());
if (!Rem) continue;
bool isSigned = Rem->getOpcode() == Instruction::SRem;
if (!isSigned && Rem->getOpcode() != Instruction::URem)
continue;
// We're only interested in the case where we know something about
// the numerator.
if (UI.getOperandValToReplace() != Rem->getOperand(0))
continue;
// Get the SCEVs for the ICmp operands.
const SCEV *S = SE->getSCEV(Rem->getOperand(0));
const SCEV *X = SE->getSCEV(Rem->getOperand(1));
// Simplify unnecessary loops away.
const Loop *ICmpLoop = LI->getLoopFor(Rem->getParent());
S = SE->getSCEVAtScope(S, ICmpLoop);
X = SE->getSCEVAtScope(X, ICmpLoop);
// i % n --> i if i is in [0,n).
if ((!isSigned || SE->isKnownNonNegative(S)) &&
SE->isKnownPredicate(isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
S, X))
Rem->replaceAllUsesWith(Rem->getOperand(0));
else {
// (i+1) % n --> (i+1)==n?0:(i+1) if i is in [0,n).
const SCEV *LessOne =
SE->getMinusSCEV(S, SE->getConstant(S->getType(), 1));
if ((!isSigned || SE->isKnownNonNegative(LessOne)) &&
SE->isKnownPredicate(isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
LessOne, X)) {
ICmpInst *ICmp = new ICmpInst(Rem, ICmpInst::ICMP_EQ,
Rem->getOperand(0), Rem->getOperand(1),
"tmp");
SelectInst *Sel =
SelectInst::Create(ICmp,
ConstantInt::get(Rem->getType(), 0),
Rem->getOperand(0), "tmp", Rem);
Rem->replaceAllUsesWith(Sel);
} else
continue;
}
// Inform IVUsers about the new users.
if (Instruction *I = dyn_cast<Instruction>(Rem->getOperand(0)))
IU->AddUsersIfInteresting(I);
DEBUG(dbgs() << "INDVARS: Simplified rem: " << *Rem << '\n');
DeadInsts.push_back(Rem);
}
}
// Insert an intrinsic for fast fdiv for safe math situations where we can
// reduce precision. Leave fdiv for situations where the generic node is
// expected to be optimized.
bool AMDGPUCodeGenPrepare::visitFDiv(BinaryOperator &FDiv) {
Type *Ty = FDiv.getType();
// TODO: Handle half
if (!Ty->getScalarType()->isFloatTy())
return false;
MDNode *FPMath = FDiv.getMetadata(LLVMContext::MD_fpmath);
if (!FPMath)
return false;
const FPMathOperator *FPOp = cast<const FPMathOperator>(&FDiv);
float ULP = FPOp->getFPAccuracy();
if (ULP < 2.5f)
return false;
FastMathFlags FMF = FPOp->getFastMathFlags();
bool UnsafeDiv = HasUnsafeFPMath || FMF.unsafeAlgebra() ||
FMF.allowReciprocal();
if (ST->hasFP32Denormals() && !UnsafeDiv)
return false;
IRBuilder<> Builder(FDiv.getParent(), std::next(FDiv.getIterator()), FPMath);
Builder.setFastMathFlags(FMF);
Builder.SetCurrentDebugLocation(FDiv.getDebugLoc());
const AMDGPUIntrinsicInfo *II = TM->getIntrinsicInfo();
Function *Decl
= II->getDeclaration(Mod, AMDGPUIntrinsic::amdgcn_fdiv_fast, {});
Value *Num = FDiv.getOperand(0);
Value *Den = FDiv.getOperand(1);
Value *NewFDiv = nullptr;
if (VectorType *VT = dyn_cast<VectorType>(Ty)) {
NewFDiv = UndefValue::get(VT);
// FIXME: Doesn't do the right thing for cases where the vector is partially
// constant. This works when the scalarizer pass is run first.
for (unsigned I = 0, E = VT->getNumElements(); I != E; ++I) {
Value *NumEltI = Builder.CreateExtractElement(Num, I);
Value *DenEltI = Builder.CreateExtractElement(Den, I);
Value *NewElt;
if (shouldKeepFDivF32(NumEltI, UnsafeDiv)) {
NewElt = Builder.CreateFDiv(NumEltI, DenEltI);
} else {
NewElt = Builder.CreateCall(Decl, { NumEltI, DenEltI });
}
NewFDiv = Builder.CreateInsertElement(NewFDiv, NewElt, I);
}
} else {
if (!shouldKeepFDivF32(Num, UnsafeDiv))
NewFDiv = Builder.CreateCall(Decl, { Num, Den });
}
if (NewFDiv) {
FDiv.replaceAllUsesWith(NewFDiv);
NewFDiv->takeName(&FDiv);
FDiv.eraseFromParent();
}
return true;
}
// Peephole optimize the following instructions:
// %t1 = cast ? to x *
// %t2 = add x * %SP, %t1 ;; Constant must be 2nd operand
//
// Into: %t3 = getelementptr {<...>} * %SP, <element indices>
// %t2 = cast <eltype> * %t3 to {<...>}*
//
static bool HandleCastToPointer(BasicBlock::iterator BI,
const PointerType *DestPTy,
const TargetData &TD) {
CastInst &CI = cast<CastInst>(*BI);
if (CI.use_empty()) return false;
// Scan all of the uses, looking for any uses that are not add or sub
// instructions. If we have non-adds, do not make this transformation.
//
bool HasSubUse = false; // Keep track of any subtracts...
for (Value::use_iterator I = CI.use_begin(), E = CI.use_end();
I != E; ++I)
if (BinaryOperator *BO = dyn_cast<BinaryOperator>(*I)) {
if ((BO->getOpcode() != Instruction::Add &&
BO->getOpcode() != Instruction::Sub) ||
// Avoid add sbyte* %X, %X cases...
BO->getOperand(0) == BO->getOperand(1))
return false;
else
HasSubUse |= BO->getOpcode() == Instruction::Sub;
} else {
return false;
}
std::vector<Value*> Indices;
Value *Src = CI.getOperand(0);
const Type *Result = ConvertibleToGEP(DestPTy, Src, Indices, TD, &BI);
if (Result == 0) return false; // Not convertible...
// Cannot handle subtracts if there is more than one index required...
if (HasSubUse && Indices.size() != 1) return false;
PRINT_PEEPHOLE2("cast-add-to-gep:in", *Src, CI);
// If we have a getelementptr capability... transform all of the
// add instruction uses into getelementptr's.
while (!CI.use_empty()) {
BinaryOperator *I = cast<BinaryOperator>(*CI.use_begin());
assert((I->getOpcode() == Instruction::Add ||
I->getOpcode() == Instruction::Sub) &&
"Use is not a valid add instruction!");
// Get the value added to the cast result pointer...
Value *OtherPtr = I->getOperand((I->getOperand(0) == &CI) ? 1 : 0);
Instruction *GEP = new GetElementPtrInst(OtherPtr, Indices, I->getName());
PRINT_PEEPHOLE1("cast-add-to-gep:i", *I);
// If the instruction is actually a subtract, we are guaranteed to only have
// one index (from code above), so we just need to negate the pointer index
// long value.
if (I->getOpcode() == Instruction::Sub) {
Instruction *Neg = BinaryOperator::createNeg(GEP->getOperand(1),
GEP->getOperand(1)->getName()+".neg", I);
GEP->setOperand(1, Neg);
}
if (GEP->getType() == I->getType()) {
// Replace the old add instruction with the shiny new GEP inst
ReplaceInstWithInst(I, GEP);
} else {
// If the type produced by the gep instruction differs from the original
// add instruction type, insert a cast now.
//
// Insert the GEP instruction before the old add instruction...
I->getParent()->getInstList().insert(I, GEP);
PRINT_PEEPHOLE1("cast-add-to-gep:o", *GEP);
GEP = new CastInst(GEP, I->getType());
// Replace the old add instruction with the shiny new GEP inst
ReplaceInstWithInst(I, GEP);
}
PRINT_PEEPHOLE1("cast-add-to-gep:o", *GEP);
}
return true;
}
/// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
/// instruction.
bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
SelectInst *SI = cast<SelectInst>(I.getOperand(1));
// div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
int NonNullOperand = -1;
if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
if (ST->isNullValue())
NonNullOperand = 2;
// div/rem X, (Cond ? Y : 0) -> div/rem X, Y
if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
if (ST->isNullValue())
NonNullOperand = 1;
if (NonNullOperand == -1)
return false;
Value *SelectCond = SI->getOperand(0);
// Change the div/rem to use 'Y' instead of the select.
I.setOperand(1, SI->getOperand(NonNullOperand));
// Okay, we know we replace the operand of the div/rem with 'Y' with no
// problem. However, the select, or the condition of the select may have
// multiple uses. Based on our knowledge that the operand must be non-zero,
// propagate the known value for the select into other uses of it, and
// propagate a known value of the condition into its other users.
// If the select and condition only have a single use, don't bother with this,
// early exit.
if (SI->use_empty() && SelectCond->hasOneUse())
return true;
// Scan the current block backward, looking for other uses of SI.
BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
while (BBI != BBFront) {
--BBI;
// If we found a call to a function, we can't assume it will return, so
// information from below it cannot be propagated above it.
if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
break;
// Replace uses of the select or its condition with the known values.
for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
I != E; ++I) {
if (*I == SI) {
*I = SI->getOperand(NonNullOperand);
Worklist.Add(BBI);
} else if (*I == SelectCond) {
*I = NonNullOperand == 1 ? ConstantInt::getTrue(BBI->getContext()) :
ConstantInt::getFalse(BBI->getContext());
Worklist.Add(BBI);
}
}
// If we past the instruction, quit looking for it.
if (&*BBI == SI)
SI = 0;
if (&*BBI == SelectCond)
SelectCond = 0;
// If we ran out of things to eliminate, break out of the loop.
if (SelectCond == 0 && SI == 0)
break;
}
return true;
}
// NegateValue - Insert instructions before the instruction pointed to by BI,
// that computes the negative version of the value specified. The negative
// version of the value is returned, and BI is left pointing at the instruction
// that should be processed next by the reassociation pass.
//
static Value *NegateValue(Value *V, Instruction *BI) {
if (Constant *C = dyn_cast<Constant>(V))
return ConstantExpr::getNeg(C);
// We are trying to expose opportunity for reassociation. One of the things
// that we want to do to achieve this is to push a negation as deep into an
// expression chain as possible, to expose the add instructions. In practice,
// this means that we turn this:
// X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
// so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
// the constants. We assume that instcombine will clean up the mess later if
// we introduce tons of unnecessary negation instructions.
//
if (Instruction *I = dyn_cast<Instruction>(V))
if (I->getOpcode() == Instruction::Add && I->hasOneUse()) {
// Push the negates through the add.
I->setOperand(0, NegateValue(I->getOperand(0), BI));
I->setOperand(1, NegateValue(I->getOperand(1), BI));
// We must move the add instruction here, because the neg instructions do
// not dominate the old add instruction in general. By moving it, we are
// assured that the neg instructions we just inserted dominate the
// instruction we are about to insert after them.
//
I->moveBefore(BI);
I->setName(I->getName()+".neg");
return I;
}
// Okay, we need to materialize a negated version of V with an instruction.
// Scan the use lists of V to see if we have one already.
for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
User *U = *UI;
if (!BinaryOperator::isNeg(U)) continue;
// We found one! Now we have to make sure that the definition dominates
// this use. We do this by moving it to the entry block (if it is a
// non-instruction value) or right after the definition. These negates will
// be zapped by reassociate later, so we don't need much finesse here.
BinaryOperator *TheNeg = cast<BinaryOperator>(U);
// Verify that the negate is in this function, V might be a constant expr.
if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
continue;
BasicBlock::iterator InsertPt;
if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
InsertPt = II->getNormalDest()->begin();
} else {
InsertPt = InstInput;
++InsertPt;
}
while (isa<PHINode>(InsertPt)) ++InsertPt;
} else {
InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
}
TheNeg->moveBefore(InsertPt);
return TheNeg;
}
// Insert a 'neg' instruction that subtracts the value from zero to get the
// negation.
return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI);
}