RPCS3/llvm-mirror
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-10-19 02:52:53 +02:00
Code Issues Projects Releases Wiki Activity

[IR] Add WithOverflowInst class

Browse Source 
This adds a WithOverflowInst class with a few helper methods to get
the underlying binop, signedness and nowrap type and makes use of it
where sensible. There will be two more uses in D60650/D60656.

The refactorings are all NFC, though I left some TODOs where things
could be improved. In particular we have two places where add/sub are
handled but mul isn't.

Differential Revision: https://reviews.llvm.org/D60668

llvm-svn: 358512
This commit is contained in:
Nikita Popov 2019-04-16 18:55:16 +00:00
parent 0687a4b90a
commit 32304a1b52
9 changed files with 153 additions and 220 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
include/llvm/Analysis/ValueTracking.h
View File

@ -32,6 +32,7 @@ class DataLayout;
class DominatorTree;
class GEPOperator;
class IntrinsicInst;
class WithOverflowInst;
struct KnownBits;
class Loop;
class LoopInfo;
@ -454,10 +455,10 @@ class Value;
const Instruction *CxtI,
const DominatorTree *DT);
/// Returns true if the arithmetic part of the \p II 's result is
/// Returns true if the arithmetic part of the \p WO 's result is
/// used only along the paths control dependent on the computation
/// not overflowing, \p II being an <op>.with.overflow intrinsic.
bool isOverflowIntrinsicNoWrap(const IntrinsicInst *II,
/// not overflowing, \p WO being an <op>.with.overflow intrinsic.
bool isOverflowIntrinsicNoWrap(const WithOverflowInst *WO,
const DominatorTree &DT);

33
include/llvm/IR/IntrinsicInst.h
View File

@ -265,6 +265,39 @@ namespace llvm {
}
};
/// This class represents a op.with.overflow intrinsic.
class WithOverflowInst : public IntrinsicInst {
public:
static bool classof(const IntrinsicInst *I) {
switch (I->getIntrinsicID()) {
case Intrinsic::uadd_with_overflow:
case Intrinsic::sadd_with_overflow:
case Intrinsic::usub_with_overflow:
case Intrinsic::ssub_with_overflow:
case Intrinsic::umul_with_overflow:
case Intrinsic::smul_with_overflow:
return true;
default:
return false;
}
}
static bool classof(const Value *V) {
return isa<IntrinsicInst>(V) && classof(cast<IntrinsicInst>(V));
}
Value *getLHS() const { return const_cast<Value*>(getArgOperand(0)); }
Value *getRHS() const { return const_cast<Value*>(getArgOperand(1)); }
/// Returns the binary operation underlying the intrinsic.
Instruction::BinaryOps getBinaryOp() const;
/// Whether the intrinsic is signed or unsigned.
bool isSigned() const;
/// Returns one of OBO::NoSignedWrap or OBO::NoUnsignedWrap.
unsigned getNoWrapKind() const;
};
/// Common base class for all memory intrinsics. Simply provides
/// common methods.
/// Written as CRTP to avoid a common base class amongst the

55
lib/Analysis/ScalarEvolution.cpp
View File

@ -4575,52 +4575,21 @@ static Optional<BinaryOp> MatchBinaryOp(Value *V, DominatorTree &DT) {
if (EVI->getNumIndices() != 1 || EVI->getIndices()[0] != 0)
break;
auto *CI = dyn_cast<CallInst>(EVI->getAggregateOperand());
if (!CI)
auto *WO = dyn_cast<WithOverflowInst>(EVI->getAggregateOperand());
if (!WO)
break;
if (auto *F = CI->getCalledFunction())
switch (F->getIntrinsicID()) {
case Intrinsic::sadd_with_overflow:
case Intrinsic::uadd_with_overflow:
if (!isOverflowIntrinsicNoWrap(cast<IntrinsicInst>(CI), DT))
return BinaryOp(Instruction::Add, CI->getArgOperand(0),
CI->getArgOperand(1));
Instruction::BinaryOps BinOp = WO->getBinaryOp();
bool Signed = WO->isSigned();
// TODO: Should add nuw/nsw flags for mul as well.
if (BinOp == Instruction::Mul || !isOverflowIntrinsicNoWrap(WO, DT))
return BinaryOp(BinOp, WO->getLHS(), WO->getRHS());
// Now that we know that all uses of the arithmetic-result component of
// CI are guarded by the overflow check, we can go ahead and pretend
// that the arithmetic is non-overflowing.
if (F->getIntrinsicID() == Intrinsic::sadd_with_overflow)
return BinaryOp(Instruction::Add, CI->getArgOperand(0),
CI->getArgOperand(1), /* IsNSW = */ true,
/* IsNUW = */ false);
else
return BinaryOp(Instruction::Add, CI->getArgOperand(0),
CI->getArgOperand(1), /* IsNSW = */ false,
/* IsNUW*/ true);
case Intrinsic::ssub_with_overflow:
case Intrinsic::usub_with_overflow:
if (!isOverflowIntrinsicNoWrap(cast<IntrinsicInst>(CI), DT))
return BinaryOp(Instruction::Sub, CI->getArgOperand(0),
CI->getArgOperand(1));
// The same reasoning as sadd/uadd above.
if (F->getIntrinsicID() == Intrinsic::ssub_with_overflow)
return BinaryOp(Instruction::Sub, CI->getArgOperand(0),
CI->getArgOperand(1), /* IsNSW = */ true,
/* IsNUW = */ false);
else
return BinaryOp(Instruction::Sub, CI->getArgOperand(0),
CI->getArgOperand(1), /* IsNSW = */ false,
/* IsNUW = */ true);
case Intrinsic::smul_with_overflow:
case Intrinsic::umul_with_overflow:
return BinaryOp(Instruction::Mul, CI->getArgOperand(0),
CI->getArgOperand(1));
default:
break;
}
break;
// Now that we know that all uses of the arithmetic-result component of
// CI are guarded by the overflow check, we can go ahead and pretend
// that the arithmetic is non-overflowing.
return BinaryOp(BinOp, WO->getLHS(), WO->getRHS(),
/* IsNSW = */ Signed, /* IsNUW = */ !Signed);
}
default:

15
lib/Analysis/ValueTracking.cpp
View File

@ -4206,23 +4206,12 @@ OverflowResult llvm::computeOverflowForSignedSub(const Value *LHS,
return mapOverflowResult(LHSRange.signedSubMayOverflow(RHSRange));
}
bool llvm::isOverflowIntrinsicNoWrap(const IntrinsicInst *II,
bool llvm::isOverflowIntrinsicNoWrap(const WithOverflowInst *WO,
const DominatorTree &DT) {
#ifndef NDEBUG
auto IID = II->getIntrinsicID();
assert((IID == Intrinsic::sadd_with_overflow ||
IID == Intrinsic::uadd_with_overflow ||
IID == Intrinsic::ssub_with_overflow ||
IID == Intrinsic::usub_with_overflow ||
IID == Intrinsic::smul_with_overflow ||
IID == Intrinsic::umul_with_overflow) &&
"Not an overflow intrinsic!");
#endif
SmallVector<const BranchInst *, 2> GuardingBranches;
SmallVector<const ExtractValueInst *, 2> Results;
for (const User *U : II->users()) {
for (const User *U : WO->users()) {
if (const auto *EVI = dyn_cast<ExtractValueInst>(U)) {
assert(EVI->getNumIndices() == 1 && "Obvious from CI's type");

34
lib/IR/IntrinsicInst.cpp
View File

@ -21,6 +21,7 @@
//===----------------------------------------------------------------------===//
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/Operator.h"
#include "llvm/ADT/StringSwitch.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DebugInfoMetadata.h"
@ -168,3 +169,36 @@ bool ConstrainedFPIntrinsic::isTernaryOp() const {
}
}
Instruction::BinaryOps WithOverflowInst::getBinaryOp() const {
switch (getIntrinsicID()) {
case Intrinsic::uadd_with_overflow:
case Intrinsic::sadd_with_overflow:
return Instruction::Add;
case Intrinsic::usub_with_overflow:
case Intrinsic::ssub_with_overflow:
return Instruction::Sub;
case Intrinsic::umul_with_overflow:
case Intrinsic::smul_with_overflow:
return Instruction::Mul;
default:
llvm_unreachable("Invalid intrinsic");
}
}
bool WithOverflowInst::isSigned() const {
switch (getIntrinsicID()) {
case Intrinsic::sadd_with_overflow:
case Intrinsic::ssub_with_overflow:
case Intrinsic::smul_with_overflow:
return true;
default:
return false;
}
}
unsigned WithOverflowInst::getNoWrapKind() const {
if (isSigned())
return OverflowingBinaryOperator::NoSignedWrap;
else
return OverflowingBinaryOperator::NoUnsignedWrap;
}

65
lib/Transforms/InstCombine/InstructionCombining.cpp
View File

@ -2667,53 +2667,28 @@ Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
return ExtractValueInst::Create(IV->getInsertedValueOperand(),
makeArrayRef(exti, exte));
}
if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
// We're extracting from an intrinsic, see if we're the only user, which
// allows us to simplify multiple result intrinsics to simpler things that
// just get one value.
if (II->hasOneUse()) {
// Check if we're grabbing the overflow bit or the result of a 'with
// overflow' intrinsic. If it's the latter we can remove the intrinsic
if (WithOverflowInst *WO = dyn_cast<WithOverflowInst>(Agg)) {
// We're extracting from an overflow intrinsic, see if we're the only user,
// which allows us to simplify multiple result intrinsics to simpler
// things that just get one value.
if (WO->hasOneUse()) {
// Check if we're grabbing only the result of a 'with overflow' intrinsic
// and replace it with a traditional binary instruction.
switch (II->getIntrinsicID()) {
case Intrinsic::uadd_with_overflow:
case Intrinsic::sadd_with_overflow:
if (*EV.idx_begin() == 0) { // Normal result.
Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
replaceInstUsesWith(*II, UndefValue::get(II->getType()));
eraseInstFromFunction(*II);
return BinaryOperator::CreateAdd(LHS, RHS);
}
// If the normal result of the add is dead, and the RHS is a constant,
// we can transform this into a range comparison.
// overflow = uadd a, -4 --> overflow = icmp ugt a, 3
if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow)
if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getArgOperand(1)))
return new ICmpInst(ICmpInst::ICMP_UGT, II->getArgOperand(0),
ConstantExpr::getNot(CI));
break;
case Intrinsic::usub_with_overflow:
case Intrinsic::ssub_with_overflow:
if (*EV.idx_begin() == 0) { // Normal result.
Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
replaceInstUsesWith(*II, UndefValue::get(II->getType()));
eraseInstFromFunction(*II);
return BinaryOperator::CreateSub(LHS, RHS);
}
break;
case Intrinsic::umul_with_overflow:
case Intrinsic::smul_with_overflow:
if (*EV.idx_begin() == 0) { // Normal result.
Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1);
replaceInstUsesWith(*II, UndefValue::get(II->getType()));
eraseInstFromFunction(*II);
return BinaryOperator::CreateMul(LHS, RHS);
}
break;
default:
break;
if (*EV.idx_begin() == 0) {
Instruction::BinaryOps BinOp = WO->getBinaryOp();
Value *LHS = WO->getLHS(), *RHS = WO->getRHS();
replaceInstUsesWith(*WO, UndefValue::get(WO->getType()));
eraseInstFromFunction(*WO);
return BinaryOperator::Create(BinOp, LHS, RHS);
}
// If the normal result of the add is dead, and the RHS is a constant,
// we can transform this into a range comparison.
// overflow = uadd a, -4 --> overflow = icmp ugt a, 3
if (WO->getIntrinsicID() == Intrinsic::uadd_with_overflow)
if (ConstantInt *CI = dyn_cast<ConstantInt>(WO->getRHS()))
return new ICmpInst(ICmpInst::ICMP_UGT, WO->getLHS(),
ConstantExpr::getNot(CI));
}
}
if (LoadInst *L = dyn_cast<LoadInst>(Agg))

85
lib/Transforms/Scalar/CorrelatedValuePropagation.cpp
View File

@ -399,59 +399,38 @@ static bool processSwitch(SwitchInst *SI, LazyValueInfo *LVI,
}
// See if we can prove that the given overflow intrinsic will not overflow.
static bool willNotOverflow(IntrinsicInst *II, LazyValueInfo *LVI) {
using OBO = OverflowingBinaryOperator;
auto NoWrap = [&] (Instruction::BinaryOps BinOp, unsigned NoWrapKind) {
Value *RHS = II->getOperand(1);
ConstantRange RRange = LVI->getConstantRange(RHS, II->getParent(), II);
ConstantRange NWRegion = ConstantRange::makeGuaranteedNoWrapRegion(
BinOp, RRange, NoWrapKind);
// As an optimization, do not compute LRange if we do not need it.
if (NWRegion.isEmptySet())
return false;
Value *LHS = II->getOperand(0);
ConstantRange LRange = LVI->getConstantRange(LHS, II->getParent(), II);
return NWRegion.contains(LRange);
};
switch (II->getIntrinsicID()) {
default:
break;
case Intrinsic::uadd_with_overflow:
return NoWrap(Instruction::Add, OBO::NoUnsignedWrap);
case Intrinsic::sadd_with_overflow:
return NoWrap(Instruction::Add, OBO::NoSignedWrap);
case Intrinsic::usub_with_overflow:
return NoWrap(Instruction::Sub, OBO::NoUnsignedWrap);
case Intrinsic::ssub_with_overflow:
return NoWrap(Instruction::Sub, OBO::NoSignedWrap);
}
return false;
static bool willNotOverflow(WithOverflowInst *WO, LazyValueInfo *LVI) {
// TODO: Also support multiplication.
Instruction::BinaryOps BinOp = WO->getBinaryOp();
if (BinOp == Instruction::Mul)
return false;
Value *RHS = WO->getRHS();
ConstantRange RRange = LVI->getConstantRange(RHS, WO->getParent(), WO);
ConstantRange NWRegion = ConstantRange::makeGuaranteedNoWrapRegion(
BinOp, RRange, WO->getNoWrapKind());
// As an optimization, do not compute LRange if we do not need it.
if (NWRegion.isEmptySet())
return false;
Value *LHS = WO->getLHS();
ConstantRange LRange = LVI->getConstantRange(LHS, WO->getParent(), WO);
return NWRegion.contains(LRange);
}
static void processOverflowIntrinsic(IntrinsicInst *II) {
IRBuilder<> B(II);
Value *NewOp = nullptr;
switch (II->getIntrinsicID()) {
default:
llvm_unreachable("Unexpected instruction.");
case Intrinsic::uadd_with_overflow:
NewOp = B.CreateNUWAdd(II->getOperand(0), II->getOperand(1), II->getName());
break;
case Intrinsic::sadd_with_overflow:
NewOp = B.CreateNSWAdd(II->getOperand(0), II->getOperand(1), II->getName());
break;
case Intrinsic::usub_with_overflow:
NewOp = B.CreateNUWSub(II->getOperand(0), II->getOperand(1), II->getName());
break;
case Intrinsic::ssub_with_overflow:
NewOp = B.CreateNSWSub(II->getOperand(0), II->getOperand(1), II->getName());
break;
}
static void processOverflowIntrinsic(WithOverflowInst *WO) {
IRBuilder<> B(WO);
Value *NewOp = B.CreateBinOp(
WO->getBinaryOp(), WO->getLHS(), WO->getRHS(), WO->getName());
if (WO->isSigned())
cast<Instruction>(NewOp)->setHasNoSignedWrap();
else
cast<Instruction>(NewOp)->setHasNoUnsignedWrap();
Value *NewI = B.CreateInsertValue(UndefValue::get(WO->getType()), NewOp, 0);
NewI = B.CreateInsertValue(NewI, ConstantInt::getFalse(WO->getContext()), 1);
WO->replaceAllUsesWith(NewI);
WO->eraseFromParent();
++NumOverflows;
Value *NewI = B.CreateInsertValue(UndefValue::get(II->getType()), NewOp, 0);
NewI = B.CreateInsertValue(NewI, ConstantInt::getFalse(II->getContext()), 1);
II->replaceAllUsesWith(NewI);
II->eraseFromParent();
}
/// Infer nonnull attributes for the arguments at the specified callsite.
@ -459,9 +438,9 @@ static bool processCallSite(CallSite CS, LazyValueInfo *LVI) {
SmallVector<unsigned, 4> ArgNos;
unsigned ArgNo = 0;
if (auto *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) {
if (willNotOverflow(II, LVI)) {
processOverflowIntrinsic(II);
if (auto *WO = dyn_cast<WithOverflowInst>(CS.getInstruction())) {
if (willNotOverflow(WO, LVI)) {
processOverflowIntrinsic(WO);
return true;
}
}

39
lib/Transforms/Scalar/GVN.cpp
View File

@ -329,36 +329,15 @@ GVN::Expression GVN::ValueTable::createExtractvalueExpr(ExtractValueInst *EI) {
e.type = EI->getType();
e.opcode = 0;
IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
if (I != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
// EI might be an extract from one of our recognised intrinsics. If it
// is we'll synthesize a semantically equivalent expression instead on
// an extract value expression.
switch (I->getIntrinsicID()) {
case Intrinsic::sadd_with_overflow:
case Intrinsic::uadd_with_overflow:
e.opcode = Instruction::Add;
break;
case Intrinsic::ssub_with_overflow:
case Intrinsic::usub_with_overflow:
e.opcode = Instruction::Sub;
break;
case Intrinsic::smul_with_overflow:
case Intrinsic::umul_with_overflow:
e.opcode = Instruction::Mul;
break;
default:
break;
}
if (e.opcode != 0) {
// Intrinsic recognized. Grab its args to finish building the expression.
assert(I->getNumArgOperands() == 2 &&
"Expect two args for recognised intrinsics.");
e.varargs.push_back(lookupOrAdd(I->getArgOperand(0)));
e.varargs.push_back(lookupOrAdd(I->getArgOperand(1)));
return e;
}
WithOverflowInst *WO = dyn_cast<WithOverflowInst>(EI->getAggregateOperand());
if (WO != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0) {
// EI is an extract from one of our with.overflow intrinsics. Synthesize
// a semantically equivalent expression instead of an extract value
// expression.
e.opcode = WO->getBinaryOp();
e.varargs.push_back(lookupOrAdd(WO->getLHS()));
e.varargs.push_back(lookupOrAdd(WO->getRHS()));
return e;
}
// Not a recognised intrinsic. Fall back to producing an extract value

40
lib/Transforms/Scalar/NewGVN.cpp
View File

@ -1814,39 +1814,13 @@ NewGVN::performSymbolicPHIEvaluation(ArrayRef<ValPair> PHIOps,
const Expression *
NewGVN::performSymbolicAggrValueEvaluation(Instruction *I) const {
if (auto *EI = dyn_cast<ExtractValueInst>(I)) {
auto *II = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
if (II && EI->getNumIndices() == 1 && *EI->idx_begin() == 0) {
unsigned Opcode = 0;
// EI might be an extract from one of our recognised intrinsics. If it
// is we'll synthesize a semantically equivalent expression instead on
// an extract value expression.
switch (II->getIntrinsicID()) {
case Intrinsic::sadd_with_overflow:
case Intrinsic::uadd_with_overflow:
Opcode = Instruction::Add;
break;
case Intrinsic::ssub_with_overflow:
case Intrinsic::usub_with_overflow:
Opcode = Instruction::Sub;
break;
case Intrinsic::smul_with_overflow:
case Intrinsic::umul_with_overflow:
Opcode = Instruction::Mul;
break;
default:
break;
}
if (Opcode != 0) {
// Intrinsic recognized. Grab its args to finish building the
// expression.
assert(II->getNumArgOperands() == 2 &&
"Expect two args for recognised intrinsics.");
return createBinaryExpression(Opcode, EI->getType(),
II->getArgOperand(0),
II->getArgOperand(1), I);
}
}
auto *WO = dyn_cast<WithOverflowInst>(EI->getAggregateOperand());
if (WO && EI->getNumIndices() == 1 && *EI->idx_begin() == 0)
// EI is an extract from one of our with.overflow intrinsics. Synthesize
// a semantically equivalent expression instead of an extract value
// expression.
return createBinaryExpression(WO->getBinaryOp(), EI->getType(),
WO->getLHS(), WO->getRHS(), I);
}
return createAggregateValueExpression(I);