Files
swift-mirror/lib/SILOptimizer/Transforms/RedundantOverflowCheckRemoval.cpp
practicalswift 492f5cd35a [gardening] Remove redundant repetition of type names (DRY): RepeatedTypeName foo = dyn_cast<RepeatedTypeName>(bar)
Replace `NameOfType foo = dyn_cast<NameOfType>(bar)` with DRY version `auto foo = dyn_cast<NameOfType>(bar)`.

The DRY auto version is by far the dominant form already used in the repo, so this PR merely brings the exceptional cases (redundant repetition form) in line with the dominant form (auto form).

See the [C++ Core Guidelines](https://github.com/isocpp/CppCoreGuidelines/blob/master/CppCoreGuidelines.md#es11-use-auto-to-avoid-redundant-repetition-of-type-names) for a general discussion on why to use `auto` to avoid redundant repetition of type names.
2017-05-05 09:45:53 +02:00

703 lines
25 KiB
C++

//===--- RedundantOverflowCheckRemoval.cpp --------------------------------===//
//
// This source file is part of the Swift.org open source project
//
// Copyright (c) 2014 - 2017 Apple Inc. and the Swift project authors
// Licensed under Apache License v2.0 with Runtime Library Exception
//
// See https://swift.org/LICENSE.txt for license information
// See https://swift.org/CONTRIBUTORS.txt for the list of Swift project authors
//
//===----------------------------------------------------------------------===//
// Remove overflow checks that are guarded by control flow or other
// overflow checks.
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "remove-redundant-overflow-checks"
#include "swift/SIL/Dominance.h"
#include "swift/SILOptimizer/Analysis/DominanceAnalysis.h"
#include "swift/SILOptimizer/Analysis/PostOrderAnalysis.h"
#include "swift/SILOptimizer/Analysis/Analysis.h"
#include "swift/SILOptimizer/PassManager/Passes.h"
#include "swift/SILOptimizer/PassManager/Transforms.h"
#include "swift/SILOptimizer/Utils/Local.h"
#include "swift/SIL/SILInstruction.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Support/Debug.h"
using namespace swift;
STATISTIC(NumCondFailRemoved, "Number of cond_fail instructions removed");
namespace {
class RedundantOverflowCheckRemovalPass : public SILFunctionTransform {
public:
RedundantOverflowCheckRemovalPass() {}
/// This enum represents a relationship between two operands.
/// The relationship represented by arithmetic operators represents the
/// information that the operation did not trap.
///
/// The following code translate (with the correct signedness prefix):
///
/// if (x > 2) { x } -> LT(2, x)
/// if (x > 2) {} else { x } -> LE(x, 2)
/// x - 2 -> Sub(x, 2)
/// 2 - x -> Sub(2, x)
/// 2 * x -> Mul(2, x)
/// x + y -> Add(x, y)
enum class ValueRelation {EQ, ULT, ULE, UAdd, USub, UMul,
SLT, SLE, SAdd, SSub, SMul};
/// This struct represents a constraint on the range of some values in some
/// basic blocks in the program.
/// For example, it can represent the constraint "X < 2" for some blocks in
/// the function.
struct Constraint {
Constraint(SILBasicBlock *BB, SILValue L, SILValue R, ValueRelation Rel) :
DominatingBlock(BB), Left(L), Right(R), Relationship(Rel) {
DEBUG(dump());
}
/// The constraint is valid in blocks dominated by this block.
SILBasicBlock *DominatingBlock;
/// The first operand.
SILValue Left;
/// The second operand.
SILValue Right;
/// Describes the relationship between the operands.
ValueRelation Relationship;
/// Print the content of the constraint.
void dump() {
llvm::dbgs()<<"Constraint [" << DominatingBlock <<"]\n";
llvm::dbgs()<<" Relationship:";
switch (Relationship) {
case ValueRelation::EQ: llvm::dbgs()<<"Equal\n"; break;
case ValueRelation::SLT: llvm::dbgs()<<"SLT\n"; break;
case ValueRelation::ULT: llvm::dbgs()<<"ULT\n"; break;
case ValueRelation::SLE: llvm::dbgs()<<"SLE\n"; break;
case ValueRelation::ULE: llvm::dbgs()<<"ULE\n"; break;
case ValueRelation::SMul: llvm::dbgs()<<"SMul\n"; break;
case ValueRelation::SSub: llvm::dbgs()<<"SSub\n"; break;
case ValueRelation::SAdd: llvm::dbgs()<<"SAdd\n"; break;
case ValueRelation::UMul: llvm::dbgs()<<"UMul\n"; break;
case ValueRelation::USub: llvm::dbgs()<<"USub\n"; break;
case ValueRelation::UAdd: llvm::dbgs()<<"UAdd\n"; break;
}
llvm::dbgs()<<" Left:"; Left->dump();
llvm::dbgs()<<" Right:"; Right->dump();
}
};
typedef SmallVector<Constraint, 16> ConstraintList;
typedef SmallVector<CondFailInst*, 16> CondFailList;
/// A list of constraints that represent the value relationships.
ConstraintList Constraints;
/// A list of cond_fail instructions to remove.
CondFailList ToRemove;
// Dominators info.
DominanceInfo *DT;
/// Remove the instructions that were marked as redundant
/// and return True if and instructions were removed.
bool removeCollectedRedundantInstructions() {
if (ToRemove.size()) {
DEBUG(llvm::dbgs()<<"Removing "<<ToRemove.size()<<" condfails in "
<<getFunction()->getName()<<"\n");
for (auto *CF : ToRemove) {
CF->eraseFromParent();
NumCondFailRemoved++;
}
ToRemove.clear();
return true;
}
return false;
}
void run() override {
DT = PM->getAnalysis<DominanceAnalysis>()->get(getFunction());
auto *PO = getAnalysis<PostOrderAnalysis>()->get(getFunction());
Constraints.clear();
ToRemove.clear();
auto ReversePostOrder = PO->getReversePostOrder();
// Perform a forward scan and use control flow and previously detected
// overflow checks to remove the overflow checks.
// For each block in a Reverse Post Order scan:
for (auto &BB : ReversePostOrder) {
// For each instruction:
for (auto Inst = BB->begin(), End = BB->end(); Inst != End; Inst++) {
// Use branch information for eliminating condfails.
if (auto *CBI = dyn_cast<CondBranchInst>(Inst))
registerBranchFormula(CBI);
// Handle cond_fail instructions.
if (auto *CFI = dyn_cast<CondFailInst>(Inst)) {
if (tryToRemoveCondFail(CFI)) {
ToRemove.push_back(CFI);
continue;
}
// We were not able to remove the condfail. Try to use this
// information to remove other cond_fails.
registerCondFailFormula(CFI);
}
}
}
// If we've collected redundant cond_fails then remove them now.
bool Changed = removeCollectedRedundantInstructions();
// Perform another check, this time in reverse and use future overflow
// checks that must be executed to eliminate earlier overflow checks.
// Notice that this scan is only block local because at this point we
// don't use post-dominators.
for (auto &BB : ReversePostOrder) {
// Clear the list of constraint on every block.
Constraints.clear();
// Notice: we scan the basic block in reverse.
for (auto Inst = --BB->end(), End = BB->begin(); Inst != End; --Inst) {
if (auto *CFI = dyn_cast<CondFailInst>(Inst)) {
// Try to remove the cond_fail based on previous overflow checks.
if (tryToRemoveCondFail(CFI)) {
ToRemove.push_back(CFI);
continue;
}
// Record the overflow check and try to optimize other checks.
registerCondFailFormula(CFI);
continue;
}
// We do not optimize overflow checks across instructions with side
// effects because we don't want to delay the trap past user-visible
// changes.
if (Inst->mayHaveSideEffects()) {
Constraints.clear();
continue;
}
}
}
// If we've collected more redundant cond_fails then remove them now.
Changed |= removeCollectedRedundantInstructions();
if (Changed)
PM->invalidateAnalysis(getFunction(),
SILAnalysis::InvalidationKind::Instructions);
}
/// Return True if the relationship \p Rel describes a known relation
/// between A and B.
static bool knownRelation(SILValue A, SILValue B, ValueRelation Rel) {
// Identical values are known to be equal, or less than or equal.
if ((A == B) && (Rel == ValueRelation::EQ ||
Rel == ValueRelation::SLE ||
Rel == ValueRelation::ULE))
return true;
// Evaluate literal integers.
auto *AI = dyn_cast<IntegerLiteralInst>(A);
auto *BI = dyn_cast<IntegerLiteralInst>(B);
if (AI && BI) {
APInt Ap = AI->getValue();
APInt Bp = BI->getValue();
if (Ap.getBitWidth() != Bp.getBitWidth())
return false;
switch (Rel) {
case ValueRelation::EQ: return Ap.eq(Bp);
case ValueRelation::SLE: return Ap.sle(Bp);
case ValueRelation::ULE: return Ap.ule(Bp);
case ValueRelation::SLT: return Ap.slt(Bp);
case ValueRelation::ULT: return Ap.ult(Bp);
default: llvm_unreachable("Invalid value relation");
}
}
return false;
}
/// Return True if we can deduct that \p N is always positive (N > 0).
static bool isKnownPositive(SILValue N) {
if (auto *NI = dyn_cast<IntegerLiteralInst>(N))
return NI->getValue().isStrictlyPositive();
return false;
}
/// Return true if the absolute value of \p A is smaller than the
/// absolute value of \p B. In other words, check if \p A known to be closer
/// to zero.
static bool isKnownAbsLess(SILValue A, SILValue B) {
auto *AI = dyn_cast<IntegerLiteralInst>(A);
auto *BI = dyn_cast<IntegerLiteralInst>(B);
if (AI && BI)
return AI->getValue().abs().ult(BI->getValue().abs());
return false;
}
/// Return true if the constraint \p F can prove that the overflow check
/// for \p BI is not needed.
static bool isOverflowCheckRemovedByConstraint(Constraint &F,
BuiltinInst *BI) {
// L and R are the righthand and lefthand sides of the constraint.
SILValue L = F.Left;
SILValue R = F.Right;
assert(L->getType() == R->getType() && "Invalid constraint type");
// Make sure that the types of the constraints match the types of the
// arithmetic operation.
switch (BI->getBuiltinInfo().ID) {
default: return false;
case BuiltinValueKind::SAddOver:
case BuiltinValueKind::UAddOver:
case BuiltinValueKind::SMulOver:
case BuiltinValueKind::UMulOver:
case BuiltinValueKind::USubOver:
case BuiltinValueKind::SSubOver:
if (L->getType() != BI->getOperand(0)->getType())
return false;
}
switch (BI->getBuiltinInfo().ID) {
default: return false;
case BuiltinValueKind::SAddOver:
// A + B traps unless:
if (F.Relationship == ValueRelation::SAdd) {
// L + R already known to not trap at this point in the program.
// And the following applies:
// L >= A and R >= B or (commutatively) R >= A and L >= B.
SILValue A = BI->getOperand(0);
SILValue B = BI->getOperand(1);
if (knownRelation(A, L, ValueRelation::SLE) &&
knownRelation(B, R, ValueRelation::SLE))
return true;
if (knownRelation(B, L, ValueRelation::SLE) &&
knownRelation(A, R, ValueRelation::SLE))
return true;
}
// A + 1 does not trap if A is smaller than anything.
if (F.Relationship == ValueRelation::SLT) {
SILValue A = BI->getOperand(0);
SILValue B = BI->getOperand(1);
auto *AI = dyn_cast<IntegerLiteralInst>(A);
auto *BI = dyn_cast<IntegerLiteralInst>(B);
if (L == A && BI && BI->getValue().getSExtValue() == 1)
return true;
if (L == B && AI && AI->getValue().getSExtValue() == 1)
return true;
}
return false;
case BuiltinValueKind::UAddOver:
// A + B traps unless:
if (F.Relationship == ValueRelation::UAdd) {
// L + R already known to not trap at this point in the program.
// And the following applies:
// L >= A and R >= B or (commutatively) R >= A and L >= B.
SILValue A = BI->getOperand(0);
SILValue B = BI->getOperand(1);
if (knownRelation(A, L, ValueRelation::ULE) &&
knownRelation(B, R, ValueRelation::ULE))
return true;
if (knownRelation(B, L, ValueRelation::ULE) &&
knownRelation(A, R, ValueRelation::ULE))
return true;
}
// A + 1 does not trap if A is smaller than anything.
if (F.Relationship == ValueRelation::ULT) {
SILValue A = BI->getOperand(0);
SILValue B = BI->getOperand(1);
auto *AI = dyn_cast<IntegerLiteralInst>(A);
auto *BI = dyn_cast<IntegerLiteralInst>(B);
if (L == A && BI && BI->getValue().getZExtValue() == 1)
return true;
if (L == B && AI && AI->getValue().getZExtValue() == 1)
return true;
}
return false;
case BuiltinValueKind::SMulOver:
// A * B traps unless:
if (F.Relationship == ValueRelation::SMul) {
// L * R already known to not trap at this point in the program and
// the following rules apply:
//
// A is closer zero than L and B == R,
// or A == L, and B is closer to zero than R.
//
// We do not allow removing the overflow checks when one of the
// multipliers just switches the sign (abs(L) == abs(A)) because
// there are more negative numbers and (-MIN_INT * -1 overflows).
// In other words X * -1 does not does not guard X * 1.
SILValue A = BI->getOperand(0);
SILValue B = BI->getOperand(1);
if (isKnownAbsLess(A, L) &&
knownRelation(B, R, ValueRelation::EQ))
return true;
if (knownRelation(A, L, ValueRelation::EQ) &&
isKnownAbsLess(B, R))
return true;
// And commutatively, swapping A and B.
if (isKnownAbsLess(B, L) &&
knownRelation(A, R, ValueRelation::EQ))
return true;
if (knownRelation(B, L, ValueRelation::EQ) &&
isKnownAbsLess(A, R))
return true;
}
return false;
case BuiltinValueKind::UMulOver:
// A * B traps unless:
if (F.Relationship == ValueRelation::UMul) {
// L * R already known to not trap at this point in the program.
// And the following applies:
// L >= A and R >= B or (commutatively) R >= A and L >= B.
SILValue A = BI->getOperand(0);
SILValue B = BI->getOperand(1);
if (knownRelation(A, L, ValueRelation::ULE) &&
knownRelation(B, R, ValueRelation::ULE))
return true;
if (knownRelation(B, L, ValueRelation::ULE) &&
knownRelation(A, R, ValueRelation::ULE))
return true;
}
return false;
case BuiltinValueKind::USubOver:
// A - B traps unless:
if (F.Relationship == ValueRelation::ULE ||
F.Relationship == ValueRelation::ULT) {
// A >= B.
// Given the constraint L < R check if:
// 1. R == A
// 2. B <= L (subtracting less than L is okay)
//
// Example: Given 2<X we know that X-2 can't trap.
SILValue A = BI->getOperand(0);
SILValue B = BI->getOperand(1);
if (knownRelation(R, A, ValueRelation::EQ) &&
knownRelation(B, L, ValueRelation::ULE)) {
return true;
}
}
if (F.Relationship == ValueRelation::EQ) {
SILValue A = BI->getOperand(0);
SILValue B = BI->getOperand(1);
// A - B, L == R and known that L >= A and R == B.
if (knownRelation(R, B, ValueRelation::EQ) &&
knownRelation(A, L, ValueRelation::ULE)) {
return true;
}
// Swap L and R because equality is commutative.
if (knownRelation(L, B, ValueRelation::EQ) &&
knownRelation(A, R, ValueRelation::ULE)) {
return true;
}
}
if (F.Relationship == ValueRelation::USub) {
// L - R already known to not trap at this point in the program.
// And the following applies:
// L <= A and B <= R.
SILValue A = BI->getOperand(0);
SILValue B = BI->getOperand(1);
if (knownRelation(L, A, ValueRelation::ULE) &&
knownRelation(B, R, ValueRelation::ULE))
return true;
}
// A - 1 does not trap if A is greater than some other number.
if (F.Relationship == ValueRelation::ULT) {
SILValue A = BI->getOperand(0);
SILValue B = BI->getOperand(1);
auto *BI = dyn_cast<IntegerLiteralInst>(B);
if (R == A && BI && BI->getValue().getZExtValue() == 1)
return true;
}
return false;
case BuiltinValueKind::SSubOver:
// A - B traps unless:
if (F.Relationship == ValueRelation::SLE ||
F.Relationship == ValueRelation::SLT) {
// A >= B and B is positive.
// Notice that we need to handle underflow and overflow.
// Given the constraint L < R check if:
// 1. L is positive (because double negative can overflow)
// 2. R == A
// 3. B <= L (subtracting less than L is okay)
//
// Example: Given 2<X we know that X-2 can't trap.
SILValue A = BI->getOperand(0);
SILValue B = BI->getOperand(1);
if (isKnownPositive(L) &&
knownRelation(R, A, ValueRelation::EQ) &&
knownRelation(B, L, ValueRelation::SLE)) {
return true;
}
}
if (F.Relationship == ValueRelation::SSub) {
// L - R already known to not trap at this point in the program.
// And the following applies:
// L <= A and B <= R.
SILValue A = BI->getOperand(0);
SILValue B = BI->getOperand(1);
if (knownRelation(L, A, ValueRelation::SLE) &&
knownRelation(B, R, ValueRelation::SLE))
return true;
}
if (F.Relationship == ValueRelation::EQ) {
SILValue A = BI->getOperand(0);
SILValue B = BI->getOperand(1);
// A - B, L == R and known that L >= A and R == B.
if (knownRelation(R, B, ValueRelation::EQ) &&
knownRelation(A, L, ValueRelation::SLE)) {
return true;
}
// Swap L and R because equality is commutative.
if (knownRelation(L, B, ValueRelation::EQ) &&
knownRelation(A, R, ValueRelation::SLE)) {
return true;
}
}
// A - 1 does not trap if A is greater than some other number.
if (F.Relationship == ValueRelation::SLT) {
SILValue A = BI->getOperand(0);
SILValue B = BI->getOperand(1);
auto *BI = dyn_cast<IntegerLiteralInst>(B);
if (R == A && BI && BI->getValue().getSExtValue() == 1)
return true;
}
return false;
}
}
bool tryToRemoveCondFail(CondFailInst *CFI) {
// Extract the arithmetic operation from the condfail.
auto *TEI = dyn_cast<TupleExtractInst>(CFI->getOperand());
if (!TEI) return false;
auto *BI = dyn_cast<BuiltinInst>(TEI->getOperand());
if (!BI) return false;
for (auto &F : Constraints) {
// If we are dominated by a constraint:
if (DT->dominates(F.DominatingBlock, CFI->getParent())) {
// Try to use the constraint to remove the overflow check.
if (isOverflowCheckRemovedByConstraint(F, BI)) {
return true;
}
}
}
// Was not able to remove this branch.
return false;
}
Optional<ValueRelation> getArithOpRelation(BuiltinInst *BI) {
ValueRelation Rel;
switch (BI->getBuiltinInfo().ID) {
default:
return None;
case BuiltinValueKind::SAddOver:
Rel = ValueRelation::SAdd;
break;
case BuiltinValueKind::UAddOver:
Rel = ValueRelation::UAdd;
break;
case BuiltinValueKind::SSubOver:
Rel = ValueRelation::SSub;
break;
case BuiltinValueKind::USubOver:
Rel = ValueRelation::USub;
break;
case BuiltinValueKind::SMulOver:
Rel = ValueRelation::SMul;
break;
case BuiltinValueKind::UMulOver:
Rel = ValueRelation::UMul;
break;
}
return Rel;
}
void addComparisonRelation(BuiltinInst *CMP, SILBasicBlock *TrueBB,
SILBasicBlock *FalseBB) {
// The relationship expressed in the builtin.
ValueRelation TrueRel;
ValueRelation FalseRel;
bool Swap = false;
switch (CMP->getBuiltinInfo().ID) {
default:
return;
case BuiltinValueKind::ICMP_NE: {
SILValue Left = CMP->getOperand(0);
SILValue Right = CMP->getOperand(1);
if (FalseBB)
Constraints.push_back(
Constraint(FalseBB, Left, Right, ValueRelation::EQ));
return;
}
case BuiltinValueKind::ICMP_EQ: {
SILValue Left = CMP->getOperand(0);
SILValue Right = CMP->getOperand(1);
if (TrueBB)
Constraints.push_back(
Constraint(TrueBB, Left, Right, ValueRelation::EQ));
return;
}
case BuiltinValueKind::ICMP_SLE:
TrueRel = ValueRelation::SLE;
FalseRel = ValueRelation::SLT;
break;
case BuiltinValueKind::ICMP_SLT:
TrueRel = ValueRelation::SLT;
FalseRel = ValueRelation::SLE;
break;
case BuiltinValueKind::ICMP_SGE:
TrueRel = ValueRelation::SLT;
FalseRel = ValueRelation::SLE;
Swap = true;
break;
case BuiltinValueKind::ICMP_SGT:
TrueRel = ValueRelation::SLE;
FalseRel = ValueRelation::SLT;
Swap = true;
break;
case BuiltinValueKind::ICMP_ULE:
TrueRel = ValueRelation::ULE;
FalseRel = ValueRelation::ULT;
break;
case BuiltinValueKind::ICMP_ULT:
TrueRel = ValueRelation::ULT;
FalseRel = ValueRelation::ULE;
break;
case BuiltinValueKind::ICMP_UGT:
TrueRel = ValueRelation::ULE;
FalseRel = ValueRelation::ULT;
Swap = true;
break;
case BuiltinValueKind::ICMP_UGE:
TrueRel = ValueRelation::ULT;
FalseRel = ValueRelation::ULE;
Swap = true;
break;
}
SILValue Left = CMP->getOperand(0);
SILValue Right = CMP->getOperand(1);
if (Swap)
std::swap(Left, Right);
// Set the constraints for both side of the conditional branch, if
// that the condition is dominating the dest block (see comment above).
if (TrueBB) Constraints.push_back(Constraint(TrueBB, Left, Right, TrueRel));
if (FalseBB) Constraints.push_back(Constraint(FalseBB, Right, Left, FalseRel));
}
void registerCondFailFormula(CondFailInst *CFI) {
// Extract the arithmetic operation from the condfail.
if (auto *TEI = dyn_cast<TupleExtractInst>(CFI->getOperand())) {
auto *BI = dyn_cast<BuiltinInst>(TEI->getOperand());
if (!BI)
return;
// The relationship expressed in the builtin.
Optional<ValueRelation> Rel = getArithOpRelation(BI);
if (!Rel.hasValue())
return;
// Construct and register the constraint.
SILBasicBlock *Dom = CFI->getParent();
SILValue Left = BI->getOperand(0);
SILValue Right = BI->getOperand(1);
Constraint F = Constraint(Dom, Left, Right, *Rel);
Constraints.push_back(F);
}
// Handle patterns like this:
// %cmp_result = builtin "cmp_ult_Int64"
// (%x : $Builtin.Int64, %y : $Builtin.Int64) : $Builtin.Int1
// This cond_fail formula should be registered!
// cond_fail %cmp_result : $Builtin.Int1
// %check_underflow = integer_literal $Builtin.Int1, -1
// At this point we know that x >= y
// %usub_result = builtin "usub_with_overflow_Int64"
// (%x : $Builtin.Int64, %y : $Builtin.Int64, %check_underflow : $Builtin.Int1):
// $(Builtin.Int64, Builtin.Int1)
// %usub_val = tuple_extract %usub_result : $(Builtin.Int64, Builtin.Int1), 0
// We can figure out that x - y will not underflow because of x >= y
// %usub_underflow = tuple_extract %usub_result : $(Builtin.Int64, Builtin.Int1), 1
// cond_fail %usub_underflow : $Builtin.Int1
if (auto *CMP = dyn_cast<BuiltinInst>(CFI->getOperand())) {
SILBasicBlock *TrueBB = nullptr;
SILBasicBlock *FalseBB = CMP->getParent();
addComparisonRelation(CMP, TrueBB, FalseBB);
}
}
void registerBranchFormula(CondBranchInst *BI) {
// Extract the arithmetic operation from the Branch.
auto *CMP = dyn_cast<BuiltinInst>(BI->getCondition());
if (!CMP) return;
SILBasicBlock *TrueBB = BI->getTrueBB();
SILBasicBlock *FalseBB = BI->getFalseBB();
// Notice that we need to handle control-flow programs such as the one
// below. The rule here is that only blocks with a single predecessor
// and blocks that are dominated by them can rely on branch information.
// The reason is that if there is not a single predecessor then the code
// that is dominated by the block can be reachable from other blocks.
//
// [ x > 2 ]
// / |
// / |
// / |
// [ .. ] |
// \ |
// \ |
// \ |
// \ v
// [use(x)]
if (!TrueBB->getSinglePredecessorBlock())
TrueBB = nullptr;
if (!FalseBB->getSinglePredecessorBlock())
FalseBB = nullptr;
addComparisonRelation(CMP, TrueBB, FalseBB );
}
};
} // end anonymous namespace
SILTransform *swift::createRedundantOverflowCheckRemoval() {
return new RedundantOverflowCheckRemovalPass();
}