[llvm-branch-commits] [llvm] [InstCombine][JumpTableToSwitch] Make a helper for iterating global elements (PR #242822)
Ömer Sinan Ağacan via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Sun Nov 15 06:36:41 PST 2026
================
@@ -1068,3 +1069,80 @@ LinearExpression llvm::decomposeLinearExpression(const DataLayout &DL,
return Expr;
}
+
+std::optional<ConstGlobalIterator>
+ConstGlobalIterator::fromLoadInst(LoadInst *LI) {
+ const DataLayout &DL = LI->getDataLayout();
+ Value *Ptr = LI->getPointerOperand();
+ auto *GV = dyn_cast<GlobalVariable>(getUnderlyingObject(Ptr));
+ if (LI->isVolatile() || !GV || !GV->isConstant() ||
+ !GV->hasDefinitiveInitializer())
+ return std::nullopt;
+
+ Type *EltTy = LI->getType();
+ TypeSize EltSize = DL.getTypeStoreSize(EltTy);
+ if (EltSize.isScalable())
+ return std::nullopt;
+
+ LinearExpression Expr = decomposeLinearExpression(DL, Ptr);
+ if (!Expr.Index || Expr.BasePtr != GV || Expr.Offset.getBitWidth() > 64)
+ return std::nullopt;
+
+ Constant *Init = GV->getInitializer();
+ TypeSize GlobalSize = DL.getTypeAllocSize(Init->getType());
+
+ const APInt &Stride = Expr.Scale;
+ const APInt &ConstOffset = Expr.Offset;
+
+ // Allow an additional context offset, but only within the stride.
+ if (!ConstOffset.ult(Stride))
+ return std::nullopt;
+
+ // Don't handle overlapping loads for now.
+ if (!Stride.uge(EltSize.getFixedValue()))
+ return std::nullopt;
+
+ // - With nuw: `Idx * Stride` does not wrap and all strides are valid.
+ // - With nusw: `Idx * Stride` does not wrap as signed, and the following
+ // enumeration over ArrayElementCount is valid so long as the stride is
+ // positive.
+ // - Without nusw or nuw: `Idx * Stride` is computed modulo 2^N (see
+ // getIndex), and the masking in getIndex is only exact when the stride is
+ // a power of two.
+ bool StrideIsValid =
+ Expr.Flags.hasNoUnsignedWrap() ||
+ (Expr.Flags.hasNoUnsignedSignedWrap() ? !Stride.isNegative()
+ : Stride.isPowerOf2());
+ if (!StrideIsValid)
+ return std::nullopt;
+
+ uint64_t ArrayElementCount =
+ divideCeil((GlobalSize.getFixedValue() - ConstOffset.getZExtValue()),
+ Stride.getZExtValue());
----------------
osa1 wrote:
This can wrap around and generate a large number. I think this is only working as expected today because all the call sites check the global size against a limit.
I also don't quite understand why the zext here is correct, given that the stride can be negative. We should document it.
https://github.com/llvm/llvm-project/pull/242822
More information about the llvm-branch-commits
mailing list