[clang] [CIR] Keep CIR global types visible across lowering replacements (PR #238698)
Henrich Lauko via cfe-commits
cfe-commits at lists.llvm.org
Thu Nov 5 01:53:49 PST 2026
================
@@ -3420,21 +3460,66 @@ CIRToLLVMGlobalOpLowering::matchAndRewriteRegionInitializedGlobal(
// to the appropriate value.
const mlir::Location loc = op.getLoc();
mlir::ModuleOp moduleOp = op->getParentOfType<mlir::ModuleOp>();
- if (failed(setupRegionInitializedLLVMGlobalOp(op, rewriter)))
+ // Build the cached table before the replacement exists, so the swap below
+ // drops a real entry instead of depending on duplicate tolerance.
+ auto &symbolTable = symbolTables.getSymbolTable(moduleOp);
+ // Create the replacement beside the root op, not via replaceOp: the table
+ // swap below and the visit both need the old op alive.
+ rewriter.setInsertionPoint(op);
+ mlir::LLVM::GlobalOp newGlobalOp =
+ setupRegionInitializedLLVMGlobalOp(op, rewriter);
+ if (!newGlobalOp)
return mlir::failure();
-
+ // Keep the old op in the table during the visit so self-references
+ // resolve to the CIR type, not the possibly adjusted LLVM type.
// Pass blockInfoAddr so that block address initializers (either as the whole
// initializer or nested inside an aggregate) can be resolved by the
// BlockAddrInfoAttr visitor.
CIRAttrToValue valueConverter(loc, moduleOp, rewriter, symbolTables,
- typeConverter, &blockInfoAddr);
+ typeConverter, preConversionTypes,
+ &blockInfoAddr);
mlir::Value value = valueConverter.visit(init);
- if (!value)
+ if (!value) {
+ // Drop the speculative replacement; the table still holds the live
+ // root op for the driver to report on. Everything created so far
+ // lives inside the new global's region, so erasing it restores the
+ // module exactly: module-level ops must wait for a successful visit
+ // (see the comdat below).
+ rewriter.eraseOp(newGlobalOp);
return mlir::failure();
+ }
mlir::LLVM::ReturnOp::create(rewriter, loc, value);
+ // Attach comdat only after the visit succeeds, so the failure path
+ // above only has the new global itself to drop.
+ if (mlir::SymbolRefAttr comdatAttr = getComdatAttr(op, rewriter))
+ newGlobalOp.setComdatAttr(comdatAttr);
+ symbolTable.remove(op);
+ symbolTable.insert(newGlobalOp);
+ rewriter.eraseOp(op);
----------------
xlauko wrote:
The global and alias paths now do remove/insert on the cached symbol table, but this path doesn’t. Without rollback, a later lookupSymbolIn for this function would return the erased cir.func. Can you handle it the same way, or check why it’s safe?
https://github.com/llvm/llvm-project/pull/238698
More information about the cfe-commits
mailing list