================
@@ -3420,8 +3412,15 @@ CIRGenModule::createCIRFunction(mlir::Location loc,
StringRef name,
//
// Be sure to insert a new function before a current one.
CIRGenFunction *cgf = this->curCGF;
- if (cgf)
+ if (cgf) {
builder.setInsertionPoint(cgf->curFn);
+ } else {
+ // No CIRGenFunction is active, but the builder's insertion point may
+ // still be inside another op (e.g. a function materialized on demand
+ // while generating a vtable thunk). Insert at module scope so the new
+ // function is not parented under the ambient insertion point.
+ builder.setInsertionPointToEnd(theModule.getBody());
----------------
adams381 wrote:
No good reason, and it is unconditional now. The fallout is 86 test files and
it is all one shape: an on-demand declaration used to be emitted immediately
before the definition that triggered it, and now it follows it.
@andykaylor, your ordering hunch was right. Three tests carried a comment
recording that CIR and classic disagreed on emission order,
`array-init-loop-exprs.cpp`, `delegating-ctor.cpp` and
`lambda-generic-in-cxx11abi-lambda.cpp`. All three now emit in the same order
as classic, so the comments are gone. `setjmp.c` had a second check prefix
that existed only to tolerate the disagreement, and that is gone too.
https://github.com/llvm/llvm-project/pull/207260
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
