On Thu, Mar 20, 2025 at 09:08:25AM +0100, Richard Biener wrote:
> Hmm. I do wonder whether your earlier patch was more "correct" in the
> sense that a tail call does not return to the calling function but its caller.
> That means it should not have a fallthru edge, so our representation
> with feeding a return value to a function-local return stmt isn't a good one.
Tail call certainly doesn't.
But what we have in the IL right now, i.e. musttail marked call in the
middle, isn't actually a tail call, it is a call that should be turned into
a tail call or error should be reported.
And that certainly returns to the caller (if it is not non-marked noreturn
or say exits/aborts/longjmp/throws) and performs the statements in there and
only then returns.
So pretending it doesn't return is IMHO risky. We could only do that if
we had in the IL something that would be a GIMPLE_RETURN combined with the
GIMPLE_CALL. I bet that is how LLVM implements it, but then the discovery
needs to be done in the FE, not late, clearly they don't bother with
discovering address escaped automatic vars still in scope and just declare
that UB, ...
All we should ensure is that we don't add after such calls code which would
turn those calls from successfully tail-callable into ones on which we
error.
Jakub
