[HawkForge]

Fixes #3489. Thanks to @lwz23 for the report.

For void (__attribute__((noreturn)) *take(int t))(int); libclang spells the function type as void (*(int))(int) __attribute__((noreturn)). FunctionSig::from_ty looks for __attribute__((noreturn)) at parenthesis depth 0 of that spelling, but here the attribute belongs to the pointed-to function type, not to take, so take was emitted as -> !.

When the return type is spelled as a raw function pointer, the function's own attributes follow its parameters inside the (*...) group, e.g. void (*(int) __attribute__((noreturn)))(int) __attribute__((noreturn)) for a function that is noreturn itself. So in that case the check now looks at depth 1 instead of 0. A return type spelled through a typedef (fp (int) __attribute__((noreturn))) and plain functions keep the depth-0 check, and argument attributes (the #2715/#2716 case) stay at depth 2+.

The header test covers take, a noreturn function returning a noreturn function pointer, the typedef'd return with and without noreturn, and a plain noreturn function. Only take changes compared to main; the test fails without the change.

[emilio]

Is this the right check tho? What if I return an alias or so? What if I have both? E.g:

void (__attribute__((noreturn)) *take(int t))(int) __attribute__((noreturn));
void plain(int t) __attribute__((noreturn));

View changes since the review

[HawkForge]

good catch, both of those broke. with the first version take_noreturn-style decls and a noreturn function returning a typedef'd fn pointer both lost -> !.

looked at what libclang actually spells:

• void (__attribute__((noreturn)) *take(int))(int) → void (*(int))(int) __attribute__((noreturn))
• the same plus a trailing __attribute__((noreturn)) → void (*(int) __attribute__((noreturn)))(int) __attribute__((noreturn))
• fp f(int) __attribute__((noreturn)) with fp a typedef → fp (int) __attribute__((noreturn))

so when the return type is spelled as a raw function pointer, the function's own attributes end up right after its params inside the (*...) group (depth 1), and depth 0 is the pointee's. for an alias (or anything else) it's still depth 0. pushed that: it now checks ret.kind() == CXType_Pointer on the non-canonical type (so aliases keep the old path) and looks at depth 1 in that case. args still sit at depth 2+, so the #2715 case is unchanged.

the test now covers plain, both, alias and alias + noreturn. all match main except take, which is the fix.