On Thu, Sep 10, 2020 at 11:51:51PM +0100, Kwok Cheung Yeung wrote:
> simd_clone_vector_of_formal_parm_types prefers to use the TYPE_ARG_TYPES of
> a function decl if available, but they do not contain entries for the hidden
> caf_* parameters. Replacements are therefore not generated for them, and the
> ICE occurs in ipa_simd_modify_function_body due to an assert firing when a
> replacement is not found for these hidden arguments.
So, isn't it a bug that TYPE_ARG_TYPES is inconsistent with DECL_ARGUMENTS
on the hidden arguments? That looks like lying to the middle-end to me,
one doesn't have always a declaration with DECL_ARGUMENTS around, so
sometimes TYPE_ARG_TYPES is the only way to get at the argument types.
We had e.g. PR92305 in the past...
Jakub
