Re: [PATCH] Use DECL_ARG_TYPE for argument types in fnsplitting (PR tree-optimization/47148)

[Jakub Jelinek <jakub at redhat dot com>]

On Sun, Jan 02, 2011 at 06:43:34PM +0100, Richard Guenther wrote:
> Ick ...  I don't like
> 
> > + ? ? ? arg = gimple_default_def (cfun, parm);
> > + ? ? ? if (!useless_type_conversion_p (DECL_ARG_TYPE (parm), TREE_TYPE (arg))
> > + ? ? ? ? ? && fold_convertible_p (DECL_ARG_TYPE (parm), arg))
> 
> this check.  Can we instead not do function splitting here?  And maybe

You're worried just about the fold_convertible_p check, right?
If the conversion is useless, there is no point to add it.

It was added there just in case, I guess it could be an gcc_checking_assert
too, or such test could be moved to execute_split_function.  split_function
is already meant to be executed when we are committed to do the splitting.

> fix the x86 backend as well for a start?

x86_64?  Maybe.  I don't think we want to touch i?86 ABI (where the test
fails too), haven't tried other targets to see where else it fails as well.
I think DECL_ARG_TYPE isn't a target thing, it is something the FE decided
and the middle-end should honor it.  So it isn't something that should
be controlled by a target hook.

	Jakub