Re: [PATCH] Fix eipa_sra AAPCS issue (PR target/65956)

[Richard Earnshaw <Richard.Earnshaw@foss.arm.com>]

On 05/05/15 15:29, Jakub Jelinek wrote:
> On Tue, May 05, 2015 at 02:20:43PM +0100, Richard Earnshaw wrote:
>> On 05/05/15 14:06, Jakub Jelinek wrote:
>>> On Tue, May 05, 2015 at 02:02:19PM +0100, Richard Earnshaw wrote:
>>>> In a literal sense, yes.  However, even K&R & stdarg have standard
>>>> promotion and conversion rules (size < int => int, floats promoted to
>>>> double, etc).  What are those rules for GCC's overaligned types (ie
>>>> where in the docs does it say what happens and how a back-end should
>>>> interpret them)?  Shouldn't the mid-end be doing that work so as to
>>>
>>> For the middle-end, the TYPE_ALIGN info on expression types is considered
>>> useless, you can get there anything.  There is no conversion rule to what
>>> you get for myalignedint + int, or (myalignedint) int, or (int)
>>> myalignedint, etc.
>>>
>>>> create a consistent view of the values passed into the back-end?  It
>>>> seems to me that at present the back-end has to be psychic as to what is
>>>> really happening.
>>>
>>> No, the backend just shouldn't consider TYPE_ALIGN on the scalars, and it
>>> seems only arm ever looks at that.
>>>
>>
>> Nothing in the specification for TARGET_FUNCTION_ARG (or any of the
>> related functions) makes any mention of this...
> 
> While this requirement isn't documented, it is clearly common sense or at
> least something any kind of testing would reveal immediately.

Then clearly no such tests exist in the testsuite :-(

R.

> And it is nothing broken recently (except that with the SRA changes it hits
> much more often), looking e.g. at GCC 3.2, I'm seeing that expand_call is on
> that testcase also called with pretty random TYPE_ALIGN on the argument
> types; we didn't have GIMPLE then, so it is nothing that GIMPLE brought in.
> 
> 	Jakub
>