Re: [rfc] stack alignment macro cleanup

["H.J. Lu" <hjl dot tools at gmail dot com>]

On Mon, Oct 4, 2010 at 11:27 AM, Richard Henderson <rth@redhat.com> wrote:
> On 10/02/2010 04:03 PM, H.J. Lu wrote:
>>> MIN_STACK_BOUNDARY
>>> ?(undocumented; local to i386 atm)
>>> ?-- appears to be the ABI specified stack boundary, i.e.
>>> ?the minimum that must be in place at a call site. ?This
>>> ?somehow differs from I_S_B due to proliferation of
>>> ?command-line options.
>>
>> It is used to implement -mstackreliagn. I think you should just
>> move it to i386.c. Otherwise, you have to copy the same logic
>> to where it is used in i386.c.
>
> Well, the problem here is that there's stuff in i386.c related
> to enforcing outgoing stack boundary that shouldn't be there.
> It should be in the generic code where we manage the other bits
> of the stack boundary code.

Ia32 is a special case where the incoming/outgoing stack boundary
was changed from 4 byte to 16byte.  If you want to move that part
to generic, you need to define it for generic.

>> On ia32, ?INCOMING_STACK_BOUNDARY may be different
>> from PREFERRED_STACK_BOUNDARY. It is used to support
>> linking against libraries with 4byte incoming stack boundary and
>> generate 16byte outgoing stack boundary.
>
> As an aside let me say that I hate the word "preferred" in this
> context. ?It implies that the boundary is optional, and it's not.

Change it to OUTGOING_STACK_BOUNDARY.

> Second, we're providing too many confusing options to our users.
> While it's fine for LTO to manage all sorts of differing stack
> alignment requirements for the functions its compiling, it's not
> ok for us to expose all that complexity to the user.
>
> There should only be the stack boundary as defined by the ABI,
> plus the increased boundary discovered for local functions and
> managed by cgraph.
>
> I suspect that we have to keep the x86 backend command-line
> option that adjusts the ABI boundary for backward compatibility
> with previous gcc versions.

You need MIN_STACK_BOUNDARY, INCOMING_STACK_BOUNDARY
and PREFERRED_STACK_BOUNDARY to support those.

>> -mdrap is mainly for testing purpose and used in testsuite.
>> It has caught many bugs. Removing it means regressions
>> may become latent.
>
> That simply means that the set of test cases is incomplete.
> You get DRAP with highly aligned local variables plus alloca.
> These same conditions ought to be testable with the code that
> I'm adding for generic alignment of user variables, but do not
> use a DRAP register.
>

As I remembered, -mforce-drap exposed issues with register allocator.
ix86_force_drap is only referenced in one place in i386.c. I'd like to keep
it.  I don't see why it can't be moved to generic. It may expose problems
for other targets.

Thanks.


-- 
H.J.