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

[Alan Lawrence <alan.lawrence@arm.com>]

Richard Earnshaw wrote:
> On 01/06/15 13:07, Jakub Jelinek wrote:
>> On Thu, May 07, 2015 at 12:16:32PM +0100, Alan Lawrence wrote:
>>> So for my two cents, or perhaps three:
>> Any progress on this PR?
>> A P1 bug that affects several packages stalled for a month isn't a very good
>> thing... (not to mention broken profiledbootstrap on ARM due to the same
>> issue).
>> I've checked and llvm on ARM ignores the alignment on the scalar
>> arguments...
>>
>> 	Jakub
>>
> 
> We're working on some updates to the ABI documents.  If we're going to
> break ABI compatibility, even in some corner cases, it would make sense
> to only do this once.

One question is whether to treat structs differently from scalars in the ABI 
specification. Structs raise lots of corner cases! I notice the following 
oddity, and wonder if anyone can shed any light on this:

typedef __attribute__((aligned(8))) struct { int x; int y; } foo;
typedef struct __attribute__((aligned(8))) { int x; int y; } bar;
typedef struct { int x; int y; } __attribute__((aligned(8))) baz;
typedef struct { int x; int y; } qux __attribute__((aligned(8)));

create typedefs (foo, bar, baz, qux) all with alignment 8, as expected. However, 
the TYPE_MAIN_VARIANT (an anonymous struct type) has alignment 4 for foo and qux 
(so the attribute has been applied only to the typedef), but alignment 8 for bar 
and baz (i.e. the attribute has been applied to the underlying struct).

--Alan