Problem with malloc on PA

Problem with malloc on PA

[Matthew Wilcox]

The PA-RISC architecture requires a lock to be 16-byte aligned.  So we
add __attribute__((aligned (16))) to our lock type and that aligns it
to a 16-byte offset within the struct.  But if the struct isn't 16-byte
aligned, this loses.

malloc() currently returns pointers which are 8-byte aligned.  Would it
be possible to change it so that they're 16-byte aligned?  I took a quick
look at the code, and I saw MALLOC_ALIGN_MASK, but I'm not sure exactly
what effects simply changing this would have.  I suspect I would want
to change MALLOC_ALIGNMENT as well, but not SIZE_SZ?

help greatly appreciated.

Re: Problem with malloc on PA

[Roland McGrath]

> The PA-RISC architecture requires a lock to be 16-byte aligned.  So we
> add __attribute__((aligned (16))) to our lock type and that aligns it
> to a 16-byte offset within the struct.  But if the struct isn't 16-byte
> aligned, this loses.

That sure confer the 16-byte alignment requirement onto the containing struct.
Then you can replace malloc(sizeof(struct foo)) with
memalign(__alignof(struct foo), sizefo(struct foo)).

Re: Problem with malloc on PA

[Ulrich Drepper]

willy@ldl.fc.hp.com (Matthew Wilcox) writes:

> malloc() currently returns pointers which are 8-byte aligned.  Would it
> be possible to change it so that they're 16-byte aligned?

Not in general, but for PA, yes.

> I took a quick look at the code, and I saw MALLOC_ALIGN_MASK, but
> I'm not sure exactly what effects simply changing this would have.
> I suspect I would want to change MALLOC_ALIGNMENT as well, but not
> SIZE_SZ?

What type does gcc return __alignof__(type) == 16 for?  We should be
able to define

  #define MALLOC_ALIGNMENT MAX(SIZE_SZ + SIZE_SZ, __alignof__(type))

The rest should then fall into place.

-- 
---------------.                          ,-.   1325 Chesapeake Terrace
Ulrich Drepper  \    ,-------------------'   \  Sunnyvale, CA 94089 USA
Red Hat          `--' drepper at redhat.com   `------------------------

Re: Problem with malloc on PA

[Matthew Wilcox]

On Wed, Apr 25, 2001 at 03:38:26PM -0700, Ulrich Drepper wrote:
> Not in general, but for PA, yes.

Great!

> What type does gcc return __alignof__(type) == 16 for?  We should be
> able to define
> 
>   #define MALLOC_ALIGNMENT MAX(SIZE_SZ + SIZE_SZ, __alignof__(type))
> 
> The rest should then fall into place.

_lt_spinlock_t in our tree, see my other mail.

Re: Problem with malloc on PA

[Ulrich Drepper]

willy@ldl.fc.hp.com (Matthew Wilcox) writes:

> _lt_spinlock_t in our tree, see my other mail.

I generally available type.  From the above I assume there is none and
this is an issue with the locking hardware support.  In this case
we'll do something else.  We allow MALLOC_ALIGNMENT to be defined on
the compiler command line and only use the default if it isn't.  The
PA Makefiles will then define it as 16.  I'll make this change now.

-- 
---------------.                          ,-.   1325 Chesapeake Terrace
Ulrich Drepper  \    ,-------------------'   \  Sunnyvale, CA 94089 USA
Red Hat          `--' drepper at redhat.com   `------------------------

Re: Problem with malloc on PA

[Matthew Wilcox]

On Wed, Apr 25, 2001 at 05:00:27PM -0700, Ulrich Drepper wrote:
> willy@ldl.fc.hp.com (Matthew Wilcox) writes:
> 
> > _lt_spinlock_t in our tree, see my other mail.
> 
> I generally available type.  From the above I assume there is none and
> this is an issue with the locking hardware support.  In this case
> we'll do something else.  We allow MALLOC_ALIGNMENT to be defined on
> the compiler command line and only use the default if it isn't.  The
> PA Makefiles will then define it as 16.  I'll make this change now.

The other patch would make _lt_spinlock_t a generally available type.
For all other architectures, it would be an int, for PA, a 16-byte-aligned
int.