Hi!
On Fri, Nov 06, 2020 at 04:18:00PM -0600, Peter Bergner wrote:
> As part of the MMA base support, we incremented BIGGEST_ALIGNMENT in
> order to align the __vector_pair and __vector_quad types to 256 and 512
> bits respectively. This had the unintended effect of changing the
> default alignment used by __attribute__ ((__aligned__)) which causes
> an ABI break because of some dodgy code in GLIBC's struct pthread
> (GLIBC is going to fix that too). The fix in GCC is to revert the
> BIGGEST_ALIGNMENT change and to force the alignment on the type itself
> rather than the mode used by the type.
The ABI break is all GCC's faultt, but it is exposed by that glibc code,
sure :-)
> This passes bootstrap and regtesting on powerpc64le-linux with no regressions.
> Ok for mainline and the GCC 10 release branch after some burn in?
Yes, okay for both. Thanks!
(See note below though.)
> gcc/
> * config/rs6000/rs6000.h (BIGGEST_ALIGNMENT): Revert previous commit
> so as not to break the ABI.
> * config/rs6000/rs6000-call.c (rs6000_init_builtins): Set the ABI
> mandated alignment for __vector_pair and __vector_quad types.
>
> gcc/testsuite/
> * gcc.target/powerpc/mma-alignment.c: New test.
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/powerpc/mma-alignment.c
> @@ -0,0 +1,41 @@
> +/* { dg-do run } */
> +/* { dg-require-effective-target hard_float } */
> +/* { dg-options "-O2 -mhard-float" } */
> +
> +#include <stdlib.h>
> +
> +/* The MMA types below are enabled for pre-power10 compiles, because the
> + built-ins that use them must always be initialized in case the user has
> + a target attribute or pragma on a function that uses the MMA built-ins.
> + Since the test below doesn't need any other MMA support, we can enable
> + this test case on basically any cpu that has hard floating point
> + registers. */
Good comment, thanks :-) That "basically" does not sound super
convincing, but this is just a testcase, we can fix it later if need be.
Segher
