On Wed, Feb 01, 2017 at 06:26:04PM +0100, Torvald Riegel wrote:
> On Mon, 2017-01-30 at 19:54 +0100, Torvald Riegel wrote:
> > This patch fixes the __atomic builtins to not implement supposedly
> > lock-free atomic loads based on just a compare-and-swap operation.
>
> After an off-list OK by Jakub, I have committed this as r245098.
> Jakub will take care of the OpenMP side in a follow-up patch.
Here it is. It is an ABI change for -fopenmp -mcx16 on x86_64, but -mcx16
used to be ABI incompatible with -mno-cx16, and it affects mostly just
__int128 and long double atomics.
Bootstrapped/regtested on x86_64-linux and i686-linux, committed to trunk.
2017-02-06 Jakub Jelinek <ja...@redhat.com>
* omp-expand.c (oxpand_omp_atomic_fetch_op,
expand_omp_atomic_pipeline): Return false if can_atomic_load_p is
false.
--- gcc/omp-expand.c.jj 2017-01-25 17:17:52.000000000 +0100
+++ gcc/omp-expand.c 2017-02-06 13:41:22.703372812 +0100
@@ -6241,7 +6241,7 @@ expand_omp_atomic_fetch_op (basic_block
matter is that (with the exception of i486 vs i586 and xadd) all targets
that support any atomic operaton optab also implements compare-and-swap.
Let optabs.c take care of expanding any compare-and-swap loop. */
- if (!can_compare_and_swap_p (imode, true))
+ if (!can_compare_and_swap_p (imode, true) || !can_atomic_load_p (imode))
return false;
gsi = gsi_last_bb (load_bb);
@@ -6318,7 +6318,8 @@ expand_omp_atomic_pipeline (basic_block
type = TYPE_MAIN_VARIANT (TREE_TYPE (TREE_TYPE (addr)));
itype = TREE_TYPE (TREE_TYPE (cmpxchg));
- if (!can_compare_and_swap_p (TYPE_MODE (itype), true))
+ if (!can_compare_and_swap_p (TYPE_MODE (itype), true)
+ || !can_atomic_load_p (TYPE_MODE (itype)))
return false;
/* Load the initial value, replacing the GIMPLE_OMP_ATOMIC_LOAD. */
Jakub
