[RFC][nvptx, libgomp] Add 128-bit atomic support

[Tom de Vries]

[ was: Re: [patch][nvptx] libgomp: Split testcase in order to XFAIL
__sync_val_compare_and_swap_16 (was: [PATCH] nvptx: Add support for
subword compare-and-swap) ]

On 9/2/20 9:56 AM, Tom de Vries wrote:
> On 9/1/20 2:58 PM, Tom de Vries wrote:
>> On 9/1/20 1:41 PM, Tobias Burnus wrote:
>>> Hi Tom, hello all,
>>>
>>> it turned out that the testcase fails on PowerPC (but not x86_64)
>>> as the nvptx lto complains: unresolved symbol
>>> __sync_val_compare_and_swap_16
>>>
>>> The testcase uses int128 – and that's the culprit, but I have no idea
>>> why it only fails with PowerPC and not with x86-64.
>>>
>>
> 
> Reproduced on x86_64 using trigger patch:
> ...
> $ git diff
> diff --git a/gcc/config/i386/sync.md b/gcc/config/i386/sync.md
> index ed17bb00205..eccedac192f 100644
> --- a/gcc/config/i386/sync.md
> +++ b/gcc/config/i386/sync.md
> @@ -153,9 +153,15 @@
>      (DI "TARGET_64BIT || (TARGET_CMPXCHG8B && (TARGET_80387 ||
> TARGET_SSE))")
>     ])
> 
> + (define_mode_iterator ATOMIC2
> +    [QI HI SI
> +     (DI "TARGET_64BIT || (TARGET_CMPXCHG8B && (TARGET_80387 ||
> TARGET_SSE))")
> +    TI
> +    ])
> +
>  (define_expand "atomic_load<mode>"
> -  [(set (match_operand:ATOMIC 0 "nonimmediate_operand")
> -       (unspec:ATOMIC [(match_operand:ATOMIC 1 "memory_operand")
> +  [(set (match_operand:ATOMIC2 0 "nonimmediate_operand")
> +       (unspec:ATOMIC2 [(match_operand:ATOMIC2 1 "memory_operand")
>                         (match_operand:SI 2 "const_int_operand")]
>                        UNSPEC_LDA))]
>    ""
> diff --git a/libgomp/testsuite/libgomp.c-c++-common/reduction-16.c
> b/libgomp/testsuite/libgomp.c-c++-common/reduction-16.c
> index d0e82b04790..62b0e032c33 100644
> --- a/libgomp/testsuite/libgomp.c-c++-common/reduction-16.c
> +++ b/libgomp/testsuite/libgomp.c-c++-common/reduction-16.c
> @@ -1,4 +1,5 @@
>  /* { dg-do run } */
> +/* { dg-additional-options "-mcx16" } */
> 
>  #include <stdlib.h>
> 
> ...
> 

And test-case passes on x86_64 with this patch (obviously, in
combination with trigger patch above).

Jakub, WDYT?

Tobias, can you try on powerpc?

Thanks,
- Tom


[nvptx, libgomp] Add 128-bit atomic support

---
 libgomp/config/nvptx/atomic.c | 34 ++++++++++++++++++++++++++++++++++
 1 file changed, 34 insertions(+)

diff --git a/libgomp/config/nvptx/atomic.c b/libgomp/config/nvptx/atomic.c
new file mode 100644
index 00000000000..49a6d350827
--- /dev/null
+++ b/libgomp/config/nvptx/atomic.c
@@ -0,0 +1,34 @@
+#include <stdbool.h>
+
+#include "../../atomic.c"
+
+unsigned __int128
+__sync_val_compare_and_swap_16 (volatile void *vptr, unsigned __int128 oldval,
+				unsigned __int128 newval)
+{
+  volatile unsigned __int128 *ptr = vptr;
+  GOMP_atomic_start ();
+  unsigned __int128 val = *ptr;
+  if (val == oldval)
+    *ptr = newval;
+  GOMP_atomic_end ();
+  return val;
+}
+
+bool
+__sync_bool_compare_and_swap_16 (volatile void *vptr, unsigned __int128 oldval,
+				 unsigned __int128 newval)
+{
+  return __sync_val_compare_and_swap_16 (vptr, oldval, newval) == oldval;
+}
+
+unsigned __int128
+__atomic_load_16 (const volatile void *vptr,
+		  int memorder __attribute__((unused)))
+{
+  const volatile unsigned __int128 *ptr = vptr;
+  GOMP_atomic_start ();
+  unsigned __int128 val = *ptr;
+  GOMP_atomic_end ();
+  return val;
+}