The asm insert explicitly set the EXEC register, but that's illegal in SIMT
mode. The new code does the same, but is SIMT safe. The EXEC set is still
needed for SIMD mode.
libgomp/ChangeLog:
* libgomp.h (team_malloc): Add SIMT variant of the asm.
---
libgomp/libgomp.h | 8 ++++++++
1 file changed, 8 insertions(+)
diff --git a/libgomp/libgomp.h b/libgomp/libgomp.h
index c4bab109736..1c25f09277f 100644
--- a/libgomp/libgomp.h
+++ b/libgomp/libgomp.h
@@ -131,8 +131,16 @@ team_malloc (size_t size)
/* Allocate directly from the arena.
The compiler does not support DS atomics, yet. */
void *result;
+#ifndef __SIMT__
asm ("ds_add_rtn_u64 %0, %1, %2\n\ts_waitcnt 0"
: "=v"(result) : "v"(TEAM_ARENA_FREE), "v"(size), "e"(1L) : "memory");
+#else
+ /* SIMT should not touch EXEC.
+ FIXME: this will allocate SIZE bytes for each active thread... is that
+ what we want to happen? Usually, only one will be active here. */
+ asm ("ds_add_rtn_u64 %0, %1, %2\n\ts_waitcnt 0"
+ : "=v"(result) : "v"(TEAM_ARENA_FREE), "v"(size) : "memory");
+#endif
/* Handle OOM. */
if (result + size > *(void * __lds *)TEAM_ARENA_END)
--
2.54.0
