As noted in the PR, expand_atomic_compare_and_swap can generate wrong code when
libcalls are emitted
for the sync_compare_and_swap and the result comparison test. This is fixed by
emitting a move insn to copy
the result rtx of the sync_compare_and_swap libcall to target_oval instead of
directly assigning it.
Tested on hppa-unknown-linux-gnu. Okay for trunk and gcc-5?
With this change I can enable __sync_compare_and_swap_8 on hppa-linux.
Dave
--
John David Anglin dave.ang...@bell.net
2015-09-10 John David Anglin <dang...@gcc.gnu.org>
PR middle-end/67401
* optabs.c (expand_atomic_compare_and_swap): Move result of emitting
sync_compare_and_swap_optab libcall to target_oval.
Index: optabs.c
===================================================================
--- optabs.c (revision 226978)
+++ optabs.c (working copy)
@@ -7510,9 +7510,10 @@
if (libfunc != NULL)
{
rtx addr = convert_memory_address (ptr_mode, XEXP (mem, 0));
- target_oval = emit_library_call_value (libfunc, NULL_RTX, LCT_NORMAL,
- mode, 3, addr, ptr_mode,
- expected, mode, desired, mode);
+ rtx target = emit_library_call_value (libfunc, NULL_RTX, LCT_NORMAL,
+ mode, 3, addr, ptr_mode,
+ expected, mode, desired, mode);
+ emit_move_insn (target_oval, target);
/* Compute the boolean return value only if requested. */
if (ptarget_bool)
