Andy Wingo in IRC pointed out that the reason the patch appears to work
is that the `movr` isn't idempotent! By the time it comes round again,
`val` has already been overwritten by LDAXR in the case that `dst == val`.
On 22/04/2024 10:18, Tony Garnock-Jones wrote:
Oh man. This little patch all by itself makes the problem behaviour go
away. No switching to SWPAL/CASAL, just tightening the spinloop. (And no
changes at all to the CAS code, so nothing to do with the fibers bug I
guess.)
With the patch, the spinloop goes LDAXR-STLXR-CBNZ (which is what GCC
does when SWPAL isn't there) instead of potentially MOV-LDAXR-STLXR-CBNZ
(which isn't).
Could the machine really be so sensitive to the target of the CBNZ?
---
libguile/lightening/lightening/aarch64-cpu.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/libguile/lightening/lightening/aarch64-cpu.c
b/libguile/lightening/lightening/aarch64-cpu.c
index 13aa351e9..4df712a0e 100644
--- a/libguile/lightening/lightening/aarch64-cpu.c
+++ b/libguile/lightening/lightening/aarch64-cpu.c
@@ -2532,10 +2532,10 @@ str_atomic(jit_state_t *_jit, int32_t loc,
int32_t val)
static void
swap_atomic(jit_state_t *_jit, int32_t dst, int32_t loc, int32_t val)
{
- void *retry = jit_address(_jit);
int32_t result = jit_gpr_regno(get_temp_gpr(_jit));
int32_t val_or_tmp = dst == val ? jit_gpr_regno(get_temp_gpr(_jit))
: val;
movr(_jit, val_or_tmp, val);
+ void *retry = jit_address(_jit);
LDAXR(_jit, dst, loc);
STLXR(_jit, val_or_tmp, loc, result);
jit_patch_there(_jit, bnei(_jit, result, 0), retry);
