As mentioned in the PR, the fallback path when LSE is unavailable writes
incorrect registers to the memory if the previous content compares equal
to x0, x1 - it writes copy of x0, x1 from the start of function, but it
should write x2, x3.
2020-08-03 Jakub Jelinek <jakub@redhat.com>
PR target/96402
* config/aarch64/lse.S (__aarch64_cas16_acq_rel): Use x2, x3 instead
of x(tmp0), x(tmp1) in STXP arguments.
* gcc.target/aarch64/pr96402.c: New test.
--- /dev/null
+/* PR target/96402 */
+/* { dg-do run { target int128 } } */
+/* { dg-options "-moutline-atomics" } */
+
+int
+main ()
+{
+ __int128 a = 0;
+ __sync_val_compare_and_swap (&a, (__int128) 0, (__int128) 1);
+ if (a != 1)
+ __builtin_abort ();
+ __sync_val_compare_and_swap (&a, (__int128) 1, (((__int128) 0xdeadbeeffeedbac1ULL) << 64) | 0xabadcafe00c0ffeeULL);
+ if (a != ((((__int128) 0xdeadbeeffeedbac1ULL) << 64) | 0xabadcafe00c0ffeeULL))
+ __builtin_abort ();
+ return 0;
+}
cmp x0, x(tmp0)
ccmp x1, x(tmp1), #0, eq
bne 1f
- STXP w(tmp2), x(tmp0), x(tmp1), [x4]
+ STXP w(tmp2), x2, x3, [x4]
cbnz w(tmp2), 0b
1: ret