The following testcase ICEs, because aarch64_gen_compare_reg_maybe_ze emits
invalid RTL.
For y_mode [QH]Imode it expects y to be of that mode (or CONST_INT that fits
into that mode) and x being SImode; for non-CONST_INT y it zero extends y
into SImode and compares that against x, for CONST_INT y it zero extends y
into SImode. The problem is that when the zero extended constant isn't
usable directly, it forces it into a REG, but with y_mode mode, and then
compares against y. That is wrong, because it should force it into a SImode
REG and compare that way.
2020-04-02 Jakub Jelinek <jakub@redhat.com>
PR target/94435
* config/aarch64/aarch64.c (aarch64_gen_compare_reg_maybe_ze): For
y_mode E_[QH]Imode and y being a CONST_INT, change y_mode to SImode.
* gcc.target/aarch64/pr94435.c: New test.
if (y_mode == E_QImode || y_mode == E_HImode)
{
if (CONST_INT_P (y))
- y = GEN_INT (INTVAL (y) & GET_MODE_MASK (y_mode));
+ {
+ y = GEN_INT (INTVAL (y) & GET_MODE_MASK (y_mode));
+ y_mode = SImode;
+ }
else
{
rtx t, cc_reg;
--- /dev/null
+/* PR target/94435 */
+/* { dg-do compile } */
+/* { dg-options "-march=armv8-a+nolse -moutline-atomics" } */
+
+int b, c, d, e, f, h;
+short g;
+int foo (int) __attribute__ ((__const__));
+
+void
+bar (void)
+{
+ while (1)
+ {
+ while (1)
+ {
+ __atomic_load_n (&e, 0);
+ if (foo (2))
+ __sync_val_compare_and_swap (&c, 0, f);
+ b = 1;
+ if (h == e)
+ break;
+ }
+ __sync_val_compare_and_swap (&g, -1, f);
+ }
+}