The following testcase ICEs starting with r15-3213 in
decompose_normal_address and starting with r15-3288 ICEs
in lra_rtx_hash, which since r8-5466 can't handle SUBREG
(previously SUBREG was "ei" and lra_rtx_hash can handle
that through
val += lra_rtx_hash (XEXP (x, i));
for e and
val += XINT (x, i);
for i, now it is "ep" where p stands for poly_uint16).
The following patch fixes it by handling SUBREG directly, a variant
could be instead add
case 'p':
for (int i = 0; i < NUM_POLY_INT_COEFFS; ++i)
val += SUBREG_BYTE (x).coeffs[i];
break;
if you prefer that more (p is used solely for SUBREG and e.g. rtx_equal_p
has
case 'p':
if (maybe_ne (SUBREG_BYTE (x), SUBREG_BYTE (y)))
return false;
break;
). Given the above rtx_equal_p snippet and that lra_rtx_hash
is solely used in invariant_hash (and recursion) and invariant_eq_p
uses rtx_equal_p we'll never consider different SUBREGs of the same thing
as the same invariant.
2025-03-18 Jakub Jelinek <jakub@redhat.com>
PR rtl-optimization/119307
* lra.cc (lra_rtx_hash): Handle SUBREG.
* gcc.target/i386/pr119307.c: New test.
case CONST_INT:
return val + UINTVAL (x);
+ case SUBREG:
+ val += lra_rtx_hash (SUBREG_REG (x));
+ for (int i = 0; i < NUM_POLY_INT_COEFFS; ++i)
+ val += SUBREG_BYTE (x).coeffs[i];
+ return val;
+
default:
break;
}
--- /dev/null
+/* PR rtl-optimization/119307 */
+/* { dg-do compile { target x32 } } */
+/* { dg-require-profiling "-fprofile-generate" } */
+/* { dg-options "-Os -maddress-mode=long -fprofile-generate -ftrapv" } */
+
+_Complex int x;
+__int128 y;
+long n;
+
+void
+foo (void)
+{
+ x *= *(__int128 *) __builtin_memmove (&y, &x, n);
+}