https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111901
--- Comment #14 from Richard Sandiford <rsandifo at gcc dot gnu.org> ---
ISTM that the problem is that cselib doesn't consider:
(clobber (mem:BLK (scratch)))
to be a read from memory (unlike note_uses, which gets this right). cselib
instead assumes that the SET_SRC of the asm is self-contained, and doesn't
depend on anything outside the SET_SRC.
This seems like a representational issue. It would be better if the SET_SRC
contained (mem:BLK (scratch)), which is what I'd expect an md insn to do.
So as it stands, I think the equivalence recorded by cselib_record_sets is
simply bogus: the instruction itself invalidates the equivalence that is being
recorded. All the other breakage seems to be a downstream effect of that.
One hacky fix would be:
diff --git a/gcc/cselib.cc b/gcc/cselib.cc
index 7f1991b09c3..673fda30f3a 100644
--- a/gcc/cselib.cc
+++ b/gcc/cselib.cc
@@ -3024,6 +3024,7 @@ cselib_record_sets (rtx_insn *insn)
int n_sets_before_autoinc;
int n_strict_low_parts = 0;
struct cselib_record_autoinc_data data;
+ bool has_mem_clobber = false;
rtx body = PATTERN (insn);
if (GET_CODE (body) == COND_EXEC)
@@ -3053,9 +3054,14 @@ cselib_record_sets (rtx_insn *insn)
sets[n_sets].dest = SET_DEST (x);
n_sets++;
}
+ else if (GET_CODE (x) == CLOBBER && MEM_P (XEXP (x, 0)))
+ has_mem_clobber = true;
}
}
+ if (has_mem_clobber)
+ n_sets = 0;
+
if (n_sets == 1
&& MEM_P (sets[0].src)
&& !cselib_record_memory
but that doesn't feel like the best approach…
