https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83129
Richard Biener changed:
What|Removed |Added
Status|UNCONFIRMED |ASSIGNED
Last reconfirmed||2017-11-23
Version|unknown |8.0
Assignee|unassigned at gcc dot gnu.org |rguenth at gcc dot
gnu.org
Ever confirmed|0 |1
--- Comment #1 from Richard Biener ---
Confirmed - mine.
This one is interesting because alias-wise we consider allocation functions
not clobbering anything -- dependences for the pointer will prevent any
invalid transform.
That of course means we'll never visit it via vn_reference_lookup_3.
One trick we don't pull is that if there's a must-alloc in the path we can
stop walking. Another trick we don't pull is that if we don't find any
may-def the memory we load is uninitialized (or in this case zero-initialized
as we don't find the calloc as may-def either - an issue for calloc and
-Wuninit maybe). So we could in theory optimistically value-number each
of those to zero (conservatively to treat calloc "right").
Need to think about that -Wuninitialized case. Ah, we only ever warn
using the oracle walking for decls, so never for allocated memory.
So in theory the following but it doesn't trigger because of the above.
Index: gcc/tree-ssa-sccvn.c
===
--- gcc/tree-ssa-sccvn.c(revision 255093)
+++ gcc/tree-ssa-sccvn.c(working copy)
@@ -2412,6 +2417,21 @@ vn_reference_lookup_3 (ao_ref *ref, tree
return NULL;
}
+ /* 7) calloc, optimize ref to all zeros. */
+ else if (is_gimple_reg_type (vr->type)
+ && gimple_call_builtin_p (def_stmt, BUILT_IN_CALLOC)
+ && TREE_CODE (base) == MEM_REF
+ && TREE_CODE (TREE_OPERAND (base, 0)) == SSA_NAME
+ && SSA_VAL (TREE_OPERAND (base, 0)) != VN_TOP
+ && SSA_VAL (TREE_OPERAND (base, 0)) == gimple_call_lhs (def_stmt))
+{
+ /* We do not bother to check whether the ref is in range because
+ we'll usually not be able to anyway. */
+ tree val = build_zero_cst (vr->type);
+ return vn_reference_lookup_or_insert_for_pieces
+ (vuse, vr->set, vr->type, vr->operands, val);
+}
+
/* Bail out and stop walking. */
return (void *)-1;
}
