When set_mem_attributes_minus_bitpos computes the MEM_EXPR for MEM RTXen
it strips outermost ARRAY_REFs (for historic efficiency reasons).
Nowadays a MEM_EXPR has to be conservative with respect to alias
analysis which means it has to cover the original access. This means
we may _not_ end up with a MEM_EXPR accessing a trailing array as its
size is not conservative when used in overlap analysis.
The following patch fixes that (by dropping the MEM_EXPR). This
fix should be suitable for release branches and hopefully not
pessimize things too much.
In the end my idea always was to use the full memory reference tree
and not stripping anything but I never came to implementing this.
Bootstrap and regtest running on x86_64-unknown-linux-gnu.
No testcase, I failed to create one on x86_64 and the ppc one in
the PR is not an execute testcase so quite useless. I tried
struct q { int n; long o[100]; };
struct r { int n; long o[0]; };
union {
struct r r;
struct q q;
} u;
int foo (int i, int j)
{
long *q = u.r.o;
u.r.o[i/j] = 1;
return q[2];
}
but nothing convinced scheduling to move the load before the store ;)
The two memory references are seen as not aliasing though. Stupid
scheduler.
Richard.
2017-04-27 Richard Biener <rguent...@suse.de>
PR middle-end/80533
* emit-rtl.c (set_mem_attributes_minus_bitpos): When
stripping ARRAY_REFs from MEM_EXPR make sure we're not
keeping a reference to a trailing array.
Index: gcc/emit-rtl.c
===================================================================
--- gcc/emit-rtl.c (revision 247277)
+++ gcc/emit-rtl.c (working copy)
@@ -1954,7 +1954,10 @@ set_mem_attributes_minus_bitpos (rtx ref
while (TREE_CODE (t2) == ARRAY_REF);
if (DECL_P (t2)
- || TREE_CODE (t2) == COMPONENT_REF)
+ || (TREE_CODE (t2) == COMPONENT_REF
+ /* For trailing arrays t2 doesn't have a size that
+ covers all valid accesses. */
+ && ! array_at_struct_end_p (t, false)))
{
attrs.expr = t2;
attrs.offset_known_p = false;
