https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94734
--- Comment #11 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
If we don't want to revert the change completely, could we perhaps do:
--- gcc/tree-ssa-phiopt.c.jj 2020-03-19 10:23:50.542872359 +0100
+++ gcc/tree-ssa-phiopt.c 2020-04-24 10:54:10.341716841 +0200
@@ -2237,10 +2237,26 @@ cond_store_replacement (basic_block midd
whose value is not available readily, which we want to avoid. */
if (!nontrap->contains (lhs))
{
- /* If LHS is a local variable without address-taken, we could
- always safely move down the store. */
- tree base = get_base_address (lhs);
- if (!auto_var_p (base) || TREE_ADDRESSABLE (base))
+ /* If LHS is an access to a local variable without address-taken
+ or its part and the access is provably within the bounds of the
+ local variable, we could always safely move down the store. */
+ HOST_WIDE_INT offset, size, decl_size;
+ bool reverse;
+ tree base = get_ref_base_and_extent_hwi (lhs, &offset, &size,
+ &reverse);
+ if (base == NULL_TREE || !auto_var_p (base) || TREE_ADDRESSABLE (base))
+ return false;
+ if (!DECL_SIZE (base)
+ || !tree_fits_shwi_p (DECL_SIZE (base)))
+ return false;
+ decl_size = tree_to_shwi (DECL_SIZE (base));
+ if (offset < 0
+ || size < 0
+ || decl_size < 0
+ || offset >= decl_size
+ || size > decl_size
+ || ((unsigned HOST_WIDE_INT) offset + size
+ > (unsigned HOST_WIDE_INT) decl_size))
return false;
}
+ xfail the tests from the PR89430 because they all need the nontrap ARRAY_REF
etc. handling?
With that, in the above testcase baz is still cselim optimized , but foo and
bar are not. bar is still miscompiled by some other optimization though (and
GCC 9 didn't do that), so we have some other regression.
