On 12/17/13 13:45, Jan Hubicka wrote:
Hi,
this patch contines on tract of making us to handle OBJ_TYPE_REF in expressions
more gracefully. The change in gimple_fold_stmt_to_constant_1 makes
it to skip the wrapper, so valuelize functions of passes never see it,
and in addition OBJ_TYPE_REF is stripped once argument becomes constant.
In gimple-fold we also ask type devirtualization machinery for call target the
same way as we do for call statements. I think this is rather important, since
still we resolve good part of devirtualization only after inlining. Often we
however manage to produce a speculative edge and we really do not want to keep
the speculation in code after we found real target.
Bootstrapped/regtested x86_64-linux, OK?
Honza
* gimple-fold.c (fold_gimple_assign): Attempt to devirtualize
OBJ_TYPE_REF.
(gimple_fold_stmt_to_constant_1): Bypass OBJ_TYPE_REF wrappers.
Index: gimple-fold.c
===================================================================
--- gimple-fold.c (revision 206042)
+++ gimple-fold.c (working copy)
@@ -374,6 +375,30 @@ fold_gimple_assign (gimple_stmt_iterator
if (REFERENCE_CLASS_P (rhs))
return maybe_fold_reference (rhs, false);
+ else if (TREE_CODE (rhs) == OBJ_TYPE_REF)
+ {
+ tree val = OBJ_TYPE_REF_EXPR (rhs);
+ if (is_gimple_min_invariant (val))
+ return val;
+ else if (flag_devirtualize && virtual_method_call_p (val))
+ {
+ bool final;
+ vec <cgraph_node *>targets
+ = possible_polymorphic_call_targets (val, &final);
+ if (final && targets.length () <= 1)
+ {
+ tree fndecl;
+ if (targets.length () == 1)
+ fndecl = targets[0]->decl;
+ else
+ fndecl = builtin_decl_implicit (BUILT_IN_UNREACHABLE);
+ val = fold_convert (TREE_TYPE (val), fndecl);
+ STRIP_USELESS_TYPE_CONVERSION (val);
+ return val;
+ }
+ }
+
+ }
else if (TREE_CODE (rhs) == ADDR_EXPR)
{
tree ref = TREE_OPERAND (rhs, 0);
@@ -2525,6 +2550,13 @@ gimple_fold_stmt_to_constant_1 (gimple s
return build_vector (TREE_TYPE (rhs), vec);
}
+ if (subcode == OBJ_TYPE_REF)
+ {
+ tree val = (*valueize) (OBJ_TYPE_REF_EXPR (rhs));
+ /* If calle is constant, we can fold away the wrapper. */
s/calle/callee/?
OK with that fix. Please add reference to the right (35545) PR number
in the ChangeLog/commit message.
jeff
