ipa: Adjust references to identify read-only globals

this patch has been motivated by SPEC 2017's 544.nab_r in which there is
a static variable which is never written to and so zero throughout the
run-time of the benchmark.  However, it is passed by reference to a
function in which it is read and (after some multiplications) passed
into __builtin_exp which in turn unnecessarily consumes almost 10% of
the total benchmark run-time.  The situation is illustrated by the added
testcase remref-3.c.

The patch adds a flag to ipa-prop descriptor of each parameter to mark
such parameters.  IPA-CP and inling then take the effort to remove
IPA_REF_ADDR references in the caller and only add IPA_REF_LOAD
reference to the clone/overall inlined function.  This is sufficient
for subsequent symbol table analysis code to identify the read-only
variable as such and optimize the code.

There are two changes from the RFC version posted to the list earlier.
First, three missing calls to get_base_address were added (there was
another one in an assert).  Second, references are not stripped off
the callers if the cloned function cannot change the signature.  The
second change reveals a real shortcoming stemming from the fact we
cannot adjust function prototypes with fnspecs.  But that is a more
general problem.

gcc/ChangeLog:

2021-07-20  Martin Jambor  <mjambor@suse.cz>

	* cgraph.h (ipa_replace_map): New field force_load_ref.
	* ipa-prop.h (ipa_param_descriptor): Reduce precision of move_cost,
	aded new flag load_dereferenced, adjusted comments.
	(ipa_get_param_dereferenced): New function.
	(ipa_set_param_dereferenced): Likewise.
	* cgraphclones.c (cgraph_node::create_virtual_clone): Follow it.
	* ipa-cp.c: Include gimple.h.
	(ipcp_discover_new_direct_edges): Take into account dereferenced flag.
	(get_replacement_map): New parameter force_load_ref, set the
	appropriate flag in ipa_replace_map if set.
	(struct symbol_and_index_together): New type.
	(adjust_refs_in_act_callers): New function.
	(adjust_references_in_caller): Likewise.
	(create_specialized_node): When appropriate, call
	adjust_references_in_caller and force only load references.
	* ipa-prop.c (load_from_dereferenced_name): New function.
	(ipa_analyze_controlled_uses): Also detect loads from a
	dereference, harden testing of call statements.
	(ipa_write_node_info): Stream the dereferenced flag.
	(ipa_read_node_info): Likewise.
	(ipa_set_jf_constant): Also create refdesc when jump function
	references a variable.
	(cgraph_node_for_jfunc): Rename to symtab_node_for_jfunc, work
	also on references of variables and return a symtab_node.  Adjust
	all callers.
	(propagate_controlled_uses): Also remove references to VAR_DECLs.

gcc/testsuite/ChangeLog:

2021-06-29  Martin Jambor  <mjambor@suse.cz>

	* gcc.dg/ipa/remref-3.c: New test.
	* gcc.dg/ipa/remref-4.c: Likewise.
	* gcc.dg/ipa/remref-5.c: Likewise.
	* gcc.dg/ipa/remref-6.c: Likewise.

1 files changed, 9 insertions, 1 deletions

diff --git a/gcc/cgraphclones.c b/gcc/cgraphclones.c
index fef4a23..b16e681 100644
--- a/gcc/cgraphclones.c
+++ b/gcc/cgraphclones.c
@@ -633,7 +633,15 @@ cgraph_node::create_virtual_clone (const vec<cgraph_edge *> &redirect_callers,
       || in_lto_p)
     new_node->unique_name = true;
   FOR_EACH_VEC_SAFE_ELT (tree_map, i, map)
-    new_node->maybe_create_reference (map->new_tree, NULL);
+    {
+      tree repl = map->new_tree;
+      if (map->force_load_ref)
+	{
+	  gcc_assert (TREE_CODE (repl) == ADDR_EXPR);
+	  repl = get_base_address (TREE_OPERAND (repl, 0));
+	}
+      new_node->maybe_create_reference (repl, NULL);
+    }
 
   if (ipa_transforms_to_apply.exists ())
     new_node->ipa_transforms_to_apply