> This is the "real" fix for PR46076 that I wanted to persue. Make
> function pointer type conversions useless as to more aggressively
> be able to turn indirect into direct calls. This requires that we
> preserve the original type signature of the called function as
> presented by the frontend. The patch does that by adding a fntype
> field to every call stmt in GIMPLE and extract this information
> during gimplification.
The patch is incomplete though: the type is preserved at the GIMPLE level but
is dropped at the RTL level, so you can get call convention mismatches. The
attached patch is needed to cure the 3 ACATS failures on x86:
FAIL: c431001
FAIL: c731001
FAIL: ca11c02
I think the GIMPLE->Tree->RTL interface would need to be audited here.
* cfgexpand.c (expand_call_stmt): Rematerialize original function type.
--
Eric Botcazou
Index: cfgexpand.c
===================================================================
--- cfgexpand.c (revision 172371)
+++ cfgexpand.c (working copy)
@@ -1844,7 +1844,10 @@ expand_call_stmt (gimple stmt)
exp = build_vl_exp (CALL_EXPR, gimple_call_num_args (stmt) + 3);
- CALL_EXPR_FN (exp) = gimple_call_fn (stmt);
+ CALL_EXPR_FN (exp)
+ = fold_build1 (NOP_EXPR, build_pointer_type (gimple_call_fntype (stmt)),
+ gimple_call_fn (stmt));
+
decl = gimple_call_fndecl (stmt);
builtin_p = decl && DECL_BUILT_IN (decl);
