------- Comment #8 from jakub at gcc dot gnu dot org 2007-05-31 10:00 ------- I saw was_readonly as well and wondered if it might be a problem or not, then incorrectly assumed REFERENCE_TYPEs can't be an issue, but as PR31806 shows, they apparently are.
2007-05-31 Jakub Jelinek <[EMAIL PROTECTED]> PR c++/31806 * decl.c (cp_finish_decl): Also clear was_readonly if a static var needs runtime initialization. --- gcc/cp/decl.c.jj 2007-05-30 21:03:07.000000000 +0200 +++ gcc/cp/decl.c 2007-05-31 11:32:48.000000000 +0200 @@ -5361,8 +5361,12 @@ cp_finish_decl (tree decl, tree init, bo /* If a TREE_READONLY variable needs initialization at runtime, it is no longer readonly and we need to avoid MEM_READONLY_P being set on RTL created for it. */ - if (init && TREE_READONLY (decl)) - TREE_READONLY (decl) = 0; + if (init) + { + if (TREE_READONLY (decl)) + TREE_READONLY (decl) = 0; + was_readonly = 0; + } expand_static_init (decl, init); } } should cure this. Regarding the performance degradation, to cure this, we'd either have to revert to RTX_UNCHANGING_P (which we know just leads to tremendous amount of bugs), or introduce a way to say "this tree resp. this rtl is readonly in all functions but <this set of fndecls>". For filescope C++ vars that need runtime the set is certainly static initialization and destruction, all functions inlined into it and then depends on how aggresive IPA optimizations want to be, if one day they choose to clone some constructor of a initialized once, then read-only variable and the optimizers would assume that var is readonly, we'd again miscompile things. For function-scope static vars we'd need to put the containing function into the set, so it probably won't help there at all. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31809