Bernd Edlinger
2014-10-07 09:34:29 UTC
Hi Honza,
as you know, we have a wrong code bug, when a pure or const method is called via a virtual thunk.
I had some more Ideas, how to fix that, but all of them had some serious draw-backs, so I leave the details out...
But now I have a new insight, why the "obvious" fix for this serious code generation bug did not work
in the first place.
And the reason was, that if ipa-pure-const.c calls set_const_flag or set_pure_flag for a thunk, it calls the same
function later for the called method, and this overwrites the flags of _all_ associated thunks and aliases.
However that should at least not be done for virtual thunks, as these need to be IPA_NEITHER, even if
the method itself has different attributes, that is because the assembler thunk accesses the vtable, while
other thunks do not.
So I re-factored set_const_flag and set_pure_flag to exclude the virtual thunks, taking care that other
users of call_for_symbol_thunks_and_aliases do not get a different behavior than before this patch.
The attached patch was boot-strapped and
regression-tested on x86_64-linux-gnu.
Ok for trunk?
PS: As a side-note, there are two identical functions, named "call_for_symbol_and_aliases", in
class symtab_node and in class cgraph_node, which inherits from symtab_node. Both functions are
not declared virtual. Is that what's intended? Usually this could lead to errors, or at least some serious
compiler warnings.
Thanks
Bernd.
as you know, we have a wrong code bug, when a pure or const method is called via a virtual thunk.
I had some more Ideas, how to fix that, but all of them had some serious draw-backs, so I leave the details out...
But now I have a new insight, why the "obvious" fix for this serious code generation bug did not work
in the first place.
And the reason was, that if ipa-pure-const.c calls set_const_flag or set_pure_flag for a thunk, it calls the same
function later for the called method, and this overwrites the flags of _all_ associated thunks and aliases.
However that should at least not be done for virtual thunks, as these need to be IPA_NEITHER, even if
the method itself has different attributes, that is because the assembler thunk accesses the vtable, while
other thunks do not.
So I re-factored set_const_flag and set_pure_flag to exclude the virtual thunks, taking care that other
users of call_for_symbol_thunks_and_aliases do not get a different behavior than before this patch.
The attached patch was boot-strapped and
regression-tested on x86_64-linux-gnu.
Ok for trunk?
PS: As a side-note, there are two identical functions, named "call_for_symbol_and_aliases", in
class symtab_node and in class cgraph_node, which inherits from symtab_node. Both functions are
not declared virtual. Is that what's intended? Usually this could lead to errors, or at least some serious
compiler warnings.
Thanks
Bernd.