[Bug target/41246] should "sorry" when regparm=3 and nested functions are encountered

[phcoder at gmail dot com]

------- Comment #13 from phcoder at gmail dot com  2009-09-05 20:34 -------
(In reply to comment #2)
> Created an attachment (id=18480)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=18480&action=view) [edit]
> gcc45-pr41246.patch
> 
> I'm afraid there is nothing that can make that unmodified testcase work for
> -mregparm=3.
> It is true that gcc is buggy and doesn't force regparm(2) for nested functions
> with -mregparm=3 (regparm(3) is impossible, due to the static chain in %ecx),
Could this chain perhaps be moved to %edx?
> we only error if explicit __attribute__((regparm (3))) is used on nested
> functions.
> But if we silently change the call to regparm(2), there is no way for the
> compiler to know whether the hook is using regparm(3) or regparm(2) calling
> convention 
If you pass a pointer to nested function you actually pass a pointer to a
trampoline. Could trampoline just push %ecx as its first operation? this way
trampoline would "convert" regparm(2) to regparm(3) function
(if you pass it a non-nested function, it will be regparm(3), if you
> pass a nested function, it will be regparm(2).  So, perhaps we should just
> error out whenever we see in -m32 mode on i?86/x86_64 a nested function in
> -mregparm=3 mode.
> 


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=41246