Message-Id: <199604181352.JAA16827@delorie.com>
Date:         Thu, 18 Apr 96 16:36:28 LIT
From: Martynas Kunigelis <martynas DOT kunigelis AT VM DOT KTU DOT LT>
Subject:      GCC bug
To: DJGPP mailing list <djgpp AT delorie DOT com>

 I know this is not DJGPP related, but seems like I am not able to report
bugs directly to FSF, since my messages come back with a weird note, kind
of "unknown mailer". Well, guess that's what I have to pay for using VMS
these days. Anyway..

I found that if I declare a function like this:

void foo(int a, int b, int c, int d) __attribute__ ((regparm(3), stdcall));

then the `stdcall' attribute overrides `regparm', i.e. the function returns
with an `rtd' instruction, but parameters are passed through stack as usual.
If I swap the attributes in the declaration [ ((stdcall, regparm(3))) ],
`regparm' overcomes `stdcall', i.e. a, b, and c are passed in registers. but
the function returns with a usual `ret' instruction. However, if I compile
with -mregparm=3 and -mrtd options, everything's OK, i.e. a, b and c are
passed in registers and the function returns with an `rtd $4' instruction.
I think those two attributes combioned with the -fomit-frame-pointer option
could result in *very* optimal code, thus GCC behaviour with __attribute__
missatisfies me. I consider this a bug, because if it wasn't, the stuff
would not work with cmd. line options either.
  Now, can anyone PLEASE report this to bug-gcc AT prep DOT ai DOT mit DOT edu for me [and
all of us]?? I really have no way to do this..

P.S. I think it is a *very* good idea to add some new option to gcc by
editing the `specs' file which, once specified, would involve:

-mregparm=3
-mrtd
-fomit-frame-pointer
-Wimplicit
-D__FASTCALL__

and also appropriate libs and startup files.

The C library [and others] could be recompiled with the new calling convention
( asm parts should check for the __FASTCALL__ macro). The code resulting from
this would BLOW *any* other known compilers in terms of speed. Good idea, eh?.
Charlie?


Martynas