Another code generation bug in EGCS-1.0.1

Another code generation bug in EGCS-1.0.1

[Gary Thomas]

EGCS-1.0.1 (and GCC-2.8.0) can generate bad code when configured
for PowerPC and SYSV4-ABI (Linux/PPC).  In this configuration, the
stack is managed differently when varargs are present (a stack space
optimization).  However, for inlined functions that end up not
generated inline and nested functions, this optimization is lost and
the generated functions generate incorrect preamble code which will
destroy the runtime stack.

Code exists in the current version to try and address this for the nested
functions case, but it was a little wrong (patch to "function.c").  Also,
the case of inline functions was not present at all.

A patch is attached which solves both problems.

------------------------------------------------------------------------
Gary Thomas                              |
The Open Group / Research Institute      | "Fine wine is a necessity of
2 Avenue de Vignate                      |        life for me"
38610 Gieres - FRANCE                    |
+33 4 76 63 48 74                        |      Thomas Jefferson
email: g.thomas@opengroup.org            |
< http://www.opengroup.org/~gdt >          |
   ... opinions expressed here are mine  |
       and no one else would claim them! |
------------------------------------------------------------------------

Re: Another code generation bug in EGCS-1.0.1

[Jeffrey A Law]

  In message < XFMail.980208101002.g.thomas@opengroup.org >you write:
  > Content-Type: text/plain; charset=us-ascii
  > 
  > EGCS-1.0.1 (and GCC-2.8.0) can generate bad code when configured
  > for PowerPC and SYSV4-ABI (Linux/PPC).  In this configuration, the
  > stack is managed differently when varargs are present (a stack space
  > optimization).  However, for inlined functions that end up not
  > generated inline and nested functions, this optimization is lost and
  > the generated functions generate incorrect preamble code which will
  > destroy the runtime stack.
  > 
  > Code exists in the current version to try and address this for the nested
  > functions case, but it was a little wrong (patch to "function.c").  Also,
  > the case of inline functions was not present at all.
  > 
  > A patch is attached which solves both problems.
Can you please send a testcase which exposes this problem?

Thanks,
jeff

Re: Another code generation bug in EGCS-1.0.1

[Gary Thomas]

On 09-Feb-98 Jeffrey A Law wrote:
> 
>   In message < XFMail.980208101002.g.thomas@opengroup.org >you write:
>   > Content-Type: text/plain; charset=us-ascii
>   > 
>   > EGCS-1.0.1 (and GCC-2.8.0) can generate bad code when configured
>   > for PowerPC and SYSV4-ABI (Linux/PPC).  In this configuration, the
>   > stack is managed differently when varargs are present (a stack space
>   > optimization).  However, for inlined functions that end up not
>   > generated inline and nested functions, this optimization is lost and
>   > the generated functions generate incorrect preamble code which will
>   > destroy the runtime stack.
>   > 
>   > Code exists in the current version to try and address this for the nested
>   > functions case, but it was a little wrong (patch to "function.c").  Also,
>   > the case of inline functions was not present at all.
>   > 
>   > A patch is attached which solves both problems.
> Can you please send a testcase which exposes this problem?
> 
> Thanks,
> jeff

The crux of the bug is that on the RS6000/PowerPC, the stack has a
slightly different setup if there are varargs present or not. (This is
an optimization; trying to save stack space and some register saves if
there are no varargs).  Anyway, the code generator needs to have information
about the stack when the preamble is generated.  Given that for inline
functions this may be quite separated from the definition, that information
needs to be saved at declaration time and restored when the code for the
function is actually generated.  This is related to the problem of nested
functions where the inside and outside functions do not have the same
stack setup requirements.

There was already code in place to handle the nested functions (although it was
a little broken).  However, I couldn't tell how to use that functionality for
the inlined function case, so I just mimicked it.  If these can be combined, so
much the better. 

Seeing what is wrong can come from a very simple test case:

#include <stdio.h>
#include <stdarg.h>
inline void debug(const char *msg,...){
    va_list ap;
    va_start( ap, msg );                // use variable arg list
    vfprintf( stderr, msg, ap );
    va_end( ap );                    
}    

void fun(void)
{
  debug("%s.%d\n", __FUNCTION__, __LINE__);
}

The generated code is [excerpt]:


        .align 2
        .weak   debug__FPCce
        .type    debug__FPCce,@function
debug__FPCce:
        stwu 1,-32(1)
        mflr 0
        stw 0,36(1)
        stw 3,8(1)
        stw 4,12(1)
        stw 5,16(1)
        stw 6,20(1)
        stw 7,24(1)
        stw 8,28(1)
        stw 9,32(1)
        stw 10,36(1)
        bc 4,6,.L19
        stfd 1,40(1)
        stfd 2,48(1)
        stfd 3,56(1)
        stfd 4,64(1)
        stfd 5,72(1)
        stfd 6,80(1)
        stfd 7,88(1)
        stfd 8,96(1)
        stfd 9,104(1)

You'll notice that the inline function 'debug' only reserves 32 bytes
on the stack and then proceeds to save 112!  This is because the code
generator has lost the "context" for the function which told it that
there were varargs.

I hope that this clears this up.  Perhaps Mike Meissner will have 
something to say about this since he added the original code to handle
the nested function case.

Note: AFAIK this currently only affects the RS6000/PowerPC when using
the SYSV4 ABI (Linux/PPC).  However, there could be other platforms
with similar cases in the future.  I also noticed that there is a use
of the nested function (save_machine_state) for the i386 already.

------------------------------------------------------------------------
Gary Thomas                              |
The Open Group / Research Institute      | "Fine wine is a necessity of
2 Avenue de Vignate                      |        life for me"
38610 Gieres - FRANCE                    |
+33 4 76 63 48 74                        |      Thomas Jefferson
email: g.thomas@opengroup.org            |
< http://www.opengroup.org/~gdt >          |
   ... opinions expressed here are mine  |
       and no one else would claim them! |
------------------------------------------------------------------------