From: pavenis AT lanet DOT lv
Message-ID: <B0000101116@stargate.astr.lu.lv>
To: "Rob Kramer" <robk AT cyberway DOT com DOT sg>, djgpp-workers AT delorie DOT com
Date: Fri, 10 Sep 1999 11:04:27 +0300
MIME-Version: 1.0
Content-type: text/plain; charset=US-ASCII
Content-transfer-encoding: 7BIT
Subject: Bug in CPP_CPU_SPECS for i[3456]86 [Was: Re: GCC 2.95 global string crashes on 386..]
CC: gcc-bugs AT gcc DOT gnu DOT org
In-reply-to: <199909100328.LAA11335@westgate.cyberway.com.sg>
References: <B0000101000 AT stargate DOT astr DOT lu DOT lv>
X-mailer: Pegasus Mail for Win32 (v3.12a)
Reply-To: djgpp-workers AT delorie DOT com
X-Mailing-List: djgpp-workers AT delorie DOT com
X-Unsubscribes-To: listserv AT delorie DOT com
Precedence: bulk

On 10 Sep 99, at 11:27, Rob Kramer wrote:

> I've disassembled the executable, and it seems the compiler issues a locked 
> 'xaddl' instruction. I couldn't find opcode 'c1' in my 386 reference manual, 
> so this may be a 486++ instruction?   (I can't find anything but Intel's 386 
> manual online, sigh)
> 
>     2052:	f0 0f c1 02    	lock xaddl %eax,(%edx)

It's from libstdc++. File bastring.h (for DJGPP it's 
lang/cxx/std/bastring.h) uses such inline assembler when __i486__,
__i586__ or __i686__ is defined 

> 
> > Confirmed with DJGPP ports of egcs-1.1.2 and gcc-2.95.1 (I don't more 
> > gcc-2.95 installed, but that doesn't matter). 2 possible cures:
> >  - currently use option '-march=i386' on 386, it should help
> 
> I tried that already, to no avail.. Tried again, and the compiler outputs the 
> same 'xaddl' at a slighty different location (other parts of the code are 
> different)

One can use -march=i386 -mcpu=i386 now.

It seems to be bug in gcc. I built gcc-2.95.1 for DJGPP with
 -march=pentium -mcpu=pentium as default (target=i586-pc-
msdosdjgpp). Specifying -march=i386 doesn't prevent macros
__i486__, __i586__, __pentium__ from being defined, so bastring.h
from libstdc++ uses inline assembler incorrect for i386. If I'm specifying
-mcpu=i386 then these macros are not defined. 

The reason is that CPP_CPU_SPEC in gcc/config/i386/i386.h is broken
--------------------------------------------------------------------------
#ifndef CPP_CPU_SPEC
#define CPP_CPU_SPEC "\
-Acpu(i386) -Amachine(i386) \
%{!ansi:-Di386} -D__i386 -D__i386__ \
%{mcpu=i486:%(cpp_486)} %{m486:%(cpp_486)} \
%{mpentium:%(cpp_586)} %{mcpu=pentium:%(cpp_586)} \
%{mpentiumpro:%(cpp_686)} %{mcpu=pentiumpro:%(cpp_686)} \
%{mcpu=k6:%(cpp_k6)} \
%{!mcpu*:%{!m486:%{!mpentium*:%(cpp_cpu_default)}}}"
#endif
--------------------------------------------------------------------------

There are at least following problems:
	
1) no check is done for -m386 so option -m386 is silently ignored if the 
default is something else (for example I have built gcc for i586-*)

2) I think we should check for -march=foo instead of -mcpu=foo in
CPP_CPU_SPEC


> >  - in future: I must rebuild gcc-2.95.1 for target i386-pc-msdosdjgpp
> 
> Hmm. How come that in thousands of lines of code, the compiler only fails in 
> a global string definition..
> 

Andris