i386 - subtle bug with asm volatile constraints

[Markus F.X.J. Oberhumer]

On 13-Apr-99 Jeffrey A Law wrote:
> > >   > below you will find a new test.c and the output from
> > >   >  gcc '-DEAX="a"' -O2 -S -o test_a.s test.c
> > >   > and
> > >   >  gcc '-DEAX="0"' -O2 -S -o test_0.s test.c
> > >   > 
> > >   > You will see that missing `int $80' problem when you run a diff
> > >   > between test_a.s and test_0.s.
> > > Still no good.  I don't get any missing int $80 instructions.
> > 
> > This is somewhat surprising - does this mean that your version
> > generates identical code (as it should be) ?
>
> No.  But the int80s are the same.  The differences were not relavent to the
> correct operation of the code.

My previous mail to egcs-bugs also contains both assembler output files as
generated on my machine. They are identical except 2 missing `int $80'.

If you can't reproduce this behaviour using the compile statements above
then you should get identical output because hte EAX="0" and EAX="a"
constraints are equivalent in this test program.

> > I can spot the missing basic blocks in test.c.jump2 and test.c.stack, but
> > as I'm completely unfamiliar with gcc internals (and RTL) this about
> > all I can tell. test.c.greg also seems to contain some interesting
> > differences.
>
> It's not the differences that you need to look for.  You need to look for the
> first dump which does not have the proper number of int80 instructions.  The
> differences between dumps are not interesting for this problem.

I've compiled with -dap -DEAX="0" in one directory and with -dap -DEAX="a" in
another directory and run a diff on the two directories. And I can spot
the missing `int $80' basic blocks in the -DEAX="0" case in the .jump2 and
.stack files (as compared to the -DEAX="a" version).

> > Well, I could mail you the intermediate -dap files of both versions, but
> > this might not be very useful unless you can reproduce the problem.
> > Any suggestions on how to procede ?
>
> You're going to have to do the debugging on this -- I've already wasted too
> much time on a wild goose chase.

I'm sorry to waste anybodys time, either by unclear instructions or because
the bug might be fixed in the latest snapshot, but unfortunately I am not
able to debug this any deeper as I don't know anything about gcc internals.

All I can tell you that this code generation problem is reproducible
on my machine with both egcs 1.1.1 and gcc 2.7.2, and reportedly with
a March egcs snapshot as well. Maybe someone else on this list can
find out what is happening here.

Markus

-----   Markus F.X.J. Oberhumer <markus.oberhumer@jk.uni-linz.ac.at>   -----
-----             http://wildsau.idv.uni-linz.ac.at/mfx/               -----
-----        5E CB 5C 85 DE AF 9E BF  E9 DA 7E 6A 39 F8 CC 67          -----

                             3 WARPS TO URANUS