Re: [PATCH,RFC] Disallow reordering of x87 insns while scheduling

[Roger Sayle <roger at eyesopen dot com>]

Hi Vlad,

On Mon, 18 Apr 2005, Vladimir N. Makarov wrote:
> I think the most important platform is Pentium-M right now.  So I
> ran SPECFP95 on my 1.6Ghz Pentium-M notebook (sorry SPECFp2000 is too
> much for the notebook).
>
> A big degradation for wave5 (in size and rate) is surprise for me.
> Text segment size (base is the first, peak is the second).
>
> ----------------CFP95-----------------
>  11.395%         111099         123759 146.wave5


Could you describe which compiler flags and options you used in
your SPEC benchmarking, and against which branch you tested?

The reason is that my compiling of wave5_data.f from SPECfp95 with
mainline's gfortran and the compiler flags "-O2 -mtune=pentiumpro"
I currently see a small decrease in object file size, as
expected/predicted from all the other SPEC executables and CSiBE,
and certainly not the mysterious 11% code growth you report above.

crush% wc wave5_data*.s
  43597  101160  664861 wave5_data.old.s
  42314   98586  648211 wave5_data.new.s

crush% ls -al wave5_data*.o
-rw-r--r--    1 roger    users      321508 Apr 18 21:07 wave5_data.old.o
-rw-r--r--    1 roger    users      318980 Apr 18 21:07 wave5_data.new.o


Unfortunately, adding the "-ffast-math" flag to gfortran's command
line options results in an ICE in the tree-ssa optimizers that looks
like it's related to garbage collection both with and without my change.
#0  is_gimple_stmt (t=0xa5a5a5a5) at ../../gcc/gcc/tree-gimple.c:196


Could you double check your numbers and see if they're reproduceable?
I suspect that the incomprehensible 10% code growth from simply
disabling scheduling is probably likely to be the result of some
silly mistake.

Thanks in advance,

Roger
--