Decoupling -O from -fvia-C

[Simon Marlow]

Wolfgang Thaller wrote:
>> may be it's better to start with speed tests?
>
> What would that achieve, given that the decision does not depend on
> speed measurements? The argument "implicit -fvia-C makes things
> refuse to work" is stronger than the argument "-fvia-C makes things x
> % faster on platform Y, so it should be on by default", for all
> values of x and y.
>
> That said, speed tests are always a good thing, as long as I don't
> have to do them myself ;-).

Here are the results from last night's 6.6 build, on x86_64, over the whole nofib suite:

           Size    Allocs   Runtime
          -0.3%     -0.0%     -1.0%

That is, the NCG makes binaries 0.3% smaller and 1% faster.  The speed measurements are well within the boundaries of error on this machine, so don't take that 1% as significant.

On x86:

            -0.0%     -0.0%     -3.6%

That %3.6 speed improvement for the NCG is a little more significant.  It reflects the fact that -fvia-C uses -ffloat-store, but the NCG doesn't do anything equivalent.  The NCG does however have terrible floating point performance on x86, but it clearly isn't as bad as -ffloat-store.

Oh, I just noticed that -fexcess-precision is broken in 6.6.  Grrr.

Cheers,
        Simon