I took a closer look at some of the worst offenders, and in fact,
when I use the old code generator, the difference in sizes is not all
that great.  So I suspect that I (or someone else) introduced
an unrelated regression in the backend.  I'll have some better stats
soon...

Edward

Excerpts from Edward Z. Yang's message of Thu Apr 07 11:37:05 -0400 2011:
> I've been working on new optimization passes for Hoopl to get
> the code back to the quality of the old code generator.
> With the new optimization pass, the binaries in the test suite
> are, on average, 7% larger than the old code generator.
> 
> Here are the worst offenders, with T3294 clocking in with a whopping
> 112% increase in size.
> 
>     %change oldsize delta  filename
>     -----------------------------------------------------
>     0.3233   6224   2012 ./deSugar/should_compile/ds036.o
>     0.3235   2068    669 ./ffi/should_compile/cc005.o
>     0.3333  11616   3872 ./arrows/should_compile/arrowcase1.o
>     0.3371  35228  11876 ./typecheck/should_compile/tc095.o
>     0.3374 111300  37556 ./indexed-types/should_compile/T3787.o
>     0.3394   2684    911 ./ffi/should_run/ffi002.o
>     0.3443   2916   1004 ./simplCore/should_compile/simpl016.o
>     0.3581   3664   1312 ./haddock/should_compile_flag_haddock/haddockA033.o
>     0.4439   7948   3528 ./gadt/red-black.o
>     0.4791 209128 100200 ./deSugar/should_run/dsrun023.o
>     0.5381   1548    833 ./driver/B027/F.o
>     0.5822   1520    885 ./driver/F016.o
>     0.5822   1520    885 ./driver/F034.o
>     0.5822   1520    885 ./driver/obj018/TestStub018.o
>     0.5822   1520    885 ./driver/out019/TestStub019.o
>     0.6001   1488    893 ./driver/B044/F.o
>     0.6111   1476    902 ./ffi/should_compile/cc012.o
>     0.7677   3512   2696 ./ffi/should_compile/cc013.o
>     1.1209 700884 785632 ./perf/compiler/T3294.o
> 
> I am planning on doing similar tests on nofib.
> 
> Cheers,
> Edward

_______________________________________________
Cvs-ghc mailing list
[email protected]
http://www.haskell.org/mailman/listinfo/cvs-ghc

Reply via email to