Re: PATCH: Use edge probabilities in interblock and ebb schedulers

[Vladimir Makarov <vmakarov at redhat dot com>]

Maxim Kuvyrkov wrote:

> Hi,
>
> As it seems that patch by Peter Steinmetz was never applied, here is its
> fixed version (we could have got division by zero in sched-rgn.c:
> compute_trg_info() if probability of a parent block is zero).
>
> This is the link to the original patch:
> http://gcc.gnu.org/ml/gcc-patches/2005-09/msg00370.html
>
> Also, please find attached patch that similarly computes basic block
> probability information in sched-ebb.c.  Currently it is possible
> that CFG has wrong values in bb->frequency fields of some basic blocks.
> The proposed patch calculates basic block frequencies for each EBB
> separately relying on edge probabilities information. It does not
> correct frequency values in CFG.

Could you be more specific what is wong with bb->frequency.

Even if you found wrong values of bb->frequency, it should be fixed in 
the first place not in just one place.

So you should use bb->frequency in the patch even if their values are 
wrong and you can not fix them in the 1st place or you have no time for 
this.

Send me the modified patch and I approve it.  And please do not forget 
comments for new static variables (and functions although in the 
modified patch there will be no new functions).

> SPECCPU2000 shows 2 improvements when compiled with gcc -O2 with both of
> these patches: 181.mcf: +4.19%, 188.ammp: 11.37%. And several
> regressions: 197.parser: -1.25%, 300.twolf: -1.67%, 179.art: -1.41%. But
> still, the patches do more good than bad: SPECint_base: +0.14%,
> SPECfp_base: +0.72%.
>
> Both patches were bootstraped and regtested on ia64-unknown-linux-gnu on
> mainline weekly snapshot 20060107 (as all later snapshots, including
> trunk, fail to compile because of libgomp).
>
>
>