Re: [PATCH][dbgcnt] New debug counter global_alloc_at_func, global_alloc_at_reg, and make sched_insn work

["Seongbae Park (박성배, 朴成培)" <seongbae dot park at gmail dot com>]

On 6/27/07, Ayal Zaks <ZAKS@il.ibm.com> wrote:
> We should use dbg_cnt instead of MAX_SMS_LOOP_NUMBER in modulo-sched.c.

Probably, as debug counter provides slightly better abstraction.
Ultimately we should be able to write a script that does the binary search
using various counters automatically, even a script that finds
the correct counter to use. One option I haven't implemented
is to dump the counters at the end of the compilation
- this in itself is interesting as we can see how many times
each counter is triggered, but it's actually very useful
because we can use those values as the upper bound for the binary search.

> Perhaps these counters should be part of every pass, and use the inherent
> pass names. In some cases a pass may want more than one counter.

I haven't put a lot of thought into naming convention for the counters,
I'm sure having some convention would be nice - any idea ?
Most likely, those counters have different granularity -
function/block/instruction/
transformation/register/ or whatever the unit of each optimization.
The finer control we have, the better.

> Are there
> any optimizations that cannot make use of such a counter?

I'm sure there are. But once we have enough passes
that have the finer control, even passes that can't have such a counter
will benefit, as long as we can control passes before the problematic one
to trigger/not trigger the buggy condition.
--
#pragma ident "Seongbae Park, compiler, http://seongbae.blogspot.com";