>
> As noted above, the change to avoid the RDIV by runs decreases the
> effect of the min_run_ratio (now unlikely_count_fraction) in half. So
> we really need to increase this parameter to 32 to compare to what the
> old version of the code was doing.
>
> Indeed, using 32 does fix the same set o
Ping on this one since the new param would be useful in the other
profile related patch I'm working on (the one to handle the dropped
profile counts in tree-profile).
Thanks,
Teresa
On Thu, Oct 10, 2013 at 2:35 PM, Teresa Johnson wrote:
> On Mon, Oct 7, 2013 at 10:45 PM, Teresa Johnson wrote:
>
On Mon, Oct 7, 2013 at 10:45 PM, Teresa Johnson wrote:
> On Sun, Oct 6, 2013 at 6:43 AM, Jan Hubicka wrote:
>>> 2013-10-03 Teresa Johnson
>>>
>>> * params.def (PARAM_MIN_HOT_RUN_RATIO): New parameter.
>>
>> I would not mention HOT in the parameter name. Blocks are
>> hot/normal/unlik
On Mon, Oct 7, 2013 at 10:49 PM, Teresa Johnson wrote:
> On Sun, Oct 6, 2013 at 6:51 AM, Jan Hubicka wrote:
>>> The second part ensures that frequencies are correctly updated after
>>> inlining. The problem is that after inlining the frequencies were
>>> being recomputed directly from the corresp
On Sun, Oct 6, 2013 at 6:51 AM, Jan Hubicka wrote:
>> The second part ensures that frequencies are correctly updated after
>> inlining. The problem is that after inlining the frequencies were
>> being recomputed directly from the corresponding bb counts in
>> rebuild_frequencies. If the counts had
On Fri, Oct 4, 2013 at 8:54 AM, Xinliang David Li wrote:
> On Thu, Oct 3, 2013 at 10:15 PM, Teresa Johnson wrote:
>> This patch handles the fact that small profile count values sometimes
>> get truncated down to 0 during updates due to integer arithmetic. This
>> causes sub-optimal function split
On Sun, Oct 6, 2013 at 6:43 AM, Jan Hubicka wrote:
>> 2013-10-03 Teresa Johnson
>>
>> * params.def (PARAM_MIN_HOT_RUN_RATIO): New parameter.
>
> I would not mention HOT in the parameter name. Blocks are hot/normal/unlikely
> and we HOT_BB_FREQUENCY_FRACTION.
>
> So perhaps UNLIKELY_BB_
> The second part ensures that frequencies are correctly updated after
> inlining. The problem is that after inlining the frequencies were
> being recomputed directly from the corresponding bb counts in
> rebuild_frequencies. If the counts had been truncated to 0, then the
> recomputed frequencies
> 2013-10-03 Teresa Johnson
>
> * params.def (PARAM_MIN_HOT_RUN_RATIO): New parameter.
I would not mention HOT in the parameter name. Blocks are hot/normal/unlikely
and we HOT_BB_FREQUENCY_FRACTION.
So perhaps UNLIKELY_BB_COUNT_FRACTION with an explanation that it is relative
to numb
On Thu, Oct 3, 2013 at 10:15 PM, Teresa Johnson wrote:
> This patch handles the fact that small profile count values sometimes
> get truncated down to 0 during updates due to integer arithmetic. This
> causes sub-optimal function splitting under
> -freorder-blocks-and-partition.
>
> The first part
This patch handles the fact that small profile count values sometimes
get truncated down to 0 during updates due to integer arithmetic. This
causes sub-optimal function splitting under
-freorder-blocks-and-partition.
The first part fixes the logic in probably_never_executed that looks
at the bb fr
11 matches
Mail list logo