Remove profiling overhead from profiles

[MasterDuke17]

By calculating the average cost of executing the profiling functions and
then subtracting that times the number of entries from the exclusive
cost of an HLL block.

This gist (https://gist.github.com/MasterDuke17/ab7108070054f92e4cab5313033ea3f8) has some data for what the values were when https://github.com/MasterDuke17/MoarVM/blob/remove_profiling_overhead_from_profile/src/profiler/instrument.c#L399 and https://github.com/MasterDuke17/MoarVM/blob/remove_profiling_overhead_from_profile/src/profiler/instrument.c#L418 branches were taken when profiling (^5_000).grep(*.is-prime).tail.say; my $a = (^100).pick(2).join; my $b; for ^100_000 { ++$b if $a ~~ / $_ / }; say $b.

[jnthn]

What's the * 0.9 here?

[MasterDuke17]

Just a scaling factor in case the measurement is a little high.

…

Sent from my iPhone

On Jun 7, 2019, at 4:26 PM, Jonathan Worthington ***@***.***> wrote: @jnthn commented on this pull request. In src/profiler/profile.c: > MVM_profile_instrumented_start(tc, config); + + /* Call the profiling functions a bunch of times and record how long they took. */ + s = uv_hrtime(); + for (i = 0; i < 1000; i++) { + MVM_profile_log_enter(tc, tc->cur_frame->static_info, MVM_PROFILE_ENTER_NORMAL); + MVM_profile_log_exit(tc); + } + e = uv_hrtime(); + tc->instance->profiling_overhead = (MVMuint64) ((e - s) / 1000) * 0.9; What's the * 0.9 here? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

[jnthn]

Just a scaling factor in case the measurement is a little high.

Hm, from the loop overhead, perhaps? It's quite hard to reason about, since the C compiler may also do inlining and cancel the cost of the function call overhead.

I guess let's give this a try, though; I think even slightly off compensation is going to give better numbers than none at all.