[prev in list] [next in list] [prev in thread] [next in thread] 

List:       gdb
Subject:    Re: JIT debugging (Attach and speed)
From:       Yichao Yu <yyc1992 () gmail ! com>
Date:       2016-03-24 3:14:32
Message-ID: CAMvDr+S4pwdPwdRK71uwTqcCc5CoNHHsUuZQKAqbF4JG+y3Evg () mail ! gmail ! com
[Download RAW message or body]

> The prototype patch below (also in users/palves/jit-speedup, along
> with the attach fixes), simply delays computing a frame until
> it's necessary, and it cuts down the time for:

(The patch seems to have some white space breakage but the branch works great.)

> gdb ---args ./testsuite/outputs/gdb.base/jit/jit-main \
> ./testsuite/outputs/gdb.base/jit/jit-solib.so 2000 
> from "takes too long so I canceled it" minutes, to around 2 seconds, for me.

There's indeed a huge improvement. The test I was using with ~2000 JIT
objects goes drops from ~400s to 7.5s (~3s without gdb). However a
rough scaling test suggests that the scaling is still ~O(n^2) 20k JIT
objects takes ~500s (~130s without gdb). It seems that the master was
actually super quadratic (if this is a word) when dealing with
multiple JIT objects.

> Having a pending breakpoint that is never resolved slows things down
> significantly again though...  The perf profile shows that breakpoint_re_set
> ends up calling parse_linespec which wants the selected block, which wants
> the selected frame, and then we're back at the same...  Maybe we
> could reconsider how pending breakpoints are re-parsed, and which
> context they're parsed in.  Not sure.
> 
> Still, very promising.
> 

Indeed. It's already so much better.


[prev in list] [next in list] [prev in thread] [next in thread]