Re: [tip:perf/core] perf ui annotate browser: Allow toggling addr offset view

[Arnaldo Carvalho de Melo <acme@redhat.com>]

Em Fri, Apr 13, 2012 at 11:25:50AM -0700, Linus Torvalds escreveu:
> On Fri, Apr 13, 2012 at 11:07 AM, tip-bot for Arnaldo Carvalho de Melo
> <acme@redhat.com> wrote:
> >
> > The offset view will be the default as soon as operations that deal with
> > offsets in a function are handled accodringly, i.e. in offset view the
> > above will become:
> >
> >     2f:       jne    __list_del_entry+0x84
> > <SNIP>
> >     84:       mov    %rdi,%rcx
> 
> Oh, please please please make this happen.

I will in the next days. Glad you like it.
 
> Also, if at all possible, please mark offsets that are referenced by a
> branch some way. I do think that the "size of instruction" can be a
> very valid piece of information when looking at instruction-level
> profiles, but quite frankly, I'm also 100% sure that the fact that an
> instruction is a branch target is *more* important.
> 
> Now, fancy arrows etc might be too hard to do (especially without
> cluttering things up too much), but marking just the targets would
> already be lovely. IOW, turn this mess:
> 
> 
>          :        ffffffff810d7e80 <kmem_cache_free>:
>     1.91 :        ffffffff810d7e80:       push   %rbp
>     0.05 :        ffffffff810d7e81:       mov    %rsp,%rbp
>     2.02 :        ffffffff810d7e84:       push   %r12
>     0.00 :        ffffffff810d7e86:       mov    %rdi,%r12
>     0.55 :        ffffffff810d7e89:       push   %rbx
>     0.00 :        ffffffff810d7e8a:       mov    %rsi,%rdi
>     1.20 :        ffffffff810d7e8d:       mov    %rsi,%rbx
>     1.64 :        ffffffff810d7e90:       callq  ffffffff8102a730 <__phys_addr>
>     0.05 :        ffffffff810d7e95:       movabs $0xffffea0000000000,%rdi
>     0.22 :        ffffffff810d7e9f:       shr    $0xc,%rax
>     0.11 :        ffffffff810d7ea3:       shl    $0x6,%rax
>     1.64 :        ffffffff810d7ea7:       lea    (%rax,%rdi,1),%rdi
>    42.80 :        ffffffff810d7eab:       mov    (%rdi),%rax
>     2.02 :        ffffffff810d7eae:       test   $0x80,%ah
>     0.00 :        ffffffff810d7eb1:       jne    ffffffff810d7ef6
> <kmem_cache_free+0x76>
>     0.22 :        ffffffff810d7eb3:       mov    0x8(%rbp),%r9
>     7.14 :        ffffffff810d7eb7:       mov    (%r12),%rax
>     0.65 :        ffffffff810d7ebb:       add    %gs:0xcb88,%rax
>     4.58 :        ffffffff810d7ec4:       mov    0x8(%rax),%rdx
> 
> (which has tons of unnecessary white-space too - shades of G+) into

Grin ;-)

Yeah, even the : column isn't needed I think.
 
> <kmem_cache_free>:
>     1.91 :          push   %rbp
>     0.05 :          mov    %rsp,%rbp
>     2.02 :          push   %r12
>     0.00 :          mov    %rdi,%r12
>     0.55 :          push   %rbx
>     0.00 :          mov    %rsi,%rdi
>     1.20 :          mov    %rsi,%rbx
>     1.64 :          callq  ffffffff8102a730 <__phys_addr>
>     0.05 :          movabs $0xffffea0000000000,%rdi
>     0.22 :          shr    $0xc,%rax
>     0.11 :          shl    $0x6,%rax
>     1.64 :          lea    (%rax,%rdi,1),%rdi
>    42.80 :          mov    (%rdi),%rax
>     2.02 :          test   $0x80,%ah
>     0.00 :          jne    kmem_cache_free+0x76
>     0.22 : 0x33:    mov    0x8(%rbp),%r9
>     7.14 : 0x37:    mov    (%r12),%rax
>     0.65 :          add    %gs:0xcb88,%rax
>     4.58 :          mov    0x8(%rax),%rdx
>    ....
> 
> or something. I think the "callq  ffffffff8102a730 <__phys_addr>"
> could also be prettified with *zero* downside.

Yeah, no need to see the address in default mode.
 
> Sure, leave some magic keycombination to get the "full information", but ..

Yeah, we have o to toggle offset view + s to toggle source code view,
etc. 

>                       Linus