Re: [PATCH v2 3/3] perf report: Implement visual marker for macro fusion in annotate

From: Arnaldo Carvalho de Melo
Date: Mon Jun 19 2017 - 21:38:02 EST


Em Tue, Jun 20, 2017 at 09:25:35AM +0800, Jin, Yao escreveu:
>
> > Ok, thanks for making this per-arch! Some comments:
> >
> > I think we should have this marked permanently, i.e. not just when we go
> > to the jump line, something like this (testing here in a t450s
> > broadwell, function hc_find_func, /usr/lib64/liblzma.so.5.2.2):
> >
> > It is like this now, when we are not on the jne jump line:
> >
> > 0.71 â mov %r14d,%r10d â
> > â movzbl (%rdx,%r10,1),%ebp â
> > 1.06 â 70: mov (%r9,%rcx,4),%ecx â
> > 77.98 â 74: cmp %bpl,(%rbx,%r10,1) â
> > â â jne 70 â
> > 0.85 â movzbl (%rdx),%r10d â
> > 0.99 â cmp %r10b,(%rbx) â
> >
> > I think it should be augmented to:
> >
> > 0.71 â mov %r14d,%r10d â
> > â movzbl (%rdx,%r10,1),%ebp â
> > 1.06 â 70: ââmov (%r9,%rcx,4),%ecx â
> > 77.98 â 74: ââcmp %bpl,(%rbx,%r10,1) â
> > â â jne 70 â
> > 0.85 â movzbl (%rdx),%r10d â
> > 0.99 â cmp %r10b,(%rbx) â
> >
> > I.e. no arrow, the two instructions that end up as one micro-op being
> > connected.
>
> The fused instruction pairs are:
> cmp + jcc
> test + jcc
> add + jcc
> sub + jcc
> and + jcc
> inc + jcc
> dec + jcc
>
> Mov and cmp are not the fused instruction pair. So we don't need to connect

Right, my bad, what I was trying to say was to have a marker for fused
instructions, not just when we go with the cursor over it like with this
patchset.

> mov and cmp. I guess what Arnaldo wants is to connect two fused instructions
> even we don't go to the jcc line. For example: a line is connected between
> cmp and jne in above case.

Right

> I have thought about that. While the visualization may be not very good
> because the original arrow before jne would be overwritten. So now I just
> implement a way that joins the jump arrow when we go to the jcc line.
> Another consideration is the fused instruction pairs are very common
> instructions in code, if we mark them all, there may be too much.

perhaps

> > And then this:
> >
> > â âââcmpl $0x0,argp_program_version_hook
> > 81.93 â âââje 20
> > â â lock cmpxchg %esi,0x38a9a4(%rip)
> > â ââ jne 29
> > â ââ jmp 43
> > 11.47 â20:âââcmpxch %esi,0x38a999(%rip)
> >
> > Would look better as:
> >
> > â âââcmpl $0x0,argp_program_version_hook
> > 81.93 â âââje 20
> > â â lock cmpxchg %esi,0x38a9a4(%rip)
> > â ââ jne 29
> > â ââ jmp 43
> > 11.47 â20:âââcmpxch %esi,0x38a999(%rip)
> >
> > Patch below, please test/ack :-)
>
> I have tested. It's better! There is no space in the line. Thanks!
>
> > This was the low hanging fruit, having the:
> >
> > 1.06 â 70: ââmov (%r9,%rcx,4),%ecx â
> > 77.98 â 74: ââcmp %bpl,(%rbx,%r10,1) â
> >
> > Marker always there, not just when we have the cursor on top of one of
> > those lines remains to be coded.
>
> My comment is as above.
>
> > But you state:
> >
> > ------------
> > Macro fusion merges two instructions to a single micro-op. Intel core
> > platform performs this hardware optimization under limited
> > circumstances.
> > ------------
> >
> > "Intel core", what about older arches, etc, don't you have to look at:
> >
> > # cpudesc : Intel(R) Core(TM) i7-5600U CPU @ 2.60GHz
> > # cpuid : GenuineIntel,6,61,4
> >
> > present in the perf.data header (or in the running system, for things
> > like 'perf top') to make sure that this is a machine where such "macro
> > fusion" takes place?
> >
> > - Arnaldo
>
> Reference for macro fusion is the optimization guide,
> http://www.intel.com/content/www/us/en/architecture-and-technology/64-ia-32-architectures-optimization-manual.html
> 2.3.2.1
> â In Intel microarchitecture code name Nehalem: CMP, TEST.
> â In Intel microarchitecture code name Sandy Bridge: CMP, TEST, ADD, SUB,
> AND, INC, DEC
> â These instructions can fuse if The first source / destination operand is a
> register.
>
> The second source operand (if exists) is one of: immediate, register, or non
> RIP-relative memory.
> The second instruction of the macro-fusable pair is a conditional branch.
>
> We probably don't need the full rules, just a simple test for
> CMP/TEST/ADD/SUB/AND/INC/DEC and second instruction a Jcc condition branch.
> Also I don't think we need to distinguish Nehalem/Sandy Bridge and other
> core platforms. A simple test may be acceptable.

Humm, then we need to make sure somehow that this may or may not be
happening, with the above rules and optimization guide URL and pages
mentioned in the documentation.

I think that as we improve the disassembler, the more precise we can go
the better. If we know that the machine is x86 _and_ Nehalem, then we
should do this fusing visual cue onlyu for CMP and TEST, etc.

- Arnaldo

> Thanks!
> Jin Yao
>
> > diff --git a/tools/perf/ui/browser.c b/tools/perf/ui/browser.c
> > index acba636bd165..9ef7677ae14f 100644
> > --- a/tools/perf/ui/browser.c
> > +++ b/tools/perf/ui/browser.c
> > @@ -756,8 +756,10 @@ void ui_browser__mark_fused(struct ui_browser *browser, unsigned int column,
> > ui_browser__gotorc(browser, end_row, column);
> > SLsmg_draw_hline(2);
> > ui_browser__gotorc(browser, end_row + 1, column - 1);
> > - SLsmg_draw_vline(1);
> > + SLsmg_write_char(SLSMG_LTEE_CHAR);
> > } else {
> > + ui_browser__gotorc(browser, end_row, column - 1);
> > + SLsmg_write_char(SLSMG_LTEE_CHAR);
> > ui_browser__gotorc(browser, end_row, column);
> > SLsmg_draw_hline(2);
> > }