Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
From:   Vineet Gupta <vineet.gupta1@synopsys.com>
To:     Eugeniy Paltsev <eugeniy.paltsev@synopsys.com>,
        "linux-arm-kernel@lists.infradead.org" 
        <linux-arm-kernel@lists.infradead.org>,
        "linux-snps-arc@lists.infradead.org" 
        <linux-snps-arc@lists.infradead.org>
CC:     "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
        "peterz@infradead.org" <peterz@infradead.org>,
        "jolsa@redhat.com" <jolsa@redhat.com>,
        "acme@kernel.org" <acme@kernel.org>,
        Alexey Brodkin <alexey.brodkin@synopsys.com>,
        "namhyung@kernel.org" <namhyung@kernel.org>,
        "mark.rutland@arm.com" <mark.rutland@arm.com>,
        "will.deacon@arm.com" <will.deacon@arm.com>,
        "mingo@redhat.com" <mingo@redhat.com>,
        "alexander.shishkin@linux.intel.com" 
        <alexander.shishkin@linux.intel.com>
Subject: Re: 'branches' perf event mapping differs on ARC and ARM
Thread-Topic: 'branches' perf event mapping differs on ARC and ARM
Thread-Index: AQHUhl6UMI9IICATOUqAlhDukujZmA==
Date:   Tue, 27 Nov 2018 17:06:10 +0000
Message-ID: <C2D7FE5348E1B147BCA15975FBA2307501464172A6@US01WEMBX2.internal.synopsys.com>
References: <1543329386.13651.13.camel@synopsys.com>
Accept-Language: en-US
Content-Language: en-US
Content-Type: text/plain; charset="us-ascii"
Content-Transfer-Encoding: quoted-printable
MIME-Version: 1.0
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On 11/27/18 6:36 AM, Eugeniy Paltsev wrote:=0A=
> Hi,=0A=
>=0A=
> While playing with perf tool on ARMv7 and ARCv2 processors and profiling =
the=0A=
> same application I got interesting results. Even if we got pretty=0A=
> similar total=0A=
> execution time and instructions number the number of branches on ARC is a=
bout=0A=
> three times more then on ARM.=0A=
>=0A=
> I dug into architecture=0A=
> specific perf sources and found that we map different=0A=
> HW counters into generic 'branches' event on ARC and ARM.=0A=
> - We use "ijmp" event on ARC which=0A=
> counts all jump and branch instructions (regardless=0A=
> of real execution flow - even if no real jump happens)=0A=
=0A=
That doesn't seem correct IMO. A NOT taken conditional branch doesn't chang=
e=0A=
control flow, so semantically doesn't qualify as a branch.=0A=
On ARC, the generic branches event should be mapped to "actually taken bran=
ches"=0A=
condition i.e.  ijmptak=0A=
=0A=
=0A=
> - We use "pc_write_retired" event on ARM=0A=
> which counts only taken branches (Instruction=0A=
> architecturally executed, condition check pass - software change of the P=
C)=0A=
=0A=
That seems correct.=0A=
=0A=
> I guess counting all jump and branch instructions is correct because we u=
se=0A=
> 'branches' event value to calculate relative value of 'branch-misses' usi=
ng=0A=
>=0A=
> following formula:=0A=
> ----------------------------8----------------------------=0A=
> branch-misses-ration =3D 'branch-misses' / 'branches' * 100.0=0A=
> ----------------=0A=
> ------------8----------------------------=0A=
> And using only taken branches here is incorrect IMHO.=0A=
=0A=
Why ? branch-misses is a CPU specific micro-arch state where the a changed =
control=0A=
flow was NOT predicted. If an implementation mispredicts NOT taken branches=
, those=0A=
should actually get counted and be fed to hardware folks to improve the mic=
ro-arch.=0A=
=0A=
=0A=
>  So I guess we should=0A=
> map 'br_immed_retired' instead of=0A=
> "pc_write_retired" into generic 'branches'=0A=
> event on ARM.=0A=
=0A=