Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <202106281231.E99B92BB13@keescook>
In-Reply-To: <202106281231.E99B92BB13@keescook>
From:   Linus Torvalds <torvalds@linux-foundation.org>
Date:   Mon, 28 Jun 2021 19:49:04 -0700
Message-ID: <CAHk-=whqCT0BeqBQhW8D-YoLLgp_eFY=8Y=9ieREM5xx0ef08w@mail.gmail.com>
Subject: Re: [GIT PULL] Clang feature updates for v5.14-rc1
To:     Kees Cook <keescook@chromium.org>
Cc:     Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        Bill Wendling <morbo@google.com>,
        Bill Wendling <wcw@google.com>,
        Catalin Marinas <catalin.marinas@arm.com>,
        clang-built-linux <clang-built-linux@googlegroups.com>,
        Fangrui Song <maskray@google.com>,
        Heiko Carstens <hca@linux.ibm.com>,
        Jarmo Tiitto <jarmo.tiitto@gmail.com>,
        Lukas Bulwahn <lukas.bulwahn@gmail.com>,
        Mark Rutland <mark.rutland@arm.com>,
        Masahiro Yamada <masahiroy@kernel.org>,
        Miguel Ojeda <ojeda@kernel.org>,
        Nathan Chancellor <nathan@kernel.org>,
        Nick Desaulniers <ndesaulniers@google.com>,
        Peter Oberparleiter <oberpar@linux.ibm.com>,
        Peter Zijlstra <peterz@infradead.org>,
        Sami Tolvanen <samitolvanen@google.com>,
        Will Deacon <will@kernel.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Mon, Jun 28, 2021 at 12:32 PM Kees Cook <keescook@chromium.org> wrote:
>
> The big addition for this merge window is the core support for Clang's
> Profile Guided Optimization, which lets Clang build the kernel for
> improved performance when running specific kernel workloads. This
> currently covers only vmlinux, but module support is under active
> development. (Sami Tolvanen, Bill Wendling, Kees Cook, Jarmo Tiitto,
> Lukas Bulwahn)

Am I misreading this?

The PGO data seems to be done by using clang instrumentation, instead
of done sanely using sample data from a regular "perf" run?

That odd decision seems to not be documented anywhere, and it seems
odd and counter-productive, and causes all that odd special buffer
handling and that vmlinux.profraw file etc.

And it causes the kernel to be bigger and run slower.

The actual link to the clang pgo documentation even says that there is
already support for converting regular "perf" profile output to pgo
data, yet that model isn't actually used even though it appears vastly
superior.

Why use an inferior compiler instructmentation profile when we have
the much better tools?

                  Linus