On Thu, Sep 10, 2020 at 10:18:05AM +0900, Masahiro Yamada wrote: > On Wed, Sep 9, 2020 at 8:46 AM Sami Tolvanen <samitolvanen@xxxxxxxxxx> wrote: > > > > On Sun, Sep 06, 2020 at 09:24:38AM +0900, Masahiro Yamada wrote: > > > On Fri, Sep 4, 2020 at 5:30 AM Sami Tolvanen <samitolvanen@xxxxxxxxxx> wrote: > > > > > > > > This patch series adds support for building x86_64 and arm64 kernels > > > > with Clang's Link Time Optimization (LTO). > > > > > > > > In addition to performance, the primary motivation for LTO is > > > > to allow Clang's Control-Flow Integrity (CFI) to be used in the > > > > kernel. Google has shipped millions of Pixel devices running three > > > > major kernel versions with LTO+CFI since 2018. > > > > > > > > Most of the patches are build system changes for handling LLVM > > > > bitcode, which Clang produces with LTO instead of ELF object files, > > > > postponing ELF processing until a later stage, and ensuring initcall > > > > ordering. > > > > > > > > Note that patches 1-4 are not directly related to LTO, but are > > > > needed to compile LTO kernels with ToT Clang, so I'm including them > > > > in the series for your convenience: > > > > > > > > - Patches 1-3 are required for building the kernel with ToT Clang, > > > > and IAS, and patch 4 is needed to build allmodconfig with LTO. > > > > > > > > - Patches 3-4 are already in linux-next, but not yet in 5.9-rc. > > > > > > > > > > > > > I still do not understand how this patch set works. > > > (only me?) > > > > > > Please let me ask fundamental questions. > > > > > > > > > > > > I applied this series on top of Linus' tree, > > > and compiled for ARCH=arm64. > > > > > > I compared the kernel size with/without LTO. > > > > > > > > > > > > [1] No LTO (arm64 defconfig, CONFIG_LTO_NONE) > > > > > > $ llvm-size vmlinux > > > text data bss dec hex filename > > > 15848692 10099449 493060 26441201 19375f1 vmlinux > > > > > > > > > > > > [2] Clang LTO (arm64 defconfig + CONFIG_LTO_CLANG) > > > > > > $ llvm-size vmlinux > > > text data bss dec hex filename > > > 15906864 10197445 490804 26595113 195cf29 vmlinux > > > > > > > > > I compared the size of raw binary, arch/arm64/boot/Image. > > > Its size increased too. > > > > > > > > > > > > So, in my experiment, enabling CONFIG_LTO_CLANG > > > increases the kernel size. > > > Is this correct? > > > > Yes. LTO does produce larger binaries, mostly due to function > > inlining between translation units, I believe. The compiler people > > can probably give you a more detailed answer here. Without -mllvm > > -import-instr-limit, the binaries would be even larger. > > > > > One more thing, could you teach me > > > how Clang LTO optimizes the code against > > > relocatable objects? > > > > > > > > > > > > When I learned Clang LTO first, I read this document: > > > https://llvm.org/docs/LinkTimeOptimization.html > > > > > > It is easy to confirm the final executable > > > does not contain foo2, foo3... > > > > > > > > > > > > In contrast to userspace programs, > > > kernel modules are basically relocatable objects. > > > > > > Does Clang drop unused symbols from relocatable objects? > > > If so, how? > > > > I don't think the compiler can legally drop global symbols from > > relocatable objects, but it can rename and possibly even drop static > > functions. > > > Compilers can drop static functions without LTO. > Rather, it is a compiler warning > (-Wunused-function), so the code should be cleaned up. > > > > > This is why we need global wrappers for initcalls, for > > example, to have stable symbol names. > > > > Sami > > > > At first, I thought the motivation of LTO > was to remove unused global symbols, and > to perform further optimization. > > > It is true for userspace programs. > In fact, the example of > https://llvm.org/docs/LinkTimeOptimization.html > produces a smaller binary. > > > In contrast, this patch set produces a bigger kernel > because LTO cannot remove any unused symbol. > > So, I do not understand what the benefit is. > > > Is inlining beneficial? > I am not sure. > > > Documentation/process/coding-style.rst > "15) The inline disease" > mentions that inlining is not always > a good thing. > > > As a whole, I still do not understand > the motivation of this patch set. Clang produces faster code with LTO even if unused functions are not removed, and I'm not sure how many unused globals there really are in the kernel that aren't exported for modules. However, as I mentioned in the cover letter, we also need LTO for Control-Flow Integrity (CFI), which we have used in Pixel kernels for a couple of years now, and plan to use in more Android devices in future: https://clang.llvm.org/docs/ControlFlowIntegrity.html Sami
