This is an attempt at cleaning up some ARM inline assembly, namely: the
MRC/MCR P15 instructions, with two objectives:
1. Getting rid of some changes that are suboptimal, including a recent
workaround for kirkwood. I am referring to commits forcing -marm (arm32
instructions) on whole C files just because they use the MCR/MRC
instruction in inline assembly and those instructions are not supported
in Thumb-1, as well as patches disabling LTO on select files for pretty
much the same reason. Those commits are the following:
62e92077a893 ("arm: support Thumb-1 with CONFIG_SYS_THUMB_BUILD")
8f9696510afc ("ARM: make LTO available")
e5fc9037dd33 ("ARM: fix LTO build for some thumb-interwork cases")
410d59095a9f ("arm: kirkwood: fix freeze on boot")
2. Hopefully get smaller binaries for Arm platforms that use Thumb and/
or LTO, at the cost of marginal increases (say a few dozens of bytes at
most) for the other aArm boards due to not inlining a few functions. And
of course no change for non-arm architectures.
There are four patches in this series:
- The first one reverts a change which does solve a linker error but
does not work as [1] has shown.
- The second patch moves inline assembly into separate .S files
- With that in place, it is possible to remove the -marm overrides as
well as the C flags overrides selectively disabling LTO. This is what
the third patch does.
- The fourth patch is not directly related, but is trying to reduce the
binary size further for platforms that compile C with
-ffunction-sections -fdata-sections so that the linker can eliminate
unreferenced code and data with --gc-sections. I noticed that .S files
do not follow the same convention. By modifying the ENTRY, WEAK and
ENDPROC macros, each assembly function is emitted in its own section
(.text.<name>) like the C compiler would do if it were C code compiled
with -ffunction-sections, thus allowing dead code elimination for .S
too. The savings are not much though (100 or 200 bytes most of the
time), and it makes a couple of boards significantly bigger
(imx8mn_beacon_fspi: all +16196 data +16304; stm32* as well as
imxrt1170-evk: all +2K text +2K). This needs further investigation.
See [6] for details.
Now for the results:
- CI status can be seen at [2]. All good.
- buildman output showing size differences before and after the series:
$ unbuffer tools/buildman/buildman -b lto-fixes-squashed -c 2 -sS \
| ansi2html >buildman-series-squashed-1e2c64f1537-size-summary.html
See [3] (arch averages)
$ unbuffer tools/buildman/buildman -b lto-fixes-squashed -c 2 -sSd \
| ansi2html >buildman-series-squashed-1e2c64f1537-size-details.html
See [4] (per board)
- buildman output showing size differences for each patch:
$ unbuffer tools/buildman/buildman -b lto-fixes -c 4 -sS \
| ansi2html >buildman-series-1e2c64f1537-size-summary.html
See [5] (arch averages)
$ unbuffer tools/buildman/buildman -b lto-fixes -c 4 -sSd \
| ansi2html >buildman-series-1e2c64f1537-size-details.html
See [6] (per board)
[1] https://lists.denx.de/pipermail/u-boot/2025-June/592682.html
[2] https://source.denx.de/u-boot/custodians/u-boot-net/-/pipelines/26972
[3]
https://people.linaro.org/~jerome.forissier/u-boot-07-Jul-2025/buildman-series-squashed-1e2c64f1537-size-summary.html
[4]
https://people.linaro.org/~jerome.forissier/u-boot-07-Jul-2025/buildman-series-squashed-1e2c64f1537-size-details.html
[5]
https://people.linaro.org/~jerome.forissier/u-boot-07-Jul-2025/buildman-series-1e2c64f1537-size-summary.html
[6]
https://people.linaro.org/~jerome.forissier/u-boot-07-Jul-2025/buildman-series-1e2c64f1537-size-details.html
Jerome Forissier (4):
Revert "arm: asm/system.h: mrc and mcr need .arm if __thumb2__ is not
set"
arm: move inline assembly CP15 instructions to separate .S files
arm: do not force -marm on some C files and allow LTO everywhere
linkage: use per-function section in ENTRY, WEAK and ENDPROC
arch/arm/cpu/arm926ejs/Makefile | 7 +-
arch/arm/cpu/arm926ejs/cache.c | 32 ++----
arch/arm/cpu/arm926ejs/cp15.S | 46 ++++++++
arch/arm/cpu/arm926ejs/cpu.c | 13 +--
arch/arm/cpu/armv7/lowlevel_init.S | 4 -
arch/arm/cpu/armv7/ls102xa/psci.S | 2 +-
arch/arm/cpu/armv7/nonsec_virt.S | 16 +--
arch/arm/cpu/armv7/psci.S | 118 ++++++++++-----------
arch/arm/cpu/armv7/start.S | 20 ++--
arch/arm/cpu/armv8/cache.S | 26 -----
arch/arm/cpu/armv8/psci.S | 22 ++--
arch/arm/cpu/armv8/tlb.S | 2 -
arch/arm/cpu/armv8/transition.S | 8 --
arch/arm/include/asm/system.h | 36 ++-----
arch/arm/lib/Makefile | 17 ++-
arch/arm/lib/ashldi3.S | 8 +-
arch/arm/lib/ashrdi3.S | 8 +-
arch/arm/lib/bitops.S | 8 --
arch/arm/lib/cache-cp15.c | 62 ++++-------
arch/arm/lib/cache.c | 7 +-
arch/arm/lib/cp15.S | 92 ++++++++++++++++
arch/arm/lib/crt0.S | 4 +-
arch/arm/lib/div64.S | 2 -
arch/arm/lib/lib1funcs.S | 36 ++-----
arch/arm/lib/lshrdi3.S | 8 +-
arch/arm/lib/muldi3.S | 8 +-
arch/arm/lib/relocate.S | 6 +-
arch/arm/lib/semihosting.S | 2 -
arch/arm/lib/setjmp.S | 6 --
arch/arm/lib/setjmp_aarch64.S | 6 --
arch/arm/lib/uldivmod.S | 2 -
arch/arm/mach-imx/mx5/lowlevel_init.S | 4 +-
arch/arm/mach-kirkwood/Makefile | 8 +-
arch/arm/mach-kirkwood/cp15.S | 13 +++
arch/arm/mach-kirkwood/include/mach/cpu.h | 13 +--
arch/arm/mach-omap2/omap3/lowlevel_init.S | 36 +++----
arch/arm/mach-renesas/lowlevel_init_gen3.S | 6 +-
arch/arm/mach-tegra/psci.S | 12 +--
arch/riscv/lib/memcpy.S | 6 +-
arch/riscv/lib/memmove.S | 7 +-
arch/riscv/lib/memset.S | 6 +-
arch/riscv/lib/semihosting.S | 2 -
arch/riscv/lib/setjmp.S | 6 --
common/Makefile | 4 -
include/linux/linkage.h | 20 +++-
45 files changed, 387 insertions(+), 390 deletions(-)
create mode 100644 arch/arm/cpu/arm926ejs/cp15.S
create mode 100644 arch/arm/lib/cp15.S
create mode 100644 arch/arm/mach-kirkwood/cp15.S
--
2.43.0
base-commit: 7027b445cc0bfb86204ecb1f1fe596f5895048d9
branch: lto-fixes
