Impact of RISC-V C Extension on Linux Kernel Size

[Chunyan Zhang]

Hi,

I was recently asked how much the RISC-V C extensions can shrink the
Linux kernel size. I looked it up online and could only roughly
predict a 25% reduction in theory. But I couldn't find more specific
data on the Linux kernel program. So I compiled a few kernel images
with or without RISC-V C extension and would like to share the data
and analysis. Maybe someone will also be interested.

The kernel version is v6.18-rc1, toolchain is
riscv64-glibc-ubuntu-22.04-gcc-nightly-2025.09.28-nightly.

Compile with C extension is the default, so to disable C extension on
riscv64, I hacked two Makefiles like below:

diff --git a/arch/riscv/Makefile b/arch/riscv/Makefile
index ecf2fcce2d92..e51810608695 100644
--- a/arch/riscv/Makefile
+++ b/arch/riscv/Makefile
@@ -61,7 +61,6 @@ endif
 riscv-march-$(CONFIG_ARCH_RV32I)       := rv32ima
 riscv-march-$(CONFIG_ARCH_RV64I)       := rv64ima
 riscv-march-$(CONFIG_FPU)              := $(riscv-march-y)fd
-riscv-march-$(CONFIG_RISCV_ISA_C)      := $(riscv-march-y)c
 riscv-march-$(CONFIG_RISCV_ISA_V)      := $(riscv-march-y)v

 ifneq ($(CONFIG_RISCV_ISA_C),y)
diff --git a/arch/riscv/kernel/Makefile b/arch/riscv/kernel/Makefile
index f60fce69b725..0c93c397fed0 100644
--- a/arch/riscv/kernel/Makefile
+++ b/arch/riscv/kernel/Makefile
@@ -46,6 +46,7 @@ endif
 always-$(KBUILD_BUILTIN) += vmlinux.lds

 obj-y  += head.o
+AFLAGS_head.o  += -march=rv64gc
 obj-y  += soc.o
 obj-$(CONFIG_RISCV_ALTERNATIVE) += alternative.o
 obj-y  += cpu.o


First with C extension (default):

    $ make defconfig
    $ time make vmlinux -j32
       real 4m27.219s
       user 57m28.761s
       sys 5m19.494s
    $ size vmlinux
        text     data     bss       dec      hex  filename
    13294359  6288266  423589  20006214  1314546  vmlinux

Without C extension:

    $ make clean
    $ time make vmlinux -j32
       real 4m34.450s
       user 58m52.848s
       sys 5m25.042s
    $ size vmlinux
        text     data     bss       dec      hex  filename
    17728589  6288586  423589  24440764  174efbc  vmlinux

Build time for both is close, and we can get 25% and 18% for text
section and vmlinux size reduction respectively with C extension
support.

I also get performance test results with perf on Pioneer Board, which
is running kernel 6.6.111,

With C extension,

Performance counter stats for 'netperf' (3 runs):

          9,966.81 msec task-clock                       #    0.995
CPUs utilized               ( +-  0.04% )
               780      context-switches                 #   78.260
/sec                        ( +-  1.74% )
                35      cpu-migrations                   #    3.512
/sec                        ( +-  3.30% )
               121      page-faults                      #   12.140
/sec                        ( +-  5.86% )
    19,924,390,238      cycles                           #    1.999
GHz                         ( +-  0.04% )
     7,681,821,013      instructions                     #    0.39
insn per cycle              ( +-  0.88% )
                 0      branches                         #    0.000
/sec
                 0      branch-misses

          10.01376 +- 0.00290 seconds time elapsed  ( +-  0.03% )

Without C extension,

Performance counter stats for 'netperf' (3 runs):

          9,967.98 msec task-clock                       #    0.995
CPUs utilized               ( +-  0.02% )
               772      context-switches                 #   77.448
/sec                        ( +-  1.78% )
                33      cpu-migrations                   #    3.311
/sec                        ( +-  4.63% )
               126      page-faults                      #   12.640
/sec                        ( +-  2.12% )
    19,925,181,546      cycles                           #    1.999
GHz                         ( +-  0.02% )
     6,639,648,698      instructions                     #    0.33
insn per cycle              ( +-  0.96% )
                 0      branches                         #    0.000
/sec
                 0      branch-misses

          10.01589 +- 0.00309 seconds time elapsed  ( +-  0.03% )

IPC from 0.33 (without C) to 0.39 (with C), IPC is improved by 18%.

Thanks,
Chunyan