Currently, cmd/internal/obj/arm optimizes this instruction in sys_linux_arm.s:

sys_linux_arm.s:202          MOVW    $8(R13), R1  // timespec

into this ADD instruction:

    sys_linux_arm.s:202     0xe87b4 e28d1008        ADD $8, R13, R1

but misses an opportunity to optimize this move:

sys_linux_arm.s:261          MOVW    $-16(R1), R1

into a SUB instruction:

    sys_linux_arm.s:261     0xe8870 e59fb0b4        MOVW 0xb4(R15), R11
    sys_linux_arm.s:261     0xe8874 e081100b        ADD R11, R1, R1
    ...
    sys_linux_arm.s:324     0xe892c fffffff0        ?

Alternatively, there's not a lot of explicit "MOVW $-N(Rx), Ry" instructions in the ARM assembly, and the only compiler generated ones I've found are from cmd/internal/obj/arm's stacksplit function, so it's feasible to just change them all to use explicit SUB instructions instead.

Changing just stacksplit to use ASUB instead of AMOVW shaves ~5kB of text off cmd/go:

   text    data     bss     dec     hex filename
6666309  103960   85812 6856081  689d91 go.after
6671670  103960   85812 6861442  68b282 go.before