Re: [PATCH bpf-next] bpf: arm64: Redefine MOV consistent with arch insn

[Will Deacon]

On Wed, Mar 31, 2021 at 05:22:18PM +0800, Jianlin Lv wrote:
> On Tue, Mar 30, 2021 at 5:31 PM Will Deacon <will@xxxxxxxxxx> wrote:
> >
> > On Tue, Mar 30, 2021 at 03:42:35PM +0800, Jianlin Lv wrote:
> > > A64_MOV is currently mapped to Add Instruction. Architecturally MOV
> > > (register) is an alias of ORR (shifted register) and MOV (to or from SP)
> > > is an alias of ADD (immediate).
> > > This patch redefines A64_MOV and uses existing functionality
> > > aarch64_insn_gen_move_reg() in insn.c to encode MOV (register) instruction.
> > > For moving between register and stack pointer, rename macro to A64_MOV_SP.
> >
> > What does this gain us? There's no requirement for a BPF "MOV" to match an
> > arm64 architectural "MOV", so what's the up-side of aligning them like this?
> 
> According to the description in the Arm Software Optimization Guide,
> Arithmetic(basic) and Logical(basic) instructions have the same
> Exec Latency and Execution Throughput.
> This change did not bring about a performance improvement.
> The original intention was to make the instruction map more 'natively'.

I think we should leave the code as-is, then. Having a separate MOV_SP
macro s confusing and, worse, I worry that somebody passing A64_SP to
A64_MOV will end up using the zero register.

Will