Re: [Qemu-devel] [PATCH 02/13] target-arm: A64: add support for logical (shifted register)

[Richard Henderson]

On 12/06/2013 10:36 PM, Alex Bennée wrote:
>> While correct, surely better to work with tcg and select on opc:invert to
>> generate andc/orc/eqv?
> 
> Shouldn't the TCG optimiser/back-end just be smart enough to figure it
> out? It seems clearer to express the tcg ops in terms of the front-end's
> meaning?

No, the TCG optimizer is really quite stupid.  It only does constant folding
and dead code elimination.  No peepholing or combination sorts of opts.

>> Also, isn't MOV (register) canonical for ORR (rn=31 && shift_amount=0), and 
>> MVN
>> (register) canonical for ORN (rn=31 && shift_amount=0), and both therefore 
>> also
>> worth a special case?
> 
> I suspect I'm being overly cheeky to expect the optimiser to detect and
> optimise for that case as the ZR is a const ;-)

It would.  But since register-register move is a rather common operation, it
will pay off to not require the optimizer to clean up that special case.

Thus I only recommend special casing the official aliases, not any operation
that could mathematically be considered an identity.


r~