On 6/22/19 7:55 AM, Jason Duerstock wrote:
> I was starting at the assembly from some of the Python source, and
> came across this (simplified) comparison:
>
> if (x > 2305843009213693951) {...}
>
> This is the same as:
>
> if (x > 0x1fffffffffffffff) {...}
>
> This is equivalent to:
>
> if (x >> 61) {...}
>
> More generally, we can rewrite
>
> if ( x > ((1 << z) -1)) { ...}
>
> as
>
> if ( x >> z ) { ... }
>
> This does not appear to currently be a gcc optimization. What is
> involved in adding it?
So first, when discussing this kind of stuff it's usually best to
actually have a compilable example. Fragments usually are insufficient.
Adding it to the RTL optimizers would be painful because of the need for
type information (this is only valid when X is unsigned, right?)
Adding it to the gimple optimizers is painful because the optimized form
is actually de-optimized on some targets (say embedded targets with weak
shifters) and querying the target costs/capabilities is generally
frowned upon in the gimple optimizers.
I think the combination of those two factors would tend to argue for an
implementation in the gimple->rtl expanders. You've still got type
information and you can query the backend for costing and capabilities.
cfgexpand::expand_gimple_cond might be a good place to start.
Another place to poke around would be expr:expand_expr_real_2, case
COND_EXPR.
Jeff
