Re: [PATCH] LoongArch: improve 64-bit bitwise AND operation

在 2026/8/10 09:33, Ben Shi 写道:
> Shall we still use `define_insn_and_split` at first ? At least it does 
> emit more efficient assembly.

It may not as I've explained.  If the AND operation is in a loop, 
materializing the mask out before
the loop and then reuse it in the loop with a normal and instruction 
would be more efficient.  And
we don't know if the uarch has a shunt allowing lu12i.w + addi.d + 
lu32i.d + lu52i.d + and to finish
in one cycle instead of two cycles of the bstrins pair.

We may need some benchmark number (of real or well-known workloads like 
SPEC CPU) to
demonstrate which is better, instead of using eyes.

> The solution you suggested involved changes to immediate 
> materializing, which seems tricky, and I need more time to calculate a 
> best form.

We didn't suggest that.  Jeff said it may be better to do that in the 
expand pass and I said the
cost model does not like your change and will undo that (if it's done in 
expand).

Using split just bypasses the code model and it will make the GCC code 
base logically inconsistent
(the cost model says materializing the const is better, but the split 
rule says otherwise).  Yes it may
be already inconsistent at some places now but don't worsen the situation.

And we still have 5 months before feature freeze (stage 4) of GCC 17.  
To me there's plenty of time
to gather the benchmark number and perhaps adjust the cost model if needed.