Re: [rvm-research] missed optimization opportunity in SimpleEscape/ObjectReplacer

[David P G. <gr...@us...>]

Ondrej Lhotak <ol...@uw...> wrote on 10/07/2011 11:04:29 PM:
>
> OK, I think I've gotten to the bottom of this.
>
> After the first scalar replacement, the code pattern looks like this:
> t495a = null
> t495a = t28sa
> ...
> use t495a
>
> In all of my small examples, the use was in the *same* basic block as
> the defs. Therefore, LocalCopyProp replaced the use with a use of t28sa.
> This made *both* of the defs dead, and they were both subsequently
eliminated.
>
> In the interesting Scala code, however, the use was in a *different*
basic
> block than the two defs. LocalCopyProp therefore did not change the
> use. None of the transformations that are performed between passes of
scalar
> replacement remove the first def. As a result, t495a never becomes SSA,
> and the second pass of scalar replacement always ignores t495a, and
> therefore does nothing.
>
> To check this, I modified one of my small examples to force the use into
> a different basic block. With this modification, the second scalar
> replacement pass no longer did anything.
>
> The bottom line is that the multiple passes of scalar replacement are
> useful only in the case when all getfields of the replaced field are in
> the same basic block as the allocation and all of the putfields.
>
> That leaves me with one more question: is there a cheap and easy way to
> put just a single variable into SSA form (when the rest of the code is
> already mostly SSA)?

Cool.  Thanks for pushing on this.

Are the two defs in the same basic block? If so a quick and dirty solution
would be to do a local dead assignment elimination.

A better long term solution is probably for us to implement a classical
dense data flow global dead assignment elimination pass.  We can probably
afford to do this at O1.

We used to get this by going in & out of SSA form at O2, but the SSA
transformation was considered to be buggy enough that it got disabled
(moved to O3) in the stability push we did a few years back.  The intent
was always to bring it back, but we never found the time :(

--dave