Re: [Patch,tree-optimization]: Add new path Splitting pass on tree ssa representation

[Jeff Law <law at redhat dot com>]

On 01/14/2016 01:55 AM, Jeff Law wrote:
[ Replying to myself again, mostly to make sure we've got these thoughts 
in the archives. ]
> Anyway, going back to adpcm_decode, we do end up splitting this path:
>
>   # vpdiff_12 = PHI <vpdiff_11(12), vpdiff_50(13)>
>    if (sign_41 != 0)
>      goto <bb 15>;
>    else
>      goto <bb 16>;
> ;;    succ:       15
> ;;                16
>
> ;;   basic block 15, loop depth 1
> ;;    pred:       14
>    valpred_51 = valpred_76 - vpdiff_12;
>    goto <bb 17>;
> ;;    succ:       17
>
> ;;   basic block 16, loop depth 1
> ;;    pred:       14
>    valpred_52 = vpdiff_12 + valpred_76;
> ;;    succ:       17
>
> ;;   basic block 17, loop depth 1
> ;;    pred:       15
> ;;                16
>    # valpred_7 = PHI <valpred_51(15), valpred_52(16)>
>    _85 = MAX_EXPR <valpred_7, -32768>;
>    valpred_13 = MIN_EXPR <_85, 32767>;
>    step_53 = stepsizeTable[index_62];
>    outp_54 = outp_69 + 2;
>    _55 = (short int) valpred_13;
>    MEM[base: outp_54, offset: -2B] = _55;
>    if (outp_54 != _74)
>      goto <bb 20>;
>    else
>      goto <bb 18>;
>
> This doesn't result in anything particularly interesting/good AFAICT. We
> propagate valpred_51/52 into the use in the MAX_EXPR in the duplicate
> paths, but that doesn't allow any further simplification.
So with the heuristic I'm poking at, this gets rejected.  Essentially it 
doesn't think it's likely to expose CSE/DCE opportunities (and it's 
correct).  The number of statements in predecessor blocks that feed 
operands in the to-be-copied-block is too small relative to the size of 
the to-be-copied-block.


> Ajit, can you confirm which of adpcm_code or adpcm_decode where path
> splitting is showing a gain?  I suspect it's the former but would like
> to make sure so that I can adjust the heuristics properly.
I'd still like to have this answered when you can Ajit, just to be 100% 
 that it's the path splitting in adpcm_code that's responsible for the 
improvements you're seeing in adpcm.

jeff