[llvm-dev] Disable if with store to select optimization pass

[Craig Topper via llvm-dev]

If its the conversion I know of, try  -simplifycfg-merge-cond-stores=false

~Craig

On Mon, Apr 24, 2017 at 3:34 AM, Micha Reiser via llvm-dev <
llvm-dev at lists.llvm.org> wrote:

> Hy everyone
>
> I'm writing an LLVM frontend that compiles a subset of TypeScript to
> WebAssembly using the new experimental Web Assembly Target. and LLVM opt to
> optimize the output. As part of my work, I compare the runtime of the
> generated Web Assembly implementation to the plain JavaScript version.
> During this evaluation, I noticed that LLVM opt converts if statements with
> a store to a select statement to reduce the number of branches. E.g., the
> following
>
> for (let j = i; j < points.length; j += 2) {
>   const distance = euclideanDistance(currentX, currentY, points[j],
> points[j+1]);
>
>   if (distance < shortestDistance) {
>     shortestDistance = distance;
>     nearestIndex = j;
>   }
> }
>
> collapses to (the if is removed)
>
> for (let j = i; j < points.length; j += 2) {
>   const distance = euclideanDistance(currentX, currentY, points[j],
> points[j+1]);
>
>   let shorter = distance < shortestDistance;
>   shortestDistance shorter ? distance : shortestDistance;
>   nearestIndex = shorter ? j : nearestIndex;
> }
>
> My benchmarks show that this optimization has a negative impact on
> performance (at least if there are two or more assignments). Therefore, I
> would like to disable this optimization pass or at least tune it only to
> collapse the if statement if there is only a single assignment.
>
> Is there a way to disable this optimization pass or fine tune it to define
> how many statements are allowed to be collapsed?
>
> Thank you  very much,
> Micha
>
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170424/dc2d997d/attachment-0001.html>