JIT: more likelihood checking #99541
Conversation
Check has likelihood and likelihood sum consistency all the way until rationalize. Several phases required a bit of revision, notably the multi-guess type test expansion (similar to GDV) and optimize bools. Fix carry-over issue from previous round where block and edge scaling in loop cloning used different factors. Contributes to dotnet#93020
ghost
added
the
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch |
|
Arm failure: This PR passed in draft form, so issue is likely unrelated. |
Seems like an instance of #86273, already known to build analysis. |
| FlowEdge* oldEdge = bSrc->GetFalseEdge(); | ||
| jmpBlk = fgNewBBafter(BBJ_ALWAYS, bSrc, true); | ||
| FlowEdge* const oldEdge = bSrc->GetFalseEdge(); | ||
| // Access the likelihood of oldEdge before |
There was a problem hiding this comment.
Good catch, I probably should've paid more attention to the diffs this likelihood weirdness would've caused when I changed SetTargetEdge to set the likelihood.
|
|
||
| // review: we assume length check always succeeds?? | ||
| trueEdge->setLikelihood(1.0); | ||
| falseEdge->setLikelihood(0.0); |
There was a problem hiding this comment.
cc @dotnet/jit-contrib -- what do you think?
| // Predecessor is the nullcheck, control reaches on false. | ||
| // | ||
| curTypeCheckBb->inheritWeight(nullcheckBb); | ||
| curTypeCheckBb->scaleBBWeight(nullcheckBb->GetFalseEdge()->getLikelihood()); |
There was a problem hiding this comment.
Once edge likelihood propagation is done, we'll probably want to combine these two methods into one, right?
There was a problem hiding this comment.
I will probably add a float overload for inherit weight. I might try and redo the old GDV likelihoods so they're [0..1] instead of [0..100].
Not sure yet. There are some. Looking at lower's switch expansion now. |
|
Only a handful of diffs. |
Check has likelihood and likelihood sum consistency all the way until rationalize. Several phases required a bit of revision, notably the multi-guess type test expansion (similar to GDV) and optimize bools. Fix carry-over issue from previous round where block and edge scaling in loop cloning used different factors. Contributes to dotnet#93020
2 participants
Check has likelihood and likelihood sum consistency all the way until rationalize.
Several phases required a bit of revision, notably the multi-guess type test expansion (similar to GDV) and optimize bools.
Fix carry-over issue from previous round where block and edge scaling in loop cloning used different factors.
Contributes to #93020
(Continuation of #99460)
Diffs