JIT: Factor SSA's DFS and profile synthesis's loop finding #95251
Conversation
Factor out SSA's general DFS (that takes EH into account) and encapsulate it in a `FlowGraphDfsTree` class. Factor out profile synthesis's loop finding and encapsulate it in a `FlowGraphNaturalLoops` class. Switch construction of it to use the general DFS instead of the restricted one (that does not account for exceptional flow). Optimize a few things in the process: * Avoid storing loop blocks in a larger than necessary bit vector; store them starting from the loop header's postorder index instead. * Provide post-order and reverse post-order visitors for the loop blocks; switch profile synthesis to use this in a place No diffs are expected. A small amount of diffs are expected when profile synthesis is enabled due to the modelling of exceptional flow and also from handling unreachable predecessors (which would reject some loops as unnatural loops before). My future plans are to proceed to replace the loop representation of loops with this factored version, removing the lexicality requirement in the process, and hopefully fixing some of our deficiencies.
ghost
assigned 
dotnet-issue-labeler
bot
added
the
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch Issue DetailsFactor out SSA's general DFS (that takes EH into account) and encapsulate it in a Factor out profile synthesis's loop finding and encapsulate it in a Optimize a few things in the process:
No diffs are expected. A small amount of diffs are expected when profile synthesis is enabled due to the modelling of exceptional flow and also from handling unreachable predecessors (which would reject some loops as unnatural loops before). My future plans are to proceed to replace the loop representation of loops with this factored version, removing the lexicality requirement in the process, and hopefully fixing some of our deficiencies.
|
| // Find the exit edges | ||
| // | ||
| loop->VisitLoopBlocks([=](BasicBlock* loopBlock) { | ||
| loopBlock->VisitRegularSuccs(comp, [=](BasicBlock* succBlock) { |
There was a problem hiding this comment.
This probably should be VisitAllSuccs -- if you have a loop enclosed in a handler, then every block becomes an exit block. But I think I will look at it separately -- once we start using this elsewhere I think it becomes unlikely that we'll want to compute the exit and entry edges proactively here, and presumably profile synthesis only wants to consider the non EH exit edges as part of assigning probabilities. So at that point I think this will end up refactored a bit.
There was a problem hiding this comment.
Yeah, makes sense -- for profile purposes we will assume exceptions never happen.
| BasicBlock* const loopBlock = worklist.back(); | ||
| worklist.pop_back(); | ||
|
|
||
| for (FlowEdge* predEdge = comp->BlockPredsWithEH(loopBlock); predEdge != nullptr; |
There was a problem hiding this comment.
This one is the important one to switch to BlockPredsWithEH to recognize the full SCC of loops containing EH constructs.
With more general cycles the loop below that looks at predecessors could see uninitialized weights, so this is definitely needed.
|
cc @dotnet/jit-contrib PTAL @AndyAyersMS @BruceForstall Diffs. Just some small TP regressions that seem to come down to some different enregistration in the DFS lambda I added. Diffs with |
| // Find the exit edges | ||
| // | ||
| loop->VisitLoopBlocks([=](BasicBlock* loopBlock) { | ||
| loopBlock->VisitRegularSuccs(comp, [=](BasicBlock* succBlock) { |
There was a problem hiding this comment.
Yeah, makes sense -- for profile purposes we will assume exceptions never happen.
|
I'm going to merge this, but @BruceForstall please feel free to take a look when you have time and I'll try to address your feedback in a follow-up. |
Factor out SSA's general DFS (that takes EH into account) and encapsulate it in a
FlowGraphDfsTreeclass.Factor out profile synthesis's loop finding and encapsulate it in a
FlowGraphNaturalLoopsclass. Switch construction of it to use the general DFS instead of the restricted one (that does not account for exceptional flow).Optimize a few things in the process:
No diffs are expected. A small amount of diffs are expected when profile synthesis is enabled due to the modelling of exceptional flow and also from handling unreachable predecessors (which would reject some loops as unnatural loops before).
My future plans are to proceed to replace the loop representation of loops with this factored version, removing the lexicality requirement in the process, and hopefully fixing some of our deficiencies.