JIT: Use post order computed by SSA in VN #94623
Conversation
VN tries hard to dynamically compute a reverse post-order to visit the flow graph in. However, SSA has already computed such an order, so simply pass this along to VN instead. A few positive diffs are expected. As Andy has pointed out recently the "dynamic RPO" VN was doing does not necessarily result in an actual RPO, so using SSA's order is expected to be a better order than what VN ends up with today.
dotnet-issue-labeler
bot
added
the
area-CodeGen-coreclr
ghost
assigned 
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch Issue DetailsVN tries hard to dynamically compute a reverse post-order to visit the flow graph in. However, SSA has already computed such an order, so simply pass this along to VN instead. A few positive diffs are expected. As Andy has pointed out recently the "dynamic RPO" VN was doing does not necessarily result in an actual RPO, so using SSA's order is expected to be a better order than what VN ends up with today.
|
| @@ -769,7 +769,6 @@ bool Compiler::fgRemoveDeadBlocks() | |||
| // Notes: | |||
| // Each block's `bbPreorderNum` and `bbPostorderNum` is set. | |||
| // The `fgBBReversePostorder` array is filled in with the `BasicBlock*` in reverse post-order. | |||
| // This algorithm only pays attention to the actual blocks. It ignores any imaginary entry block. | |||
There was a problem hiding this comment.
I wasn't totally sure what this was referring to... but today we do ensure all blocks are visited below, so it seemed odd.
There was a problem hiding this comment.
SSA creates a stack-allocated BB0 when it computes dominators, I suspect this is the "imaginary entry block."
There was a problem hiding this comment.
Hmm, might be. Still, not sure it is too relevant for this computation.
|
/azp run runtime-coreclr jitstress, runtime-coreclr libraries-jitstress, Fuzzlyn |
|
Azure Pipelines successfully started running 3 pipeline(s). |
|
cc @dotnet/jit-contrib PTAL @AndyAyersMS Some minor positive diffs, both in codegen and TP. |
Good idea, will think about how to reuse that. |
|
Possibly related improvement: dotnet/perf-autofiling-issues#24588. |
VN tries hard to dynamically compute a reverse post-order to visit the flow graph in. However, SSA has already computed such an order, so simply pass this along to VN instead.
A few positive diffs are expected. As Andy has pointed out recently the "dynamic RPO" VN was doing does not necessarily result in an actual RPO, so using SSA's order is expected to be a better order than what VN ends up with today.