optimize compute_basic_blocks a bit
#45364
Conversation
By scalar-folding `basic_blocks_starts(stmts)::BitSet`
|
Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. |
| block_for_inst(cfg::CFG, inst::Int) = block_for_inst(cfg.index, inst) | ||
|
|
||
| function basic_blocks_starts(stmts::Vector{Any}) | ||
| @inline function basic_blocks_starts(stmts::Vector{Any}) |
There was a problem hiding this comment.
Out of curiosity, why is it a good idea to force inlining of such a large function?
There was a problem hiding this comment.
We can perform SROA optimization by replacing BitSet created by basic_blocks_starts with "scalar"-values (internal fields of BitSet in this case) if we inline basic_blocks_starts into compute_basic_blocks.
Ideally this inlining should be performed automatically with some advanced inlining cost model, but at this moment our inlining cost model only accounts for simple code complexity and doesn't reason about SROA-profitability at all. With EscapeAnalysis.jl merged this kind of optimization is now possible, but it still doesn't happen yet.
By scalar-folding
basic_blocks_starts(stmts)::BitSet.@nanosoldier
runbenchmarks("inference", vs=":master")