optimize (a little bit) faster #31338
Conversation
| @@ -780,8 +780,10 @@ function renumber_ssa2(val::SSAValue, ssanums::Vector{Any}, used_ssa::Vector{Int | |||
| if do_rename_ssa | |||
| val = ssanums[id] | |||
| end | |||
| if isa(val, SSAValue) && used_ssa !== nothing | |||
There was a problem hiding this comment.
Is there a performance advantage to avoiding the && here?
There was a problem hiding this comment.
yeah, inference can't see through &&, so it's best not to combine isa with anything else in the same expression
|
Optimizer parts look fine to me. @JeffBezanson should look at the subtyping improvements. |
| } | ||
| if (jl_obvious_subtype((jl_value_t*)temp, y, subtype) && *subtype) | ||
| return 1; | ||
| return 0; |
There was a problem hiding this comment.
Couldn't this be simplified to
return (jl_obvious_subtype((jl_value_t*)temp, y, subtype) && *subtype)?
There was a problem hiding this comment.
yes, but it would break the symmetry with the rest of the function
|
@nanosoldier |
|
Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. cc @ararslan |
This is a new subtype pre-process step that attempts to look for easily computable reasons that the result must be false (or true). Since it can avoid allocating an `stenv` and otherwise setting up the state needed for handling TypeVar constraints precisely, it is hoped to save time (on average) by filtering out easy cases.
help performance by removing bits of code highlighted by profiling as possible hot-spots.
|
@nanosoldier |
|
Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. cc @ararslan |
Simplifies some code patterns that the compiler can't optimize.
Adds a fast-path to subtype to attempt to handle common simple cases. We also still probably should duplicate parts of this fast path in the compiler (to handle the pattern
x isa AnySSAValue), since it is now heavily used by the optimizer code in various hot-paths.Also updates a few of the inlining costs to reflect that certain operations have non-zero expected cost in aggregate. This helps to put an eventual cap on the inlining of large functions that consist mostly of these operations.