-
-
Notifications
You must be signed in to change notification settings - Fork 5.5k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
inlining: bail out unless match.spec_types <: match.method.sig
#53720
Conversation
@nanosoldier |
Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. |
As Jameson pointed out in the link below, while the union-split handles cases when there are uncovered matches, sometimes the expected condition `spec_types <: method.sig` that the union-split algorithm relies on isn't met in such cases, which caused issues like #52644. This commit fixes the problem by adding explicit checks for such cases. Note that this is based on #52092. The extra handling for single method match unmatched static parameters based on `only_method` that was removed in JuliaLang/#52092 has been ineffective and would otherwise cause problematic inlining on this PR. We'll likely need to come back to this later and figure out a safe and effective way to deal with such cases in the future when the handling for either case turns out to be necessary. - closes #52644 - xref: <#53600 (review)>
end | ||
elseif !(match.spec_types <: match.method.sig) # the requirement for correct union-split | ||
handled_all_cases = false | ||
else | ||
handled_all_cases &= handle_any_const_result!(cases, | ||
result, match, argtypes, info, flag, state; allow_abstract=true, allow_typevars=false) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
result, match, argtypes, info, flag, state; allow_abstract=true, allow_typevars=false) | |
result, match, argtypes, info, flag, state; allow_abstract=handled_all_cases, allow_typevars=false) |
This can only permit later abstract calls if we proved that all previous calls were disjoint from this. One conservative way to express that requirement is to only allow isdispatchtuple here.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Isn't this filtering enough for this?
julia/base/compiler/ssair/inlining.jl
Lines 1391 to 1393 in 63e75e8
elseif !isempty(cases) | |
# if we've not seen all candidates, union split is valid only for dispatch tuples | |
filter!(case::InliningCase->isdispatchtuple(case.sig), cases) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Additionally, if nunion != 1
, then the list of matches cannot be assumed to be sorted, so we can only allow the abstract calls that we can prove are disjoint from the other union elements, since we do not attempt to merge-sort them to combine them (the merge sort might fail) but we simply concat the lists.
result, match, argtypes, info, flag, state; allow_abstract=true, allow_typevars=false) | |
result, match, argtypes, info, flag, state; allow_abstract=handled_all_cases && nunion == 1, allow_typevars=false) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Ah, okay, I wasn't expecting we would over-estimate this then filter it later. That late filter is actually a bit more strict than necessary then, since it doesn't consider the ordering. But that does at least seem valid.
) As Jameson pointed out in the link below, while the union-split handles cases when there are uncovered matches, sometimes the expected condition `spec_types <: method.sig` that the union-split algorithm relies on isn't met in such cases, which caused issues like #52644. This commit fixes the problem by adding explicit checks for such cases. Note that this is based on #52092. The extra handling for single method match unmatched static parameters based on `only_method` that was removed in JuliaLang/#52092 has been ineffective and would otherwise cause problematic inlining on this PR. We'll likely need to come back to this later and figure out a safe and effective way to deal with such cases in the future when the handling for either case turns out to be necessary. - closes #52644 - xref: <#53600 (review)>
Backported PRs: - [x] #39071 <!-- Add a lazy `logrange` function and `LogRange` type --> - [x] #51802 <!-- Allow AnnotatedStrings in log messages --> - [x] #53369 <!-- Orthogonalize re-indexing for FastSubArrays --> - [x] #48050 <!-- improve `--heap-size-hint` arg handling --> - [x] #53482 <!-- add IR encoding for EnterNode --> - [x] #53499 <!-- Avoid compiler warning about redefining jl_globalref_t --> - [x] #53507 <!-- update staled `Core.Compiler.Effects` documentation --> - [x] #53408 <!-- task splitting: change additive accumulation to multiplicative --> - [x] #53523 <!-- add back an alias for `check_top_bit` --> - [x] #53377 <!-- add _readdirx for returning more object info gathered during dir scan --> - [x] #53525 <!-- fix InteractiveUtils call in Base.runtests on failure --> - [x] #53540 <!-- use more efficient `_readdirx` for tab completion --> - [x] #53545 <!-- use `_readdirx` for `walkdir` --> - [x] #53551 <!-- revert "Add @create_log_macro for making custom styled logging macros (#52196)" --> - [x] #53554 <!-- Always return a value in 1-d circshift! of abstractarray.jl --> - [x] #53424 <!-- yet more atomics & cache-line fixes on work-stealing queue --> - [x] #53571 <!-- Update Documenter to v1.3 for inventory writing --> - [x] #53403 <!-- Move parallel precompilation to Base --> - [x] #53589 <!-- add back `unsafe_convert` to pointer for arrays --> - [x] #53596 <!-- build: remove extra .a file --> - [x] #53606 <!-- fix error path in `precompilepkgs` --> - [x] #53004 <!-- Unexport with, at_with, and ScopedValue from Base --> - [x] #53629 <!-- typo fix in scoped values docs --> - [x] #53630 <!-- sroa: Fix incorrect scope counting --> - [x] #53598 <!-- Use Base parallel precompilation to build stdlibs --> - [x] #53649 <!-- precompilepkgs: package in boths deps and weakdeps are in fact only weak --> - [x] #53671 <!-- Fix bootstrap Base precompile in cross compile configuration --> - [x] #52125 <!-- Load Pkg if not already to reinstate missing package add prompt --> - [x] #53602 <!-- Handle zero on arrays of unions of number types and missings --> - [x] #53516 <!-- permit NamedTuple{<:Any, Union{}} to be created --> - [x] #53643 <!-- Bump CSL to 1.1.1 to fix libgomp bug --> - [x] #53679 <!-- move precompile workload back from Base --> - [x] #53663 <!-- add isassigned methods for reinterpretarray --> - [x] #53662 <!-- [REPL] fix incorrectly cleared line after completions accepted --> - [x] #53611 <!-- Linalg: matprod_dest for Diagonal and adjvec --> - [x] #53659 <!-- fix #52025, re-allow all implicit pointer casts in cconvert for Array --> - [x] #53631 <!-- LAPACK: validate input parameters to throw informative errors --> - [x] #53628 <!-- Make some improvements to the Scoped Values documentation. --> - [x] #53655 <!-- Change tbaa of ptr_phi to tbaa_value --> - [x] #53391 <!-- Default to the medium code model in x86 linux --> - [x] #53699 <!-- Move `isexecutable, isreadable, iswritable` to `filesystem.jl` --> - [x] #41232 <!-- Fix linear indexing for ReshapedArray if the parent has offset axes --> - [x] #53527 <!-- Enable analyzegc checks for try catch and fix found issues --> - [x] #52092 - [x] #53682 <!-- Increase build precompilation --> - [x] #53720 - [x] #53553 <!-- typeintersect: fix `UnionAll` unaliasing bug caused by innervars. --> Contains multiple commits, manual intervention needed: - [ ] #53305 <!-- Propagate inbounds in isassigned with CartesianIndex indices --> Non-merged PRs with backport label: - [ ] #53736 <!-- fix literal-pow to return the right type when the base is -1 --> - [ ] #53707 <!-- Make ScopedValue public --> - [ ] #53696 <!-- add invokelatest to on_done callback in bracketed paste --> - [ ] #53660 <!-- put Logging back in default sysimage --> - [ ] #53509 <!-- revert moving "creating packages" from Pkg.jl --> - [ ] #53452 <!-- RFC: allow Tuple{Union{}}, returning Union{} --> - [ ] #53402 <!-- Add `jl_getaffinity` and `jl_setaffinity` --> - [ ] #52694 <!-- Reinstate similar for AbstractQ for backward compatibility --> - [ ] #51928 <!-- Styled markdown, with a few tweaks --> - [ ] #51816 <!-- User-themable stacktraces --> - [ ] #51811 <!-- Make banner size depend on terminal size --> - [ ] #51479 <!-- prevent code loading from lookin in the versioned environment when building Julia -->
) As Jameson pointed out in the link below, while the union-split handles cases when there are uncovered matches, sometimes the expected condition `spec_types <: method.sig` that the union-split algorithm relies on isn't met in such cases, which caused issues like #52644. This commit fixes the problem by adding explicit checks for such cases. Note that this is based on #52092. The extra handling for single method match unmatched static parameters based on `only_method` that was removed in JuliaLang/#52092 has been ineffective and would otherwise cause problematic inlining on this PR. We'll likely need to come back to this later and figure out a safe and effective way to deal with such cases in the future when the handling for either case turns out to be necessary. - closes #52644 - xref: <#53600 (review)> (cherry picked from commit b730d33)
As Jameson pointed out in the link below, while the union-split handles cases when there are uncovered matches, sometimes the expected condition
spec_types <: method.sig
that the union-split algorithm relies on isn't met in such cases, which caused issues like #52644. This commit fixes the problem by adding explicit checks for such cases. Note that this is based on #52092. The extra handling for single method match unmatched static parameters based ononly_method
that was removed in JuliaLang/#52092 has been ineffective and would otherwise cause problematic inlining on this PR. We'll likely need to come back to this later and figure out a safe and effective way to deal with such cases in the future when the handling for either case turns out to be necessary.DataType
method withUnionAll
argument causes segfault #52644