fix some invalidations caused by FixedPointNumbers

[ranocha]

This is based on the following code (on Julia v1.8.3):

julia> using Pkg; Pkg.activate(temp=true); Pkg.add("FixedPointNumbers"); Pkg.develop("LoopVectorization")

julia> using SnoopCompileCore; invalidations = @snoopr(using LoopVectorization, FixedPointNumbers); using SnoopCompile

julia> length(uinvalidated(invalidations))
613

julia> trees = invalidation_trees(invalidations);

julia> trees[end]
inserting reduce_first(::typeof(Base.add_sum), x::FixedPoint) in FixedPointNumbers at ~/.julia/packages/FixedPointNumbers/HAGk2/src/FixedPointNumbers.jl:295 invalidated:
   backedges: 1: superseding reduce_first(::typeof(Base.add_sum), x) in Base at reduce.jl:402 with MethodInstance for Base.reduce_first(::typeof(Base.add_sum), ::Any) (166 children)

julia> ascend(trees[end].backedges[end])

With this PR,

julia> length(uinvalidated(invalidations))
513

The initial number of invalidations coming from this is bigger when loading something doing more precompilation involving LoopVectorization.jl.

[codecov]

Codecov Report

Base: 85.97% // Head: 86.08% // Increases project coverage by +0.10% 🎉

Coverage data is based on head (e01278b) compared to base (0bf0080).
Patch coverage: 66.66% of modified lines in pull request are covered.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #448      +/-   ##
==========================================
+ Coverage   85.97%   86.08%   +0.10%     
==========================================
  Files          37       37              
  Lines        9379     9379              
==========================================
+ Hits         8064     8074      +10     
+ Misses       1315     1305      -10     

Impacted Files	Coverage Δ
src/parse/memory_ops_common.jl	88.94% <0.00%> (ø)
src/modeling/graphs.jl	88.96% <100.00%> (ø)
src/codegen/lower_constant.jl	72.76% <0.00%> (-2.35%)	⬇️
src/codegen/lower_memory_common.jl	86.23% <0.00%> (+0.36%)	⬆️
src/modeling/determinestrategy.jl	93.29% <0.00%> (+0.51%)	⬆️
src/modeling/costs.jl	53.05% <0.00%> (+0.93%)	⬆️
src/simdfunctionals/mapreduce.jl	90.32% <0.00%> (+7.52%)	⬆️

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[chriselrod]

Would making vals a Number[] instead of an Any[] also help at all?

[chriselrod]

The test failures are likely related to StrideArraysCore v0.4.0. I pushed it out prematurely.

Neither the DiffEq stack or StrideArrays.jl have releases supporting it yet (for good reason), so it's mostly been unnoticed.

[ranocha]

Would making vals a Number[] instead of an Any[] also help at all?

No, I don't think so. The invalidations come from FixedPointNumbers.jl, which provides new Number types with specializations.

[ranocha]

@chriselrod Would you mind making a new release of LoopVectorization.jl including these invalidation fixes?