Fix string cat in show method invalidations

[charleskawczynski]

This PR fixes a small typo for concatenating strings in the show method for invalidations.

Thanks for the great package!

[codecov[bot]]

Codecov Report

Merging #268 (71ab836) into master (d18b722) will decrease coverage by 4.41%. The diff coverage is 0.00%.

@@            Coverage Diff             @@
##           master     #268      +/-   ##
==========================================
- Coverage   86.87%   82.46%   -4.42%     
==========================================
  Files          16       16              
  Lines        2004     1893     -111     
==========================================
- Hits         1741     1561     -180     
- Misses        263      332      +69     

Impacted Files	Coverage Δ
src/invalidations.jl	87.04% <0.00%> (-3.12%)	:arrow_down:
src/invalidation_and_inference.jl	0.00% <0.00%> (-100.00%)	:arrow_down:
src/parcel_snoopi_deep.jl	87.74% <0.00%> (-1.84%)	:arrow_down:
src/deep_demos.jl	57.89% <0.00%> (-1.08%)	:arrow_down:
src/write.jl	91.30% <0.00%> (-1.01%)	:arrow_down:
src/parcel_snoopl.jl	89.47% <0.00%> (-0.53%)	:arrow_down:
SnoopCompileCore/src/snoopi_deep.jl	91.66% <0.00%> (-0.34%)	:arrow_down:
src/parcel_snoopi.jl	93.83% <0.00%> (-0.31%)	:arrow_down:
src/visualizations.jl	0.00% <0.00%> (ø)
... and 1 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update d18b722...71ab836. Read the comment docs.

[timholy]

Thanks! Do you have a reproducer for this? I generally attempt to capture the problem in a test before making the fix, to ensure no similar problems come up in the future.

[charleskawczynski]

I don't, but I can try to cook one up

[timholy]

Don't kill yourself for it, because it is indeed an obvious fix, but I thought I'd ask.

[charleskawczynski]

This is as far as I've gotten 😅:

using SnoopCompileCore

invalidations = @snoopr begin
    import BlockArrays
    import CUDA
end;

using SnoopCompile
trees = invalidation_trees(invalidations)

Also, this only "works" (results in an error) on Julia 1.6. I suppose inference on Julia 1.7 is better and we no longer hit this code path.

[timholy]

Thanks for the tip! I added a test that doesn't require those big packages. Here's how I captured the error: with SnoopCompile devved, I identified method that threw the error from the stacktrace and added this (Revise.jl makes this easy):

diff --git a/src/invalidations.jl b/src/invalidations.jl
index f8fd2db..37df189 100644
--- a/src/invalidations.jl
+++ b/src/invalidations.jl
@@ -173,7 +173,9 @@ end

 # We could use AbstractTrees here, but typically one is not interested in the full tree,
 # just the top method and the number of children it has
+const minvs = Ref{MethodInvalidations}()
 function Base.show(io::IO, methinvs::MethodInvalidations)
+    minvs[] = methinvs
     iscompact = get(io, :compact, false)::Bool

     print(io, methinvs.reason, " ")

This allowed me to extract (in the REPL0 the most recent input to the function via minvs = SnoopCompile.minvs[]. Then I was able to inspect it, identify which method definition in CUDA was responsible for the invalidation, and just mimic that.