remove let from @setup_workload

[KristofferC]

This seems to have a bad effect on precompilation somehow.

I investigated this from JuliaSIMD/LoopVectorization.jl#490. Removing these let statements makes the regression go away.

[KristofferC]

I don't really understand the test error on nightly.

[timholy]

This seems to have a bad effect on precompilation somehow.

I think it's because the let must trigger inference and/or codegen before reaching the @compile_workload. If runtime-dispatched callees get compiled outside the "scope" of the snooping that it turns on, they are not guaranteed to be cached. (They might be, by the ordinary rules, but it's not guaranteed.)

I don't really understand the test error on nightly.

Nightly infers both a specialized and an unspecialized version without the let. The let is presumably helping inference avoid the need for the unspecialized version. Not sure what change made it specific to nightly, though.

[codecov]

Codecov Report

Merging #16 (90b1597) into main (302eeae) will decrease coverage by 0.08%.
The diff coverage is 100.00%.

❗ Current head 90b1597 differs from pull request most recent head 8dec582. Consider uploading reports for the commit 8dec582 to get more accurate results

@@            Coverage Diff             @@
##             main      #16      +/-   ##
==========================================
- Coverage   93.25%   93.18%   -0.08%     
==========================================
  Files           3        3              
  Lines          89       88       -1     
==========================================
- Hits           83       82       -1     
  Misses          6        6              

Impacted Files	Coverage Δ
src/workloads.jl	87.75% <100.00%> (-0.25%)	⬇️

[timholy]

Thanks for getting this out the door!