potential fix for try - branch - catch

[maartenvd]

I must admit I don't fully understand the ssa! pass, but there is an issue where try ... can branch. handlers/catch_branches will be incorrect for those branches. Example:

met = IRTools.meta(Tuple{typeof(Base.Partr.multiq_size), Int8})
ir = IRTools.IR(met)

Again, I don't fully understand the pass, and this 'fix' may break other things!

I imagine that it is possible for a branch to jump back to an earlier branch, and catch_branches will then be completely wrong!

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 75.19%. Comparing base (8f5f50e) to head (8ca9f1f).
Report is 11 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #128      +/-   ##
==========================================
+ Coverage   73.61%   75.19%   +1.58%     
==========================================
  Files          16       16              
  Lines        1512     1532      +20     
==========================================
+ Hits         1113     1152      +39     
+ Misses        399      380      -19     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[Pangoraw]

but there is an issue where try ... can branch. handlers/catch_branches

yeah, the ssa! pass seems to be written with the assumption that there is only one leave per try region (+1 leave in the catch region) which is not the case when there is a return in the try.

Here is a minimal reproducer which you can add as a test:

function f(cond)
   try
       if cond return end
   finally
   end
end

I am not sure about this fix, though