Make the optimization level configurable.

[maleadt]

We also need to make sure that we have an optimization mode (probably -O1?) that only lightly optimizes, e.g. not doing loop unrolling or unswitching, but just optimizing Julia's typical IR patterns away. This is for use with shader compilers (i.e., Metal, SPIR-V, NVVM) which do their own optimization.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 82.74%. Comparing base (0c9292f) to head (fc36b0f).
Report is 1 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #577      +/-   ##
==========================================
- Coverage   82.86%   82.74%   -0.12%     
==========================================
  Files          24       24              
  Lines        3361     3362       +1     
==========================================
- Hits         2785     2782       -3     
- Misses        576      580       +4     

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

[maleadt]

Turns out both O0 and O1 introduce additional allocations in the generated IR. Guess we're more dependent on the alloc_otp pass than I assumed...