concerning a few points of the discussion about code-size reduction
(Minutes available at
LLVM 2020 Code Size BoF - Google Docs):
Do we need different levels like -Os1, -Os2, -Os3?
Together with a colleague, Anderson Faustino, we spend some time
finding optimization sequences that are good for code-size reduction.
This report gives you 22 such sequences, all short, that tend to
improve upon Os or Oz on typical benchmarks:
What code size benchmarks can we use to measure patches?
I'd suggest AnghaBench (PageAngha). I've talked
about it here before. It's a collection with One Million compilable C
programs. That's the collection that we have used to find the
optimization sequences in that report. We shall be presenting a paper
about it at CGO next February.