flang-new driver does not enable some LLVM optimizations

[vzakhari]

LLVM PassBuilder relies on PipelineTuningOptions to enable some LLVM optimizations. For example, clang enables SLP vectorization at higher optimization levels. flang-new does not do this.

I think we need to align the optimization pipelines between clang and flang-new.

[llvmbot]

@llvm/issue-subscribers-flang-driver

Author: Slava Zakharin (vzakhari)

LLVM [PassBuilder](https://llvm.org/doxygen/classllvm_1_1PassBuilder.html) relies on `PipelineTuningOptions` to enable some LLVM optimizations. For example, `clang` [enables](https://github.com/llvm/llvm-project/blob/a8450619358c3de23b72aef7c1fd932bca112257/clang/lib/CodeGen/BackendUtil.cpp#L808) SLP vectorization at higher optimization levels. `flang-new` [does not](https://github.com/llvm/llvm-project/blob/a8450619358c3de23b72aef7c1fd932bca112257/flang/lib/Frontend/FrontendActions.cpp#L970) do this.

I think we need to align the optimization pipelines between clang and flang-new.

[banach-space]

Thanks for bringing this up!

I think we need to align the optimization pipelines between clang and flang-new.

Yes, but we also need to be careful as Clang and Flang might have different needs. SLP vectorisation is just one specific example, but there's actually a lot of fine tuning in Clang as well. Should Flang inherit that? Or should things be added incrementally? What if running SLP for code generated by Flang exposes bugs in SLP?

-Andrzej

[vzakhari]

As a matter of fact, the SLP changes have been a problem several times for my downstream testing, because I had it running with SLP. I think this should not be the reason not to enable it in Flang. I believe it improves performance of some benchmarks.

I view Clang's pipeline as finely tuned, i.e. that it provides the best performance overall. I agree that Fortran and C apps may exploit different optimization paths and "the best" Clang's pipeline may behave differently on Fortran apps, but I think we should at least try. This will probably imply enabling the optimizations one by one or in groups and collecting performance data for each change. So being careful sounds right :)

[banach-space]

I strongly support this :) And will try my best to review any incoming patches in a timely manner 🤞🏻 .

I expect this requiring multiple patches and some fine-tuning based on benchmarking. It would be great if one person could drive this and keep track of the delta between clang and flang-new.

[vzakhari]

I think we need to start with adding -fdebug-pass-structure support in flang-new so that we can get the two lists and get a diff: echo "void test() {}" | clang -O2 -fdebug-pass-structure -c -x c -