Optimisation pipelines

[banach-space]

Hi everyone,

I’m trying to identify where Clang defines optimsation pipelines (middle-end and back-end). For the sake of simplicity, I’m mostly interested in plain -O{0|1|2|3} (I appreciate that LTO, sanitizers and various code-gen flags make this much more nuanced).

I’ve always been under the impression that this is totally custom in Clang (i.e. defined from scratch), but on closer inspection I see that this is done through LLVM’s API instead (i.e. Clang relies on the pipelines defined in LLVM). More specifically, in BackendUtil.cpp:

• EmitAssemblyHelper::RunOptimizationPipeline configures middle-end optimisation pipelines via LLVM’s PassBuilder API
• EmitAssemblyHelper::CreateTargetMachine configures back-end optimisation pipeline through LLVM’s TargetMachine API.

So, the basic pipelines are in fact defined in LLVM. Clang simply forwards the optimisation level (e.g. -O3) to LLVM. Did I get this right?

For context, I would like to define optimisation flags in Flang. It would be great to make Clang and Flang behave consistently here!

Thank you,
-Andrzej

[mshockwave]

@banach-space

So, the basic pipelines are in fact defined in LLVM. Clang simply forwards the optimisation level (e.g. -O3 ) to LLVM. Did I get this right?

Correct, majority of the pipeline body in any optimization level is dictated by PassBuilder, with only few differences, like sanitizer (when needed) + its preparation Passes.

[banach-space]

PassBuilder belongs to the new PM infra, right? Or is it also used for the legacy PM pipelines? IIRC, the backend still uses the legacy PM.

[jvstech]

Indeed, the backend still uses the legacy PM.

The equivalent legacy functions for optimization level passes are defined in PassManagerBuilder.cpp (populateModulePassManager, populateLTOPassManager, etc.)

[banach-space]

Thank you both!