[llvm] Add support for the new pass manager API

[gloaming2dawn]

🚀 Feature

Motivation

The new pass manager (now the default pass manager for the optimization pipeline in LLVM 13) leads to better compile times and benefits runtime of many programs. It requires explicit pass nesting. For example, to run a function pass, then a module pass, we need to wrap the function pass in a module adaptor:
$ opt -passes='function(no-op-function),no-op-module' /tmp/a.ll -S

The trend is that LLVM will abandon the legacy pass manager (which is now used by LLVM 10 of Compilergym). As shown in https://llvm.org/docs/ReleaseNotes.html, using the legacy pass manager for the optimization pipeline is deprecated and will be removed after LLVM 14.

So if Compilergym could support for the newest LLVM which use new pass manager, it could be more attractive.

[ChrisCummins]

Hi @gloaming2dawn, thanks for the feature request! I agree this would be nice. I haven't used the new pass manager API, but from what I understand it is supported in LLVM 10, so I would happily take a patch to port the current LLVM environment to the new pass manager API.

Here is the code which uses the legacy pass manager:

CompilerGym/compiler_gym/envs/llvm/service/LlvmSession.cc

Lines 237 to 254 in 9734103

	bool LlvmSession::runPass(llvm::Pass* pass) {
	llvm::legacy::PassManager passManager;
	setupPassManager(&passManager, pass);
	return passManager.run(benchmark().module());
	}
	bool LlvmSession::runPass(llvm::FunctionPass* pass) {
	llvm::legacy::FunctionPassManager passManager(&benchmark().module());
	setupPassManager(&passManager, pass);
	bool changed = passManager.doInitialization();
	for (auto& function : benchmark().module()) {
	changed |= (passManager.run(function) ? 1 : 0);
	}
	changed |= (passManager.doFinalization() ? 1 : 0);
	return changed;
	}

I believe that swapping from LLVM 10 to 13 should be easier using the new CMake build system, though it may require a couple of API updates.

Left to my own devices, I would probably get around to taking a look at this in Q2, so patches welcome!

Cheers,
Chris

[gloaming2dawn]

Thank you for your quick reply. Here are some supplements from https://llvm.org/docs/NewPassManager.html. The following code

ModulePassManager MPM;
MPM.addPass(createModuleToFunctionPassAdaptor(FunctionPass1()));
MPM.addPass(createModuleToFunctionPassAdaptor(FunctionPass2()));
MPM.run();

will run FunctionPass1 on each function in a module, then run FunctionPass2 on each function in the module. In contrast,

ModulePassManager MPM;
FunctionPassManager FPM;
FPM.addPass(FunctionPass1());
FPM.addPass(FunctionPass2());
MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));

will run FunctionPass1 and FunctionPass2 on the first function in a module, then run both passes on the second function in the module, and so on. This is better for cache locality around LLVM data structures. For example, running all loop passes on a loop may cause a later loop to be able to be optimized more than if each loop pass were run separately.

So, the new pass manager may have a larger search space than the legacy pass manager, which can be quite different from the current action space (the current action space: each pass is an action, separately applying pass1 and pass2 is the same as applying pass1 and pass2 in the pass manager).

[ChrisCummins]

So, the new pass manager may have a larger search space than the legacy pass manager, which can be quite different from the current action space (the current action space: each pass is an action, separately applying pass1 and pass2 is the same as applying pass1 and pass2 in the pass manager).

Interesting! I think this could be exposed pretty neatly by creating a new action space that takes a list of passes, that are applied in a single step. Something like:

env = compiler_gym.make(
    "llvm-v0",
    combine_function_passes_into_single_adaptor=True,
)
env.reset()
# "normal" use:
env.step(0)
env.step(1)
# run pass 0 -> 1 -> 2, combining adjacent function passes into a single function pass adaptor where possible:
env.step([0, 1, 2])

What do you think?

Cheers,
Chris

[libin049]

@gloaming2dawn @ChrisCummins Do you have any progress on new pass manager API?

[ChrisCummins]

Nothing on my end