How to get all fine grained optimization options of the LLVM optimizer opt -Ox (O1/O2/Os/O3...)

[cpython-java]

I found an opt-Ox bug, but I don't know which optimization step is causing the bug, in order to figure this out, I need to know all the fine-grained optimization options for opt-Ox, and then verify which optimization step the bug is in by trying to disable some of the fine-grained options.

I tried one way:

opt -Ox -debug -debug-pass=Arguments

However, this command does not work with all versions of opt, for some versions of opt it will report errors, and even for versions that do not report errors, I am not sure whether the output of this command is all optimization options and does not include non-optimization necessary execution options.

Hope a friend can help me!

[topperc]

You can use -opt-bisect-limit to narrow down the optimization with a binary search https://llvm.org/docs/OptBisect.html

[paperchalice]

Try opt -p='default<Ox>' --print-pipeline-passes /dev/null if your llvm distribution is new enough.

[cpython-java]

@topperc Thank you very much, but I still have doubts about this method. For example, I found through testing that -opt-bisect-limit=11 triggers a bug, while -opt-bisect-limit=10 does not. Based on this, I want to further narrow down the range of suspicious stages. For example, among the first 10 stages, perhaps the 2nd, 4th, and 8th stages do not have any bugs. How can I specify skipping these three non-contiguous stages?
In this way, if I find that specifying -opt-bisect-limit=11 and skipping these three stages still triggers the bug, I would know that I don't need to consider the code of these three stages during debugging.

[cpython-java]

@paperchalice Thank you!! However, the version of LLVM I'm studying spans a wide range, and this command doesn't seem to work with some of the older versions. And I don't know which of the passes output by this command can be disabled and which must be executed?

[topperc]

@topperc Thank you very much, but I still have doubts about this method. For example, I found through testing that -opt-bisect-limit=11 triggers a bug, while -opt-bisect-limit=10 does not. Based on this, I want to further narrow down the range of suspicious stages. For example, among the first 10 stages, perhaps the 2nd, 4th, and 8th stages do not have any bugs. How can I specify skipping these three non-contiguous stages?
In this way, if I find that specifying -opt-bisect-limit=11 and skipping these three stages still triggers the bug, I would know that I don't need to consider the code of these three stages during debugging.

In my experience the pass it points to is usually the one you should debug.

One exception to this is sometimes the problem is SelectionDAG which opt-bisect can't turn off. So opt-bisect will instead find the pass that generated the code that triggered a SelectionDAG bug. I've mainly seen this with the vectorizers.

What pass is number 11?

[cpython-java]

@topperc Thank you very much, but I still have doubts about this method. For example, I found through testing that -opt-bisect-limit=11 triggers a bug, while -opt-bisect-limit=10 does not. Based on this, I want to further narrow down the range of suspicious stages. For example, among the first 10 stages, perhaps the 2nd, 4th, and 8th stages do not have any bugs. How can I specify skipping these three non-contiguous stages?
In this way, if I find that specifying -opt-bisect-limit=11 and skipping these three stages still triggers the bug, I would know that I don't need to consider the code of these three stages during debugging.

In my experience the pass it points to is usually the one you should debug.

One exception to this is sometimes the problem is SelectionDAG which opt-bisect can't turn off. So opt-bisect will instead find the pass that generated the code that triggered a SelectionDAG bug. I've mainly seen this with the vectorizers.

What pass is number 11?

The number 11 is just a random example, not a number observed in my real experiments. However, I thought about my experiment carefully in the light of your answer, and got a great inspiration now, thank you very much!!!