Good job. This problem also blocks #128802 where's NOP isn't taken into account in between.
I'll take a closer look tomorrow.

One thing to keep in mind about NOPs is that they are not necessarily going to be removed. When we replace an instruction by a NOP it retains the source location of the instruction we overwrote. If there is no other instruction with the same line, the NOP stays so that tracing/debugging sees that an instruction executed on that line. For example:

>>> def f():
...     1
...     2
...     3
...     return 4
...     
>>> dis.dis(f)
  1           RESUME                   0

  2           NOP

  3           NOP

  4           NOP

  5           LOAD_SMALL_INT           4
              RETURN_VALUE

I don't see a problem related to this here, just want you to be aware of this.

I don't think adding complexity to the individual optimizations is the way to go.
All these optimizations can expose the opportunity for other optimizations, so ultimately we want to iterate to a fixed point.
We already have a pass to remove NOPs from a basic block, so we should incorporate that into the loop.

Take the constant 3 if (1 if True else 0, 1)[0 if 2 == 2 else 1] else 2 which evaluates to 3. To preform that evaluation in the CFG requires tuple folding, flow control simplification, comparison folding and subscript folding. All of which depend on each other to get the final result. The only way to fold this is iteratively.

I don't think adding complexity to the individual optimizations is the way to go.
All these optimizations can expose the opportunity for other optimizations, so ultimately we want to iterate to a fixed point.
We already have a pass to remove NOPs from a basic block, so we should incorporate that into the loop.

The issue here is skipping NOPs that are not going to be removed (because source locations), but they are in the middle of a const sequence construction.

@markshannon

3 if (1 if True else 0, 1)[0 if 2 == 2 else 1] else 2 which evaluates to 3

Current Python version does not fold it to constant 3. Constant folding is not done across basic block boundaries in cases where control flow present (like you posted) IIRC, and AST optimizations being moved to CFG operate on single blocks.

Constant folding is not done across basic block boundaries

Do you suggest this is something we will be adding?

The example I give could be folded using AST transforms, even if it happens not to be.
So I think we should do the same in the CFG optimizer, which will need iteration.

@iritkatriel what's your take on this?

But this particular logic does not interfere with potential control flow simplification. We just fold inside individual basic blocks. Sometimes multiple such foldings happen together and we need to take into account intermediate NOPs they leave so this is all this PR is about. Or am I missing something?

The example I give could be folded using AST transforms, even if it happens not to be. So I think we should do the same in the CFG optimizer, which will need iteration.

@iritkatriel what's your take on this?

We have some of that in inline_small_or_no_lineno_blocks. We can put more steps into that loop.
But as I said, the issue here is NOPs that are not going to be removed, so fixed point won't help them.

Oh, turns out @iritkatriel merged a1417b2, so tests are failing because BINARY_SUBSCR does not exist anymore. I will rebase & update the PR.

Done.

LGTM. Thanks, Yan! Just to make sure there are no refleaks, I will run the test suite locally in huntrefleak mode.

Results:

== Tests result: SUCCESS ==

22 tests skipped:
    test.test_asyncio.test_windows_events
    test.test_asyncio.test_windows_utils test_android test_apple
    test_devpoll test_free_threading test_idle test_kqueue
    test_launcher test_msvcrt test_smtpnet test_ssl test_startfile
    test_tcl test_tkinter test_ttk test_ttk_textonly test_turtle
    test_winapi test_winconsoleio test_winreg test_wmi

8 tests skipped (resource denied):
    test_curses test_peg_generator test_pyrepl test_socketserver
    test_urllib2net test_urllibnet test_winsound test_zipfile64

454 tests OK.

Total duration: 13 min 49 sec
Total tests: run=44,920 skipped=2,362
Total test files: run=476/484 skipped=22 resource_denied=8
Result: SUCCESS