dis: Show names of intrinsics

[JelleZijlstra]

For CALL_INTRINSIC_* opcodes, dis currently just shows the number, which makes it hard to figure out what the code actually does:

              2 LOAD_CONST               0 ('T')
              4 CALL_INTRINSIC_1         7

It would be nice if this instead said

              2 LOAD_CONST               0 ('T')
              4 CALL_INTRINSIC_1         7 (INTRINSIC_TYPEVAR)

Possible implementation strategy:

• instrinsic names are defined in opcode.py
• pycore_intrinsics.h is generated from opcode.py (bonus: we no longer have to manually update MAX_INTRINSIC_1)
• dis learns to read the intrinsic names from opcode.py

Linked PRs

• GH-103963: Make dis display names of args for intrinsics opcodes #104029
• gh-103963: fix 'make regen-opcode' in out-of-tree builds #104177

[iritkatriel]

@jkchandalia are you interested in this?

[jkchandalia]

@iritkatriel I can work on this.

[iritkatriel]

Great. Have a look at /Tools/build/generate_opcode_h.py
(re generating the .h file).

[iritkatriel]

Great. Have a look at /Tools/build/generate_opcode_h.py
(re generating the .h file).

Use make regen-opcode to run this script.

[JelleZijlstra]

@jkchandalia thanks! It would be nice for me if this can make it in before the 3.12 feature freeze (May 8), since #103764 relies heavily on intrinsics, so this feature will make the bytecode generated by that PR much easier to understand. If that timeline is too tight for you, I'm happy to have a go at implementing this myself.

[jkchandalia]

@JelleZijlstra that timeline is fine for me.

[JelleZijlstra]

Great! Let me know if you need any help.

FYI I think these are some of the easiest ways to trigger intrinsics on current main for testing:

>>> dis.dis('(*a,)')
  0           0 RESUME                   0

  1           2 BUILD_LIST               0
              4 LOAD_NAME                0 (a)
              6 LIST_EXTEND              1
              8 CALL_INTRINSIC_1         6
             10 RETURN_VALUE
>>> dis.dis('+a')
  0           0 RESUME                   0

  1           2 LOAD_NAME                0 (a)
              4 CALL_INTRINSIC_1         5
              6 RETURN_VALUE

[jkchandalia]

Absolutely, thank you. And I appreciate the examples.

[jkchandalia]

@JelleZijlstra I've pushed my branch (https://github.com/jkchandalia/cpython/tree/display_intrinsics_names) and I wasn't quite sure if it's ready for a PR but can definitely open one if that's easier for feedback/discussion.

1. I added two lists to opcode.py for the intrinsics names (I added an extra option for not valid, not sure if this is needed)
2. I use these lists in dis.py to get the argrepr to pass to the Instruction class.
3. I added tests for args 2, 5, and 6 for CALL_INTRINSIC_1 and updated an existing test for arg 3. I put the three new tests into one test function but can split out if better. I had some trouble getting something to trigger the other codes (0?, 1, 4) and also CALL_INTRINSIC_2 for additional testing.
4. I updated /Tools/build/generate_opcode_h.py to include the new intrinsic names sections (not sure if this is desired?)
5. I regenerated opcode.h with the new intrinsics-related sections
6. I could definitely write the code to generate pycore_intrinsics.h as you suggested but I wasn’t quite sure where to put it? A separate script or another internal file output from /Tools/build/generate_opcode_h.py or something else?

[JelleZijlstra]

Looks great! Feel free to open a draft PR and we can get it over the finish line.

You can get the CALL_INTRINSIC_2 with except* (position 38):

>>> dis.dis("try: pass\nexcept* Exception: x")
  0           0 RESUME                   0

  1           2 RETURN_CONST             0 (None)
              4 PUSH_EXC_INFO

  2           6 BUILD_LIST               0
              8 COPY                     2
             10 LOAD_NAME                0 (Exception)
             12 CHECK_EG_MATCH
             14 COPY                     1
             16 POP_JUMP_IF_NONE         8 (to 34)
             18 POP_TOP
             20 LOAD_NAME                1 (x)
             22 POP_TOP
             24 JUMP_FORWARD             3 (to 32)
        >>   26 LIST_APPEND              3
             28 POP_TOP
             30 JUMP_FORWARD             2 (to 36)
        >>   32 JUMP_FORWARD             1 (to 36)
        >>   34 POP_TOP
        >>   36 LIST_APPEND              1
             38 CALL_INTRINSIC_2         1
             40 COPY                     1
             42 POP_JUMP_IF_NOT_NONE     3 (to 50)
             44 POP_TOP
             46 POP_EXCEPT
             48 RETURN_CONST             0 (None)
        >>   50 SWAP                     2
             52 POP_EXCEPT
             54 RERAISE                  0
        >>   56 COPY                     3
             58 POP_EXCEPT
             60 RERAISE                  1
ExceptionTable:
  4 to 18 -> 56 [1] lasti
  20 to 22 -> 26 [4] lasti
  24 to 44 -> 56 [1] lasti

I don't think it's necessary to test every single one of the intrinsics in test_dis.

I would suggest simply expanding generate_opcode_h.py to also generate pycore_intrinsics.h. It already generates multiple files, so it doesn't seem too bad to make it generate another closely related file. We'll add a header to pycore_intrinsics.h to make it clear how it is generated. Perhaps @iritkatriel has a different opinion, though. We definitely shouldn't add the intrinsics to opcode.h, as that's a public header file and we don't want the intrinsics to be part of the public API.

[iritkatriel]

Sounds good.