GH-98831: Typed stack effects, and more instructions converted #99764

gvanrossum · 2022-11-25T05:51:02Z

This started out as just converting a bunch of instructions (a few small families) to the new format. I then ended up adding:

Support for typed stack effects (e.g. jump: size_t)
Some general refactoring of the super/macro code
Reversing the temp variable numbering

Issue: Generate the interpreter #98831

bedevere-bot · 2022-11-25T05:51:51Z

🤖 New build scheduled with the buildbot fleet by @gvanrossum for commit 522dc9a 🤖

If you want to schedule another build, you need to add the ":hammer: test-with-buildbots" label again.

bedevere-bot · 2022-11-26T05:30:54Z

🤖 New build scheduled with the buildbot fleet by @gvanrossum for commit 1717003 🤖

If you want to schedule another build, you need to add the ":hammer: test-with-buildbots" label again.

gvanrossum · 2022-11-28T00:11:41Z

Trying to debug the Windows crashes, my only idea (while I'm traveling without my Windows laptop) is to bisect by hand, pushing each time until I've identified the commit that broke it. It passed in GH-99601 which is a direct ancestor.

gvanrossum · 2022-11-28T04:05:21Z

The culprit appears to be commit 5e433d6, "LIST_APPEND (a bit unhappy with it)". I'll have to closely read that commit again.

gvanrossum · 2022-11-28T04:21:17Z

Well d'oh, the generated code has this beauty:

            PREDICT(JUMP_BACKWARD);
            STACK_SHRINK(1);
            DISPATCH();

I need to go back to the drawing board for this. Presumably if the C code block ends with PREDICT() we need to insert the STACK_SHRINK() before that.

gvanrossum · 2022-11-28T04:36:09Z

So this only failed on Windows because the other popular platforms use computed goto and then PREDICT() is a no-op.

Maybe the answer is faster-cpython/ideas#496 (proposal to get rid of PREDICT()). There are only 10 PREDICT() calls left. We should benchmark this on Windows first though.

bedevere-bot · 2022-11-28T23:16:15Z

🤖 New build scheduled with the buildbot fleet by @gvanrossum for commit aff6852 🤖

If you want to schedule another build, you need to add the ":hammer: test-with-buildbots" label again.

gvanrossum · 2022-12-02T22:54:04Z

The one failing test seems to be unrelated -- several other builds on that machine are also failing. https://buildbot.python.org/all/#/builders/387

This doesn't really make things much cleaner, but it works.

This was more convoluted than I expected.

Also removed some dead code from the former (somehow the case for CacheEffect had crept back in).

This should fix the build crash on Windows.

brandtbucher

Looks mostly good, just a few questions and other things I noticed:

brandtbucher · 2022-12-08T00:58:53Z

Tools/cases_generator/parser.py

+            type = ""
+            if self.expect(lx.COLON):
+                type = self.require(lx.IDENTIFIER).text
+            return StackEffect(tkn.text, type)


Not a huge fan of shadowing type here:

Suggested change

type = ""

if self.expect(lx.COLON):

type = self.require(lx.IDENTIFIER).text

return StackEffect(tkn.text, type)

type_ = ""

if self.expect(lx.COLON):

type_ = self.require(lx.IDENTIFIER).text

return StackEffect(tkn.text, type_)

Not a fan of suffix _, but I hear you, and I can change this to typ. I'll make a pass.

brandtbucher · 2022-12-08T01:01:39Z

Python/generated_cases.c.h

-                goto error;
+            PyObject *v = PEEK(1);
+            PyObject *list = PEEK(oparg + 1);  // +1 to account for v staying on stack
+            if (_PyList_AppendTakeRef((PyListObject *)list, v) < 0) goto pop_1_error;


Any plans to make these conform to PEP 7 (braces and newline for body)?

I know this isn't exactly meant for human consumption, but as a human who will definitely be consuming this output, I think it might be worth improving the readability of the generated code a bit.

Is the issue that you easily miss the goto, or is your brain just yelling "missing braces" at you? I intentionally put it all on one line so the reason why PEP 7 requires braces doesn't apply. I think there's also a technical issue where we suddenly need to know the indentation level (but that's recovered easily enough).

All in all my preferred answer is "no". :-) But I feel you. My brain yells at me a little bit too.

brandtbucher · 2022-12-08T01:05:17Z

Python/bytecodes.c

            if (hook == NULL) {
                _PyErr_SetString(tstate, PyExc_RuntimeError,
                                 "lost sys.displayhook");
                Py_DECREF(value);
-                goto error;
+                ERROR_IF(1, error);


This took me a second to parse. Maybe it's worth adding another ERROR(LABEL) macro for unconditional errors?

The problem with ERROR(error) is that the generator would need to intercept that too (since possibly it needs to go to _pop_1_error), and the intercept code is slightly hairy already. Maybe ERROR_IF(true, error) works? I'll try that.

brandtbucher · 2022-12-08T01:06:46Z

Python/bytecodes.c

-                goto error;
-            }
-            JUMPBY(INLINE_CACHE_ENTRIES_STORE_SUBSCR);
+            ERROR_IF(err != 0, error);


Minor: the != 0 is just adding noise here (I had to double-check the condition when reading). Maybe just:

Suggested change

ERROR_IF(err != 0, error);

ERROR_IF(err, error);

(Elsewhere too.)

Huh, the original had it too. Agreed it's redundant, my brain doesn't see it as noise though (the original never gave me a second thought).

brandtbucher · 2022-12-08T01:08:29Z

Python/bytecodes.c

@@ -81,8 +81,19 @@ do { \
 // Dummy variables for stack effects.
 static PyObject *value, *value1, *value2, *left, *right, *res, *sum, *prod, *sub;
 static PyObject *container, *start, *stop, *v, *lhs, *rhs;
-static PyObject *list, *tuple, *dict;
-static PyObject *exit_func, *lasti, *val;
+static PyObject *list, *tuple, *dict, *owner;


Would it be possible to have the generator also build a header with all of these lines? You would need to run the generator at least once before the red squiggles go away, but then we don't have to worry about manually maintaining this blob of stuff.

Yeah, this part of the tooling is kind of annoying. But having the generator also update bytecodes.c seems sneaky (and slightly dangerous).

brandtbucher · 2022-12-08T01:31:29Z

Python/bytecodes.c

+            _COMPARE_OP_FLOAT,
+            _COMPARE_OP_INT,
+            _COMPARE_OP_STR,


Hm, so families can't contain superinstructions? I'm not sure I would have been able to figure out this trick on my own... :(

Also see faster-cpython/ideas#495 (comment). This is definitely suboptimal.

brandtbucher · 2022-12-08T01:33:17Z

Python/bytecodes.c

+            jump = sign_ish & when_to_jump_mask;
+        }
+        // The input is an int disguised as an object pointer!
+        op(_JUMP_ON_SIGN, (jump: size_t --)) {


Maybe call this _JUMP_IF? It's more general than on "sign" alone, and might be useful elsewhere...

Suggested change

op(_JUMP_ON_SIGN, (jump: size_t --)) {

op(_JUMP_IF, (jump: uintptr_t --)) {

Oh, you're right, that uop needs a better name. I like _JUMP_IF. Will do.

brandtbucher · 2022-12-08T01:34:50Z

Python/generated_cases.c.h

            STACK_SHRINK(3);
+            next_instr += 1;


Might be more readable to spell all these as JUMPBY(...)?

Oh, good idea. Although there will still be explicit mentions of next_instr in the read_uNN() calls.

brandtbucher · 2022-12-08T01:36:18Z

Python/generated_cases.c.h

+            PyObject *_tmp_2 = PEEK(2);
+            PyObject *_tmp_1 = PEEK(1);


Minor: these got better, but still seem "backwards". Not sure if you noticed.

Oh, you'd want the PEEK(1) to come before the PEEK(2). I could do that.

brandtbucher · 2022-12-08T01:40:57Z

Tools/cases_generator/generate_cases.py

+    def declare(self, dst: StackEffect, src: StackEffect | None):
+        if dst.name == UNUSED:
+            return
+        type = f"{dst.type} " if dst.type else "PyObject *"


Should we just set the default value of type to "PyObject *" in the parser to avoid this move here and elsewhere?

Then we'd be generating slightly un-idiomatic code where there's a space after the star. I'm not sure what bugs me more. I think the code I already wrote wins though. :-)

gvanrossum

Sorry for doing one-by-one responses to the first few comments. If I do the things you've asked for, do you want another chance at review?

gvanrossum · 2022-12-08T02:15:59Z

Python/bytecodes.c

+        // Part 1's output effect is a lie -- it has no result.
+        // Part 2's input effect is equally a lie, and the two lies
+        // cancel each other out.
+        op(_BINARY_OP_INPLACE_ADD_UNICODE_PART_1, (left, right, unused/1 -- unused)) {


The lies are so that we can add _BINARY_OP_INPLACE_ADD_UNICODE_PART_1 to the family and define the thing as a proper super-instruction.

Since you react rather strongly to this I think maybe I should just roll that back. See also faster-cpython/ideas#495 (comment).

gvanrossum · 2022-12-08T02:16:35Z

Python/bytecodes.c

-            PyObject *list = PEEK(oparg);
-            if (_PyList_AppendTakeRef((PyListObject *)list, v) < 0)
-                goto error;
+        // Alternative: (list, unused[oparg], v -- list, unused[oparg])


Yeah, but those are still only a faint glimmer on the horizon. Duly noted though.

Python/bytecodes.c

gvanrossum · 2022-12-08T02:18:13Z

Python/bytecodes.c

-                goto error;
-            }
-            JUMPBY(INLINE_CACHE_ENTRIES_STORE_SUBSCR);
+            ERROR_IF(err != 0, error);


Huh, the original had it too. Agreed it's redundant, my brain doesn't see it as noise though (the original never gave me a second thought).

gvanrossum · 2022-12-08T02:20:16Z

Python/bytecodes.c

            assert(frame == &entry_frame);
            assert(_PyFrame_IsIncomplete(frame));
-            PyObject *retval = POP();
+            STACK_SHRINK(1);  // Since we're not going to DISPATCH()


Without it the assert(EMPTY()) on the next line fails. I'm kind of hoping that when asserts are off, the compiler sees that the stack pointer is dead at this point and won't bother to generate code for it. Then again, the original had POP() which also adjusts the stack pointer.

gvanrossum · 2022-12-08T02:25:36Z

Python/bytecodes.c

+            _COMPARE_OP_FLOAT,
+            _COMPARE_OP_INT,
+            _COMPARE_OP_STR,


Also see faster-cpython/ideas#495 (comment). This is definitely suboptimal.

gvanrossum · 2022-12-08T02:26:15Z

Python/bytecodes.c

+            jump = sign_ish & when_to_jump_mask;
+        }
+        // The input is an int disguised as an object pointer!
+        op(_JUMP_ON_SIGN, (jump: size_t --)) {


Oh, you're right, that uop needs a better name. I like _JUMP_IF. Will do.

gvanrossum · 2022-12-08T02:28:05Z

Python/generated_cases.c.h

            STACK_SHRINK(3);
+            next_instr += 1;


Oh, good idea. Although there will still be explicit mentions of next_instr in the read_uNN() calls.

gvanrossum · 2022-12-08T02:29:26Z

Python/generated_cases.c.h

+            PyObject *_tmp_2 = PEEK(2);
+            PyObject *_tmp_1 = PEEK(1);


Oh, you'd want the PEEK(1) to come before the PEEK(2). I could do that.

gvanrossum · 2022-12-08T02:30:49Z

Tools/cases_generator/generate_cases.py

+    def declare(self, dst: StackEffect, src: StackEffect | None):
+        if dst.name == UNUSED:
+            return
+        type = f"{dst.type} " if dst.type else "PyObject *"


Then we'd be generating slightly un-idiomatic code where there's a space after the star. I'm not sure what bugs me more. I think the code I already wrote wins though. :-)

gvanrossum · 2022-12-08T05:37:41Z

netlify · 2022-12-08T05:56:46Z

✅ Deploy Preview for python-cpython-preview ready!

Name	Link
🔨 Latest commit	`c6dfeec`
🔍 Latest deploy log	https://app.netlify.com/sites/python-cpython-preview/deploys/639209a248c88400083c9c50
😎 Deploy Preview	https://deploy-preview-99764--python-cpython-preview.netlify.app
📱 Preview on mobile	Toggle QR Code... Use your smartphone camera to open QR code link.

To edit notification comments on pull requests, go to your Netlify site settings.

It was too fancy, and I have another idea for super instructions and effects anyways. This reverts commit 6860ed7.

gvanrossum

I'm going to do the following:

It'll probably take all mornning. Stop me if you want me to change anything; if I don't hear from you I'll just go ahead with merging.

bedevere-bot · 2022-12-08T15:59:07Z

🤖 New build scheduled with the buildbot fleet by @gvanrossum for commit c6dfeec 🤖

If you want to schedule another build, you need to add the ":hammer: test-with-buildbots" label again.

gvanrossum · 2022-12-08T17:52:01Z

It looks like the ARM64 buildbot failure is not this PR's fault, its other runs also failed in the last day or so.

brandtbucher

Thanks for the changes!

bedevere-bot mentioned this pull request Nov 25, 2022

Generate the interpreter #98831

Closed

bedevere-bot added the awaiting core review label Nov 25, 2022

gvanrossum added skip news 🔨 test-with-buildbots Test PR w/ buildbots; report in status section labels Nov 25, 2022

bedevere-bot removed the 🔨 test-with-buildbots Test PR w/ buildbots; report in status section label Nov 25, 2022

gvanrossum added the 🔨 test-with-buildbots Test PR w/ buildbots; report in status section label Nov 26, 2022

bedevere-bot removed the 🔨 test-with-buildbots Test PR w/ buildbots; report in status section label Nov 26, 2022

gvanrossum force-pushed the typed-stack-effects branch 2 times, most recently from fdb0dbd to 8f52910 Compare November 28, 2022 00:06

gvanrossum force-pushed the typed-stack-effects branch from 141bdb9 to d24cc5a Compare November 28, 2022 00:44

gvanrossum force-pushed the typed-stack-effects branch from fa0df3e to aff6852 Compare November 28, 2022 06:48

gvanrossum added the 🔨 test-with-buildbots Test PR w/ buildbots; report in status section label Nov 28, 2022

bedevere-bot removed the 🔨 test-with-buildbots Test PR w/ buildbots; report in status section label Nov 28, 2022

gvanrossum mentioned this pull request Nov 29, 2022

GH-98831: Support cache effects in super- and macro instructions #99601

Merged

brandtbucher self-requested a review December 3, 2022 01:04

gvanrossum added 7 commits December 2, 2022 21:24

Make BINARY_OP_INPLACE_ADD_UNICODE a legit super instruction

6860ed7

This doesn't really make things much cleaner, but it works.

COMPARE_OP

dcba32e

COMPARE_OP_FLOAT_JUMP

30cb8cd

COMPARE_OP_INT_JUMP

6af0a5d

COMPARE_OP_STR_JUMP

039efbf

Support typed stack effects

2fd8822

This was more convoluted than I expected.

Refactor common code of analyze_{super,macro}

14bbd50

Also removed some dead code from the former (somehow the case for CacheEffect had crept back in).

gvanrossum added 3 commits December 2, 2022 21:24

Emit PREDICT() macros right before DISPATCH()

f7a54d7

This should fix the build crash on Windows.

LIST_APPEND (a bit unhappy with it)

6ae9e6e

SET_ADD (also a bit unhappy with it)

34aa393

gvanrossum force-pushed the typed-stack-effects branch from aff6852 to 34aa393 Compare December 3, 2022 06:12

gvanrossum marked this pull request as ready for review December 4, 2022 00:34

brandtbucher approved these changes Dec 8, 2022

View reviewed changes

bedevere-bot added awaiting merge and removed awaiting core review labels Dec 8, 2022

gvanrossum commented Dec 8, 2022

View reviewed changes

gvanrossum added 4 commits December 7, 2022 21:50

Move all cache effects to the front of the effects list

e144db3

Rename _JUMP_ON_SIGN to _JUMP_IF

029bf07

Change ERROR_IF(err != 0, ...) to ERROR_IF(err, ...)

fa3135f

Change ERROR_IF(1, ...) to ERROR_IF(true, ...)

445f20b

gvanrossum added 5 commits December 7, 2022 22:07

Revert "Make BINARY_OP_INPLACE_ADD_UNICODE a legit super instruction"

32b4ac0

It was too fancy, and I have another idea for super instructions and effects anyways. This reverts commit 6860ed7.

Emit PEEKs in super/macro instructions in correct order

0ef4da7

Use JUMPBY(n) instead of next_instr += n

7d19466

Rename local variables type to typ

99e8ec6

Remove unused op names

c08dbdb

gvanrossum commented Dec 8, 2022

View reviewed changes

Merge branch 'main' into typed-stack-effects

c6dfeec

gvanrossum added the 🔨 test-with-buildbots Test PR w/ buildbots; report in status section label Dec 8, 2022

bedevere-bot removed the 🔨 test-with-buildbots Test PR w/ buildbots; report in status section label Dec 8, 2022

brandtbucher approved these changes Dec 8, 2022

View reviewed changes

gvanrossum merged commit c85be73 into python:main Dec 8, 2022

bedevere-bot removed the awaiting merge label Dec 8, 2022

gvanrossum deleted the typed-stack-effects branch December 8, 2022 23:45

	op(_JUMP_ON_SIGN, (jump: size_t --)) {
	op(_JUMP_IF, (jump: uintptr_t --)) {

GH-98831: Typed stack effects, and more instructions converted #99764

GH-98831: Typed stack effects, and more instructions converted #99764

Conversation

gvanrossum commented Nov 25, 2022 • edited Loading

bedevere-bot commented Nov 25, 2022

bedevere-bot commented Nov 26, 2022

gvanrossum commented Nov 28, 2022

gvanrossum commented Nov 28, 2022

gvanrossum commented Nov 28, 2022

gvanrossum commented Nov 28, 2022

bedevere-bot commented Nov 28, 2022

gvanrossum commented Dec 2, 2022

brandtbucher left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

gvanrossum left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

gvanrossum commented Dec 8, 2022 • edited Loading

netlify bot commented Dec 8, 2022 • edited Loading

✅ Deploy Preview for python-cpython-preview ready!

gvanrossum left a comment • edited Loading

Choose a reason for hiding this comment

bedevere-bot commented Dec 8, 2022

gvanrossum commented Dec 8, 2022

brandtbucher left a comment

Choose a reason for hiding this comment

gvanrossum commented Nov 25, 2022 •

edited

Loading

gvanrossum commented Dec 8, 2022 •

edited

Loading

netlify bot commented Dec 8, 2022 •

edited

Loading

gvanrossum left a comment •

edited

Loading