GH-104584: Remove `ip_offset` and simplify `EXIT_TRACE` #108961

brandtbucher · 2023-09-05T20:11:57Z

We don't need to keep ip_offset in a local, since it's known statically during trace construction. The only time that we don't know it is when pushing a frame for an unknown code object, so modify _PUSH_FRAME to set the prev_instr of the frame being pushed (this doesn't have any overhead, since it was always being followed by a SAVE_IP anyways). Otherwise, we just set the operand of SAVE_IP to the actual pointer that should be stored.

Also, EXIT_TRACE behaves the same as an unconditional goto deoptimize. Changing it to just do that instead makes things simpler, since all returns from tier two happen at labels (rather than tied up in the uop handlers).

Issue: Pluggable optimizer API #104584

gvanrossum

Sorry for the chatty review -- I have two angles, one about still being able to make sense of the debug output, another about whether _PUSH_FRAME is the right place to set prev_instr from the new code object. And some mumblings about wanting @markshannon to review this too (I'm not sure I 100% follow his reasoning for the LOAD/STORE macros he recently added to ceval_macros.h.)

gvanrossum · 2023-09-05T20:36:58Z

Python/optimizer.c

@@ -668,7 +667,7 @@ translate_bytecode_to_trace(
                                oparg = orig_oparg & 0xF;
                                break;
                            case OPARG_SAVE_IP:  // op==SAVE_IP; oparg=next instr


Please fix the comment.

gvanrossum · 2023-09-05T20:40:10Z

Python/optimizer.c

@@ -503,7 +502,7 @@ translate_bytecode_to_trace(
 top:  // Jump here after _PUSH_FRAME
    for (;;) {
        RESERVE_RAW(2, "epilogue");  // Always need space for SAVE_IP and EXIT_TRACE
-        ADD_TO_TRACE(SAVE_IP, INSTR_IP(instr, code), 0);
+        ADD_TO_TRACE(SAVE_IP, 0, (uintptr_t)instr);


May I ask to keep (redundantly) the INSTR_IP(...) argument in all cases? It's helpful for debugging -- the 64-bit hex operand is useless, but the code-relative offset can be related directly to the output from dis, and I regularly do that.

If you worry about the cost of computing INSTR_IP() when no debugging is selected, maybe define the macro differently (or define another macro)?

Also, I think the operand may have to be printed in hex (both here and in executor.c) -- debug output like 106858786520180 is not very helpful.

I think we will ultimately want to change from instruction pointer back to the offset (as we had in 3.9 and earlier), so let's keep the redundant (for now) macro.

The reason for wanting the offset, not the pointer is that we expect writes to frame->instr_ptr to be rare, thanks to optimization, most reads actually want the offset, and it means we only need a 32 bit operand (less copying and cache pressure).

gvanrossum · 2023-09-05T22:21:01Z

Python/optimizer.c

@@ -707,15 +706,13 @@ translate_bytecode_to_trace(
                                            PyUnicode_AsUTF8(new_code->co_qualname),
                                            PyUnicode_AsUTF8(new_code->co_filename),
                                            new_code->co_firstlineno);
-                                    ADD_TO_TRACE(SAVE_IP, 0, 0);


Interesting. Because you added an assignment to frame->prev_instr to _PUSH_FRAME, you don't need this dummy SAVE_IP any more. Same when bailing on version mismatch below and ditto if func == NULL. Right?

gvanrossum · 2023-09-05T22:24:08Z

Python/optimizer.c

@@ -731,7 +728,6 @@ translate_bytecode_to_trace(
                                    2 * INSTR_IP(instr, code));
                                goto top;
                            }
-                            ADD_TO_TRACE(SAVE_IP, 0, 0);


I had to look twice before understanding how we even get here. :-) I think we could use a comment here explaining that here func == NULL. (And/or a DPRINT() call?)

gvanrossum · 2023-09-05T22:34:03Z

Python/ceval_macros.h

 #define LOAD_IP() \
-do { ip_offset = (_Py_CODEUNIT *)_PyFrame_GetCode(frame)->co_code_adaptive; } while (0)
+do {} while (0)


I'd be curious what @markshannon thinks of this change. It makes me think that the "abstractions" embodied by these LOAD/STORE macros are not as solid as we thought. And I wonder if the line that sets prev_instr should also be a macro. (Note that the _SP macros don't actually differ by tier.)

This seems good to me. LOAD_IP should be a no-op in tier 2 execution. The IP should be virtual.

(In this context "virtual" means tracked by the region selector (translate_bytecode_to_trace) not at runtime)

The _SP may not differ by tier, but the copy-and-patch compiler will need to treat them differently from the tier 2 interpreter.

gvanrossum · 2023-09-05T22:39:08Z

Python/bytecodes.c

-            _PyFrame_SetStackPointer(frame, stack_pointer);
-            Py_DECREF(self);
-            return frame;
+            goto deoptimize;


In uops debug mode (PYTHONUOPSDEBUG=2 or higher) this would (confusingly, IMO) print DEOPT: [Opcode ...]. Doesn't that bother you? Or do you consider EXIT_TRACE as just another form of deoptimization?

It definitely is annoying. Actual deoptimizations are rare, but exiting a trace is common. So I get a lot of log spam containing endless repetitions of

DEOPT: [Opcode 300, operand 0]

This is tier 2 only, so it doesn't matter much what the implementation is. I can see it changing quite a few times in the near future.
Whatever it is, it shouldn't have the context dependent frame->prev_instr--; // Back up to just before destination

gvanrossum · 2023-09-05T22:40:36Z

Python/bytecodes.c

@@ -3006,6 +3006,7 @@ dummy_func(
            tstate->py_recursion_remaining--;
            LOAD_SP();
            LOAD_IP();
+            frame->prev_instr = _PyCode_CODE(_PyFrame_GetCode(frame));


See my comment in ceval_macros.c -- perhaps this should also be a macro?

This is unnecessary, frame->prev_instr is set during frame creation.

And if we want to reuse this code for SEND we don't want to be messing with frame->prev_instr

This is unnecessary, frame->prev_instr is set during frame creation.

It's set to an index of -1 (since the frame hasn't started yet), which is invalid. So we either need to set it like this, or at least do an increment.

Remove ip_offset and simplify EXIT_TRACE

089af02

brandtbucher added skip news interpreter-core (Objects, Python, Grammar, and Parser dirs) labels Sep 5, 2023

brandtbucher requested review from gvanrossum and markshannon September 5, 2023 20:11

brandtbucher self-assigned this Sep 5, 2023

bedevere-bot added the awaiting core review label Sep 5, 2023

bedevere-bot mentioned this pull request Sep 5, 2023

Pluggable optimizer API #104584

Closed

gvanrossum reviewed Sep 5, 2023

View reviewed changes

brandtbucher closed this Nov 20, 2023

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

GH-104584: Remove `ip_offset` and simplify `EXIT_TRACE` #108961

GH-104584: Remove `ip_offset` and simplify `EXIT_TRACE` #108961

brandtbucher commented Sep 5, 2023 •

edited

Loading

gvanrossum left a comment

gvanrossum Sep 5, 2023

gvanrossum Sep 5, 2023

markshannon Sep 7, 2023

gvanrossum Sep 5, 2023

gvanrossum Sep 5, 2023

gvanrossum Sep 5, 2023

markshannon Sep 7, 2023

markshannon Sep 7, 2023

gvanrossum Sep 5, 2023

gvanrossum Sep 5, 2023

markshannon Sep 7, 2023

gvanrossum Sep 5, 2023

markshannon Sep 7, 2023

markshannon Sep 7, 2023

brandtbucher Sep 7, 2023 •

edited

Loading

GH-104584: Remove ip_offset and simplify EXIT_TRACE #108961

GH-104584: Remove ip_offset and simplify EXIT_TRACE #108961

Conversation

brandtbucher commented Sep 5, 2023 • edited Loading

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

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

brandtbucher Sep 7, 2023 • edited Loading

Choose a reason for hiding this comment

GH-104584: Remove `ip_offset` and simplify `EXIT_TRACE` #108961

GH-104584: Remove `ip_offset` and simplify `EXIT_TRACE` #108961

brandtbucher commented Sep 5, 2023 •

edited

Loading

brandtbucher Sep 7, 2023 •

edited

Loading