During superblock generation, a JUMP_BACKWARD instruction is translated to either a JUMP_TO_TOP micro-op (when the target of the jump is exactly the beginning of the superblock, closing the loop), or a SAVE_IP + EXIT_TRACE pair, when the jump goes elsewhere.
The new JUMP_TO_TOP instruction includes a CHECK_EVAL_BREAKER() call, so a closed loop can still be interrupted.
- Hand-written uops JUMP_IF_{TRUE,FALSE}.
These peek at the top of the stack.
The jump target (in superblock space) is absolute.
- Hand-written translation for POP_JUMP_IF_{TRUE,FALSE},
assuming the jump is unlikely.
Once we implement jump-likelihood profiling,
we can implement the jump-unlikely case (in another PR).
- Tests (including some test cleanup).
- Improvements to len(ex) and ex[i] to expose the whole trace.
This adds several of unspecialized opcodes to superblocks:
TO_BOOL, BINARY_SUBSCR, STORE_SUBSCR,
UNPACK_SEQUENCE, LOAD_GLOBAL, LOAD_ATTR,
COMPARE_OP, BINARY_OP.
While we may not want that eventually, for now this helps finding bugs.
There is a rudimentary test checking for UNPACK_SEQUENCE.
Once we're ready to undo this, that would be simple:
just replace the call to variable_used_unspecialized
with a call to variable_used (as shown in a comment).
Or add individual opcdes to FORBIDDEN_NAMES_IN_UOPS.
Instead of special-casing specific instructions,
we add a few more special values to the 'size' field of expansions,
so in the future we can automatically handle
additional super-instructions in the generator.
- Tweak uops debugging output
- Fix the bug from gh-106290
- Rename `SET_IP` to `SAVE_IP` (per https://github.com/faster-cpython/ideas/issues/558)
- Add a `SAVE_IP` uop at the start of the trace (ditto)
- Allow `unbound_local_error`; this gives us uops for `LOAD_FAST_CHECK`, `LOAD_CLOSURE`, and `DELETE_FAST`
- Longer traces
- Support `STORE_FAST_LOAD_FAST`, `STORE_FAST_STORE_FAST`
- Add deps on pycore_uops.h to Makefile(.pre.in)
This produces longer traces (superblocks?).
Also improved debug output (uop names are now printed instead of numeric opcodes). This would be simpler if the numeric opcode values were generated by generate_cases.py, but that's another project.
Refactored some code in generate_cases.py so the essential algorithm for cache effects is only run once. (Deciding which effects are used and what the total cache size is, regardless of what's used.)
Added a new, experimental, tracing optimizer and interpreter (a.k.a. "tier 2"). This currently pessimizes, so don't use yet -- this is infrastructure so we can experiment with optimizing passes. To enable it, pass ``-Xuops`` or set ``PYTHONUOPS=1``. To get debug output, set ``PYTHONUOPSDEBUG=N`` where ``N`` is a debug level (0-4, where 0 is no debug output and 4 is excessively verbose).
All of this code is likely to change dramatically before the 3.13 feature freeze. But this is a first step.
This behavior is optional, because in some extreme cases it
may just make debugging harder. The tool defaults it to off,
but it is on in Makefile.pre.in.
Also note that this makes diffs to generated_cases.c.h noisier,
since whenever you insert or delete a line in bytecodes.c,
all subsequent #line directives will change.
* Write output and metadata in a single run
This halves the time to run the cases generator
(most of the time goes into parsing the input).
* Declare or define opcode metadata based on NEED_OPCODE_TABLES
* Use generated metadata for stack_effect()
* compile.o depends on opcode_metadata.h
* Return -1 from _PyOpcode_num_popped/pushed for unknown opcode
New generator feature: Generate useful glue for output arrays, so you can just write values to the output array (no bounds checking). Rewrote UNPACK_SEQUENCE_TWO_TUPLE to use this, and also UNPACK_SEQUENCE_{TUPLE,LIST}.
You can now write things like this:
```
inst(BUILD_STRING, (pieces[oparg] -- str)) { ... }
inst(LIST_APPEND, (list, unused[oparg-1], v -- list, unused[oparg-1])) { ... }
```
Note that array output effects are only partially supported (they must be named `unused` or correspond to an input effect).
For these the instr_format field uses IX instead of IB.
Register instructions use IX, IB, IBBX, IBBB, etc.
Also: Include the closing '}' in Block.tokens, for completeness
(These aren't used yet, but may be coming soon,
and it's easier to keep this tool the same between branches.)
Added a sanity check for all this to compile.c.
Co-authored-by: Irit Katriel <iritkatriel@yahoo.com>
The presence of this macro indicates that a particular instruction
may be considered for conversion to a register-based format
(see https://github.com/faster-cpython/ideas/issues/485).
An invariant (currently unchecked) is that `DEOPT_IF()` may only
occur *before* `DECREF_INPUTS()`, and `ERROR_IF()` may only occur
*after* it. One reason not to check this is that there are a few
places where we insert *two* `DECREF_INPUTS()` calls, in different
branches of the code. The invariant checking would have to be able
to do some flow control analysis to understand this.
Note that many instructions, especially specialized ones,
can't be converted to use this macro straightforwardly.
This is because the generator currently only generates plain
`Py_DECREF(variable)` statements, and cannot generate
things like `_Py_DECREF_SPECIALIZED()` let alone deal with
`_PyList_AppendTakeRef()`.
Stack effects can now have a type, e.g. `inst(X, (left, right -- jump/uint64_t)) { ... }`.
Instructions converted to the non-legacy format:
* COMPARE_OP
* COMPARE_OP_FLOAT_JUMP
* COMPARE_OP_INT_JUMP
* COMPARE_OP_STR_JUMP
* STORE_ATTR
* DELETE_ATTR
* STORE_GLOBAL
* STORE_ATTR_INSTANCE_VALUE
* STORE_ATTR_WITH_HINT
* STORE_ATTR_SLOT, and complete the store_attr family
* Complete the store_subscr family: STORE_SUBSCR{,DICT,LIST_INT}
(STORE_SUBSCR was alread half converted,
but wasn't using cache effects yet.)
* DELETE_SUBSCR
* PRINT_EXPR
* INTERPRETER_EXIT (a bit weird, ends in return)
* RETURN_VALUE
* GET_AITER (had to restructure it some)
The original had mysterious `SET_TOP(NULL)` before `goto error`.
I assume those just account for `obj` having been decref'ed,
so I got rid of them in favor of the cleanup implied by `ERROR_IF()`.
* LIST_APPEND (a bit unhappy with it)
* SET_ADD (also a bit unhappy with it)
Various other improvements/refactorings as well.
Newly supported interpreter definition syntax:
- `op(NAME, (input_stack_effects -- output_stack_effects)) { ... }`
- `macro(NAME) = OP1 + OP2;`
Also some other random improvements:
- Convert `WITH_EXCEPT_START` to use stack effects
- Fix lexer to balk at unrecognized characters, e.g. `@`
- Fix moved output names; support object pointers in cache
- Introduce `error()` method to print errors
- Introduce read_uint16(p) as equivalent to `*p`
Co-authored-by: Brandt Bucher <brandtbucher@gmail.com>
The switch cases (really TARGET(opcode) macros) have been moved from ceval.c to generated_cases.c.h. That file is generated from instruction definitions in bytecodes.c (which impersonates a C file so the C code it contains can be edited without custom support in e.g. VS Code).
The code generator lives in Tools/cases_generator (it has a README.md explaining how it works). The DSL used to describe the instructions is a work in progress, described in https://github.com/faster-cpython/ideas/blob/main/3.12/interpreter_definition.md.
This is surely a work-in-progress. An easy next step could be auto-generating super-instructions.
**IMPORTANT: Merge Conflicts**
If you get a merge conflict for instruction implementations in ceval.c, your best bet is to port your changes to bytecodes.c. That file looks almost the same as the original cases, except instead of `TARGET(NAME)` it uses `inst(NAME)`, and the trailing `DISPATCH()` call is omitted (the code generator adds it automatically).
Mark Shannon
Irit Katriel
Guido van Rossum
Kirill Podoprigora
Carl Meyer
Jacob Bower
Steve Dower
Brandt Bucher