Split opcodes into micro-ops #104909
Description
See faster-cpython/ideas#592. This project can be parallelized, and I could use help!
Note that initially we don't have a tier-2 interpreter or instruction format defined yet, but the first stage of splitting doesn't require that.
Remaining non-viable ops
With usage stats from the latest weekly pyperformance run (not all are uop candidates):
956,835,460 YIELD_VALUE
490,121,077 SEND_GEN (specialization of SEND)
429,220,003 JUMP_FORWARD
410,623,920 EXTENDED_ARG
390,537,200 JUMP_BACKWARD_NO_INTERRUPT
250,483,954 RETURN_GENERATOR
239,793,181 CALL_LIST_APPEND (specialization of CALL)
168,297,543 CALL_KW
162,922,780 FOR_ITER_GEN (specialization of FOR_ITER)
157,442,920 CALL_PY_WITH_DEFAULTS (specialization of CALL)
145,986,780 BINARY_SUBSCR_GETITEM (specialization of BINARY_SUBSCR)
135,636,840 STORE_FAST_LOAD_FAST
83,118,452 MAKE_CELL
74,149,898 CALL_FUNCTION_EX
68,587,076 CALL_ALLOC_AND_ENTER_INIT (specialization of CALL)
49,897,724 STORE_ATTR_WITH_HINT (specialization of STORE_ATTR)
49,846,886 LOAD_ATTR_PROPERTY (specialization of LOAD_ATTR)
8,224,500 RERAISE
6,000,000 END_ASYNC_FOR
5,801,385 BEFORE_WITH
2,892,780 RAISE_VARARGS
1,850,040 IMPORT_FROM
1,813,620 IMPORT_NAME
240 CLEANUP_THROW
120 BEFORE_ASYNC_WITH
ENTER_EXECUTOR
LOAD_ATTR_GETATTRIBUTE_OVERRIDDEN (specialization of LOAD_ATTR)
Linked PRs
- gh-104909: Split BINARY_OP into micro-ops #104910
- gh-104909: Implement conditional stack effects for macros #105748
- GH-104909: Break LOAD_GLOBAL specializations in micro-ops. #106677
- GH-104909: Split
LOAD_ATTR_INSTANCE_VALUEinto micro-ops #106678 - GH-104909: Move unused cache entries from uops to macros #107444
- GH-104909: Break instrumented instructions into micro-ops. #109316
- gh-104909: Split some more insts into ops #109943
- GH-104909: Implement some instrumented instructions with micro-ops. #110025
- gh-104909: Split more LOAD_ATTR specializations #110317
- gh-104909: Make LOAD_ATTR_PROPERTY a viable uop #110560