Repository navigation
gh-154124: Reserve frame pointer on Clang FP+TC builds - #159204
Fidget-Spinner wants to merge 15 commits into
Conversation
# Conflicts: # Makefile.pre.in
pablogsal
left a comment
pablogsal
left a comment
There was a problem hiding this comment.
a first pass late at night :)
|
|
||
| def function_proto(name: str) -> str: | ||
| return f"static PyObject *Py_PRESERVE_NONE_CC _TAIL_CALL_{name}(TAIL_CALL_PARAMS)" | ||
| def function_signature(name: str) -> str: |
There was a problem hiding this comment.
This makes every handler and both tables external symbols, and neither _TAIL_CALL_* nor instruction_funcptr_handler_table is in our namespace. Unless I am missing something, make smelly will fail on a --with-tail-call-interp build because these show up as global symbols in libpython3.X.a (CI only runs it on a non tail-call build so we don't see it).
_testinternalcapi/interpreter.c now also defines the same names, so we are relying on -fvisibility=hidden to not interpose them, and a static build of that module would fail to link.
Can we give these a _Py prefix, and use a different one for the test interpreter?
| @@ -0,0 +1,13 @@ | |||
| #include "pyconfig.h" // _Py_TAIL_CALL_INTERP | |||
There was a problem hiding this comment.
We should add this file to the paths filter in .github/workflows/tail-call.yml, otherwise changes here will not trigger the tail-call jobs. Probably Python/opcode_targets.h as well now that the tables are defined from it.
| @@ -0,0 +1 @@ | |||
| Reserve frame pointer on tail calling builds using newer Clang. This recovers most of the performance hit from frame pointers on tail calling builds using recent versions of Clang | |||
There was a problem hiding this comment.
Nit: can we say which Clang version is needed instead of "newer"/"recent", and reference :option:--with-tail-call-interp? This also says "tail calling builds using ... Clang" twice and is missing the final period.
|
|
||
|
|
||
| def function_proto(name: str) -> str: | ||
| return f"extern {function_signature(name)}" |
There was a problem hiding this comment.
MODULE_BUILDTYPE=static now links duplicate TAIL_CALL* functions and dispatch tables from the core and _testinternalcapi. Can we keep the test interpreter's symbols static?
| Python/bltinmodule.o \ | ||
| Python/brc.o \ | ||
| Python/ceval.o \ | ||
| Python/ceval_tailcall.o \ |
There was a problem hiding this comment.
We also need ceval_tailcall.o in DTRACE_DEPS. Its probe-enabled checks require dtrace -G processing on FreeBSD, otherwise tail-call builds with DTrace fail to link.
2 participants
So both TC and Computed goto interpreter take a 2% hit when enabling frame pointers.
However, on TC interpreter, if we just let the first handler (
start_frame) push a frame pointer, then reserve it throughout the opcode-opcode tail calls, we remove all overhead of the prologue/epilogue on each tail call handler except for the first one entering the interpreter.I measured on my machine, and it's a 2% speedup on x86-64 for fastmark. In other words, this makes frame pointer builds zero-overhead for tail calling interpreter!
To do this without horrible pragma hacks, we need to move the frame-pointer reserved opcode handlers to a separate compilation unit and pass in separate flags for that compilation unit. The rest of the CPython still builds with default frame pointer flags.