- true
- false
llvm:
- - 18
+ - 19
include:
- target: i686-pc-windows-msvc/msvc
architecture: Win32
choco install llvm --allow-downgrade --no-progress --version ${{ matrix.llvm }}.1.0
./PCbuild/build.bat --experimental-jit ${{ matrix.debug && '-d' || '' }} -p ${{ matrix.architecture }}
+ # The `find` line is required as a result of https://github.com/actions/runner-images/issues/9966.
+ # This is a bug in the macOS runner image where the pre-installed Python is installed in the same
+ # directory as the Homebrew Python, which causes the build to fail for macos-13. This line removes
+ # the symlink to the pre-installed Python so that the Homebrew Python is used instead.
- name: Native macOS
if: runner.os == 'macOS'
run: |
brew update
+ find /usr/local/bin -lname '*/Library/Frameworks/Python.framework/*' -delete
brew install llvm@${{ matrix.llvm }}
SDKROOT="$(xcrun --show-sdk-path)" \
./configure --enable-experimental-jit ${{ matrix.debug && '--with-pydebug' || '--enable-optimizations --with-lto' }}
name: Free-Threaded (Debug)
needs: interpreter
runs-on: ubuntu-latest
+ strategy:
+ matrix:
+ llvm:
+ - 19
steps:
- uses: actions/checkout@v4
- uses: actions/setup-python@v5
python-version: '3.11'
- name: Build with JIT enabled and GIL disabled
run: |
- sudo bash -c "$(wget -O - https://apt.llvm.org/llvm.sh)" ./llvm.sh 18
- export PATH="$(llvm-config-18 --bindir):$PATH"
+ sudo bash -c "$(wget -O - https://apt.llvm.org/llvm.sh)" ./llvm.sh ${{ matrix.llvm }}
+ export PATH="$(llvm-config-${{ matrix.llvm }} --bindir):$PATH"
./configure --enable-experimental-jit --with-pydebug --disable-gil
make all --jobs 4
- name: Run tests
--- /dev/null
+Update JIT compilation to use LLVM 19
--- /dev/null
+The JIT has been updated to leverage Clang 19’s new ``preserve_none`` attribute,
+which supports more platforms and is more useful than LLVM's existing ``ghccc``
+calling convention. This also removes the need to manually patch the calling
+convention in LLVM IR, simplifying the JIT compilation process.
The JIT compiler does not require end users to install any third-party dependencies, but part of it must be *built* using LLVM[^why-llvm]. You are *not* required to build the rest of CPython using LLVM, or even the same version of LLVM (in fact, this is uncommon).
-LLVM version 18 is required. Both `clang` and `llvm-readobj` need to be installed and discoverable (version suffixes, like `clang-18`, are okay). It's highly recommended that you also have `llvm-objdump` available, since this allows the build script to dump human-readable assembly for the generated code.
+LLVM version 19 is required. Both `clang` and `llvm-readobj` need to be installed and discoverable (version suffixes, like `clang-19`, are okay). It's highly recommended that you also have `llvm-objdump` available, since this allows the build script to dump human-readable assembly for the generated code.
It's easy to install all of the required tools:
### Linux
-Install LLVM 18 on Ubuntu/Debian:
+Install LLVM 19 on Ubuntu/Debian:
```sh
wget https://apt.llvm.org/llvm.sh
chmod +x llvm.sh
-sudo ./llvm.sh 18
+sudo ./llvm.sh 19
```
-Install LLVM 18 on Fedora Linux 40 or newer:
+Install LLVM 19 on Fedora Linux 40 or newer:
```sh
-sudo dnf install 'clang(major) = 18' 'llvm(major) = 18'
+sudo dnf install 'clang(major) = 19' 'llvm(major) = 19'
```
### macOS
-Install LLVM 18 with [Homebrew](https://brew.sh):
+Install LLVM 19 with [Homebrew](https://brew.sh):
```sh
-brew install llvm@18
+brew install llvm@19
```
Homebrew won't add any of the tools to your `$PATH`. That's okay; the build script knows how to find them.
### Windows
-Install LLVM 18 [by searching for it on LLVM's GitHub releases page](https://github.com/llvm/llvm-project/releases?q=18), clicking on "Assets", downloading the appropriate Windows installer for your platform (likely the file ending with `-win64.exe`), and running it. **When installing, be sure to select the option labeled "Add LLVM to the system PATH".**
+Install LLVM 19 [by searching for it on LLVM's GitHub releases page](https://github.com/llvm/llvm-project/releases?q=19), clicking on "Assets", downloading the appropriate Windows installer for your platform (likely the file ending with `-win64.exe`), and running it. **When installing, be sure to select the option labeled "Add LLVM to the system PATH".**
Alternatively, you can use [chocolatey](https://chocolatey.org):
```sh
-choco install llvm --version=18.1.6
+choco install llvm --version=19.1.0
```
-### Dev Containers
-
-If you are working CPython in a [Codespaces instance](https://devguide.python.org/getting-started/setup-building/#using-codespaces), there's no need to install LLVM as the Fedora 40 base image includes LLVM 18 out of the box.
## Building
import subprocess
import typing
-_LLVM_VERSION = 18
+_LLVM_VERSION = 19
_LLVM_VERSION_PATTERN = re.compile(rf"version\s+{_LLVM_VERSION}\.\d+\.\d+\S*\s+")
_P = typing.ParamSpec("_P")
import dataclasses
import enum
+import sys
import typing
import _schema
def __post_init__(self) -> None:
self.func = _PATCH_FUNCS[self.kind]
- def fold(self, other: typing.Self) -> typing.Self | None:
+ def fold(self, other: typing.Self, body: bytes) -> typing.Self | None:
"""Combine two holes into a single hole, if possible."""
+ instruction_a = int.from_bytes(
+ body[self.offset : self.offset + 4], byteorder=sys.byteorder
+ )
+ instruction_b = int.from_bytes(
+ body[other.offset : other.offset + 4], byteorder=sys.byteorder
+ )
+ reg_a = instruction_a & 0b11111
+ reg_b1 = instruction_b & 0b11111
+ reg_b2 = (instruction_b >> 5) & 0b11111
+
if (
self.offset + 4 == other.offset
and self.value == other.value
and self.addend == other.addend
and self.func == "patch_aarch64_21rx"
and other.func == "patch_aarch64_12x"
+ and reg_a == reg_b1 == reg_b2
):
# These can *only* be properly relaxed when they appear together and
# patch the same value:
PYTHON_EXECUTOR_CASES_C_H = CPYTHON / "Python" / "executor_cases.c.h"
TOOLS_JIT_TEMPLATE_C = TOOLS_JIT / "template.c"
-
_S = typing.TypeVar("_S", _schema.COFFSection, _schema.ELFSection, _schema.MachOSection)
_R = typing.TypeVar(
"_R", _schema.COFFRelocation, _schema.ELFRelocation, _schema.MachORelocation
_: dataclasses.KW_ONLY
alignment: int = 1
args: typing.Sequence[str] = ()
- ghccc: bool = False
prefix: str = ""
stable: bool = False
debug: bool = False
sections: list[dict[typing.Literal["Section"], _S]] = json.loads(output)
for wrapped_section in sections:
self._handle_section(wrapped_section["Section"], group)
- # The trampoline's entry point is just named "_ENTRY", since on some
- # platforms we later assume that any function starting with "_JIT_" uses
- # the GHC calling convention:
- entry_symbol = "_JIT_ENTRY" if "_JIT_ENTRY" in group.symbols else "_ENTRY"
- assert group.symbols[entry_symbol] == (_stencils.HoleValue.CODE, 0)
+ assert group.symbols["_JIT_ENTRY"] == (_stencils.HoleValue.CODE, 0)
if group.data.body:
line = f"0: {str(bytes(group.data.body)).removeprefix('b')}"
group.data.disassembly.append(line)
async def _compile(
self, opname: str, c: pathlib.Path, tempdir: pathlib.Path
) -> _stencils.StencilGroup:
- # "Compile" the trampoline to an empty stencil group if it's not needed:
- if opname == "trampoline" and not self.ghccc:
- return _stencils.StencilGroup()
o = tempdir / f"{opname}.o"
args = [
f"--target={self.triple}",
f"-I{CPYTHON / 'Include' / 'internal'}",
f"-I{CPYTHON / 'Include' / 'internal' / 'mimalloc'}",
f"-I{CPYTHON / 'Python'}",
+ f"-I{CPYTHON / 'Tools' / 'jit'}",
"-O3",
"-c",
# This debug info isn't necessary, and bloats out the JIT'ed code.
# Don't call stack-smashing canaries that we can't find or patch:
"-fno-stack-protector",
"-std=c11",
+ "-o",
+ f"{o}",
+ f"{c}",
*self.args,
]
- if self.ghccc:
- # This is a bit of an ugly workaround, but it makes the code much
- # smaller and faster, so it's worth it. We want to use the GHC
- # calling convention, but Clang doesn't support it. So, we *first*
- # compile the code to LLVM IR, perform some text replacements on the
- # IR to change the calling convention(!), and then compile *that*.
- # Once we have access to Clang 19, we can get rid of this and use
- # __attribute__((preserve_none)) directly in the C code instead:
- ll = tempdir / f"{opname}.ll"
- args_ll = args + [
- # -fomit-frame-pointer is necessary because the GHC calling
- # convention uses RBP to pass arguments:
- "-S",
- "-emit-llvm",
- "-fomit-frame-pointer",
- "-o",
- f"{ll}",
- f"{c}",
- ]
- await _llvm.run("clang", args_ll, echo=self.verbose)
- ir = ll.read_text()
- # This handles declarations, definitions, and calls to named symbols
- # starting with "_JIT_":
- ir = re.sub(
- r"(((noalias|nonnull|noundef) )*ptr @_JIT_\w+\()", r"ghccc \1", ir
- )
- # This handles calls to anonymous callees, since anything with
- # "musttail" needs to use the same calling convention:
- ir = ir.replace("musttail call", "musttail call ghccc")
- # Sometimes *both* replacements happen at the same site, so fix it:
- ir = ir.replace("ghccc ghccc", "ghccc")
- ll.write_text(ir)
- args_o = args + ["-Wno-unused-command-line-argument", "-o", f"{o}", f"{ll}"]
- else:
- args_o = args + ["-o", f"{o}", f"{c}"]
- await _llvm.run("clang", args_o, echo=self.verbose)
+ await _llvm.run("clang", args, echo=self.verbose)
return await self._parse(o)
async def _build_stencils(self) -> dict[str, _stencils.StencilGroup]:
def get_target(host: str) -> _COFF | _ELF | _MachO:
"""Build a _Target for the given host "triple" and options."""
- # ghccc currently crashes Clang when combined with musttail on aarch64. :(
target: _COFF | _ELF | _MachO
if re.fullmatch(r"aarch64-apple-darwin.*", host):
target = _MachO(host, alignment=8, prefix="_")
]
target = _ELF(host, alignment=8, args=args)
elif re.fullmatch(r"i686-pc-windows-msvc", host):
- args = ["-DPy_NO_ENABLE_SHARED"]
- target = _COFF(host, args=args, ghccc=True, prefix="_")
+ args = [
+ "-DPy_NO_ENABLE_SHARED",
+ # __attribute__((preserve_none)) is not supported
+ "-Wno-ignored-attributes",
+ ]
+ target = _COFF(host, args=args, prefix="_")
elif re.fullmatch(r"x86_64-apple-darwin.*", host):
- target = _MachO(host, ghccc=True, prefix="_")
+ target = _MachO(host, prefix="_")
elif re.fullmatch(r"x86_64-pc-windows-msvc", host):
args = ["-fms-runtime-lib=dll"]
- target = _COFF(host, args=args, ghccc=True)
+ target = _COFF(host, args=args)
elif re.fullmatch(r"x86_64-.*-linux-gnu", host):
args = ["-fpic"]
- target = _ELF(host, args=args, ghccc=True)
+ target = _ELF(host, args=args)
else:
raise ValueError(host)
return target
if skip:
skip = False
continue
- if pair and (folded := hole.fold(pair)):
+ if pair and (folded := hole.fold(pair, stencil.body)):
skip = True
hole = folded
yield f" {hole.as_c(part)}"
--- /dev/null
+// To use preserve_none in JIT builds, we need to declare a separate function
+// pointer with __attribute__((preserve_none)), since this attribute may not be
+// supported by the compiler used to build the rest of the interpreter.
+typedef jit_func __attribute__((preserve_none)) jit_func_preserve_none;
#include "ceval_macros.h"
+#include "jit.h"
+
#undef CURRENT_OPARG
#define CURRENT_OPARG() (_oparg)
do { \
OPT_STAT_INC(traces_executed); \
__attribute__((musttail)) \
- return ((jit_func)((EXECUTOR)->jit_side_entry))(frame, stack_pointer, tstate); \
+ return ((jit_func_preserve_none)((EXECUTOR)->jit_side_entry))(frame, stack_pointer, tstate); \
} while (0)
#undef GOTO_TIER_ONE
do { \
PyAPI_DATA(void) ALIAS; \
__attribute__((musttail)) \
- return ((jit_func)&ALIAS)(frame, stack_pointer, tstate); \
+ return ((jit_func_preserve_none)&ALIAS)(frame, stack_pointer, tstate); \
} while (0)
#undef JUMP_TO_JUMP_TARGET
#define TIER_TWO 2
-_Py_CODEUNIT *
+__attribute__((preserve_none)) _Py_CODEUNIT *
_JIT_ENTRY(_PyInterpreterFrame *frame, _PyStackRef *stack_pointer, PyThreadState *tstate)
{
// Locals that the instruction implementations expect to exist:
#include "pycore_frame.h"
#include "pycore_jit.h"
-// This is where the calling convention changes, on platforms that require it.
-// The actual change is patched in while the JIT compiler is being built, in
-// Tools/jit/_targets.py. On other platforms, this function compiles to nothing.
+#include "jit.h"
+
_Py_CODEUNIT *
-_ENTRY(_PyInterpreterFrame *frame, _PyStackRef *stack_pointer, PyThreadState *tstate)
+_JIT_ENTRY(_PyInterpreterFrame *frame, _PyStackRef *stack_pointer, PyThreadState *tstate)
{
// This is subtle. The actual trace will return to us once it exits, so we
// need to make sure that we stay alive until then. If our trace side-exits
Py_INCREF(executor);
// Note that this is *not* a tail call:
PyAPI_DATA(void) _JIT_CONTINUE;
- _Py_CODEUNIT *target = ((jit_func)&_JIT_CONTINUE)(frame, stack_pointer, tstate);
+ _Py_CODEUNIT *target = ((jit_func_preserve_none)&_JIT_CONTINUE)(frame, stack_pointer, tstate);
Py_SETREF(tstate->previous_executor, executor);
return target;
}
projects / thirdparty / Python / cpython.git / commitdiff
