GH-125498: Update JIT builds to use LLVM 19 and preserve_none (GH-125499)

Author: savannahostrowski

.github/workflows/jit.yml

@@ -61,7 +61,7 @@ jobs:
           - true
           - false
         llvm:
-          - 18
+          - 19
         include:
           - target: i686-pc-windows-msvc/msvc
             architecture: Win32
@@ -121,10 +121,15 @@ jobs:
           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 ${{ matrix.debug && '--with-pydebug' || '--enable-optimizations --with-lto' }}
@@ -165,15 +170,19 @@ jobs:
     name: Free-Threaded (Debug)
     needs: interpreter
     runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        llvm:
+          - 19
     steps:
       - uses: actions/checkout@v4
       - uses: actions/setup-python@v5
         with:
           python-version: '3.11'
       - name: Build with JIT enabled and GIL disabled
         run: |
-          sudo bash -c "$(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

Misc/NEWS.d/next/Core_and_Builtins/2024-09-14-20-09-39.gh-issue-123714.o1mbe4.rst

@@ -0,0 +1 @@
+Update JIT compilation to use LLVM 19

Misc/NEWS.d/next/Core_and_Builtins/2024-10-22-04-18-53.gh-issue-125498.cFjPIn.rst

@@ -0,0 +1,4 @@
+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.

Tools/jit/README.md

@@ -7,49 +7,46 @@ This version of CPython can be built with an experimental just-in-time compiler[
 
 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
 

Tools/jit/_llvm.py

@@ -8,7 +8,7 @@
 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")

Tools/jit/_stencils.py

@@ -2,6 +2,7 @@
 
 import dataclasses
 import enum
+import sys
 import typing
 
 import _schema
@@ -132,15 +133,26 @@ class Hole:
     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.symbol == other.symbol
             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:

Tools/jit/_targets.py

@@ -26,7 +26,6 @@
 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
@@ -39,7 +38,6 @@ class _Target(typing.Generic[_S, _R]):
     _: dataclasses.KW_ONLY
     alignment: int = 1
     args: typing.Sequence[str] = ()
-    ghccc: bool = False
     prefix: str = ""
     stable: bool = False
     debug: bool = False
@@ -88,11 +86,7 @@ async def _parse(self, path: pathlib.Path) -> _stencils.StencilGroup:
         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)
@@ -112,9 +106,6 @@ def _handle_relocation(
     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}",
@@ -128,6 +119,7 @@ async def _compile(
             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.
@@ -143,44 +135,12 @@ async def _compile(
             # 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]:
@@ -519,7 +479,6 @@ def _handle_relocation(
 
 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="_")
@@ -535,16 +494,20 @@ def get_target(host: str) -> _COFF | _ELF | _MachO:
         ]
         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

Tools/jit/_writer.py

@@ -65,7 +65,7 @@ def _dump_stencil(opname: str, group: _stencils.StencilGroup) -> typing.Iterator
             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)}"

Tools/jit/jit.h

@@ -0,0 +1,4 @@
+// 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;

Tools/jit/template.c

@@ -21,6 +21,8 @@
 
 #include "ceval_macros.h"
 
+#include "jit.h"
+
 #undef CURRENT_OPARG
 #define CURRENT_OPARG() (_oparg)
 
@@ -49,7 +51,7 @@
 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
@@ -72,7 +74,7 @@ do {  \
 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
@@ -86,7 +88,7 @@ do {                                                         \
 
 #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: