Method _compile

Tools/jit/_targets.py:133–201

(
        self, opname: str, c: pathlib.Path, tempdir: pathlib.Path
    )

        raise NotImplementedError(type(self))

    async def _compile(
        self, opname: str, c: pathlib.Path, tempdir: pathlib.Path
    ) -> _stencils.StencilGroup:
        s = tempdir / f"{opname}.s"
        o = tempdir / f"{opname}.o"
        args_s = [
            f"--target={self.triple}",
            "-DPy_BUILD_CORE_MODULE",
            "-D_DEBUG" if self.debug else "-DNDEBUG",
            f"-DSUPPORTS_SMALL_CONSTS={1 if self.optimizer.supports_small_constants else 0}",
            f"-D_JIT_OPCODE={opname}",
            "-D_PyJIT_ACTIVE",
            "-D_Py_JIT",
            f"-I{self.pyconfig_dir}",
            f"-I{CPYTHON / 'Include'}",
            f"-I{CPYTHON / 'Include' / 'internal'}",
            f"-I{CPYTHON / 'Include' / 'internal' / 'mimalloc'}",
            f"-I{CPYTHON / 'Python'}",
            f"-I{CPYTHON / 'Tools' / 'jit'}",
            # -O2 and -O3 include some optimizations that make sense for
            # standalone functions, but not for snippets of code that are going
            # to be laid out end-to-end (like ours)... common examples include
            # passes like tail-duplication, or aligning jump targets with nops.
            # -Os is equivalent to -O2 with many of these problematic passes
            # disabled. Based on manual review, for *our* purposes it usually
            # generates better code than -O2 (and -O2 usually generates better
            # code than -O3). As a nice benefit, it uses less memory too:
            "-Os",
            "-S",
            # Shorten full absolute file paths in the generated code (like the
            # __FILE__ macro and assert failure messages) for reproducibility:
            f"-ffile-prefix-map={CPYTHON}=.",
            f"-ffile-prefix-map={tempdir}=.",
            # This debug info isn't necessary, and bloats out the JIT'ed code.
            # We *may* be able to re-enable this, process it, and JIT it for a
            # nicer debugging experience... but that needs a lot more research:
            "-fno-asynchronous-unwind-tables",
            # Don't call built-in functions that we can't find or patch:
            "-fno-builtin",
            # Don't call stack-smashing canaries that we can't find or patch:
            "-fno-stack-protector",
            "-std=c11",
            "-o",
            f"{s}",
            f"{c}",
        ]
        is_shim = opname == "shim"
        if self.frame_pointers:
            frame_pointer = "all" if is_shim else "reserved"
            args_s += ["-Xclang", f"-mframe-pointer={frame_pointer}"]
        args_s += self.args
        # Allow user-provided CFLAGS to override any defaults
        args_s += shlex.split(self.cflags)
        await _llvm.run(
            "clang", args_s, echo=self.verbose, llvm_version=self.llvm_version
        )
        if not is_shim:
            self.optimizer(