Merge pull request #399 from faster-cpython/clang-separate

Separate clang from tail calling

Author: mdboom

CHANGELOG.md

@@ -2,12 +2,63 @@
 
 ## v2.0.0
 
-Most of the work has moved from GitHub Actions `.yml` files to Python code in `workflow.py`.
-In the future, this will allow supporting more workflow engines beyond just GitHub Actions.
+### Moving more code to Python
 
-**Migration note**: After running `python -m bench_runner install` to update your local files, but sure to add the new `workflow_bootstrap.py` file to your git repository.
+Most of the code to orchestrate the benchmarks at a high level has moved from
+GitHub Actions `.yml` files to Python code in `workflow.py`. In the future, this
+will allow supporting more workflow engines beyond just GitHub Actions.
 
-### New configuration
+**Migration note**: After running `python -m bench_runner install` to update
+your local files, be sure to add the new `workflow_bootstrap.py` file to your
+git repository.
+
+### Decoupling compiler configuration from tail-calling
+
+Previously, bench_runner had a flag, `CLANG`, that both (a) built with clang 19
+or later and (b) built the tail calling interpreter. This has been replaced with
+a new flag `TAILCALL` that only sets the flags to build with the tail calling
+interpreter. It is now up to the user to select a machine configured with a
+compiler that supports the tail calling interpreter.
+
+For machines with a clang-19 or later installed these example machine
+configurations could be used in your `bench_runner.toml`. The exact details will
+depend on your distribution and method of installing clang, etc.:
+
+```toml
+[runners.linux_clang]
+os = "linux"
+arch = "x86_64"
+hostname = "pyperf"
+github_runner_name = "linux-x86_64-linux"
+include_in_all = false
+[runners.linux_clang.env]
+CC = "$(which clang-19)"
+LLVM_AR = "$(which llvm-ar-19)"
+LLVM_PROFDATA = "$(which llvm-profdata-19)"
+
+[runners.darwin_clang19]
+os = "darwin"
+arch = "arm64"
+hostname = "CPythons-Mac-mini.local"
+github_runner_name = "darwin-arm64-darwin"
+[runners.darwin_clang19.env]
+PATH = "$(brew --prefix llvm)/bin:$PATH"
+CC = "$(brew --prefix llvm)/bin/clang"
+LDFLAGS = "-L$(brew --prefix llvm)/lib"
+CFLAGS = "-L$(brew --prefix llvm)/include"
+
+[runners.pythonperf1_clang]
+os = "windows"
+arch = "x86_64"
+hostname = "WIN-5FKPU9U7KDT"
+github_runner_name = "windows-x86_64-pythonperf1"
+include_in_all = false
+[runners.pythonperf1_clang.env]
+BUILD_DEST = "PCBuild/amd64"
+PYTHON_CONFIGURE_FLAGS = '`"/p:PlatformToolset=clangcl`" `"/p:LLVMInstallDir=C:\Program Files\LLVM`" `"/p:LLVMToolsVersion=19.1.6`"'
+```
+
+### Controlling number of cores used for builds
 
 Runners have a new configuration `use_cores` to control the number of CPU cores
 used to build CPython. By default, this will use all available cores, but some

bench_runner/flags.py

@@ -17,7 +17,7 @@ class Flag:
     Flag("tier2", "PYTHON_UOPS", "tier 2 interpreter", "T2"),
     Flag("jit", "JIT", "JIT", "JIT"),
     Flag("nogil", "NOGIL", "free threading", "NOGIL"),
-    Flag("clang", "CLANG", "build with latest clang and tailcall", "CLANG"),
+    Flag("tailcall", "TAILCALL", "tail-calling interpreter", "TAILCALL"),
 ]
 
 

bench_runner/plot.py

@@ -437,7 +437,16 @@ def get_comparison_value(ref, r):
             cfg["runners"], cfg["names"], cfg["colors"], cfg["styles"], cfg["markers"]
         ):
             runner_results = commits.get(runner, {})
-            base_results = runner_results.get("", {})
+            # For tailcall, we want to compare against the default compiler
+            # which is a different "machine"
+            if flag == "TAILCALL":
+                # It's ok if the removesuffix fails -- it's fine to compare with
+                # the same machine if the "default" equivalent isn't available.
+                base_results = commits.get(runner.removesuffix("_clang"), {}).get(
+                    "", {}
+                )
+            else:
+                base_results = runner_results.get("", {})
 
             line = []
             for cpython_hash, r in runner_results.get(flag, {}).items():

bench_runner/scripts/workflow.py

@@ -135,18 +135,6 @@ def compile_unix(cpython: PathLike, flags: list[str], pgo: bool, pystats: bool)
     cfg = config.get_config_for_current_runner()
 
     env = os.environ.copy()
-    if "CLANG" in flags:
-        match util.get_simple_platform():
-            case "linux":
-                env["CC"] = util.safe_which("clang-19")
-                env["LLVM_AR"] = util.safe_which("llvm-ar-19")
-                env["LLVM_PROFDATA"] = util.safe_which("llvm-profdata-19")
-            case "macos":
-                llvm_prefix = util.get_brew_prefix("llvm")
-                env["PATH"] = f"{llvm_prefix}/bin:{env['PATH']}"
-                env["CC"] = f"{llvm_prefix}/bin/clang"
-                env["LDFLAGS"] = f"-L{llvm_prefix}/lib"
-                env["CFLAGS"] = f"-I{llvm_prefix}/include"
 
     if util.get_simple_platform() == "macos":
         openssl_prefix = util.get_brew_prefix("[email protected]")
@@ -165,7 +153,7 @@ def compile_unix(cpython: PathLike, flags: list[str], pgo: bool, pystats: bool)
         args.append("--enable-experimental-jit=yes")
     if "NOGIL" in flags:
         args.append("--disable-gil")
-    if "CLANG" in flags:
+    if "TAILCALL" in flags:
         args.append("--with-tail-call-interp")
     args.append("--enable-option-checking=fatal")
     if configure_flags := os.environ.get("PYTHON_CONFIGURE_FLAGS"):
@@ -199,15 +187,10 @@ def compile_windows(
         args.append("--experimental-jit-interpreter")
     if "NOGIL" in flags:
         args.append("--disable-gil")
-    if "CLANG" in flags:
-        args.extend(
-            [
-                "--tail-call-interp",
-                '"/p:PlatformToolset=clangcl"',
-                '"/p:LLVMInstallDir=C:\\Program Files\\LLVM"',
-                '"/p:LLVMToolsVersion=19.1.6"',
-            ]
-        )
+    if "TAILCALL" in flags:
+        args.append("--tail-call-interp")
+    if configure_flags := os.environ.get("PYTHON_CONFIGURE_FLAGS"):
+        args.extend(shlex.split(configure_flags))
 
     with contextlib.chdir(cpython):
         subprocess.check_call(

bench_runner/templates/_weekly.src.yml

@@ -79,7 +79,7 @@ jobs:
       tier2: false
       jit: false
       nogil: false
-      clang: true
+      tailcall: true
     secrets: inherit
 
   pystats:

tests/test_gh.py

@@ -59,5 +59,5 @@ def get_args(args, **kwargs):
         "-f",
         "nogil=false",
         "-f",
-        "clang=false",
+        "tailcall=false",
     ]