unix: enable BOLT optimizations on Linux x86-64

BOLT appears to only work on Linux. And it only supports x86-64 and
aarch64. And since we don't have PGO builds for aarch64, that means we
can only ship it on x86-64. (We can't even enable support in the build
system since we don't have a sufficiently modern Clang to use on aarch64.)

In order to get this to work, I had to force DWARF version 4 to avoid
BOLT crashing. This is related to #155. I plan to file an upstream issue
on the LLVM crash.

cpython-unix/build-cpython.sh

@@ -333,6 +333,9 @@ fi
 
 if [ -n "${CPYTHON_OPTIMIZED}" ]; then
     CONFIGURE_FLAGS="${CONFIGURE_FLAGS} --enable-optimizations"
+    if [[ -n "${PYTHON_MEETS_MINIMUM_VERSION_3_12}" && -n "${BOLT_CAPABLE}" ]]; then
+        CONFIGURE_FLAGS="${CONFIGURE_FLAGS} --enable-bolt"
+    fi
 fi
 
 if [ -n "${CPYTHON_LTO}" ]; then

cpython-unix/build.py

@@ -93,6 +93,9 @@ def add_target_env(env, build_platform, target_triple, build_env):
     env["HOST_CXX"] = settings["host_cxx"]
     env["CC"] = settings["target_cc"]
 
+    if settings.get("bolt_capable"):
+        env["BOLT_CAPABLE"] = "1"
+
     env["PYBUILD_PLATFORM"] = build_platform
     env["TOOLS_PATH"] = build_env.tools_path
 

cpython-unix/targets.yml

@@ -45,6 +45,9 @@
 #
 # openssl_target
 #    Name of OpenSSL platform build target.
+#
+# bolt_capable
+#    Whether the target is capable of performing BOLT optimizations.
 
 ---
 
@@ -720,6 +723,8 @@ x86_64-unknown-linux-gnu:
   target_cc: clang
   target_cflags:
     - '-fvisibility=hidden'
+    # Needed to prevent BOLT from crashing.
+    - '-fdebug-default-version=4'
   needs:
     - autoconf
     - bdb
@@ -745,6 +750,7 @@ x86_64-unknown-linux-gnu:
     - xz
     - zlib
   openssl_target: linux-x86_64
+  bolt_capable: true
 
 x86_64_v2-unknown-linux-gnu:
   host_platforms:
@@ -762,6 +768,8 @@ x86_64_v2-unknown-linux-gnu:
   target_cflags:
     - '-march=x86-64-v2'
     - '-fvisibility=hidden'
+    # Needed to prevent BOLT from crashing.
+    - '-fdebug-default-version=4'
   needs:
     - autoconf
     - bdb
@@ -787,6 +795,7 @@ x86_64_v2-unknown-linux-gnu:
     - xz
     - zlib
   openssl_target: linux-x86_64
+  bolt_capable: true
 
 x86_64_v3-unknown-linux-gnu:
   host_platforms:
@@ -804,6 +813,8 @@ x86_64_v3-unknown-linux-gnu:
   target_cflags:
     - '-march=x86-64-v3'
     - '-fvisibility=hidden'
+    # Needed to prevent BOLT from crashing.
+    - '-fdebug-default-version=4'
   needs:
     - autoconf
     - bdb
@@ -829,6 +840,7 @@ x86_64_v3-unknown-linux-gnu:
     - xz
     - zlib
   openssl_target: linux-x86_64
+  bolt_capable: true
 
 x86_64_v4-unknown-linux-gnu:
   host_platforms:
@@ -846,6 +858,8 @@ x86_64_v4-unknown-linux-gnu:
   target_cflags:
     - '-march=x86-64-v4'
     - '-fvisibility=hidden'
+    # Needed to prevent BOLT from crashing.
+    - '-fdebug-default-version=4'
   needs:
     - autoconf
     - bdb
@@ -871,6 +885,7 @@ x86_64_v4-unknown-linux-gnu:
     - xz
     - zlib
   openssl_target: linux-x86_64
+  bolt_capable: true
 
 x86_64-unknown-linux-musl:
   host_platforms:

docs/running.rst

@@ -138,6 +138,9 @@ Common configurations include:
 ``pgo``
    Profile guided optimization.
 
+   Starting with CPython 3.12, BOLT is also applied alongside traditional
+   PGO on platforms supporting BOLT. (Currently just Linux x86-64.)
+
 ``lto``
    Link-time optimization.