bolt -instrument

.github/workflows/artifact.yaml

@@ -71,19 +71,20 @@ jobs:
     env:
       CC: "clang"
       CFLAGS: "-Os -fstrict-aliasing -fno-plt -flto=full -emit-llvm"
-      LDFLAGS: "-fuse-ld=lld -Wl,-plugin-opt=also-emit-llvm -Wl,--as-needed -Wl,-zrelro,-znow"
-      RUSTFLAGS: "-C linker=clang -C link-arg=-fuse-ld=lld -C linker-plugin-lto -C lto=fat -C link-arg=-Wl,-zrelro,-znow -Z mir-opt-level=4 -Z virtual-function-elimination -Z threads=2 -D warnings"
+      LDFLAGS: "-fuse-ld=lld -Wl,-plugin-opt=also-emit-llvm -Wl,--as-needed -Wl,-zrelro,-znow -Wl,--emit-relocs"
+      RUSTFLAGS: "-C linker=clang -C link-arg=-fuse-ld=lld -C linker-plugin-lto -C lto=fat -C link-arg=-Wl,-zrelro,-znow -C link-arg=-Wl,--emit-relocs -Z mir-opt-level=4 -Z virtual-function-elimination -Z threads=4 -D warnings"
       PATH: "/__w/orjson/orjson/.venv/bin:/github/home/.cargo/bin:/root/.local/bin:/root/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin"
     container:
       image: fedora:41
+      options: --cap-add SYS_ADMIN
     steps:
 
     - name: cpuinfo
       run: cat /proc/cpuinfo
 
     - name: Build environment pre-clone
       run: |
-        dnf install -y rustup clang lld python${{ matrix.python.version }} git
+        dnf install -y rustup clang lld python${{ matrix.python.version }} git llvm-bolt zip unzip perf
         rustup-init --default-toolchain "${RUST_TOOLCHAIN}-x86_64-unknown-linux-gnu" --profile minimal --component rust-src -y
 
     - uses: actions/checkout@v4
@@ -113,6 +114,13 @@ jobs:
       env:
         PYTHONMALLOC: "debug"
 
+    - name: bolt
+      run: |
+        source .venv/bin/activate
+        ./script/bolt
+        auditwheel show $(find target/wheels/*.whl) | grep '"manylinux_2_17_x86_64"'
+        uv pip install target/wheels/orjson*.whl
+
     - run: source .venv/bin/activate && ./integration/run thread
     - run: source .venv/bin/activate && ./integration/run http
     - run: source .venv/bin/activate && ./integration/run init
@@ -124,7 +132,7 @@ jobs:
       if: matrix.python.version != '3.8'
 
     - name: Store wheels
-      if: "startsWith(github.ref, 'refs/tags/')"
+      # if: "startsWith(github.ref, 'refs/tags/')"
       uses: actions/upload-artifact@v4
       with:
         name: orjson_manylinux_2_17_amd64_${{ matrix.python.version }}

.gitignore

@@ -8,5 +8,6 @@
 /include/cargo
 /perf.*
 /target
+/tmp
 /vendor
 __pycache__

script/bolt

@@ -0,0 +1,77 @@
+#!/usr/bin/env bash
+
+set -eou pipefail
+
+rm -f ./perf.fdata
+rm -rf tmp
+
+wheel=$(find "${CARGO_TARGET_DIR}/wheels" -type f -name "orjson*.whl" -exec basename {} \;)
+
+unzip -q -o "${CARGO_TARGET_DIR}/wheels/${wheel}" -d tmp
+sharedob=$(find tmp/orjson -type f -name "orjson*.so" -exec basename {} \;)
+
+installedobj=$(find .venv -type f -name "orjson*.so")
+llvm-bolt "${installedobj}" --instrument --o orjson.so --jump-tables=none --align-macro-fusion=none
+mv orjson.so "${installedobj}"
+
+echo "collecting perf data..."
+
+./script/profdata
+mv /tmp/prof.fdata ./perf.fdata
+
+llvm-bolt \
+    "tmp/orjson/${sharedob}" \
+    --o "${sharedob}" \
+    --data=perf.fdata \
+    --strict \
+    --use-gnu-stack \
+    --assume-abi \
+    --use-gnu-stack \
+    --deterministic-debuginfo \
+    --sequential-disassembly \
+    --update-debug-sections \
+    --use-old-text \
+    --relocs \
+    --hot-data \
+    --cg-from-perf-data \
+    --cg-ignore-recursive-calls \
+    --inline-ap \
+    --inline-small-functions \
+    --inline-memcpy \
+    --eliminate-unreachable \
+    --plt=all \
+    --peepholes=all \
+    --indirect-call-promotion=all \
+    --icf \
+    --icp-eliminate-loads \
+    --jump-tables=move \
+    --reg-reassign \
+    --use-aggr-reg-reassign \
+    --align-macro-fusion=all \
+    --frame-opt=all \
+    --frame-opt-rm-stores \
+    --sctc-mode=always \
+    --simplify-rodata-loads \
+    --simplify-conditional-tail-calls \
+    --split-all-cold \
+    --reorder-data-inplace \
+    --reorder-blocks=ext-tsp \
+    --reorder-functions=hfsort+ \
+    --reorder-functions-use-hot-size \
+    --strip-rep-ret \
+    --x86-align-branch-boundary-hot-only \
+    --x86-strip-redundant-address-size \
+    --dyno-stats
+
+if [[ "$@" == *"strip"* ]]; then
+    strip --strip-debug "${sharedob}"
+fi
+
+mv -f "${sharedob}" tmp/orjson
+
+./script/fix-dist-info $(find tmp -type f -name "RECORD") $(find tmp -type f -name "orjson*.so")
+
+cd tmp
+zip -9 -r "${wheel}" -xi *
+cd ..
+mv -f "tmp/${wheel}" "${CARGO_TARGET_DIR}/wheels"

script/develop

@@ -12,8 +12,8 @@ export CARGO_TARGET_DIR="${CARGO_TARGET_DIR:-target}"
 echo "CC: ${CC}, LD: ${LD}, LD_LIBRARY_PATH: ${LD_LIBRARY_PATH}"
 
 export CFLAGS="-Os -fstrict-aliasing -fno-plt -flto=full -emit-llvm"
-export LDFLAGS="-fuse-ld=${LD} -Wl,-plugin-opt=also-emit-llvm -Wl,--as-needed -Wl,-zrelro,-znow"
-export RUSTFLAGS="-C linker=${CC} -C link-arg=-fuse-ld=${LD} -C linker-plugin-lto -C lto=fat -C link-arg=-Wl,-zrelro,-znow -Z mir-opt-level=4 -Z virtual-function-elimination -Z threads=8"
+export LDFLAGS="-fuse-ld=${LD} -Wl,-plugin-opt=also-emit-llvm -Wl,--as-needed -Wl,-zrelro,-znow -Wl,--emit-relocs"
+export RUSTFLAGS="-C linker=${CC} -C link-arg=-fuse-ld=${LD} -C linker-plugin-lto -C lto=fat -C link-arg=-Wl,-zrelro,-znow -C link-arg=-Wl,--emit-relocs -Z mir-opt-level=4 -Z virtual-function-elimination -Z threads=8"
 
 maturin build --target="${TARGET}" "$@"
 

script/fix-dist-info

@@ -0,0 +1,38 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: (Apache-2.0 OR MIT)
+# usage:
+# ./script/fix-dist-info $(find tmp -type f -name "RECORD") $(find tmp -type f -name "orjson*.so")
+
+import base64
+import hashlib
+import pathlib
+import sys
+
+record_path = pathlib.Path(sys.argv[1])
+so_path = pathlib.Path(sys.argv[2])
+so_name = sys.argv[2].split("/")[-1]
+
+print(f"Fixing hash of {so_path} in {record_path} ...")
+
+so_bytes = so_path.read_bytes()
+so_len = len(so_bytes)
+
+# URL_SAFE_NO_PAD.encode(Sha256::digest(bytes));
+so_hash = (
+    base64.urlsafe_b64encode(hashlib.sha256(so_bytes).digest())
+    .rstrip(b"=")
+    .decode("utf-8")
+)
+
+record_original = record_path.read_bytes().decode("utf-8")
+record_fixed = []
+for line in record_original.split("\n"):
+    print(line)
+    if line.startswith(f"orjson/{so_name}"):
+        new_line = f"orjson/{so_name},sha256={so_hash},{so_len}"
+        print(f"fixed: {new_line}")
+        record_fixed.append(new_line)
+    else:
+        record_fixed.append(line)
+
+record_path.write_bytes("\n".join(record_fixed).encode("utf-8"))

script/lint

@@ -2,8 +2,10 @@
 
 set -eou pipefail
 
-to_lint="./bench/*.py ./pysrc/orjson/__init__.pyi ./test/*.py script/pydataclass script/pymem
-script/pysort script/pynumpy script/pynonstr script/pycorrectness script/graph integration/init
+to_lint="./bench/*.py ./pysrc/orjson/__init__.pyi ./test/*.py
+script/fix-dist-info script/profdata
+script/pydataclass script/pymem script/pysort script/pynumpy
+script/pynonstr script/pycorrectness script/graph integration/init
 integration/wsgi.py integration/typestubs.py integration/thread"
 
 ruff check ${to_lint} --fix

script/profdata

@@ -0,0 +1,106 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: (Apache-2.0 OR MIT)
+
+import datetime
+import gc
+import lzma
+import os
+import sys
+
+import pytest
+from faker import Faker
+
+from orjson import (
+    OPT_APPEND_NEWLINE,
+    OPT_INDENT_2,
+    OPT_NAIVE_UTC,
+    OPT_OMIT_MICROSECONDS,
+    dumps,
+    loads,
+)
+
+os.sched_setaffinity(os.getpid(), {0, 1})
+gc.disable()
+
+NUM_LOOPS = 100
+NUM_ENTRIES = 5
+
+FAKER_LOCALES = [
+    "ar_AA",
+    "fi_FI",
+    "fil_PH",
+    "he_IL",
+    "ja_JP",
+    "th_TH",
+    "tr_TR",
+    "uk_UA",
+    "vi_VN",
+]
+
+NOW = datetime.datetime.now()
+
+
+def fake():
+    fake = Faker(FAKER_LOCALES)
+    profile_keys = tuple(set(fake.profile().keys()) - {"current_location"})
+
+    DATA = [
+        [
+            {
+                "id": "jzoijxzzoijx",
+                "created_at": NOW,
+                "person": fake.profile(profile_keys),
+                "emoji": fake.emoji(),
+                "😾": "",
+                "a": 12123,
+                "int": 19298012910,
+            }
+            for _ in range(0, NUM_ENTRIES)
+        ]
+        for _ in range(0, NUM_LOOPS)
+    ]
+
+    orjson_opt = OPT_NAIVE_UTC | OPT_OMIT_MICROSECONDS
+    for each in DATA:
+        loads(dumps(each, option=orjson_opt))
+
+
+def _run_fixture(fixture: str, iterations: int):
+    with lzma.open(fixture, "r") as fileh:
+        file_bytes = fileh.read()
+
+    for _ in range(iterations):
+        loads(file_bytes)
+
+    file_obj = loads(file_bytes)
+    for _ in range(iterations * 2):
+        dumps(file_obj)
+
+
+N = 5
+
+
+def fixture():
+    _run_fixture("data/github.json.xz", 40000 * N)
+    _run_fixture("data/twitter.json.xz", 10000 * N)
+    _run_fixture("data/citm_catalog.json.xz", 500 * N)
+    _run_fixture("data/canada.json.xz", 50 * N)
+
+
+def run_pytest():
+    pytest.main(
+        [
+            "--quiet",
+            "--ignore=test/test_api.py",
+            "--ignore=test/test_fixture.py",
+            "--ignore=test/test_fake.py",
+            "--ignore=test/test_ujson.py",
+            "test",
+        ]
+    )
+
+
+if __name__ == "__main__":
+    run_pytest()
+    fake()
+    fixture()