Reland PR #1900 "Add platform-specific release configuration"

The original PR failed to apply the `release` configuration to
release_unix, leading to a binary size regression, especially on macOS.
This was caused by bazel using fastbuild mode instead of opt mode,
Rust code was not being optimized and on macOS -DDEBUG was defined
through the bazel toolchain defaults.

- Fix a typo and clarify Windows release config comment

Original commit message:
- Match macOS wd_benchmark link options with wd_binary options
- Correct spelling of WIN32_LEAN_AND_MEAN
- /await and /GF are ignored by clang-cl, no need to pass them in the
  Windows build.

.bazelrc

@@ -101,10 +101,8 @@ build --@rules_rust//:extra_exec_rustc_flags=-C,panic=abort,-C,debug-assertions=
 build:fastdbg --@rules_rust//:extra_rustc_flags=-C,panic=unwind,-C,debug-assertions=y,-C,debuginfo=1,-C,embed-bitcode=n
 build:fastdbg --@rules_rust//:extra_exec_rustc_flags=-C,panic=unwind,-C,debug-assertions=y,-C,debuginfo=1,-C,embed-bitcode=n
 
-# TODO(later): -C codegen-units=1 improves code size and quality, should be enabled in a future
-# release configuration even if lto is off. Similarly, adding -C lto=thin would improve binary size
-# significantly – disable it for now due to compile errors and wrong code generation when bazel and
-# rust use different LLVM versions.
+# Adding -C lto=thin here would improve binary size significantly – disable it for now due to
+# compile errors and wrong code generation when bazel and rust use different LLVM versions.
 build:thin-lto --@rules_rust//:extra_rustc_flags=-C,panic=abort,-C,codegen-units=1,-C,embed-bitcode=n
 
 # common sanitizers options
@@ -178,13 +176,6 @@ build:linux --linkopt="-Wl,--gc-sections" --linkopt="-Wl,-O2"
 build:linux --cxxopt="-ffunction-sections" --host_cxxopt="-ffunction-sections"
 build:linux --cxxopt="-fdata-sections" --host_cxxopt="-fdata-sections"
 
-# TODO(later): rustc's -C,codegen-units=1 and ld/lld's -Wl,-O2 on Linux/possibly Windows should
-# bring an additional binary size improvement. These should only be enabled in opt mode which won't
-# be possible in bazelrc without establishing a release configuration, but it can be done using the
-# extra_rustc_flags parameter in rust_register_toolchains() and by modifying linkopts in kj_test()
-# and wd_cc_binary() similar to how the use_dead_strip flag is used. The binary size improvement is
-# modest (~ 300kb), but if we encounter size concerns again this would be a good place to start.
-
 #
 # Windows
 #
@@ -215,19 +206,53 @@ build:windows --extra_execution_platforms=//:x64_windows-clang-cl
 build:windows_no_dbg -c opt
 build:windows_no_dbg --copt='-O0' --host_copt='-O0'
 build:windows_no_dbg --copt='/Od' --host_copt='/Od'
-build:windows_no_dbg --copt='/INCREMENTAL:NO' --host_copt='/INCREMENTAL:NO'
+build:windows_no_dbg --linkopt='/INCREMENTAL:NO' --host_linkopt='/INCREMENTAL:NO'
 build:windows_no_dbg --features=-smaller_binary --features=-disable_assertions_feature
 
 # Mitigate the large size impact of Windows debug binaries somewhat by applying string merging and
 # linker garbage collection.
 build:windows_dbg --config=debug
-build:windows_dbg --copt='/Gy' --copt='/Gw' --copt='/GF'
+build:windows_dbg --copt='/Gy' --copt='/Gw'
 build:windows_dbg --linkopt='/OPT:REF'
+build:windows_dbg --linkopt='/OPT:LLDTAILMERGE' --linkopt='/OPT:SAFEICF'
+
+# This hides inline symbols in classes that are marked to be exported, similar to
+# -fvisibility-inlines-hidden on Unix systems (https://blog.llvm.org/2018/11/30-faster-windows-builds-with-clang-cl_14.html)
+# Currently this only reduces object sizes slightly, larger gains are possible if we compile V8 as
+# a shared library.
+build:windows --copt='/Zc:dllexportInlines-'
+
+# Release configuration.
+build:release -c opt
+build:release --@rules_rust//:extra_rustc_flags=-C,panic=abort,-C,codegen-units=1,-C,embed-bitcode=n
+
+# enable -O3 for the release configuration. Based on benchmarking there is little difference in
+# performance, but -O3 should generally be expected to be faster for at least parts of the workerd API.
+build:release_unix --copt='-O3'
+build:release_unix --config=release
+build:release_linux --config=release_unix
+
+build:release_macos --config=release_unix
+# Disable generating LC_DATA_IN_CODE and LC_FUNCTION_STARTS binary sections, two rarely used types
+# of macOS debug info. These sections are largely undocumented, but are used by LLDB to improve
+# debugging on binaries that are otherwise stripped. There should be no need to include this
+# data in releases.
+build:release_macos --linkopt="-Wl,-no_data_in_code_info"
+build:release_macos --linkopt="-Wl,-no_function_starts"
+
+build:release_windows --config=release
+# Windows uses /O2 as its preferred optimization setting and enabled by bazel in the opt
+# configuration, but for clang-cl this is equivalent to only -O2 and a few other things. -O3 is
+# not exposed directly in the clang-cl driver, but we can access regular clang
+# flags using the /clang prefix anyway. https://clang.llvm.org/docs/UsersManual.html#the-clang-option
+build:release_windows --copt="/clang:-O3"
+# clang-cl does not enable strict aliasing by default to match MSVC's approach, unlike clang on
+# Unix which turns it on for opt builds.
+build:release_windows --copt="-fstrict-aliasing"
 
 build:windows --cxxopt='/std:c++20' --host_cxxopt='/std:c++20'
-build:windows --cxxopt='/await' --host_cxxopt='/await'
 build:windows --copt='/D_CRT_USE_BUILTIN_OFFSETOF' --host_copt='/D_CRT_USE_BUILTIN_OFFSETOF'
-build:windows --copt='/DWINDOWS_LEAN_AND_MEAN' --host_copt='/DWINDOWS_LEAN_AND_MEAN'
+build:windows --copt='/DWIN32_LEAN_AND_MEAN' --host_copt='/DWIN32_LEAN_AND_MEAN'
 # The `/std:c++20` argument is unused during boringssl compilation and we don't
 # want a warning when compiling each file.
 build:windows --copt='-Wno-unused-command-line-argument' --host_copt='-Wno-unused-command-line-argument'

.github/workflows/release.yml

@@ -60,10 +60,13 @@ jobs:
         include:
           - os-name: linux
             os: ubuntu-20.04
+            bazel-config: release_linux
           - os-name: macOS
             os: macos-13
+            bazel-config: release_macos
           - os-name: windows
             os: windows-2022
+            bazel-config: release_windows
     runs-on: ${{ matrix.os }}
     name: build (${{ matrix.os-name }})
     steps:
@@ -106,7 +109,7 @@ jobs:
           fi
       - name: Bazel build
         run: |
-          bazelisk build --disk_cache=~/bazel-disk-cache  --remote_cache=https://bazel:${{ secrets.BAZEL_CACHE_KEY }}@bazel-remote-cache.devprod.cloudflare.dev -c opt //src/workerd/server:workerd
+          bazelisk build --disk_cache=~/bazel-disk-cache --remote_cache=https://bazel:${{ secrets.BAZEL_CACHE_KEY }}@bazel-remote-cache.devprod.cloudflare.dev --config=${{ matrix.bazel-config }} //src/workerd/server:workerd
       - name: Strip debug symbols
         if: runner.os != 'Windows'
         run: |

.github/workflows/test.yml

@@ -90,7 +90,7 @@ jobs:
         # Not enabled because symbolication does not work on workerd macOS builds yet and running
         # llvm-symbolizer in the currently broken state causes some tests to time out on the
         # runner.
-        # Use latest available Xcode version – runner still defaults to 14.3.1.
+        # Use latest available Xcode version – runner still defaults to 15.0.1.
         run: |
           sudo xcode-select -s "/Applications/Xcode_15.1.app"
           # export LLVM_SYMBOLIZER=$(brew --prefix llvm@15)/bin/llvm-symbolizer
@@ -115,12 +115,17 @@ jobs:
         # don't need to run within a debugger, so information needed for symbolication is
         # sufficient. The host configuration compiles in opt mode/without debug info by default, so
         # there's no need to set host_copt here.
+        # LLVM produces a bit more debug info on macOS by default to facilitate debugging with
+        # LLDB. This is not needed for this use case, so disable it using -fno-standalone-debug –
+        # this is already the default for Linux/Windows.
         if: contains(matrix.config.suffix, 'debug') || contains(matrix.config.suffix, 'asan')
         shell: bash
         run: |
           cat <<EOF >> .bazelrc
           build:limit-storage --build_tag_filters=-off-by-default,-benchmark
           build:limit-storage --copt="-g1"
+          build:limit-storage --copt="-fno-standalone-debug"
+          build:limit-storage --@rules_rust//:extra_rustc_flags=-C,panic=unwind,-C,debug-assertions=y,-C,debuginfo=limited,-C,embed-bitcode=n
           build:asan --config=limit-storage
           build:debug --config=limit-storage
           EOF

Dockerfile.release

@@ -19,7 +19,7 @@ COPY .bazel-cache /bazel-disk-cache
 RUN npm install -g pnpm@latest-7
 RUN pnpm install
 
-RUN pnpm exec bazelisk build --disk_cache=/bazel-disk-cache -c opt //src/workerd/server:workerd
+RUN pnpm exec bazelisk build --disk_cache=/bazel-disk-cache --config=release_linux //src/workerd/server:workerd
 
 FROM scratch as artifact
 COPY --from=builder /workerd/bazel-bin/src/workerd/server/workerd /workerd-linux-arm64

build-releases.sh

@@ -15,7 +15,7 @@ TAG_NAME=$(curl -sL https://api.github.com/repos/cloudflare/workerd/releases/lat
 git checkout $TAG_NAME
 
 # Build macOS binary
-pnpm exec bazelisk build --disk_cache=./.bazel-cache -c opt //src/workerd/server:workerd
+pnpm exec bazelisk build --disk_cache=./.bazel-cache --config=release_macos //src/workerd/server:workerd
 
 cp bazel-bin/src/workerd/server/workerd ./workerd-darwin-arm64
 

build/wd_cc_benchmark.bzl

@@ -20,7 +20,7 @@ def wd_cc_benchmark(
           "//conditions:default": 1,
         }),
         linkopts = linkopts + select({
-          "@//:use_dead_strip": ["-Wl,-dead_strip"],
+          "@//:use_dead_strip": ["-Wl,-dead_strip", "-Wl,-no_exported_symbols"],
           "//conditions:default": [""],
         }),
         visibility = visibility,