@@ -124,50 +124,6 @@ failure precisely because a killed run writes no profile. The training runs *are
124124representative workload that makes PGO help instead of hurt; the 256 kHz gate is a hard
125125release gate on the shipped PGO binary. See [[ ss-hotpath]] .
126126
127- ### Post-link layout (ENABLE_POST_LINK_LAYOUT, spec 0090)
128-
129- OFF by default; only the release CI pipelines pass it, at ** both** PGO configures (the
130- stage-identity rule governs the ` -D ` values, so the option never varies between GENERATE and
131- USE). One mechanism per platform, all downstream of PGO:
132-
133- - ** Linux ELF (x86-64/aarch64; armv7l excluded -- BOLT has no 32-bit ARM port):** the module
134- only makes the link BOLT-ready: ` -Wl,--emit-relocs ` , plus GCC
135- ` -fno-reorder-blocks-and-partition ` (pre-split functions are ones BOLT must skip; its own
136- ` -split-functions ` replaces that optimization). The rewrite itself is CI's
137- ` .github/actions/bolt-linux ` : drop orphan ` .cold ` symbols (the prebuilt gRPC carries
138- ICF-orphaned fragments BOLT hard-fails on) -> instrument -> re-run both training loads
139- (seat already active; the trainer's exit code is advisory and a crash prints a gdb
140- backtrace -- the profile is dumped every 5 s during the run, so a teardown crash loses
141- nothing) -> ` merge-fdata ` -> two `llvm-bolt -reorder-blocks=ext-tsp
142- -reorder-functions=cdsort -split-functions -split-all-cold -dyno-stats` rewrites, with
143- ` -skip-funcs ` covering the LuaJIT VM asm and the gRPC/upb/protobuf-C families. The BOLT
144- pin must stay >= 22: Hardening.cmake's ` -mbranch-protection=standard ` (pac-ret, Linux
145- aarch64) needs BOLT's OpNegateRAState CFI support (llvm/llvm-project #120064
146- drops the RA-state markers and the first unwind through a rewritten frame (a routine
147- LuaJIT trace abort suffices) segfaults in libgcc on a still-signed return address. BOLT is
148- the only post-link writer: objcopy/strip corrupt a rewritten ELF (llvm/llvm-project #89336
149- so the shipping rewrite uses ` -remove-symtab ` (the spec-0084 hardening strip) and BOLT's
150- built-in debug/` --emit-relocs ` -payload stripping, while a layout-identical
151- ` -update-debug-sections ` twin exists only so the ` .debug ` sidecar can be cut from it
152- (` objcopy --only-keep-debug ` reads the twin, never the shipped binary). The POST_BUILD
153- split/strip in ` app/CMakeLists.txt ` is disabled entirely when the option is ON. Fail-hard:
154- there is no fallback to the un-rewritten binary, and every gate/package/signature below
155- the stage consumes the rewritten one.
156- - ** macOS (AppleClang, PGO-USE only, arm64 slice only):** no Mach-O rewriter and no linker
157- swap; the module derives a hot-function order file from ` merged.profdata `
158- (` llvm-profdata show --topn ` , ` _ ` -prefixed for Mach-O) into ` hot-order.txt ` and links with
159- ` -Wl,-order_file,... ` . An empty order file is a configure-time FATAL_ERROR. The x86_64 leg
160- builds without PGO and never sets the option.
161- - ** Windows (clang-cl):** no new mechanism -- lld-link already sorts functions by the PGO
162- call-graph profile. The module adds the native lld-link flag
163- ` /print-symbol-order:${CMAKE_BINARY_DIR}/symbol-order.txt ` and ci.yml asserts the file is
164- present and non-trivial after the optimized link; weakening that assertion is not an option
165- (the documented fallback is a generated ` /call-graph-ordering-file ` ).
166-
167- Layout/metadata only: the IEEE-math and unwind-table invariants are untouched by
168- construction, and the rewritten binary's unwind is proven per push by the ` script-unwind `
169- pre-root selftest suite (a Lua error raised through a C frame must land back in the host).
170-
171127### Flags deliberately NOT enabled
172128` -fcomplete-member-pointers ` (clang family). Rejects a pointer-to-member whose class is still
173129incomplete -- under the MSVC ABI the representation follows the inheritance model, so an
0 commit comments