mirror of
https://github.com/hanzoai/datastore.git
synced 2026-09-26 13:55:07 +00:00
Konstantin BogdanovandClaude Fable 5
44f64dd3f0
The x86_64 mem* dispatchers in contrib/libllvmlibc-cmake rely on `flatten`
to inline the whole LLVM-libc helper tree into the `target("arch=x86-64-v4")`
clones. At -O0 (Debug build with -DDEBUG_O_LEVEL=0) clang lowers `flatten`
as a single level of inlining, so the v4-attributed `memmove_overlap_avx512`
calls out-of-line baseline-target (x86-64-v3) helpers that pass 64-byte
vectors by value. That is an ABI mismatch - the v4 caller reads the return
value from zmm0 while the v3 callee returns it in ymm0:ymm1 - which silently
corrupts every overlapping `memmove` of more than 128 bytes on AVX-512
machines. Clang rejects this pattern with a hard error when the mismatched
call is visible in the source, but the inliner recreates it silently.
The corruption made llvm-tblgen crash (SIGSEGV in -gen-x86-fold-tables,
SmallVector `idx < size()` assertion in -gen-asm-matcher) during -O0 builds
on AVX-512 hardware. CI debug builds did not catch it: they run with the
default -Og, and CI machines have no AVX-512.
Fix it in the source so the code is correct at any optimization level:
upgrade LIBC_INLINE to `always_inline` for this TU (honored even at -O0),
and redeclare the generic `load`/`store`/`splat` - which lack LIBC_INLINE
upstream yet are exactly the functions passing vectors by value - with
`always_inline` as well. The helper tree then always collapses into the
exported functions and no cross-target call boundary survives.
Caused by: https://github.com/ClickHouse/ClickHouse/pull/107566
Co-Authored-By: Claude Fable 5 <[email protected]>