-fno-semantic-interposition
can significantly improve performance of code in shared libraries but may change semantics in some cases.
By default GCC respects the ELF symbol interposition semantics. In short, any exported library function (i.e. any library function if compiled with default compiler flags) can be replaced at runtime via LD_PRELOAD
or simply by function with the same name in another shared library which happens to be loaded earlier by dynamic linker. This prevents compiler from doing a lot of useful analyses and optimizations (most notably inlining and cloning) because they may break interposition.
-fno-semantic-interposition
gives compiler permission to ignore potential interposition and optimize much more aggressively.
As I said, there are some caveats in using -fno-semantic-interposition
:
- it might change behavior of your program (when it was actually relying on interposition, sometimes without you realizing this)
- it's only relevant for shared libraries (not executables)
- it's much less useful if you already do proper optimization of your libraries (i.e. compile with
-fvisibility=hidden
and explicitly annotate all exported symbols with __attribute__((visibility("default")))
)
The first item prevents wide deployment of -fno-semantic-interposition
. E.g. to my knowledge no Linux distro uses it at wide scale (it would be a great project btw).
BTW note that Clang compiler has -fno-semantic-interposition
enabled by default, presumably for the sake of performance. They have an inverse -fsemantic-interposition
flag to enable GCC-compatible ELF interposition semantics.