On Wed, Oct 20, 2021 at 03:22:10PM +0200, Richard Biener via Gcc-patches wrote:
> This maps -ftrapv to -fsanitize=signed-integer-overflow
> -fsanitize-undefined-trap-on-error, effectively removing
> flag_trapv (or rather making it always false).
>
> This has implications on language support - while -ftrapv
> was formerly universally available the mapping restricts it
> to the C family of frontends.
>
> It also raises questions on mixing -ftrapv with -fsanitize
> flags, specifically with other recovery options for the
> undefined sanitizer since -fsanitize-undefined-trap-on-error
> cannot be restricted to the signed-integer-overflow part at
> the moment. To more closely map behavior we could add
> -fsanitize=trapv where with a single option we could also
> simply alias -ftrapv to that.
I think we shouldn't do it this way.
There is no reason not to support it in all FEs, not just C family,
the instrumentation is done during in this case in the ubsan pass anyway.
And it also should cope well with different sanitizers, while
-ftrapv vs. -fsanitize=signed-integer-overflow probably needs to be
either/or, so one of those should take precedence over the other,
e.g. -fsanitize=shift -fsanitize-recover=shift -ftrapv should result
in recovering from shift UBs, not trap on them.
My preference would be new set of ifns for -ftrapv, similar to
.UBSAN_CHECK_{ADD,SUB,MUL}, say .TRAPV_CHECK_{ADD,SUB,MUL,DIV},
that uses moreless the same internal-fn.c expansion as .UBSAN_CHECK_*,
but doesn't call ubsan_build_overflow_builtin and rely on
flag_sanitize_undefined_trap_on_error, instead either emits
the trap call directly, or also uses libcalls if optab isn't
available and libcall is (or for -Os cases if libcall is smaller).
Because some of the -fsanitize=signed-integer-overflow emitted code
for multiplication at least on some architectures is very large...
Jakub
