These patches fix various cases in the x86 backend and the runtime TB-jump-patching code where we cast an unaligned pointer to a uint32_t* and store into it. Unaligned accesses are OK on x86 hardware, of course, but this is still undefined behaviour in C, and the clang sanitizer complains.
Sent out as an RFC to get a feel for whether we want to go down some road like this. Personally I think it is worthwhile for two reasons: (1) it's easier to see sanitizer warnings that actually matter if they're not hidden in among a lot of warnings that don't. (2) it's increasingly clear that it's a bad idea to trust compiler engineers, who will happily throw real applications under the bus for a 0.05% improvement in Dhrystone scores; so if we can reasonably avoid undefined behaviour we should. It would probably be better to split the tcg_out functions into separate ones for "I know this pointer is aligned" and "may be unaligned", to avoid overhead on hosts which don't have cheap unaligned stores. Also I haven't tried to extend the tcg_patch* usage to the other backends. Peter Maydell (3): exec-all.h: Use stl_p to avoid undefined behaviour patching x86 jumps tcg: Avoid stores to unaligned addresses tcg: Avoid undefined behaviour patching code at unaligned addresses include/exec/exec-all.h | 2 +- tcg/i386/tcg-target.c | 12 ++++++------ tcg/tcg.c | 26 +++++++++++++++++++++++--- 3 files changed, 30 insertions(+), 10 deletions(-) -- 1.9.0
