On Tue, Mar 12, 2024 at 10:46:42AM +0100, Jan Hubicka wrote:
> I am sorry for delaying this. I made the variant that simply compares
> value range of functions and prevents merging if they diverge and wanted
> to make some bigger statistics. This made me notice some performance
> problems on clang performance and libstdc++ RB-trees which disrailed me
> from the original PR. I will finish the statistics today.
With the posted patch, perhaps if we don't want to union jump_tables etc.,
all we could punt on is differences in the jump_table VRs rather than just
any SSA_NAME_RANGE_INFO differences.
But I agree that for GCC 15 we really want to merge rather than punt.
> For next stage1 we definitly want to move ahead with merging metadata
> (not only value ranges, but hopefully also TBAA). Currently the only
> thing we merge aggressively is the edge profile (and we have PR on that
> merging functions which differs only but likely/unlikely attribute).
> However for backportability and for this avoiding merging may be safer
> solution depending on the stats.A
>
> I was looking into ipa-vrp useability and there are several tests on
> Firefox talos where ipa-prop VRP makes difference. It boils down to
> propagating fact that parameter is alway snon-NULL. This is commonly
> tested in C++ casts.
Jakub
