On Tue, Jan 8, 2019 at 5:17 PM H.J. Lu <hjl.to...@gmail.com> wrote: > > On Tue, Jan 8, 2019 at 6:54 AM Uros Bizjak <ubiz...@gmail.com> wrote: > > > > On Tue, Jan 8, 2019 at 3:39 PM H.J. Lu <hjl.to...@gmail.com> wrote: > > > > > > On Mon, Jan 7, 2019 at 11:12 PM Uros Bizjak <ubiz...@gmail.com> wrote: > > > > > > > > On Mon, Jan 7, 2019 at 6:40 PM H.J. Lu <hongjiu...@intel.com> wrote: > > > > > > > > > > There is no need to generate vzeroupper if caller uses upper bits of > > > > > AVX/AVX512 registers, We track caller's avx_u128_state and avoid > > > > > vzeroupper when caller's avx_u128_state is AVX_U128_DIRTY. > > > > > > > > > > Tested on i686 and x86-64 with and without --with-arch=native. > > > > > > > > > > OK for trunk? > > > > > > > > In principle OK, but I think we don't have to cache the result of > > > > ix86_avx_u128_mode_entry. Simply call the function from > > > > ix86_avx_u128_mode_exit; it is a simple function, so I guess we can > > > > afford to re-call it one more time per function. > > > > > > Do we really need ix86_avx_u128_mode_entry? We can just > > > set entry state to AVX_U128_CLEAN and set exit state to > > > AVX_U128_DIRTY if caller returns AVX/AVX512 register or passes > > > AVX/AVX512 registers to callee. > > > > > > Does this patch look OK? > > > > No, the compiler is then free to move optimal insertion point at the > > beginning of the function. > > > > Here is the updated patch. OK for trunk?
OK with the comment fix. Thanks, Uros. - return AVX_U128_CLEAN; + /* Entry mode is set to AVX_U128_DIRTY if there are 256bit or 512bit s/Entry/Exit/ + modes used in function arguments. */ ... , otherwise return AVX_U128_CLEAN. + return ix86_avx_u128_mode_entry (); }
