http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60418
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |rguenth at gcc dot gnu.org
--- Comment #13 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to H.J. Lu from comment #10)
> Sources have many FP loops contains codes like:
>
> rsq11 = dx11*dx11+dy11*dy11+dz11*dz11;
>
> When they are compiled with
>
> -O3 -funroll-loops -ffast-math -fwhole-program -flto=jobserver
> -fuse-linker-plugin
>
> LTO input IRs contain statements like
>
> powmult_241 = dy11_71 * dy11_71;
> powmult_240 = dz11_72 * dz11_72;
> _699 = powmult_240 + powmult_80;
> rsq11_77 = _699 + powmult_241;
>
> During the final LTO link, lto1 repeatedly removes loop a preheader
> in one pass and adds it back in the next pass. Each removal/add
> changes the statements to
>
> powmult_213 = dy11_71 * dy11_71;
> _75 = powmult_213 + powmult_80;
> powmult_244 = dz11_72 * dz11_72;
> rsq11_77 = _75 + powmult_244;
>
> Each such re-order may change the FP result slightly. They
> can accumulate to such a degree that the end result is
> outside of tolerance.
Huh, adding a pre-header should _never_ do sth like that. Can you produce
a small testcase that exhibits these kind of changes with adding/removing
a preheader?
