https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82004
--- Comment #14 from Jakub Jelinek <jakub at gcc dot gnu.org> --- If glibc implements a faster exp10/exp10f, we could do what I've done for PR84309 and expand pow (10, x) as exp10 (x) and pow (C, x) where C is pow (10, N) for integral N as exp10 (log10 (C) * x). Even if it doesn't, there is some chance that perhaps 628.pop2_s propagates a constant into that x of pow (10, x), so in theory even just deferring the folding to after vectorization like we do for C which is pow (2, N) for integral N might help. Looking at http://oceans11.lanl.gov/trac/POP/browser/branches/ice_shelf_partial_cells/source/sw_absorption.F90?rev=373&order=name&desc=True I think that actually is the case here. >From that source (haven't looked at SPEC2k17) it seems it is essentially: ! PR middle-end/82004 ! { dg-do run } ! { dg-options "-Ofast" } integer, parameter :: r8 = selected_real_kind(13), i4 = kind(1) integer (i4), parameter :: a = 400, b = 2 real (r8), parameter, dimension(b) :: c = (/ .001_r8, 10.00_r8 /) real (r8) :: d, e, f, g, h, j d = c(1) e = c(b) f = (log10(e)-log10(d))/real(a) g = log10(d) - f h = 10**(g) j = 10 j = j**(g) if (h.ne.j) stop 1 end (with the j and j = 10; j = j**(g); if (h.ne.j) stop 1 added by me to make it fail when miscompiled).
