https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80131
--- Comment #3 from joseph at codesourcery dot com <joseph at codesourcery dot com> --- On Tue, 21 Mar 2017, segher at gcc dot gnu.org wrote: > If we have d = a << (b - c); and a << b does not truncate in the > original mode, write it as d := (a << b) >> c; instead (and then > see if it is actually cheaper, as combine always does). If b is not 31 (for shifts of 32-bit values), a << (b - c) may be valid for some negative values of c, where (a << b) >> c isn't valid for negative c (as C code, so doing such an optimization in such a case would depend on the semantics of the machine instructions available and associated RTL).
