On Thu, Feb 06, 2014 at 08:23:00AM -0800, Richard Henderson wrote: > On 02/06/2014 08:02 AM, Jakub Jelinek wrote: > > On Thu, Feb 06, 2014 at 06:53:55AM -0800, Richard Henderson wrote: > >> On 02/04/2014 04:40 AM, Jakub Jelinek wrote: > >>> - tem = expand_shift (LSHIFT_EXPR, mode, lopart, hprec, NULL_RTX, 1); > >>> + tem = gen_rtx_SUBREG (mode, lopart, 0); > >>> + tem = expand_shift (LSHIFT_EXPR, mode, tem, hprec, NULL_RTX, 1); > >> > >> I would be happier with gen_lowpart rather than the explicit > >> gen_rtx_subreg. > > > > I need a paradoxical subreg, gen_lowpart ICEs in that case > > It does? Since when? I've certainly used it for paradoxicals in the past. > > Oh, I see, yes, it would ICE for a multi-word paradoxical subreg. But that > sort of thing is ... skirting the bounds of validity at best. > > Surely we should be able to optimize away a zero-extension in all cases?
All I know is that the generated code with the ZERO_EXTEND has been larger than with the paradoxical subreg. But if you prefer, I can surely emit a ZERO_EXTEND and open a PR for GCC 5.0 that we should investigate why we generate worse code then. Jakub