On Mon, Apr 04, 2016 at 03:10:34PM +0200, Richard Biener wrote: > On Mon, 4 Apr 2016, Jakub Jelinek wrote: > > > On Mon, Apr 04, 2016 at 02:56:51PM +0200, Richard Biener wrote: > > > The testcase gcc.target/i386/pr37870.c will already ICE with that > > > patch, so no additional testcase. > > > > In theory you could validate_subreg first and use that code if validation > > went ok, otherwise go through memory. > > But I admit I don't have anything in particular in mind where it would > > trigger this code and the subreg would successfully validate. > > Not sure if it would help as that has > > /* ??? Similarly, e.g. with (subreg:DF (reg:TI)). Though > store_bit_field > is the culprit here, and not the backends. */ > else if (osize >= UNITS_PER_WORD && isize >= osize) > ; > > and thus we'd return true anyway for (subreg:XF (reg:TI) 0)
If XFmode subreg of TImode reg passes validation, where does it ICE then? Jakub