On Fri, Feb 26, 2016 at 01:35:10PM -0800, Richard Henderson wrote: > On 02/26/2016 01:03 PM, Segher Boessenkool wrote: > > On Thu, Feb 25, 2016 at 09:08:32PM -0800, Richard Henderson wrote: > >> + /* Perform rematerialization if only all operands are registers and > >> + all operations are PLUS. */ > >> + for (i = 0; i < n_ops; i++) > >> + if (ops[i].neg || !REG_P (ops[i].op)) > >> + return NULL_RTX; > >> + goto gen_result; > >> + } > > > > If you check for fixed registers as well here, does that work for you? > > Maybe. It prevents canonicalization of reg+fp vs fp+reg, which could well > occur via arithmetic on locally allocated arrays.
Where are these canonicalization rules described? > I guess for the purposes of stage4 I'd be willing to do > > if (ops[i].neg > || !REG_P (ops[i].op) > || (REGNO (ops[i].op) < FIRST_PSEUDO_REGISTER > && fixed_regs[REGNO (ops[i].op)] > && !global_regs[REGNO (ops[i].op)] > && ops[i].op != frame_pointer_rtx > && ops[i].op != arg_pointer_rtx > && ops[i].op != stack_pointer_rtx)) > > It's pretty ugly though, and I wouldn't want to keep this forever. Yeah. > The rs6000 change really ought to be evaluated at some point. Given its > scope, > I see little difference to doing that now vs putting it off to gcc7. It is stage 4. This rs6000 change has almost 100% chance of introducing regressions. Segher