On Thu, Aug 10, 2017 at 02:17:40PM +0930, Alan Modra wrote:
> > > Ick, looks like papering over the real problem to me, and will no
> > > doubt cause -Os size regressions.
> > I think it is very directly solving the real problem. It isn't likely
> > to cause size regressions (look how long it took for this PR to show
> > up, so this cannot happen often).
> > This only happens if r30 (the PIC reg) is used but r31 isn't; which means
> > that a) there are no other registers to save, or b) the function is marked
> > as needing a hard frame pointer but eventually doesn't need one.
> > (RA picks the registers r31, r30, ... in sequence).
> > In the case in the PR, this patch replaces one insn by one (cheaper)
> > insn.
> And in other cases your patch will prevent stmw when it should be
> used. Testcase attached. It shows the wrong use of lmw too.
I dunno... If you change r25 to r14 in that testcase it will use stmw 14
with my patch reverted. Not very reasonable imho (but then again, people
using register asm like this get what they deserve anyway).