On Tue, Jun 28, 2016 at 12:49:12PM -0500, Segher Boessenkool wrote: > On Tue, Jun 28, 2016 at 01:27:45PM -0400, Michael Meissner wrote: > > <snip> > > Thanks for the explanation. > > > I think the thing to do is create yet another memory constraint, that is > > just > > an offsetable address, with the bottom 2 bits 0, and no PRE_MODIFY, etc. > > That sounds best yes. The current patch seemed fragile and a bit > confusing / surprising to me, but you now found an actual problem as > well. > > > > 20599 lines, can you minimize this a bit? If not, maybe we should just > > > do without testcase here. > > > > I doubt I could minimize it, since it is only this one source so far that > > has > > shown to be failure. I can delete it if you prefer. > > Huge testcases for simple problems does not really scale.
But it is not a simple problem. You need to have allocated a DImode to an Altivec register, which typically means you would otherwise be spilling registers because the GPRs and FPRs are used for other values. But I will delete the test case, since we do build Spec 2006 quite often. -- Michael Meissner, IBM IBM, M/S 2506R, 550 King Street, Littleton, MA 01460-6245, USA email: meiss...@linux.vnet.ibm.com, phone: +1 (978) 899-4797
