On Apr 12, 2016 10:16 AM, "Connor Abbott" <<a href="mailto:cwabbott0@gmail.com">cwabbott0@gmail.com</a>> wrote: > > I'm not sure I'm so comfortable with this. For one, this is the sort > of thing the backend can (and should) do. Why should be be adding > stuff at the NIR level to help the backend's copy propagation pass? > And perhaps more importantly, why should we be adding stuff at the NIR > level to workaround the lack of spilling support for doubles? I know > it's more difficult to implement than normal spilling, but it just > feels dirty to workaround it because some tests happen to use too many > registers. > > Jason, what do you think? Agreed. The solution to spilling problems is to fix spilling. At first I thought this *might* be a good idea but thinking a bit more, I have to agree with Connor. The only thing this does is short-cut part of the use-def graph in such a way as to make RA happier. > On Tue, Apr 12, 2016 at 4:05 AM, Samuel Iglesias Gonsálvez > <<a href="mailto:siglesias@igalia.com">siglesias@igalia.com</a>> wrote: > > From: Iago Toral Quiroga <<a href="mailto:itoral@igalia.com">itoral@igalia.com</a>> > > > > This is useful when we only need to modify the high 32-bit chunk of a double. > > This is a common case, because this is the part that encodes the exponent > > which we manipulate in some double lowering passes. Although we can accomplish > > the same by using pack_double_2x32, this new opcode is better for register > > pressure, since we don't have to unpack both parts of the double and keep > > the low 32-bits around until we can recombine the new exponent. > > > > We will use this opcode in the set_exponent() function of the double lowering > > pass and with that we will fix spilling issues in some dmat4 divide > > piglit tests on Intel. > > --- > > src/compiler/nir/nir_opcodes.py | 14 ++++++++++++++ > > 1 file changed, 14 insertions(+) > > > > diff --git a/src/compiler/nir/nir_opcodes.py b/src/compiler/nir/nir_opcodes.py > > index 9f62e08..f92dd8d 100644 > > --- a/src/compiler/nir/nir_opcodes.py > > +++ b/src/compiler/nir/nir_opcodes.py > > @@ -276,6 +276,20 @@ di.i2 = src0.y; > > dst.x = di.u64; > > """) > > > > +opcode("pack_double_2x32_split_y", 0, tuint64, [0, 0], [tuint64, tuint32], "", """ > > +union { > > + uint64_t u64; > > + struct { > > + uint32_t i1; > > + uint32_t i2; > > + }; > > +} di; > > + > > +di.u64 = src0; > > +di.i1 = src1; > > +dst = di.u64; > > +""") > > + > > unop_horiz("unpack_double_2x32", 2, tuint32, 1, tuint64, """ > > union { > > uint64_t u64; > > -- > > 2.5.0 > > > > _______________________________________________ > > mesa-dev mailing list > > <a href="mailto:mesa-dev@lists.freedesktop.org">mesa-dev@lists.freedesktop.org</a> > > <a href="https://lists.freedesktop.org/mailman/listinfo/mesa-dev">https://lists.freedesktop.org/mailman/listinfo/mesa-dev</a>