Comment # 7
on bug 90887
from Ilia Mirkin
OK, so... as I suspected, the phi nodes are actually fine. I added this to the
nv50 printer:
for (Graph::EdgeIterator ei = bb->cfg.incident(); !ei.end(); ei.next())
INFO(" <- BB:%i (%s)\n",
BasicBlock::get(ei.getNode())->getId(),
ei.getEdge()->typeStr());
Which makes it much easier to look at things. So right before RA we have:
15: texlod 2D $r0 $s0 f32 { %r52 %r53 %r54 } %r45 %r47 %r49 (0)
BB:5 (7 instructions) - idom = BB:4, df = { }
<- BB:4 (tree)
-> BB:3 (forward)
-> BB:2 (tree)
16: join (0)
17: mov u32 %r61 0x3ffb4a23 (0)
18: mad f32 %r62 %r53 %r61 %r52 (0)
19: mov u32 %r64 0x3f600000 (0)
20: joinat BB:3 (0)
21: set u8 %c68 lt f32 %r62 %r64 (0)
22: eq %c68 bra BB:3 (0)
...
BB:9 (1 instructions) - idom = BB:2, df = { BB:3 }
<- BB:13 (forward)
<- BB:12 (forward)
<- BB:11 (forward)
<- BB:2 (forward)
-> BB:10 (tree)
32: texlod 2D $r0 $s0 f32 { %r76 %r77 %r78 } %r73 %r73 %r73 (0)
BB:10 (2 instructions) - idom = BB:9, df = { BB:3 }
<- BB:9 (tree)
-> BB:3 (forward)
33: join ind BB:0 (0)
34: bra BB:3 (0)
BB:3 (5 instructions) - idom = BB:5, df = { }
<- BB:10 (forward)
<- BB:5 (forward)
-> BB:1 (tree)
35: phi u32 %r88 %r78 %r52 (0)
36: phi u32 %r89 %r77 %r53 (0)
37: phi u32 %r90 %r76 %r54 (0)
38: join (0)
Which is perfectly correct. Although it *is* a bit odd that BB:3 doesn't have
any incoming TREE edges. Definitely a bit worrying. But let's hope that the
thing doesn't ACTUALLY care about that, and that TREE is merely an optimization
and it'd be just as happy with s/tree/forward everywhere.
And then after the various RA passes run:
20: texlod 2D $r0 $s0 f32 %r107t %r108t (0)
21: split b96 { %r52 %r53 %r54 } %r107t (0)
22: bra BB:5 (0)
BB:5 (7 instructions) - idom = BB:4, df = { }
<- BB:4 (tree)
-> BB:17 (tree)
-> BB:2 (tree)
23: join ind BB:0 (0)
24: mov u32 %r61 0x3ffb4a23 (0)
25: mad f32 %r62 %r53 %r61 %r52 (0)
26: mov u32 %r64 0x3f600000 (0)
27: joinat BB:3 (0)
28: set u8 %c68 lt f32 %r62 %r64 (0)
29: eq %c68 bra BB:17 (0)
...
BB:9 (7 instructions) - idom = BB:2, df = { BB:3 }
<- BB:20 (forward)
<- BB:19 (forward)
<- BB:18 (forward)
<- BB:13 (forward)
-> BB:10 (tree)
42: mov u32 %r112 %r73 (0)
43: mov u32 %r113 %r73 (0)
44: mov u32 %r114 %r73 (0)
45: merge b96 %r110t %r112 %r113 %r114 (0)
46: texlod 2D $r0 $s0 f32 %r109t %r110t (0)
47: split b96 { %r76 %r77 %r78 } %r109t (0)
48: bra BB:10 (0)
BB:10 (5 instructions) - idom = BB:9, df = { BB:3 }
<- BB:9 (tree)
-> BB:3 (forward)
49: join ind BB:0 (0)
50: mov u32 %r118 %r52 (0)
51: mov u32 %r119 %r53 (0)
52: mov u32 %r120 %r54 (0)
53: bra BB:3 (0)
BB:17 (4 instructions) - df = { }
<- BB:5 (tree)
-> BB:3 (forward)
54: mov u32 %r115 %r78 (0)
55: mov u32 %r116 %r77 (0)
56: mov u32 %r117 %r76 (0)
57: bra BB:3 (0)
BB:3 (6 instructions) - idom = BB:5, df = { }
<- BB:17 (forward)
<- BB:10 (forward)
-> BB:1 (tree)
58: phi u32 %r88 %r115 %r118 (0)
59: phi u32 %r89 %r116 %r119 (0)
60: phi u32 %r90 %r117 %r120 (0)
61: join (0)
So..... the phi nodes are fine :) %r115 does indeed come from BB:17. HOWEVER!
Note that the BB10/BB17 stuff is reversed! %r78 comes from BB:9, but the
incident edge into BB:17 is BB:5. Same for BB:10. So the issue is somewhere in
there.
Perhaps that's what you were saying all along and I was too dense to understand
it? It picks the wrong mov source because the incident edge order is different
now?
[As an aside, this whole quadop thing is *soooo* unnecessary. It's only needed
for non-uniform lod/bias values, and in this case, they're clearly uniform.
However this uniformity check is done pre-ssa, since doing the bb splitting
post-ssa would be a huge pain. Ugh!]