rbyd-rr: Limping along, fixed test_files_many, all tests are passing now
The issue, found in test_files_many:h1g4j10l18, occurs when a SUBWIDE
tag follows a compaction.
When this happens, it's possible for our stitched diverging alt to be
followed/flipped when it shouldn't be. This is because the new
lower_rid/upper_rid window can make the stitched alt ambiguous.
I don't think this is strictly an issue with compaction, as much as
compaction is giving us a tree structure that's not reachable through
only appendattrs.
Here are the three culprit trunks:
altrle 0x300 w8 0x2c8
altble 0x203 w6 0x2d4 <- diverge
null
altrle 0x300 w8 0x2c8
altbgt 0x203 w0 0x2e8 <- diverge
altble 0x300 w4 0x2b4
altbn w0 0x0
altble 0x300 w1 0x228
altbn w0 0x0
null
altrle 0x300 w8 0x2c8
altbgt 0x300 w0 0x2e8 <- stitch
altbn w0 0x0
altbn w0 0x0
altbn w0 0x0
altbgt 0x201 w0 0x164
reg w1
And here is a simplified view, after compaction, before we do a subwide
append/replace:
.-> reg w1 a
.-------b-> data
| .-> reg w1 b
| .---b-> data
r-b-r-b-b-> orphan w1 <- removed as a part of our subwide op
^
diverging
'-+-'
weight=3
altrle data w1
altble orphan w1
First, as a part of our subwide append, we're going to write out the
lower trunk. We diverge on the first altble since the entire orphan is
inside our subwide range.
It may seem a bit strange to diverge on a null tag, but this isn't
actually an issue, we're allowed a single null tag to terminate our
tree:
.-> reg w1 a
.-------b-> data
| .-> reg w1 b
r-b-----b-> data
^
diverging
'-+-'
weight=3
altrle data w1
altbgt data w0
Nothing wrong so far. The weight of our leaves (2) don't match our
tree's weight (3), but this is normal for the lower trunk. We fix this
when we stitch the diverging alt on the upper trunk.
Speaking of the upper trunk, let's start writing it out, but pause at
the stitching alt:
.-> reg w1 a
.-------b-> data
| .-> reg w1 b
| .-----b-> data
r-b-?
^
stitching
'-+-'
weight=3
altrle data w1
altble data w1
Note we've flipped the altbgt data into an altble data, since we're
going down the other diverged path now.
But before we continue, as a part of stitching, we need to adjust our
tree weight to account for the weight of the orphan we deleted as a part
of our range operation:
.-> reg w1 a
.-------b-> data
| .-> reg w1 b
| .-----b-> data
r-b-?
^
stitching
'-+-'
weight=2
altrle data w1
altble data w1
Uh oh. Weight is 2 and both our alts add up to 2? All of a sudden it
looks like we should follow the stitched alt.
Our follow/flip logic kicks in, and disaster!
.-> reg w1 a
.-------b-> data
r-b-----b-> reg w1 c <- added as a part of our subwide op
'-> data <- somehow data survives
'-+-' but where did b go?
weight=2
altrle data w1
altbgt data w0
We go down the wrong path, and because our state machine thinks we've
diverged, we prune all le alts, destroying our tree.
---
So what's is going wrong?
The problem is that when we update our window, the stitched diverging
alt can become ambiguous.
Which sort of makes sense. The reason we update our window is so we can
continue down the tree veiwing it as it was _before_ the range
operation. But the stitched alt belongs to the tree _after_ the range
operation.
The solution here is to just make sure we never follow the stitched alt.
This is a bit annoying, as it makes the stitched alt a rather special
case, but as far as I can tell it's necessary to avoid ambiguity.
This commit is contained in:
Christopher Haster
@@ -2980,7 +2980,7 @@ again:;
|
||||
p_alts[0],
|
||||
p_weights[0]);
|
||||
if (d_upper) {
|
||||
//alt &= ~LFSR_TAG_R;
|
||||
alt &= ~LFSR_TAG_R;
|
||||
diverged = true;
|
||||
|
||||
// alt = LFSR_TAG_ALT(
|
||||
@@ -2992,11 +2992,20 @@ again:;
|
||||
// } else {
|
||||
// weight = (upper_rid - lower_rid) - weight;
|
||||
// }
|
||||
|
||||
if (lfsr_tag_follow2(
|
||||
alt, weight,
|
||||
p_alts[0], p_weights[0],
|
||||
lower_rid, upper_rid,
|
||||
a_rid, a_tag)) {
|
||||
lfsr_tag_flip2(
|
||||
&alt, &weight,
|
||||
p_alts[0], p_weights[0],
|
||||
lower_rid, upper_rid);
|
||||
lfs_swap32(&jump, &branch_);
|
||||
}
|
||||
|
||||
if (lfsr_tag_isle(alt)) {
|
||||
alt = LFSR_TAG_ALT(
|
||||
LFSR_TAG_LE,
|
||||
alt & LFSR_TAG_R,
|
||||
d_tag);
|
||||
printf("%04x->%04x: dle 0x%x %d w%d (%d %d)\n",
|
||||
branch,
|
||||
rbyd->eoff,
|
||||
@@ -3005,43 +3014,55 @@ again:;
|
||||
weight,
|
||||
lower_rid,
|
||||
upper_rid);
|
||||
lower_rid += weight;
|
||||
weight = d_rid - lower_rid + weight;
|
||||
if (lfsr_tag_isred(p_alts[0])
|
||||
&& lfsr_tag_isle(p_alts[0])) {
|
||||
weight -= p_weights[0];
|
||||
}
|
||||
lower_rid -= weight;
|
||||
// jump = d_branch;
|
||||
} else {
|
||||
lfsr_tag_flip2(
|
||||
&alt, &weight,
|
||||
p_alts[0], p_weights[0],
|
||||
lower_rid, upper_rid);
|
||||
alt = LFSR_TAG_ALT(
|
||||
LFSR_TAG_LE,
|
||||
alt & LFSR_TAG_R,
|
||||
d_tag);
|
||||
printf("%04x->%04x: dgt 0x%x %d w%d (%d %d)\n",
|
||||
branch,
|
||||
rbyd->eoff,
|
||||
d_tag,
|
||||
d_rid,
|
||||
weight,
|
||||
lower_rid,
|
||||
upper_rid);
|
||||
lower_rid += weight;
|
||||
weight = d_rid - lower_rid + weight;
|
||||
lfsr_rid_t weight_ = d_rid - lower_rid;
|
||||
if (lfsr_tag_isred(p_alts[0])
|
||||
&& lfsr_tag_isle(p_alts[0])) {
|
||||
weight -= p_weights[0];
|
||||
weight_ -= p_weights[0];
|
||||
}
|
||||
lower_rid -= weight;
|
||||
lfsr_tag_flip2(
|
||||
&alt, &weight,
|
||||
p_alts[0], p_weights[0],
|
||||
lower_rid, upper_rid);
|
||||
// branch_ = d_branch;
|
||||
lower_rid += weight - weight_;
|
||||
weight = weight_;
|
||||
// lower_rid += weight;
|
||||
// weight = d_rid - lower_rid + weight;
|
||||
// if (lfsr_tag_isred(p_alts[0])
|
||||
// && lfsr_tag_isle(p_alts[0])) {
|
||||
// weight -= p_weights[0];
|
||||
// }
|
||||
// lower_rid -= weight;
|
||||
// jump = d_branch;
|
||||
} else {
|
||||
LFS_UNREACHABLE();
|
||||
// printf("%04x->%04x: dgt 0x%x %d w%d (%d %d)\n",
|
||||
// branch,
|
||||
// rbyd->eoff,
|
||||
// d_tag,
|
||||
// d_rid,
|
||||
// weight,
|
||||
// lower_rid,
|
||||
// upper_rid);
|
||||
// lfsr_tag_flip2(
|
||||
// &alt, &weight,
|
||||
// p_alts[0], p_weights[0],
|
||||
// lower_rid, upper_rid);
|
||||
// alt = LFSR_TAG_ALT(
|
||||
// LFSR_TAG_LE,
|
||||
// alt & LFSR_TAG_R,
|
||||
// d_tag);
|
||||
// lower_rid += weight;
|
||||
// weight = d_rid - lower_rid + weight;
|
||||
// if (lfsr_tag_isred(p_alts[0])
|
||||
// && lfsr_tag_isle(p_alts[0])) {
|
||||
// weight -= p_weights[0];
|
||||
// }
|
||||
// lower_rid -= weight;
|
||||
// lfsr_tag_flip2(
|
||||
// &alt, &weight,
|
||||
// p_alts[0], p_weights[0],
|
||||
// lower_rid, upper_rid);
|
||||
//// branch_ = d_branch;
|
||||
}
|
||||
|
||||
printf("%04x->%04x: dtag 0x%x w%d (%d %d)\n",
|
||||
@@ -3051,8 +3072,11 @@ again:;
|
||||
weight,
|
||||
lower_rid,
|
||||
upper_rid);
|
||||
|
||||
goto push;
|
||||
|
||||
} else {
|
||||
//alt &= ~LFSR_TAG_R;
|
||||
alt &= ~LFSR_TAG_R;
|
||||
//d_will_diverge = true;
|
||||
diverged = true;
|
||||
}
|
||||
@@ -3526,6 +3550,7 @@ again:;
|
||||
}
|
||||
}
|
||||
|
||||
push:;
|
||||
// trim alts from our current bounds
|
||||
lfsr_tag_trim2(
|
||||
alt, weight,
|
||||
|
||||
Reference in New Issue
Block a user