Python: Fix bad join order in TPythonTuple

TL;DR: Something introduced the following bad join order:
```
(227s) Tuple counts for dom#TObject::TPythonTuple#ff/2@i2#8f58670w after 3m46s:
25000      ~0%     {2} r1 = SCAN PointsToContext::PointsToContext::appliesToScope_dispred#ff#prev_delta OUTPUT In.1, In.0 'context'
24000      ~1%     {2} r2 = JOIN r1 WITH @py_scope#f ON FIRST 1 OUTPUT Lhs.1 'context', Lhs.0
1076876712 ~6%     {3} r3 = JOIN r2 WITH Flow::TupleNode#class#f CARTESIAN PRODUCT OUTPUT Rhs.0, Lhs.0 'context', Lhs.1
870129666  ~0%     {3} r4 = JOIN r3 WITH Flow::ControlFlowNode::isLoad_dispred#f ON FIRST 1 OUTPUT Lhs.1 'context', Lhs.2, Lhs.0 'origin'
870129000  ~0%     {3} r5 = r4 AND NOT dom#TObject::TPythonTuple#ff#prev(Lhs.2 'origin', Lhs.0 'context')
870129000  ~1%     {3} r6 = SCAN r5 OUTPUT In.2 'origin', In.1, In.0 'context'
9000       ~0%     {2} r7 = JOIN r6 WITH Flow::ControlFlowNode::getScope_dispred#ff ON FIRST 2 OUTPUT Lhs.0 'origin', Lhs.2 'context'
                    return r7
```
(...the above being the tuple counts _at the point when I cancelled the
query_!)

Rewriting the code to force a join between `TupleNode#class` and
`getScope` results in the following join orders:

```
(0s) Tuple counts for TObject::scope_loads_tuplenode#ff/2@b3cf0bo5 after 13ms:
37369 ~3%     {1} r1 = JOIN Flow::TupleNode#class#f WITH Flow::ControlFlowNode::isLoad_dispred#f ON FIRST 1 OUTPUT Lhs.0 'origin'
37369 ~3%     {2} r2 = JOIN r1 WITH Flow::ControlFlowNode::getScope_dispred#ff ON FIRST 1 OUTPUT Rhs.1 's', Lhs.0 'origin'
            return r2
```
and
```
(78s) Tuple counts for dom#TObject::TPythonTuple#ff/2@i53#121c440w after 6ms:
34736 ~3%     {2} r1 = SCAN PointsToContext::PointsToContext::appliesToScope_dispred#ff#prev_delta OUTPUT In.1, In.0 'context'
7370  ~5%     {2} r2 = JOIN r1 WITH TObject::scope_loads_tuplenode#ff ON FIRST 1 OUTPUT Lhs.1 'context', Rhs.1 'origin'
7370  ~5%     {2} r3 = r2 AND NOT dom#TObject::TPythonTuple#ff#prev(Lhs.1 'origin', Lhs.0 'context')
7370  ~1%     {2} r4 = SCAN r3 OUTPUT In.1 'origin', In.0 'context'
            return r4
```
the latter being the largest iteration of `dom#TPythonTuple` throughout
the log.

No other major performance issues were observed.
This commit is contained in:
Taus
2022-01-31 16:59:50 +00:00
committed by GitHub
parent 56ac99039f
commit 4a29095e3b

View File

@@ -150,8 +150,10 @@ newtype TObject =
TBuiltinTuple(Builtin bltn) { bltn.getClass() = Builtin::special("tuple") } or
/** Represents a tuple in the Python source */
TPythonTuple(TupleNode origin, PointsToContext context) {
origin.isLoad() and
context.appliesTo(origin)
exists(Scope s |
context.appliesToScope(s) and
scope_loads_tuplenode(s, origin)
)
} or
/** Varargs tuple */
TVarargsTuple(CallNode call, PointsToContext context, int offset, int length) {
@@ -201,6 +203,13 @@ newtype TObject =
Expressions::subscriptPartsPointsTo(_, _, generic, index)
}
/** Join-order helper for TPythonTuple */
pragma[nomagic]
private predicate scope_loads_tuplenode(Scope s, TupleNode origin) {
origin.isLoad() and
origin.getScope() = s
}
/** Holds if the object `t` is a type. */
predicate isType(ObjectInternal t) {
t.isClass() = true