This change updates `FlowModel.promotionInfo`, the primary data
structure used by flow analysis to track program state, so that
instead of being a `Map<int, PromotionModel<Type>>`, it is represented
by a new data structure called a `FlowLink`, an immutable data
structure describing program state in a way that's particularly
optimized for flow analysis's usage patterns.
Like a map, a `FlowLink` data structure represents a collection of
key/value pairs (where the keys are integers), however instead of
storing the keys and values in a hashtable, each `FlowLink` object
contains a single key/value pair and a pointer to a previous
`FlowLink` object. The value associated with a given key can be looked
up by starting with the current `FlowLink` and walking backwards
through the linked list of `previous` pointers until a matching key is
found. (An empty map is represented by `null`). This makes it an
`O(1)` operation to update the promotion state associated with a
single promotion key (an operation that flow analysis performs
frequently), since all that is required is a single allocation.
If the `previous` pointers are regarded as parent pointers, all the
`FlowLink` objects produced by a given run of flow analysis form a
tree that mirrors the dominator tree of the code being analyzed.
To optimize reads of `FlowLink` data structures, there is a
`FlowLinkReader` class that keeps track of a lookup table reflecting
the implicit map represented by a given `FlowLink` object; this table
can be updated to reflect a different `FlowLink` object in `O(n)`
time, where `n` is the number of edges between the two `FlowLink`
objects in the tree. Since flow analysis is based on a depth-first
traversal of the syntax tree of the code being analyzed, it has a high
degree of tree locality in the `FlowLink` objects it needs to be able
to read, so these `O(n)` updates do not consume much CPU.
The `FlowLinkReader` class is also able to compute a difference
between the program states represented by two `FlowLink` objects, in
`O(n)` time, where `n` is the number of edges between the two
`FlowLink` objects in the tree. This is used by flow analysis to
compute the program state after a control flow join, so that it does
not need to spend any time examining promotion keys that are unchanged
since the corresponding control flow split.
For more information about the `FlowLink` data structure and how it
works, see the comments in `flow_link.dart`.
This change improves the performance of CFE compilation fairly
substantially:
instructions:u: -0.8167% +/- 0.0007% (-158214865.67 +/- 130252.25)
branches:u: -0.4694% +/- 0.0009% (-18575169.00 +/- 37220.97)
branch-misses:u: -1.0009% +/- 0.7189% (-575742.67 +/- 413521.70)
Change-Id: Ia87458ee599977e6efdc9f0e7aa283a41f84f616
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/326900
Reviewed-by: Johnni Winther <johnniwinther@google.com>
Reviewed-by: Morgan :) <davidmorgan@google.com>
Commit-Queue: Paul Berry <paulberry@google.com>
Paul Berry
2887fe0b2c