diff --git a/pkg/compiler/doc/split_constraints.md b/pkg/compiler/doc/split_constraints.md
new file mode 100644
index 0000000..e0b47c6
--- /dev/null
+++ b/pkg/compiler/doc/split_constraints.md
@@ -0,0 +1,385 @@
+# Guide to Program Split Constraints in Dart2js
+
+## Introduction
+Deferred loading can be a very powerful tool for improving IPL in
+Dart2js. However, if Dart2js generates too many part files, the user
+experience can degrade significantly. This degradation occurs
+because each part file has to be downloaded and initialized. While
+the download overhead can be minimized via bundling, there is
+currently no way to alleviate initialization overhead.
+
+The Dart2js team has explored ways to reduce part file
+initialization overhead, and we have ideas for further incremental
+improvements. However, ‘zero’ overhead is impossible as a
+significant source of initialization overhead can be attributed to
+parsing and compiling JS, and a naive restructuring of part files
+for reduced overhead will significantly harm steady state
+performance.
+
+Though someday initializing part files may be as close to ‘free’ as
+possible, in the meantime clients want to ship more granular apps,
+and thus there is an urgent need for addressing initialization
+overhead. One obvious way to address initialization overhead is to
+reduce the number of part files. Unfortunately, doing this
+automatically turns out to be non-trivial. Dart2js simply does not
+have enough information to intelligently reduce the number of part
+files, and a naive reduction will do more harm than good, bloating
+load lists with code that may be totally unrelated.
+
+In order to reduce the number of part files efficiently, Dart2js
+needs accurate load order information. Program split constraints are
+a way for user’s of Dart2js to supply load order information to the
+compiler. With accurate information about the relative ordering of
+loadLibrary calls, Dart2js is able to constrain the deferred loading
+graph, and reduce the number of part files all without bloating load
+lists unnecessarily.
+
+One final point, program split constraints are only useful if a
+program is structured in a hierarchical manner, i.e. where
+loadLibrary calls frequently dominate other loadLibrary calls. If
+programs are structured hierarchically, and a complete load order
+graph can be provided, then Dart2js can reduce the number of part
+files, in many cases very significantly.
+
+
+## Constraints
+Practically, a program split constraints file is just a yaml list of
+constraint nodes. There are different types of constraint nodes,
+with different properties, but nearly all of them reduce the number
+of part files. Writing constraint files requires a good
+understanding of the ordering of loadLibrary calls in a program. It
+is worth pointing out that constraints always affect performance,
+and never program correctness. Thus, incorrect use of constraints
+should never break a valid Dart program, but can negatively impact
+performance.
+
+### Reference Nodes
+The most basic node type is a reference node. A reference node is
+just a way to create a symbol which represents a deferred import.
+
+For example:
+**foo.dart:**
+
+```dart
+import ‘...’ deferred as baz;
+```
+
+**constraints.yaml:**
+
+```yaml
+  ...
+  -  type: reference
+     name: baz
+     import: /path/to/foo.dart#baz
+  ...
+```
+
+Creates a reference ‘baz’ which can be used in other nodes. We could
+support references inline with the body of a node, but explicit
+reference nodes do help keep constraints organized.
+
+### Order Nodes
+The most important constraint node is the order node. An order node
+indicates that the ‘predecessor’ temporally dominates a given
+‘successor’ and thus Dart2js should ensure any code shared between
+predecessor and successor loads with the predecessor.
+
+Sequencing even just two nodes will reduce the total number of
+output units if those nodes share code but because sequencing is a
+transitive operation the real benefit comes from deep hierarchies.
+
+For example:
+
+**foo.dart:**
+
+```dart
+import ‘...’ deferred as step1;
+import ‘...’ deferred as step2;
+
+do() {
+  step1.loadLibrary().then((_) { step2.loadLibrary().then(...) } );
+}
+```
+
+**constraints.yaml:**
+
+```yaml
+  ...
+   - type: order
+     predecessor: step1
+     successor: step2
+  ...
+```
+
+### Combiner Nodes
+Order nodes support both fan in and fan out, that is multiple
+predecessors mapping to the same successor and multiple successors
+mapping to the same predecessor. In addition to fan in / out, other
+ways of combining constraints are supported via explicit combiner
+nodes.
+
+The primary purpose of combiner nodes is to allow users to propagate
+ordering information deeper into the deferred graph. However,
+combiner nodes will also reduce the number of part files in many
+cases.
+
+For v0, combiner nodes must have only reference nodes as children,
+though in the longer term we may relax this restriction and allow
+combiners to nest under certain situations. A further limitation is
+that we do not currently support cycles in the constraints graph,
+and thus every reference to a node references the exact same node.
+
+#### ‘And’ Combiner Node
+‘And’ nodes are used for cases where multiple loadLibrary calls are
+guaranteed to occur at a certain point in time, but the relative
+ordering of those nodes may change.
+
+In the case of an ‘and’ node, because all of the nodes in an ‘and’
+node are guaranteed to load at a given point in time, Dart2js can
+merge certain part files. In the case of the below example, any code
+shared between step1a and its successors can be merged into step1a.
+And code shared between step1a and its predecessors can be merged
+into its predecessors. The same optimizations will also apply to
+step1b.  However step1a and step1b themselves will not merge.
+
+For example:
+
+**foo.dart**:
+
+```dart
+import ‘...’ deferred as step1a;
+import ‘...’ deferred as step1b;
+
+do() {
+  if (...) {
+    step1a.loadLibrary().then((_) { step1b.loadLibrary().then(...) } );
+  } else {
+    step1b.loadLibrary().then((_) { step1a.loadLibrary().then(...) } );
+  }
+}
+```
+
+**constraints.yaml**:
+
+```yaml
+  ...
+  - type: and
+    name: step1
+    nodes:
+      - step1a
+      - step1b
+  ...
+```
+
+#### ‘Or’ Combiner Node
+‘Or’ nodes are used for cases where at least one of multiple
+loadLibrary calls will occur at a certain point in time. Nodes
+within an ‘or’ node need not be mutually exclusive.
+
+Because at least one of the nodes within the ‘or’ is guaranteed to
+load at a certain point in time, Dart2js can perform optimizations
+with the code shared between all of the nodes in the ‘or.’
+Specifically, code shared between successors of the ‘or’ node and
+all of the nodes in the ‘or’ node can merge with the shared ‘or’
+code, and code shared between all of the nodes in the ‘or’ and
+predecessors can merge with predecessors without bloating load
+lists.
+
+For example:
+
+**foo.dart:**
+
+```dart`
+import ‘...’ deferred as step1a;
+import ‘...’ deferred as step1b;
+
+do() {
+  if (...) {
+    step1a.loadLibrary().then((_) { ... });
+  } else {
+    step1b.loadLibrary().then((_) { ... } );
+  }
+}
+```
+
+**constraints.yaml:**
+
+```yaml
+  ...
+  - type: or
+    name: step1
+    nodes:
+      - step1a
+      - step1b
+  ...
+```
+
+#### Fuse Combiner Node
+Fuse nodes combine multiple nodes into a strongly connected
+component. This is very useful in cases where two loadLibrary calls
+almost always happen together. Fuse nodes can greatly reduce the
+number of part files.
+
+For example:
+
+**foo.dart:**
+```dart
+import ‘...’ deferred as step1a;
+import ‘...’ deferred as step1b;
+
+do() {
+  ...
+}
+```
+
+**constraints.yaml:**
+```yaml
+  ...
+  - type: fuse
+    name: step1
+    nodes:
+      - step1a
+      - step1b
+  ...
+```
+
+## Examples
+Below is a more complete example. The part files impact section is
+based on a worst case scenario where code is shared between all
+combinations of deferred imports.
+
+**foo.dart**:
+
+```dart
+import ‘...’ as deferred S1;
+import ‘...’ as deferred S2a;
+import ‘...’ as deferred S2b;
+import ‘...’ as deferred S3;
+
+main() {
+  S1.loadLibrary().then((_) {
+    if (...) {
+      S2a.loadLibrary().then((_) {
+        S2b.loadLibrary().then((_) { S3.loadLibrary().then((_) {...}); });
+      });
+    } else {
+      S2b.loadLibrary().then((_) {
+        S2a.loadLibrary().then((_) { S3.loadLibrary().then((_) {...}); });
+      });
+    }
+  });
+}
+```
+
+**constraints.yaml**:
+
+```yaml
+  - type: reference
+    name: s1
+    import: /path/to/foo.dart#S1
+  - type: reference
+    name: s2a
+    import: /path/to/foo.dart#S2a
+  - type: reference
+    name: s2b
+    import: /path/to/foo.dart#S2b
+  - type: reference
+    name: s3
+    import: /path/to/foo.dart#S3
+  - type: $COMBINER_TYPE
+    name: s2
+    nodes: [ s2a, s2b ]
+  - type: order
+    predecessor: s1
+    successor: s2
+  - type: order
+    predecessor: s2
+    successor: s3
+```
+
+**part files impact**:
+
+**unconstrained**
+* {S1}
+* {S2a}
+* {S2b}
+* {S3}
+* {S1, S2a}
+* {S1, S2b}
+* {S1, S3}
+* {S2a, S2b}
+* {S2a, S3}
+* {S2b, S3}
+* {S1, S2a, S2b}
+* {S1, S2a, S3}
+* {S1, S2b, S3}
+* {S2a, S2b, S3}
+* {S1, S2a, S2b, S3}
+
+**COMBINER\_TYPE = or**
+* {S2a}
+* {S2b}
+* {S3}
+* {S2a, S3}
+* {S2b, S3}
+* {S2a, S2b, S3}
+* {S1, S2a, S2b, S3}
+
+**COMBINER\_TYPE = and**
+* {S3}
+* {S2a, S3}
+* {S2b, S3}
+* {S2a, S2b, S3}
+* {S1, S2a, S2b, S3}
+
+**COMBINER\_TYPE = fuse**
+* {S3}
+* {S2a, S2b, S3}
+* {S1, S2a, S2b, S3}
+
+**load list impact(redundant loads are ~~crossed out~~):**
+
+**unconstrained**
+* S1 : {S1, S2a, S2b, S3}, {S1, S2a, S2b}, {S1, S2a, S3}, {S1, S2b, S3}, {S1, S2a}, {S1,  S2b}, {S1, S3}, {S1}
+* S2a: ~~{S1, S2a, S2b, S3}, {S1, S2a, S2b}, {S1, S2a, S3}~~, {S2a, S2b, S3}, ~~{S1, S2a}~~, {S2a, S2b}, {S2a, S3}, {S2a}
+* S2b: ~~{S1, S2a, S2b, S3}, {S1, S2a, S2b}, {S1, S2b, S3}~~, {S2a, S2b, S3}, ~~{S1, S2b}~~, {S2a, S2b}, {S2b, S3}, {S2b}
+* S3 : ~~{S1, S2a, S2b, S3}, {S2a, S2b, S3}, {S1, S2a, S3}, {S1, S2b, S3}, {S2a, S3}, {S2b, S3}, {S1, S3}~~, {S3}
+
+**COMBINER_TYPE = or**
+* S1 : {S1, S2a, S2b, S3}
+* S2a: ~~{S1, S2a, S2b, S3}~~, {S2a, S2b, S3}, {S2a, S3}, {S2a}
+* S2b: ~~{S1, S2a, S2b, S3}~~, {S2a, S2b, S3}, {S2b, S3}, {S2b}
+* S3 : ~~{S1, S2a, S2b, S3}, {S2a, S2b, S3}, {S2a, S3}, {S2b, S3}~~, {S3}
+
+**COMBINER_TYPE = and**
+* S1: {S1, S2a, S2b, S3}
+* S2a: ~~{S1, S2a, S2b, S3}~~, {S2a, S2b, S3}, {S2a, S3}
+* S2b: ~~{S1, S2a, S2b, S3}~~, {S2a, S2b, S3}, {S2b, S3}
+* S3 : ~~{S1, S2a, S2b, S3}, {S2a, S2b, S3}, {S2a, S3}, {S2b, S3}~~, {S3}
+
+**COMBINER_TYPE = fuse**
+* S1: {S1, S2a, S2b, S3}
+* S2a: ~~{S1, S2a, S2b, S3}~~, {S2a, S2b, S3}
+* S2b: ~~{S1, S2a, S2b, S3}~~, {S2a, S2b, S3}
+* S3 : ~~{S1, S2a, S2b, S3}, {S2a, S2b, S3}~~, {S3}
+
+## TODO
+* Documentation for Inline references, glob support, groups, set
+  operations on groups, nested combiners
+
+## Glossary
+* deferred import: A method of asynchronously loading code. Deferred
+  imports are used to reduce the size of the main part file and thus
+  improve IPL.
+* deferred load: The runtime implementation of a deferred import.
+* load list: A list of part files which must be loaded before a
+  loadLibrary call completes.
+* main part file: The part file representing the initial chunk of
+  code which must be downloaded and initialized before a given
+  program can run.
+* part file: A chunk of JS code representing some subset of the
+  compiled output which results from compiling a Dart program to JS
+  via Dart2js.
+* IPL: Stands for initial page load, i.e. the time it takes for a
+  given web page to complete its first load.
+