1 files changed, 133 insertions, 0 deletions
diff --git a/docs/chapter16/README.md b/docs/chapter16/README.md
new file mode 100644
index 00000000..a9d9422d
--- /dev/null
+++ b/docs/chapter16/README.md
@@ -0,0 +1,133 @@
+# Chapter 16: Constructors
+
+English | [日本語](README.ja.md)
+
+[Previous: Chapter 15](../chapter15/README.md) |
+[Next: Chapter 17a](../chapter17a/README.md)
+
+# Table of Contents
+
+1. [Constructors](#constructors)
+2. [Initialization code](#initialization-code)
+3. [Multiple Decls](#multiple-declarations-of-the-same-type)
+4. [Memory through constructors](#memory-through-constructors)
+
+This chapter adds field defaults and constructor blocks. Chapter 17a will add
+fixed bindings and read-only access, using these constructors to initialize them.
+
+You can also read [this chapter](https://github.com/SeaOfNodes/Simple/tree/linear-chapter16) in a linear Git revision history on the [linear](https://github.com/SeaOfNodes/Simple/tree/linear) branch and [compare](https://github.com/SeaOfNodes/Simple/compare/linear-chapter15...linear-chapter16) it to the previous chapter.
+
+## Constructors
+
+A big problem with the [Chapter 12](../chapter12/README.md) refs is that
+not-null fields always start out `null`.  This is fixed in this chapter, where
+a *constructor* syntax is required to initialize not-null fields.
+
+Fields are initialized in three ways:
+
+1. A default zero/null initialization for fields that allow it.  No special
+   syntax is required.
+```
+int x;                    // x is initialized to 0
+struct Ref { Ref? ptr; }; // Instances of Ref will have ptr initialized to null
+return new Ref.ptr;       // Returns a null
+```
+
+2. The type declaration can specify an initial value.  Later allocations will
+   start with this value.
+```
+int x = 5;                // x is initialized to 5
+struct Point { int x=1; int y=1; }; // Point x and y will start as 1, not 0
+return new Point.x;       // Returns a 1
+```
+
+3. The allocation can specify an initial value.
+
+```
+struct Point { int x=1; int y=1; }; // Point x and y will start as 1, not 0
+return new Point { x=3; }.x;        // Returns a 3
+```
+
+Not-null fields *must* be initialized before first use and before the
+end of the allocation.  They can be initialized in either the declaration or
+allocation.  They do not start with the default value, although the
+initialization can be to the default.
+
+```
+struct Person { u8[] name; };
+return new Person; // ERROR: 'Person' is not fully initialized, field 'name' needs to be set in a constructor
+```
+
+
+## Initialization code
+
+The parser allows a full Block statement to be parsed in either the type
+declaration or the allocation.  Any amount of code is legal, including
+`while` loops and `return`s.
+
+```
+struct Square {
+    flt side = arg;
+    flt diag = arg*arg/2;
+    // Newton's approximation to the square root, computed in a constructor.
+    // The actual allocation will copy in this result as the initial
+    // value for 'diag'.
+    while( 1 ) {
+        // The next-guess variable "next" is not a field in Square,
+        // because it does not appear at the top level
+        flt next = (side/diag + diag)/2;
+        if( next == diag ) break;
+        diag = next;
+    }
+};
+return new Square;
+```
+
+Another example:
+
+```
+struct Buffer {
+    if( arg < 0 || arg > 1000000 )
+      return null; // Size out of bounds
+    u8[] buffer = new u8[arg];
+};
+return new Buffer;
+```
+
+
+### Multiple declarations of the same type
+
+Also new this chapter is allowing multiple declarations with the same type, as
+is commonly seen in other languages:
+
+```
+int x,y; // Two int variables declared
+struct Point { int x,y,z; }; // Three fields declared
+```
+
+## Memory through constructors
+
+The lazy memory partitioning from [Chapter 11](../chapter11/README.md) continues
+through constructors. The parser tracks one `$mem` variable alongside its scalar
+variables. Branches and loops merge that binding with BulkMemPhi; field access
+splits out precise MemPhis when needed. The variable records for declared types live in ScopeNode. There is no separate parser table of memory aliases.
+
+A constructor computes its field values before the allocation. Reads, writes,
+and loops in its body update the same `$mem` binding as ordinary code. New then
+takes `{ctrl, $mem, size, field values...}` and produces `{ptr, $mem}`. The input
+memory is a partial MemMerge containing the struct's aliases; the output memory
+covers those same aliases. A whole-memory MemMerge preserves unrelated slices.
+Arrays use the same layout, supplying their length and default element value as
+initializer inputs.
+
+A Load from its own New can use the matching initializer input. Alias contents
+combine the initializer's type with incoming contents from older objects, so a
+new object does not erase facts or effects for existing objects. Loads bypass
+only allocations proven to be distinct. Scheduling follows partial aggregates
+when establishing read-before-write order.
+
+Sharing one allocation memory result makes Store-to-New folding more
+conservative: the existing sole-use check now sees users of every covered
+alias. This can leave a field Load where separate memory projections previously
+allowed a constant fold. Phi factoring also retains the simple one-step Load
+safety check.