fix(compiler): == / != on references is identity; only text compares by content

emit_bin_vals lowered every pointer-typed ==/!= to @lp_str_eq, so two
distinct records compared their bytes up to the first zero byte. Content
compare now needs both sides to be text (string, or the untyped
pointer/ptr runtime code carries text in); other references use
icmp eq ptr, and string vs a non-text reference is a compile error.
Adds selfhost/tests/identity.ludic; both seeds regenerated.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

LANGUAGE.md

@@ -1011,7 +1011,8 @@ are **`& | ^ << >> ~`** (`>>` is a logical/unsigned shift).
 **Strings are values.** `a + b` concatenates two strings, and `a == b` / `a != b`
 compare them **by content** (not by pointer). `"go" + dir == "goleft"` works as
 written. (Under the hood these call a small emitted string runtime; a `==`/`!=`
-against `null` is still a pointer test.)
+against `null` is still a pointer test. Every other reference — records, slices,
+enums — compares by identity, and comparing a string with one is a compile error.)
 
 **Interpolation is the readable way to build them.** A backtick string
 `` `text {expr} text` `` embeds any expression in `{…}` — numbers, bools and

changes/reference-identity.md

@@ -0,0 +1,10 @@
+bump: patch
+type: fix
+**`==` / `!=` on records is identity again** — only two strings compare by content.
+
+Every pointer-typed comparison used to be lowered to a C-string content compare,
+so two distinct records (properties, slices, enums) compared their bytes up to
+the first zero byte: `a == b` could be true for different objects that shared a
+leading field, depending on layout. References now compare by identity
+(`icmp eq ptr`); `string` (and untyped `pointer`/`pointers` text) still compares by content,
+and comparing a `string` with a non-string reference is a compile error.

selfhost/backend/emit_expr.ludic

@@ -161,19 +161,31 @@ function emit_bin(e: Node) -> Val {
   return r
 }
 
+# text types: `==` on two of these compares content, not the pointers. Besides
+# `string`, that is the untyped raw pointer — `pointer`, and `ptr`, the element of
+# a `pointers` buffer — which runtime code uses to carry text.
+function is_textish(t: pointer) -> bool { return (t == "string") or (t == "pointer") or (t == "ptr") }
+
 # lower `a <op> b` on two already-evaluated operands. Shared by binary expressions
 # and compound assignment (`x += y` is exactly `x = x + y`), so both agree on
 # string concatenation, Q16.16 multiply/divide, and int->long promotion.
 # `isnull` marks a comparison against the literal null (pointer identity, not
 # string content).
 function emit_bin_vals(op: pointer, a: Val, b: Val, isnull: bool) -> Val {
-  # strings are pointer-typed, so any `+` with a pointer operand is concatenation,
-  # and `==`/`!=` between pointers is content comparison — except `x == null`,
-  # which is a pointer-identity test and falls through to the icmp below.
+  # strings are pointer-typed, so any `+` with a pointer operand is concatenation.
+  # `==`/`!=` compares by content only when BOTH sides are text (`string`, or an
+  # untyped raw pointer, see is_textish); every other reference — records,
+  # slices, enums, buffers — compares by identity in the icmp below, as does
+  # `x == null`. A string against a non-text reference is a type error.
   let ptrish = (llty(a.ty) == "ptr") or (llty(b.ty) == "ptr")
   if ptrish {
     if (op == ("+")) { return emit_str_op("+", a, b) }
-    if ((op == ("==")) or (op == ("!="))) and not isnull { return emit_str_op(op, a, b) }
+    if ((op == ("==")) or (op == ("!="))) and not isnull {
+      if is_textish(a.ty) and is_textish(b.ty) { return emit_str_op(op, a, b) }
+      if ((a.ty == "string") and (llty(b.ty) == "ptr")) or ((b.ty == "string") and (llty(a.ty) == "ptr")) {
+        perr(`cannot compare {a.ty} with {b.ty} using {op}: strings compare by content, other references by identity`)
+      }
+    }
   }
   let fx = (a.ty == "fixed") or (b.ty == "fixed")
   # a 64-bit operand (and no fixed/ptr involved) promotes the whole expression to

selfhost/tests/identity.ludic

@@ -0,0 +1,21 @@
+program T {
+  # `==` / `!=` on two references is identity; only strings compare by content.
+  # The records below start with a zero/equal field, which a byte compare would
+  # have called equal.
+  property Thing { kind: int = 0, name: string = "" }
+  entry {
+    let a = new Thing
+    let b = new Thing
+    print(a == b)        # 0: distinct records, same (zero) leading field
+    print(a == a)        # 1: the same record
+    a.kind = 3
+    b.kind = 3
+    print(a != b)        # 1: equal fields, still distinct
+    let c = a
+    print(c == a)        # 1: an alias is the same record
+    let xs: []Thing = [a, b]
+    print(xs[1] == b)    # 1: identity through a slice
+    let s = "ab"
+    print(s + "c" == "abc")   # 1: strings still compare by content
+  }
+}

tools/ludic-cli/selfhost_test.ludic

@@ -82,6 +82,7 @@ function cmd_selfhost_test() -> int {
 
   print("== the self-host compiler is correct ==")
   sh_case("structs", "7 9 109 2 42")
+  sh_case("identity", "0 1 1 1 1 1")
   sh_case("slices", "0 20 361 777")
   sh_case("sort", "0 39 20 1 2 3 3 1 30 10 20 40 2 8")
   sh_case("control", "55 4 15 1")