From b92fb630435e4850e719729eda255164eb4d185c Mon Sep 17 00:00:00 2001
From: Chris Tate <chris@ctate.dev>
Date: Fri, 23 Oct 2026 11:49:45 -0500
Subject: [PATCH] Mark typed-array header loads invariant

A typed array's length, data pointer and shared backing are fixed when
the view is constructed. Each element access loads the data pointer
beside the length, ahead of its bounds branch. LLVM then sinks the data
pointer load into the in-bounds block, and LICM refuses to hoist a load
that runs only on some iterations. A loop whose body branches or stores
an element therefore reloaded the data pointer for every access.

Header loads now carry `!invariant.load` next to their alias tag, so
the data pointer and length leave such loops. Debug builds emit neither.
---
 packages/compiler/src/backend/llvm/bytes-alias.ts      | 10 ++++++++--
 .../src/backend/llvm/bytes-loop-versions.test.ts       |  5 +++++
 packages/compiler/src/backend/llvm/emit-statements.ts  |  6 ++----
 3 files changed, 15 insertions(+), 6 deletions(-)

diff --git a/packages/compiler/src/backend/llvm/bytes-alias.ts b/packages/compiler/src/backend/llvm/bytes-alias.ts
index 9aaeaa6c7..d43620fd9 100644
--- a/packages/compiler/src/backend/llvm/bytes-alias.ts
+++ b/packages/compiler/src/backend/llvm/bytes-alias.ts
@@ -7,9 +7,15 @@ import type { LlvmEmitterContext } from "./expr-context.js";
  * buffer. Neither overlaps the other, a native field, array storage or a
  * module global's slot, so element stores leave header loads and global
  * receiver slots loop-invariant. Untagged accesses and calls still alias
- * every category. */
+ * every category.
+ *
+ * Header loads are also `!invariant.load`: no store or call between the
+ * construction of a view and its release changes them. Without it, a data
+ * pointer load ahead of a bounds branch is sunk into the in-bounds block,
+ * and LICM keeps a conditionally executed load inside the loop. */
 export function bytesHeaderAlias(host: LlvmEmitterContext): string {
-  return host.memoryAliasAttachment("bytes:header");
+  const tag = host.memoryAliasAttachment("bytes:header");
+  return tag === "" ? "" : `${tag}, !invariant.load !{}`;
 }
 
 export function bytesElementAlias(host: LlvmEmitterContext): string {
diff --git a/packages/compiler/src/backend/llvm/bytes-loop-versions.test.ts b/packages/compiler/src/backend/llvm/bytes-loop-versions.test.ts
index cee91c66d..34b75861c 100644
--- a/packages/compiler/src/backend/llvm/bytes-loop-versions.test.ts
+++ b/packages/compiler/src/backend/llvm/bytes-loop-versions.test.ts
@@ -249,6 +249,11 @@ test("emitted loops version the body and tag typed-array memory", () => {
   const elements = /store double [^\n]*!tbaa !(\d+)/.exec(body)![1];
   const header = /load ptr, ptr [^\n]*!tbaa !(\d+)/.exec(body)![1];
   expect(elements).not.toBe(header);
+  // Header loads are invariant for the view's lifetime; element accesses
+  // are not.
+  expect(body).toMatch(/load ptr, ptr [^\n]*, !tbaa !\d+, !invariant\.load !\{\}$/m);
+  expect(body).toMatch(/load i64, ptr [^\n]*, !tbaa !\d+, !invariant\.load !\{\}$/m);
+  expect(body).not.toMatch(/(load double|store double)[^\n]*invariant\.load/);
 });
 
 test("loops to a typed array's length version the other typed arrays they index", () => {
diff --git a/packages/compiler/src/backend/llvm/emit-statements.ts b/packages/compiler/src/backend/llvm/emit-statements.ts
index 48f7a1d5f..aef5ccf30 100644
--- a/packages/compiler/src/backend/llvm/emit-statements.ts
+++ b/packages/compiler/src/backend/llvm/emit-statements.ts
@@ -64,6 +64,7 @@ import {
   emitPairedLoopSafety,
 } from "./counted-loops.js";
 import { emitBytesLoopVersionGuard, type BytesLoopVersion } from "./bytes-loop-versions.js";
+import { bytesHeaderAlias } from "./bytes-alias.js";
 import { mangleLocal } from "../mangle.js";
 import { LlvmUnsupportedError } from "./unsupported.js";
 import type { LlValue } from "./expr-context.js";
@@ -1229,10 +1230,7 @@ export function emitStmtBody(emitter: LlEmitter, s: IrStmt): void {
             }
             B.line(`${inBounds} = icmp ult i32 ${narrow}, ${len}`);
           } else {
-            emitter.markMemoryPointer(lenPtr, "bytes:header");
-            B.line(
-              `${len} = load ${emitter.sizeType}, ptr ${lenPtr}${emitter.fieldAliasAttachment(lenPtr)}`,
-            );
+            B.line(`${len} = load ${emitter.sizeType}, ptr ${lenPtr}${bytesHeaderAlias(emitter)}`);
             B.line(`${inBounds} = icmp ult ${emitter.sizeType} ${index}, ${len}`);
           }
           B.condBr(inBounds, lb, exit);