From 35cf7efe551f8d07fec0620db12620abe5d70be7 Mon Sep 17 00:00:00 2001
From: Chris Tate <chris@ctate.dev>
Date: Tue, 13 Oct 2026 07:29:48 -0500
Subject: [PATCH] Make short async calls and awaits cheap

---
 packages/compiler/src/backend/llvm/emitter.ts | 144 ++++++++-
 .../compiler/src/backend/llvm/expr-async.ts   |  12 +
 .../compiler/src/backend/llvm/expr-calls.ts   |  20 ++
 packages/compiler/src/backend/mangle.ts       |   5 +
 .../compiler/test/await-call-emission.test.ts |  73 +++++
 .../corpus/4880-await-call-turn-order.ts.json |   6 +
 .../corpus/4881-await-call-exceptions.ts.json |   6 +
 .../4882-await-call-request-chain.ts.json     |   6 +
 .../corpus/4883-await-values-and-hops.ts.json |   6 +
 .../4884-await-call-unhandled-sibling.ts.json |   6 +
 .../4885-await-call-checkpoints.ts.json       |   6 +
 packages/runtime/src/scr_async.c              | 282 +++++++++++++++---
 packages/runtime/src/scr_runtime.h            |  11 +
 tests/corpus/4880-await-call-turn-order.ts    |  88 ++++++
 tests/corpus/4881-await-call-exceptions.ts    |  82 +++++
 tests/corpus/4882-await-call-request-chain.ts |  53 ++++
 tests/corpus/4883-await-values-and-hops.ts    |  32 ++
 .../4884-await-call-unhandled-sibling.ts      |  22 ++
 tests/corpus/4885-await-call-checkpoints.ts   |  40 +++
 19 files changed, 847 insertions(+), 53 deletions(-)
 create mode 100644 packages/compiler/test/await-call-emission.test.ts
 create mode 100644 packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4880-await-call-turn-order.ts.json
 create mode 100644 packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4881-await-call-exceptions.ts.json
 create mode 100644 packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4882-await-call-request-chain.ts.json
 create mode 100644 packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4883-await-values-and-hops.ts.json
 create mode 100644 packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4884-await-call-unhandled-sibling.ts.json
 create mode 100644 packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4885-await-call-checkpoints.ts.json
 create mode 100644 tests/corpus/4880-await-call-turn-order.ts
 create mode 100644 tests/corpus/4881-await-call-exceptions.ts
 create mode 100644 tests/corpus/4882-await-call-request-chain.ts
 create mode 100644 tests/corpus/4883-await-values-and-hops.ts
 create mode 100644 tests/corpus/4884-await-call-unhandled-sibling.ts
 create mode 100644 tests/corpus/4885-await-call-checkpoints.ts

diff --git a/packages/compiler/src/backend/llvm/emitter.ts b/packages/compiler/src/backend/llvm/emitter.ts
index b01d12c29..526cc9f78 100644
--- a/packages/compiler/src/backend/llvm/emitter.ts
+++ b/packages/compiler/src/backend/llvm/emitter.ts
@@ -95,6 +95,7 @@ import { RUNTIME_ABI_MARKER } from "../runtime-abi.js";
 import { computeMayThrow } from "../may-throw.js";
 import {
   mangleArgPack,
+  mangleAsyncAwaitCall,
   mangleAsyncSpawn,
   mangleBorrowedFunction,
   mangleClassObj,
@@ -659,6 +660,9 @@ export class LlEmitter {
   /** Method names with at least one may-throw implementation — the
    * virtualCall pending check's key. */
   readonly mayThrowMethods = new Set<string>();
+  /** Async functions awaited directly at some `await f(...)` site; each
+   * gets an await-call entry (emitAsyncAwaitCall). */
+  readonly awaitCallees = new Set<string>();
   /** setTimeout and friends appeared somewhere: main must run the event
    * loop even in programs with no async functions. */
   usesTimers = false;
@@ -2996,8 +3000,14 @@ export class LlEmitter {
     const definitions = [`%${pack} = type { ${fieldTys.join(", ") || "i8"} } ; ${fn.name} args`];
     const sizeOf = `ptrtoint (ptr getelementptr (%${pack}, ptr null, i32 1) to ${this.sizeType})`;
 
-    this.declare(`declare void @free(ptr)`);
-    this.declare(`declare ptr @malloc(${this.sizeType})`);
+    // Async argument packs live from the spawn to the trampoline's first
+    // instructions, so they take the runtime's small-block allocator.
+    // Generator packs keep malloc: the runtime frees never-started ones.
+    const smallPack = fn.generator === undefined;
+    const packAlloc = smallPack ? "scr_rt_calloc" : "malloc";
+    const packFree = smallPack ? "scr_rt_free" : "free";
+    this.declare(`declare void @${packFree}(ptr)`);
+    this.declare(`declare ptr @${packAlloc}(${this.sizeType})`);
 
     const tr: string[] = [
       `define internal void @${mangleTrampoline(fn.name)}(ptr %self, ptr %ap) ${FN_ATTRS} {`,
@@ -3011,7 +3021,7 @@ export class LlEmitter {
       );
       loads.push(`${ty} %a${i}`);
     });
-    tr.push(`  call void @free(ptr %ap)`);
+    tr.push(`  call void @${packFree}(ptr %ap)`);
     const ret = fn.returnType;
     const retTy = this.llType(ret);
     const bodyCall = `call ${retTy} @${mangleFunction(fn.name)}(${loads.join(", ")})`;
@@ -3022,7 +3032,7 @@ export class LlEmitter {
 
     const spawnParams = fieldTys.map((ty, i) => `${ty} %a${i}`);
     const argPackLines = [
-      `  %ap = call ptr @malloc(${this.sizeType} ${sizeOf})`,
+      `  %ap = call ptr @${packAlloc}(${this.sizeType} ${sizeOf})`,
       `  %isnull = icmp eq ptr %ap, null`,
       `  br i1 %isnull, label %oom, label %ok`,
       `oom:`,
@@ -3094,7 +3104,7 @@ export class LlEmitter {
     const eligible = new Set<string>();
     if (this.wasi || this.mod.workers === true) return eligible;
     const suspends =
-      /@scr_(?:await_|module_await\b|promise_await_settled\b|test_sub\b|gen_yield_|async_gen_|fiber_)/;
+      /@(?:scr_(?:await_|module_await\b|promise_await_settled\b|test_sub\b|gen_yield_|async_gen_|fiber_)|sc_aw_)/;
     for (const decl of this.decls) if (/@scr_test_sub\b/.test(decl)) return eligible;
     for (const text of this.fnDefText.values()) if (/@scr_test_sub\b/.test(text)) return eligible;
     const byMangled = new Map<string, string>();
@@ -3180,6 +3190,118 @@ export class LlEmitter {
     return out;
   }
 
+  /** Whether `await f(...)` may enter f through its await-call entry. The
+   * entry's result is the awaited value of f's own body type; direct calls
+   * never carry a lifted environment. */
+  awaitCallEligible(fnName: string): boolean {
+    if (this.wasi || this.mod.workers === true) return false;
+    const fn = this.fnByName.get(fnName);
+    if (fn === undefined || fn.async !== true || fn.generator !== undefined) return false;
+    if (fn.captures !== undefined || fn.returnType.kind === "dyn") return false;
+    return fn.asyncCacheGlobal === undefined && fn.asyncCycleCacheGlobal === undefined;
+  }
+
+  /** The await-call entry of an async function (`sc_aw_*`): one direct
+   * `await f(...)` as a call returning the awaited value, so the call site
+   * needs no promise of its own. A fiberless body runs on the caller's
+   * stack as in its spawn wrapper; a clean result then takes the await's
+   * one microtask turn (scr_await_hop) and is returned without ever
+   * materializing the promise, which nothing else could observe. A throw
+   * builds the rejected promise exactly as the spawn wrapper would and
+   * awaits it, so the rejection reaches the awaiter on the same turn. Other
+   * bodies pack their arguments as the spawn wrapper does and enter
+   * scr_async_spawn_await_*, which parks the caller on the promise as soon
+   * as the body first suspends. */
+  private emitAsyncAwaitCall(fn: IrFunction, argPackLines: string[] | null): string[] {
+    const ret = fn.returnType;
+    const retTy = this.llType(ret);
+    const params = fn.params.map((p, i) => `${this.llType(p.type)} %a${i}`);
+    const awaitLine = (dest: string, promise: string): string => {
+      switch (ret.kind) {
+        case "void":
+          this.declare(`declare void @scr_await_void(ptr)`);
+          return `  call void @scr_await_void(ptr ${promise})`;
+        case "f64":
+        case "date":
+          this.declare(`declare double @scr_await_f64(ptr)`);
+          return `  ${dest} = call double @scr_await_f64(ptr ${promise})`;
+        case "bool":
+          this.declare(`declare zeroext i1 @scr_await_bool(ptr)`);
+          return `  ${dest} = call zeroext i1 @scr_await_bool(ptr ${promise})`;
+        case "string":
+          this.declare(`declare ptr @scr_await_str(ptr)`);
+          return `  ${dest} = call ptr @scr_await_str(ptr ${promise})`;
+        default:
+          this.declare(`declare ptr @scr_await_ref(ptr)`);
+          return `  ${dest} = call ptr @scr_await_ref(ptr ${promise})`;
+      }
+    };
+    const retLine = (v: string): string =>
+      retTy === "void" ? `  ret void` : `  ret ${retTy} ${v}`;
+    const head = `define internal ${retTy} @${mangleAsyncAwaitCall(fn.name)}(${params.join(", ")}) ${FN_ATTRS} { ; await ${fn.name}`;
+    const args = params.join(", ");
+    if (argPackLines !== null) {
+      const kind =
+        ret.kind === "void"
+          ? "void"
+          : ret.kind === "f64" || ret.kind === "date"
+            ? "f64"
+            : ret.kind === "bool"
+              ? "bool"
+              : ret.kind === "string"
+                ? "str"
+                : "ref";
+      const cret = kind === "bool" ? "zeroext i1" : retTy;
+      this.declare(`declare ${cret} @scr_async_spawn_await_${kind}(ptr, ptr)`);
+      this.needOom();
+      const call = `call ${cret} @scr_async_spawn_await_${kind}(ptr @${mangleTrampoline(fn.name)}, ptr %ap)`;
+      return [
+        head,
+        `entry:`,
+        ...argPackLines,
+        kind === "void" ? `  ${call}` : `  %v = ${call}`,
+        retLine("%v"),
+        `}`,
+        ``,
+      ];
+    }
+    this.declare(`declare void @scr_async_inline_enter(ptr)`);
+    this.declare(`declare void @scr_async_inline_exit(ptr)`);
+    this.declare(`declare void @scr_await_hop()`);
+    this.declare(`declare ptr @scr_promise_new()`);
+    this.declare(`declare void @scr_promise_reject_pending(ptr)`);
+    this.declare(`declare void @scr_promise_release(ptr)`);
+    const bodyCall = `call ${retTy} @${mangleFunction(fn.name)}(${args})`;
+    const out = [
+      head,
+      `entry:`,
+      `  %frame = alloca { ptr, ptr }`,
+      `  call void @scr_async_inline_enter(ptr %frame)`,
+      retTy === "void" ? `  ${bodyCall}` : `  %r = ${bodyCall}`,
+      ...this.pendingTestLines("%pend"),
+      `  call void @scr_async_inline_exit(ptr %frame)`,
+      `  br i1 %pend, label %thrown, label %clean`,
+      `clean:`,
+      `  call void @scr_await_hop()`,
+      retLine("%r"),
+      `thrown:`,
+    ];
+    if (ret.kind !== "void" && isRefCounted(ret)) {
+      // An escaping throw means %r is the never-read dummy (NULL).
+      out.push(`  call void ${releaseSym(this.shapeHost, ret)}(ptr %r)`);
+    }
+    out.push(
+      `  %p = call ptr @scr_promise_new()`,
+      `  call void @scr_promise_reject_pending(ptr %p)`,
+      awaitLine("%v", "%p"),
+      `  call void @scr_promise_release(ptr %p)`,
+      retLine("%v"),
+      `}`,
+      ``,
+    );
+    return out;
+  }
+
   /** Per-async-function machinery — async.ts's scaffolding, .ll
    * flavored: an argument-pack struct type, a fiber trampoline (unpacks,
    * frees the pack, runs the ordinary compiled body, settles the
@@ -3209,6 +3331,9 @@ export class LlEmitter {
     const fiberless = this.fiberlessAsyncFunctions();
     for (const fn of this.mod.functions) {
       if (fn.async !== true || fn.generator !== undefined) continue;
+      if (fiberless.has(fn.name) && this.awaitCallees.has(fn.name)) {
+        for (const line of this.emitAsyncAwaitCall(fn, null)) out.push(line);
+      }
       if (fiberless.has(fn.name)) {
         const fieldTys = [
           ...(fn.captures !== undefined ? ["ptr"] : []),
@@ -3221,6 +3346,9 @@ export class LlEmitter {
       const { definitions, ret, tr, spawnParams, argPackLines } =
         this.emitArgPackAndTrampolinePrologue(fn);
       for (const line of definitions) out.push(line);
+      if (this.awaitCallees.has(fn.name)) {
+        for (const line of this.emitAsyncAwaitCall(fn, argPackLines)) out.push(line);
+      }
       this.declare(`declare ptr @scr_fiber_promise(ptr)`);
       this.declare(`declare ptr @scr_async_spawn(ptr, ptr)`);
       this.needOom();
@@ -7443,6 +7571,12 @@ export class LlEmitter {
     return emitCallExpr(this, e);
   }
 
+  /** A direct async call an enclosing `await` consumes whole: emitted
+   * through the callee's await-call entry, yielding the awaited value. */
+  emitAwaitedCall(e: ExprOf<"call">): LlValue {
+    return emitCallExpr(this, e, true);
+  }
+
   emitRecordExpr(
     e: ExprOf<
       | "fieldGet"
diff --git a/packages/compiler/src/backend/llvm/expr-async.ts b/packages/compiler/src/backend/llvm/expr-async.ts
index 15e272796..004f8ae24 100644
--- a/packages/compiler/src/backend/llvm/expr-async.ts
+++ b/packages/compiler/src/backend/llvm/expr-async.ts
@@ -626,6 +626,18 @@ export function emitAsyncExpr(
       // re-throw here (hence the pending check). Promise temp borrowed;
       // refcounted results arrive +1 and join the frame pre-check so an
       // unwind releases the dummy (NULL) harmlessly.
+      // `await f(...)` of a direct async call: the call and the await go
+      // through f's await-call entry, which returns the awaited value
+      // (owned, with the pending check done by the call emission).
+      if (e.value.kind === "call" && host.awaitCallEligible(e.value.callee)) {
+        const callee = host.fnByName.get(e.value.callee)!;
+        if (
+          typeEquals(callee.returnType, e.type) &&
+          !host.callLifetimes.borrowed.has(e.value.callee)
+        ) {
+          return host.emitAwaitedCall(e.value);
+        }
+      }
       const pr = host.emitExpr(e.value);
       if (host.wasi) host.emitWasiSuspend(pr.name);
       if (e.type.kind === "void") {
diff --git a/packages/compiler/src/backend/llvm/expr-calls.ts b/packages/compiler/src/backend/llvm/expr-calls.ts
index 00b766795..62a00103a 100644
--- a/packages/compiler/src/backend/llvm/expr-calls.ts
+++ b/packages/compiler/src/backend/llvm/expr-calls.ts
@@ -9,6 +9,7 @@ import {
 } from "../../ir/ir.js";
 import { collectFfiRetainedOps, parseFfiCallbackKey } from "../ffi-callbacks.js";
 import {
+  mangleAsyncAwaitCall,
   mangleBorrowedFunction,
   mangleClassNew,
   mangleClassRetain,
@@ -51,6 +52,7 @@ export function emitCallExpr(
     | "instanceOf"
     | "virtualCall"
   >,
+  awaitEntry = false,
 ): LlValue {
   const B = host.B;
   switch (e.kind) {
@@ -120,6 +122,24 @@ export function emitCallExpr(
       const argList = args
         .map((a, i) => `${host.llType(callee.params[i]!.type)} ${a.name}`)
         .join(", ");
+      if (awaitEntry) {
+        // The enclosing await routes this call through the callee's
+        // await-call entry: the result is the awaited value, and the
+        // await's rejection re-throw is the pending check.
+        host.awaitCallees.add(e.callee);
+        const awaited = callee.returnType;
+        const entry = `@${mangleAsyncAwaitCall(e.callee)}`;
+        if (awaited.kind === "void") {
+          B.line(`call void ${entry}(${argList})`);
+          host.emitPendingCheck();
+          return { name: "", type: awaited };
+        }
+        const t = B.tmp();
+        B.line(`${t} = call ${host.llType(awaited)} ${entry}(${argList})`);
+        const out = host.own({ name: t, type: awaited });
+        host.emitPendingCheck();
+        return out;
+      }
       // Async callee: the spawn wrapper runs the body eagerly to its
       // first suspension and returns the promise (+1). The call itself
       // never unwinds — rejections surface at await (computeMayThrow's
diff --git a/packages/compiler/src/backend/mangle.ts b/packages/compiler/src/backend/mangle.ts
index d23e4efd5..e2fc58146 100644
--- a/packages/compiler/src/backend/mangle.ts
+++ b/packages/compiler/src/backend/mangle.ts
@@ -166,6 +166,11 @@ export function mangleGlobal(globalId: string): string {
 export function mangleAsyncSpawn(fnName: string): string {
   return `sc_as_${sanitize(fnName)}`;
 }
+/** `await f(...)` of a direct async call: the call and its await as one
+ * entry that returns the awaited value. */
+export function mangleAsyncAwaitCall(fnName: string): string {
+  return `sc_aw_${sanitize(fnName)}`;
+}
 export function mangleTrampoline(fnName: string): string {
   return `sc_tr_${sanitize(fnName)}`;
 }
diff --git a/packages/compiler/test/await-call-emission.test.ts b/packages/compiler/test/await-call-emission.test.ts
new file mode 100644
index 000000000..4051f6d5c
--- /dev/null
+++ b/packages/compiler/test/await-call-emission.test.ts
@@ -0,0 +1,73 @@
+import { mkdtemp, readFile, rm, writeFile } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { afterEach, expect, test } from "vitest";
+import { compile } from "../src/index.js";
+
+const directories: string[] = [];
+afterEach(async () => {
+  await Promise.all(directories.splice(0).map((dir) => rm(dir, { recursive: true, force: true })));
+});
+
+async function emit(source: string): Promise<string> {
+  const outDir = await mkdtemp(join(tmpdir(), "scriptc-await-call-"));
+  directories.push(outDir);
+  const entry = join(outDir, "main.ts");
+  await writeFile(entry, source);
+  const outPath = join(outDir, "program.ll");
+  const result = await compile(entry, { entry, outDir, outPath, outputKind: "llvm" });
+  expect(result.ok, JSON.stringify(result)).toBe(true);
+  return readFile(outPath, "utf8");
+}
+
+function definition(llvm: string, symbol: string): string {
+  const match = new RegExp(`^define internal [^\\n]*@${symbol}\\([^]*?^}`, "m").exec(llvm);
+  expect(match, symbol).not.toBeNull();
+  return match![0];
+}
+
+const source = `
+const cache = new Map<number, number>();
+async function leaf(x: number): Promise<number> {
+  const hit = cache.get(x);
+  return hit ?? x + 1;
+}
+async function middle(x: number): Promise<number> {
+  const y = await leaf(x);
+  return y * 2;
+}
+async function main(): Promise<void> {
+  const pending = leaf(1);
+  let sum = await middle(2);
+  sum += await pending;
+  console.log(sum);
+}
+void main();
+`;
+
+test("awaiting a direct call to a body that never suspends skips the promise", async () => {
+  const llvm = await emit(source);
+  expect(definition(llvm, "sc_f_middle")).toContain("call double @sc_aw_leaf(double");
+  const entry = definition(llvm, "sc_aw_leaf");
+  expect(entry).toContain("call double @sc_f_leaf(double %a0)");
+  expect(entry).toContain("call void @scr_async_inline_exit(ptr %frame)");
+  // The clean result takes the await's microtask turn and returns directly;
+  // only the throw path builds the rejected promise.
+  const clean = entry.slice(entry.indexOf("clean:"), entry.indexOf("thrown:"));
+  expect(clean).toContain("call void @scr_await_hop()");
+  expect(clean).not.toContain("@scr_promise_new");
+  expect(entry.slice(entry.indexOf("thrown:"))).toContain("call void @scr_promise_reject_pending");
+});
+
+test("awaiting a direct call to a suspending body spawns and awaits in one call", async () => {
+  const llvm = await emit(source);
+  const main = definition(llvm, "sc_f_main");
+  expect(main).toContain("call double @sc_aw_middle(double");
+  // A call whose promise is kept keeps the promise-returning entry.
+  expect(main).toContain("call ptr @sc_as_leaf(double");
+  expect(definition(llvm, "sc_aw_middle")).toContain(
+    "call double @scr_async_spawn_await_f64(ptr @sc_tr_middle, ptr %ap)",
+  );
+  // middle awaits, so it keeps its fiber.
+  expect(definition(llvm, "sc_as_middle")).toContain("@scr_async_spawn(");
+});
diff --git a/packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4880-await-call-turn-order.ts.json b/packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4880-await-call-turn-order.ts.json
new file mode 100644
index 000000000..f1ce33b19
--- /dev/null
+++ b/packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4880-await-call-turn-order.ts.json
@@ -0,0 +1,6 @@
+{
+ "order": [
+  "<repo>/tests/corpus/4880-await-call-turn-order.ts"
+ ],
+ "diags": []
+}
diff --git a/packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4881-await-call-exceptions.ts.json b/packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4881-await-call-exceptions.ts.json
new file mode 100644
index 000000000..38a16af52
--- /dev/null
+++ b/packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4881-await-call-exceptions.ts.json
@@ -0,0 +1,6 @@
+{
+ "order": [
+  "<repo>/tests/corpus/4881-await-call-exceptions.ts"
+ ],
+ "diags": []
+}
diff --git a/packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4882-await-call-request-chain.ts.json b/packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4882-await-call-request-chain.ts.json
new file mode 100644
index 000000000..b9888c118
--- /dev/null
+++ b/packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4882-await-call-request-chain.ts.json
@@ -0,0 +1,6 @@
+{
+ "order": [
+  "<repo>/tests/corpus/4882-await-call-request-chain.ts"
+ ],
+ "diags": []
+}
diff --git a/packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4883-await-values-and-hops.ts.json b/packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4883-await-values-and-hops.ts.json
new file mode 100644
index 000000000..f9a318b5c
--- /dev/null
+++ b/packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4883-await-values-and-hops.ts.json
@@ -0,0 +1,6 @@
+{
+ "order": [
+  "<repo>/tests/corpus/4883-await-values-and-hops.ts"
+ ],
+ "diags": []
+}
diff --git a/packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4884-await-call-unhandled-sibling.ts.json b/packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4884-await-call-unhandled-sibling.ts.json
new file mode 100644
index 000000000..91f046058
--- /dev/null
+++ b/packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4884-await-call-unhandled-sibling.ts.json
@@ -0,0 +1,6 @@
+{
+ "order": [
+  "<repo>/tests/corpus/4884-await-call-unhandled-sibling.ts"
+ ],
+ "diags": []
+}
diff --git a/packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4885-await-call-checkpoints.ts.json b/packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4885-await-call-checkpoints.ts.json
new file mode 100644
index 000000000..c9614b2bb
--- /dev/null
+++ b/packages/compiler/test/ts7/baselines/order-parity/tests/corpus/4885-await-call-checkpoints.ts.json
@@ -0,0 +1,6 @@
+{
+ "order": [
+  "<repo>/tests/corpus/4885-await-call-checkpoints.ts"
+ ],
+ "diags": []
+}
diff --git a/packages/runtime/src/scr_async.c b/packages/runtime/src/scr_async.c
index e1606a2e7..c75877fd3 100644
--- a/packages/runtime/src/scr_async.c
+++ b/packages/runtime/src/scr_async.c
@@ -163,6 +163,8 @@ typedef struct { size_t rc; } ScrPromiseIdentity;
 
 struct ScrPromise {
   size_t rc; /* The cycle collector reads the reference count at offset zero. */
+  /* Shared identity cell for promises adopted across inner types (NULL
+   * until scr_promise_share_identity first links this promise). */
   ScrPromiseIdentity *identity;
   int state;
   bool top_level_handled; /* uncaughtException handled entry evaluation */
@@ -178,8 +180,9 @@ struct ScrPromise {
   void (*release_fn)(void *);
   ScrTraceFn trace_fn;
   /* Fibers parked on this promise. */
-  ScrFiber **waiters;
+  ScrFiber **waiters; /* waiter_inline until a second fiber parks */
   size_t nwaiters, waiters_cap;
+  ScrFiber *waiter_inline;
   /* Combinator callbacks (Promise.race / Promise.all): at settle,
    * fulfilled promises run adapt(dst, this) — an emitted adapter
    * converting the payload into the destination's inner type — and
@@ -235,6 +238,37 @@ static void scr_oom(void) {
   abort();
 }
 
+static void scr_promise_free_side(ScrPromise *p) {
+  if (p->identity && --p->identity->rc == 0) free(p->identity);
+  if (p->waiters != &p->waiter_inline) free(p->waiters);
+}
+
+static void scr_promise_add_waiter(ScrPromise *p, ScrFiber *f) {
+  if (p->nwaiters == p->waiters_cap) {
+    if (p->waiters_cap == 0) {
+      p->waiters = &p->waiter_inline;
+      p->waiters_cap = 1;
+    } else {
+      size_t cap = p->waiters_cap < 4 ? 4 : p->waiters_cap * 2;
+      ScrFiber **grown = p->waiters == &p->waiter_inline
+                             ? malloc(cap * sizeof *grown)
+                             : realloc(p->waiters, cap * sizeof *grown);
+      if (!grown) scr_oom();
+      if (p->waiters == &p->waiter_inline) grown[0] = p->waiter_inline;
+      p->waiters = grown;
+      p->waiters_cap = cap;
+    }
+  }
+  p->waiters[p->nwaiters++] = f;
+}
+
+static bool scr_promise_has_traced_children(const ScrPromise *p) {
+  return p->ncbs != 0 ||
+         ((p->payload_kind == SCR_EXC_REF || p->payload_kind == SCR_EXC_OBJ ||
+           p->payload_kind == SCR_EXC_PRIMITIVE_REF) &&
+          p->trace_fn && p->payload);
+}
+
 static void scr_promise_trace(void *o, ScrTraceVisit visit, void *ctx) {
   ScrPromise *p = (ScrPromise *)o;
   /* Waiters are fibers (not refcounted objects) — the settled payload and
@@ -266,8 +300,7 @@ static void scr_promise_gcfree(void *o) {
   for (size_t i = 0; i < p->ncbs; i++) {
     if (p->cbs[i].all) scr_promise_all_state_release(p->cbs[i].all);
   }
-  if (--p->identity->rc == 0) free(p->identity);
-  free(p->waiters);
+  scr_promise_free_side(p);
   free(p->cbs);
 #ifdef SCR_RC_AUDIT
   scr_live_promises--;
@@ -279,9 +312,6 @@ ScrPromise *scr_promise_new(void) {
   static const ScrCycType cyc_type = { &scr_promise_trace, &scr_promise_gcfree, SCR_CYC_KIND_PROMISE };
   ScrPromise *p = scr_cyc_alloc(sizeof *p, &cyc_type);
   p->rc = 1;
-  p->identity = malloc(sizeof *p->identity);
-  if (!p->identity) scr_oom();
-  p->identity->rc = 1;
 #ifdef SCR_RC_AUDIT
   scr_live_promises++;
 #endif
@@ -289,14 +319,20 @@ ScrPromise *scr_promise_new(void) {
 }
 
 void scr_promise_share_identity(ScrPromise *dst, ScrPromise *src) {
+  if (dst == src) return;
+  if (!src->identity) {
+    src->identity = malloc(sizeof *src->identity);
+    if (!src->identity) scr_oom();
+    src->identity->rc = 1;
+  }
   if (dst->identity == src->identity) return;
-  if (--dst->identity->rc == 0) free(dst->identity);
+  if (dst->identity && --dst->identity->rc == 0) free(dst->identity);
   dst->identity = src->identity;
   dst->identity->rc++;
 }
 
 bool scr_promise_identity_equal(ScrPromise *a, ScrPromise *b) {
-  return a->identity == b->identity;
+  return a == b || (a->identity != NULL && a->identity == b->identity);
 }
 
 /* The 'rejectionHandled' hook (scr_async_dyn.c installs it at listener
@@ -352,8 +388,7 @@ void scr_promise_release(ScrPromise *p) {
   if (--p->rc == 0) {
     scr_cyc_dead(p);
     scr_promise_release_payload(p);
-    if (--p->identity->rc == 0) free(p->identity);
-    free(p->waiters);
+    scr_promise_free_side(p);
     for (size_t i = 0; i < p->ncbs; i++) {
       scr_promise_release(p->cbs[i].dst);
       if (p->cbs[i].all) scr_promise_all_state_release(p->cbs[i].all);
@@ -363,9 +398,14 @@ void scr_promise_release(ScrPromise *p) {
     scr_live_promises--;
 #endif
     scr_cyc_free(p);
-  } else {
+  } else if (scr_promise_has_traced_children(p)) {
     scr_cyc_released(p); /* possible cycle root; may collect — p is done */
   }
+  /* A promise the trace walks nothing from cannot be on a cycle at this
+   * decrement, so it is not a candidate root. If it later gains traced
+   * children (a cycle-capable payload or a combinator destination) while
+   * still referenced, the decrement that could strand a cycle through it
+   * sees them and buffers it then. */
 }
 
 void *scr_promise_retain_v(void *p) { return scr_promise_retain((ScrPromise *)p); }
@@ -557,6 +597,14 @@ struct ScrFiber {
   /* Async-generator settlement jobs: stackless READY-queue envelopes.
    * The retained generator is pumped on the main stack. */
   ScrGen *gen_job;
+  /* Set during the eager prefix of a fiber spawned by an await-call entry
+   * (scr_async_spawn_await_*): the spawner that awaits this fiber's promise
+   * as soon as the spawn returns. The first suspension parks it on the
+   * promise right away (scr_fiber_park_awaiters) instead of switching back
+   * just so it can park itself. */
+  struct ScrFiber *awaiter;
+  /* This fiber was parked by such a child; set before the switch away. */
+  bool await_parked;
 #ifdef SCR_ASAN_FIBERS
   void *fake_stack;
 #endif
@@ -599,6 +647,8 @@ void scr_als_ctx_release(ScrAlsCtx *c) {
 
 static SCR_TL ScrCtx scr_loop_ctx; /* the main stack (scheduler home) */
 static SCR_TL ScrFiber *scr_current = NULL;
+/* A finished fiber whose last switch went to the loop (scr_resume_fiber). */
+static SCR_TL ScrFiber *scr_loop_reap = NULL;
 static SCR_TL long scr_fibers_live = 0;
 static SCR_TL long scr_fibers_abandoned = 0;
 
@@ -1548,6 +1598,10 @@ static void scr_trampoline(void) {
   ScrFiber *self = scr_current;
   self->entry(self, self->argpack);
   scr_fiber_finish(self);
+  /* The loop reaps the fiber that comes back to it finished, which need
+   * not be the one it resumed: a fiber that continued in place for the
+   * awaiters it parked (scr_await_yield) returns to the loop too. */
+  if (self->gen == NULL && self->return_to == &scr_loop_ctx) scr_loop_reap = self;
   /* Dead fiber: hop back to whoever resumed us. The loop frees us. */
   scr_switch(&self->ctx, self->return_to, NULL);
   /* unreachable */
@@ -1626,13 +1680,24 @@ static void scr_fiber_destroy(ScrFiber *f) {
 #else
   free(f->stack);
 #endif
-  free(f);
+  scr_mem_free(f);
 }
 
+#if !defined(_WIN32) && !defined(__wasi__) && !defined(SCR_WORKERS)
+#define SCR_SPAWN_AWAIT 1
+#endif
+
 /* Spawns and EAGERLY runs an async body until its first suspension (JS's
- * synchronous-prefix rule). Returns the fiber's promise, +1. */
-ScrPromise *scr_async_spawn(void (*entry)(ScrFiber *, void *), void *argpack) {
-  ScrFiber *f = calloc(1, sizeof *f);
+ * synchronous-prefix rule). Returns the fiber's promise, +1. A non-NULL
+ * `awaiter` (the current fiber, SCR_SPAWN_AWAIT only) awaits that promise
+ * immediately: the child saves into and returns to the awaiter's own
+ * context, so a child suspension can park the awaiter and switch past
+ * it (scr_fiber_park_awaiters). The promise is retained before the child
+ * runs, as the child may then finish and be reaped by the loop before the
+ * awaiter resumes. */
+static ScrPromise *scr_async_spawn_impl(void (*entry)(ScrFiber *, void *), void *argpack,
+                                        ScrFiber *awaiter) {
+  ScrFiber *f = scr_mem_calloc(sizeof *f);
   if (!f) scr_oom();
   scr_fiber_track(f);
   f->promise = scr_promise_new();
@@ -1651,6 +1716,7 @@ ScrPromise *scr_async_spawn(void (*entry)(ScrFiber *, void *), void *argpack) {
   f->ctx = CreateFiberEx(4096, 0, 0, scr_trampoline, NULL);
   if (f->ctx == NULL) scr_oom();
 #elif defined(__wasi__)
+  (void)awaiter;
   ScrFiber *spawner = scr_current;
   scr_current = f;
   scr_exc_swap_cell(&f->exc);
@@ -1672,9 +1738,21 @@ ScrPromise *scr_async_spawn(void (*entry)(ScrFiber *, void *), void *argpack) {
   ScrCtx here;
 #endif
 #ifndef __wasi__
-  f->return_to = &here;
+  ScrCtx *home = &here;
+  ScrPromise *awaited = NULL;
+#ifdef SCR_SPAWN_AWAIT
+  if (awaiter != NULL) {
+    f->awaiter = awaiter;
+    awaiter->await_parked = false;
+    home = &awaiter->ctx;
+    awaited = scr_promise_retain(f->promise);
+  }
+#else
+  (void)awaiter;
+#endif
+  f->return_to = home;
   ScrFiber *spawner = scr_current;
-  scr_switch(&here, &f->ctx, f);
+  scr_switch(home, &f->ctx, f);
   /* back: the fiber suspended or finished. The switch back targeted NULL
    * (the child doesn't know who spawned it), which pointed both the current
    * fiber AND the exception machinery at main — restore the spawner's cell
@@ -1684,7 +1762,12 @@ ScrPromise *scr_async_spawn(void (*entry)(ScrFiber *, void *), void *argpack) {
   scr_current = spawner;
   scr_exc_swap_cell(spawner ? &spawner->exc : NULL);
   scr_als_active = spawner ? &spawner->als : &scr_als_main_slot;
-  ScrPromise *result = scr_promise_retain(f->promise);
+  if (awaited != NULL) {
+    /* Parked by the child: the loop owns (and may have reaped) it. */
+    if (awaiter->await_parked) return awaited;
+    f->awaiter = NULL;
+  }
+  ScrPromise *result = awaited != NULL ? awaited : scr_promise_retain(f->promise);
   if (f->done) {
     scr_promise_release(f->promise);
     scr_fiber_destroy(f);
@@ -1694,6 +1777,81 @@ ScrPromise *scr_async_spawn(void (*entry)(ScrFiber *, void *), void *argpack) {
 #endif
 }
 
+ScrPromise *scr_async_spawn(void (*entry)(ScrFiber *, void *), void *argpack) {
+  return scr_async_spawn_impl(entry, argpack, NULL);
+}
+
+static void scr_await_park(ScrPromise *p);
+static void scr_promise_rethrow(ScrPromise *p);
+
+/* `await f(...)` of a fiber-backed async f in one call (the compiler's
+ * await-call entry): spawn, then await the promise, with the same turns
+ * as the two steps. When the child suspends in its eager prefix, the
+ * awaiter is already parked on the pending promise, so the settle wake
+ * resumes it with the result and no further hop; a child that finished
+ * eagerly leaves a settled promise, which the ordinary await hops on.
+ * Returns the promise +1; `ok` false means the rejection was re-thrown. */
+
+static ScrPromise *scr_async_spawn_await(void (*entry)(ScrFiber *, void *), void *argpack,
+                                         bool *ok) {
+#ifdef SCR_SPAWN_AWAIT
+  ScrFiber *self = scr_current;
+  if (self != NULL && self->gen == NULL && self->inline_depth == 0) {
+    ScrPromise *p = scr_async_spawn_impl(entry, argpack, self);
+    if (self->await_parked) {
+      self->await_parked = false;
+      while (p->state == SCR_PROM_PENDING) scr_await_park(p);
+      scr_prom_observe(p);
+      if (p->state == SCR_PROM_REJECTED) {
+        scr_promise_rethrow(p);
+        *ok = false;
+      } else {
+        *ok = true;
+      }
+      return p;
+    }
+    *ok = scr_await_settled(p);
+    return p;
+  }
+#endif
+  ScrPromise *p = scr_async_spawn(entry, argpack);
+  *ok = scr_await_settled(p);
+  return p;
+}
+
+void scr_async_spawn_await_void(void (*entry)(ScrFiber *, void *), void *argpack) {
+  bool ok;
+  scr_promise_release(scr_async_spawn_await(entry, argpack, &ok));
+}
+double scr_async_spawn_await_f64(void (*entry)(ScrFiber *, void *), void *argpack) {
+  bool ok;
+  ScrPromise *p = scr_async_spawn_await(entry, argpack, &ok);
+  double v = ok ? p->f64 : 0;
+  scr_promise_release(p);
+  return v;
+}
+bool scr_async_spawn_await_bool(void (*entry)(ScrFiber *, void *), void *argpack) {
+  bool ok;
+  ScrPromise *p = scr_async_spawn_await(entry, argpack, &ok);
+  bool v = ok ? p->b : false;
+  scr_promise_release(p);
+  return v;
+}
+ScrStr *scr_async_spawn_await_str(void (*entry)(ScrFiber *, void *), void *argpack) {
+  bool ok;
+  ScrPromise *p = scr_async_spawn_await(entry, argpack, &ok);
+  ScrStr *v = ok && p->payload ? scr_str_retain((ScrStr *)p->payload) : NULL;
+  scr_promise_release(p);
+  return v;
+}
+void *scr_async_spawn_await_ref(void (*entry)(ScrFiber *, void *), void *argpack) {
+  bool ok;
+  ScrPromise *p = scr_async_spawn_await(entry, argpack, &ok);
+  void *v = ok && p->payload ? p->retain_fn(p->payload) : NULL;
+  scr_promise_release(p);
+  return v;
+}
+
 /* ── fiberless async calls ─────────────────────────────────────────────
  * An async function whose body the compiler proved can never park its
  * execution context (no await, no other suspending runtime call, directly
@@ -1721,7 +1879,7 @@ void scr_async_inline_enter(ScrAsyncInline *frame) {
   if (scr_current != NULL) scr_current->inline_depth++;
 }
 
-ScrPromise *scr_async_inline_leave(ScrAsyncInline *frame) {
+void scr_async_inline_exit(ScrAsyncInline *frame) {
   if (scr_current != NULL) scr_current->inline_depth--;
   scr_exc_current_cell()->stack = (ScrStackFrame *)frame->stack;
   ScrAlsCtx *saved = (ScrAlsCtx *)frame->als;
@@ -1732,6 +1890,10 @@ ScrPromise *scr_async_inline_leave(ScrAsyncInline *frame) {
   } else {
     scr_als_ctx_release(saved);
   }
+}
+
+ScrPromise *scr_async_inline_leave(ScrAsyncInline *frame) {
+  scr_async_inline_exit(frame);
   return scr_promise_new();
 }
 
@@ -1747,7 +1909,7 @@ ScrPromise *scr_async_spawn_after(ScrPromise *dependency,
   (void)dependency;
   return scr_async_spawn(entry, argpack);
 #else
-  ScrFiber *f = calloc(1, sizeof *f);
+  ScrFiber *f = scr_mem_calloc(sizeof *f);
   if (!f) scr_oom();
   f->promise = scr_promise_new();
   f->entry = entry;
@@ -1756,13 +1918,7 @@ ScrPromise *scr_async_spawn_after(ScrPromise *dependency,
   scr_fibers_live++;
 
   if (dependency->state == SCR_PROM_PENDING) {
-    if (dependency->nwaiters == dependency->waiters_cap) {
-      dependency->waiters_cap = dependency->waiters_cap ? dependency->waiters_cap * 2 : 4;
-      dependency->waiters = realloc(
-          dependency->waiters, dependency->waiters_cap * sizeof *dependency->waiters);
-      if (!dependency->waiters) scr_oom();
-    }
-    dependency->waiters[dependency->nwaiters++] = f;
+    scr_promise_add_waiter(dependency, f);
   } else {
     scr_ready_push(f);
   }
@@ -1770,6 +1926,29 @@ ScrPromise *scr_async_spawn_after(ScrPromise *dependency,
 #endif
 }
 
+/* Where a suspending async fiber switches to. During the eager prefix
+ * under an await-call entry (awaiter set), control would return to the
+ * awaiter only for it to park on this fiber's still-pending promise and
+ * switch to its own resumer: park it here, for each such awaiter up the
+ * chain, and switch there directly. Nothing runs in between either way. */
+static ScrCtx *scr_fiber_park_awaiters(ScrFiber *self) {
+  ScrCtx *to = self->return_to;
+#ifdef SCR_SPAWN_AWAIT
+  for (ScrFiber *cur = self; cur->awaiter != NULL;) {
+    ScrFiber *aw = cur->awaiter;
+    cur->awaiter = NULL;
+    scr_promise_add_waiter(cur->promise, aw);
+    aw->await_parked = true;
+    to = aw->return_to;
+    cur = aw;
+  }
+  /* Whoever resumes this fiber next (the loop) becomes its return target;
+   * also right when the caller continues in place without switching. */
+  self->return_to = to;
+#endif
+  return to;
+}
+
 /* Parks the current fiber on `p` until it settles. Only fibers await. */
 static void scr_await_park(ScrPromise *p) {
   ScrFiber *self = scr_current;
@@ -1778,16 +1957,11 @@ static void scr_await_park(ScrPromise *p) {
     abort();
   }
   if (self->inline_depth != 0) scr_inline_park_trap();
-  if (p->nwaiters == p->waiters_cap) {
-    p->waiters_cap = p->waiters_cap ? p->waiters_cap * 2 : 4;
-    p->waiters = realloc(p->waiters, p->waiters_cap * sizeof *p->waiters);
-    if (!p->waiters) scr_oom();
-  }
-  p->waiters[p->nwaiters++] = self;
+  scr_promise_add_waiter(p, self);
 #ifdef SCR_WORKERS
   self->waiting = p;
 #endif
-  scr_switch(&self->ctx, self->return_to, NULL);
+  scr_switch(&self->ctx, scr_fiber_park_awaiters(self), NULL);
 #ifdef SCR_WORKERS
   /* Shutdown resumes suspended stacks without settling their dependencies.
    * Remove the borrowed waiter before the stack releases its promise. */
@@ -1809,12 +1983,7 @@ static void scr_await_park(ScrPromise *p) {
  * the continuation through their waiter list. */
 void scr_wasi_await_prepare(ScrFiber *self, ScrPromise *p) {
   if (p->state == SCR_PROM_PENDING) {
-    if (p->nwaiters == p->waiters_cap) {
-      p->waiters_cap = p->waiters_cap ? p->waiters_cap * 2 : 4;
-      p->waiters = realloc(p->waiters, p->waiters_cap * sizeof *p->waiters);
-      if (!p->waiters) scr_oom();
-    }
-    p->waiters[p->nwaiters++] = self;
+    scr_promise_add_waiter(p, self);
   } else {
     scr_ready_push(self);
   }
@@ -1851,8 +2020,22 @@ static void scr_await_yield(void) {
     abort();
   }
   if (self->inline_depth != 0) scr_inline_park_trap();
+  if (self->gen != NULL) {
+    scr_ready_push(self);
+    scr_switch(&self->ctx, self->return_to, NULL);
+    return;
+  }
+  /* The hop would return control to the loop's microtask drain with
+   * nothing else queued (this async fiber, or the awaiters parked for it,
+   * were resumed by the drain), so this fiber would be the very next job:
+   * the drain runs to exhaustion with no checkpoint in between. Continuing
+   * in place is the same turn order without the two switches. A fiber
+   * resumed by its spawner (the eager prefix) must still hop: the spawner
+   * has code to run first. */
+  ScrCtx *to = scr_fiber_park_awaiters(self);
+  if (to == &scr_loop_ctx && scr_ready_len == 0) return;
   scr_ready_push(self);
-  scr_switch(&self->ctx, self->return_to, NULL);
+  scr_switch(&self->ctx, to, NULL);
 }
 
 /* The emitted promise-or-absent await's unit arm (`await u` where u holds
@@ -2782,7 +2965,7 @@ ScrDyn *scr_set_immediate_dyn_value(void) {
  * main stack; a throw is an uncaught exception (Node's queueMicrotask),
  * never a rejection. */
 void scr_queue_microtask(ScrClosure *cb) {
-  ScrFiber *f = calloc(1, sizeof *f);
+  ScrFiber *f = scr_mem_calloc(sizeof *f);
   if (!f) scr_oom();
   f->micro_cb = cb; /* ownership moves in */
   scr_ready_push(f);
@@ -2840,7 +3023,7 @@ void scr_loop_set_events(bool (*pending)(void), bool (*watching)(void),
 static void scr_resume_fiber(ScrFiber *f) {
   if (f->gen_job != NULL) {
     ScrGen *g = f->gen_job;
-    free(f);
+    scr_mem_free(f);
     scr_async_gen_job(g);
     scr_gen_release(g);
     return;
@@ -2851,7 +3034,7 @@ static void scr_resume_fiber(ScrFiber *f) {
    * semantics (uncaughtException, never an unhandled rejection). */
   if (f->micro_cb != NULL) {
     ScrClosure *cb = f->micro_cb;
-    free(f);
+    scr_mem_free(f);
     ((void (*)(ScrClosure *))cb->fn)(cb);
     scr_closure_release(cb);
     return;
@@ -2884,6 +3067,9 @@ static void scr_resume_fiber(ScrFiber *f) {
   scr_als_active = &scr_als_main_slot;
 #else
   scr_switch(&scr_loop_ctx, &f->ctx, f);
+  f = scr_loop_reap;
+  if (f == NULL) return;
+  scr_loop_reap = NULL;
 #endif
   if (f->done) {
     scr_promise_release(f->promise);
@@ -3868,7 +4054,7 @@ static ScrGen *scr_gen_new_common(void (*entry)(ScrFiber *, void *), void *argpa
   g->settle_async = settle_async;
   scr_obj_alloc_note();
 
-  ScrFiber *f = calloc(1, sizeof *f);
+  ScrFiber *f = scr_mem_calloc(sizeof *f);
   if (!f) scr_oom();
   f->gen = g;
   scr_fiber_track(f);
@@ -4273,7 +4459,7 @@ static void scr_async_gen_start_queued(ScrGen *g);
 
 static void scr_async_gen_schedule_settle(ScrGen *g) {
   if (g->settle_queued) return;
-  ScrFiber *job = calloc(1, sizeof *job);
+  ScrFiber *job = scr_mem_calloc(sizeof *job);
   if (!job) scr_oom();
   job->gen_job = scr_gen_retain(g);
   g->settle_queued = true;
@@ -4485,12 +4671,12 @@ static void scr_context_ready_discard(void) {
     scr_ready_len--;
     if (fiber->micro_cb) {
       scr_closure_release(fiber->micro_cb);
-      free(fiber);
+      scr_mem_free(fiber);
     } else if (fiber->gen_job) {
       ScrGen *generator = fiber->gen_job;
       generator->settle_queued = false;
       scr_gen_release(generator);
-      free(fiber);
+      scr_mem_free(fiber);
     }
     /* Real fibers belong to the context registry, including queued ones. */
   }
diff --git a/packages/runtime/src/scr_runtime.h b/packages/runtime/src/scr_runtime.h
index 176c2be65..695dd2c5f 100644
--- a/packages/runtime/src/scr_runtime.h
+++ b/packages/runtime/src/scr_runtime.h
@@ -5351,6 +5351,13 @@ void scr_promise_release_v(void *p);
 typedef struct ScrFiber ScrFiber;
 /* Spawn + run eagerly to the first suspension; returns the promise, +1. */
 ScrPromise *scr_async_spawn(void (*entry)(ScrFiber *, void *), void *argpack);
+/* `await f(...)` of a fiber-backed async f: spawn and await in one call,
+ * returning the awaited value (a rejection re-throws into the caller). */
+void scr_async_spawn_await_void(void (*entry)(ScrFiber *, void *), void *argpack);
+double scr_async_spawn_await_f64(void (*entry)(ScrFiber *, void *), void *argpack);
+bool scr_async_spawn_await_bool(void (*entry)(ScrFiber *, void *), void *argpack);
+ScrStr *scr_async_spawn_await_str(void (*entry)(ScrFiber *, void *), void *argpack);
+void *scr_async_spawn_await_ref(void (*entry)(ScrFiber *, void *), void *argpack);
 /* Runtime-authored C waiters cannot themselves become LLVM switched
  * coroutines on wasm32. Spawn after `dependency` settles so their first
  * (and only) await is extraction from an already-settled promise, while
@@ -5370,6 +5377,10 @@ typedef struct ScrAsyncInline {
 } ScrAsyncInline;
 void scr_async_inline_enter(ScrAsyncInline *frame);
 ScrPromise *scr_async_inline_leave(ScrAsyncInline *frame);
+/* Leaves the frame without creating the call's promise: `await f()` of a
+ * fiberless f consumes the result directly (the emitted await-call entry
+ * still takes the await's microtask turn). */
+void scr_async_inline_exit(ScrAsyncInline *frame);
 
 /* wasm32-wasi lowers async functions and generators through LLVM switched
  * coroutines instead of the native stack-switching implementations. The
diff --git a/tests/corpus/4880-await-call-turn-order.ts b/tests/corpus/4880-await-call-turn-order.ts
new file mode 100644
index 000000000..e183c4596
--- /dev/null
+++ b/tests/corpus/4880-await-call-turn-order.ts
@@ -0,0 +1,88 @@
+// Directly awaited async calls (bodies that never suspend, bodies that do,
+// and chains of them) keep Node's microtask order against then-callbacks,
+// queueMicrotask, and settled and pending promises.
+const log: string[] = [];
+const cache = new Map<number, number>();
+
+async function hit(key: number): Promise<number> {
+  const v = cache.get(key);
+  if (v !== undefined) return v;
+  cache.set(key, key * 10);
+  return key * 10;
+}
+async function flag(n: number): Promise<boolean> {
+  return n % 2 === 0;
+}
+async function label(n: number): Promise<string> {
+  return "L" + n;
+}
+async function nothing(n: number): Promise<void> {
+  log.push("nothing " + n);
+}
+async function suspends(n: number): Promise<number> {
+  log.push("suspends enter " + n);
+  const a = await hit(n);
+  log.push("suspends mid " + n);
+  const b = await hit(n + 1);
+  return a + b;
+}
+async function chain(n: number): Promise<number> {
+  log.push("chain enter " + n);
+  const v = await suspends(n);
+  log.push("chain got " + n + " " + v);
+  return v + 1;
+}
+async function deep(n: number): Promise<number> {
+  return (await chain(n)) * 2;
+}
+
+async function worker(name: string, n: number): Promise<number> {
+  log.push(name + " start");
+  const a = await hit(n);
+  log.push(name + " a=" + a);
+  const f = await flag(n);
+  log.push(name + " flag=" + f);
+  const s = await label(n);
+  log.push(name + " label=" + s);
+  await nothing(n);
+  log.push(name + " after nothing");
+  const d = await deep(n);
+  log.push(name + " deep=" + d);
+  return a + d;
+}
+
+async function main(): Promise<void> {
+  const settled = Promise.resolve(7);
+  const pending = new Promise<number>((resolve) => {
+    setTimeout(() => {
+      log.push("timer resolves");
+      resolve(9);
+    }, 0);
+  });
+  settled.then((v) => log.push("then settled " + v));
+  queueMicrotask(() => log.push("microtask 1"));
+  const w1 = worker("w1", 1);
+  const w2 = worker("w2", 2);
+  queueMicrotask(() => log.push("microtask 2"));
+  Promise.resolve()
+    .then(() => log.push("then A1"))
+    .then(() => log.push("then A2"))
+    .then(() => log.push("then A3"))
+    .then(() => log.push("then A4"));
+  log.push("main sync end");
+  const x = await settled;
+  log.push("main settled " + x);
+  const t = await (async (): Promise<number> => await hit(5))();
+  log.push("main arrow " + t);
+  const both: number[] = await Promise.all([w1, w2]);
+  log.push("workers " + both[0] + " " + both[1]);
+  const p = await pending;
+  log.push("main pending " + p);
+  const r = await deep(3);
+  log.push("main deep " + r);
+}
+
+main().then(() => {
+  for (const line of log) console.log(line);
+});
+console.log("top-level sync end");
diff --git a/tests/corpus/4881-await-call-exceptions.ts b/tests/corpus/4881-await-call-exceptions.ts
new file mode 100644
index 000000000..52574b97f
--- /dev/null
+++ b/tests/corpus/4881-await-call-exceptions.ts
@@ -0,0 +1,82 @@
+// Exceptions thrown by directly awaited async calls, before and after the
+// callee's first await, reach the awaiter's catch on Node's turn, and
+// finally blocks and rethrows keep their order against other microtasks.
+const log: string[] = [];
+
+async function failsNow(n: number): Promise<number> {
+  if (n > 0) throw new Error("now " + n);
+  return n;
+}
+async function failsString(n: number): Promise<string> {
+  if (n > 0) throw "plain " + n;
+  return "ok";
+}
+async function failsLater(n: number): Promise<number> {
+  log.push("later enter " + n);
+  await failsNow(0);
+  log.push("later resumed " + n);
+  throw new Error("later " + n);
+}
+async function failsVoid(): Promise<void> {
+  throw new RangeError("void failure");
+}
+async function rethrows(n: number): Promise<number> {
+  try {
+    return await failsLater(n);
+  } catch (e) {
+    log.push("rethrows caught " + (e as Error).message);
+    throw new Error("wrapped " + (e as Error).message);
+  } finally {
+    log.push("rethrows finally " + n);
+  }
+}
+
+async function probe(name: string): Promise<void> {
+  try {
+    await failsNow(1);
+  } catch (e) {
+    log.push(name + " caught " + (e as Error).message);
+  }
+  try {
+    await failsString(2);
+  } catch (e) {
+    log.push(name + " caught " + String(e));
+  }
+  try {
+    await failsVoid();
+  } catch (e) {
+    log.push(name + " caught " + (e as Error).name + " " + (e as Error).message);
+  }
+  try {
+    await rethrows(3);
+  } catch (e) {
+    log.push(name + " caught " + (e as Error).message);
+  }
+  const ok = await failsNow(0);
+  log.push(name + " ok " + ok);
+}
+
+async function main(): Promise<void> {
+  queueMicrotask(() => log.push("microtask before"));
+  const a = probe("a");
+  const b = probe("b");
+  Promise.resolve()
+    .then(() => log.push("tick 1"))
+    .then(() => log.push("tick 2"))
+    .then(() => log.push("tick 3"));
+  await a;
+  await b;
+  let caught = "";
+  try {
+    await (async (): Promise<number> => {
+      throw new Error("inline arrow");
+    })();
+  } catch (e) {
+    caught = (e as Error).message;
+  }
+  log.push("arrow " + caught);
+}
+
+main().then(() => {
+  for (const line of log) console.log(line);
+});
diff --git a/tests/corpus/4882-await-call-request-chain.ts b/tests/corpus/4882-await-call-request-chain.ts
new file mode 100644
index 000000000..4148c0935
--- /dev/null
+++ b/tests/corpus/4882-await-call-request-chain.ts
@@ -0,0 +1,53 @@
+// A request handler chain of small async helpers that usually complete
+// without suspending (cache hits), awaited directly, with concurrent
+// requests interleaving on the microtask queue.
+interface Session {
+  user: string;
+  role: string;
+}
+const sessions = new Map<string, Session>();
+const quotas = new Map<string, number>();
+const order: string[] = [];
+
+async function lookupSession(token: string): Promise<Session | undefined> {
+  return sessions.get(token);
+}
+async function loadSession(token: string): Promise<Session> {
+  const hit = await lookupSession(token);
+  if (hit !== undefined) return hit;
+  await new Promise<void>((resolve) => setTimeout(resolve, 0));
+  const fresh = { user: "u-" + token, role: token.length % 2 === 0 ? "admin" : "user" };
+  sessions.set(token, fresh);
+  return fresh;
+}
+async function canWrite(session: Session): Promise<boolean> {
+  return session.role === "admin";
+}
+async function quota(user: string): Promise<number> {
+  const q = quotas.get(user) ?? 3;
+  quotas.set(user, q - 1);
+  return q;
+}
+async function render(status: number, body: string): Promise<string> {
+  return status + " " + body;
+}
+async function handle(id: number, token: string): Promise<string> {
+  order.push("start " + id);
+  const session = await loadSession(token);
+  if (!(await canWrite(session))) return await render(403, session.user);
+  const left = await quota(session.user);
+  if (left <= 0) return await render(429, session.user);
+  order.push("write " + id);
+  return await render(200, session.user + " left " + left);
+}
+
+async function main(): Promise<void> {
+  const tokens = ["ab", "abc", "ab", "abcd", "ab", "abc", "abcd", "ab", "ab", "ab"];
+  const first = await Promise.all(tokens.map((t, i) => handle(i, t)));
+  for (const r of first) console.log(r);
+  const second: string[] = [];
+  for (let i = 0; i < tokens.length; i++) second.push(await handle(100 + i, tokens[i]!));
+  for (const r of second) console.log(r);
+  console.log(order.join(","));
+}
+void main();
diff --git a/tests/corpus/4883-await-values-and-hops.ts b/tests/corpus/4883-await-values-and-hops.ts
new file mode 100644
index 000000000..4d12ccbd5
--- /dev/null
+++ b/tests/corpus/4883-await-values-and-hops.ts
@@ -0,0 +1,32 @@
+// await of non-promise values and of directly called async functions each
+// take exactly one microtask turn, interleaved with two counters.
+const log: string[] = [];
+async function one(): Promise<number> {
+  return 1;
+}
+async function ticker(name: string, n: number): Promise<void> {
+  for (let i = 0; i < n; i++) {
+    log.push(name + i);
+    await null;
+  }
+}
+async function mixed(): Promise<void> {
+  log.push("m0");
+  const a = await 5;
+  log.push("m1 " + a);
+  const b = await one();
+  log.push("m2 " + b);
+  const c = await "text";
+  log.push("m3 " + c);
+  const d = await undefined;
+  log.push("m4 " + d);
+  const e = await one();
+  log.push("m5 " + e);
+  const f = await Promise.resolve(await one());
+  log.push("m6 " + f);
+}
+async function main(): Promise<void> {
+  await Promise.all([ticker("t", 10), mixed(), ticker("s", 4)]);
+  console.log(log.join(" "));
+}
+void main();
diff --git a/tests/corpus/4884-await-call-unhandled-sibling.ts b/tests/corpus/4884-await-call-unhandled-sibling.ts
new file mode 100644
index 000000000..330d3ce24
--- /dev/null
+++ b/tests/corpus/4884-await-call-unhandled-sibling.ts
@@ -0,0 +1,22 @@
+// @exit: 1
+// A rejection from a directly awaited async call is handled by that await,
+// while the same function's rejected promise left unawaited still reports
+// as an unhandled rejection after the microtask queue drains.
+async function fails(n: number): Promise<number> {
+  if (n > 0) throw new Error("fail " + n);
+  return n;
+}
+async function main(): Promise<void> {
+  try {
+    await fails(1);
+  } catch (e) {
+    console.log("handled", (e as Error).message);
+  }
+  process.on("unhandledRejection", (reason) => {
+    console.log("unhandled", (reason as Error).message);
+    process.exitCode = 1;
+  });
+  void fails(2);
+  console.log("after", await fails(0));
+}
+void main();
diff --git a/tests/corpus/4885-await-call-checkpoints.ts b/tests/corpus/4885-await-call-checkpoints.ts
new file mode 100644
index 000000000..269dfcd91
--- /dev/null
+++ b/tests/corpus/4885-await-call-checkpoints.ts
@@ -0,0 +1,40 @@
+// Directly awaited async calls against process.nextTick, setImmediate,
+// a timer and queueMicrotask: ticks run before promise jobs at each
+// checkpoint, and every await still waits for the jobs queued before it.
+const log: string[] = [];
+async function quick(n: number): Promise<number> {
+  return n * 3;
+}
+async function once(n: number): Promise<number> {
+  const v = await quick(n);
+  process.nextTick(() => log.push("tick from once " + n));
+  return v + 1;
+}
+async function runner(name: string): Promise<void> {
+  log.push(name + " begin");
+  for (let i = 0; i < 3; i++) {
+    const v = await once(i);
+    log.push(name + " step " + i + " = " + v);
+    if (i === 1) {
+      queueMicrotask(() => log.push(name + " microtask"));
+      process.nextTick(() => log.push(name + " tick"));
+    }
+  }
+  await new Promise<void>((resolve) => setImmediate(resolve));
+  log.push(name + " after immediate " + (await quick(7)));
+  await new Promise<void>((resolve) => setTimeout(resolve, 1));
+  log.push(name + " after timer " + (await once(9)));
+}
+async function main(): Promise<void> {
+  process.nextTick(() => log.push("first tick"));
+  setImmediate(() => log.push("immediate"));
+  const a = runner("a");
+  const b = runner("b");
+  log.push("main sync");
+  await Promise.all([a, b]);
+  const p = quick(1);
+  const q = Promise.resolve(p);
+  log.push("same promise " + (p === q));
+  log.push("value " + (await q));
+}
+main().then(() => console.log(log.join("\n")));