diff --git a/src/interp.ts b/src/interp.ts
index d285417..5cd3dae 100644
--- a/src/interp.ts
+++ b/src/interp.ts
@@ -4,11 +4,11 @@ import {
   INITIALIZED, JArray, JClass, JMethod, JObject, isWide, longToFloat, newArrayData,
 } from "./runtime.ts";
 import {
-  Frame, JThread, JavaThrow, NativeFrame, PendingThrow, RUNNABLE, TERMINATED, WAITING, fromNative,
+  Frame, JThread, JavaThrow, MAX_DIRECT, NativeFrame, PendingThrow, RUNNABLE, TERMINATED, WAITING, fromNative,
 } from "./thread.ts";
 import type { NativeGen } from "./thread.ts";
 import type { JVM } from "./jvm.ts";
-import { compileIfHot, hooks, noteBackedge } from "./jit.ts";
+import { INTERPRET, PUSHED, compileIfHot, hooks, noteBackedge } from "./jit.ts";
 
 const MAX_FRAMES = 6000;
 
@@ -86,6 +86,7 @@ Object.assign(hooks, {
   popFrame: (t: JThread, f: Frame) => popFrame(t, f),
   findHandler: findHandlerAt,
   checkStack: (t: JThread) => checkStack(t),
+  callNested,
 });
 
 export function interpret(t: JThread, f: Frame, budget: number): number {
@@ -290,16 +291,19 @@ export function interpret(t: JThread, f: Frame, budget: number): number {
         case 0xac: case 0xae: case 0xb0:
           v = stack[sp - 1];
           popFrame(t, f);
-          deliver(t, v, m.ret, false);
+          if (f.direct) t.ret = v;
+          else deliver(t, v, m.ret, false);
           return budget;
         case 0xad: case 0xaf:
           v = stack[sp - 2];
           popFrame(t, f);
-          deliver(t, v, m.ret, false);
+          if (f.direct) t.ret = v;
+          else deliver(t, v, m.ret, false);
           return budget;
         case 0xb1:
           popFrame(t, f);
-          deliver(t, undefined, "V", false);
+          if (f.direct) t.ret = undefined;
+          else deliver(t, undefined, "V", false);
           return budget;
         // fields
         case 0xb2: case 0xb3: {
@@ -633,6 +637,8 @@ export function raise(t: JThread, exc: JObject) {
       return;
     }
     popFrame(t, f);
+    // the rest of the stack belongs to the compiled code that called f
+    if (f.direct) throw new JavaThrow(exc);
   }
 }
 
@@ -697,6 +703,47 @@ export function handleError(t: JThread, e: unknown) {
   else t.vm.internalError(t, e);
 }
 
+// Interprets a call from compiled code nested on the JS stack, as a direct
+// call runs, so the caller need not spill. Returns m's result, PUSHED when
+// the call must finish in the run loop (its frames stay and the caller
+// spills), or INTERPRET when the run loop must make the call.
+function callNested(t: JThread, m: JMethod, args: any[]): any {
+  if (m.isSynchronized || t.frames.length - t.base >= MAX_DIRECT || t.frames.length >= t.stackLimit) return INTERPRET;
+  if (m.jit === null && !m.noJit) compileIfHot(t, m);
+  const f = new Frame(m);
+  const locals = f.locals;
+  let slot = 0, i = 0;
+  if (!m.isStatic) locals[slot++] = args[i++];
+  for (const p of m.params) {
+    locals[slot] = args[i++];
+    slot += isWide(p) ? 2 : 1;
+  }
+  f.direct = true;
+  t.frames.push(f);
+  const depth = t.frames.length;
+  for (;;) {
+    if (t.frames.length < depth) return t.ret;
+    const top = t.top;
+    if (!f.direct || t.slice <= 0 || t.state !== RUNNABLE || top instanceof NativeFrame || top.needLock) {
+      f.direct = false;
+      return PUSHED;
+    }
+    const jit = top.m.jit;
+    if (jit === null) t.slice = interpret(t, top, t.slice);
+    else {
+      const n = t.frames.length;
+      try {
+        jit(t, top);
+      } catch (e) {
+        if (t.frames.length < depth) throw e;
+        handleError(t, e);
+      }
+      // a call that returned is cheap; one left part done used up its own budget
+      t.slice -= t.frames[n - 1] === top ? 4000 : 1;
+    }
+  }
+}
+
 // Feeds a native frame its pending input and acts on what it yields.
 function stepNative(t: JThread, f: NativeFrame) {
   let r: IteratorResult<any, any>;
@@ -759,7 +806,7 @@ function stepNative(t: JThread, f: NativeFrame) {
 
 const entryStats = typeof process !== "undefined" && process.env?.JW_JIT_STATS ? new Map<string, number>() : null;
 if (entryStats) process.on("exit", () => {
-  console.error("-- run loop entries into compiled code");
+  console.error("-- run loop entries into compiled code", [...entryStats.values()].reduce((a, b) => a + b, 0));
   for (const [k, v] of [...entryStats].sort((a, b) => b[1] - a[1]).slice(0, 25)) console.error(String(v).padStart(8), k);
 });
 
@@ -794,12 +841,13 @@ export function runThread(t: JThread, budget: number) {
     if (jit !== null) {
       if (entryStats) countEntry(t, f);
       t.base = t.frames.length;
+      t.slice = budget;
       try {
         jit(t, f);
       } catch (e) {
         handleError(t, e);
       }
-      budget -= 4000;
+      budget = t.slice - 4000;
     } else budget = interpret(t, f, budget);
   }
   if (t.frames.length === 0 && t.state !== TERMINATED) terminate(t);
diff --git a/src/jit.ts b/src/jit.ts
index 769c29c..4c4dbb0 100644
--- a/src/jit.ts
+++ b/src/jit.ts
@@ -48,6 +48,7 @@ export const hooks = {
   popFrame: null as unknown as (t: JThread, f: Frame) => void,
   findHandler: null as unknown as (t: JThread, f: Frame, pc: number, exc: JObject) => number,
   checkStack: null as unknown as (t: JThread) => void,
+  callNested: null as unknown as (t: JThread, m: JMethod, args: any[]) => any,
 };
 
 let disabled = typeof process !== "undefined" && !!process.env?.JW_NOJIT;
@@ -129,7 +130,7 @@ function at(e: PendingThrow, pc: number | undefined): PendingThrow {
 }
 
 // results of runtime.slowCall
-const PUSHED = Object.freeze({ pushed: true }), INTERPRET = Object.freeze({ interpret: true });
+export const PUSHED = Object.freeze({ pushed: true }), INTERPRET = Object.freeze({ interpret: true });
 
 const STRING_BUILDER = "java/lang/StringBuilder";
 
@@ -189,10 +190,12 @@ const runtime = {
     f.sp = sp;
     return f;
   },
-  // A call that is not direct: a native's result, PUSHED once a native that
-  // must wait is on the stack, or INTERPRET for the run loop to make the call.
+  // A call that is not direct: the result of a native or an interpreted
+  // method, PUSHED once the callee's frames must finish in the run loop, or
+  // INTERPRET for the run loop to make the call.
   slowCall(t: JThread, m: JMethod, args: any[]): any {
-    if (m.native === null || m.isAbstract) return INTERPRET;
+    if (m.isAbstract) return INTERPRET;
+    if (m.native === null) return m.code ? hooks.callNested(t, m, args) : INTERPRET;
     let r;
     try {
       r = m.native(t, args);
@@ -580,7 +583,7 @@ function generate(t: JThread, m: JMethod): { src: string; consts: any[] } {
   const out: string[] = [];
   // direct calls pass arguments as parameters and a frame without arrays
   out.push(`return function ${fname}(t,f${params.map((p) => "," + p).join("")}){let L=f.locals,S=f.stack;const vm=t.vm,H=rt.hooks;`);
-  out.push(`let pc=f.pc,bud=4000,tm,cf,r,a,v,ic,ifs=0,dpc=0,dsp=0,x=0,xs=0,xe=null,${[...g.vars].join(",")};`);
+  out.push(`let pc=f.pc,tm,cf,r,a,v,ic,ifs=0,dpc=0,dsp=0,x=0,xs=0,xe=null,${[...g.vars].join(",")};`);
   out.push(`if(L===null){${params.map((p, i) => `${L(i)}=${p};`).join("")}}else{${Array.from({ length: nl }, (_, i) => `${L(i)}=L[${i}];`).join("")}${Array.from({ length: ns }, (_, i) => `${S(i)}=S[${i}];`).join("")}}`);
   out.push(`X:{for(;;){try{switch(pc){`);
   out.push(...top.out);
@@ -688,7 +691,9 @@ function insn(g: Gen, b: Body, pc: number) {
   const top = (k = 0) => S(d - 1 - k);
   const jump = (target: number) => {
     if (region) return `{${region.pcVar}=${target};continue R${region.id};}`;
-    if (target <= pc) return `{if(--bud===0)${g.exit(target, b.a.depth.get(target)!, EXIT_SUSPEND)}pc=${target};continue;}`;
+    // loops spend the thread's slice wherever they are on the stack, so a loop
+    // calling others' loops still yields
+    if (target <= pc) return `{if(--t.slice<=0)${g.exit(target, b.a.depth.get(target)!, EXIT_SUSPEND)}pc=${target};continue;}`;
     return `{pc=${target};continue;}`;
   };
   const idx2 = u16(c, pc + 1);
diff --git a/src/thread.ts b/src/thread.ts
index 80227d3..59e1cee 100644
--- a/src/thread.ts
+++ b/src/thread.ts
@@ -102,7 +102,7 @@ export const STATUS_SLEEPING = 0x00e1, STATUS_WAITING = 0x0191, STATUS_TIMED_WAI
 export const STATUS_PARKED = 0x0291, STATUS_PARKED_TIMED = 0x02a1, STATUS_BLOCKED = 0x0401;
 
 // compiled frames nested on one JS stack before calls go through the run loop
-const MAX_DIRECT = 120;
+export const MAX_DIRECT = 120;
 
 let nextId = 1;
 
@@ -129,6 +129,9 @@ export class JThread {
   // run loop last entered compiled code
   ret: any = undefined;
   base = 0;
+  // what is left of the run loop's budget while compiled code runs: its
+  // loops and the calls it interprets in place spend it
+  slice = 0;
   // wakes the thread early from sleep, park or wait, for interrupt/unpark
   wake: (() => void) | null = null;
   // set when the bottom frame finishes
diff --git a/test/java/NestedCalls.java b/test/java/NestedCalls.java
new file mode 100644
index 0000000..e105f6d
--- /dev/null
+++ b/test/java/NestedCalls.java
@@ -0,0 +1,140 @@
+import java.util.Arrays;
+
+// Compiled code calling methods still too cold to compile: results of each
+// type, exceptions caught on either side with their stack traces, monitors,
+// recursion, stack overflow and a callee that waits on another thread. Each
+// callee calls pad, so it is too big to be inlined.
+public class NestedCalls {
+  static int counter, noise;
+  static final Object LOCK = new Object();
+
+  static void pad(int x) {
+    noise += x * 3 + (x >>> 2) - (x ^ 5) + (x | 9) + (x & 0x77) + x % 13 + x / 7 + (x << 3) - (x >> 1);
+  }
+
+  static int i(int x) { pad(x); pad(x + 1); return x + 1; }
+  static long j(long x) { pad((int) x); pad(1); return x * 3; }
+  static double d(double x) { pad((int) x); pad(2); return x / 2; }
+  static boolean z(int x) { pad(x); pad(3); return (x & 1) == 0; }
+  static Object a(int x) { pad(x); pad(4); return x % 3 == 0 ? null : new int[] {x}; }
+  static void v() { pad(5); pad(6); counter++; }
+
+  static int throwsOut(int x) {
+    pad(x);
+    pad(7);
+    if (x % 7 == 0) throw new IllegalStateException("x" + x);
+    return x;
+  }
+
+  static int catchesInside(int x) {
+    try {
+      return throwsOut(x);
+    } catch (IllegalStateException e) {
+      return -1;
+    }
+  }
+
+  static int npe(int[] a) { pad(8); pad(9); return a[0]; }
+
+  // cold, called by compiled hotCaller, which is called by cold outerCatch:
+  // its exception passes compiled code to reach outerCatch's handler
+  static int coldThrow(int x) {
+    pad(x);
+    pad(10);
+    if (x < 0) throw new IllegalArgumentException("neg " + x);
+    return x;
+  }
+
+  static int hotCaller(int x, boolean call) {
+    pad(x);
+    pad(11);
+    return call ? coldThrow(x) : x;
+  }
+
+  static int outerCatch(int x) {
+    pad(x);
+    try {
+      return hotCaller(x, true);
+    } catch (IllegalArgumentException e) {
+      return e.getMessage().length() - 100;
+    }
+  }
+
+  static int locked(int x) {
+    synchronized (LOCK) {
+      return x * 2;
+    }
+  }
+
+  static synchronized int syncMethod(int x) { return x - 1; }
+
+  static int depth(int n) { return n == 0 ? 0 : 1 + depth(n - 1); }
+
+  static int forever(int n) { return forever(n + 1) + 1; }
+
+  static long sum(int n) {
+    long s = 0;
+    for (int k = 0; k < n; k++) s += i(k) + depth(k & 15);
+    return s;
+  }
+
+  static int joined(int x) throws InterruptedException {
+    Thread t = new Thread(() -> counter += 1000);
+    t.start();
+    t.join();
+    return x;
+  }
+
+  static String where(Throwable e) {
+    StackTraceElement[] st = e.getStackTrace();
+    return st[0].getMethodName() + ":" + st[0].getLineNumber() + " < " + st[1].getMethodName() + ":" + st[1].getLineNumber();
+  }
+
+  // compiled by its loop after the first pass resolves the calls and before
+  // the others start
+  static void run() throws InterruptedException {
+    long acc = 0;
+    double dacc = 0;
+    int thrown = 0, inside = 0, nulls = 0;
+    String firstThrow = null, firstNpe = null;
+    for (int k = 0; k < 3000; k++) {
+      acc += k & 3;
+      if (k != 0 && (k < 2000 || k >= 2300)) continue;
+      acc += i(k) + j(k) + (z(k) ? 1 : 0) + locked(k) + syncMethod(k) + depth(k & 31);
+      dacc += d(k);
+      v();
+      Object o = a(k);
+      if (o == null) nulls++;
+      try {
+        acc += throwsOut(k);
+      } catch (IllegalStateException e) {
+        thrown++;
+        if (firstThrow == null) firstThrow = e.getMessage() + " at " + where(e);
+      }
+      if (catchesInside(k) < 0) inside++;
+      try {
+        acc += npe(k % 50 == 0 ? null : new int[] {k});
+      } catch (NullPointerException e) {
+        if (firstNpe == null) firstNpe = where(e);
+      }
+      if (k % 100 == 0) acc += joined(k);
+      if (k % 10 == 0) acc += outerCatch(k % 20 == 0 ? -k : k);
+    }
+    System.out.println(acc + " " + dacc + " " + counter + " " + thrown + " " + inside + " " + nulls + " " + noise);
+    System.out.println(firstThrow);
+    System.out.println(firstNpe);
+    System.out.println(sum(5000));
+    try {
+      forever(0);
+    } catch (StackOverflowError e) {
+      System.out.println("StackOverflowError");
+    }
+    System.out.println(Arrays.toString((int[]) a(4)));
+  }
+
+  public static void main(String[] args) throws InterruptedException {
+    hotCaller(1, true);
+    for (int k = 0; k < 600; k++) hotCaller(k, false);
+    run();
+  }
+}